wrote:\\r\\n> The main motivation is provide a way to archive without shelling out.\\r\\n> This reduces the amount of overhead, which can improve archival rate\\r\\n> significantly. It should also make it easier to archive more safely.\\r\\n> For example, many of the common shell commands used for archiving\\r\\n> won't fsync the data, but it isn't too hard to do so via C. The\\r\\n> current proposal doesn't introduce any extra infrastructure for\\r\\n> batching or parallelism, but it is probably still possible. I would\\r\\n> like to eventually add batching, but for now I'm only focused on\\r\\n> introducing basic archive module support.\\r\\n\\r\\nAs noted above, the latest patch set (v11) doesn't add any batching or\\r\\nparallelism. Now that beb4e9b is committed (which causes the archiver\\r\\nto gather multiple files to archive in each scan of archive_status),\\r\\nit seems like a good time to discuss this a bit further. I think\\r\\nthere are some interesting design considerations.\\r\\n\\r\\nAs is, the current archive module infrastructure in the v11 patch set\\r\\nshould help reduce the amount of overhead per-file quite a bit, and I\\r\\nobserved a noticeable speedup with a basic file-copying archive\\r\\nstrategy (although this is likely not representative of real-world\\r\\nworkloads). I believe it would be possible for archive module authors\\r\\nto implement batching/parallelism, but AFAICT it would still require\\r\\nhacks similar to what folks do today with archive_command. For\\r\\nexample, you could look ahead in archive_status, archive a bunch of\\r\\nfiles in a batch or in parallel with background workers, and then\\r\\nquickly return true when the archive_library is called for later files\\r\\nin the batch.\\r\\n\\r\\nAlternatively, we could offer some kind of built-in batching support\\r\\nin the archive module infrastructure. One simple approach would be to\\r\\njust have pgarch_readyXlog() optionally return the entire list of\\r\\nfiles gathered from the directory scan of archive_status (presently up\\r\\nto 64 files). Or we could provide a GUC like archive_batch_size that\\r\\nwould allow users to limit how many files are sent to the\\r\\narchive_library each time. This list would be given to\\r\\npgarch_archiveXlog(), which would return which files were successfully\\r\\narchived and which failed. I think this could be done for\\r\\narchive_command as well, although it might be tricky to determine\\r\\nwhich files were archived successfully. To handle that, we might just\\r\\nneed to fail the whole batch if the archive_command return value\\r\\nindicates failure.\\r\\n\\r\\nAnother interesting change is that the special timeline file handling\\r\\nadded in beb4e9b becomes less useful. Presently, if a timeline\\r\\nhistory file is marked ready for archival, we force pgarch_readyXlog()\\r\\nto do a new scan of archive_status the next time it is called in order\\r\\nto pick it up as soon as possible (ordinarily it just returns the\\r\\nfiles gathered in a previous scan until it runs out). If we are\\r\\nsending a list of files to the archive module, it will be more\\r\\ndifficult to ensure timeline history files are picked up so quickly.\\r\\nPerhaps this is a reasonable tradeoff to make when archive batching is\\r\\nenabled.\\r\\n\\r\\nI think the retry logic can stay roughly the same. If any files in a\\r\\nbatch cannot be archived, wait a second before retrying. If that\\r\\nhappens a few times in a row, stop archiving for a bit. It wouldn't\\r\\nbe quite as precise as what's there today because the failures could\\r\\nbe for different files each time, but I don't know if that is terribly\\r\\nimportant.\\r\\n\\r\\nFinally, I wonder if batching support is something we should bother\\r\\nwith at all for the first round of archive module support. I believe\\r\\nit is something that could be easily added later, although it might\\r\\nrequire archive modules to adjust the archiving callback to accept and\\r\\nreturn a list of files. IMO the archive modules infrastructure is\\r\\nstill an improvement even without batching, and it seems to fit nicely\\r\\ninto the existing behavior of the archiver process. I'm curious what\\r\\nothers think about all this.\\r\\n\\r\\nNathan\\r\\n\\r\\n\",\n \"msg_date\": \"Wed, 15 Dec 2021 18:32:02 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 11/22/21, 10:01 AM, \\\"Bossart, Nathan\\\" <bossartn@amazon.com> wrote:\\r\\n> On 11/19/21, 11:24 AM, \\\"Bossart, Nathan\\\" <bossartn@amazon.com> wrote:\\r\\n>> I went ahead and split the patch into 4 separate patches in an effort\\r\\n>> to ease review. 0001 just refactors the shell archiving logic to its\\r\\n>> own file. 0002 introduces the archive modules infrastructure. 0003\\r\\n>> introduces the basic_archive test module. And 0004 is the docs.\\r\\n>\\r\\n> Here is a rebased patch set (1b06d7b broke v10).\\r\\n\\r\\nI'm attempting to make cfbot happy again with v12. It looked like\\r\\nthere was a missing #include for Windows.\\r\\n\\r\\nNathan\",\n \"msg_date\": \"Wed, 5 Jan 2022 23:10:50 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 1/5/22, 3:14 PM, \\\"Bossart, Nathan\\\" <bossartn@amazon.com> wrote:\\r\\n> On 11/22/21, 10:01 AM, \\\"Bossart, Nathan\\\" <bossartn@amazon.com> wrote:\\r\\n>> On 11/19/21, 11:24 AM, \\\"Bossart, Nathan\\\" <bossartn@amazon.com> wrote:\\r\\n>>> I went ahead and split the patch into 4 separate patches in an effort\\r\\n>>> to ease review. 0001 just refactors the shell archiving logic to its\\r\\n>>> own file. 0002 introduces the archive modules infrastructure. 0003\\r\\n>>> introduces the basic_archive test module. And 0004 is the docs.\\r\\n>>\\r\\n>> Here is a rebased patch set (1b06d7b broke v10).\\r\\n>\\r\\n> I'm attempting to make cfbot happy again with v12. It looked like\\r\\n> there was a missing #include for Windows.\\r\\n\\r\\nHere is another rebase for cfbot.\\r\\n\\r\\nNathan\",\n \"msg_date\": \"Thu, 13 Jan 2022 19:38:43 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 13, 2022 at 2:38 PM Bossart, Nathan <bossartn@amazon.com> wrote:\\n> Here is another rebase for cfbot.\\n\\nI've committed 0001 now. I don't see anything particularly wrong with\\nthe rest of this either, but here are a few comments:\\n\\n- I wonder whether it might be better to promote the basic archiving\\nmodule to contrib (as compared with src/test/modules) and try to\\nharden it to the extent that such hardening is required. I think a lot\\nof people would get good use out of that. It might not be a completely\\nbaked solution, but a solution doesn't have to be completely baked to\\nbe a massive improvement over the stupidity endorsed by our current\\ndocumentation.\\n\\n- I wonder whether it's a good idea to silently succeed if the file\\nexists and has the same contents as the file we're trying to archive.\\nISTR that being necessary behavior for robustness, because what if we\\narchive the file and then die before recording the fact that we\\narchived it?\\n\\n- If we load a new archive library, should we give the old one a\\ncallback so it can shut down? And should the archiver considering\\nexiting since we can't unload? It isn't necessary but it might be\\nnicer.\\n\\n- I believe we decided some time back to invoke function pointers\\n(*like)(this) rather than like(this) for clarity. It was judged that\\nsomething->like(this) was fine because that can only be a function\\npointer, so no need to write (*(something->like))(this), but\\nlike(this) could make you think that \\\"like\\\" is a plain function rather\\nthan a function pointer.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 28 Jan 2022 14:06:50 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 28, 2022 at 02:06:50PM -0500, Robert Haas wrote:\\n> I've committed 0001 now. I don't see anything particularly wrong with\\n> the rest of this either, but here are a few comments:\\n\\nThanks!\\n\\n> - I wonder whether it might be better to promote the basic archiving\\n> module to contrib (as compared with src/test/modules) and try to\\n> harden it to the extent that such hardening is required. I think a lot\\n> of people would get good use out of that. It might not be a completely\\n> baked solution, but a solution doesn't have to be completely baked to\\n> be a massive improvement over the stupidity endorsed by our current\\n> documentation.\\n\\nThis has been suggested a few times in this thread, so I'll go ahead and\\nmove it to contrib. I am clearly outnumbered! :)\\n\\nI discussed the two main deficiencies I'm aware of with basic_archive\\nearlier [0]. The first one is the issue with \\\"incovenient\\\" server crashes\\n(mentioned below). The second is that there is no handling for multiple\\nservers writing to the same location since the temporary file is always\\nnamed \\\"archtemp.\\\" I thought about a few ways to pick a unique file name\\n(or at least one that is _probably_ unique), but that began adding a lot of\\ncomplexity for something I intended as a test module. I can spend some\\nmore time on this if you think it's worth fixing for a contrib module.\\n\\n> - I wonder whether it's a good idea to silently succeed if the file\\n> exists and has the same contents as the file we're trying to archive.\\n> ISTR that being necessary behavior for robustness, because what if we\\n> archive the file and then die before recording the fact that we\\n> archived it?\\n\\nYes. The only reason I didn't proceed with this earlier is because the\\nlogic became a sizable chunk of the module. I will add this in the next\\nrevision.\\n\\n> - If we load a new archive library, should we give the old one a\\n> callback so it can shut down? And should the archiver considering\\n> exiting since we can't unload? It isn't necessary but it might be\\n> nicer.\\n\\nGood idea. I'll look into this.\\n\\n> - I believe we decided some time back to invoke function pointers\\n> (*like)(this) rather than like(this) for clarity. It was judged that\\n> something->like(this) was fine because that can only be a function\\n> pointer, so no need to write (*(something->like))(this), but\\n> like(this) could make you think that \\\"like\\\" is a plain function rather\\n> than a function pointer.\\n\\nWill fix.\\n\\n[0] https://postgr.es/m/A30D8D33-8944-4898-BCA8-C77C18258247%40amazon.com\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com/\\n\\n\\n\",\n \"msg_date\": \"Fri, 28 Jan 2022 12:01:38 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 28, 2022 at 3:01 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n> I discussed the two main deficiencies I'm aware of with basic_archive\\n> earlier [0]. The first one is the issue with \\\"incovenient\\\" server crashes\\n> (mentioned below).\\n\\nSeems easy enough to rectify, if it's just a matter of silently-succeed-if-same.\\n\\n> The second is that there is no handling for multiple\\n> servers writing to the same location since the temporary file is always\\n> named \\\"archtemp.\\\" I thought about a few ways to pick a unique file name\\n> (or at least one that is _probably_ unique), but that began adding a lot of\\n> complexity for something I intended as a test module. I can spend some\\n> more time on this if you think it's worth fixing for a contrib module.\\n\\nWell, my first thought was to wonder whether we even care about that\\nscenario, but I guess we probably do, at least a little bit.\\n\\nHow about:\\n\\n1. Name temporary files like\\narchive_temp.${FINAL_NAME}.${PID}.${SOME_RANDOM_NUMBER}. Create them\\nwith O_EXCL. If it fails, die.\\n\\n2. Try not to leave files like this behind, perhaps installing an\\non_proc_exit callback or similar, but accept that crashes and unlink()\\nfailures will make it inevitable in some cases.\\n\\n3. Document that crashes or other strange failure cases may leave\\narchive_temp.* files behind in the archive directory, and recommend\\nthat users remove them before restarting the database after a crash\\n(or, with caution, removing them while the database is running if the\\nuser is sure that the files are old and unrelated to any archiving\\nstill in progress).\\n\\nI'm not arguing that this is exactly the right idea. But I am arguing\\nthat it shouldn't take a ton of engineering to come up with something\\nreasonable here.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 28 Jan 2022 15:20:41 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 28, 2022 at 03:20:41PM -0500, Robert Haas wrote:\\n> On Fri, Jan 28, 2022 at 3:01 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n>> I discussed the two main deficiencies I'm aware of with basic_archive\\n>> earlier [0]. The first one is the issue with \\\"incovenient\\\" server crashes\\n>> (mentioned below).\\n> \\n> Seems easy enough to rectify, if it's just a matter of silently-succeed-if-same.\\n\\nYes.\\n\\n>> The second is that there is no handling for multiple\\n>> servers writing to the same location since the temporary file is always\\n>> named \\\"archtemp.\\\" I thought about a few ways to pick a unique file name\\n>> (or at least one that is _probably_ unique), but that began adding a lot of\\n>> complexity for something I intended as a test module. I can spend some\\n>> more time on this if you think it's worth fixing for a contrib module.\\n> \\n> Well, my first thought was to wonder whether we even care about that\\n> scenario, but I guess we probably do, at least a little bit.\\n> \\n> How about:\\n> \\n> 1. Name temporary files like\\n> archive_temp.${FINAL_NAME}.${PID}.${SOME_RANDOM_NUMBER}. Create them\\n> with O_EXCL. If it fails, die.\\n> \\n> 2. Try not to leave files like this behind, perhaps installing an\\n> on_proc_exit callback or similar, but accept that crashes and unlink()\\n> failures will make it inevitable in some cases.\\n> \\n> 3. Document that crashes or other strange failure cases may leave\\n> archive_temp.* files behind in the archive directory, and recommend\\n> that users remove them before restarting the database after a crash\\n> (or, with caution, removing them while the database is running if the\\n> user is sure that the files are old and unrelated to any archiving\\n> still in progress).\\n> \\n> I'm not arguing that this is exactly the right idea. But I am arguing\\n> that it shouldn't take a ton of engineering to come up with something\\n> reasonable here.\\n\\nThis is roughly what I had in mind. I'll give it a try.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 28 Jan 2022 15:39:35 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Here is a new revision. I've moved basic_archive to contrib, hardened it\\nas suggested, and added shutdown support for archive modules.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Sat, 29 Jan 2022 12:50:18 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Sat, Jan 29, 2022 at 12:50:18PM -0800, Nathan Bossart wrote:\\n> Here is a new revision. I've moved basic_archive to contrib, hardened it\\n> as suggested, and added shutdown support for archive modules.\\n\\ncfbot was unhappy with v14, so here's another attempt. One other change I\\nam pondering is surrounding pgarch_MainLoop() with PG_TRY/PG_FINALLY so\\nthat we can also call the shutdown callback in the event of an ERROR. This\\nmight be necessary for an archive module that uses background workers.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Sat, 29 Jan 2022 16:31:48 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Sat, Jan 29, 2022 at 04:31:48PM -0800, Nathan Bossart wrote:\\n> On Sat, Jan 29, 2022 at 12:50:18PM -0800, Nathan Bossart wrote:\\n>> Here is a new revision. I've moved basic_archive to contrib, hardened it\\n>> as suggested, and added shutdown support for archive modules.\\n> \\n> cfbot was unhappy with v14, so here's another attempt. One other change I\\n> am pondering is surrounding pgarch_MainLoop() with PG_TRY/PG_FINALLY so\\n> that we can also call the shutdown callback in the event of an ERROR. This\\n> might be necessary for an archive module that uses background workers.\\n\\nUgh. Apologies for the noise. cfbot still isn't happy, so here's yet\\nanother attempt. This new patch set also ensures the shutdown callback is\\ncalled when the archiver process exits.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Sat, 29 Jan 2022 21:01:41 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Sat, Jan 29, 2022 at 09:01:41PM -0800, Nathan Bossart wrote:\\n> On Sat, Jan 29, 2022 at 04:31:48PM -0800, Nathan Bossart wrote:\\n>> On Sat, Jan 29, 2022 at 12:50:18PM -0800, Nathan Bossart wrote:\\n>>> Here is a new revision. I've moved basic_archive to contrib, hardened it\\n>>> as suggested, and added shutdown support for archive modules.\\n>> \\n>> cfbot was unhappy with v14, so here's another attempt. One other change I\\n>> am pondering is surrounding pgarch_MainLoop() with PG_TRY/PG_FINALLY so\\n>> that we can also call the shutdown callback in the event of an ERROR. This\\n>> might be necessary for an archive module that uses background workers.\\n> \\n> Ugh. Apologies for the noise. cfbot still isn't happy, so here's yet\\n> another attempt. This new patch set also ensures the shutdown callback is\\n> called when the archiver process exits.\\n\\nIf basic_archive is to be in contrib, we probably want to avoid restarting\\nthe archiver every time the module ERRORs. I debated trying to add a\\ngeneric exception handler that all archive modules could use, but I suspect\\nmany will have unique cleanup requirements. Plus, AFAICT restarting the\\narchiver isn't terrible, it just causes most of the normal retry logic to\\nbe skipped.\\n\\nI also looked into rewriting basic_archive to avoid ERRORs and return false\\nfor all failures, but this was rather tedious. Instead, I just introduced\\na custom exception handler for basic_archive's archive callback. This\\nallows us to ERROR as necessary (which helps ensure that failures show up\\nin the server logs), and the archiver can treat it like a normal failure\\nand avoid restarting.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Mon, 31 Jan 2022 17:36:08 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 31, 2022 at 8:36 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n> If basic_archive is to be in contrib, we probably want to avoid restarting\\n> the archiver every time the module ERRORs. I debated trying to add a\\n> generic exception handler that all archive modules could use, but I suspect\\n> many will have unique cleanup requirements. Plus, AFAICT restarting the\\n> archiver isn't terrible, it just causes most of the normal retry logic to\\n> be skipped.\\n>\\n> I also looked into rewriting basic_archive to avoid ERRORs and return false\\n> for all failures, but this was rather tedious. Instead, I just introduced\\n> a custom exception handler for basic_archive's archive callback. This\\n> allows us to ERROR as necessary (which helps ensure that failures show up\\n> in the server logs), and the archiver can treat it like a normal failure\\n> and avoid restarting.\\n\\nI think avoiding ERROR is going to be impractical. Catching it in the\\ncontrib module seems OK. Catching it in the generic code is probably\\nalso possible to do in a reasonable way. Not catching the error also\\nseems like it would be OK, since we expect errors to be infrequent.\\nI'm not objecting to anything you did here, but I'm uncertain why\\nadding basic_archive along shell_archive changes the calculus here in\\nany way. It just seems like a separate problem.\\n\\n+ /* Perform logging of memory contexts of this process */\\n+ if (LogMemoryContextPending)\\n+ ProcessLogMemoryContextInterrupt();\\n\\nAny special reason for moving this up higher? Not really an issue, just curious.\\n\\n+ gettext_noop(\\\"This is unused if\\n\\\\\\\"archive_library\\\\\\\" does not indicate archiving via shell is\\nenabled.\\\")\\n\\nThis contains a double negative. We could describe it more positively:\\nThis is used only if \\\\\\\"archive_library\\\\\\\" specifies archiving via\\nshell. But that's actually a little confusing, because the way you've\\nset it up, archiving via shell can be specified by writing either\\narchive_library = '' or archive_library = 'shell'. I don't see any\\nparticularly good reason to allow that to be spelled in two ways.\\nLet's pick one. Then here we can write either:\\n\\n(a) This is used only if archive_library = 'shell'.\\n-or-\\n(b) This is used only if archive_library is not set.\\n\\nIMHO, either of those would be clearer than what you have right now,\\nand it would definitely be shorter.\\n\\n+/*\\n+ * Callback that gets called to determine if the archive module is\\n+ * configured.\\n+ */\\n+typedef bool (*ArchiveCheckConfiguredCB) (void);\\n+\\n+/*\\n+ * Callback called to archive a single WAL file.\\n+ */\\n+typedef bool (*ArchiveFileCB) (const char *file, const char *path);\\n+\\n+/*\\n+ * Called to shutdown an archive module.\\n+ */\\n+typedef void (*ArchiveShutdownCB) (void);\\n\\nI think that this is the wrong amount of comments. One theory is that\\nthe reader should refer to the documentation to understand how these\\ncallbacks work. In that case, having a separate comment for each one\\nthat doesn't really say anything is just taking up space. It would be\\nbetter to have one comment for all three lines referring the reader to\\nthe documentation. Alternatively, one could take the position that the\\nexplanation should go into these comments, and then perhaps we don't\\neven really need documentation. A one-line comment that doesn't really\\nsay anything non-obvious seems like the worst amount.\\n\\n+ <warning>\\n+ <para>\\n+ There are considerable robustness and security risks in using\\narchive modules\\n+ because, being written in the <literal>C</literal> language, they\\nhave access\\n+ to many server resources. Administrators wishing to enable archive modules\\n+ should exercise extreme caution. Only carefully audited modules should be\\n+ loaded.\\n+ </para>\\n+ </warning>\\n\\nMaybe I'm just old and jaded, but do we really need this? I know we\\nhave the same thing for background workers, but if anything that seems\\nlike an argument against duplicating it elsewhere. Lots of copies of\\nessentially identical warnings aren't the way to great documentation;\\nif we copy this here, we'll probably copy it to more places. And also,\\nit seems a bit like warning people that they shouldn't give their\\ncomplete financial records to total strangers about whom they have no\\nlittle or no information. Do tell.\\n\\n+<programlisting>\\n+typedef bool (*ArchiveCheckConfiguredCB) (void);\\n+</programlisting>\\n+\\n+ If <literal>true</literal> is returned, the server will proceed with\\n+ archiving the file by calling the <function>archive_file_cb</function>\\n+ callback. If <literal>false</literal> is returned, archiving will not\\n+ proceed. In the latter case, the server will periodically call this\\n+ function, and archiving will proceed if it eventually returns\\n+ <literal>true</literal>.\\n\\nIt's not obvious from reading this why anyone would want to provide\\nthis callback, or have it do anything other than 'return true'. But\\nthere actually is a behavior difference if you provide this and have\\nit return false, vs. just having archiving itself fail. At least, the\\nmessage \\\"archive_mode enabled, yet archiving is not configured\\\" will\\nbe emitted. So that's something we could mention here.\\n\\nI would suggest s/if it eventually/only when it/\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Feb 2022 13:42:55 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Thanks for the review!\\n\\nOn Wed, Feb 02, 2022 at 01:42:55PM -0500, Robert Haas wrote:\\n> I think avoiding ERROR is going to be impractical. Catching it in the\\n> contrib module seems OK. Catching it in the generic code is probably\\n> also possible to do in a reasonable way. Not catching the error also\\n> seems like it would be OK, since we expect errors to be infrequent.\\n> I'm not objecting to anything you did here, but I'm uncertain why\\n> adding basic_archive along shell_archive changes the calculus here in\\n> any way. It just seems like a separate problem.\\n\\nThe main scenario I'm thinking about is when there is no space left for\\narchives. The shell archiving logic is pretty good about avoiding ERRORs,\\nso when there is a problem executing the command, the archiver will retry\\nthe command a few times before giving up for a while. If basic_archive\\njust ERROR'd due to ENOSPC, it would cause the archiver to restart. Until\\nspace frees up, I believe the archiver will end up restarting every 10\\nseconds.\\n\\nI thought some more about adding a generic exception handler for the\\narchiving callback. I think we'd need to add a new callback function that\\nwould perform any required cleanup (e.g., closing any files that might be\\nleft open). That part isn't too bad. However, module authors will also\\nneed to keep in mind that the archiving callback runs in its own transient\\nmemory context. If the module needs to palloc() something that needs to\\nstick around for a while, it will need to do so in a different memory\\ncontext. With sufficient documentation, maybe this part isn't too bad\\neither, but in the end, all of this is to save an optional ~15 lines of\\ncode in the module. It's not crucial to do your own ERROR handling in your\\narchive module, but if you want to, you can use basic_archive as a good\\nstarting point.\\n\\ntl;dr - I left it the same.\\n\\n> + /* Perform logging of memory contexts of this process */\\n> + if (LogMemoryContextPending)\\n> + ProcessLogMemoryContextInterrupt();\\n> \\n> Any special reason for moving this up higher? Not really an issue, just curious.\\n\\nSince archive_library changes cause the archiver to restart, I thought it\\nmight be good to move this before the process might exit in case\\nLogMemoryContextPending and ConfigReloadPending are both true.\\n\\n> + gettext_noop(\\\"This is unused if\\n> \\\\\\\"archive_library\\\\\\\" does not indicate archiving via shell is\\n> enabled.\\\")\\n> \\n> This contains a double negative. We could describe it more positively:\\n> This is used only if \\\\\\\"archive_library\\\\\\\" specifies archiving via\\n> shell. But that's actually a little confusing, because the way you've\\n> set it up, archiving via shell can be specified by writing either\\n> archive_library = '' or archive_library = 'shell'. I don't see any\\n> particularly good reason to allow that to be spelled in two ways.\\n> Let's pick one. Then here we can write either:\\n> \\n> (a) This is used only if archive_library = 'shell'.\\n> -or-\\n> (b) This is used only if archive_library is not set.\\n> \\n> IMHO, either of those would be clearer than what you have right now,\\n> and it would definitely be shorter.\\n\\nI went with (b). That felt a bit more natural to me, and it was easier to\\ncode because I don't have to add error checking for an empty string.\\n\\n> +/*\\n> + * Callback that gets called to determine if the archive module is\\n> + * configured.\\n> + */\\n> +typedef bool (*ArchiveCheckConfiguredCB) (void);\\n> +\\n> +/*\\n> + * Callback called to archive a single WAL file.\\n> + */\\n> +typedef bool (*ArchiveFileCB) (const char *file, const char *path);\\n> +\\n> +/*\\n> + * Called to shutdown an archive module.\\n> + */\\n> +typedef void (*ArchiveShutdownCB) (void);\\n> \\n> I think that this is the wrong amount of comments. One theory is that\\n> the reader should refer to the documentation to understand how these\\n> callbacks work. In that case, having a separate comment for each one\\n> that doesn't really say anything is just taking up space. It would be\\n> better to have one comment for all three lines referring the reader to\\n> the documentation. Alternatively, one could take the position that the\\n> explanation should go into these comments, and then perhaps we don't\\n> even really need documentation. A one-line comment that doesn't really\\n> say anything non-obvious seems like the worst amount.\\n\\nIn my quest to write well-commented code, I sometimes overdo it. I\\nadjusted these comments in the new revision.\\n\\n> + <warning>\\n> + <para>\\n> + There are considerable robustness and security risks in using\\n> archive modules\\n> + because, being written in the <literal>C</literal> language, they\\n> have access\\n> + to many server resources. Administrators wishing to enable archive modules\\n> + should exercise extreme caution. Only carefully audited modules should be\\n> + loaded.\\n> + </para>\\n> + </warning>\\n> \\n> Maybe I'm just old and jaded, but do we really need this? I know we\\n> have the same thing for background workers, but if anything that seems\\n> like an argument against duplicating it elsewhere. Lots of copies of\\n> essentially identical warnings aren't the way to great documentation;\\n> if we copy this here, we'll probably copy it to more places. And also,\\n> it seems a bit like warning people that they shouldn't give their\\n> complete financial records to total strangers about whom they have no\\n> little or no information. Do tell.\\n\\nI removed this in the new revision.\\n\\n> +<programlisting>\\n> +typedef bool (*ArchiveCheckConfiguredCB) (void);\\n> +</programlisting>\\n> +\\n> + If <literal>true</literal> is returned, the server will proceed with\\n> + archiving the file by calling the <function>archive_file_cb</function>\\n> + callback. If <literal>false</literal> is returned, archiving will not\\n> + proceed. In the latter case, the server will periodically call this\\n> + function, and archiving will proceed if it eventually returns\\n> + <literal>true</literal>.\\n> \\n> It's not obvious from reading this why anyone would want to provide\\n> this callback, or have it do anything other than 'return true'. But\\n> there actually is a behavior difference if you provide this and have\\n> it return false, vs. just having archiving itself fail. At least, the\\n> message \\\"archive_mode enabled, yet archiving is not configured\\\" will\\n> be emitted. So that's something we could mention here.\\n\\nThe blurb just above this provides a bit more information, but I tried to\\nadd some additional context in the new revision anyway.\\n\\n> I would suggest s/if it eventually/only when it/\\n\\nDone.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Wed, 2 Feb 2022 14:44:33 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 2, 2022 at 5:44 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n> > I would suggest s/if it eventually/only when it/\\n>\\n> Done.\\n\\nCommitted. I'm going to be 0% surprised if the buildfarm turns pretty\\ncolors, but I don't know how to know what it's going to be unhappy\\nabout except by trying it, so here goes.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Feb 2022 14:11:18 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 03, 2022 at 02:11:18PM -0500, Robert Haas wrote:\\n> Committed. I'm going to be 0% surprised if the buildfarm turns pretty\\n> colors, but I don't know how to know what it's going to be unhappy\\n> about except by trying it, so here goes.\\n\\nThanks! I'll keep an eye on the buildfarm and will send any new patches\\nthat are needed.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Feb 2022 11:27:47 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 3, 2022 at 2:27 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n> On Thu, Feb 03, 2022 at 02:11:18PM -0500, Robert Haas wrote:\\n> > Committed. I'm going to be 0% surprised if the buildfarm turns pretty\\n> > colors, but I don't know how to know what it's going to be unhappy\\n> > about except by trying it, so here goes.\\n>\\n> Thanks! I'll keep an eye on the buildfarm and will send any new patches\\n> that are needed.\\n\\nAndres just pointed out to me that thorntail is unhappy:\\n\\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=thorntail&dt=2022-02-03%2019%3A54%3A42\\n\\nIt says:\\n\\n==~_~===-=-===~_~==\\npgsql.build/contrib/basic_archive/log/postmaster.log\\n==~_~===-=-===~_~==\\n2022-02-03 23:17:49.019 MSK [1253623:1] FATAL: WAL archival cannot be\\nenabled when wal_level is \\\"minimal\\\"\\n\\nThe notes for the machine say:\\n\\nUBSan; force_parallel_mode; wal_level=minimal; OS bug breaks truncate()\\n\\nSo apparently we need to either skip this test when wal_level=minimal,\\nor force a higher wal_level to be used for this particular test. Not\\nsure what the existing precedents are, if any.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Feb 2022 16:04:33 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 03, 2022 at 04:04:33PM -0500, Robert Haas wrote:\\n> So apparently we need to either skip this test when wal_level=minimal,\\n> or force a higher wal_level to be used for this particular test. Not\\n> sure what the existing precedents are, if any.\\n\\nThe only precedent I've found so far is test_decoding, which sets wal_level\\nto \\\"logical.\\\" Perhaps we can just set it to \\\"replica\\\" in\\nbasic_archive.conf.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Feb 2022 13:11:21 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 3, 2022 at 4:11 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n> On Thu, Feb 03, 2022 at 04:04:33PM -0500, Robert Haas wrote:\\n> > So apparently we need to either skip this test when wal_level=minimal,\\n> > or force a higher wal_level to be used for this particular test. Not\\n> > sure what the existing precedents are, if any.\\n>\\n> The only precedent I've found so far is test_decoding, which sets wal_level\\n> to \\\"logical.\\\" Perhaps we can just set it to \\\"replica\\\" in\\n> basic_archive.conf.\\n\\nYeah, that seems to make sense.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Feb 2022 16:15:30 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 03, 2022 at 04:15:30PM -0500, Robert Haas wrote:\\n> On Thu, Feb 3, 2022 at 4:11 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n>> On Thu, Feb 03, 2022 at 04:04:33PM -0500, Robert Haas wrote:\\n>> > So apparently we need to either skip this test when wal_level=minimal,\\n>> > or force a higher wal_level to be used for this particular test. Not\\n>> > sure what the existing precedents are, if any.\\n>>\\n>> The only precedent I've found so far is test_decoding, which sets wal_level\\n>> to \\\"logical.\\\" Perhaps we can just set it to \\\"replica\\\" in\\n>> basic_archive.conf.\\n> \\n> Yeah, that seems to make sense.\\n\\n024_archive_recovery.pl seems to do something similar. Patch attached.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Thu, 3 Feb 2022 13:25:39 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 3, 2022 at 4:25 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n> 024_archive_recovery.pl seems to do something similar. Patch attached.\\n\\nCommitted. I think this is mostly an issue for pg_regress tests, as\\nopposed to 024_archive_recovery.pl, which is a TAP test. Maybe I'm\\nwrong about that, but it looks to me like most TAP tests choose what\\nthey want explicitly, while pg_regress tests tend to inherit the\\nvalue.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Feb 2022 16:45:52 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 03, 2022 at 04:45:52PM -0500, Robert Haas wrote:\\n> On Thu, Feb 3, 2022 at 4:25 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n>> 024_archive_recovery.pl seems to do something similar. Patch attached.\\n> \\n> Committed. I think this is mostly an issue for pg_regress tests, as\\n> opposed to 024_archive_recovery.pl, which is a TAP test. Maybe I'm\\n> wrong about that, but it looks to me like most TAP tests choose what\\n> they want explicitly, while pg_regress tests tend to inherit the\\n> value.\\n\\nThanks!\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Feb 2022 13:55:37 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nI am not sure why, but I can't find \\\"basic_archive.archive_directory\\\" in\\npg_settings the same way I would find for example :\\n\\\"pg_stat_statements.max\\\".\\n\\n[local]:5656 benoit@postgres=# SELECT count(*) FROM pg_settings WHERE name\\n= 'basic_archive.archive_directory';\\n count\\n-------\\n 0\\n(1 row)\\n\\nshow can find it if I use the complete name but tab completion can't find\\nthe guc:\\n\\n[local]:5656 benoit@postgres=# show basic_archive.archive_directory;\\n basic_archive.archive_directory\\n---------------------------------\\n /home/benoit/tmp/tmp/archives\\n(1 row)\\n\\nThe archiver is configured with \\\"basic_archive\\\" and is working fine. I use\\nthis version of pg:\\n\\nPostgreSQL 15beta2 on x86_64-pc-linux-gnu, compiled by gcc (GCC) 11.3.1\\n20220421 (Red Hat 11.3.1-2), 64-bit\\n\\nHi,I am not sure why, but I can't find \\\"basic_archive.archive_directory\\\" in pg_settings the same way I would find for example : \\\"pg_stat_statements.max\\\".[local]:5656 benoit@postgres=# SELECT count(*) FROM pg_settings WHERE name = 'basic_archive.archive_directory'; count ------- 0(1 row)show can find it if I use the complete name but tab completion can't find the guc:[local]:5656 benoit@postgres=# show basic_archive.archive_directory; basic_archive.archive_directory --------------------------------- /home/benoit/tmp/tmp/archives(1 row)The archiver is configured with \\\"basic_archive\\\" and is working fine. I use this version of pg:PostgreSQL 15beta2 on x86_64-pc-linux-gnu, compiled by gcc (GCC) 11.3.1 20220421 (Red Hat 11.3.1-2), 64-bit\",\n \"msg_date\": \"Wed, 6 Jul 2022 18:21:24 +0200\",\n \"msg_from\": \"talk to ben <blo.talkto@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Wed, Jul 06, 2022 at 06:21:24PM +0200, talk to ben wrote:\\n> I am not sure why, but I can't find \\\"basic_archive.archive_directory\\\" in\\n> pg_settings the same way I would find for example :\\n> \\\"pg_stat_statements.max\\\".\\n> \\n> [local]:5656 benoit@postgres=# SELECT count(*) FROM pg_settings WHERE name\\n> = 'basic_archive.archive_directory';\\n> count\\n> -------\\n> 0\\n> (1 row)\\n> \\n> show can find it if I use the complete name but tab completion can't find\\n> the guc:\\n> \\n> [local]:5656 benoit@postgres=# show basic_archive.archive_directory;\\n> basic_archive.archive_directory\\n> ---------------------------------\\n> /home/benoit/tmp/tmp/archives\\n> (1 row)\\n\\nI think the reason is that only the archiver process loads the library, so\\nthe GUC isn't registered at startup like you'd normally see with\\nshared_preload_libraries. IIUC the server will still create a placeholder\\nGUC during startup for custom parameters, which is why it shows up for SHOW\\ncommands.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Jul 2022 09:41:05 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"> I think the reason is that only the archiver process loads the library, so\\n> the GUC isn't registered at startup like you'd normally see with\\n> shared_preload_libraries. IIUC the server will still create a placeholder\\n> GUC during startup for custom parameters, which is why it shows up for SHOW\\n> commands.\\n>\\n\\nThanks for the quick answer !\\nThat's a little surprising at first but I understand better now.\\n\\nWill there be a facility to check archive_library gucs later on ? It might\\ncome in handy with more\\nguc rich archive modules.\\n\\n\\n\\nI think the reason is that only the archiver process loads the library, so\\nthe GUC isn't registered at startup like you'd normally see with\\nshared_preload_libraries. IIUC the server will still create a placeholder\\nGUC during startup for custom parameters, which is why it shows up for SHOW\\ncommands.Thanks for the quick answer ! That's a little surprising at first but I understand better now.Will there be a facility to check archive_library gucs later on ? It might come in handy with moreguc rich archive modules.\",\n \"msg_date\": \"Thu, 7 Jul 2022 09:08:39 +0200\",\n \"msg_from\": \"talk to ben <blo.talkto@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"(Sorry for the spam Nathan)\\n\\nWith the list in CC and additional information :\\n\\nThe modified archive module parameters are visible in pg_file_settings.\\nThey don't show up in \\\\dconfig+, which I understand given the query used by\\nthe\\nmeta command, but I find a little confusing from an end user POV.\\n\\n(Sorry for the spam Nathan)With the list in CC and additional information :The modified archive module parameters are visible in pg_file_settings.They don't show up in \\\\dconfig+, which I understand given the query used by the meta command, but I find a little confusing from an end user POV.\",\n \"msg_date\": \"Thu, 7 Jul 2022 09:45:46 +0200\",\n \"msg_from\": \"talk to ben <blo.talkto@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 2022-Jul-07, talk to ben wrote:\\n\\n> The modified archive module parameters are visible in pg_file_settings.\\n> They don't show up in \\\\dconfig+, which I understand given the query used by\\n> the meta command, but I find a little confusing from an end user POV.\\n\\nWell, this does sound unsatisfactory. I suppose one answer would be to\\nload the module in all backends, in case the user wants to look at the\\nvalue. But that would be wasteful. Maybe we should have a warning\\nabout it in the docs -- tell people to LOAD the library if they want to\\nexamine the configuration?\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Jul 2022 17:11:57 +0200\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> Well, this does sound unsatisfactory. I suppose one answer would be to\\n> load the module in all backends, in case the user wants to look at the\\n> value. But that would be wasteful. Maybe we should have a warning\\n> about it in the docs -- tell people to LOAD the library if they want to\\n> examine the configuration?\\n\\nThe underlying issue here is that the pg_settings view doesn't show\\nplaceholder GUCs because we lack satisfactory values to put in\\nmost of the columns. I don't know if revisiting that conclusion\\nwould be appropriate or not. The purist approach would be to show\\nNULL for any unknown column, but how many applications would that\\nbreak? And even the \\\"known\\\" values could change unexpectedly when\\nthe module does get loaded, for example if the GUC has units and\\nthe value in the config file is expressed in a non-canonical way.\\n(To say nothing of what a show hook might do...)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Jul 2022 11:48:20 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Jul 07, 2022 at 11:48:20AM -0400, Tom Lane wrote:\\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n>> Well, this does sound unsatisfactory. I suppose one answer would be to\\n>> load the module in all backends, in case the user wants to look at the\\n>> value. But that would be wasteful. Maybe we should have a warning\\n>> about it in the docs -- tell people to LOAD the library if they want to\\n>> examine the configuration?\\n\\nYeah, for something like basic_archive, it should be fine to load it via\\nshared_preload_libraries or LOAD as well as archive_library, but not all\\narchive libraries might be written to handle that correctly. And this is\\nnot the most user-friendly.\\n\\n> The underlying issue here is that the pg_settings view doesn't show\\n> placeholder GUCs because we lack satisfactory values to put in\\n> most of the columns. I don't know if revisiting that conclusion\\n> would be appropriate or not. The purist approach would be to show\\n> NULL for any unknown column, but how many applications would that\\n> break? And even the \\\"known\\\" values could change unexpectedly when\\n> the module does get loaded, for example if the GUC has units and\\n> the value in the config file is expressed in a non-canonical way.\\n> (To say nothing of what a show hook might do...)\\n\\nPerhaps the \\\"category\\\" could indicate that this is a placeholder value, and\\nthe pg_settings documentation could explain exactly what that means (i.e.,\\nunknown to any libraries loaded in the current process, but may have\\nmeaning to others).\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Jul 2022 09:59:08 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Would this addition to the documentation be ok ? I hope the english is not\\ntoo broken ..\",\n \"msg_date\": \"Tue, 23 Aug 2022 16:18:52 +0200\",\n \"msg_from\": \"talk to ben <blo.talkto@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 23, 2022 at 04:18:52PM +0200, talk to ben wrote:\\n> --- a/doc/src/sgml/basic-archive.sgml\\n> +++ b/doc/src/sgml/basic-archive.sgml\\n> @@ -68,6 +68,19 @@ basic_archive.archive_directory = '/path/to/archive/directory'\\n> to any archiving still in progress, but users should use extra caution when\\n> doing so.\\n> </para>\\n> +\\n> + <para>\\n> + The archive module is loaded by the archiver process. Therefore, the\\n> + parameters defined in the module are not set outside this process and cannot\\n> + be seen from the <structname>pg_settings</structname> view or the\\n> + \\\\dconfig meta-command.\\n> + These parameters values can be shown from the server's configuration\\n> + file(s) through the <structname>pg_file_settings</structname> view.\\n> + If you want to check the actual values applied by the archiver, you can\\n> + <command>LOAD</command> the module before reading\\n> + <structname>pg_settings</structname>. It's also possible to search\\n> + for the options directly with the <command>SHOW</command> command.\\n> + </para>\\n\\nI don't know if it makes sense to document this in basic_archive. On one\\nhand, it seems like folks will commonly encounter this behavior with this\\nmodule, so this feels like a natural place for such a note. But on the\\nother hand, this is generic behavior for any library that is dynamically\\nloaded in a separate process.\\n\\nOverall, I think I'm +1 for this patch. I haven't thought too much about\\nthe exact wording, but provided others support it as well, I will try to\\ntake a deeper look soon.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 23 Aug 2022 10:39:30 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Nathan Bossart <nathandbossart@gmail.com> writes:\\n> I don't know if it makes sense to document this in basic_archive. On one\\n> hand, it seems like folks will commonly encounter this behavior with this\\n> module, so this feels like a natural place for such a note. But on the\\n> other hand, this is generic behavior for any library that is dynamically\\n> loaded in a separate process.\\n\\nYeah, I don't think this material is at all specific to basic_archive.\\nMaybe it could be documented near the relevant views, if it isn't already.\\n\\nAlso, I think the proposed text neglects the case of including the\\nmodule in shared_preload_libraries.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 23 Aug 2022 13:47:40 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Nathan Bossart <nathandbossart@gmail.com> writes:\\n> On one hand, it seems like folks will commonly encounter this behavior\\nwith this\\n> module, so this feels like a natural place for such a note.\\n\\nYes, I looked there first.\\n\\nWould this addition to the pg_settings description be better ?\",\n \"msg_date\": \"Wed, 24 Aug 2022 10:05:55 +0200\",\n \"msg_from\": \"talk to ben <blo.talkto@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 24, 2022 at 10:05:55AM +0200, talk to ben wrote:\\n> This view does not display <link linkend=\\\"runtime-config-custom\\\">customized options</link>\\n> - until the extension module that defines them has been loaded.\\n> + until the extension module that defines them has been loaded. Therefore, any\\n> + option defined in a library that is dynamically loaded in a separate process\\n> + will not be visible in the view, unless the module is manually loaded\\n> + beforehand. This case applies for example to an archive module loaded by the\\n> + archiver process.\\n\\nI would suggest something like:\\n\\n\\tThis view does not display customized options until the extension\\n\\tmodule that defines them has been loaded by the backend process\\n\\texecuting the query (e.g., via shared_preload_libraries, the LOAD\\n\\tcommand, or a call to a user-defined C function). For example, since\\n\\tthe archive_library is only loaded by the archiver process, this view\\n\\twill not display any customized options defined by archive modules\\n\\tunless special action is taken to load them into the backend process\\n\\texecuting the query.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 24 Aug 2022 15:17:23 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Here is a patch with the proposed wording.\",\n \"msg_date\": \"Thu, 25 Aug 2022 15:29:41 +0200\",\n \"msg_from\": \"talk to ben <blo.talkto@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 25, 2022 at 03:29:41PM +0200, talk to ben wrote:\\n> Here is a patch with the proposed wording.\\n\\nHere is the same patch with a couple more links.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Thu, 25 Aug 2022 13:06:00 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 25, 2022 at 01:06:00PM -0700, Nathan Bossart wrote:\\n> On Thu, Aug 25, 2022 at 03:29:41PM +0200, talk to ben wrote:\\n>> Here is a patch with the proposed wording.\\n> \\n> Here is the same patch with a couple more links.\\n\\nI would advise that you create a commitfest entry for your patch so that it\\nisn't forgotten.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Aug 2022 15:46:21 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 30, 2022 at 12:46 AM Nathan Bossart <nathandbossart@gmail.com>\\nwrote:\\n\\n> I would advise that you create a commitfest entry for your patch so that it\\n> isn't forgotten.\\n>\\n\\nOk done, https://commitfest.postgresql.org/39/3856/ (is that fine if I\\nadded you as a reviewer ?)\\n\\nand thanks for the added links in the patch.\\n\\nOn Tue, Aug 30, 2022 at 12:46 AM Nathan Bossart <nathandbossart@gmail.com> wrote:\\nI would advise that you create a commitfest entry for your patch so that it\\nisn't forgotten.Ok done, https://commitfest.postgresql.org/39/3856/ (is that fine if I added you as a reviewer ?)and thanks for the added links in the patch.\",\n \"msg_date\": \"Tue, 30 Aug 2022 09:49:20 +0200\",\n \"msg_from\": \"=?UTF-8?Q?Benoit_Lobr=C3=A9au?= <benoit.lobreau@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 30, 2022 at 09:49:20AM +0200, Benoit Lobr�au wrote:\\n> Ok done, https://commitfest.postgresql.org/39/3856/ (is that fine if I\\n> added you as a reviewer ?)\\n\\nOf course. I've marked it as ready-for-committer.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 30 Aug 2022 09:30:48 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Nathan Bossart <nathandbossart@gmail.com> writes:\\n> Of course. I've marked it as ready-for-committer.\\n\\nPushed with a bit of additional wordsmithing.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 10 Sep 2022 16:44:16 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Sat, Sep 10, 2022 at 04:44:16PM -0400, Tom Lane wrote:\\n> Nathan Bossart <nathandbossart@gmail.com> writes:\\n>> Of course. I've marked it as ready-for-committer.\\n> \\n> Pushed with a bit of additional wordsmithing.\\n\\nThanks!\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Sat, 10 Sep 2022 14:10:46 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"I noticed that this patch has gone around and mostly purged mentions of\\narchive_command from the documentation and replaced them with\\narchive_library. I don't think this is helpful, since people still use\\narchive_command and will want to see what the documentation has to say\\nabout it. I suggest we rewind that a bit and for example replace things\\nlike\\n\\n removed or recycled until they are archived. If WAL archiving cannot keep up\\n- with the pace that WAL is generated, or if <varname>archive_command</varname>\\n+ with the pace that WAL is generated, or if <varname>archive_library</varname>\\n fails repeatedly, old WAL files will accumulate in <filename>pg_wal</filename>\\n\\nwith\\n\\n removed or recycled until they are archived. If WAL archiving cannot keep up\\n with the pace that WAL is generated, or if <varname>archive_command</varname>\\n with the pace that WAL is generated, or if <varname>archive_command</varname>\\n or <varname>archive_library</varname>\\n fail repeatedly, old WAL files will accumulate in <filename>pg_wal</filename>\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Sep 2022 06:37:38 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 14, 2022 at 06:37:38AM +0200, Peter Eisentraut wrote:\\n> I noticed that this patch has gone around and mostly purged mentions of\\n> archive_command from the documentation and replaced them with\\n> archive_library. I don't think this is helpful, since people still use\\n> archive_command and will want to see what the documentation has to say\\n> about it. I suggest we rewind that a bit and for example replace things\\n> like\\n> \\n> removed or recycled until they are archived. If WAL archiving cannot keep up\\n> - with the pace that WAL is generated, or if <varname>archive_command</varname>\\n> + with the pace that WAL is generated, or if <varname>archive_library</varname>\\n> fails repeatedly, old WAL files will accumulate in <filename>pg_wal</filename>\\n> \\n> with\\n> \\n> removed or recycled until they are archived. If WAL archiving cannot keep up\\n> with the pace that WAL is generated, or if <varname>archive_command</varname>\\n> with the pace that WAL is generated, or if <varname>archive_command</varname>\\n> or <varname>archive_library</varname>\\n> fail repeatedly, old WAL files will accumulate in <filename>pg_wal</filename>\\n\\nYep. Some references to archive_library have been changed by 31e121\\nto do exactly that. There seem to be more spots in need of an\\nupdate.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 14 Sep 2022 14:25:30 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 14.09.22 07:25, Michael Paquier wrote:\\n>> removed or recycled until they are archived. If WAL archiving cannot keep up\\n>> - with the pace that WAL is generated, or if <varname>archive_command</varname>\\n>> + with the pace that WAL is generated, or if <varname>archive_library</varname>\\n>> fails repeatedly, old WAL files will accumulate in <filename>pg_wal</filename>\\n>>\\n>> with\\n>>\\n>> removed or recycled until they are archived. If WAL archiving cannot keep up\\n>> with the pace that WAL is generated, or if <varname>archive_command</varname>\\n>> with the pace that WAL is generated, or if <varname>archive_command</varname>\\n>> or <varname>archive_library</varname>\\n>> fail repeatedly, old WAL files will accumulate in <filename>pg_wal</filename>\\n> \\n> Yep. Some references to archive_library have been changed by 31e121\\n> to do exactly that. There seem to be more spots in need of an\\n> update.\\n\\nI don't see anything in 31e121 about that.\\n\\nHere is a patch that addresses this.\",\n \"msg_date\": \"Wed, 14 Sep 2022 21:31:04 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Another question on this feature: Currently, if archive_library is set, \\narchive_command is ignored. I think if both are set, it should be an \\nerror. Compare for example what happens if you set multiple \\nrecovery_target_xxx settings. I don't think silently turning off one \\nsetting by setting another is a good behavior.\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Sep 2022 21:33:46 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 14, 2022 at 09:33:46PM +0200, Peter Eisentraut wrote:\\n> Another question on this feature: Currently, if archive_library is set,\\n> archive_command is ignored. I think if both are set, it should be an error.\\n> Compare for example what happens if you set multiple recovery_target_xxx\\n> settings. I don't think silently turning off one setting by setting another\\n> is a good behavior.\\n\\nI originally did it this way, but changed it based on this feedback [0]. I\\nhave no problem with the general idea, but the recovery_target_* logic does\\nhave the following note:\\n\\n\\t * XXX this code is broken by design. Throwing an error from a GUC assign\\n\\t * hook breaks fundamental assumptions of guc.c. So long as all the variables\\n\\t * for which this can happen are PGC_POSTMASTER, the consequences are limited,\\n\\t * since we'd just abort postmaster startup anyway. Nonetheless it's likely\\n\\t * that we have odd behaviors such as unexpected GUC ordering dependencies.\\n\\n[0] https://postgr.es/m/CA%2BTgmoaf4Y7_U%2B_W%2BSg5DoAta_FMssr%3D52mx7-_tJnfaD1VubQ%40mail.gmail.com\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Sep 2022 13:03:05 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 14.09.22 22:03, Nathan Bossart wrote:\\n> On Wed, Sep 14, 2022 at 09:33:46PM +0200, Peter Eisentraut wrote:\\n>> Another question on this feature: Currently, if archive_library is set,\\n>> archive_command is ignored. I think if both are set, it should be an error.\\n>> Compare for example what happens if you set multiple recovery_target_xxx\\n>> settings. I don't think silently turning off one setting by setting another\\n>> is a good behavior.\\n> \\n> I originally did it this way, but changed it based on this feedback [0]. I\\n> have no problem with the general idea, but the recovery_target_* logic does\\n> have the following note:\\n> \\n> \\t * XXX this code is broken by design. Throwing an error from a GUC assign\\n> \\t * hook breaks fundamental assumptions of guc.c. So long as all the variables\\n> \\t * for which this can happen are PGC_POSTMASTER, the consequences are limited,\\n> \\t * since we'd just abort postmaster startup anyway. Nonetheless it's likely\\n> \\t * that we have odd behaviors such as unexpected GUC ordering dependencies.\\n\\nAh yes, that won't work. But maybe we can just check it at run time, \\nlike in LoadArchiveLibrary().\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Sep 2022 22:07:11 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> On 14.09.22 22:03, Nathan Bossart wrote:\\n>> I originally did it this way, but changed it based on this feedback [0]. I\\n>> have no problem with the general idea, but the recovery_target_* logic does\\n>> have the following note:\\n>> \\n>> * XXX this code is broken by design. Throwing an error from a GUC assign\\n>> * hook breaks fundamental assumptions of guc.c. So long as all the variables\\n>> * for which this can happen are PGC_POSTMASTER, the consequences are limited,\\n>> * since we'd just abort postmaster startup anyway. Nonetheless it's likely\\n>> * that we have odd behaviors such as unexpected GUC ordering dependencies.\\n\\n> Ah yes, that won't work. But maybe we can just check it at run time, \\n> like in LoadArchiveLibrary().\\n\\nYeah, the objection there is only to trying to enforce such\\ninterrelationships in GUC hooks. In this case it seems to me that\\nwe could easily check and complain at the point where we're about\\nto use the GUC values.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Sep 2022 16:47:23 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 14, 2022 at 09:31:04PM +0200, Peter Eisentraut wrote:\\n> Here is a patch that addresses this.\\n\\nMy intent was to present archive_command as the built-in archive library,\\nbut I can see how this might cause confusion, so this change seems\\nreasonable to me.\\n\\n> + <para>\\n> + It is important that the archive command return zero exit status if and\\n> + only if it succeeds. Upon getting a zero result,\\n> + <productname>PostgreSQL</productname> will assume that the file has been\\n> + successfully archived, and will remove or recycle it. However, a nonzero\\n> + status tells <productname>PostgreSQL</productname> that the file was not archived;\\n> + it will try again periodically until it succeeds.\\n> + </para>\\n> +\\n> + <para>\\n> + When the archive command is terminated by a signal (other than\\n> + <systemitem>SIGTERM</systemitem> that is used as part of a server\\n> + shutdown) or an error by the shell with an exit status greater than\\n> + 125 (such as command not found), the archiver process aborts and gets\\n> + restarted by the postmaster. In such cases, the failure is\\n> + not reported in <xref linkend=\\\"pg-stat-archiver-view\\\"/>.\\n> + </para>\\n\\nThis wording is very similar to the existing wording in the archive library\\nsection below it. I think the second paragraph covers the shell command case\\nexplicitly, too. Perhaps these should be combined.\\n\\n> + <varname>archive_mode</varname> and <varname>archive_command</varname> are\\n> + separate variables so that <varname>archive_command</varname> can be\\n> + changed without leaving archiving mode.\\n\\nI believe this applies to archive_library, too.\\n\\n> - for segments to complete like <xref linkend=\\\"guc-archive-library\\\"/> does.\\n> + for segments to complete like <xref linkend=\\\"guc-archive-command\\\"/> and\\n> + <xref linkend=\\\"guc-archive-library\\\"/> does.\\n\\nnitpick: s/does/do\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Sep 2022 14:09:33 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 14, 2022 at 04:47:23PM -0400, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>> On 14.09.22 22:03, Nathan Bossart wrote:\\n>>> I originally did it this way, but changed it based on this feedback [0]. I\\n>>> have no problem with the general idea, but the recovery_target_* logic does\\n>>> have the following note:\\n>>> \\n>>> * XXX this code is broken by design. Throwing an error from a GUC assign\\n>>> * hook breaks fundamental assumptions of guc.c. So long as all the variables\\n>>> * for which this can happen are PGC_POSTMASTER, the consequences are limited,\\n>>> * since we'd just abort postmaster startup anyway. Nonetheless it's likely\\n>>> * that we have odd behaviors such as unexpected GUC ordering dependencies.\\n> \\n>> Ah yes, that won't work. But maybe we can just check it at run time, \\n>> like in LoadArchiveLibrary().\\n> \\n> Yeah, the objection there is only to trying to enforce such\\n> interrelationships in GUC hooks. In this case it seems to me that\\n> we could easily check and complain at the point where we're about\\n> to use the GUC values.\\n\\nI think the cleanest way to do something like that would be to load a\\ncheck_configured_cb that produces a WARNING. IIRC failing in\\nLoadArchiveLibrary() would just cause the archiver process to restart over\\nand over. HandlePgArchInterrupts() might need some work as well.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Sep 2022 14:17:46 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Nathan Bossart <nathandbossart@gmail.com> writes:\\n> On Wed, Sep 14, 2022 at 04:47:23PM -0400, Tom Lane wrote:\\n>> Yeah, the objection there is only to trying to enforce such\\n>> interrelationships in GUC hooks. In this case it seems to me that\\n>> we could easily check and complain at the point where we're about\\n>> to use the GUC values.\\n\\n> I think the cleanest way to do something like that would be to load a\\n> check_configured_cb that produces a WARNING. IIRC failing in\\n> LoadArchiveLibrary() would just cause the archiver process to restart over\\n> and over. HandlePgArchInterrupts() might need some work as well.\\n\\nHm. Maybe consistency-check these settings in the postmaster, sometime\\nafter we've absorbed all GUC settings but before we launch any children?\\nThat could provide a saner implementation for the recovery_target_*\\nvariables too.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Sep 2022 18:12:09 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 14, 2022 at 06:12:09PM -0400, Tom Lane wrote:\\n> Nathan Bossart <nathandbossart@gmail.com> writes:\\n>> On Wed, Sep 14, 2022 at 04:47:23PM -0400, Tom Lane wrote:\\n>>> Yeah, the objection there is only to trying to enforce such\\n>>> interrelationships in GUC hooks. In this case it seems to me that\\n>>> we could easily check and complain at the point where we're about\\n>>> to use the GUC values.\\n> \\n>> I think the cleanest way to do something like that would be to load a\\n>> check_configured_cb that produces a WARNING. IIRC failing in\\n>> LoadArchiveLibrary() would just cause the archiver process to restart over\\n>> and over. HandlePgArchInterrupts() might need some work as well.\\n> \\n> Hm. Maybe consistency-check these settings in the postmaster, sometime\\n> after we've absorbed all GUC settings but before we launch any children?\\n> That could provide a saner implementation for the recovery_target_*\\n> variables too.\\n\\nBoth archive_command and archive_library are PGC_SIGHUP, so IIUC that\\nwouldn't be sufficient. I attached a quick sketch that seems to provide\\nthe desired behavior. It's nowhere near committable yet, but it\\ndemonstrates what I'm thinking.\\n\\nFor recovery_target_*, something like you are describing seems reasonable.\\nI believe PostmasterMain() already performs some similar checks.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Wed, 14 Sep 2022 15:27:36 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 14.09.22 23:09, Nathan Bossart wrote:\\n> On Wed, Sep 14, 2022 at 09:31:04PM +0200, Peter Eisentraut wrote:\\n>> Here is a patch that addresses this.\\n> \\n> My intent was to present archive_command as the built-in archive library,\\n> but I can see how this might cause confusion, so this change seems\\n> reasonable to me.\\n\\nWhile working on this, I noticed that in master this conflicts with \\ncommit 3cabe45a819f8a2a282d9d57e45f259c84e97c3f. I have posted a \\nmessage in that thread looking for a resolution.\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 17 Sep 2022 11:49:01 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 17.09.22 11:49, Peter Eisentraut wrote:\\n> On 14.09.22 23:09, Nathan Bossart wrote:\\n>> On Wed, Sep 14, 2022 at 09:31:04PM +0200, Peter Eisentraut wrote:\\n>>> Here is a patch that addresses this.\\n>>\\n>> My intent was to present archive_command as the built-in archive library,\\n>> but I can see how this might cause confusion, so this change seems\\n>> reasonable to me.\\n> \\n> While working on this, I noticed that in master this conflicts with \\n> commit 3cabe45a819f8a2a282d9d57e45f259c84e97c3f. I have posted a \\n> message in that thread looking for a resolution.\\n\\nI have received clarification there, so I went ahead with this patch \\nhere after some adjustments in master around that other patch.\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 22 Sep 2022 07:54:33 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 14, 2022 at 09:33:46PM +0200, Peter Eisentraut wrote:\\n> Another question on this feature: Currently, if archive_library is set,\\n> archive_command is ignored. I think if both are set, it should be an error.\\n> Compare for example what happens if you set multiple recovery_target_xxx\\n> settings. I don't think silently turning off one setting by setting another\\n> is a good behavior.\\n\\nPeter, would you like to proceed with something like [0] to resolve this?\\nIf so, I will work on cleaning the patch up.\\n\\n[0] https://postgr.es/m/20220914222736.GA3042279%40nathanxps13\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 22 Sep 2022 11:46:13 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 15.09.22 00:27, Nathan Bossart wrote:\\n> Both archive_command and archive_library are PGC_SIGHUP, so IIUC that\\n> wouldn't be sufficient. I attached a quick sketch that seems to provide\\n> the desired behavior. It's nowhere near committable yet, but it\\n> demonstrates what I'm thinking.\\n\\nWhat is the effect of issuing a warning like in this patch? Would it \\njust not archive anything until the configuration is fixed? I'm not \\nsure what behavior you are going for; it's a bit hard to imagine from \\njust reading the patch.\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Sep 2022 05:58:42 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Fri, Sep 23, 2022 at 05:58:42AM -0400, Peter Eisentraut wrote:\\n> On 15.09.22 00:27, Nathan Bossart wrote:\\n>> Both archive_command and archive_library are PGC_SIGHUP, so IIUC that\\n>> wouldn't be sufficient. I attached a quick sketch that seems to provide\\n>> the desired behavior. It's nowhere near committable yet, but it\\n>> demonstrates what I'm thinking.\\n> \\n> What is the effect of issuing a warning like in this patch? Would it just\\n> not archive anything until the configuration is fixed? I'm not sure what\\n> behavior you are going for; it's a bit hard to imagine from just reading the\\n> patch.\\n\\nYes, it will halt archiving and emit a WARNING, just like what happens on\\nreleased versions when you leave archive_command empty.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Sep 2022 09:14:02 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 23.09.22 18:14, Nathan Bossart wrote:\\n> On Fri, Sep 23, 2022 at 05:58:42AM -0400, Peter Eisentraut wrote:\\n>> On 15.09.22 00:27, Nathan Bossart wrote:\\n>>> Both archive_command and archive_library are PGC_SIGHUP, so IIUC that\\n>>> wouldn't be sufficient. I attached a quick sketch that seems to provide\\n>>> the desired behavior. It's nowhere near committable yet, but it\\n>>> demonstrates what I'm thinking.\\n>>\\n>> What is the effect of issuing a warning like in this patch? Would it just\\n>> not archive anything until the configuration is fixed? I'm not sure what\\n>> behavior you are going for; it's a bit hard to imagine from just reading the\\n>> patch.\\n> \\n> Yes, it will halt archiving and emit a WARNING, just like what happens on\\n> released versions when you leave archive_command empty.\\n\\nLeaving archive_command empty is an intentional configuration choice.\\n\\nWhat we are talking about here is, arguably, a misconfiguration, so it \\nshould result in an error.\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 5 Oct 2022 19:55:58 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 05, 2022 at 07:55:58PM +0200, Peter Eisentraut wrote:\\n> Leaving archive_command empty is an intentional configuration choice.\\n> \\n> What we are talking about here is, arguably, a misconfiguration, so it\\n> should result in an error.\\n\\nOkay. What do you think about something like the attached?\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Wed, 5 Oct 2022 11:57:16 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 05.10.22 20:57, Nathan Bossart wrote:\\n> On Wed, Oct 05, 2022 at 07:55:58PM +0200, Peter Eisentraut wrote:\\n>> Leaving archive_command empty is an intentional configuration choice.\\n>>\\n>> What we are talking about here is, arguably, a misconfiguration, so it\\n>> should result in an error.\\n> \\n> Okay. What do you think about something like the attached?\\n\\nThat looks like the right solution to me.\\n\\nLet's put that into PG 16, and maybe we can consider backpatching it.\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 10 Oct 2022 09:47:24 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 10, 2022 at 1:17 PM Peter Eisentraut\\n<peter.eisentraut@enterprisedb.com> wrote:\\n>\\n> On 05.10.22 20:57, Nathan Bossart wrote:\\n> >> What we are talking about here is, arguably, a misconfiguration, so it\\n> >> should result in an error.\\n> >\\n> > Okay. What do you think about something like the attached?\\n\\nThe intent here looks reasonable to me. However, why should the user\\nbe able to set both archive_command and archive_library in the first\\nplace only to later fail in LoadArchiveLibrary() per the patch? IMO,\\nthe check_hook() is the right way to disallow any sorts of GUC\\nmisconfigurations, no?\\n\\nFWIW, I'm attaching a small patch that uses check_hook().\\n\\n> That looks like the right solution to me.\\n>\\n> Let's put that into PG 16, and maybe we can consider backpatching it.\\n\\n+1 to backpatch to PG 15 where the archive modules feature was introduced.\\n\\n--\\nBharath Rupireddy\\nPostgreSQL Contributors Team\\nRDS Open Source Databases\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Thu, 13 Oct 2022 15:25:27 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 13, 2022 at 03:25:27PM +0530, Bharath Rupireddy wrote:\\n> The intent here looks reasonable to me. However, why should the user\\n> be able to set both archive_command and archive_library in the first\\n> place only to later fail in LoadArchiveLibrary() per the patch? IMO,\\n> the check_hook() is the right way to disallow any sorts of GUC\\n> misconfigurations, no?\\n\\nThere was some discussion upthread about using the GUC hooks to enforce\\nthis [0]. In general, it doesn't seem to be a recommended practice. One\\nbasic example of the problems with this approach is the following:\\n\\n 1. Set archive_command and leave archive_library unset and restart\\n the server.\\n 2. Unset archive_command and set archive_library and call 'pg_ctl\\n reload'.\\n\\nAfter these steps, you'll see the following log messages:\\n\\n\\t2022-10-13 10:58:42.112 PDT [1562524] LOG: received SIGHUP, reloading configuration files\\n\\t2022-10-13 10:58:42.114 PDT [1562524] LOG: cannot set \\\"archive_library\\\" when \\\"archive_command\\\" is specified\\n\\t2022-10-13 10:58:42.114 PDT [1562524] DETAIL: Only one of \\\"archive_library\\\" or \\\"archive_command\\\" can be specified.\\n\\t2022-10-13 10:58:42.114 PDT [1562524] LOG: parameter \\\"archive_command\\\" changed to \\\"\\\"\\n\\t2022-10-13 10:58:42.114 PDT [1562524] LOG: configuration file \\\"/home/nathan/pgdata/postgresql.conf\\\" contains errors; unaffected changes were applied\\n\\n[0] https://postgr.es/m/20220914200305.GA2984249%40nathanxps13\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 13 Oct 2022 11:24:08 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"Nathan Bossart <nathandbossart@gmail.com> writes:\\n> On Thu, Oct 13, 2022 at 03:25:27PM +0530, Bharath Rupireddy wrote:\\n>> The intent here looks reasonable to me. However, why should the user\\n>> be able to set both archive_command and archive_library in the first\\n>> place only to later fail in LoadArchiveLibrary() per the patch? IMO,\\n>> the check_hook() is the right way to disallow any sorts of GUC\\n>> misconfigurations, no?\\n\\n> There was some discussion upthread about using the GUC hooks to enforce\\n> this [0]. In general, it doesn't seem to be a recommended practice.\\n\\nYeah, it really does not work to use GUC hooks to enforce multi-variable\\nconstraints. We've learned that the hard way (more than once, if memory\\nserves).\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 13 Oct 2022 14:53:38 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 13, 2022 at 02:53:38PM -0400, Tom Lane wrote:\\n> Yeah, it really does not work to use GUC hooks to enforce multi-variable\\n> constraints. We've learned that the hard way (more than once, if memory\\n> serves).\\n\\n414c2fd is one of the most recent ones. Its thread is about the same\\nthing.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 14 Oct 2022 09:30:30 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 14, 2022 at 6:00 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Thu, Oct 13, 2022 at 02:53:38PM -0400, Tom Lane wrote:\\n> > Yeah, it really does not work to use GUC hooks to enforce multi-variable\\n> > constraints. We've learned that the hard way (more than once, if memory\\n> > serves).\\n>\\n> 414c2fd is one of the most recent ones. Its thread is about the same\\n> thing.\\n\\nGot it. Thanks. Just thinking if we must move below comment somewhere\\nto guc related files?\\n\\n * XXX this code is broken by design. Throwing an error from a GUC assign\\n * hook breaks fundamental assumptions of guc.c. So long as all the variables\\n * for which this can happen are PGC_POSTMASTER, the consequences are limited,\\n * since we'd just abort postmaster startup anyway. Nonetheless it's likely\\n * that we have odd behaviors such as unexpected GUC ordering dependencies.\\n */\\n\\nFWIW, I see check_stage_log_stats() and check_log_stats() that set\\nerrdetail and return false causing the similar error:\\n\\npostgres=# alter system set log_statement_stats = true;\\npostgres=# select pg_reload_conf();\\npostgres=# alter system set log_statement_stats = false;\\npostgres=# alter system set log_parser_stats = true;\\nERROR: invalid value for parameter \\\"log_parser_stats\\\": 1\\nDETAIL: Cannot enable parameter when \\\"log_statement_stats\\\" is true.\\n\\nOn Thu, Oct 13, 2022 at 11:54 PM Nathan Bossart\\n<nathandbossart@gmail.com> wrote:\\n>\\n> On Thu, Oct 13, 2022 at 03:25:27PM +0530, Bharath Rupireddy wrote:\\n> > The intent here looks reasonable to me. However, why should the user\\n> > be able to set both archive_command and archive_library in the first\\n> > place only to later fail in LoadArchiveLibrary() per the patch? IMO,\\n> > the check_hook() is the right way to disallow any sorts of GUC\\n> > misconfigurations, no?\\n>\\n> There was some discussion upthread about using the GUC hooks to enforce\\n> this [0]. In general, it doesn't seem to be a recommended practice. One\\n> basic example of the problems with this approach is the following:\\n>\\n> 1. Set archive_command and leave archive_library unset and restart\\n> the server.\\n> 2. Unset archive_command and set archive_library and call 'pg_ctl\\n> reload'.\\n\\nThanks. And yes, if GUC 'foo' is reset but not reloaded and the\\ncheck_hook() in the GUC 'bar' while setting it uses the old value of\\n'foo' and fails.\\n\\nI'm re-attaching Nathan's patch as-is from [1] here again, just to\\nmake CF bot test the correct patch. Few comments on that patch:\\n\\n1)\\n+ if (XLogArchiveLibrary[0] != '\\\\0' && XLogArchiveCommand[0] != '\\\\0')\\n+ ereport(ERROR,\\n+ (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n+ errmsg(\\\"both archive_command and archive_library specified\\\"),\\n+ errdetail(\\\"Only one of archive_command,\\narchive_library may be set.\\\")));\\n\\nThe above errmsg looks informational. Can we just say something like\\nbelow? It doesn't require errdetail as the errmsg says it all. See\\nthe other instances elsewhere [2].\\n\\n ereport(ERROR,\\n (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n errmsg(\\\"cannot specify both \\\\\\\"archive_command\\\\\\\" and\\n\\\\\\\"archive_library\\\\\\\"\\\")));\\n\\n2) I think we have a problem - set archive_mode and archive_library\\nand start the server, then set archive_command, reload the conf, see\\n[3] - the archiver needs to error out right? The archiver gets\\nrestarted whenever archive_library changes but not when\\narchive_command changes. I think the right place for the error is\\nafter or at the end of HandlePgArchInterrupts().\\n\\n[1] https://www.postgresql.org/message-id/20221005185716.GB201192%40nathanxps13\\n[2] errmsg(\\\"cannot specify both PARSER and COPY options\\\")));\\nerrmsg(\\\"cannot specify both %s and %s\\\",\\nerrmsg(\\\"cannot specify both %s and %s\\\",\\n[3]\\n./psql -c \\\"alter system set archive_mode='on'\\\" postgres\\n./psql -c \\\"alter system set\\narchive_library='/home/ubuntu/postgres/contrib/basic_archive/basic_archive.so'\\\"\\npostgres\\n./pg_ctl -D data -l logfile restart\\n./psql -c \\\"alter system set archive_command='cp %p\\n/home/ubuntu/archived_wal/%f'\\\" postgres\\n./psql -c \\\"select pg_reload_conf();\\\" postgres\\npostgres=# show archive_mode;\\n archive_mode\\n--------------\\n on\\n(1 row)\\npostgres=# show archive_command ;\\n archive_command\\n------------------------------------\\n cp %p /home/ubuntu/archived_wal/%f\\n(1 row)\\npostgres=# show archive_library ;\\n archive_library\\n--------------------------------------------------------------\\n /home/ubuntu/postgres/contrib/basic_archive/basic_archive.so\\n(1 row)\\npostgres=# select pid, wait_event_type, backend_type from\\npg_stat_activity where backend_type = 'archiver';\\n pid | wait_event_type | backend_type\\n---------+-----------------+--------------\\n 2116760 | Activity | archiver\\n(1 row)\\n\\n--\\nBharath Rupireddy\\nPostgreSQL Contributors Team\\nRDS Open Source Databases\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Fri, 14 Oct 2022 12:10:18 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 14, 2022 at 12:10:18PM +0530, Bharath Rupireddy wrote:\\n> + ereport(ERROR,\\n> + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n> + errmsg(\\\"both archive_command and archive_library specified\\\"),\\n> + errdetail(\\\"Only one of archive_command,\\n> archive_library may be set.\\\")));\\n> \\n> The above errmsg looks informational. Can we just say something like\\n> below? It doesn't require errdetail as the errmsg says it all. See\\n> the other instances elsewhere [2].\\n> \\n> ereport(ERROR,\\n> (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n> errmsg(\\\"cannot specify both \\\\\\\"archive_command\\\\\\\" and\\n> \\\\\\\"archive_library\\\\\\\"\\\")));\\n\\nI modeled this after the ERROR that error_multiple_recovery_targets()\\nemits. I don't think there's really any material difference between your\\nproposal and mine, but I don't have a strong opinion.\\n\\n> 2) I think we have a problem - set archive_mode and archive_library\\n> and start the server, then set archive_command, reload the conf, see\\n> [3] - the archiver needs to error out right? The archiver gets\\n> restarted whenever archive_library changes but not when\\n> archive_command changes. I think the right place for the error is\\n> after or at the end of HandlePgArchInterrupts().\\n\\nGood catch. You are right, this is broken. I believe that we need to\\ncheck for the misconfiguration in HandlePgArchInterrupts() in addition to\\nLoadArchiveLibrary(). I will work on fixing this.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 14 Oct 2022 11:51:30 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 14, 2022 at 11:51:30AM -0700, Nathan Bossart wrote:\\n> On Fri, Oct 14, 2022 at 12:10:18PM +0530, Bharath Rupireddy wrote:\\n>> 2) I think we have a problem - set archive_mode and archive_library\\n>> and start the server, then set archive_command, reload the conf, see\\n>> [3] - the archiver needs to error out right? The archiver gets\\n>> restarted whenever archive_library changes but not when\\n>> archive_command changes. I think the right place for the error is\\n>> after or at the end of HandlePgArchInterrupts().\\n> \\n> Good catch. You are right, this is broken. I believe that we need to\\n> check for the misconfiguration in HandlePgArchInterrupts() in addition to\\n> LoadArchiveLibrary(). I will work on fixing this.\\n\\nAs promised...\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Fri, 14 Oct 2022 14:42:56 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Sat, Oct 15, 2022 at 3:13 AM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n>\\n> On Fri, Oct 14, 2022 at 11:51:30AM -0700, Nathan Bossart wrote:\\n> > On Fri, Oct 14, 2022 at 12:10:18PM +0530, Bharath Rupireddy wrote:\\n> >> 2) I think we have a problem - set archive_mode and archive_library\\n> >> and start the server, then set archive_command, reload the conf, see\\n> >> [3] - the archiver needs to error out right? The archiver gets\\n> >> restarted whenever archive_library changes but not when\\n> >> archive_command changes. I think the right place for the error is\\n> >> after or at the end of HandlePgArchInterrupts().\\n> >\\n> > Good catch. You are right, this is broken. I believe that we need to\\n> > check for the misconfiguration in HandlePgArchInterrupts() in addition to\\n> > LoadArchiveLibrary(). I will work on fixing this.\\n>\\n> As promised...\\n\\nThanks. I think that if the condition can be simplified something like\\nin the attached. It's okay to call shutdown callback twice by getting\\nrid of the comment [1] as it doesn't add any extra value or\\ninformation, it just says that we're calling shutdown callback\\nfunction. With the attached, the code is more readable and the\\nfootprint of the changes are reduced.\\n\\n[1]\\n /*\\n * Call the currently loaded archive module's shutdown callback,\\n * if one is defined.\\n */\\n call_archive_module_shutdown_callback(0, 0);\\n\\n-- \\nBharath Rupireddy\\nPostgreSQL Contributors Team\\nRDS Open Source Databases\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Sun, 16 Oct 2022 13:05:54 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"At Fri, 14 Oct 2022 14:42:56 -0700, Nathan Bossart <nathandbossart@gmail.com> wrote in \\n> As promised...\\n\\nAs the code written, when archive library is being added while archive\\ncommand is already set, archiver first emits seemingly positive\\nmessage \\\"restarting archive process because of..\\\", then errors out\\nafter the resatart and keep restarting with complaining for the wrong\\nsetting. I think we don't need the first message.\\n\\nThe ERROR always turns into FATAL, so FATAL would less confusing here,\\nmaybe.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Mon, 17 Oct 2022 13:46:39 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 17, 2022 at 01:46:39PM +0900, Kyotaro Horiguchi wrote:\\n> As the code written, when archive library is being added while archive\\n> command is already set, archiver first emits seemingly positive\\n> message \\\"restarting archive process because of..\\\", then errors out\\n> after the resatart and keep restarting with complaining for the wrong\\n> setting. I think we don't need the first message.\\n> \\n> The ERROR always turns into FATAL, so FATAL would less confusing here,\\n> maybe.\\n\\nYou mean the second message in HandlePgArchInterrupts() when\\narchiveLibChanged is false? An ERROR or a FATAL would not change much\\nas there is a proc_exit() anyway down the road.\\n\\n+ if (XLogArchiveLibrary[0] != '\\\\0' && XLogArchiveCommand[0] != '\\\\0')\\n+ ereport(ERROR,\\n+ (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n+ errmsg(\\\"both archive_command and archive_library specified\\\"),\\n+ errdetail(\\\"Only one of archive_command, archive_library may be set.\\\")));\\n\\nSo, backpedalling from upthread where Peter mentions that we should\\ncomplain if both archive_command and archive_library are set (creating\\na parallel with recovery parameters), I'd like to think that pgarch.c\\nshould have zero knowledge of what an archive_command is and should\\njust handle the library part. This makes the whole reasoning around\\nwhat pgarch.c should be much simpler, aka it just needs to know about\\narchive *libraries*, not *commands*. That's the kind of business that\\ncheck_configured_cb() is designed for, actually, as far as I\\nunderstand, or this callback could just be removed entirely for the\\nsame effect, as there would be no point in having pgarch.c do its\\nthing without archive_library or archive_command where a WARNING is\\nissued in the default case (shell_archive with no archive_command).\\n\\nAnd, by the way, this patch would prevent the existence of archive\\nmodules that need to be loaded but *want* an archive_command with\\nwhat they want to achieve. That does not strike me as a good idea if\\nwe want to have a maximum of flexibility with this facility. I think\\nthat for all that, we should put the responsability of what should be\\nset or not set directly to the modules, aka basic_archive could\\ncomplain if archive_command is set, but that does not strike me as a\\nmandatory requirement, either. It is true that archive_library has\\nbeen introduced as a way to avoid using archive_command, but the point\\nof creating a stronger dependency between both would be IMO annoying\\nin the long-term.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 17 Oct 2022 14:49:51 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 17, 2022 at 11:20 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Mon, Oct 17, 2022 at 01:46:39PM +0900, Kyotaro Horiguchi wrote:\\n> > As the code written, when archive library is being added while archive\\n> > command is already set, archiver first emits seemingly positive\\n> > message \\\"restarting archive process because of..\\\", then errors out\\n> > after the resatart and keep restarting with complaining for the wrong\\n> > setting. I think we don't need the first message.\\n> >\\n> > The ERROR always turns into FATAL, so FATAL would less confusing here,\\n> > maybe.\\n>\\n> You mean the second message in HandlePgArchInterrupts() when\\n> archiveLibChanged is false? An ERROR or a FATAL would not change much\\n> as there is a proc_exit() anyway down the road.\\n\\nYes, ERROR or FATAL it really doesn't matter, the process exits see,\\npg_re_throw(), for archiver PG_exception_stack is null.\\n2022-10-18 09:57:41.869 UTC [2479104] FATAL: both archive_command and\\narchive_library specified\\n2022-10-18 09:57:41.869 UTC [2479104] DETAIL: Only one of\\narchive_command, archive_library may be set.\\n\\nI think Kyotaro-san's concern is to place errmsg(\\\"both archive_command\\nand archive_library specified\\\"), before errmsg(\\\"restarting archiver\\nprocess because value of \\\\\\\"archive_library\\\\\\\" was changed\\\", something\\nlike the attached v4 patch.\\n\\n> + if (XLogArchiveLibrary[0] != '\\\\0' && XLogArchiveCommand[0] != '\\\\0')\\n> + ereport(ERROR,\\n> + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n> + errmsg(\\\"both archive_command and archive_library specified\\\"),\\n> + errdetail(\\\"Only one of archive_command, archive_library may be set.\\\")));\\n>\\n> So, backpedalling from upthread where Peter mentions that we should\\n> complain if both archive_command and archive_library are set (creating\\n> a parallel with recovery parameters), I'd like to think that pgarch.c\\n> should have zero knowledge of what an archive_command is and should\\n> just handle the library part. This makes the whole reasoning around\\n> what pgarch.c should be much simpler, aka it just needs to know about\\n> archive *libraries*, not *commands*.\\n\\nAre you saying that we make/treat/build shell_archive.c as a separate\\nshared library/module (instead of just an object file) and load it in\\npgarc.c? If yes, this can make pgarch.c simple.\\n\\n> That's the kind of business that\\n> check_configured_cb() is designed for, actually, as far as I\\n> understand, or this callback could just be removed entirely for the\\n> same effect, as there would be no point in having pgarch.c do its\\n> thing without archive_library or archive_command where a WARNING is\\n> issued in the default case (shell_archive with no archive_command).\\n\\nIf it's done as said above, the corresponding check_configured_cb()\\ncan deal with allowing/disallowing/misconfiguring various parameters.\\n\\n> And, by the way, this patch would prevent the existence of archive\\n> modules that need to be loaded but *want* an archive_command with\\n> what they want to achieve. That does not strike me as a good idea if\\n> we want to have a maximum of flexibility with this facility. I think\\n> that for all that, we should put the responsability of what should be\\n> set or not set directly to the modules, aka basic_archive could\\n> complain if archive_command is set, but that does not strike me as a\\n> mandatory requirement, either. It is true that archive_library has\\n> been introduced as a way to avoid using archive_command, but the point\\n> of creating a stronger dependency between both would be IMO annoying\\n> in the long-term.\\n\\nGreat thought! If the responsibility of\\nallowing/disallowing/misconfiguring various parameters is given to\\ncheck_configured_cb(), the modules can decide whether to error out or\\ndeal with it or use it.\\n\\n--\\nBharath Rupireddy\\nPostgreSQL Contributors Team\\nRDS Open Source Databases\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 18 Oct 2022 16:22:50 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"2022年10月16日(日) 16:36 Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>:\\n>\\n> On Sat, Oct 15, 2022 at 3:13 AM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n> >\\n> > On Fri, Oct 14, 2022 at 11:51:30AM -0700, Nathan Bossart wrote:\\n> > > On Fri, Oct 14, 2022 at 12:10:18PM +0530, Bharath Rupireddy wrote:\\n> > >> 2) I think we have a problem - set archive_mode and archive_library\\n> > >> and start the server, then set archive_command, reload the conf, see\\n> > >> [3] - the archiver needs to error out right? The archiver gets\\n> > >> restarted whenever archive_library changes but not when\\n> > >> archive_command changes. I think the right place for the error is\\n> > >> after or at the end of HandlePgArchInterrupts().\\n> > >\\n> > > Good catch. You are right, this is broken. I believe that we need to\\n> > > check for the misconfiguration in HandlePgArchInterrupts() in addition to\\n> > > LoadArchiveLibrary(). I will work on fixing this.\\n> >\\n> > As promised...\\n>\\n> Thanks. I think that if the condition can be simplified something like\\n> in the attached. It's okay to call shutdown callback twice by getting\\n> rid of the comment [1] as it doesn't add any extra value or\\n> information, it just says that we're calling shutdown callback\\n> function. With the attached, the code is more readable and the\\n> footprint of the changes are reduced.\\n>\\n> [1]\\n> /*\\n> * Call the currently loaded archive module's shutdown callback,\\n> * if one is defined.\\n> */\\n> call_archive_module_shutdown_callback(0, 0);\\n\\nHi\\n\\ncfbot reports the patch no longer applies [1]. As CommitFest 2022-11 is\\ncurrently underway, this would be an excellent time to update the patch.\\n\\n[1] http://cfbot.cputube.org/patch_40_3933.log\\n\\nThanks\\n\\nIan Barwick\\n\\n\\n\",\n \"msg_date\": \"Fri, 4 Nov 2022 12:05:26 +0900\",\n \"msg_from\": \"Ian Lawrence Barwick <barwick@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 04, 2022 at 12:05:26PM +0900, Ian Lawrence Barwick wrote:\\n> cfbot reports the patch no longer applies [1]. As CommitFest 2022-11 is\\n> currently underway, this would be an excellent time to update the patch.\\n\\nIndeed. Here is a new version of the patch.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\",\n \"msg_date\": \"Sat, 5 Nov 2022 13:51:54 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 17, 2022 at 02:49:51PM +0900, Michael Paquier wrote:\\n> And, by the way, this patch would prevent the existence of archive\\n> modules that need to be loaded but *want* an archive_command with\\n> what they want to achieve. That does not strike me as a good idea if\\n> we want to have a maximum of flexibility with this facility.\\n\\nSuch a module could define a custom GUC that accepts a shell command. I\\ndon't think we should overload the meaning of archive_command based on the\\nwhims of whatever archive module is loaded. Besides the potential end-user\\nconfusion, your archive_command might be unexpectedly used incorrectly if\\nyou forget to set archive_library.\\n\\nPerhaps we could eventually move the archive_command functionality to a\\ncontrib module (i.e., \\\"shell_archive\\\") so that users must always set\\narchive_library. But until then, I suspect it's better to treat modules\\nand commands as two separate interfaces to ease migration from older major\\nversions (even though archive_command is now essentially a built-in archive\\nmodule).\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Sat, 5 Nov 2022 14:08:58 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 05, 2022 at 02:08:58PM -0700, Nathan Bossart wrote:\\n> Such a module could define a custom GUC that accepts a shell command. I\\n> don't think we should overload the meaning of archive_command based on the\\n> whims of whatever archive module is loaded. Besides the potential end-user\\n> confusion, your archive_command might be unexpectedly used incorrectly if\\n> you forget to set archive_library.\\n\\nWhile mostly copying the logic from shell_archive.c to build the\\ncommand to execute (aka shell_archive_file), which is not great as\\nwell. But well, perhaps my whole line of argument is just moot.. \\n\\n> Perhaps we could eventually move the archive_command functionality to a\\n> contrib module (i.e., \\\"shell_archive\\\") so that users must always set\\n> archive_library. But until then, I suspect it's better to treat modules\\n> and commands as two separate interfaces to ease migration from older major\\n> versions (even though archive_command is now essentially a built-in archive\\n> module).\\n\\nI agree that this is a fine long-term goal, removing all traces of the\\narchive_command from the backend core code. This is actually an\\nargument in favor of having no traces of XLogArchiveCommand in\\npgarch.c, no? ;p\\n\\nI am not sure how long we should wait before being able to do that,\\nperhaps a couple of years of least? I'd like to think the sooner the\\nbetter (like v17?) but we are usually conservative, and the removal of\\nthe exclusive backup mode took 5~6 years if I recall correctly..\\n--\\nMichael\",\n \"msg_date\": \"Mon, 7 Nov 2022 15:20:31 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 07, 2022 at 03:20:31PM +0900, Michael Paquier wrote:\\n> On Sat, Nov 05, 2022 at 02:08:58PM -0700, Nathan Bossart wrote:\\n>> Perhaps we could eventually move the archive_command functionality to a\\n>> contrib module (i.e., \\\"shell_archive\\\") so that users must always set\\n>> archive_library. But until then, I suspect it's better to treat modules\\n>> and commands as two separate interfaces to ease migration from older major\\n>> versions (even though archive_command is now essentially a built-in archive\\n>> module).\\n> \\n> I agree that this is a fine long-term goal, removing all traces of the\\n> archive_command from the backend core code. This is actually an\\n> argument in favor of having no traces of XLogArchiveCommand in\\n> pgarch.c, no? ;p\\n\\nIndeed.\\n\\n> I am not sure how long we should wait before being able to do that,\\n> perhaps a couple of years of least? I'd like to think the sooner the\\n> better (like v17?) but we are usually conservative, and the removal of\\n> the exclusive backup mode took 5~6 years if I recall correctly..\\n\\nYeah, I imagine we'd need to mark it as deprecated-and-to-be-removed for\\nseveral years first.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Nov 2022 13:15:47 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 05.11.22 21:51, Nathan Bossart wrote:\\n> On Fri, Nov 04, 2022 at 12:05:26PM +0900, Ian Lawrence Barwick wrote:\\n>> cfbot reports the patch no longer applies [1]. As CommitFest 2022-11 is\\n>> currently underway, this would be an excellent time to update the patch.\\n> \\n> Indeed. Here is a new version of the patch.\\n\\nI have committed this to master.\\n\\nThe surrounding code has changed a bit between PG15 and master, so if we \\nwanted to backpatch this, we'd need another patch from you. However, at \\nthis point, I'm content to just leave it be in PG15.\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 15 Nov 2022 10:31:44 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 15, 2022 at 10:31:44AM +0100, Peter Eisentraut wrote:\\n> I have committed this to master.\\n\\nThanks!\\n\\n> The surrounding code has changed a bit between PG15 and master, so if we\\n> wanted to backpatch this, we'd need another patch from you. However, at\\n> this point, I'm content to just leave it be in PG15.\\n\\nSounds good to me.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 15 Nov 2022 08:45:13 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On 2022-Nov-15, Nathan Bossart wrote:\\n\\n> On Tue, Nov 15, 2022 at 10:31:44AM +0100, Peter Eisentraut wrote:\\n\\n> > The surrounding code has changed a bit between PG15 and master, so if we\\n> > wanted to backpatch this, we'd need another patch from you. However, at\\n> > this point, I'm content to just leave it be in PG15.\\n> \\n> Sounds good to me.\\n\\nHmm, really? It seems to me that we will have two slightly different\\nbehaviors in 15 and master, which may be confusing later on. I think\\nit'd be better to make them both work identically.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\\"Industry suffers from the managerial dogma that for the sake of stability\\nand continuity, the company should be independent of the competence of\\nindividual employees.\\\" (E. Dijkstra)\\n\\n\\n\",\n \"msg_date\": \"Tue, 15 Nov 2022 18:14:25 +0100\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 15, 2022 at 06:14:25PM +0100, Alvaro Herrera wrote:\\n> On 2022-Nov-15, Nathan Bossart wrote:\\n> \\n>> On Tue, Nov 15, 2022 at 10:31:44AM +0100, Peter Eisentraut wrote:\\n> \\n>> > The surrounding code has changed a bit between PG15 and master, so if we\\n>> > wanted to backpatch this, we'd need another patch from you. However, at\\n>> > this point, I'm content to just leave it be in PG15.\\n>> \\n>> Sounds good to me.\\n> \\n> Hmm, really? It seems to me that we will have two slightly different\\n> behaviors in 15 and master, which may be confusing later on. I think\\n> it'd be better to make them both work identically.\\n\\nI don't have a strong opinion either way. While consistency between v15\\nand master seems nice, the behavior change might not be appropriate for a\\nminor release. BTW I was able to cherry-pick the committed patch to v15\\nwithout any changes. Peter, could you clarify what changes you'd like to\\nsee in a back-patched version?\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 15 Nov 2022 12:57:49 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 15, 2022 at 12:57:49PM -0800, Nathan Bossart wrote:\\n> On Tue, Nov 15, 2022 at 06:14:25PM +0100, Alvaro Herrera wrote:\\n>> Hmm, really? It seems to me that we will have two slightly different\\n>> behaviors in 15 and master, which may be confusing later on. I think\\n>> it'd be better to make them both work identically.\\n> \\n> I don't have a strong opinion either way. While consistency between v15\\n> and master seems nice, the behavior change might not be appropriate for a\\n> minor release. BTW I was able to cherry-pick the committed patch to v15\\n> without any changes. Peter, could you clarify what changes you'd like to\\n> see in a back-patched version?\\n\\nFWIW, I am not sure that I would have done d627ce3 as I already\\nmentioned upthread as the library loading should not be related to\\narchive_command. If there is support more support in doing that, I am\\nfine to withdraw, but the behavior between HEAD and REL_15_STABLE\\nought to be consistent.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 16 Nov 2022 06:51:03 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: archive modules\"\n }\n]","isConversational":false}}},{"rowIdx":53036,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nThe global variable ThisTimeLineID is rather confusing. During\nrecovery, in the startup process, when we're reading a WAL record, it\nis the timeline of the WAL record we are trying to read or have just\nread, except when we're trying to read the initial checkpoint record,\nwhen it's zero. In other processes, it's typically 0 throughout\nrecovery, but sometimes it's not, because for example\nCreateRestartPoint temporarily sets it to the timeline from which\nwe're replaying WAL, since there's no other way to get\nRemoveOldXlogFiles() to recycle files onto the right timeline, or\nPreallocXlogFiles() to preallocate onto the right timeline. Similarly,\nwalreceiver and walsender find it convenient to make ThisTimeLineID\nthe timeline from which WAL is being replayed or at least the one from\nwhich it was being replayed at last check. logicalfuncs.c and\nslotfuncs.c set the global variable sometimes too, apparently just for\nfun, as the value doesn't seem to get used for anything. In normal\nrunning, once the startup process exits, the next call to\nRecoveryInProgress() sets the value to the timeline into which new WAL\ncan be inserted. Note also that all of this is different from\nXLogCtl->ThisTimeLineID, which is set at the end of recovery and thus,\nin the startup process, generally different from ThisTImeLineID, but\nin other processes, generally the same as ThisTimeLineID.\n\nAt the risk of stating the obvious, using the same variable for\ndifferent purposes at different times is not a great programming\npractice. Commits 2f5c4397c39dea49c5608ba583868e26d767fc32 and\n902a2c280012557b85c7e0fce3f6f0e355cb2d69 show that the possibility of\nprogrammer error is not zero, even though neither of those issues are\nserious. Moreover, the global variable itself seems to do nothing but\ninvite programming errors. The name itself is a disaster. What is\n\"this\" timeline ID? Well, uh, the right one, of course! We should be\nmore clear about what we mean: the timeline into which we are\ninserting, the timeline from which we are replaying, the timeline from\nwhich we are performing logical decoding. I suspect that having a\nglobal variable here isn't even really saving us anything much, as a\nvalue that does not change can be read from shared memory without a\nlock.\n\nI would like to clean this up. Attached is a series of patches which\ntry to do that. For ease of review, this is separated out into quite a\nfew separate patches, but most likely we'd combine some of them for\ncommit. Patches 0001 through 0004 eliminate all use of the global\nvariable ThisTimeLineID outside of xlog.c, and patch 0005 makes it\n\"static\" so that it ceases to be visible outside of xlog.c. Patches\n0006 to 0010 clean up uses of ThisTimeLineID itself, passing it around\nas a function parameter, or otherwise arranging to fetch the relevant\ntimeline ID from someplace sensible rather than using the global\nvariable as a scratchpad. Finally, patch 0011 deletes the global\nvariable.\n\nI have not made a serious attempt to clear up all of the\nterminological confusion created by the term ThisTimeLineID, which\nalso appears as part of some structs, so you'll still see that name in\nthe source code after applying these patches, just not as the name of\na global variable. I have, however, used names like insertTLI and\nreplayTLI in many places changed by the patch, and I think that makes\nit significantly more clear which timeline is being discussed in each\ncase. In some places I have endeavored to improve the comments. There\nis doubtless more that could be done here, but I think this is a\nfairly respectable start.\n\nOpinions welcome.\n\nThanks,\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com","msg_date":"Mon, 1 Nov 2021 15:16:27 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"removing global variable ThisTimeLineID"},{"msg_contents":"On Mon, Nov 01, 2021 at 03:16:27PM -0400, Robert Haas wrote:\n> At the risk of stating the obvious, using the same variable for\n> different purposes at different times is not a great programming\n> practice. Commits 2f5c4397c39dea49c5608ba583868e26d767fc32 and\n> 902a2c280012557b85c7e0fce3f6f0e355cb2d69 show that the possibility of\n> programmer error is not zero, even though neither of those issues are\n> serious.\n\nThis can lead to more serious issues, see 595b9cba as recent example.\nI agree that getting rid of it in the long term is appealing.\n\n> Moreover, the global variable itself seems to do nothing but\n> invite programming errors. The name itself is a disaster. What is\n> \"this\" timeline ID? Well, uh, the right one, of course! We should be\n> more clear about what we mean: the timeline into which we are\n> inserting, the timeline from which we are replaying, the timeline from\n> which we are performing logical decoding. I suspect that having a\n> global variable here isn't even really saving us anything much, as a\n> value that does not change can be read from shared memory without a\n> lock.\n\nSome callbacks related to the WAL reader to read a patch have\nintroduced a dependency to ThisTimelineID as a mean to be used for\nlogical decoding on standbys. This is discussed a bit on this thread,\nfor example:\nhttps://www.postgresql.org/message-id/ddc31aa9-8083-58b7-0b47-e371cd4c705b@oss.nttdata.com\n\nTracking properly the TLI in logical decoding was mentioned by Andres\nmultiple times on -hackers, and this has not been really done. Using\nThisTimelineID for this purpose is something I've found confusing\nmyself, FWIW, so this is a step in the good direction.\n\n> I would like to clean this up. Attached is a series of patches which\n> try to do that. For ease of review, this is separated out into quite a\n> few separate patches, but most likely we'd combine some of them for\n> commit. Patches 0001 through 0004 eliminate all use of the global\n> variable ThisTimeLineID outside of xlog.c, and patch 0005 makes it\n> \"static\" so that it ceases to be visible outside of xlog.c. Patches\n> 0006 to 0010 clean up uses of ThisTimeLineID itself, passing it around\n> as a function parameter, or otherwise arranging to fetch the relevant\n> timeline ID from someplace sensible rather than using the global\n> variable as a scratchpad. Finally, patch 0011 deletes the global\n> variable.\n> \n> I have not made a serious attempt to clear up all of the\n> terminological confusion created by the term ThisTimeLineID, which\n> also appears as part of some structs, so you'll still see that name in\n> the source code after applying these patches, just not as the name of\n> a global variable. I have, however, used names like insertTLI and\n> replayTLI in many places changed by the patch, and I think that makes\n> it significantly more clear which timeline is being discussed in each\n> case. In some places I have endeavored to improve the comments. There\n> is doubtless more that could be done here, but I think this is a\n> fairly respectable start.\n\nThanks for splitting things this way. This helps a lot.\n\n0001 looks fair.\n\n+ /*\n+ * If we're writing and flushing WAL, the time line can't be changing,\n+ * so no lock is required.\n+ */\n+ if (insertTLI)\n+ *insertTLI = XLogCtl->ThisTimeLineID;\nIn 0002, there is no downside in putting that in the spinlock section\neither, no? It seems to me that we should be able to call this one\neven in recovery, as flush LSNs exist while in recovery.\n\n+TimeLineID\n+GetWALInsertionTimeLine(void)\n+{\n+ Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);\nOkay, that's cheaper.\n\nThe series 0003-0006 seemed rather fine, at quick glance.\n\nIn 0007, XLogFileClose() should reset openLogTLI. The same can be\nsaid in BootStrapXLOG() before creating the control file. It may be\ncleaner here to introduce an InvalidTimelineId, as discussed a couple\nof days ago.\n\nSome comments at the top of recoveryTargetTLIRequested need a refresh\nwith their references to ThisTimelineID.\n\nThe handling of XLogCtl->ThisTimeLineID still seems a bit blurry when\nit comes to timeline switches because we don't update it while in\nrecovery when replaying a end-of-recovery record. This could at least\nbe documented better.\n--\nMichael","msg_date":"Tue, 2 Nov 2021 12:33:32 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Mon, Nov 1, 2021 at 11:33 PM Michael Paquier <michael@paquier.xyz> wrote:\n> + /*\n> + * If we're writing and flushing WAL, the time line can't be changing,\n> + * so no lock is required.\n> + */\n> + if (insertTLI)\n> + *insertTLI = XLogCtl->ThisTimeLineID;\n> In 0002, there is no downside in putting that in the spinlock section\n> either, no? It seems to me that we should be able to call this one\n> even in recovery, as flush LSNs exist while in recovery.\n\nI don't think it matters very much from a performance point of view,\nalthough putting a branch inside of a spinlock-protected section may\nbe undesirable. My bigger issues with this kind of idea is that we\ndon't want to use spinlocks as a \"security blanket\" (as in Linus from\nPeanuts). Every time we use a spinlock, or any other kind of locking\nmechanism, we should have a clear idea what we're protecting ourselves\nagainst. I'm quite suspicious that there are a number of places where\nwe're taking spinlocks in xlog.c just to feel virtuous, and not\nbecause it does anything useful, and I don't particularly want to\nincrease the number of such places.\n\nAlso, XLogCtl->ThisTimeLineID isn't set until the end of recovery, so\ncalling this function during recovery would be a mistake. There seem\nto be a number of interface functions in xlog.c that should only be\ncalled when not in recovery, and neither their names nor their\ncomments make that clear. I wasn't bold enough to go label all the\nones I think fall into that category, but maybe we should. Honestly,\nthe naming of things in this file is annoyingly bad in general. My\nfavorite example, found during this project, is:\n\n#define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))\n\nIt's not good enough to have one name that's difficult to understand\nor remember or capitalize properly -- let's have TWO -- for the same\nthing!\n\n> In 0007, XLogFileClose() should reset openLogTLI. The same can be\n> said in BootStrapXLOG() before creating the control file. It may be\n> cleaner here to introduce an InvalidTimelineId, as discussed a couple\n> of days ago.\n\nI thought about that, but I think it's pointless. I think we can just\nsay that if openLogFile == -1, openLogSegNo and openLogTLI are\nmeaningless. I don't think we're currently resetting openLogSegNo to\nan invalid value either, so I don't think openLogTLI should be treated\ndifferently. I'm not opposed to introducing InvalidTimeLineID on\nprinciple, and I looked for a way of doing it in this patch set, but\nit didn't seem all that valuable, and I feel that someone who cares\nabout it can do it as a separate effort after I commit these.\nHonestly, I think these patches greatly reduce the need to be afraid\nof leaving the TLI unset, because they remove the crutch of hoping\nthat ThisTimeLineID is initialized to the proper value. You actually\nhave to think about which TLI you are going to pass. Now I'm very sure\nfurther improvement is possible, but I think this patch set is complex\nenough already, and I'd like to get it committed before contemplating\nany other work in this area.\n\n> Some comments at the top of recoveryTargetTLIRequested need a refresh\n> with their references to ThisTimelineID.\n\nAfter thinking this over, I think we should just delete the entire\nfirst paragraph, so that the comments start like this:\n\n/*\n * recoveryTargetTimeLineGoal: what the user requested, if any\n *\n * recoveryTargetTLIRequested: numeric value of requested timeline, if constant\n\nI don't see that we'd be losing anything that way. When you look at\nthat paragraph now, the first sentence is just confusing the issue,\nbecause we actually don't care about XLogCtl->ThisTimeLineID during\nrecovery, because it's not set. We do care about plain old\nThisTimeLineID, but that's going to be demoted to a local variable by\nthese patches, and there just doesn't seem to be any reason to discuss\nit here. The second sentence is wrong already, because the rmgr\nroutines do not rely on ThisTimeLineID. And the third sentence is also\nwrong and/or confusing, because not all of the variables that follow\nare TLIs -- only 3 of 5 are. I don't actually think there's any useful\nintroduction we need to provide here; we just need to tell people what\nthe variables we're about to declare do.\n\n> The handling of XLogCtl->ThisTimeLineID still seems a bit blurry when\n> it comes to timeline switches because we don't update it while in\n> recovery when replaying a end-of-recovery record. This could at least\n> be documented better.\n\nWe don't update it during recovery at all. There's exactly one\nassignment statement for that variable and it's here:\n\n /* Save the selected TimeLineID in shared memory, too */\n XLogCtl->ThisTimeLineID = ThisTimeLineID;\n XLogCtl->PrevTimeLineID = PrevTimeLineID;\n\nThat's after the main redo loop completes.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 2 Nov 2021 08:45:57 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Tue, Nov 2, 2021 at 12:46 AM Robert Haas <robertmhaas@gmail.com> wrote:\n>\n> [...]\n> I would like to clean this up. Attached is a series of patches which\n> try to do that. For ease of review, this is separated out into quite a\n> few separate patches, but most likely we'd combine some of them for\n> commit. Patches 0001 through 0004 eliminate all use of the global\n> variable ThisTimeLineID outside of xlog.c, and patch 0005 makes it\n> \"static\" so that it ceases to be visible outside of xlog.c. Patches\n> 0006 to 0010 clean up uses of ThisTimeLineID itself, passing it around\n> as a function parameter, or otherwise arranging to fetch the relevant\n> timeline ID from someplace sensible rather than using the global\n> variable as a scratchpad. Finally, patch 0011 deletes the global\n> variable.\n>\n> I have not made a serious attempt to clear up all of the\n> terminological confusion created by the term ThisTimeLineID, which\n> also appears as part of some structs, so you'll still see that name in\n> the source code after applying these patches, just not as the name of\n> a global variable. I have, however, used names like insertTLI and\n> replayTLI in many places changed by the patch, and I think that makes\n> it significantly more clear which timeline is being discussed in each\n> case. In some places I have endeavored to improve the comments. There\n> is doubtless more that could be done here, but I think this is a\n> fairly respectable start.\n>\n> Opinions welcome.\n>\n\nI had a plan to look at these patches closely, but unfortunately,\ndidn't get enough time; and might not be able to spend time in the\nrest of this week. Here are a few thoughts that I had in the initial\ngo-through, but that may or may not sound very interesting:\n\n0002:\n-GetFlushRecPtr(void)\n+GetFlushRecPtr(TimeLineID *insertTLI)\n\nShould we rename this argument to more generic as \"tli\", like\nGetStandbyFlushRecPtr, since the caller in 003 patch trying to fetch a\nTLI that means different for them, e.g. currTLI, FlushTLI, etc.\n\n\n0004:\n static int\n-XLogFileInitInternal(XLogSegNo logsegno, bool *added, char *path)\n+XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,\n+ bool *added, char *path)\n\n int\n-XLogFileInit(XLogSegNo logsegno)\n+XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)\n {\n\nHow about logsegtli instead, to be inlined with logsegno ?\n\n\n0007:\n static XLogSegNo openLogSegNo = 0;\n+static TimeLineID openLogTLI = 0;\n\nSimilarly, openLogSegTLI ?\n\n\n0008:\n+ /*\n+ * Given that we're not in recovery, ThisTimeLineID is set and can't\n+ * change, so we can read it without a lock.\n+ */\n+ insertTLI = XLogCtl->ThisTimeLineID;\n\nCan't GetWALInsertionTimeLine() called instead? I guess the reason is\nto avoid the unnecessary overhead involved in the function call. (Same\nat a few other places.)\n\nRegards,\nAmul\n\n\n","msg_date":"Wed, 3 Nov 2021 18:42:22 +0530","msg_from":"Amul Sul <sulamul@gmail.com>","msg_from_op":false,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Wed, Nov 3, 2021 at 9:12 AM Amul Sul <sulamul@gmail.com> wrote:\n> 0002:\n> -GetFlushRecPtr(void)\n> +GetFlushRecPtr(TimeLineID *insertTLI)\n>\n> Should we rename this argument to more generic as \"tli\", like\n> GetStandbyFlushRecPtr, since the caller in 003 patch trying to fetch a\n> TLI that means different for them, e.g. currTLI, FlushTLI, etc.\n\nIt's possible that more consistency would be better here, but I don't\nthink making this name more generic is going in the right direction.\nIf somebody gets the TLI using this function, it's the timeline into\nwhich a system not in recovery is inserting WAL, which is the same\ntimeline to which WAL is being flushed. I think it's important for\nthis name to try to make that clear. It doesn't really matter if\nsomeone treats the insertTLI as the writeTLI or the flushTLI since on\na system in production they're all the same - but they must not\nconfuse it with, say, the replayTLI.\n\n> How about logsegtli instead, to be inlined with logsegno ?\n> Similarly, openLogSegTLI ?\n\nHmm, well I guess it depends on how you think the words group. If you\nlogsegno means \"the number of the log segment\" then it would make\nsense to have logsegli to mean \"the tli of the log segment.\" But I\nthink of logsegno as meaning \"the segment number of the log file,\" so\nit makes more sense to have logtli to mean \"the TLI of the log file\".\nI think it's definitely not a \"log segment file\" - just a \"log file\".\n\n> Can't GetWALInsertionTimeLine() called instead? I guess the reason is\n> to avoid the unnecessary overhead involved in the function call. (Same\n> at a few other places.)\n\nThat, plus to avoid failing the assertion in that function. As we've\ndiscussed on the ALTER SYSTEM READ ONLY thread, xlog.c itself inserts\nsome WAL before recovery is officially over.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 3 Nov 2021 10:13:55 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"I looked at these in a very quick skim. I agree that this is a\ngood step in a good direction.\n\nI 'git rebase -x'd this series in order to compile and run the tests on\neach patch individually. There are no compiler warnings anywhere and no\ntest failures.\n\n\n-- \nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\n\"La victoria es para quien se atreve a estar solo\"\n\n\n","msg_date":"Fri, 5 Nov 2021 10:49:44 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Fri, Nov 5, 2021 at 9:49 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> I looked at these in a very quick skim. I agree that this is a\n> good step in a good direction.\n\nCool. I have now committed these. I grouped them into just 3 commits\nrather than having 11 separate commits as I did in my earlier posting,\nbut the content is the same.\n\n> I 'git rebase -x'd this series in order to compile and run the tests on\n> each patch individually. There are no compiler warnings anywhere and no\n> test failures.\n\nThanks for checking. It's hard to tell whether the stuff that works on\nmy machine will work everywhere, unless someone tests it.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 5 Nov 2021 12:58:04 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Tue, Nov 02, 2021 at 08:45:57AM -0400, Robert Haas wrote:\n> Also, XLogCtl->ThisTimeLineID isn't set until the end of recovery, so\n> calling this function during recovery would be a mistake. There seem\n> to be a number of interface functions in xlog.c that should only be\n> called when not in recovery, and neither their names nor their\n> comments make that clear. I wasn't bold enough to go label all the\n> ones I think fall into that category, but maybe we should. \n\nI got to wonder whether it would be better to add in GetFlushRecPtr()\nthe same assertion as GetWALInsertionTimeLine(). All the in-core\ncallers of this routine already assume that, and it would be buggy if\none passes down insertTLI to get the current TLI.\n\n> I thought about that, but I think it's pointless. I think we can just\n> say that if openLogFile == -1, openLogSegNo and openLogTLI are\n> meaningless. I don't think we're currently resetting openLogSegNo to\n> an invalid value either, so I don't think openLogTLI should be treated\n> differently.\n\nWouldn't it be better to at least document that as a comment rather\nthan letting the reader guess it, then? I agree that this is a minor\npoint, though.\n\n> I'm not opposed to introducing InvalidTimeLineID on\n> principle, and I looked for a way of doing it in this patch set, but\n> it didn't seem all that valuable, and I feel that someone who cares\n> about it can do it as a separate effort after I commit these.\n\nNo issues from me. I have bumped into a case just at the end of last\nweek where I wanted a better way to tell if a TLI is invalid rather\nthan leaving things to 0, FWIW.\n\n>> The handling of XLogCtl->ThisTimeLineID still seems a bit blurry when\n>> it comes to timeline switches because we don't update it while in\n>> recovery when replaying a end-of-recovery record. This could at least\n>> be documented better.\n> \n> We don't update it during recovery at all. There's exactly one\n> assignment statement for that variable and it's here:\n\nIndeed. It looks like my reading was sloppy here.\n--\nMichael","msg_date":"Mon, 8 Nov 2021 13:24:10 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Sun, Nov 7, 2021 at 11:24 PM Michael Paquier <michael@paquier.xyz> wrote:\n> I got to wonder whether it would be better to add in GetFlushRecPtr()\n> the same assertion as GetWALInsertionTimeLine(). All the in-core\n> callers of this routine already assume that, and it would be buggy if\n> one passes down insertTLI to get the current TLI.\n\nPerhaps. That's related to the point I made before, that it might be a\ngood idea to be more clear about which of these functions are intended\nto be used in recovery and which ones are intended to be used in\nnormal running. I don't rule out the possibility of doing some more\nresearch here and tightening that up, but I don't want to go start\ntweaking things unless I'm sure I know what I'm talking about. This\nstuff is so confusing that I don't want to take any risk of making\nchanges that turn out to be incorrect.\n\n> > I thought about that, but I think it's pointless. I think we can just\n> > say that if openLogFile == -1, openLogSegNo and openLogTLI are\n> > meaningless. I don't think we're currently resetting openLogSegNo to\n> > an invalid value either, so I don't think openLogTLI should be treated\n> > differently.\n>\n> Wouldn't it be better to at least document that as a comment rather\n> than letting the reader guess it, then? I agree that this is a minor\n> point, though.\n\nI don't think it's very confusing, but if you want to improve the\ncomments, go for it!\n\n> >> The handling of XLogCtl->ThisTimeLineID still seems a bit blurry when\n> >> it comes to timeline switches because we don't update it while in\n> >> recovery when replaying a end-of-recovery record. This could at least\n> >> be documented better.\n> >\n> > We don't update it during recovery at all. There's exactly one\n> > assignment statement for that variable and it's here:\n>\n> Indeed. It looks like my reading was sloppy here.\n\nI can hardly blame you. There are comments all over the place claiming\nthat this thing and that thing are the same when they're only\nsometimes the same. Or, as in this case, things that are named the\nsame when they're actually different. I'm actually tempted to make a\npass over the comments related to this topic and just try to make\nthings more comprehensible. It's super-confusing as things stand.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 8 Nov 2021 08:27:06 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Mon, Nov 8, 2021 at 8:27 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> Perhaps. That's related to the point I made before, that it might be a\n> good idea to be more clear about which of these functions are intended\n> to be used in recovery and which ones are intended to be used in\n> normal running. I don't rule out the possibility of doing some more\n> research here and tightening that up, but I don't want to go start\n> tweaking things unless I'm sure I know what I'm talking about. This\n> stuff is so confusing that I don't want to take any risk of making\n> changes that turn out to be incorrect.\n\nWell I went through and it seems to be OK: all the existing callers of\nthat function first verify that we're not in recovery. The patch to\nmake logical decoding work in standby mode might change that, but as\nof now it's so. So in the attached patch, I have added the assertion\nand comments that you have proposed there. I also removed the\nmisleading comment we discussed before. In addition to that, I renamed\nthe local variable ThisTimeLineID to replayTLI, and I also renamed\nXLogCtl->ThisTimeLineID to InsertTimeLineID, which I think should make\nthings clearer: imagine if things that were used for different\npurposes had different names!\n\nEven with this patch, the name ThisTimeLineID is still used for\nmultiple purposes. It remains part of the CheckPoint struct, and also\npart of the xl_end_of_recovery struct. But in both of those cases, the\nname ThisTimeLineID actually makes sense, because what we're thinking\nabout is whether there's a timeline change, so we have ThisTimeLineID\nand PrevTimeLineID and it seems fairly clear what that's supposed to\nmean.\n\nThoughts?\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com","msg_date":"Mon, 8 Nov 2021 12:49:52 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Mon, Nov 08, 2021 at 12:49:52PM -0500, Robert Haas wrote:\n> Even with this patch, the name ThisTimeLineID is still used for\n> multiple purposes. It remains part of the CheckPoint struct, and also\n> part of the xl_end_of_recovery struct. But in both of those cases, the\n> name ThisTimeLineID actually makes sense, because what we're thinking\n> about is whether there's a timeline change, so we have ThisTimeLineID\n> and PrevTimeLineID and it seems fairly clear what that's supposed to\n> mean.\n\nI got to wonder if it would be better to bite the bullet here for\nthese two looking at your patch, but I am fine to keep things as they\nare, as well.\n\n> Thoughts?\n\nI think that this patch is an improvement.\n\n@@ -6686,8 +6682,8 @@ StartupXLOG(void)\n- TimeLineID ThisTimeLineID,\n- PrevTimeLineID;\n+ TimeLineID replayTLI,\n+ newTLI;\nOne problem with newTLI is that this conflicts with the declaration\naround 7663 in xlog.c, where we check after a TLI switch when\nreplaying such a record. Perhaps this could be named newInsertTLI,\nfor example.\n--\nMichael","msg_date":"Tue, 9 Nov 2021 12:30:53 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Mon, Nov 8, 2021 at 10:31 PM Michael Paquier <michael@paquier.xyz> wrote:\n> I think that this patch is an improvement.\n\nCool.\n\n> @@ -6686,8 +6682,8 @@ StartupXLOG(void)\n> - TimeLineID ThisTimeLineID,\n> - PrevTimeLineID;\n> + TimeLineID replayTLI,\n> + newTLI;\n> One problem with newTLI is that this conflicts with the declaration\n> around 7663 in xlog.c, where we check after a TLI switch when\n> replaying such a record. Perhaps this could be named newInsertTLI,\n> for example.\n\nThat's a good point. However, since I think newTLI is already in use\nin some of the functions StartupXLOG() calls, and since I think it\nwould be good to use the same name in StartupXLOG() as in the\nfunctions that it calls, what I'd prefer to do is rename the newTLI\nvariable in the inner scope to newReplayTLI, as in the attached v2.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com","msg_date":"Tue, 9 Nov 2021 14:15:51 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Tue, Nov 09, 2021 at 02:15:51PM -0500, Robert Haas wrote:\n> That's a good point. However, since I think newTLI is already in use\n> in some of the functions StartupXLOG() calls, and since I think it\n> would be good to use the same name in StartupXLOG() as in the\n> functions that it calls, what I'd prefer to do is rename the newTLI\n> variable in the inner scope to newReplayTLI, as in the attached v2.\n\nWFM. Thanks.\n--\nMichael","msg_date":"Wed, 10 Nov 2021 09:41:16 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Tue, Nov 9, 2021 at 7:41 PM Michael Paquier <michael@paquier.xyz> wrote:\n> On Tue, Nov 09, 2021 at 02:15:51PM -0500, Robert Haas wrote:\n> > That's a good point. However, since I think newTLI is already in use\n> > in some of the functions StartupXLOG() calls, and since I think it\n> > would be good to use the same name in StartupXLOG() as in the\n> > functions that it calls, what I'd prefer to do is rename the newTLI\n> > variable in the inner scope to newReplayTLI, as in the attached v2.\n>\n> WFM. Thanks.\n\nCool. Committed that way.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 10 Nov 2021 09:52:06 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"Hi,\n\nOn 11/8/21 6:49 PM, Robert Haas wrote:\n> On Mon, Nov 8, 2021 at 8:27 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> Well I went through and it seems to be OK: all the existing callers of\n> that function first verify that we're not in recovery. The patch to\n> make logical decoding work in standby mode might change that,\n\nIndeed, I think the logical decoding on standby patch just needs to \nchange the Assert in GetWALInsertionTimeLine() to check \nSharedRecoveryState is RECOVERY_STATE_DONE or RECOVERY_STATE_ARCHIVE.\n\nWould that make sense from your point of view to add the check on \nRECOVERY_STATE_ARCHIVE or do you think it would need to be more \n\"sophisticated\"?\n\nThanks\n\nBertrand\n\n\n\n","msg_date":"Tue, 23 Nov 2021 10:36:16 +0100","msg_from":"\"Drouvot, Bertrand\" <bdrouvot@amazon.com>","msg_from_op":false,"msg_subject":"Re: removing global variable ThisTimeLineID"},{"msg_contents":"On Tue, Nov 23, 2021 at 4:36 AM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\n> Indeed, I think the logical decoding on standby patch just needs to\n> change the Assert in GetWALInsertionTimeLine() to check\n> SharedRecoveryState is RECOVERY_STATE_DONE or RECOVERY_STATE_ARCHIVE.\n>\n> Would that make sense from your point of view to add the check on\n> RECOVERY_STATE_ARCHIVE or do you think it would need to be more\n> \"sophisticated\"?\n\nUnless I'm confused, that would just be incorrect.\nGetWALInsertionTimeLine() just returns InsertTimeLineID, which won't\nbe initialized during archive recovery. I discussed this point a bit\non some other thread with Andres. It's possible that all that you need\nto do here is determine the replayTLI using e.g. GetXLogReplayRecPtr.\nBut I don't really see why that should be correct:\nread_local_xlog_page() has some kind of wait-and-retry logic and I'm\nnot clear why we don't need the same thing here. After all, if we're\non the primary and we determine that sendTimeLineHistoric = false and\nsendTimeLine = whatever, neither of those values can become incorrect\nafterward. But on a standby that can certainly happen, if an upstream\nserver is promoted. Note that the comment in read_local_xlog_page()\nsays this:\n\n * If that happens after our caller determined the TLI but before\n * we actually read the xlog page, we might still try to read from the\n * old (now renamed) segment and fail. There's not much we can do\n * about this, but it can only happen when we're a leaf of a cascading\n * standby whose primary gets promoted while we're decoding, so a\n * one-off ERROR isn't too bad.\n\nThat seems to imply that the retry loop here is designed, at least in\npart, to protect against exactly this kind of scenario.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 23 Nov 2021 08:30:44 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: removing global variable ThisTimeLineID"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nThe global variable ThisTimeLineID is rather confusing. During\\nrecovery, in the startup process, when we're reading a WAL record, it\\nis the timeline of the WAL record we are trying to read or have just\\nread, except when we're trying to read the initial checkpoint record,\\nwhen it's zero. In other processes, it's typically 0 throughout\\nrecovery, but sometimes it's not, because for example\\nCreateRestartPoint temporarily sets it to the timeline from which\\nwe're replaying WAL, since there's no other way to get\\nRemoveOldXlogFiles() to recycle files onto the right timeline, or\\nPreallocXlogFiles() to preallocate onto the right timeline. Similarly,\\nwalreceiver and walsender find it convenient to make ThisTimeLineID\\nthe timeline from which WAL is being replayed or at least the one from\\nwhich it was being replayed at last check. logicalfuncs.c and\\nslotfuncs.c set the global variable sometimes too, apparently just for\\nfun, as the value doesn't seem to get used for anything. In normal\\nrunning, once the startup process exits, the next call to\\nRecoveryInProgress() sets the value to the timeline into which new WAL\\ncan be inserted. Note also that all of this is different from\\nXLogCtl->ThisTimeLineID, which is set at the end of recovery and thus,\\nin the startup process, generally different from ThisTImeLineID, but\\nin other processes, generally the same as ThisTimeLineID.\\n\\nAt the risk of stating the obvious, using the same variable for\\ndifferent purposes at different times is not a great programming\\npractice. Commits 2f5c4397c39dea49c5608ba583868e26d767fc32 and\\n902a2c280012557b85c7e0fce3f6f0e355cb2d69 show that the possibility of\\nprogrammer error is not zero, even though neither of those issues are\\nserious. Moreover, the global variable itself seems to do nothing but\\ninvite programming errors. The name itself is a disaster. What is\\n\\\"this\\\" timeline ID? Well, uh, the right one, of course! We should be\\nmore clear about what we mean: the timeline into which we are\\ninserting, the timeline from which we are replaying, the timeline from\\nwhich we are performing logical decoding. I suspect that having a\\nglobal variable here isn't even really saving us anything much, as a\\nvalue that does not change can be read from shared memory without a\\nlock.\\n\\nI would like to clean this up. Attached is a series of patches which\\ntry to do that. For ease of review, this is separated out into quite a\\nfew separate patches, but most likely we'd combine some of them for\\ncommit. Patches 0001 through 0004 eliminate all use of the global\\nvariable ThisTimeLineID outside of xlog.c, and patch 0005 makes it\\n\\\"static\\\" so that it ceases to be visible outside of xlog.c. Patches\\n0006 to 0010 clean up uses of ThisTimeLineID itself, passing it around\\nas a function parameter, or otherwise arranging to fetch the relevant\\ntimeline ID from someplace sensible rather than using the global\\nvariable as a scratchpad. Finally, patch 0011 deletes the global\\nvariable.\\n\\nI have not made a serious attempt to clear up all of the\\nterminological confusion created by the term ThisTimeLineID, which\\nalso appears as part of some structs, so you'll still see that name in\\nthe source code after applying these patches, just not as the name of\\na global variable. I have, however, used names like insertTLI and\\nreplayTLI in many places changed by the patch, and I think that makes\\nit significantly more clear which timeline is being discussed in each\\ncase. In some places I have endeavored to improve the comments. There\\nis doubtless more that could be done here, but I think this is a\\nfairly respectable start.\\n\\nOpinions welcome.\\n\\nThanks,\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Mon, 1 Nov 2021 15:16:27 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 01, 2021 at 03:16:27PM -0400, Robert Haas wrote:\\n> At the risk of stating the obvious, using the same variable for\\n> different purposes at different times is not a great programming\\n> practice. Commits 2f5c4397c39dea49c5608ba583868e26d767fc32 and\\n> 902a2c280012557b85c7e0fce3f6f0e355cb2d69 show that the possibility of\\n> programmer error is not zero, even though neither of those issues are\\n> serious.\\n\\nThis can lead to more serious issues, see 595b9cba as recent example.\\nI agree that getting rid of it in the long term is appealing.\\n\\n> Moreover, the global variable itself seems to do nothing but\\n> invite programming errors. The name itself is a disaster. What is\\n> \\\"this\\\" timeline ID? Well, uh, the right one, of course! We should be\\n> more clear about what we mean: the timeline into which we are\\n> inserting, the timeline from which we are replaying, the timeline from\\n> which we are performing logical decoding. I suspect that having a\\n> global variable here isn't even really saving us anything much, as a\\n> value that does not change can be read from shared memory without a\\n> lock.\\n\\nSome callbacks related to the WAL reader to read a patch have\\nintroduced a dependency to ThisTimelineID as a mean to be used for\\nlogical decoding on standbys. This is discussed a bit on this thread,\\nfor example:\\nhttps://www.postgresql.org/message-id/ddc31aa9-8083-58b7-0b47-e371cd4c705b@oss.nttdata.com\\n\\nTracking properly the TLI in logical decoding was mentioned by Andres\\nmultiple times on -hackers, and this has not been really done. Using\\nThisTimelineID for this purpose is something I've found confusing\\nmyself, FWIW, so this is a step in the good direction.\\n\\n> I would like to clean this up. Attached is a series of patches which\\n> try to do that. For ease of review, this is separated out into quite a\\n> few separate patches, but most likely we'd combine some of them for\\n> commit. Patches 0001 through 0004 eliminate all use of the global\\n> variable ThisTimeLineID outside of xlog.c, and patch 0005 makes it\\n> \\\"static\\\" so that it ceases to be visible outside of xlog.c. Patches\\n> 0006 to 0010 clean up uses of ThisTimeLineID itself, passing it around\\n> as a function parameter, or otherwise arranging to fetch the relevant\\n> timeline ID from someplace sensible rather than using the global\\n> variable as a scratchpad. Finally, patch 0011 deletes the global\\n> variable.\\n> \\n> I have not made a serious attempt to clear up all of the\\n> terminological confusion created by the term ThisTimeLineID, which\\n> also appears as part of some structs, so you'll still see that name in\\n> the source code after applying these patches, just not as the name of\\n> a global variable. I have, however, used names like insertTLI and\\n> replayTLI in many places changed by the patch, and I think that makes\\n> it significantly more clear which timeline is being discussed in each\\n> case. In some places I have endeavored to improve the comments. There\\n> is doubtless more that could be done here, but I think this is a\\n> fairly respectable start.\\n\\nThanks for splitting things this way. This helps a lot.\\n\\n0001 looks fair.\\n\\n+ /*\\n+ * If we're writing and flushing WAL, the time line can't be changing,\\n+ * so no lock is required.\\n+ */\\n+ if (insertTLI)\\n+ *insertTLI = XLogCtl->ThisTimeLineID;\\nIn 0002, there is no downside in putting that in the spinlock section\\neither, no? It seems to me that we should be able to call this one\\neven in recovery, as flush LSNs exist while in recovery.\\n\\n+TimeLineID\\n+GetWALInsertionTimeLine(void)\\n+{\\n+ Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);\\nOkay, that's cheaper.\\n\\nThe series 0003-0006 seemed rather fine, at quick glance.\\n\\nIn 0007, XLogFileClose() should reset openLogTLI. The same can be\\nsaid in BootStrapXLOG() before creating the control file. It may be\\ncleaner here to introduce an InvalidTimelineId, as discussed a couple\\nof days ago.\\n\\nSome comments at the top of recoveryTargetTLIRequested need a refresh\\nwith their references to ThisTimelineID.\\n\\nThe handling of XLogCtl->ThisTimeLineID still seems a bit blurry when\\nit comes to timeline switches because we don't update it while in\\nrecovery when replaying a end-of-recovery record. This could at least\\nbe documented better.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 2 Nov 2021 12:33:32 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 1, 2021 at 11:33 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> + /*\\n> + * If we're writing and flushing WAL, the time line can't be changing,\\n> + * so no lock is required.\\n> + */\\n> + if (insertTLI)\\n> + *insertTLI = XLogCtl->ThisTimeLineID;\\n> In 0002, there is no downside in putting that in the spinlock section\\n> either, no? It seems to me that we should be able to call this one\\n> even in recovery, as flush LSNs exist while in recovery.\\n\\nI don't think it matters very much from a performance point of view,\\nalthough putting a branch inside of a spinlock-protected section may\\nbe undesirable. My bigger issues with this kind of idea is that we\\ndon't want to use spinlocks as a \\\"security blanket\\\" (as in Linus from\\nPeanuts). Every time we use a spinlock, or any other kind of locking\\nmechanism, we should have a clear idea what we're protecting ourselves\\nagainst. I'm quite suspicious that there are a number of places where\\nwe're taking spinlocks in xlog.c just to feel virtuous, and not\\nbecause it does anything useful, and I don't particularly want to\\nincrease the number of such places.\\n\\nAlso, XLogCtl->ThisTimeLineID isn't set until the end of recovery, so\\ncalling this function during recovery would be a mistake. There seem\\nto be a number of interface functions in xlog.c that should only be\\ncalled when not in recovery, and neither their names nor their\\ncomments make that clear. I wasn't bold enough to go label all the\\nones I think fall into that category, but maybe we should. Honestly,\\nthe naming of things in this file is annoyingly bad in general. My\\nfavorite example, found during this project, is:\\n\\n#define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))\\n\\nIt's not good enough to have one name that's difficult to understand\\nor remember or capitalize properly -- let's have TWO -- for the same\\nthing!\\n\\n> In 0007, XLogFileClose() should reset openLogTLI. The same can be\\n> said in BootStrapXLOG() before creating the control file. It may be\\n> cleaner here to introduce an InvalidTimelineId, as discussed a couple\\n> of days ago.\\n\\nI thought about that, but I think it's pointless. I think we can just\\nsay that if openLogFile == -1, openLogSegNo and openLogTLI are\\nmeaningless. I don't think we're currently resetting openLogSegNo to\\nan invalid value either, so I don't think openLogTLI should be treated\\ndifferently. I'm not opposed to introducing InvalidTimeLineID on\\nprinciple, and I looked for a way of doing it in this patch set, but\\nit didn't seem all that valuable, and I feel that someone who cares\\nabout it can do it as a separate effort after I commit these.\\nHonestly, I think these patches greatly reduce the need to be afraid\\nof leaving the TLI unset, because they remove the crutch of hoping\\nthat ThisTimeLineID is initialized to the proper value. You actually\\nhave to think about which TLI you are going to pass. Now I'm very sure\\nfurther improvement is possible, but I think this patch set is complex\\nenough already, and I'd like to get it committed before contemplating\\nany other work in this area.\\n\\n> Some comments at the top of recoveryTargetTLIRequested need a refresh\\n> with their references to ThisTimelineID.\\n\\nAfter thinking this over, I think we should just delete the entire\\nfirst paragraph, so that the comments start like this:\\n\\n/*\\n * recoveryTargetTimeLineGoal: what the user requested, if any\\n *\\n * recoveryTargetTLIRequested: numeric value of requested timeline, if constant\\n\\nI don't see that we'd be losing anything that way. When you look at\\nthat paragraph now, the first sentence is just confusing the issue,\\nbecause we actually don't care about XLogCtl->ThisTimeLineID during\\nrecovery, because it's not set. We do care about plain old\\nThisTimeLineID, but that's going to be demoted to a local variable by\\nthese patches, and there just doesn't seem to be any reason to discuss\\nit here. The second sentence is wrong already, because the rmgr\\nroutines do not rely on ThisTimeLineID. And the third sentence is also\\nwrong and/or confusing, because not all of the variables that follow\\nare TLIs -- only 3 of 5 are. I don't actually think there's any useful\\nintroduction we need to provide here; we just need to tell people what\\nthe variables we're about to declare do.\\n\\n> The handling of XLogCtl->ThisTimeLineID still seems a bit blurry when\\n> it comes to timeline switches because we don't update it while in\\n> recovery when replaying a end-of-recovery record. This could at least\\n> be documented better.\\n\\nWe don't update it during recovery at all. There's exactly one\\nassignment statement for that variable and it's here:\\n\\n /* Save the selected TimeLineID in shared memory, too */\\n XLogCtl->ThisTimeLineID = ThisTimeLineID;\\n XLogCtl->PrevTimeLineID = PrevTimeLineID;\\n\\nThat's after the main redo loop completes.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Nov 2021 08:45:57 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 2, 2021 at 12:46 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> [...]\\n> I would like to clean this up. Attached is a series of patches which\\n> try to do that. For ease of review, this is separated out into quite a\\n> few separate patches, but most likely we'd combine some of them for\\n> commit. Patches 0001 through 0004 eliminate all use of the global\\n> variable ThisTimeLineID outside of xlog.c, and patch 0005 makes it\\n> \\\"static\\\" so that it ceases to be visible outside of xlog.c. Patches\\n> 0006 to 0010 clean up uses of ThisTimeLineID itself, passing it around\\n> as a function parameter, or otherwise arranging to fetch the relevant\\n> timeline ID from someplace sensible rather than using the global\\n> variable as a scratchpad. Finally, patch 0011 deletes the global\\n> variable.\\n>\\n> I have not made a serious attempt to clear up all of the\\n> terminological confusion created by the term ThisTimeLineID, which\\n> also appears as part of some structs, so you'll still see that name in\\n> the source code after applying these patches, just not as the name of\\n> a global variable. I have, however, used names like insertTLI and\\n> replayTLI in many places changed by the patch, and I think that makes\\n> it significantly more clear which timeline is being discussed in each\\n> case. In some places I have endeavored to improve the comments. There\\n> is doubtless more that could be done here, but I think this is a\\n> fairly respectable start.\\n>\\n> Opinions welcome.\\n>\\n\\nI had a plan to look at these patches closely, but unfortunately,\\ndidn't get enough time; and might not be able to spend time in the\\nrest of this week. Here are a few thoughts that I had in the initial\\ngo-through, but that may or may not sound very interesting:\\n\\n0002:\\n-GetFlushRecPtr(void)\\n+GetFlushRecPtr(TimeLineID *insertTLI)\\n\\nShould we rename this argument to more generic as \\\"tli\\\", like\\nGetStandbyFlushRecPtr, since the caller in 003 patch trying to fetch a\\nTLI that means different for them, e.g. currTLI, FlushTLI, etc.\\n\\n\\n0004:\\n static int\\n-XLogFileInitInternal(XLogSegNo logsegno, bool *added, char *path)\\n+XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,\\n+ bool *added, char *path)\\n\\n int\\n-XLogFileInit(XLogSegNo logsegno)\\n+XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)\\n {\\n\\nHow about logsegtli instead, to be inlined with logsegno ?\\n\\n\\n0007:\\n static XLogSegNo openLogSegNo = 0;\\n+static TimeLineID openLogTLI = 0;\\n\\nSimilarly, openLogSegTLI ?\\n\\n\\n0008:\\n+ /*\\n+ * Given that we're not in recovery, ThisTimeLineID is set and can't\\n+ * change, so we can read it without a lock.\\n+ */\\n+ insertTLI = XLogCtl->ThisTimeLineID;\\n\\nCan't GetWALInsertionTimeLine() called instead? I guess the reason is\\nto avoid the unnecessary overhead involved in the function call. (Same\\nat a few other places.)\\n\\nRegards,\\nAmul\\n\\n\\n\",\n \"msg_date\": \"Wed, 3 Nov 2021 18:42:22 +0530\",\n \"msg_from\": \"Amul Sul <sulamul@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 3, 2021 at 9:12 AM Amul Sul <sulamul@gmail.com> wrote:\\n> 0002:\\n> -GetFlushRecPtr(void)\\n> +GetFlushRecPtr(TimeLineID *insertTLI)\\n>\\n> Should we rename this argument to more generic as \\\"tli\\\", like\\n> GetStandbyFlushRecPtr, since the caller in 003 patch trying to fetch a\\n> TLI that means different for them, e.g. currTLI, FlushTLI, etc.\\n\\nIt's possible that more consistency would be better here, but I don't\\nthink making this name more generic is going in the right direction.\\nIf somebody gets the TLI using this function, it's the timeline into\\nwhich a system not in recovery is inserting WAL, which is the same\\ntimeline to which WAL is being flushed. I think it's important for\\nthis name to try to make that clear. It doesn't really matter if\\nsomeone treats the insertTLI as the writeTLI or the flushTLI since on\\na system in production they're all the same - but they must not\\nconfuse it with, say, the replayTLI.\\n\\n> How about logsegtli instead, to be inlined with logsegno ?\\n> Similarly, openLogSegTLI ?\\n\\nHmm, well I guess it depends on how you think the words group. If you\\nlogsegno means \\\"the number of the log segment\\\" then it would make\\nsense to have logsegli to mean \\\"the tli of the log segment.\\\" But I\\nthink of logsegno as meaning \\\"the segment number of the log file,\\\" so\\nit makes more sense to have logtli to mean \\\"the TLI of the log file\\\".\\nI think it's definitely not a \\\"log segment file\\\" - just a \\\"log file\\\".\\n\\n> Can't GetWALInsertionTimeLine() called instead? I guess the reason is\\n> to avoid the unnecessary overhead involved in the function call. (Same\\n> at a few other places.)\\n\\nThat, plus to avoid failing the assertion in that function. As we've\\ndiscussed on the ALTER SYSTEM READ ONLY thread, xlog.c itself inserts\\nsome WAL before recovery is officially over.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 3 Nov 2021 10:13:55 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"I looked at these in a very quick skim. I agree that this is a\\ngood step in a good direction.\\n\\nI 'git rebase -x'd this series in order to compile and run the tests on\\neach patch individually. There are no compiler warnings anywhere and no\\ntest failures.\\n\\n\\n-- \\nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\\n\\\"La victoria es para quien se atreve a estar solo\\\"\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Nov 2021 10:49:44 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 5, 2021 at 9:49 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> I looked at these in a very quick skim. I agree that this is a\\n> good step in a good direction.\\n\\nCool. I have now committed these. I grouped them into just 3 commits\\nrather than having 11 separate commits as I did in my earlier posting,\\nbut the content is the same.\\n\\n> I 'git rebase -x'd this series in order to compile and run the tests on\\n> each patch individually. There are no compiler warnings anywhere and no\\n> test failures.\\n\\nThanks for checking. It's hard to tell whether the stuff that works on\\nmy machine will work everywhere, unless someone tests it.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Nov 2021 12:58:04 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 02, 2021 at 08:45:57AM -0400, Robert Haas wrote:\\n> Also, XLogCtl->ThisTimeLineID isn't set until the end of recovery, so\\n> calling this function during recovery would be a mistake. There seem\\n> to be a number of interface functions in xlog.c that should only be\\n> called when not in recovery, and neither their names nor their\\n> comments make that clear. I wasn't bold enough to go label all the\\n> ones I think fall into that category, but maybe we should. \\n\\nI got to wonder whether it would be better to add in GetFlushRecPtr()\\nthe same assertion as GetWALInsertionTimeLine(). All the in-core\\ncallers of this routine already assume that, and it would be buggy if\\none passes down insertTLI to get the current TLI.\\n\\n> I thought about that, but I think it's pointless. I think we can just\\n> say that if openLogFile == -1, openLogSegNo and openLogTLI are\\n> meaningless. I don't think we're currently resetting openLogSegNo to\\n> an invalid value either, so I don't think openLogTLI should be treated\\n> differently.\\n\\nWouldn't it be better to at least document that as a comment rather\\nthan letting the reader guess it, then? I agree that this is a minor\\npoint, though.\\n\\n> I'm not opposed to introducing InvalidTimeLineID on\\n> principle, and I looked for a way of doing it in this patch set, but\\n> it didn't seem all that valuable, and I feel that someone who cares\\n> about it can do it as a separate effort after I commit these.\\n\\nNo issues from me. I have bumped into a case just at the end of last\\nweek where I wanted a better way to tell if a TLI is invalid rather\\nthan leaving things to 0, FWIW.\\n\\n>> The handling of XLogCtl->ThisTimeLineID still seems a bit blurry when\\n>> it comes to timeline switches because we don't update it while in\\n>> recovery when replaying a end-of-recovery record. This could at least\\n>> be documented better.\\n> \\n> We don't update it during recovery at all. There's exactly one\\n> assignment statement for that variable and it's here:\\n\\nIndeed. It looks like my reading was sloppy here.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 8 Nov 2021 13:24:10 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 7, 2021 at 11:24 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> I got to wonder whether it would be better to add in GetFlushRecPtr()\\n> the same assertion as GetWALInsertionTimeLine(). All the in-core\\n> callers of this routine already assume that, and it would be buggy if\\n> one passes down insertTLI to get the current TLI.\\n\\nPerhaps. That's related to the point I made before, that it might be a\\ngood idea to be more clear about which of these functions are intended\\nto be used in recovery and which ones are intended to be used in\\nnormal running. I don't rule out the possibility of doing some more\\nresearch here and tightening that up, but I don't want to go start\\ntweaking things unless I'm sure I know what I'm talking about. This\\nstuff is so confusing that I don't want to take any risk of making\\nchanges that turn out to be incorrect.\\n\\n> > I thought about that, but I think it's pointless. I think we can just\\n> > say that if openLogFile == -1, openLogSegNo and openLogTLI are\\n> > meaningless. I don't think we're currently resetting openLogSegNo to\\n> > an invalid value either, so I don't think openLogTLI should be treated\\n> > differently.\\n>\\n> Wouldn't it be better to at least document that as a comment rather\\n> than letting the reader guess it, then? I agree that this is a minor\\n> point, though.\\n\\nI don't think it's very confusing, but if you want to improve the\\ncomments, go for it!\\n\\n> >> The handling of XLogCtl->ThisTimeLineID still seems a bit blurry when\\n> >> it comes to timeline switches because we don't update it while in\\n> >> recovery when replaying a end-of-recovery record. This could at least\\n> >> be documented better.\\n> >\\n> > We don't update it during recovery at all. There's exactly one\\n> > assignment statement for that variable and it's here:\\n>\\n> Indeed. It looks like my reading was sloppy here.\\n\\nI can hardly blame you. There are comments all over the place claiming\\nthat this thing and that thing are the same when they're only\\nsometimes the same. Or, as in this case, things that are named the\\nsame when they're actually different. I'm actually tempted to make a\\npass over the comments related to this topic and just try to make\\nthings more comprehensible. It's super-confusing as things stand.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 Nov 2021 08:27:06 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 8, 2021 at 8:27 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> Perhaps. That's related to the point I made before, that it might be a\\n> good idea to be more clear about which of these functions are intended\\n> to be used in recovery and which ones are intended to be used in\\n> normal running. I don't rule out the possibility of doing some more\\n> research here and tightening that up, but I don't want to go start\\n> tweaking things unless I'm sure I know what I'm talking about. This\\n> stuff is so confusing that I don't want to take any risk of making\\n> changes that turn out to be incorrect.\\n\\nWell I went through and it seems to be OK: all the existing callers of\\nthat function first verify that we're not in recovery. The patch to\\nmake logical decoding work in standby mode might change that, but as\\nof now it's so. So in the attached patch, I have added the assertion\\nand comments that you have proposed there. I also removed the\\nmisleading comment we discussed before. In addition to that, I renamed\\nthe local variable ThisTimeLineID to replayTLI, and I also renamed\\nXLogCtl->ThisTimeLineID to InsertTimeLineID, which I think should make\\nthings clearer: imagine if things that were used for different\\npurposes had different names!\\n\\nEven with this patch, the name ThisTimeLineID is still used for\\nmultiple purposes. It remains part of the CheckPoint struct, and also\\npart of the xl_end_of_recovery struct. But in both of those cases, the\\nname ThisTimeLineID actually makes sense, because what we're thinking\\nabout is whether there's a timeline change, so we have ThisTimeLineID\\nand PrevTimeLineID and it seems fairly clear what that's supposed to\\nmean.\\n\\nThoughts?\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Mon, 8 Nov 2021 12:49:52 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 08, 2021 at 12:49:52PM -0500, Robert Haas wrote:\\n> Even with this patch, the name ThisTimeLineID is still used for\\n> multiple purposes. It remains part of the CheckPoint struct, and also\\n> part of the xl_end_of_recovery struct. But in both of those cases, the\\n> name ThisTimeLineID actually makes sense, because what we're thinking\\n> about is whether there's a timeline change, so we have ThisTimeLineID\\n> and PrevTimeLineID and it seems fairly clear what that's supposed to\\n> mean.\\n\\nI got to wonder if it would be better to bite the bullet here for\\nthese two looking at your patch, but I am fine to keep things as they\\nare, as well.\\n\\n> Thoughts?\\n\\nI think that this patch is an improvement.\\n\\n@@ -6686,8 +6682,8 @@ StartupXLOG(void)\\n- TimeLineID ThisTimeLineID,\\n- PrevTimeLineID;\\n+ TimeLineID replayTLI,\\n+ newTLI;\\nOne problem with newTLI is that this conflicts with the declaration\\naround 7663 in xlog.c, where we check after a TLI switch when\\nreplaying such a record. Perhaps this could be named newInsertTLI,\\nfor example.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 9 Nov 2021 12:30:53 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 8, 2021 at 10:31 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> I think that this patch is an improvement.\\n\\nCool.\\n\\n> @@ -6686,8 +6682,8 @@ StartupXLOG(void)\\n> - TimeLineID ThisTimeLineID,\\n> - PrevTimeLineID;\\n> + TimeLineID replayTLI,\\n> + newTLI;\\n> One problem with newTLI is that this conflicts with the declaration\\n> around 7663 in xlog.c, where we check after a TLI switch when\\n> replaying such a record. Perhaps this could be named newInsertTLI,\\n> for example.\\n\\nThat's a good point. However, since I think newTLI is already in use\\nin some of the functions StartupXLOG() calls, and since I think it\\nwould be good to use the same name in StartupXLOG() as in the\\nfunctions that it calls, what I'd prefer to do is rename the newTLI\\nvariable in the inner scope to newReplayTLI, as in the attached v2.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Tue, 9 Nov 2021 14:15:51 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 09, 2021 at 02:15:51PM -0500, Robert Haas wrote:\\n> That's a good point. However, since I think newTLI is already in use\\n> in some of the functions StartupXLOG() calls, and since I think it\\n> would be good to use the same name in StartupXLOG() as in the\\n> functions that it calls, what I'd prefer to do is rename the newTLI\\n> variable in the inner scope to newReplayTLI, as in the attached v2.\\n\\nWFM. Thanks.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 10 Nov 2021 09:41:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 9, 2021 at 7:41 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> On Tue, Nov 09, 2021 at 02:15:51PM -0500, Robert Haas wrote:\\n> > That's a good point. However, since I think newTLI is already in use\\n> > in some of the functions StartupXLOG() calls, and since I think it\\n> > would be good to use the same name in StartupXLOG() as in the\\n> > functions that it calls, what I'd prefer to do is rename the newTLI\\n> > variable in the inner scope to newReplayTLI, as in the attached v2.\\n>\\n> WFM. Thanks.\\n\\nCool. Committed that way.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 09:52:06 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 11/8/21 6:49 PM, Robert Haas wrote:\\n> On Mon, Nov 8, 2021 at 8:27 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> Well I went through and it seems to be OK: all the existing callers of\\n> that function first verify that we're not in recovery. The patch to\\n> make logical decoding work in standby mode might change that,\\n\\nIndeed, I think the logical decoding on standby patch just needs to \\nchange the Assert in GetWALInsertionTimeLine() to check \\nSharedRecoveryState is RECOVERY_STATE_DONE or RECOVERY_STATE_ARCHIVE.\\n\\nWould that make sense from your point of view to add the check on \\nRECOVERY_STATE_ARCHIVE or do you think it would need to be more \\n\\\"sophisticated\\\"?\\n\\nThanks\\n\\nBertrand\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 23 Nov 2021 10:36:16 +0100\",\n \"msg_from\": \"\\\"Drouvot, Bertrand\\\" <bdrouvot@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 23, 2021 at 4:36 AM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\\n> Indeed, I think the logical decoding on standby patch just needs to\\n> change the Assert in GetWALInsertionTimeLine() to check\\n> SharedRecoveryState is RECOVERY_STATE_DONE or RECOVERY_STATE_ARCHIVE.\\n>\\n> Would that make sense from your point of view to add the check on\\n> RECOVERY_STATE_ARCHIVE or do you think it would need to be more\\n> \\\"sophisticated\\\"?\\n\\nUnless I'm confused, that would just be incorrect.\\nGetWALInsertionTimeLine() just returns InsertTimeLineID, which won't\\nbe initialized during archive recovery. I discussed this point a bit\\non some other thread with Andres. It's possible that all that you need\\nto do here is determine the replayTLI using e.g. GetXLogReplayRecPtr.\\nBut I don't really see why that should be correct:\\nread_local_xlog_page() has some kind of wait-and-retry logic and I'm\\nnot clear why we don't need the same thing here. After all, if we're\\non the primary and we determine that sendTimeLineHistoric = false and\\nsendTimeLine = whatever, neither of those values can become incorrect\\nafterward. But on a standby that can certainly happen, if an upstream\\nserver is promoted. Note that the comment in read_local_xlog_page()\\nsays this:\\n\\n * If that happens after our caller determined the TLI but before\\n * we actually read the xlog page, we might still try to read from the\\n * old (now renamed) segment and fail. There's not much we can do\\n * about this, but it can only happen when we're a leaf of a cascading\\n * standby whose primary gets promoted while we're decoding, so a\\n * one-off ERROR isn't too bad.\\n\\nThat seems to imply that the retry loop here is designed, at least in\\npart, to protect against exactly this kind of scenario.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 23 Nov 2021 08:30:44 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: removing global variable ThisTimeLineID\"\n }\n]","isConversational":false}}},{"rowIdx":53037,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Meson depends on ninja, so it behooves us to look into portability\nof ninja if we're considering migrating to meson as the build system.\nI tried it out quickly on some ancient platforms, and here's what\nI found.\n\n1. The documentation is kinda sucky, eg it fails to explain how\nto run ninja's own regression tests. I ended up with this\nreverse-engineered recipe:\n\ngit clone git://github.com/ninja-build/ninja.git\ncd ninja\ngit checkout release\n./configure.py --bootstrap\n./ninja ninja_test\n./ninja_test\nsudo cp ninja /usr/local/bin # or location of choice\n\nWe'll probably have to include that in our own documentation.\n\n2. You need Python to build it this way. There is also a process\nfor bootstrapping using cmake, but that doesn't seem like a more\nattractive dependency for old platforms. Fortunately for my\npurposes here, it seems to work with fairly old Python --- I built\nsuccessfully with python 2.6.2, though not with 2.4.1.\n\n3. You also need a C++ compiler, but really old ones will work.\nI did not see any problems with g++ as old as 4.0.1, other than\nsome compiler warnings about non-virtual destructors.\n\n4. It built and passed self-test on macOS Leopard (10.5.8), which is\npretty old ... but not old enough for prairiedog, which is stuck on\n10.4.x. There, the build fails miserably for lack of <spawn.h>.\nIt looks like that was added to POSIX in 2001, so one could have\nwished for it in a 2005-vintage OS; but nope, Apple took another\ncouple of years to get around to that. I'm not even going to bother\ntrying on gaur's pre-turn-of-the-century OS.\n\n5. It built and passed self-test on Solaris 11, but failed self-test\non Solaris 10.\n\n6. While configure.py thinks it knows what to do on AIX, it fails\non AIX 7.1 and 7.2:\n\nTraceback (most recent call last):\n File \"./configure.py\", line 544, in <module>\n if platform.is_aix() and not platform.is_os400_pase():\n File \"./configure.py\", line 103, in is_os400_pase\n return self._platform == 'os400' or os.uname().sysname.startswith('OS400')\nAttributeError: 'tuple' object has no attribute 'sysname'\n\nPossibly the ninja guys would take a patch for that (or maybe\nthis is a you-need-python-3 case?). I do see /usr/include/spawn.h\non that platform, so there's room to hope it'd work.\n\n\nSo it's pretty clear that if we go this way, it'll be the end of\nthe line for support for some very old OS versions. I can't,\nhowever, argue with the idea that it's reasonable to require\nPOSIX 2001 support now. Based on these results, I doubt that\nninja will give us trouble on any platform that isn't old\nenough to get its drivers license.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 01 Nov 2021 15:25:08 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Portability report: ninja"},{"msg_contents":"Hi,\n\nOn 11/1/21 15:25, Tom Lane wrote:\n> So it's pretty clear that if we go this way, it'll be the end of\n> the line for support for some very old OS versions. I can't,\n> however, argue with the idea that it's reasonable to require\n> POSIX 2001 support now. Based on these results, I doubt that\n> ninja will give us trouble on any platform that isn't old\n> enough to get its drivers license.\n\n\nYou can also look at it as:\n\n\nIf PostgreSQL choose a newer build system then it is up to the companies \nowning the \"non-supported\" operating systems to add support for the \nbuild system in question; not the PostgreSQL community.\n\n\n+1 for POSIX.2001 and meson/ninja.\n\n\nBest regards,\n\n Jesper\n\n\n\n\n","msg_date":"Mon, 1 Nov 2021 16:06:20 -0400","msg_from":"Jesper Pedersen <jesper.pedersen@redhat.com>","msg_from_op":false,"msg_subject":"Re: Portability report: ninja"},{"msg_contents":"On Tue, Nov 2, 2021 at 8:25 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> 6. While configure.py thinks it knows what to do on AIX, it fails\n> on AIX 7.1 and 7.2:\n>\n> Traceback (most recent call last):\n> File \"./configure.py\", line 544, in <module>\n> if platform.is_aix() and not platform.is_os400_pase():\n> File \"./configure.py\", line 103, in is_os400_pase\n> return self._platform == 'os400' or os.uname().sysname.startswith('OS400')\n> AttributeError: 'tuple' object has no attribute 'sysname'\n\nYeah, the result type changed: os.uname()[0] works on python2 AND python3.\n\n\n","msg_date":"Tue, 2 Nov 2021 09:27:17 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Portability report: ninja"},{"msg_contents":"On Tue, Nov 2, 2021 at 8:25 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> 5. It built and passed self-test on Solaris 11, but failed self-test\n> on Solaris 10.\n\nI haven't had time to actually do anything with it yet, but I can\nreport that meson and ninja are, as of quite recently, available in\nthe package repository of OpenIndiana.\n\nThat's the distribution of illumos I found very convenient to work\nwith last time I turned the Solaris-family machines red in the build\nfarm. The reason I like it is that there's a rolling \"vagrant\" image\n(libvirt and virtualbox flavours), so it's possible to spin up a\nworking VM with minimum effort. My current way of keeping track of\nwhat commands are needed to install the various packages PostgreSQL\nneeds for niche OSes corresponding to our build farm is to record it\nall in Vagrant files.\n\nhttps://github.com/macdice/postgresql-dev-vagrant/blob/master/openindiana/Vagrantfile\n\n\n","msg_date":"Tue, 2 Nov 2021 09:56:10 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Portability report: ninja"},{"msg_contents":"Hi,\n\nOn 2021-11-01 15:25:08 -0400, Tom Lane wrote:\n> Meson depends on ninja, so it behooves us to look into portability\n> of ninja if we're considering migrating to meson as the build system.\n> I tried it out quickly on some ancient platforms, and here's what\n> I found.\n\nThanks, that's helpful!\n\n\n> 1. The documentation is kinda sucky, eg it fails to explain how\n> to run ninja's own regression tests. I ended up with this\n> reverse-engineered recipe:\n> \n> git clone git://github.com/ninja-build/ninja.git\n> cd ninja\n> git checkout release\n> ./configure.py --bootstrap\n> ./ninja ninja_test\n> ./ninja_test\n> sudo cp ninja /usr/local/bin # or location of choice\n> \n> We'll probably have to include that in our own documentation.\n\nOr perhaps a helper script? I was wondering whether we'd eventually want a\nwrapper ./configure / makefile. Perhaps not worth it though - there's enough\nprojects using meson and ninja that it might add more confusion than being\nhelpful.\n\n\n> There is also a process for bootstrapping using cmake, but that doesn't seem\n> like a more attractive dependency for old platforms.\n\nYea. Especially given that meson itself needs python anyway.\n\n\n> Fortunately for my\n> purposes here, it seems to work with fairly old Python --- I built\n> successfully with python 2.6.2, though not with 2.4.1.\n\nmeson will need a newer python though...\n\n\n> 4. It built and passed self-test on macOS Leopard (10.5.8), which is\n> pretty old ... but not old enough for prairiedog, which is stuck on\n> 10.4.x. There, the build fails miserably for lack of <spawn.h>.\n> It looks like that was added to POSIX in 2001, so one could have\n> wished for it in a 2005-vintage OS; but nope, Apple took another\n> couple of years to get around to that. I'm not even going to bother\n> trying on gaur's pre-turn-of-the-century OS.\n\n> 5. It built and passed self-test on Solaris 11, but failed self-test\n> on Solaris 10.\n\nI think we can live with those...\n\n\n> 6. While configure.py thinks it knows what to do on AIX, it fails\n> on AIX 7.1 and 7.2:\n> \n> Traceback (most recent call last):\n> File \"./configure.py\", line 544, in <module>\n> if platform.is_aix() and not platform.is_os400_pase():\n> File \"./configure.py\", line 103, in is_os400_pase\n> return self._platform == 'os400' or os.uname().sysname.startswith('OS400')\n> AttributeError: 'tuple' object has no attribute 'sysname'\n> \n> Possibly the ninja guys would take a patch for that (or maybe\n> this is a you-need-python-3 case?). I do see /usr/include/spawn.h\n> on that platform, so there's room to hope it'd work.\n\nThat does seem like it'd be a issue. Briefly trawling the ninja git log it\ndoes look like there's regular-ish maintenance stuff for AIX, so I'd hope we\ncould get it fixed. I do suspect it's just a python3 issue, as Thomas noted.\n\n\n> Based on these results, I doubt that ninja will give us trouble on any\n> platform that isn't old enough to get its drivers license.\n\nAgreed.\n\nThere's also alternative compatible ninja implementation in C99 ([1]). But I\nthink it's minimum requirements aren't actually lower than ninja's (it says it\nrequires posix 2008).\n\nGreetings,\n\nAndres Freund\n\n\n[1] https://github.com/michaelforney/samurai\n\n\n","msg_date":"Mon, 1 Nov 2021 21:21:23 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Portability report: ninja"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2021-11-01 15:25:08 -0400, Tom Lane wrote:\n>> Fortunately for my\n>> purposes here, it seems to work with fairly old Python --- I built\n>> successfully with python 2.6.2, though not with 2.4.1.\n\n> meson will need a newer python though...\n\nYup. I was just trying to establish what the lower bound was for\nninja, since they (ahem) don't document it.\n\nThe meson docs say it needs Python >= 3.6, which I think is going\nto be an issue, but I've not researched it yet. What I was mainly\ncurious about today was whether anything using ninja as the backend\ncould meet our portability goals. My mind's more or less at ease\non that, but I am afraid that requiring 3.6 is going to be a problem\non LTS distributions. That's a topic for another thread though.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 02 Nov 2021 00:48:14 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Portability report: ninja"},{"msg_contents":"On 2021-11-02 09:27:17 +1300, Thomas Munro wrote:\n> On Tue, Nov 2, 2021 at 8:25 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > 6. While configure.py thinks it knows what to do on AIX, it fails\n> > on AIX 7.1 and 7.2:\n> >\n> > Traceback (most recent call last):\n> > File \"./configure.py\", line 544, in <module>\n> > if platform.is_aix() and not platform.is_os400_pase():\n> > File \"./configure.py\", line 103, in is_os400_pase\n> > return self._platform == 'os400' or os.uname().sysname.startswith('OS400')\n> > AttributeError: 'tuple' object has no attribute 'sysname'\n> \n> Yeah, the result type changed: os.uname()[0] works on python2 AND python3.\n\nFWIW, with python3 ninja builds and tests successfully on AIX.\n\n\n","msg_date":"Fri, 5 Nov 2021 14:42:49 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Portability report: ninja"}],"string":"[\n {\n \"msg_contents\": \"Meson depends on ninja, so it behooves us to look into portability\\nof ninja if we're considering migrating to meson as the build system.\\nI tried it out quickly on some ancient platforms, and here's what\\nI found.\\n\\n1. The documentation is kinda sucky, eg it fails to explain how\\nto run ninja's own regression tests. I ended up with this\\nreverse-engineered recipe:\\n\\ngit clone git://github.com/ninja-build/ninja.git\\ncd ninja\\ngit checkout release\\n./configure.py --bootstrap\\n./ninja ninja_test\\n./ninja_test\\nsudo cp ninja /usr/local/bin # or location of choice\\n\\nWe'll probably have to include that in our own documentation.\\n\\n2. You need Python to build it this way. There is also a process\\nfor bootstrapping using cmake, but that doesn't seem like a more\\nattractive dependency for old platforms. Fortunately for my\\npurposes here, it seems to work with fairly old Python --- I built\\nsuccessfully with python 2.6.2, though not with 2.4.1.\\n\\n3. You also need a C++ compiler, but really old ones will work.\\nI did not see any problems with g++ as old as 4.0.1, other than\\nsome compiler warnings about non-virtual destructors.\\n\\n4. It built and passed self-test on macOS Leopard (10.5.8), which is\\npretty old ... but not old enough for prairiedog, which is stuck on\\n10.4.x. There, the build fails miserably for lack of <spawn.h>.\\nIt looks like that was added to POSIX in 2001, so one could have\\nwished for it in a 2005-vintage OS; but nope, Apple took another\\ncouple of years to get around to that. I'm not even going to bother\\ntrying on gaur's pre-turn-of-the-century OS.\\n\\n5. It built and passed self-test on Solaris 11, but failed self-test\\non Solaris 10.\\n\\n6. While configure.py thinks it knows what to do on AIX, it fails\\non AIX 7.1 and 7.2:\\n\\nTraceback (most recent call last):\\n File \\\"./configure.py\\\", line 544, in <module>\\n if platform.is_aix() and not platform.is_os400_pase():\\n File \\\"./configure.py\\\", line 103, in is_os400_pase\\n return self._platform == 'os400' or os.uname().sysname.startswith('OS400')\\nAttributeError: 'tuple' object has no attribute 'sysname'\\n\\nPossibly the ninja guys would take a patch for that (or maybe\\nthis is a you-need-python-3 case?). I do see /usr/include/spawn.h\\non that platform, so there's room to hope it'd work.\\n\\n\\nSo it's pretty clear that if we go this way, it'll be the end of\\nthe line for support for some very old OS versions. I can't,\\nhowever, argue with the idea that it's reasonable to require\\nPOSIX 2001 support now. Based on these results, I doubt that\\nninja will give us trouble on any platform that isn't old\\nenough to get its drivers license.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 01 Nov 2021 15:25:08 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Portability report: ninja\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 11/1/21 15:25, Tom Lane wrote:\\n> So it's pretty clear that if we go this way, it'll be the end of\\n> the line for support for some very old OS versions. I can't,\\n> however, argue with the idea that it's reasonable to require\\n> POSIX 2001 support now. Based on these results, I doubt that\\n> ninja will give us trouble on any platform that isn't old\\n> enough to get its drivers license.\\n\\n\\nYou can also look at it as:\\n\\n\\nIf PostgreSQL choose a newer build system then it is up to the companies \\nowning the \\\"non-supported\\\" operating systems to add support for the \\nbuild system in question; not the PostgreSQL community.\\n\\n\\n+1 for POSIX.2001 and meson/ninja.\\n\\n\\nBest regards,\\n\\n Jesper\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 1 Nov 2021 16:06:20 -0400\",\n \"msg_from\": \"Jesper Pedersen <jesper.pedersen@redhat.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Portability report: ninja\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 2, 2021 at 8:25 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> 6. While configure.py thinks it knows what to do on AIX, it fails\\n> on AIX 7.1 and 7.2:\\n>\\n> Traceback (most recent call last):\\n> File \\\"./configure.py\\\", line 544, in <module>\\n> if platform.is_aix() and not platform.is_os400_pase():\\n> File \\\"./configure.py\\\", line 103, in is_os400_pase\\n> return self._platform == 'os400' or os.uname().sysname.startswith('OS400')\\n> AttributeError: 'tuple' object has no attribute 'sysname'\\n\\nYeah, the result type changed: os.uname()[0] works on python2 AND python3.\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Nov 2021 09:27:17 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Portability report: ninja\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 2, 2021 at 8:25 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> 5. It built and passed self-test on Solaris 11, but failed self-test\\n> on Solaris 10.\\n\\nI haven't had time to actually do anything with it yet, but I can\\nreport that meson and ninja are, as of quite recently, available in\\nthe package repository of OpenIndiana.\\n\\nThat's the distribution of illumos I found very convenient to work\\nwith last time I turned the Solaris-family machines red in the build\\nfarm. The reason I like it is that there's a rolling \\\"vagrant\\\" image\\n(libvirt and virtualbox flavours), so it's possible to spin up a\\nworking VM with minimum effort. My current way of keeping track of\\nwhat commands are needed to install the various packages PostgreSQL\\nneeds for niche OSes corresponding to our build farm is to record it\\nall in Vagrant files.\\n\\nhttps://github.com/macdice/postgresql-dev-vagrant/blob/master/openindiana/Vagrantfile\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Nov 2021 09:56:10 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Portability report: ninja\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2021-11-01 15:25:08 -0400, Tom Lane wrote:\\n> Meson depends on ninja, so it behooves us to look into portability\\n> of ninja if we're considering migrating to meson as the build system.\\n> I tried it out quickly on some ancient platforms, and here's what\\n> I found.\\n\\nThanks, that's helpful!\\n\\n\\n> 1. The documentation is kinda sucky, eg it fails to explain how\\n> to run ninja's own regression tests. I ended up with this\\n> reverse-engineered recipe:\\n> \\n> git clone git://github.com/ninja-build/ninja.git\\n> cd ninja\\n> git checkout release\\n> ./configure.py --bootstrap\\n> ./ninja ninja_test\\n> ./ninja_test\\n> sudo cp ninja /usr/local/bin # or location of choice\\n> \\n> We'll probably have to include that in our own documentation.\\n\\nOr perhaps a helper script? I was wondering whether we'd eventually want a\\nwrapper ./configure / makefile. Perhaps not worth it though - there's enough\\nprojects using meson and ninja that it might add more confusion than being\\nhelpful.\\n\\n\\n> There is also a process for bootstrapping using cmake, but that doesn't seem\\n> like a more attractive dependency for old platforms.\\n\\nYea. Especially given that meson itself needs python anyway.\\n\\n\\n> Fortunately for my\\n> purposes here, it seems to work with fairly old Python --- I built\\n> successfully with python 2.6.2, though not with 2.4.1.\\n\\nmeson will need a newer python though...\\n\\n\\n> 4. It built and passed self-test on macOS Leopard (10.5.8), which is\\n> pretty old ... but not old enough for prairiedog, which is stuck on\\n> 10.4.x. There, the build fails miserably for lack of <spawn.h>.\\n> It looks like that was added to POSIX in 2001, so one could have\\n> wished for it in a 2005-vintage OS; but nope, Apple took another\\n> couple of years to get around to that. I'm not even going to bother\\n> trying on gaur's pre-turn-of-the-century OS.\\n\\n> 5. It built and passed self-test on Solaris 11, but failed self-test\\n> on Solaris 10.\\n\\nI think we can live with those...\\n\\n\\n> 6. While configure.py thinks it knows what to do on AIX, it fails\\n> on AIX 7.1 and 7.2:\\n> \\n> Traceback (most recent call last):\\n> File \\\"./configure.py\\\", line 544, in <module>\\n> if platform.is_aix() and not platform.is_os400_pase():\\n> File \\\"./configure.py\\\", line 103, in is_os400_pase\\n> return self._platform == 'os400' or os.uname().sysname.startswith('OS400')\\n> AttributeError: 'tuple' object has no attribute 'sysname'\\n> \\n> Possibly the ninja guys would take a patch for that (or maybe\\n> this is a you-need-python-3 case?). I do see /usr/include/spawn.h\\n> on that platform, so there's room to hope it'd work.\\n\\nThat does seem like it'd be a issue. Briefly trawling the ninja git log it\\ndoes look like there's regular-ish maintenance stuff for AIX, so I'd hope we\\ncould get it fixed. I do suspect it's just a python3 issue, as Thomas noted.\\n\\n\\n> Based on these results, I doubt that ninja will give us trouble on any\\n> platform that isn't old enough to get its drivers license.\\n\\nAgreed.\\n\\nThere's also alternative compatible ninja implementation in C99 ([1]). But I\\nthink it's minimum requirements aren't actually lower than ninja's (it says it\\nrequires posix 2008).\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n[1] https://github.com/michaelforney/samurai\\n\\n\\n\",\n \"msg_date\": \"Mon, 1 Nov 2021 21:21:23 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Portability report: ninja\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2021-11-01 15:25:08 -0400, Tom Lane wrote:\\n>> Fortunately for my\\n>> purposes here, it seems to work with fairly old Python --- I built\\n>> successfully with python 2.6.2, though not with 2.4.1.\\n\\n> meson will need a newer python though...\\n\\nYup. I was just trying to establish what the lower bound was for\\nninja, since they (ahem) don't document it.\\n\\nThe meson docs say it needs Python >= 3.6, which I think is going\\nto be an issue, but I've not researched it yet. What I was mainly\\ncurious about today was whether anything using ninja as the backend\\ncould meet our portability goals. My mind's more or less at ease\\non that, but I am afraid that requiring 3.6 is going to be a problem\\non LTS distributions. That's a topic for another thread though.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 02 Nov 2021 00:48:14 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Portability report: ninja\"\n },\n {\n \"msg_contents\": \"On 2021-11-02 09:27:17 +1300, Thomas Munro wrote:\\n> On Tue, Nov 2, 2021 at 8:25 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > 6. While configure.py thinks it knows what to do on AIX, it fails\\n> > on AIX 7.1 and 7.2:\\n> >\\n> > Traceback (most recent call last):\\n> > File \\\"./configure.py\\\", line 544, in <module>\\n> > if platform.is_aix() and not platform.is_os400_pase():\\n> > File \\\"./configure.py\\\", line 103, in is_os400_pase\\n> > return self._platform == 'os400' or os.uname().sysname.startswith('OS400')\\n> > AttributeError: 'tuple' object has no attribute 'sysname'\\n> \\n> Yeah, the result type changed: os.uname()[0] works on python2 AND python3.\\n\\nFWIW, with python3 ninja builds and tests successfully on AIX.\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Nov 2021 14:42:49 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Portability report: ninja\"\n }\n]","isConversational":false}}},{"rowIdx":53038,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nTrying to insert NULL value to the Identity column defined by \"GENERATED BY\nDEFAULT\" is disallowed, but there can be use cases where the user would\nlike to have an identity column where manual NULL insertion is required(and\nit should not error-out by Postgres).\n\nHow about having a new type for the Identity column as \"GENERATED BY\nDEFAULT ON NULL\", which will allow manual NULL insertion and internally\nNULL value will be replaced by Sequence NextValue?\n\nORACLE is supporting this feature by having a similar Identity column type\nas below:\n=======================================================================\nSQL> CREATE TABLE itest1 (id1 INTEGER GENERATED BY DEFAULT ON NULL\nAS IDENTITY, id2 INTEGER);\n\nTable created.\n\nSQL> INSERT INTO itest1 VALUES (NULL, 10); --Supported with GENERATED BY\nDEFAULT ON NULL\n\n1 row created.\n\nSQL> INSERT INTO itest1 VALUES (1,30);\n\n1 row created.\n\nSQL> INSERT INTO itest1 (id2) VALUES (20);\n\n1 row created.\n\nSQL> SELECT * FROM itest1;\n\n ID1 ID2\n---------- ----------\n 1 10\n 1 30\n 2 20\n================================================================\n\nI think it is good to have support for GENERATED BY DEFAULT ON NULL in\nPostgres.\n\nThoughts?\n\nThanks,\nHimanshu\n\nHi,Trying to insert NULL value to the Identity column defined by \"GENERATED BY DEFAULT\" is disallowed, but there can be use cases where the user would like to have an identity column where manual NULL insertion is required(and it should not error-out by Postgres).How about having a new type for the Identity column as \"GENERATED BY DEFAULT ON NULL\", which will allow manual NULL insertion and internally NULL value will be replaced by Sequence NextValue?ORACLE is supporting this feature by having a similar Identity column type as below:=======================================================================SQL> CREATE TABLE itest1 (id1 INTEGER GENERATED BY DEFAULT ON NULL AS IDENTITY, id2 INTEGER);Table created.SQL> INSERT INTO itest1 VALUES (NULL, 10); --Supported with GENERATED BY DEFAULT ON NULL1 row created.SQL> INSERT INTO itest1 VALUES (1,30);1 row created.SQL> INSERT INTO itest1 (id2) VALUES (20);1 row created.SQL> SELECT * FROM itest1; ID1 ID2---------- ---------- 1 10 1 30 2 20================================================================I think it is good to have support for GENERATED BY DEFAULT ON NULL in Postgres.Thoughts?Thanks,Himanshu","msg_date":"Tue, 2 Nov 2021 16:49:57 +0530","msg_from":"Himanshu Upadhyaya <upadhyaya.himanshu@gmail.com>","msg_from_op":true,"msg_subject":"Should we support new definition for Identity column : GENERATED BY\n DEFAULT ON NULL?"},{"msg_contents":"On 11/2/21 12:19 PM, Himanshu Upadhyaya wrote:\n> Hi,\n> \n> Trying to insert NULL value to the Identity column defined by \"GENERATED BY\n> DEFAULT\" is disallowed, but there can be use cases where the user would\n> like to have an identity column where manual NULL insertion is required(and\n> it should not error-out by Postgres).\n\nWhat could possibly be the use case for this?\n\n> Thoughts?\n\nI don't like it.\n-- \nVik Fearing\n\n\n","msg_date":"Wed, 3 Nov 2021 11:07:58 +0100","msg_from":"Vik Fearing <vik@postgresfriends.org>","msg_from_op":false,"msg_subject":"Re: Should we support new definition for Identity column : GENERATED\n BY DEFAULT ON NULL?"},{"msg_contents":"On Wed, Nov 03, 2021 at 11:07:58AM +0100, Vik Fearing wrote:\n> On 11/2/21 12:19 PM, Himanshu Upadhyaya wrote:\n> > Hi,\n> > \n> > Trying to insert NULL value to the Identity column defined by \"GENERATED BY\n> > DEFAULT\" is disallowed, but there can be use cases where the user would\n> > like to have an identity column where manual NULL insertion is required(and\n> > it should not error-out by Postgres).\n> \n> What could possibly be the use case for this?\n\nUnfortunately, the PREPARE/EXECUTE infrastructure, and not just for\nPostgreSQL, has no way of passing along DEFAULT explicitly, i.e. it\ncan only explicitly pass literals and NULL, so people come up with\nworkarounds like this. I keep saying \"explicitly\" because one of the\nworkarounds is to pass DEFAULT by omission in the target list, which\nis not a great way to handle anything, being a communication by\nsilence and all. I'm thinking that lack is the problem we should\nactually address.\n\n> > Thoughts?\n> \n> I don't like it.\n\nNeither do I, but I do get some of the motivation behind it. I don't\nsuppose the standard actually says much about PREPARE, or whether we\nshould care about anything it happens to say that gets in the way of\ndoing something more helpful.\n\nBest,\nDavid.\n-- \nDavid Fetter <david(at)fetter(dot)org> http://fetter.org/\nPhone: +1 415 235 3778\n\nRemember to vote!\nConsider donating to Postgres: http://www.postgresql.org/about/donate\n\n\n","msg_date":"Wed, 3 Nov 2021 14:39:20 +0000","msg_from":"David Fetter <david@fetter.org>","msg_from_op":false,"msg_subject":"Re: Should we support new definition for Identity column : GENERATED\n BY DEFAULT ON NULL?"},{"msg_contents":"On Wed, 3 Nov 2021 at 14:39, David Fetter <david@fetter.org> wrote:\n\n> Unfortunately, the PREPARE/EXECUTE infrastructure, and not just for\n> PostgreSQL, has no way of passing along DEFAULT explicitly, i.e. it\n>\n\nMy $.02: I'd be much happier with the idea of changing those\nobviously-deficient interfaces to allow explicit DEFAULT than of having an\nexplicit NULL treated differently depending on the column type.\n\nGeoff\n\nOn Wed, 3 Nov 2021 at 14:39, David Fetter <david@fetter.org> wrote:Unfortunately, the PREPARE/EXECUTE infrastructure, and not just for\nPostgreSQL, has no way of passing along DEFAULT explicitly, i.e. it My $.02: I'd be much happier with the idea of changing those obviously-deficient interfaces to allow explicit DEFAULT than of having an explicit NULL treated differently depending on the column type.Geoff","msg_date":"Wed, 3 Nov 2021 14:54:12 +0000","msg_from":"Geoff Winkless <pgsqladmin@geoff.dj>","msg_from_op":false,"msg_subject":"Re: Should we support new definition for Identity column : GENERATED\n BY DEFAULT ON NULL?"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nTrying to insert NULL value to the Identity column defined by \\\"GENERATED BY\\nDEFAULT\\\" is disallowed, but there can be use cases where the user would\\nlike to have an identity column where manual NULL insertion is required(and\\nit should not error-out by Postgres).\\n\\nHow about having a new type for the Identity column as \\\"GENERATED BY\\nDEFAULT ON NULL\\\", which will allow manual NULL insertion and internally\\nNULL value will be replaced by Sequence NextValue?\\n\\nORACLE is supporting this feature by having a similar Identity column type\\nas below:\\n=======================================================================\\nSQL> CREATE TABLE itest1 (id1 INTEGER GENERATED BY DEFAULT ON NULL\\nAS IDENTITY, id2 INTEGER);\\n\\nTable created.\\n\\nSQL> INSERT INTO itest1 VALUES (NULL, 10); --Supported with GENERATED BY\\nDEFAULT ON NULL\\n\\n1 row created.\\n\\nSQL> INSERT INTO itest1 VALUES (1,30);\\n\\n1 row created.\\n\\nSQL> INSERT INTO itest1 (id2) VALUES (20);\\n\\n1 row created.\\n\\nSQL> SELECT * FROM itest1;\\n\\n ID1 ID2\\n---------- ----------\\n 1 10\\n 1 30\\n 2 20\\n================================================================\\n\\nI think it is good to have support for GENERATED BY DEFAULT ON NULL in\\nPostgres.\\n\\nThoughts?\\n\\nThanks,\\nHimanshu\\n\\nHi,Trying to insert NULL value to the Identity column defined by \\\"GENERATED BY DEFAULT\\\" is disallowed, but there can be use cases where the user would like to have an identity column where manual NULL insertion is required(and it should not error-out by Postgres).How about having a new type for the Identity column as \\\"GENERATED BY DEFAULT ON NULL\\\", which will allow manual NULL insertion and internally NULL value will be replaced by Sequence NextValue?ORACLE is supporting this feature by having a similar Identity column type as below:=======================================================================SQL> CREATE TABLE itest1 (id1 INTEGER GENERATED BY DEFAULT ON NULL AS IDENTITY, id2 INTEGER);Table created.SQL> INSERT INTO itest1 VALUES (NULL, 10); --Supported with GENERATED BY DEFAULT ON NULL1 row created.SQL> INSERT INTO itest1 VALUES (1,30);1 row created.SQL> INSERT INTO itest1 (id2) VALUES (20);1 row created.SQL> SELECT * FROM itest1; ID1 ID2---------- ---------- 1 10 1 30 2 20================================================================I think it is good to have support for GENERATED BY DEFAULT ON NULL in Postgres.Thoughts?Thanks,Himanshu\",\n \"msg_date\": \"Tue, 2 Nov 2021 16:49:57 +0530\",\n \"msg_from\": \"Himanshu Upadhyaya <upadhyaya.himanshu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Should we support new definition for Identity column : GENERATED BY\\n DEFAULT ON NULL?\"\n },\n {\n \"msg_contents\": \"On 11/2/21 12:19 PM, Himanshu Upadhyaya wrote:\\n> Hi,\\n> \\n> Trying to insert NULL value to the Identity column defined by \\\"GENERATED BY\\n> DEFAULT\\\" is disallowed, but there can be use cases where the user would\\n> like to have an identity column where manual NULL insertion is required(and\\n> it should not error-out by Postgres).\\n\\nWhat could possibly be the use case for this?\\n\\n> Thoughts?\\n\\nI don't like it.\\n-- \\nVik Fearing\\n\\n\\n\",\n \"msg_date\": \"Wed, 3 Nov 2021 11:07:58 +0100\",\n \"msg_from\": \"Vik Fearing <vik@postgresfriends.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should we support new definition for Identity column : GENERATED\\n BY DEFAULT ON NULL?\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 03, 2021 at 11:07:58AM +0100, Vik Fearing wrote:\\n> On 11/2/21 12:19 PM, Himanshu Upadhyaya wrote:\\n> > Hi,\\n> > \\n> > Trying to insert NULL value to the Identity column defined by \\\"GENERATED BY\\n> > DEFAULT\\\" is disallowed, but there can be use cases where the user would\\n> > like to have an identity column where manual NULL insertion is required(and\\n> > it should not error-out by Postgres).\\n> \\n> What could possibly be the use case for this?\\n\\nUnfortunately, the PREPARE/EXECUTE infrastructure, and not just for\\nPostgreSQL, has no way of passing along DEFAULT explicitly, i.e. it\\ncan only explicitly pass literals and NULL, so people come up with\\nworkarounds like this. I keep saying \\\"explicitly\\\" because one of the\\nworkarounds is to pass DEFAULT by omission in the target list, which\\nis not a great way to handle anything, being a communication by\\nsilence and all. I'm thinking that lack is the problem we should\\nactually address.\\n\\n> > Thoughts?\\n> \\n> I don't like it.\\n\\nNeither do I, but I do get some of the motivation behind it. I don't\\nsuppose the standard actually says much about PREPARE, or whether we\\nshould care about anything it happens to say that gets in the way of\\ndoing something more helpful.\\n\\nBest,\\nDavid.\\n-- \\nDavid Fetter <david(at)fetter(dot)org> http://fetter.org/\\nPhone: +1 415 235 3778\\n\\nRemember to vote!\\nConsider donating to Postgres: http://www.postgresql.org/about/donate\\n\\n\\n\",\n \"msg_date\": \"Wed, 3 Nov 2021 14:39:20 +0000\",\n \"msg_from\": \"David Fetter <david@fetter.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should we support new definition for Identity column : GENERATED\\n BY DEFAULT ON NULL?\"\n },\n {\n \"msg_contents\": \"On Wed, 3 Nov 2021 at 14:39, David Fetter <david@fetter.org> wrote:\\n\\n> Unfortunately, the PREPARE/EXECUTE infrastructure, and not just for\\n> PostgreSQL, has no way of passing along DEFAULT explicitly, i.e. it\\n>\\n\\nMy $.02: I'd be much happier with the idea of changing those\\nobviously-deficient interfaces to allow explicit DEFAULT than of having an\\nexplicit NULL treated differently depending on the column type.\\n\\nGeoff\\n\\nOn Wed, 3 Nov 2021 at 14:39, David Fetter <david@fetter.org> wrote:Unfortunately, the PREPARE/EXECUTE infrastructure, and not just for\\nPostgreSQL, has no way of passing along DEFAULT explicitly, i.e. it My $.02: I'd be much happier with the idea of changing those obviously-deficient interfaces to allow explicit DEFAULT than of having an explicit NULL treated differently depending on the column type.Geoff\",\n \"msg_date\": \"Wed, 3 Nov 2021 14:54:12 +0000\",\n \"msg_from\": \"Geoff Winkless <pgsqladmin@geoff.dj>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should we support new definition for Identity column : GENERATED\\n BY DEFAULT ON NULL?\"\n }\n]","isConversational":false}}},{"rowIdx":53039,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi!\n\nI'm here to answer your questions about contrib/pg_query_state.\n>> I only took a quick look at pg_query_state, I have some questions.\n\n>> pg_query_state seems using shm_mq to expose the plan information, but\n>> there was a discussion that this kind of architecture would be tricky\n>> to do properly [1].\n>> Does pg_query_state handle difficulties listed on the discussion?\n> [1] \n> https://www.postgresql.org/message-id/9a50371e15e741e295accabc72a41df1%40oss.nttdata.com\n\nI doubt that it was the right link.\nBut on the topic I will say that extension really use shared memory, \ninteraction is implemented by sending / receiving messages. This \narchitecture provides the required reliability and convenience.\n\n>> It seems the caller of the pg_query_state() has to wait until the\n>> target process pushes the plan information into shared memory, can it\n>> lead to deadlock situations?\n>> I came up with this question because when trying to make a view for\n>> memory contexts of other backends, we encountered deadlock situations.\n>> After all, we gave up view design and adopted sending signal and\n>> logging.\n> \n> Discussion at the following URL.\n> https://www.postgresql.org/message-id/9a50371e15e741e295accabc72a41df1%40oss.nttdata.com\n\nBefore extracting information about side process we check its state. \nInformation will only be retrieved for a process willing to provide it. \nOtherwise, we will receive an error message about impossibility of \ngetting query execution statistics + process status. Also checking fact \nof extracting your own status exists. This is even verified in tests.\n\nThanks for your attention.\nJust in case, I am ready to discuss this topic in more detail.\n\nAbout overhead:\n> I haven't measured it yet, but I believe that the overhead for backends\n> which are not called pg_log_current_plan() would be slight since the\n> patch just adds the logic for saving QueryDesc on ExecutorRun().\n> The overhead for backends which is called pg_log_current_plan() might\n> not slight, but since the target process are assumed dealing with\n> long-running query and the user want to know its plan, its overhead\n> would be worth the cost.\nI think it would be useful for us to have couple of examples with a \ndifferent number of rows compared to using without this functionality.\n\nHope this helps.\n\n-- \nEkaterina Sokolova\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company\n\n\n","msg_date":"Tue, 02 Nov 2021 14:32:59 +0300","msg_from":"Ekaterina Sokolova <e.sokolova@postgrespro.ru>","msg_from_op":true,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"This patch no longer applies on top of HEAD, please submit a rebased version.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Wed, 3 Nov 2021 13:48:51 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2021-11-02 20:32, Ekaterina Sokolova wrote:\nThanks for your response!\n\n> Hi!\n> \n> I'm here to answer your questions about contrib/pg_query_state.\n>>> I only took a quick look at pg_query_state, I have some questions.\n> \n>>> pg_query_state seems using shm_mq to expose the plan information, but\n>>> there was a discussion that this kind of architecture would be tricky\n>>> to do properly [1].\n>>> Does pg_query_state handle difficulties listed on the discussion?\n>> [1] \n>> https://www.postgresql.org/message-id/9a50371e15e741e295accabc72a41df1%40oss.nttdata.com\n> \n> I doubt that it was the right link.\n\nSorry for make you confused, here is the link.\n\n \nhttps://www.postgresql.org/message-id/CA%2BTgmobkpFV0UB67kzXuD36--OFHwz1bs%3DL_6PZbD4nxKqUQMw%40mail.gmail.com\n\n> But on the topic I will say that extension really use shared memory,\n> interaction is implemented by sending / receiving messages. This\n> architecture provides the required reliability and convenience.\n\nAs described in the link, using shared memory for this kind of work \nwould need DSM and It'd be also necessary to exchange information \nbetween requestor and responder.\n\nFor example, when I looked at a little bit of pg_query_state code, it \nlooks like the size of the queue is fixed at QUEUE_SIZE, and I wonder \nhow plans that exceed QUEUE_SIZE are handled.\n\n>>> It seems the caller of the pg_query_state() has to wait until the\n>>> target process pushes the plan information into shared memory, can it\n>>> lead to deadlock situations?\n>>> I came up with this question because when trying to make a view for\n>>> memory contexts of other backends, we encountered deadlock \n>>> situations.\n>>> After all, we gave up view design and adopted sending signal and\n>>> logging.\n>> \n>> Discussion at the following URL.\n>> https://www.postgresql.org/message-id/9a50371e15e741e295accabc72a41df1%40oss.nttdata.com\n> \n> Before extracting information about side process we check its state.\n> Information will only be retrieved for a process willing to provide\n> it. Otherwise, we will receive an error message about impossibility of\n> getting query execution statistics + process status. Also checking\n> fact of extracting your own status exists. This is even verified in\n> tests.\n> \n> Thanks for your attention.\n> Just in case, I am ready to discuss this topic in more detail.\n\nI imagined the following procedure.\nDoes it cause dead lock in pg_query_state?\n\n- session1\nBEGIN; TRUNCATE t;\n\n- session2\nBEGIN; TRUNCATE t; -- wait\n\n- session1\nSELECT * FROM pg_query_state(<pid of session>); -- wait and dead locked?\n\n> About overhead:\n>> I haven't measured it yet, but I believe that the overhead for \n>> backends\n>> which are not called pg_log_current_plan() would be slight since the\n>> patch just adds the logic for saving QueryDesc on ExecutorRun().\n>> The overhead for backends which is called pg_log_current_plan() might\n>> not slight, but since the target process are assumed dealing with\n>> long-running query and the user want to know its plan, its overhead\n>> would be worth the cost.\n> I think it would be useful for us to have couple of examples with a\n> different number of rows compared to using without this functionality.\n\nDo you have any expectaion that the number of rows would affect the \nperformance of this functionality?\nThis patch adds some codes to ExecutorRun(), but I thought the number of \nrows would not give impact on the performance.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION\n\n\n","msg_date":"Thu, 04 Nov 2021 21:49:21 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi!\n\nYou forgot my last fix to build correctly on Mac. I have added it.\n\nAbout our discussion of pg_query_state:\n\ntorikoshia писал 2021-11-04 15:49:\n>> I doubt that it was the right link.\n> Sorry for make you confused, here is the link.\n> https://www.postgresql.org/message-id/CA%2BTgmobkpFV0UB67kzXuD36--OFHwz1bs%3DL_6PZbD4nxKqUQMw%40mail.gmail.com\n\nThank you. I'll see it soon.\n\n> I imagined the following procedure.\n> Does it cause dead lock in pg_query_state?\n> \n> - session1\n> BEGIN; TRUNCATE t;\n> \n> - session2\n> BEGIN; TRUNCATE t; -- wait\n> \n> - session1\n> SELECT * FROM pg_query_state(<pid of session>); -- wait and dead \n> locked?\n\nAs I know, pg_query_state use non-blocking read and write. I have wrote \nfew tests trying to deadlock it (on 14 version), but all finished \ncorrectly.\n\nHave a nice day. Please feel free to contact me if you need any further \ninformation.\n\n-- \nEkaterina Sokolova\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company","msg_date":"Wed, 17 Nov 2021 16:44:51 +0300","msg_from":"Ekaterina Sokolova <e.sokolova@postgrespro.ru>","msg_from_op":true,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2021-11-17 22:44, Ekaterina Sokolova wrote:\n> Hi!\n> \n> You forgot my last fix to build correctly on Mac. I have added it.\n\nThanks for the notification!\nSince the patch could not be applied to the HEAD anymore, I also updated \nit.\n\n> \n> About our discussion of pg_query_state:\n> \n> torikoshia писал 2021-11-04 15:49:\n>>> I doubt that it was the right link.\n>> Sorry for make you confused, here is the link.\n>> https://www.postgresql.org/message-id/CA%2BTgmobkpFV0UB67kzXuD36--OFHwz1bs%3DL_6PZbD4nxKqUQMw%40mail.gmail.com\n> \n> Thank you. I'll see it soon.\n> \n>> I imagined the following procedure.\n>> Does it cause dead lock in pg_query_state?\n>> \n>> - session1\n>> BEGIN; TRUNCATE t;\n>> \n>> - session2\n>> BEGIN; TRUNCATE t; -- wait\n>> \n>> - session1\n>> SELECT * FROM pg_query_state(<pid of session>); -- wait and dead \n>> locked?\n> \n> As I know, pg_query_state use non-blocking read and write. I have\n> wrote few tests trying to deadlock it (on 14 version), but all\n> finished correctly.\n> \n> Have a nice day. Please feel free to contact me if you need any\n> further information.\n\nThanks for your information and help!\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Fri, 26 Nov 2021 12:39:48 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2021-11-26 12:39, torikoshia wrote:\n> Since the patch could not be applied to the HEAD anymore, I also \n> updated it.\n\nUpdated the patch for fixing compiler warning about the format on \nwindows.\n\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Fri, 07 Jan 2022 14:30:37 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2022-01-07 14:30, torikoshia wrote:\n\n> Updated the patch for fixing compiler warning about the format on \n> windows.\n\nI got another compiler warning, updated the patch again.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Fri, 07 Jan 2022 20:58:19 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"\r\n\r\nOn 2022/01/07 20:58, torikoshia wrote:\r\n> On 2022-01-07 14:30, torikoshia wrote:\r\n> \r\n>> Updated the patch for fixing compiler warning about the format on windows.\r\n> \r\n> I got another compiler warning, updated the patch again.\r\n\r\nThanks for updating the patch!\r\n\r\nI ran the following query every 0.1s by using \\watch psql command from three different sessions while make installcheck test was running.\r\n\r\n SELECT pg_log_query_plan(pid) FROM pg_stat_activity;\r\n\r\n\r\nAnd then I got the segmentation fault as follows.\r\n\r\n2022-01-07 21:40:32 JST [postmaster] LOG: server process (PID 51017) was terminated by signal 11: Segmentation fault: 11\r\n2022-01-07 21:40:32 JST [postmaster] DETAIL: Failed process was running: select description, (test_conv(inbytes, 'utf8', 'utf8')).* from utf8_verification_inputs;\r\n2022-01-07 21:40:32 JST [postmaster] LOG: terminating any other active server processes\r\n2022-01-07 21:40:32 JST [postmaster] LOG: all server processes terminated; reinitializing\r\n\r\n\r\nThe backtrace I got from the core file was:\r\n\r\n(lldb) target create --core \"/cores/core.51017\"\r\nCore file '/cores/core.51017' (x86_64) was loaded.\r\n(lldb) bt\r\n* thread #1, stop reason = signal SIGSTOP\r\n * frame #0: 0x00007fff20484552 libsystem_platform.dylib`_platform_strlen + 18\r\n frame #1: 0x000000011076e36c postgres`dopr(target=0x00007ffedfd1b1d8, format=\"\\n\", args=0x00007ffedfd1b450) at snprintf.c:444:20\r\n frame #2: 0x000000011076e133 postgres`pg_vsnprintf(str=\"Query Text: \\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\"..., count=1024, fmt=\"%s: %s\\n\", args=0x00007ffedfd1b450) at snprintf.c:195:2\r\n frame #3: 0x000000011075e51d postgres`pvsnprintf(buf=\"Query Text: \\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\"..., len=1024, fmt=\"%s: %s\\n\", args=0x00007ffedfd1b450) at psprintf.c:110:13\r\n frame #4: 0x000000011076052d postgres`appendStringInfoVA(str=0x00007fe56804bec0, fmt=\"%s: %s\\n\", args=0x00007ffedfd1b450) at stringinfo.c:149:13\r\n frame #5: 0x0000000110760449 postgres`appendStringInfo(str=0x00007fe56804bec0, fmt=\"%s: %s\\n\") at stringinfo.c:103:12\r\n frame #6: 0x000000011016f6cf postgres`ExplainProperty(qlabel=\"Query Text\", unit=0x0000000000000000, value=\"\", numeric=false, es=0x00007fe56804be28) at explain.c:4390:5\r\n frame #7: 0x000000011016eace postgres`ExplainPropertyText(qlabel=\"Query Text\", value=\"\", es=0x00007fe56804be28) at explain.c:4435:2\r\n frame #8: 0x000000011016ea97 postgres`ExplainQueryText(es=0x00007fe56804be28, queryDesc=0x00007fe568022320) at explain.c:969:3\r\n frame #9: 0x000000011016faeb postgres`ProcessLogCurrentPlanInterrupt at explain.c:5009:2\r\n frame #10: 0x00000001104dda84 postgres`ProcessInterrupts at postgres.c:3373:3\r\n frame #11: 0x0000000110c81c57 plpgsql.so`exec_stmts(estate=0x00007ffedfd1bab8, stmts=0x00007fe568050e00) at pl_exec.c:1994:3\r\n frame #12: 0x0000000110c80a78 plpgsql.so`exec_stmt_block(estate=0x00007ffedfd1bab8, block=0x00007fe56801c570) at pl_exec.c:1777:9\r\n frame #13: 0x0000000110c81c91 plpgsql.so`exec_stmts(estate=0x00007ffedfd1bab8, stmts=0x00007fe56801c5c8) at pl_exec.c:1999:10\r\n frame #14: 0x0000000110c80f21 plpgsql.so`exec_stmt_block(estate=0x00007ffedfd1bab8, block=0x00007fe56801c678) at pl_exec.c:1926:8\r\n frame #15: 0x0000000110c7ed92 plpgsql.so`exec_toplevel_block(estate=0x00007ffedfd1bab8, block=0x00007fe56801c678) at pl_exec.c:1617:7\r\n frame #16: 0x0000000110c7d3e4 plpgsql.so`plpgsql_exec_function(func=0x00007fe56a00aa20, fcinfo=0x00007fe56803c888, simple_eval_estate=0x0000000000000000, simple_eval_resowner=0x0000000000000000, procedure_resowner=0x0000000000000000, atomic=true) at pl_exec.c:611:7\r\n frame #17: 0x0000000110c9bf12 plpgsql.so`plpgsql_call_handler(fcinfo=0x00007fe56803c888) at pl_handler.c:277:13\r\n frame #18: 0x0000000110228b06 postgres`ExecInterpExpr(state=0x00007fe56803bf88, econtext=0x00007fe56803ba70, isnull=0x00007ffedfd1bf7f) at execExprInterp.c:725:8\r\n frame #19: 0x000000011024858b postgres`ExecEvalExprSwitchContext(state=0x00007fe56803bf88, econtext=0x00007fe56803ba70, isNull=0x00007ffedfd1bf7f) at executor.h:339:13\r\n frame #20: 0x00000001102482aa postgres`ExecProject(projInfo=0x00007fe56803bf80) at executor.h:373:9\r\n frame #21: 0x0000000110247e2f postgres`ExecScan(node=0x00007fe56803b958, accessMtd=(postgres`SeqNext at nodeSeqscan.c:51), recheckMtd=(postgres`SeqRecheck at nodeSeqscan.c:90)) at execScan.c:238:12\r\n frame #22: 0x000000011028c575 postgres`ExecSeqScan(pstate=0x00007fe56803b958) at nodeSeqscan.c:112:9\r\n frame #23: 0x000000011023bbe2 postgres`ExecProcNode(node=0x00007fe56803b958) at executor.h:257:9\r\n frame #24: 0x0000000110237471 postgres`ExecutePlan(estate=0x00007fe56803b720, planstate=0x00007fe56803b958, use_parallel_mode=false, operation=CMD_SELECT, sendTuples=true, numberTuples=0, direction=ForwardScanDirection, dest=0x00007fe568046ac8, execute_once=true) at execMain.c:1561:10\r\n frame #25: 0x0000000110237320 postgres`standard_ExecutorRun(queryDesc=0x00007fe568037320, direction=ForwardScanDirection, count=0, execute_once=true) at execMain.c:371:3\r\n frame #26: 0x00000001102370b8 postgres`ExecutorRun(queryDesc=0x00007fe568037320, direction=ForwardScanDirection, count=0, execute_once=true) at execMain.c:313:3\r\n frame #27: 0x00000001104e6bdb postgres`PortalRunSelect(portal=0x00007fe567825920, forward=true, count=0, dest=0x00007fe568046ac8) at pquery.c:921:4\r\n frame #28: 0x00000001104e65ba postgres`PortalRun(portal=0x00007fe567825920, count=9223372036854775807, isTopLevel=true, run_once=true, dest=0x00007fe568046ac8, altdest=0x00007fe568046ac8, qc=0x00007ffedfd1c350) at pquery.c:765:18\r\n frame #29: 0x00000001104e1861 postgres`exec_simple_query(query_string=\"select description, (test_conv(inbytes, 'utf8', 'utf8')).* from utf8_verification_inputs;\") at postgres.c:1216:10\r\n frame #30: 0x00000001104e099c postgres`PostgresMain(dbname=\"regression\", username=\"postgres\") at postgres.c:4505:7\r\n frame #31: 0x00000001103ee2a2 postgres`BackendRun(port=0x00007fe567704080) at postmaster.c:4594:2\r\n frame #32: 0x00000001103ed868 postgres`BackendStartup(port=0x00007fe567704080) at postmaster.c:4322:3\r\n frame #33: 0x00000001103ec64c postgres`ServerLoop at postmaster.c:1802:7\r\n frame #34: 0x00000001103e9e29 postgres`PostmasterMain(argc=3, argv=0x00007fe567405e70) at postmaster.c:1474:11\r\n frame #35: 0x00000001102c8b69 postgres`main(argc=3, argv=0x00007fe567405e70) at main.c:198:3\r\n frame #36: 0x00007fff2045cf3d libdyld.dylib`start + 1\r\n\r\nRegards,\r\n\r\n-- \r\nFujii Masao\r\nAdvanced Computing Technology Center\r\nResearch and Development Headquarters\r\nNTT DATA CORPORATION\r\n","msg_date":"Fri, 7 Jan 2022 21:54:31 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi,\n\nOn Fri, Jan 07, 2022 at 09:54:31PM +0900, Fujii Masao wrote:\n> \n> On 2022/01/07 20:58, torikoshia wrote:\n> > On 2022-01-07 14:30, torikoshia wrote:\n> > \n> > > Updated the patch for fixing compiler warning about the format on windows.\n> > \n> > I got another compiler warning, updated the patch again.\n> \n> Thanks for updating the patch!\n> \n> I ran the following query every 0.1s by using \\watch psql command from three different sessions while make installcheck test was running.\n> \n> SELECT pg_log_query_plan(pid) FROM pg_stat_activity;\n> \n> \n> And then I got the segmentation fault as follows.\n> \n> 2022-01-07 21:40:32 JST [postmaster] LOG: server process (PID 51017) was terminated by signal 11: Segmentation fault: 11\n> 2022-01-07 21:40:32 JST [postmaster] DETAIL: Failed process was running: select description, (test_conv(inbytes, 'utf8', 'utf8')).* from utf8_verification_inputs;\n> 2022-01-07 21:40:32 JST [postmaster] LOG: terminating any other active server processes\n> 2022-01-07 21:40:32 JST [postmaster] LOG: all server processes terminated; reinitializing\n> \n> \n> The backtrace I got from the core file was:\n> \n> (lldb) target create --core \"/cores/core.51017\"\n> Core file '/cores/core.51017' (x86_64) was loaded.\n> (lldb) bt\n> * thread #1, stop reason = signal SIGSTOP\n> * frame #0: 0x00007fff20484552 libsystem_platform.dylib`_platform_strlen + 18\n> frame #1: 0x000000011076e36c postgres`dopr(target=0x00007ffedfd1b1d8, format=\"\\n\", args=0x00007ffedfd1b450) at snprintf.c:444:20\n> frame #2: 0x000000011076e133 postgres`pg_vsnprintf(str=\"Query Text: \\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\"..., count=1024, fmt=\"%s: %s\\n\", args=0x00007ffedfd1b450) at snprintf.c:195:2\n\nOn top of the visibly use-after-pfree memory issue, the patch doesn't apply\nanymore:\n\nhttp://cfbot.cputube.org/patch_36_3142.log\n=== Applying patches on top of PostgreSQL commit ID a18b6d2dc288dfa6e7905ede1d4462edd6a8af47 ===\n=== applying patch ./v16-0001-log-running-query-plan.patch\n[...]\npatching file src/test/regress/expected/misc_functions.out\nHunk #1 FAILED at 134.\nHunk #2 succeeded at 163 (offset 7 lines).\nHunk #3 succeeded at 180 (offset 7 lines).\n1 out of 3 hunks FAILED -- saving rejects to file src/test/regress/expected/misc_functions.out.rej\npatching file src/test/regress/sql/misc_functions.sql\nHunk #1 FAILED at 31.\n1 out of 1 hunk FAILED -- saving rejects to file src/test/regress/sql/misc_functions.sql.rej\n\nI'm switching the patch to Waiting on Author.\n\n\n","msg_date":"Fri, 14 Jan 2022 14:38:46 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2022-01-14 15:38, Julien Rouhaud wrote:\n> Hi,\n> \n> On Fri, Jan 07, 2022 at 09:54:31PM +0900, Fujii Masao wrote:\n\n>> I ran the following query every 0.1s by using \\watch psql command from \n>> three different sessions while make installcheck test was running.\n>> \n>> SELECT pg_log_query_plan(pid) FROM pg_stat_activity;\n>> \n>> \n>> And then I got the segmentation fault as follows.\n>> \n>> 2022-01-07 21:40:32 JST [postmaster] LOG: server process (PID 51017) \n>> was terminated by signal 11: Segmentation fault: 11\n>> 2022-01-07 21:40:32 JST [postmaster] DETAIL: Failed process was \n>> running: select description, (test_conv(inbytes, 'utf8', 'utf8')).* \n>> from utf8_verification_inputs;\n>> 2022-01-07 21:40:32 JST [postmaster] LOG: terminating any other \n>> active server processes\n>> 2022-01-07 21:40:32 JST [postmaster] LOG: all server processes \n>> terminated; reinitializing\n>> \n>> \n>> The backtrace I got from the core file was:\n>> \n>> (lldb) target create --core \"/cores/core.51017\"\n>> Core file '/cores/core.51017' (x86_64) was loaded.\n>> (lldb) bt\n>> * thread #1, stop reason = signal SIGSTOP\n>> * frame #0: 0x00007fff20484552 \n>> libsystem_platform.dylib`_platform_strlen + 18\n>> frame #1: 0x000000011076e36c \n>> postgres`dopr(target=0x00007ffedfd1b1d8, format=\"\\n\", \n>> args=0x00007ffedfd1b450) at snprintf.c:444:20\n>> frame #2: 0x000000011076e133 postgres`pg_vsnprintf(str=\"Query \n>> Text: \n>>\n\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x\n7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\n\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x\n7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\n\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\\x7f\"..., \n>> count=1024, fmt=\"%s: %s\\n\", args=0x00007ffedfd1b450) at \n>> snprintf.c:195:2\n\nThanks for testing!\nI could reproduce the situation and confirmed almost the same backtrace.\n\nThe cause seems that the previous patch didn't do cleanup when the \ntransactions were aborted. It leaded segmentation faults when \nreferencing ActiveQueryDesc-> after the transaction were aborted.\n\nThe attached patch added cleanups when transactions and subtransactions \nare aborted.\n\nBTW, the latest patch for pg_log_backtrace() added a new function for \ndeduplicating codes for pg_log_*()[1].\nIt seems better to use it for pg_log_query_plan() after the patch is \nmerged.\n\n\n[1] \nhttps://www.postgresql.org/message-id/flat/20211115194251.GP17618%40telsasoft.com#4e681d04c78867b2cce44b3cb1791bf5\n duplicate code to a new function\nCheckPostgresProcessId which can be used by pg_log_backtrace,\n\n> On top of the visibly use-after-pfree memory issue, the patch doesn't \n> apply\n\nThanks for letting me know!\nAs mentioned above, I updated the patch.\n\nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Mon, 24 Jan 2022 14:33:05 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"\n\nOn 2022/01/24 14:33, torikoshia wrote:\n> As mentioned above, I updated the patch.\n\nThanks for updating the patch!\n\nHere are another review comments:\n\n+LOG: plan of the query running on backend with PID 17793 is:\n\nThis seems not the same as what actually logged.\n\n+\t\tereport(WARNING,\n+\t\t\t\t(errmsg(\"PID %d is not a PostgreSQL server process\", pid)));\n\nLike commit 7fa945b857 changed, this warning message should be \"PID %d is not a PostgreSQL backend process\"?\n\n+\tif (SendProcSignal(pid, PROCSIG_LOG_CURRENT_PLAN, InvalidBackendId) < 0)\n\nproc->backendId should be specified instead of InvalidBackendId, to speed up the processing in SendProcSignal()?\n\n+\tPROCSIG_LOG_CURRENT_PLAN, /* ask backend to log plan of the current query */\n+volatile sig_atomic_t LogCurrentPlanPending = false;\n+extern void HandleLogCurrentPlanInterrupt(void);\n+extern void ProcessLogCurrentPlanInterrupt(void);\n\nIsn't it better to use the names that are more consistent with the function name, i.e., pg_log_query_plan? For example, PROCSIG_LOG_QUERY_PLAN instead of PROCSIG_LOG_CURRENT_PLAN?\n\n+\t\tereport(LOG_SERVER_ONLY,\n+\t\t\t\terrmsg(\"backend with PID %d is not running a query\",\n+\t\t\t\t\tMyProcPid));\n\nerrhidestmt(true) and errhidecontext(true) need to be added, don't they? Otherwise, for example, if pg_log_query_plan() is executed after debug_query_string is set but before ActiveQueryDesc is set, STATEMENT message would be output even though the message saying \"not running a query\" is output. Which seems confusing.\n\n+\thash_seq_init(&status, GetLockMethodLocalHash());\n+\twhile ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)\n+\t{\n+\t\tif (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\n\nWhy did you use the search for local lock hash instead of IsPageLockHeld flag variable, to check whether a page lock is held or not? Because there is the corner case where the interrupt is processed after the local lock is registered into the hash but before IsPageLockHeld is enabled?\n\nThere is the case where the request to log a query plan is skipped even while the target backend is running a query. If this happens, users can just retry pg_log_query_plan(). These things should be documented?\n\n+\t\t\tereport(LOG_SERVER_ONLY,\n+\t\t\t\terrmsg(\"backend with PID %d is holding a page lock. Try again\",\n+\t\t\t\t\tMyProcPid));\n\nIt seems more proper to output this message in DETAIL or HINT, instead. So how about something like the following messages?\n\nLOG: could not log the query plan\nDETAIL: query plan cannot be logged while page level lock is being held\nHINT: Try pg_log_query_plan() after a few ....\n\nRegards,\n\n-- \nFujii Masao\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Thu, 27 Jan 2022 20:18:37 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2022-01-27 20:18, Fujii Masao wrote:\n> Here are another review comments:\n\nThanks for reviewing!\n\n> +LOG: plan of the query running on backend with PID 17793 is:\n> \n> This seems not the same as what actually logged.\n\nModified.\n\n> +\t\tereport(WARNING,\n> +\t\t\t\t(errmsg(\"PID %d is not a PostgreSQL server process\", pid)));\n> \n> Like commit 7fa945b857 changed, this warning message should be \"PID %d\n> is not a PostgreSQL backend process\"?\n\nModified.\n\n> +\tif (SendProcSignal(pid, PROCSIG_LOG_CURRENT_PLAN, InvalidBackendId) < \n> 0)\n> \n> proc->backendId should be specified instead of InvalidBackendId, to\n> speed up the processing in SendProcSignal()?\n\nAgreed. Modified.\n\n> +\tPROCSIG_LOG_CURRENT_PLAN, /* ask backend to log plan of the current \n> query */\n> +volatile sig_atomic_t LogCurrentPlanPending = false;\n> +extern void HandleLogCurrentPlanInterrupt(void);\n> +extern void ProcessLogCurrentPlanInterrupt(void);\n> \n> Isn't it better to use the names that are more consistent with the\n> function name, i.e., pg_log_query_plan? For example,\n> PROCSIG_LOG_QUERY_PLAN instead of PROCSIG_LOG_CURRENT_PLAN?\n\nAgreed.\nI removed 'current' from the variable and function names and used \n'query' instead.\n\n> +\t\tereport(LOG_SERVER_ONLY,\n> +\t\t\t\terrmsg(\"backend with PID %d is not running a query\",\n> +\t\t\t\t\tMyProcPid));\n> \n> errhidestmt(true) and errhidecontext(true) need to be added, don't\n> they? Otherwise, for example, if pg_log_query_plan() is executed after\n> debug_query_string is set but before ActiveQueryDesc is set, STATEMENT\n> message would be output even though the message saying \"not running a\n> query\" is output. Which seems confusing.\n\nAgreed. Added errhidestmt(true) and errhidecontext(true).\n\n> +\thash_seq_init(&status, GetLockMethodLocalHash());\n> +\twhile ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)\n> +\t{\n> +\t\tif (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\n> \n> Why did you use the search for local lock hash instead of\n> IsPageLockHeld flag variable, to check whether a page lock is held or\n> not? Because there is the corner case where the interrupt is processed\n> after the local lock is registered into the hash but before\n> IsPageLockHeld is enabled?\n\nAs far as I read CheckAndSetLockHeld(), IsPageLockHeld can be used only \nwhen USE_ASSERT_CHECKING is enabled.\nSince removing USE_ASSERT_CHECKING from CheckAndSetLockHeld() would give \nperformance impact on every granting/removing local lock, I used the \nsearch for local local hash.\n\n> There is the case where the request to log a query plan is skipped\n> even while the target backend is running a query. If this happens,\n> users can just retry pg_log_query_plan(). These things should be\n> documented?\n\nAgreed.\nAdded following:\n\n + Note that there is the case where the request to log a query\n + plan is skipped even while the target backend is running a\n + query due to lock conflict avoidance.\n + If this happens, users can just retry pg_log_query_plan().\n\n|\n\n> +\t\t\tereport(LOG_SERVER_ONLY,\n> +\t\t\t\terrmsg(\"backend with PID %d is holding a page lock. Try again\",\n> +\t\t\t\t\tMyProcPid));\n> \n> It seems more proper to output this message in DETAIL or HINT,\n> instead. So how about something like the following messages?\n> \n> LOG: could not log the query plan\n> DETAIL: query plan cannot be logged while page level lock is being held\n> HINT: Try pg_log_query_plan() after a few ....\n\nAgreed.\nI felt the HINT message 'after a few ...' is difficult to describe, and \nwrote as below.\n\n| HINT: Retrying pg_log_query_plan() might succeed since the lock \nduration of page level locks are usually short\n\nHow do you think?\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Fri, 28 Jan 2022 17:45:49 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"\n\nOn 2022/01/28 17:45, torikoshia wrote:\n>> There is the case where the request to log a query plan is skipped\n>> even while the target backend is running a query. If this happens,\n>> users can just retry pg_log_query_plan(). These things should be\n>> documented?\n> \n> Agreed.\n> Added following:\n> \n> + Note that there is the case where the request to log a query\n> + plan is skipped even while the target backend is running a\n> + query due to lock conflict avoidance.\n> + If this happens, users can just retry pg_log_query_plan().\n\nThis may cause users to misunderstand that pg_log_query_plan() fails while the target backend is holding *any* locks? Isn't it better to mention \"page-level locks\", instead? So how about the following?\n\n--------------------------\nNote that the request to log the query plan will be ignored if it's received during a short period while the target backend is holding a page-level lock.\n--------------------------\n\n\n>> + ereport(LOG_SERVER_ONLY,\n>> + errmsg(\"backend with PID %d is holding a page lock. Try again\",\n>> + MyProcPid));\n>>\n>> It seems more proper to output this message in DETAIL or HINT,\n>> instead. So how about something like the following messages?\n>>\n>> LOG: could not log the query plan\n>> DETAIL: query plan cannot be logged while page level lock is being held\n>> HINT: Try pg_log_query_plan() after a few ....\n> \n> Agreed.\n> I felt the HINT message 'after a few ...' is difficult to describe, and wrote as below.\n> \n> | HINT: Retrying pg_log_query_plan() might succeed since the lock duration of page level locks are usually short\n> \n> How do you think?\n\nOr we don't need HINT message?\n\n\n+\t\t\t\terrmsg(\"could not log the query plan\"),\n+\t\t\t\terrdetail(\"query plan cannot be logged while page level lock is being held\"),\n\nIn detail message, the first word of sentences should be capitalized. How about \"Cannot log the query plan while holding page-level lock.\", instead?\n\n\nThanks for updating the patch! Here are some review comments.\n\n+ <row>\n+ <entry role=\"func_table_entry\"><para role=\"func_signature\">\n+ <indexterm>\n+ <primary>pg_log_query_plan</primary>\n+ </indexterm>\n\nThis entry is placed before one for pg_log_backend_memory_contexts(). But it should be *after* that since those entries seem to be placed in alphabetical order in the table?\n\n\n+ Requests to log the plan of the query currently running on the\n+ backend with specified process ID along with the untruncated\n+ query string.\n\nOther descriptions about logging of query string seem not to mention something like \"untruncated query string\". For example, auto_explain, log_statement, etc. Why do we need to mention \"along with the untruncated query string\" specially for pg_log_query_plan()?\n\n\n+ Note that nested statements (statements executed inside a function) are not\n+ considered for logging. Only the plan of the most deeply nested query is logged.\n\nNow the plan of even nested statement can be logged. So this description needs to be updated?\n\n\n@@ -440,6 +450,7 @@ standard_ExecutorFinish(QueryDesc *queryDesc)\n \n \tMemoryContextSwitchTo(oldcontext);\n \n+\tActiveQueryDesc = NULL;\n\nActiveQueryDesc seems unnecessary. Why does ActiveQueryDesc need to be reset to NULL in standard_ExecutorFinish()?\n\n\nCurrently even during ProcessLogQueryPlanInterrupt(), pg_log_query_plan() can be call and another ProcessLogQueryPlanInterrupt() can be executed. So repeatable re-entrances to ProcessLogQueryPlanInterrupt() could cause \"stack depth limit exceeded\" error. To avoid this, shouldn't we make ProcessLogQueryPlanInterrupt() do nothing and return immediately, if it's called during another ProcessLogQueryPlanInterrupt()?\n\npg_log_backend_memory_contexts() also might have the same issue.\n\nRegards,\n\n-- \nFujii Masao\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Tue, 1 Feb 2022 01:51:11 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"At Tue, 1 Feb 2022 01:51:11 +0900, Fujii Masao <masao.fujii@oss.nttdata.com> wrote in \n> \n> \n> On 2022/01/28 17:45, torikoshia wrote:\n> >> There is the case where the request to log a query plan is skipped\n> >> even while the target backend is running a query. If this happens,\n> >> users can just retry pg_log_query_plan(). These things should be\n> >> documented?\n> > Agreed.\n> > Added following:\n> > + Note that there is the case where the request to log a\n> > query\n> > + plan is skipped even while the target backend is running a\n> > + query due to lock conflict avoidance.\n> > + If this happens, users can just retry pg_log_query_plan().\n> \n> This may cause users to misunderstand that pg_log_query_plan() fails\n> while the target backend is holding *any* locks? Isn't it better to\n> mention \"page-level locks\", instead? So how about the following?\n> \n> --------------------------\n> Note that the request to log the query plan will be ignored if it's\n> received during a short period while the target backend is holding a\n> page-level lock.\n> --------------------------\n> \n> \n> >> + ereport(LOG_SERVER_ONLY,\n> >> + errmsg(\"backend with PID %d is holding a page\n> >> lock. Try again\",\n> >> + MyProcPid));\n> >>\n> >> It seems more proper to output this message in DETAIL or HINT,\n> >> instead. So how about something like the following messages?\n> >>\n> >> LOG: could not log the query plan\n> >> DETAIL: query plan cannot be logged while page level lock is being\n> >> held\n> >> HINT: Try pg_log_query_plan() after a few ....\n> > Agreed.\n> > I felt the HINT message 'after a few ...' is difficult to describe,\n> > and wrote as below.\n> > | HINT: Retrying pg_log_query_plan() might succeed since the lock\n> > | duration of page level locks are usually short\n> > How do you think?\n> \n> Or we don't need HINT message?\n> \n> \n> +\t\t\t\terrmsg(\"could not log the query plan\"),\n> + errdetail(\"query plan cannot be logged while page level lock is\n> being held\"),\n> \n> In detail message, the first word of sentences should be\n> capitalized. How about \"Cannot log the query plan while holding\n> page-level lock.\", instead?\n> \n> \n> Thanks for updating the patch! Here are some review comments.\n> \n> + <row>\n> + <entry role=\"func_table_entry\"><para role=\"func_signature\">\n> + <indexterm>\n> + <primary>pg_log_query_plan</primary>\n> + </indexterm>\n> \n> This entry is placed before one for\n> pg_log_backend_memory_contexts(). But it should be *after* that since\n> those entries seem to be placed in alphabetical order in the table?\n> \n> \n> + Requests to log the plan of the query currently running on the\n> + backend with specified process ID along with the untruncated\n> + query string.\n> \n> Other descriptions about logging of query string seem not to mention\n> something like \"untruncated query string\". For example, auto_explain,\n> log_statement, etc. Why do we need to mention \"along with the\n> untruncated query string\" specially for pg_log_query_plan()?\n> \n> \n> + Note that nested statements (statements executed inside a function)\n> are not\n> + considered for logging. Only the plan of the most deeply nested\n> query is logged.\n> \n> Now the plan of even nested statement can be logged. So this\n> description needs to be updated?\n> \n> \n> @@ -440,6 +450,7 @@ standard_ExecutorFinish(QueryDesc *queryDesc)\n> \tMemoryContextSwitchTo(oldcontext);\n> +\tActiveQueryDesc = NULL;\n> \n> ActiveQueryDesc seems unnecessary. Why does ActiveQueryDesc need to be\n> reset to NULL in standard_ExecutorFinish()?\n> \n> \n> Currently even during ProcessLogQueryPlanInterrupt(),\n> pg_log_query_plan() can be call and another\n> ProcessLogQueryPlanInterrupt() can be executed. So repeatable\n> re-entrances to ProcessLogQueryPlanInterrupt() could cause \"stack\n> depth limit exceeded\" error. To avoid this, shouldn't we make\n> ProcessLogQueryPlanInterrupt() do nothing and return immediately, if\n> it's called during another ProcessLogQueryPlanInterrupt()?\n> \n> pg_log_backend_memory_contexts() also might have the same issue.\n\nGood catch.\n\nBy the way, I'm anxious about the following part and I'd like to\nremove it.\n\n+\t * Ensure no page lock is held on this process.\n\nIt seems to me what is wrong is ginInsertCleanup(), not this feature.\nAs I read the comment for the assertion, I don't believe we want to\nallow CFI while holding a page lock. And AFAICS the function is the\nonly point where doing that. (It is the alone user of LockPage()...)\n\nThis is the assertion.\n\nlock.c: 902\n> \t/*\n> \t * We don't acquire any other heavyweight lock while holding the page lock\n> \t * except for relation extension.\n> \t */\n> \tAssert(!IsPageLockHeld ||\n> \t\t (locktag->locktag_type == LOCKTAG_RELATION_EXTEND));\n\nIt is added by a recent commit 72e78d831ab5550c39f2dcc7cc5d44c406ec3dc2.\n\n> * Similar to relation extension, page locks are also held for a short\n> * duration, so imposing such a restriction won't hurt.\n\nI don't believe a path involving vacuum_delay_point() calls is\nshort-duration'ed.\n\nhttps://www.postgresql.org/message-id/CAH2-WznzCPUKnOV%2Bre30_rHLCkqQWm2JTSVjTCAei9LySTc2pw%40mail.gmail.com\n\n> One thing that really bothers me about commit e2c79e14 is that\n> LockPage() is called, not LockBuffer(). GIN had no LockPage() calls\n> before that commit, and is now the only code in the entire system that\n> calls LockPage()/ConditionalLockPage() (the hash am no longer uses\n> page heavyweight locks following recent work there).\n\nI agree to the discussion. Can't we use other mechanism here to get\nrid of the Lockpage()?\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Tue, 01 Feb 2022 17:27:57 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"\n\nOn 2022/02/01 17:27, Kyotaro Horiguchi wrote:\n>> * Similar to relation extension, page locks are also held for a short\n>> * duration, so imposing such a restriction won't hurt.\n> \n> I don't believe a path involving vacuum_delay_point() calls is\n> short-duration'ed.\n\nYes.\n\n\n>> One thing that really bothers me about commit e2c79e14 is that\n>> LockPage() is called, not LockBuffer(). GIN had no LockPage() calls\n>> before that commit, and is now the only code in the entire system that\n>> calls LockPage()/ConditionalLockPage() (the hash am no longer uses\n>> page heavyweight locks following recent work there).\n> \n> I agree to the discussion. Can't we use other mechanism here to get\n> rid of the Lockpage()?\n\nI have no good idea for that yet, but I agree it's better to get rid of page level lock.\n\nRegards,\n\n-- \nFujii Masao\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Tue, 1 Feb 2022 23:11:03 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"At Tue, 1 Feb 2022 23:11:03 +0900, Fujii Masao <masao.fujii@oss.nttdata.com> wrote in \n> >> One thing that really bothers me about commit e2c79e14 is that\n> >> LockPage() is called, not LockBuffer(). GIN had no LockPage() calls\n> >> before that commit, and is now the only code in the entire system that\n> >> calls LockPage()/ConditionalLockPage() (the hash am no longer uses\n> >> page heavyweight locks following recent work there).\n> > I agree to the discussion. Can't we use other mechanism here to get\n> > rid of the Lockpage()?\n> \n> I have no good idea for that yet, but I agree it's better to get rid\n> of page level lock.\n\nIt's my turn?\n\nThe page lock is used to hold-off simultaneous cleanups on the same\nindex. ShareUpdateExclusive lock on the index relation works that\nway. In that path it seems like we are always holding a RowExclusive\nlock, so it seems to me we can use ShareUpdateExclusive for our\npurpose.\n\nThere might be a false blocking case when another backend is holding a\nconflicting lock on the index. They are, Share, ShareRowExclusive,\nExclusive and AccessExclusive. The last three cases don't seem worth\ndiscussion. I'm not sure about Share and Share Row cases. AFAICS\nShare lock is taken on an index in ATExecReplicaIdentity, and there no\nuse of ShareRowExclusive lock on an index. It's no use discussing about\nexplicit locking.\n\nSo aren't we able to use ShareUpdateExclusive lock for that?\n\nIn the attached patch, ginInsertCleanup has an extra check for such\nstronger locks not being held. At least \"make check\" doesn't cause\nthe extra assertion to fire.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\ndiff --git a/src/backend/access/gin/ginfast.c b/src/backend/access/gin/ginfast.c\nindex 7409fdc165..1af9a69abb 100644\n--- a/src/backend/access/gin/ginfast.c\n+++ b/src/backend/access/gin/ginfast.c\n@@ -791,20 +791,29 @@ ginInsertCleanup(GinState *ginstate, bool full_clean,\n \tbool\t\tfsm_vac = false;\n \tSize\t\tworkMemory;\n \n-\t/*\n+\t/*r\n \t * We would like to prevent concurrent cleanup process. For that we will\n \t * lock metapage in exclusive mode using LockPage() call. Nobody other\n \t * will use that lock for metapage, so we keep possibility of concurrent\n \t * insertion into pending list\n \t */\n \n+\t/*\n+\t * we use ShareUpdateExclusive lock on this relation to hold-off concurrent\n+\t * cleanup\n+\t */\n+\tAssert(!CheckRelationLockedByMe(index, ShareUpdateExclusiveLock, false));\n+\n+\t/* tentative debug-purpose assertion for stronger locks */\n+\tAssert(!CheckRelationLockedByMe(index, ShareLock, true));\n+\t\n \tif (forceCleanup)\n \t{\n \t\t/*\n \t\t * We are called from [auto]vacuum/analyze or gin_clean_pending_list()\n \t\t * and we would like to wait concurrent cleanup to finish.\n \t\t */\n-\t\tLockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock);\n+\t\tLockRelation(index, ShareUpdateExclusiveLock);\n \t\tworkMemory =\n \t\t\t(IsAutoVacuumWorkerProcess() && autovacuum_work_mem != -1) ?\n \t\t\tautovacuum_work_mem : maintenance_work_mem;\n@@ -816,7 +825,7 @@ ginInsertCleanup(GinState *ginstate, bool full_clean,\n \t\t * just exit in hope that concurrent process will clean up pending\n \t\t * list.\n \t\t */\n-\t\tif (!ConditionalLockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock))\n+\t\tif (!ConditionalLockRelation(index, ShareUpdateExclusiveLock))\n \t\t\treturn;\n \t\tworkMemory = work_mem;\n \t}\n@@ -830,7 +839,7 @@ ginInsertCleanup(GinState *ginstate, bool full_clean,\n \t{\n \t\t/* Nothing to do */\n \t\tUnlockReleaseBuffer(metabuffer);\n-\t\tUnlockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock);\n+\t\tUnlockRelation(index, ShareUpdateExclusiveLock);\n \t\treturn;\n \t}\n \n@@ -1002,7 +1011,7 @@ ginInsertCleanup(GinState *ginstate, bool full_clean,\n \t\tpage = BufferGetPage(buffer);\n \t}\n \n-\tUnlockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock);\n+\tUnlockRelation(index, ShareUpdateExclusiveLock);\n \tReleaseBuffer(metabuffer);\n \n \t/*","msg_date":"Wed, 02 Feb 2022 16:49:57 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"At Wed, 02 Feb 2022 16:49:57 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in \n> So aren't we able to use ShareUpdateExclusive lock for that?\n> \n> In the attached patch, ginInsertCleanup has an extra check for such\n> stronger locks not being held. At least \"make check\" doesn't cause\n> the extra assertion to fire.\n\nActually, the discussion is a bit dubious. What we need really to\ncheck is wheter such locks are not held on an index *elsewhere*.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Wed, 02 Feb 2022 17:04:21 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"2022-02-01 01:51, Fujii Masao wrote:\nThanks for reviewing and suggestions!\n\n>> + Note that there is the case where the request to log a \n>> query\n>> + plan is skipped even while the target backend is running a\n>> + query due to lock conflict avoidance.\n>> + If this happens, users can just retry pg_log_query_plan().\n> \n> This may cause users to misunderstand that pg_log_query_plan() fails\n> while the target backend is holding *any* locks? Isn't it better to\n> mention \"page-level locks\", instead? So how about the following?\n> \n> --------------------------\n> Note that the request to log the query plan will be ignored if it's\n> received during a short period while the target backend is holding a\n> page-level lock.\n> --------------------------\n\nAgreed.\n\n>>> + ereport(LOG_SERVER_ONLY,\n>>> + errmsg(\"backend with PID %d is holding a page lock. \n>>> Try again\",\n>>> + MyProcPid));\n>>> \n>>> It seems more proper to output this message in DETAIL or HINT,\n>>> instead. So how about something like the following messages?\n>>> \n>>> LOG: could not log the query plan\n>>> DETAIL: query plan cannot be logged while page level lock is being \n>>> held\n>>> HINT: Try pg_log_query_plan() after a few ....\n>> \n>> Agreed.\n>> I felt the HINT message 'after a few ...' is difficult to describe, \n>> and wrote as below.\n>> \n>> | HINT: Retrying pg_log_query_plan() might succeed since the lock \n>> duration of page level locks are usually short\n>> \n>> How do you think?\n> \n> Or we don't need HINT message?\n\nRemoved the HINT message.\n\n> + errmsg(\"could not log the query plan\"),\n> + errdetail(\"query plan cannot be logged \n> while page level lock is\n> being held\"),\n> \n> In detail message, the first word of sentences should be capitalized.\n> How about \"Cannot log the query plan while holding page-level lock.\",\n> instead?\n\nAgreed.\n\n> Thanks for updating the patch! Here are some review comments.\n> \n> + <row>\n> + <entry role=\"func_table_entry\"><para role=\"func_signature\">\n> + <indexterm>\n> + <primary>pg_log_query_plan</primary>\n> + </indexterm>\n> \n> This entry is placed before one for pg_log_backend_memory_contexts().\n> But it should be *after* that since those entries seem to be placed in\n> alphabetical order in the table?\n\nModified it.\n\n> + Requests to log the plan of the query currently running on the\n> + backend with specified process ID along with the untruncated\n> + query string.\n> \n> Other descriptions about logging of query string seem not to mention\n> something like \"untruncated query string\". For example, auto_explain,\n> log_statement, etc. Why do we need to mention \"along with the\n> untruncated query string\" specially for pg_log_query_plan()?\n\nModified it as below:\n\n Requests to log the plan of the query currently running on \nthe\n - backend with specified process ID along with the \nuntruncated\n - query string.\n - They will be logged at <literal>LOG</literal> message level \nand\n + backend with specified process ID.\n + It will be logged at <literal>LOG</literal> message level \nand\n\n> + Note that nested statements (statements executed inside a \n> function) are not\n> + considered for logging. Only the plan of the most deeply nested\n> query is logged.\n> \n> Now the plan of even nested statement can be logged. So this\n> description needs to be updated?\n\nModified it as below:\n\n - Note that nested statements (statements executed inside a \nfunction) are not\n - considered for logging. Only the plan of the most deeply nested \nquery is logged.\n + Note that when the statements are executed inside a function, \nonly the\n + plan of the most deeply nested query is logged.\n\n> @@ -440,6 +450,7 @@ standard_ExecutorFinish(QueryDesc *queryDesc)\n> MemoryContextSwitchTo(oldcontext);\n> + ActiveQueryDesc = NULL;\n> \n> ActiveQueryDesc seems unnecessary. Why does ActiveQueryDesc need to be\n> reset to NULL in standard_ExecutorFinish()?\n\nActiveQueryDesc should not be reset in standard_ExecutorFinish().\nRemoved it.\n\n> Currently even during ProcessLogQueryPlanInterrupt(),\n> pg_log_query_plan() can be call and another\n> ProcessLogQueryPlanInterrupt() can be executed. So repeatable\n> re-entrances to ProcessLogQueryPlanInterrupt() could cause \"stack\n> depth limit exceeded\" error. To avoid this, shouldn't we make\n> ProcessLogQueryPlanInterrupt() do nothing and return immediately, if\n> it's called during another ProcessLogQueryPlanInterrupt()?\n> \n> pg_log_backend_memory_contexts() also might have the same issue.\n\nAs you pointed out offlist, the issue could occur even when both \npg_log_backend_memory_contexts and pg_log_query_plan are called and it \nmay be better to make another patch.\n\n\nYou also pointed out offlist that it would be necessary to call \nLockErrorCleanup() if ProcessLogQueryPlanInterrupt() can incur ERROR.\nAFAICS it can call ereport(ERROR), i.e., palloc0() in NewExplainState(), \nso I added PG_TRY/CATCH and make it call LockErrorCleanup() when ERROR \noccurs.\n\n\n\nOn 2022-02-01 17:27, Kyotaro Horiguchi wrote:\n\nThanks for reviewing Horiguchi-san!\n\n> By the way, I'm anxious about the following part and I'd like to\n> remove it.\n\nI also think it would be nice if it's possible.\n> \n> + * Ensure no page lock is held on this process.\n> \n> It seems to me what is wrong is ginInsertCleanup(), not this feature.\n\n> Actually, the discussion is a bit dubious. What we need really to\ncheck is wheter such locks are not held on an index *elsewhere*.\n\nSince I'm not sure how long it will take to discuss this point, the \nattached patch is based on the current HEAD at this time.\nI also think it may be better to discuss it on another thread.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Wed, 02 Feb 2022 21:59:12 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Wed, Feb 2, 2022 at 7:59 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> 2022-02-01 01:51, Fujii Masao wrote:\n<snip>\n> > + Note that nested statements (statements executed inside a\n> > function) are not\n> > + considered for logging. Only the plan of the most deeply nested\n> > query is logged.\n> >\n> > Now the plan of even nested statement can be logged. So this\n> > description needs to be updated?\n>\n> Modified it as below:\n>\n> - Note that nested statements (statements executed inside a\n> function) are not\n> - considered for logging. Only the plan of the most deeply nested\n> query is logged.\n> + Note that when the statements are executed inside a function,\n> only the\n> + plan of the most deeply nested query is logged.\n>\n\nMinor nit, but I think the \"the\" is superfluous.. ie.\n\n\"Note that when statements are executed inside a function,\nonly the plan of the most deeply nested query is logged\"\n\n<snip>\n> On 2022-02-01 17:27, Kyotaro Horiguchi wrote:\n>\n> Thanks for reviewing Horiguchi-san!\n>\n> > By the way, I'm anxious about the following part and I'd like to\n> > remove it.\n>\n> I also think it would be nice if it's possible.\n> >\n> > + * Ensure no page lock is held on this process.\n> >\n> > It seems to me what is wrong is ginInsertCleanup(), not this feature.\n>\n> > Actually, the discussion is a bit dubious. What we need really to\n> check is wheter such locks are not held on an index *elsewhere*.\n>\n> Since I'm not sure how long it will take to discuss this point, the\n> attached patch is based on the current HEAD at this time.\n> I also think it may be better to discuss it on another thread.\n>\n\nWhile I agree on the above points, IMHO I don't believe it should be a\nshow-stopper for adding this functionality to v15, but we have a few\nmore commitments before we get to that point.\n\n\nRobert Treat\nhttps://xzilla.net\n\n\n","msg_date":"Thu, 3 Feb 2022 03:09:07 -0500","msg_from":"Robert Treat <rob@xzilla.net>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"\n\nOn 2022/02/02 21:59, torikoshia wrote:\n>> This may cause users to misunderstand that pg_log_query_plan() fails\n>> while the target backend is holding *any* locks? Isn't it better to\n>> mention \"page-level locks\", instead? So how about the following?\n>>\n>> --------------------------\n>> Note that the request to log the query plan will be ignored if it's\n>> received during a short period while the target backend is holding a\n>> page-level lock.\n>> --------------------------\n> \n> Agreed.\n\nOn second thought, this note is confusing rather than helpful? Because the users don't know when and what operation needs page-level lock. So now I'm thinking it's better to remove this note.\n\n\n> As you pointed out offlist, the issue could occur even when both pg_log_backend_memory_contexts and pg_log_query_plan are called and it may be better to make another patch.\n\nOK.\n\n\n> You also pointed out offlist that it would be necessary to call LockErrorCleanup() if ProcessLogQueryPlanInterrupt() can incur ERROR.\n> AFAICS it can call ereport(ERROR), i.e., palloc0() in NewExplainState(), so I added PG_TRY/CATCH and make it call LockErrorCleanup() when ERROR occurs.\n\nAs we discussed off-list, this treatment might not be necessary. Because when ERROR or FATAL error happens, AbortTransaction() is called and LockErrorCleanup() is also called inside there.\n\n\n> Since I'm not sure how long it will take to discuss this point, the attached patch is based on the current HEAD at this time.\n\nThanks for updating the patch!\n\n@@ -5048,6 +5055,12 @@ AbortSubTransaction(void)\n \t */\n \tPG_SETMASK(&UnBlockSig);\n \n+\t/*\n+\t * When ActiveQueryDesc is referenced after abort, some of its elements\n+\t * are freed. To avoid accessing them, reset ActiveQueryDesc here.\n+\t */\n+\tActiveQueryDesc = NULL;\n\nAbortSubTransaction() should reset ActiveQueryDesc to save_ActiveQueryDesc that ExecutorRun() set, instead of NULL? Otherwise ActiveQueryDesc of top-level statement will be unavailable after subtransaction is aborted in the nested statements.\n\nFor example, no plan is logged while the following \"select pg_sleep(test())\" is running, because the exception inside test() function aborts the subtransaction and resets ActiveQueryDesc to NULL.\n\ncreate or replace function test () returns int as $$\nbegin\n perform 1/0;\nexception when others then\n return 30;\nend;\n$$ language plpgsql;\n\nselect pg_sleep(test());\n\n\n-CREATE ROLE regress_log_memory;\n+CREATE ROLE regress_log;\n\nIsn't this name too generic? How about regress_log_function, for example?\n\nRegards,\n\n-- \nFujii Masao\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Tue, 8 Feb 2022 01:13:44 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"At Tue, 8 Feb 2022 01:13:44 +0900, Fujii Masao <masao.fujii@oss.nttdata.com> wrote in \n> \n> \n> On 2022/02/02 21:59, torikoshia wrote:\n> >> This may cause users to misunderstand that pg_log_query_plan() fails\n> >> while the target backend is holding *any* locks? Isn't it better to\n> >> mention \"page-level locks\", instead? So how about the following?\n> >>\n> >> --------------------------\n> >> Note that the request to log the query plan will be ignored if it's\n> >> received during a short period while the target backend is holding a\n> >> page-level lock.\n> >> --------------------------\n> > Agreed.\n> \n> On second thought, this note is confusing rather than helpful? Because\n> the users don't know when and what operation needs page-level lock. So\n> now I'm thinking it's better to remove this note.\n\n*I* agree to removing the note. And the following error message looks\nas mysterious as the note is, and the DETAIL doesn't help..\n\n\t\t\tereport(LOG_SERVER_ONLY,\n+\t\t\t\terrmsg(\"could not log the query plan\"),\n+\t\t\t\terrdetail(\"Cannot log the query plan while holding page-level lock.\"));\n+\t\t\thash_seq_term(&status);\n\nWe should tell the command can be retried soon, like this?\n\n\"LOG: ignored request for logging query plan due to lock confilcts\"\n\"HINT: You can try again in a moment.\"\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Tue, 08 Feb 2022 17:18:52 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2022-02-03 17:09, Robert Treat wrote:\n> On Wed, Feb 2, 2022 at 7:59 AM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> \n>> 2022-02-01 01:51, Fujii Masao wrote:\n> <snip>\n>> > + Note that nested statements (statements executed inside a\n>> > function) are not\n>> > + considered for logging. Only the plan of the most deeply nested\n>> > query is logged.\n>> >\n>> > Now the plan of even nested statement can be logged. So this\n>> > description needs to be updated?\n>> \n>> Modified it as below:\n>> \n>> - Note that nested statements (statements executed inside a\n>> function) are not\n>> - considered for logging. Only the plan of the most deeply \n>> nested\n>> query is logged.\n>> + Note that when the statements are executed inside a \n>> function,\n>> only the\n>> + plan of the most deeply nested query is logged.\n>> \n> \n> Minor nit, but I think the \"the\" is superfluous.. ie.\n> \n> \"Note that when statements are executed inside a function,\n> only the plan of the most deeply nested query is logged\"\n\nThanks! Modified it.\n\n\nOn 2022-02-08 01:13, Fujii Masao wrote:\nThanks for the comments!\n\n> On 2022/02/02 21:59, torikoshia wrote:\n>>> This may cause users to misunderstand that pg_log_query_plan() fails\n>>> while the target backend is holding *any* locks? Isn't it better to\n>>> mention \"page-level locks\", instead? So how about the following?\n>>> \n>>> --------------------------\n>>> Note that the request to log the query plan will be ignored if it's\n>>> received during a short period while the target backend is holding a\n>>> page-level lock.\n>>> --------------------------\n>> \n>> Agreed.\n> \n> On second thought, this note is confusing rather than helpful? Because\n> the users don't know when and what operation needs page-level lock. So\n> now I'm thinking it's better to remove this note.\n\nRemoved it.\n> \n> \n>> As you pointed out offlist, the issue could occur even when both \n>> pg_log_backend_memory_contexts and pg_log_query_plan are called and it \n>> may be better to make another patch.\n> \n> OK.\n> \n> \n>> You also pointed out offlist that it would be necessary to call \n>> LockErrorCleanup() if ProcessLogQueryPlanInterrupt() can incur ERROR.\n>> AFAICS it can call ereport(ERROR), i.e., palloc0() in \n>> NewExplainState(), so I added PG_TRY/CATCH and make it call \n>> LockErrorCleanup() when ERROR occurs.\n> \n> As we discussed off-list, this treatment might not be necessary.\n> Because when ERROR or FATAL error happens, AbortTransaction() is\n> called and LockErrorCleanup() is also called inside there.\n\nAgreed.\n\n>> Since I'm not sure how long it will take to discuss this point, the \n>> attached patch is based on the current HEAD at this time.\n> \n> Thanks for updating the patch!\n> \n> @@ -5048,6 +5055,12 @@ AbortSubTransaction(void)\n> \t */\n> \tPG_SETMASK(&UnBlockSig);\n> +\t/*\n> +\t * When ActiveQueryDesc is referenced after abort, some of its \n> elements\n> +\t * are freed. To avoid accessing them, reset ActiveQueryDesc here.\n> +\t */\n> +\tActiveQueryDesc = NULL;\n> \n> AbortSubTransaction() should reset ActiveQueryDesc to\n> save_ActiveQueryDesc that ExecutorRun() set, instead of NULL?\n> Otherwise ActiveQueryDesc of top-level statement will be unavailable\n> after subtransaction is aborted in the nested statements.\n> \n> For example, no plan is logged while the following \"select\n> pg_sleep(test())\" is running, because the exception inside test()\n> function aborts the subtransaction and resets ActiveQueryDesc to NULL.\n> \n> create or replace function test () returns int as $$\n> begin\n> perform 1/0;\n> exception when others then\n> return 30;\n> end;\n> $$ language plpgsql;\n> \n> select pg_sleep(test());\n\nAgreed.\n\nBTW, since the above example results in calling ExecutorRun() only once, \nthe output didn't differ even after ActiveQueryDesc is reset to \nsave_ActiveQueryDesc.\n\nThe below definition of test() worked as expected.\n\n create or replace function test () returns int as $$\n begin\n perform 1;\n perform 1/0;\n exception when others then\n return 30;\n end;\n $$ language plpgsql;\n\n> \n> \n> -CREATE ROLE regress_log_memory;\n> +CREATE ROLE regress_log;\n> \n> Isn't this name too generic? How about regress_log_function, for \n> example?\n\nAgreed.\n\n\nOn 2022-02-08 17:18, Kyotaro Horiguchi wrote:\n> At Tue, 8 Feb 2022 01:13:44 +0900, Fujii Masao\n> <masao.fujii@oss.nttdata.com> wrote in\n>> \n>> \n>> On 2022/02/02 21:59, torikoshia wrote:\n>> >> This may cause users to misunderstand that pg_log_query_plan() fails\n>> >> while the target backend is holding *any* locks? Isn't it better to\n>> >> mention \"page-level locks\", instead? So how about the following?\n>> >>\n>> >> --------------------------\n>> >> Note that the request to log the query plan will be ignored if it's\n>> >> received during a short period while the target backend is holding a\n>> >> page-level lock.\n>> >> --------------------------\n>> > Agreed.\n>> \n>> On second thought, this note is confusing rather than helpful? Because\n>> the users don't know when and what operation needs page-level lock. So\n>> now I'm thinking it's better to remove this note.\n> \n> *I* agree to removing the note. And the following error message looks\n> as mysterious as the note is, and the DETAIL doesn't help..\n> \n> \t\t\tereport(LOG_SERVER_ONLY,\n> +\t\t\t\terrmsg(\"could not log the query plan\"),\n> +\t\t\t\terrdetail(\"Cannot log the query plan while holding page-level \n> lock.\"));\n> +\t\t\thash_seq_term(&status);\n> \n> We should tell the command can be retried soon, like this?\n> \n> \"LOG: ignored request for logging query plan due to lock confilcts\"\n> \"HINT: You can try again in a moment.\"\n\nThanks, it seems better.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Wed, 09 Feb 2022 00:12:50 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"\n\nOn 2022/02/09 0:12, torikoshia wrote:\n> BTW, since the above example results in calling ExecutorRun() only once, the output didn't differ even after ActiveQueryDesc is reset to save_ActiveQueryDesc.\n> \n> The below definition of test() worked as expected.\n> \n> create or replace function test () returns int as $$\n> begin\n> perform 1;\n> perform 1/0;\n> exception when others then\n> return 30;\n> end;\n> $$ language plpgsql;\nSo in this case ActiveQueryDesc set by ExecutorRun() called only once is still valid even when AbortSubTransaction() is called. That is, that ActiveQueryDesc should NOT be reset to save_ActiveQueryDesc in this case, should it?\n\nOTOH, in your example, ActiveQueryDesc set by the second call to ExecutorRun() should be reset in AbortSubTransaction(). Then ActiveQueryDesc set by the first call should be valid.\n\nRegards,\n\n-- \nFujii Masao\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Tue, 22 Feb 2022 02:39:16 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2022-02-08 01:13, Fujii Masao wrote:\n> AbortSubTransaction() should reset ActiveQueryDesc to\n> save_ActiveQueryDesc that ExecutorRun() set, instead of NULL?\n> Otherwise ActiveQueryDesc of top-level statement will be unavailable\n> after subtransaction is aborted in the nested statements.\n\nI once agreed above suggestion and made v20 patch making \nsave_ActiveQueryDesc a global variable, but it caused segfault when \ncalling pg_log_query_plan() after FreeQueryDesc().\n\nOTOH, doing some kind of reset of ActiveQueryDesc seems necessary since \nit also caused segfault when running pg_log_query_plan() during \ninstallcheck.\n\nThere may be a better way, but resetting ActiveQueryDesc to NULL seems \nsafe and simple.\nOf course it makes pg_log_query_plan() useless after a subtransaction is \naborted.\nHowever, if it does not often happen that people want to know the \nrunning query's plan whose subtransaction is aborted, resetting \nActiveQueryDesc to NULL would be acceptable.\n\nAttached is a patch that sets ActiveQueryDesc to NULL when a \nsubtransaction is aborted.\n\nHow do you think?\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Wed, 09 Mar 2022 19:04:49 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2022-03-09 19:04, torikoshia wrote:\n> On 2022-02-08 01:13, Fujii Masao wrote:\n>> AbortSubTransaction() should reset ActiveQueryDesc to\n>> save_ActiveQueryDesc that ExecutorRun() set, instead of NULL?\n>> Otherwise ActiveQueryDesc of top-level statement will be unavailable\n>> after subtransaction is aborted in the nested statements.\n> \n> I once agreed above suggestion and made v20 patch making\n> save_ActiveQueryDesc a global variable, but it caused segfault when\n> calling pg_log_query_plan() after FreeQueryDesc().\n> \n> OTOH, doing some kind of reset of ActiveQueryDesc seems necessary\n> since it also caused segfault when running pg_log_query_plan() during\n> installcheck.\n> \n> There may be a better way, but resetting ActiveQueryDesc to NULL seems\n> safe and simple.\n> Of course it makes pg_log_query_plan() useless after a subtransaction\n> is aborted.\n> However, if it does not often happen that people want to know the\n> running query's plan whose subtransaction is aborted, resetting\n> ActiveQueryDesc to NULL would be acceptable.\n> \n> Attached is a patch that sets ActiveQueryDesc to NULL when a\n> subtransaction is aborted.\n> \n> How do you think?\n\nAttached new patch to fix patch apply failures.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Mon, 16 May 2022 17:02:40 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2022-05-16 17:02, torikoshia wrote:\n> On 2022-03-09 19:04, torikoshia wrote:\n>> On 2022-02-08 01:13, Fujii Masao wrote:\n>>> AbortSubTransaction() should reset ActiveQueryDesc to\n>>> save_ActiveQueryDesc that ExecutorRun() set, instead of NULL?\n>>> Otherwise ActiveQueryDesc of top-level statement will be unavailable\n>>> after subtransaction is aborted in the nested statements.\n>> \n>> I once agreed above suggestion and made v20 patch making\n>> save_ActiveQueryDesc a global variable, but it caused segfault when\n>> calling pg_log_query_plan() after FreeQueryDesc().\n>> \n>> OTOH, doing some kind of reset of ActiveQueryDesc seems necessary\n>> since it also caused segfault when running pg_log_query_plan() during\n>> installcheck.\n>> \n>> There may be a better way, but resetting ActiveQueryDesc to NULL seems\n>> safe and simple.\n>> Of course it makes pg_log_query_plan() useless after a subtransaction\n>> is aborted.\n>> However, if it does not often happen that people want to know the\n>> running query's plan whose subtransaction is aborted, resetting\n>> ActiveQueryDesc to NULL would be acceptable.\n>> \n>> Attached is a patch that sets ActiveQueryDesc to NULL when a\n>> subtransaction is aborted.\n>> \n>> How do you think?\nAttached new patch to fix patch apply failures again.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Thu, 08 Sep 2022 19:19:09 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi,I'm sorry,if this message is duplicated previous this one, but I'm not sure that the previous message is sent correctly. I sent it from email address a.rybakina@postgrespro.ru and I couldn't send this one email from those address.I like idea to create patch for logging query plan. After reviewing this code and notice some moments and I'd rather ask you some questions.Firstly, I suggest some editing in the comment of commit. I think, it is turned out the more laconic and the same clear. I wrote it below since I can't think of any other way to add it.```Currently, we have to wait for finishing of the query execution to check its plan.This is not so convenient in investigation long-running queries on productionenvironments where we cannot use debuggers.To improve this situation there is proposed the patch containing the pg_log_query_plan()function which request to log plan of the specified backend process.By default, only superusers are allowed to request log of the plan otherwiseallowing any users to issue this request could create cause lots of log messagesand it can lead to denial of service.At the next requesting CHECK_FOR_INTERRUPTS(), the target backend logs its plan atLOG_SERVER_ONLY level and therefore this plan will appear in the server log only,not to be sent to the client.```Secondly, I have question about deleting USE_ASSERT_CHECKING in lock.h?It supposed to have been checked in another placed of the code by matching values. I am worry about skipping errors due to untesting with assert option in the places where it (GetLockMethodLocalHash) participates and we won't able to get core file in segfault cases. I might not understand something, then can you please explain to me?Thirdly, I have incomprehension of the point why save_ActiveQueryDesc is declared in the pquery.h? I am seemed to save_ActiveQueryDesc to be used in an once time in the ExecutorRun function and its declaration superfluous. I added it in the attached patch.Fourthly, it seems to me there are not enough explanatory comments in the code. I also added them in the attached patch.Lastly, I have incomprehension about handling signals since have been unused it before. Could another signal disabled calling this signal to log query plan? I noticed this signal to be checked the latest in procsignal_sigusr1_handler function. -- Regards,Alena RybakinaPostgres Professional 19.09.2022, 11:01, \"torikoshia\" <torikoshia@oss.nttdata.com>:On 2022-05-16 17:02, torikoshia wrote: On 2022-03-09 19:04, torikoshia wrote: On 2022-02-08 01:13, Fujii Masao wrote: AbortSubTransaction() should reset ActiveQueryDesc to save_ActiveQueryDesc that ExecutorRun() set, instead of NULL? Otherwise ActiveQueryDesc of top-level statement will be unavailable after subtransaction is aborted in the nested statements. I once agreed above suggestion and made v20 patch making save_ActiveQueryDesc a global variable, but it caused segfault when calling pg_log_query_plan() after FreeQueryDesc(). OTOH, doing some kind of reset of ActiveQueryDesc seems necessary since it also caused segfault when running pg_log_query_plan() during installcheck. There may be a better way, but resetting ActiveQueryDesc to NULL seems safe and simple. Of course it makes pg_log_query_plan() useless after a subtransaction is aborted. However, if it does not often happen that people want to know the running query's plan whose subtransaction is aborted, resetting ActiveQueryDesc to NULL would be acceptable. Attached is a patch that sets ActiveQueryDesc to NULL when a subtransaction is aborted. How do you think?Attached new patch to fix patch apply failures again. --Regards,--Atsushi TorikoshiNTT DATA CORPORATION","msg_date":"Mon, 19 Sep 2022 11:47:21 +0300","msg_from":"=?utf-8?B?0JDQu9C10L3QsCDQoNGL0LHQsNC60LjQvdCw?=\n <lena.ribackina@yandex.ru>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2022-09-19 17:47, Алена Рыбакина wrote:\nThanks for your review and comments!\n\n> Hi,\n> \n> I'm sorry,if this message is duplicated previous this one, but I'm not\n> sure that the previous message is sent correctly. I sent it from email\n> address a.rybakina@postgrespro.ru and I couldn't send this one email\n> from those address.\n\nI've successfully received your mail from both a.rybakina@postgrespro.ru \nand lena.ribackina@yandex.ru.\n\n> I like idea to create patch for logging query plan. After reviewing\n> this code and notice some moments and I'd rather ask you some\n> questions.\n> \n> Firstly, I suggest some editing in the comment of commit. I think, it \n> is\n> turned out the more laconic and the same clear. I wrote it below since \n> I\n> can't think of any other way to add it.\n\n> ```\n> Currently, we have to wait for finishing of the query execution to \n> check\n> its plan.\n> This is not so convenient in investigation long-running queries on\n> production\n> environments where we cannot use debuggers.\n\n> To improve this situation there is proposed the patch containing the\n> pg_log_query_plan()\n> function which request to log plan of the specified backend process.\n\n> By default, only superusers are allowed to request log of the plan\n> otherwise\n> allowing any users to issue this request could create cause lots of log\n> messages\n> and it can lead to denial of service.\n> \n> At the next requesting CHECK_FOR_INTERRUPTS(), the target backend logs\n> its plan at\n> LOG_SERVER_ONLY level and therefore this plan will appear in the server\n> log only,\n> not to be sent to the client.\nThanks, I have incorporated your comments.\nSince the latter part of the original message comes from the commit \nmessage of pg_log_backend_memory_contexts(43620e328617c), so I left it \nas it was for consistency.\n\n\n> Secondly, I have question about deleting USE_ASSERT_CHECKING in lock.h?\n> It supposed to have been checked in another placed of the code by\n> matching values. I am worry about skipping errors due to untesting with\n> assert option in the places where it (GetLockMethodLocalHash)\n> participates and we won't able to get core file in segfault cases. I\n> might not understand something, then can you please explain to me?\nSince GetLockMethodLocalHash() is only used for assertions, this is only \ndefined when USE_ASSERT_CHECKING is enabled.\nThis patch uses GetLockMethodLocalHash() not only for the assertion \npurpose, so I removed \"ifdef USE_ASSERT_CHECKING\" for this function.\nI belive it does not lead to skip errors.\n\n> Thirdly, I have incomprehension of the point why save_ActiveQueryDesc \n> is\n> declared in the pquery.h? I am seemed to save_ActiveQueryDesc to be \n> used\n> in an once time in the ExecutorRun function and its declaration\n> superfluous. I added it in the attached patch.\nExactly.\n\n> Fourthly, it seems to me there are not enough explanatory comments in\n> the code. I also added them in the attached patch.\nThanks!\n\n| + /*\n| + * Save value of ActiveQueryDesc before having called\n| + * ExecutorRun_hook function due to having reset by\n| + * AbortSubTransaction.\n| + */\n| +\n| save_ActiveQueryDesc = ActiveQueryDesc;\n| ActiveQueryDesc = queryDesc;\n|\n| @@ -314,6 +320,7 @@ ExecutorRun(QueryDesc *queryDesc,\n| else\n| standard_ExecutorRun(queryDesc, direction, count, \nexecute_once);\n|\n| + /* We set the actual value of ActiveQueryDesc */\n| ActiveQueryDesc = save_ActiveQueryDesc;\n\nSince these processes are needed for nested queries, not only for\nAbortSubTransaction[1], added comments on it.\n\n| +/* Function to handle the signal to output the query plan. */\n| extern void HandleLogQueryPlanInterrupt(void);\n\nI feel this comment is unnecessary since the explanation of \nHandleLogQueryPlanInterrupt() is written in explain.c and no functions \nin explain.h have comments in it.\n\n> Lastly, I have incomprehension about handling signals since have been\n> unused it before. Could another signal disabled calling this signal to\n> log query plan? I noticed this signal to be checked the latest in\n> procsignal_sigusr1_handler function.\nAre you concerned that one signal will not be processed when multiple \nsignals are sent in succession?\nAFAIU both of them are processed since SendProcSignal flags \nps_signalFlags for each signal.\n\n```\n SendProcSignal(pid_t pid, ProcSignalReason reason, BackendId backendId)\n {\n volatile ProcSignalSlot *slot;\n...(snip)...\n278 if (slot->pss_pid == pid)\n279 {\n280 /* Atomically set the proper flag */\n281 slot->pss_signalFlags[reason] = true;\n282 /* Send signal */\n283 return kill(pid, SIGUSR1);\n```\n\nComments of ProcSignalReason also say 'We can cope with concurrent \nsignals for different reasons'.\n\n```C\n /*\n * Reasons for signaling a Postgres child process (a backend or an \nauxiliary\n * process, like checkpointer). We can cope with concurrent signals \nfor different\n * reasons. However, if the same reason is signaled multiple times in \nquick\n * succession, the process is likely to observe only one notification \nof it.\n * This is okay for the present uses.\n ...\n typedef enum\n {\n PROCSIG_CATCHUP_INTERRUPT, /* sinval catchup interrupt */\n PROCSIG_NOTIFY_INTERRUPT, /* listen/notify interrupt */\n PROCSIG_PARALLEL_MESSAGE, /* message from cooperating parallel \nbackend */\n PROCSIG_WALSND_INIT_STOPPING, /* ask walsenders to prepare for \nshutdown */\n PROCSIG_BARRIER, /* global barrier interrupt */\n PROCSIG_LOG_MEMORY_CONTEXT, /* ask backend to log the memory \ncontexts */\n PROCSIG_LOG_QUERY_PLAN, /* ask backend to log plan of the current \nquery */\n...\n } ProcSignalReason;\n```\n\n\n[1] \nhttps://www.postgresql.org/message-id/8b53b32f-26cc-0531-4ac0-27310e0bef4b%40oss.nttdata.com\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Tue, 20 Sep 2022 23:41:57 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Ok, I get it.\n\n> Since GetLockMethodLocalHash() is only used for assertions, this is \n> only defined when USE_ASSERT_CHECKING is enabled. This patch uses \n> GetLockMethodLocalHash() not only for the assertion purpose, so I \n> removed \"ifdef USE_ASSERT_CHECKING\" for this function. I belive it \n> does not lead to skip errors. \nAgree.\n> Since these processes are needed for nested queries, not only for \n> AbortSubTransaction[1], added comments on it.\n\nI also noticed it. However I also discovered that above function \ndeclarations to be aplied for explain command and used to be printed \ndetails of the executed query.\n\nWe have a similar task to print the plan of an interrupted process \nmaking a request for a specific pid.\n\nIn short, I think, this task is different and for separating these parts \nI added this comment.\n\n> I feel this comment is unnecessary since the explanation of \n> HandleLogQueryPlanInterrupt() is written in explain.c and no functions \n> in explain.h have comments in it. \n\nYes, I was worried about it. I understood it, thank for explaining.\n\n>\n> AFAIU both of them are processed since SendProcSignal flags \n> ps_signalFlags for each signal.\n>\n> ```\n> SendProcSignal(pid_t pid, ProcSignalReason reason, BackendId backendId)\n> {\n> volatile ProcSignalSlot *slot;\n> ...(snip)...\n> 278 if (slot->pss_pid == pid)\n> 279 {\n> 280 /* Atomically set the proper flag */\n> 281 slot->pss_signalFlags[reason] = true;\n> 282 /* Send signal */\n> 283 return kill(pid, SIGUSR1);\n> ```\n>\n> Comments of ProcSignalReason also say 'We can cope with concurrent \n> signals for different reasons'.\n>\n> ```C\n> /*\n> * Reasons for signaling a Postgres child process (a backend or an \n> auxiliary\n> * process, like checkpointer). We can cope with concurrent signals for \n> different\n> * reasons. However, if the same reason is signaled multiple times in \n> quick\n> * succession, the process is likely to observe only one notification \n> of it.\n> * This is okay for the present uses.\n> ...\n> typedef enum\n> {\n> PROCSIG_CATCHUP_INTERRUPT, /* sinval catchup interrupt */\n> PROCSIG_NOTIFY_INTERRUPT, /* listen/notify interrupt */\n> PROCSIG_PARALLEL_MESSAGE, /* message from cooperating parallel backend */\n> PROCSIG_WALSND_INIT_STOPPING, /* ask walsenders to prepare for \n> shutdown */\n> PROCSIG_BARRIER, /* global barrier interrupt */\n> PROCSIG_LOG_MEMORY_CONTEXT, /* ask backend to log the memory contexts */\n> PROCSIG_LOG_QUERY_PLAN, /* ask backend to log plan of the current \n> query */\n> ...\n> } ProcSignalReason;\n> ```\n>\n>\n> [1] \n> https://www.postgresql.org/message-id/8b53b32f-26cc-0531-4ac0-27310e0bef4b%40oss.nttdata.com\n\n\n\n\n\n Ok, I get it.\n\n Since GetLockMethodLocalHash() is only used for assertions, this\n is only defined when USE_ASSERT_CHECKING is enabled. This patch\n uses GetLockMethodLocalHash() not only for the assertion purpose,\n so I removed \"ifdef USE_ASSERT_CHECKING\" for this function. I\n belive it does not lead to skip errors. \nAgree.\n \n\n Since these processes are needed for nested queries, not only for\n AbortSubTransaction[1], added comments on it.\nI also noticed it. However I also discovered that\n above function declarations to be aplied for explain command and\n used to be printed details of the executed query.\nWe have a similar task to print the plan of an\n interrupted process making a request for a specific pid.\nIn short, I think, this task is different and for\n separating these parts I added this comment.\n\n I feel this comment is unnecessary since the explanation of\n HandleLogQueryPlanInterrupt() is written in explain.c and no\n functions\n in explain.h have comments in it.\n \n\nYes, I was worried about it. I understood it, thank\n for explaining.\n\n\n AFAIU both of them are processed since SendProcSignal flags\n ps_signalFlags for each signal.\n \n\n ```\n \n SendProcSignal(pid_t pid, ProcSignalReason reason, BackendId\n backendId)\n \n {\n \n volatile ProcSignalSlot *slot;\n \n ...(snip)...\n \n 278 if (slot->pss_pid == pid)\n \n 279 {\n \n 280 /* Atomically set the proper flag */\n \n 281 slot->pss_signalFlags[reason] = true;\n \n 282 /* Send signal */\n \n 283 return kill(pid, SIGUSR1);\n \n ```\n \n\n Comments of ProcSignalReason also say 'We can cope with concurrent\n signals for different reasons'.\n \n\n ```C\n \n /*\n \n * Reasons for signaling a Postgres child process (a backend or an\n auxiliary\n \n * process, like checkpointer). We can cope with concurrent signals\n for different\n \n * reasons. However, if the same reason is signaled multiple times\n in\n quick\n \n * succession, the process is likely to observe only one\n notification\n of it.\n \n * This is okay for the present uses.\n \n ...\n \n typedef enum\n \n {\n \n PROCSIG_CATCHUP_INTERRUPT, /* sinval catchup interrupt */\n \n PROCSIG_NOTIFY_INTERRUPT, /* listen/notify interrupt */\n \n PROCSIG_PARALLEL_MESSAGE, /* message from cooperating parallel\n backend */\n \n PROCSIG_WALSND_INIT_STOPPING, /* ask walsenders to prepare for\n shutdown */\n \n PROCSIG_BARRIER, /* global barrier interrupt */\n \n PROCSIG_LOG_MEMORY_CONTEXT, /* ask backend to log the memory\n contexts */\n \n PROCSIG_LOG_QUERY_PLAN, /* ask backend to log plan of the current\n query */\n \n ...\n \n } ProcSignalReason;\n \n ```\n \n\n\n [1]\nhttps://www.postgresql.org/message-id/8b53b32f-26cc-0531-4ac0-27310e0bef4b%40oss.nttdata.com","msg_date":"Wed, 21 Sep 2022 11:30:25 +0300","msg_from":"Alena Rybakina <lena.ribackina@yandex.ru>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2022-09-21 17:30, Alena Rybakina wrote:\n\nThanks for your reply!\n\n> I also noticed it. However I also discovered that above function\n> declarations to be aplied for explain command and used to be printed\n> details of the executed query.\n> \n> We have a similar task to print the plan of an interrupted process\n> making a request for a specific pid.\n> \n> In short, I think, this task is different and for separating these\n> parts I added this comment.\n\nI'm not sure I understand your comment correctly, do you mean \nHandleLogQueryPlanInterrupt() should not be placed in explain.c?\n\nIt may be so.\nHowever, given that ProcesLogMemoryContextInterrupt(), which similarly \nhandles interrupts for pg_log_backend_memory_contexts(), is located in \nmcxt.c, I also think current location might be acceptable.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION\n\n\n","msg_date":"Thu, 22 Sep 2022 23:12:16 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Sorry, I wrote confusingly at that time.\n\nNo, I suggested adding comment about the explanation of \nHandleLogQueryPlanInterrupt() only in the explain.h and not removing \nfrom the explain.c.\n\nI seemed to be necessary separating declaration function for 'explaining \nfeature' of executed query from our logging plan of the running query \ninterrupts function. But now, I doubt it.\n\n> I'm not sure I understand your comment correctly, do you mean \n> HandleLogQueryPlanInterrupt() should not be placed in explain.c? \n\n\nThank you for having reminded about this function and I looked at \nProcessLogMemoryContextInterrupt() declaration. I'm noticed comments in \nthe memutils.h are missed tooю\n\nDescription of this function is written only in mcxt.c.\n> However, given that ProcesLogMemoryContextInterrupt(), which similarly \n> handles interrupts for pg_log_backend_memory_contexts(), is located in \n> mcxt.c, I also think current location might be acceptable. \n\n\nSo I think you are right and the comment about the explanation of \nHandleLogQueryPlanInterrupt() written in explain.h is redundant.\n\n> I feel this comment is unnecessary since the explanation of \n> HandleLogQueryPlanInterrupt() is written in explain.c and no functions \n> in explain.h have comments in it. \n\nRegards,\n\n-- \nAlena Rybakina\nPostgres Professional\n\n\n\n","msg_date":"Fri, 23 Sep 2022 12:43:19 +0300","msg_from":"Alena Rybakina <lena.ribackina@yandex.ru>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi,\n\nThis patch does not currently build, due to a conflicting oid:\n\nhttps://cirrus-ci.com/task/4638460594618368?logs=build#L108\n[17:26:44.602] /usr/bin/perl ../src/include/catalog/../../backend/catalog/genbki.pl --include-path=../src/include --set-version=16 --output=src/include/catalog ../src/include/catalog/pg_proc.h ../src/include/catalog/pg_type.h ../src/include/catalog/pg_attribute.h ../src/include/catalog/pg_class.h ../src/include/catalog/pg_attrdef.h ../src/include/catalog/pg_constraint.h ../src/include/catalog/pg_inherits.h ../src/include/catalog/pg_index.h ../src/include/catalog/pg_operator.h ../src/include/catalog/pg_opfamily.h ../src/include/catalog/pg_opclass.h ../src/include/catalog/pg_am.h ../src/include/catalog/pg_amop.h ../src/include/catalog/pg_amproc.h ../src/include/catalog/pg_language.h ../src/include/catalog/pg_largeobject_metadata.h ../src/include/catalog/pg_largeobject.h ../src/include/catalog/pg_aggregate.h ../src/include/catalog/pg_statistic.h ../src/include/catalog/pg_statistic_ext.h ../src/include/catalog/pg_statistic_ext_data.h ../src/include/catalog/pg_rewrite.h ../src/include/catalog/pg_trigger.h ../src/include/catalog/pg_event_trigger.h ../src/include/catalog/pg_description.h ../src/include/catalog/pg_cast.h ../src/include/catalog/pg_enum.h ../src/include/catalog/pg_namespace.h ../src/include/catalog/pg_conversion.h ../src/include/catalog/pg_depend.h ../src/include/catalog/pg_database.h ../src/include/catalog/pg_db_role_setting.h ../src/include/catalog/pg_tablespace.h ../src/include/catalog/pg_authid.h ../src/include/catalog/pg_auth_members.h ../src/include/catalog/pg_shdepend.h ../src/include/catalog/pg_shdescription.h ../src/include/catalog/pg_ts_config.h ../src/include/catalog/pg_ts_config_map.h ../src/include/catalog/pg_ts_dict.h ../src/include/catalog/pg_ts_parser.h ../src/include/catalog/pg_ts_template.h ../src/include/catalog/pg_extension.h ../src/include/catalog/pg_foreign_data_wrapper.h ../src/include/catalog/pg_foreign_server.h ../src/include/catalog/pg_user_mapping.h ../src/include/catalog/pg_foreign_table.h ../src/include/catalog/pg_policy.h ../src/include/catalog/pg_replication_origin.h ../src/include/catalog/pg_default_acl.h ../src/include/catalog/pg_init_privs.h ../src/include/catalog/pg_seclabel.h ../src/include/catalog/pg_shseclabel.h ../src/include/catalog/pg_collation.h ../src/include/catalog/pg_parameter_acl.h ../src/include/catalog/pg_partitioned_table.h ../src/include/catalog/pg_range.h ../src/include/catalog/pg_transform.h ../src/include/catalog/pg_sequence.h ../src/include/catalog/pg_publication.h ../src/include/catalog/pg_publication_namespace.h ../src/include/catalog/pg_publication_rel.h ../src/include/catalog/pg_subscription.h ../src/include/catalog/pg_subscription_rel.h\n[17:26:44.602] Duplicate OIDs detected:\n[17:26:44.602] 4550\n[17:26:44.602] found 1 duplicate OID(s) in catalog data\n\nI suggest you choose a random oid out of the \"development purposes\" range:\n\n *\t\tOIDs 1-9999 are reserved for manual assignment (see .dat files in\n *\t\tsrc/include/catalog/). Of these, 8000-9999 are reserved for\n *\t\tdevelopment purposes (such as in-progress patches and forks);\n *\t\tthey should not appear in released versions.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 6 Dec 2022 10:41:01 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2022-12-07 03:41, Andres Freund wrote:\n> Hi,\n> \n> This patch does not currently build, due to a conflicting oid:\n> \n> I suggest you choose a random oid out of the \"development purposes\" \n> range:\n\nThanks for your advice!\nAttached updated patch.\n\n\nBTW, since this patch depends on ProcessInterrupts() and EXPLAIN codes \nwhich is used in auto_explain, I'm feeling that the following discussion \nalso applies to this patch.\n\n> -- \n> https://www.postgresql.org/message-id/CA%2BTgmoYW_rSOW4JMQ9_0Df9PKQ%3DsQDOKUGA4Gc9D8w4wui8fSA%40mail.gmail.com\n> \n> explaining a query is a pretty\n> complicated operation that involves catalog lookups and lots of\n> complicated stuff, and I don't think that it would be safe to do all\n> of that at any arbitrary point in the code where ProcessInterrupts()\n> happened to fire.\n\nIf I can't come up with some workaround during the next Commitfest, I'm \ngoing to withdraw this proposal.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Thu, 08 Dec 2022 14:10:32 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hello,\n\nThanks for working on this patch!\n\nOn Thu, Dec 8, 2022 at 12:10 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> BTW, since this patch depends on ProcessInterrupts() and EXPLAIN codes\n> which is used in auto_explain, I'm feeling that the following discussion\n> also applies to this patch.\n>\n> > --\n> > https://www.postgresql.org/message-id/CA%2BTgmoYW_rSOW4JMQ9_0Df9PKQ%3DsQDOKUGA4Gc9D8w4wui8fSA%40mail.gmail.com\n> >\n> > explaining a query is a pretty\n> > complicated operation that involves catalog lookups and lots of\n> > complicated stuff, and I don't think that it would be safe to do all\n> > of that at any arbitrary point in the code where ProcessInterrupts()\n> > happened to fire.\n>\n> If I can't come up with some workaround during the next Commitfest, I'm\n> going to withdraw this proposal.\n\nWhile at PGCon this week I'd brought up this idea with a few people\nwithout realizing you'd already worked on it previously, and then\nafter I discovered this thread several of us (Greg, Ronan, David,\nHeikki, and myself -- all cc'd) discussed the safety concerns over\ndinner last night.\n\nOur conclusion was that all of the concerns we could come up with (for\nexample, walking though the code for ExplainTargetRel() and discussing\nthe relevant catalog and syscache accesses) all applied specifically\nto Robert's concerns about running explain inside an aborted\ntransaction. After all, I'd originally started that thread to ask\nabout running auto-explain after a query timeout.\n\nTo put it positively: we believe that, for example, catalog accesses\ninside CHECK_FOR_INTERRUPTS() -- assuming that the CFI call is inside\nan existing valid transaction/query state, as it would be for this\npatch -- are safe. If there were problems, then those problems are\nlikely bugs we already have in other CFI cases.\n\nAnother concern Robert raised may apply here: what if a person tries\nto cancel a query when we're explaining? I believe that's a reasonable\nquestion to ask, but I believe it's a trade-off that's worth making\nfor the (significant) introspection benefits this patch would provide.\n\nOn to the patch itself: overall I think it looks like it's in pretty\ngood shape. I also noticed we don't have any tests (I assume it'd have\nto be TAP tests) of the actual output happening, and I think it would\nbe worth adding that.\n\nAre you interested in re-opening this patch? I'd be happy to provide\nfurther review and help to try to push this along.\n\nI've rebased the patch and attached as v26.\n\nThanks,\nJames Coleman","msg_date":"Fri, 2 Jun 2023 13:51:22 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-06-03 02:51, James Coleman wrote:\n> Hello,\n> \n> Thanks for working on this patch!\n> \n> On Thu, Dec 8, 2022 at 12:10 AM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> \n>> BTW, since this patch depends on ProcessInterrupts() and EXPLAIN codes\n>> which is used in auto_explain, I'm feeling that the following \n>> discussion\n>> also applies to this patch.\n>> \n>> > --\n>> > https://www.postgresql.org/message-id/CA%2BTgmoYW_rSOW4JMQ9_0Df9PKQ%3DsQDOKUGA4Gc9D8w4wui8fSA%40mail.gmail.com\n>> >\n>> > explaining a query is a pretty\n>> > complicated operation that involves catalog lookups and lots of\n>> > complicated stuff, and I don't think that it would be safe to do all\n>> > of that at any arbitrary point in the code where ProcessInterrupts()\n>> > happened to fire.\n>> \n>> If I can't come up with some workaround during the next Commitfest, \n>> I'm\n>> going to withdraw this proposal.\n> \n> While at PGCon this week I'd brought up this idea with a few people\n> without realizing you'd already worked on it previously, and then\n> after I discovered this thread several of us (Greg, Ronan, David,\n> Heikki, and myself -- all cc'd) discussed the safety concerns over\n> dinner last night.\n> \n> Our conclusion was that all of the concerns we could come up with (for\n> example, walking though the code for ExplainTargetRel() and discussing\n> the relevant catalog and syscache accesses) all applied specifically\n> to Robert's concerns about running explain inside an aborted\n> transaction. After all, I'd originally started that thread to ask\n> about running auto-explain after a query timeout.\n> \n> To put it positively: we believe that, for example, catalog accesses\n> inside CHECK_FOR_INTERRUPTS() -- assuming that the CFI call is inside\n> an existing valid transaction/query state, as it would be for this\n> patch -- are safe. If there were problems, then those problems are\n> likely bugs we already have in other CFI cases.\n\nThanks a lot for the discussion and sharing it!\nI really appreciate it.\n\nBTW I'm not sure whether all the CFI are called in valid transaction,\ndo you think we should check each of them?\n\n> Another concern Robert raised may apply here: what if a person tries\n> to cancel a query when we're explaining? I believe that's a reasonable\n> question to ask, but I believe it's a trade-off that's worth making\n> for the (significant) introspection benefits this patch would provide.\n\nIs the concern here limited to the case where explain code goes crazy\nas Robert pointed out?\nIf so, this may be a trade-off worth doing.\nI am a little concerned about whether there will be cases where the\nexplain code is not problematic but just takes long time.\n\n> On to the patch itself: overall I think it looks like it's in pretty\n> good shape. I also noticed we don't have any tests (I assume it'd have\n> to be TAP tests) of the actual output happening, and I think it would\n> be worth adding that.\n\nChecking the log output may be better, but I didn't add it since there\nis only a little content that can be checked, and similar function\npg_log_backend_memory_contexts() does not do such type of tests.\n\n> Are you interested in re-opening this patch? I'd be happy to provide\n> further review and help to try to push this along.\nSure, I'm going to re-open this.\n\n> I've rebased the patch and attached as v26.\nThanks again for your work!\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION\n\n\n","msg_date":"Mon, 05 Jun 2023 17:30:18 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Mon, Jun 5, 2023 at 4:30 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-06-03 02:51, James Coleman wrote:\n> > Hello,\n> >\n> > Thanks for working on this patch!\n\nSure thing! I'm *very interested* in seeing this available, and I\nthink it paves the way for some additional features later on...\n\n> > On Thu, Dec 8, 2022 at 12:10 AM torikoshia <torikoshia@oss.nttdata.com>\n> ...\n> > To put it positively: we believe that, for example, catalog accesses\n> > inside CHECK_FOR_INTERRUPTS() -- assuming that the CFI call is inside\n> > an existing valid transaction/query state, as it would be for this\n> > patch -- are safe. If there were problems, then those problems are\n> > likely bugs we already have in other CFI cases.\n>\n> Thanks a lot for the discussion and sharing it!\n> I really appreciate it.\n>\n> BTW I'm not sure whether all the CFI are called in valid transaction,\n> do you think we should check each of them?\n\nI kicked off the regressions tests with a call to\nProcessLogQueryPlanInterrupt() in every single CHECK_FOR_INTERRUPTS()\ncall. Several hours and 52 GB of logs later I have confirmed that\n(with the attached revision) at the very least the regression test\nsuite can't trigger any kind of failures regardless of when we trigger\nthis. The existing code in the patch for only running the explain when\nthere's an active query handling that.\n\n> > Another concern Robert raised may apply here: what if a person tries\n> > to cancel a query when we're explaining? I believe that's a reasonable\n> > question to ask, but I believe it's a trade-off that's worth making\n> > for the (significant) introspection benefits this patch would provide.\n>\n> Is the concern here limited to the case where explain code goes crazy\n> as Robert pointed out?\n> If so, this may be a trade-off worth doing.\n> I am a little concerned about whether there will be cases where the\n> explain code is not problematic but just takes long time.\n\nThe explain code could take a long time in some rare cases (e.g., we\ndiscovered a bug a few years back with the planning code that actually\ndescends an index to find the max value), but I think the trade-off is\nworth it.\n\n> > On to the patch itself: overall I think it looks like it's in pretty\n> > good shape. I also noticed we don't have any tests (I assume it'd have\n> > to be TAP tests) of the actual output happening, and I think it would\n> > be worth adding that.\n>\n> Checking the log output may be better, but I didn't add it since there\n> is only a little content that can be checked, and similar function\n> pg_log_backend_memory_contexts() does not do such type of tests.\n\nFair enough. I still think that would be an improvement here, but\nothers could also weigh in.\n\n> > Are you interested in re-opening this patch? I'd be happy to provide\n> > further review and help to try to push this along.\n> Sure, I'm going to re-open this.\n\nI've attached v27. The important change here in 0001 is that it\nguarantees the interrupt handler is re-entrant, since that was a bug\nexposed by my testing. I've also included 0002 which is only meant for\ntesting (it attempts to log in the plan in every\nCHECK_FOR_INTERRUPTS() call).\n\nRegards,\nJames","msg_date":"Mon, 5 Jun 2023 14:26:25 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-06-06 03:26, James Coleman wrote:\n> On Mon, Jun 5, 2023 at 4:30 AM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> \n>> On 2023-06-03 02:51, James Coleman wrote:\n>> > Hello,\n>> >\n>> > Thanks for working on this patch!\n> \n> Sure thing! I'm *very interested* in seeing this available, and I\n> think it paves the way for some additional features later on...\n> \n>> > On Thu, Dec 8, 2022 at 12:10 AM torikoshia <torikoshia@oss.nttdata.com>\n>> ...\n>> > To put it positively: we believe that, for example, catalog accesses\n>> > inside CHECK_FOR_INTERRUPTS() -- assuming that the CFI call is inside\n>> > an existing valid transaction/query state, as it would be for this\n>> > patch -- are safe. If there were problems, then those problems are\n>> > likely bugs we already have in other CFI cases.\n>> \n>> Thanks a lot for the discussion and sharing it!\n>> I really appreciate it.\n>> \n>> BTW I'm not sure whether all the CFI are called in valid transaction,\n>> do you think we should check each of them?\n> \n> I kicked off the regressions tests with a call to\n> ProcessLogQueryPlanInterrupt() in every single CHECK_FOR_INTERRUPTS()\n> call. Several hours and 52 GB of logs later I have confirmed that\n> (with the attached revision) at the very least the regression test\n> suite can't trigger any kind of failures regardless of when we trigger\n> this. The existing code in the patch for only running the explain when\n> there's an active query handling that.\n\nThanks for the testing!\n\n> I've attached v27. The important change here in 0001 is that it\n> guarantees the interrupt handler is re-entrant, since that was a bug\n> exposed by my testing. I've also included 0002 which is only meant for\n> testing (it attempts to log in the plan in every\n> CHECK_FOR_INTERRUPTS() call).\n\nWhen SIGINT is sent during ProcessLogQueryPlanInterrupt(),\nProcessLogQueryPlanInterruptActive can remain true.\nAfter that, pg_log_query_plan() does nothing but just returns.\n\nAFAIU, v26 logs plan for each pg_log_query_plan() even when another\npg_log_query_plan() is running, but it doesn't cause errors or critical\nproblem.\n\nConsidering the problem solved and introduced by v27, I think v26 might\nbe fine.\nHow do you think?\n\n> \n> Regards,\n> James\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION\n\n\n","msg_date":"Mon, 12 Jun 2023 13:07:34 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Sun, Jun 11, 2023 at 11:07 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-06-06 03:26, James Coleman wrote:\n> > On Mon, Jun 5, 2023 at 4:30 AM torikoshia <torikoshia@oss.nttdata.com>\n> > wrote:\n> >>\n> >> On 2023-06-03 02:51, James Coleman wrote:\n> >> > Hello,\n> >> >\n> >> > Thanks for working on this patch!\n> >\n> > Sure thing! I'm *very interested* in seeing this available, and I\n> > think it paves the way for some additional features later on...\n> >\n> >> > On Thu, Dec 8, 2022 at 12:10 AM torikoshia <torikoshia@oss.nttdata.com>\n> >> ...\n> >> > To put it positively: we believe that, for example, catalog accesses\n> >> > inside CHECK_FOR_INTERRUPTS() -- assuming that the CFI call is inside\n> >> > an existing valid transaction/query state, as it would be for this\n> >> > patch -- are safe. If there were problems, then those problems are\n> >> > likely bugs we already have in other CFI cases.\n> >>\n> >> Thanks a lot for the discussion and sharing it!\n> >> I really appreciate it.\n> >>\n> >> BTW I'm not sure whether all the CFI are called in valid transaction,\n> >> do you think we should check each of them?\n> >\n> > I kicked off the regressions tests with a call to\n> > ProcessLogQueryPlanInterrupt() in every single CHECK_FOR_INTERRUPTS()\n> > call. Several hours and 52 GB of logs later I have confirmed that\n> > (with the attached revision) at the very least the regression test\n> > suite can't trigger any kind of failures regardless of when we trigger\n> > this. The existing code in the patch for only running the explain when\n> > there's an active query handling that.\n>\n> Thanks for the testing!\n>\n> > I've attached v27. The important change here in 0001 is that it\n> > guarantees the interrupt handler is re-entrant, since that was a bug\n> > exposed by my testing. I've also included 0002 which is only meant for\n> > testing (it attempts to log in the plan in every\n> > CHECK_FOR_INTERRUPTS() call).\n>\n> When SIGINT is sent during ProcessLogQueryPlanInterrupt(),\n> ProcessLogQueryPlanInterruptActive can remain true.\n> After that, pg_log_query_plan() does nothing but just returns.\n>\n> AFAIU, v26 logs plan for each pg_log_query_plan() even when another\n> pg_log_query_plan() is running, but it doesn't cause errors or critical\n> problem.\n>\n> Considering the problem solved and introduced by v27, I think v26 might\n> be fine.\n> How do you think?\n\nThe testing I did with calling this during every CFI is what uncovered\nthe re-entrancy problem. IIRC (without running that test again) the\nproblem was a stack overflow. Now, to be sure this is a particularly\ndegenerate case because in real-world usage it'd be impossible in\npractice, I think, to trigger that many calls to this function (and by\nextension the interrupt handler).\n\nIf SIGINT is the only concern we could reset\nProcessLogQueryPlanInterruptActive in error handling code. I admit\nthat part of my thought process here is thinking ahead to an\nadditional patch I'd like to see on top of this, which is logging a\nquery plan before cleaning up when statement timeout occurs. The\nre-entrancy issue becomes more interesting then, I think, since we\nwould then have automated calling of the logging code. BTW: I'd\nthought that would make a nice follow-up patch for this, but if you'd\nprefer I could add it as another patch in the series here.\n\nWhat do you think about resetting the flag versus just not having it?\n\nRegards,\nJames Coleman\n\n\n","msg_date":"Mon, 12 Jun 2023 10:52:01 -0500","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-06-13 00:52, James Coleman wrote:\n>> \n>> > I've attached v27. The important change here in 0001 is that it\n>> > guarantees the interrupt handler is re-entrant, since that was a bug\n>> > exposed by my testing. I've also included 0002 which is only meant for\n>> > testing (it attempts to log in the plan in every\n>> > CHECK_FOR_INTERRUPTS() call).\n>> \n>> When SIGINT is sent during ProcessLogQueryPlanInterrupt(),\n>> ProcessLogQueryPlanInterruptActive can remain true.\n>> After that, pg_log_query_plan() does nothing but just returns.\n>> \n>> AFAIU, v26 logs plan for each pg_log_query_plan() even when another\n>> pg_log_query_plan() is running, but it doesn't cause errors or \n>> critical\n>> problem.\n>> \n>> Considering the problem solved and introduced by v27, I think v26 \n>> might\n>> be fine.\n>> How do you think?\n> \n> The testing I did with calling this during every CFI is what uncovered\n> the re-entrancy problem. IIRC (without running that test again) the\n> problem was a stack overflow. Now, to be sure this is a particularly\n> degenerate case because in real-world usage it'd be impossible in\n> practice, I think, to trigger that many calls to this function (and by\n> extension the interrupt handler).\n\nYeah.In addition, currently only superusers are allowed to execute\npg_log_query_plan(), I think we don't need to think about cases\nwhere users are malicious.\n\n> If SIGINT is the only concern we could reset\n> ProcessLogQueryPlanInterruptActive in error handling code. I admit\n> that part of my thought process here is thinking ahead to an\n> additional patch I'd like to see on top of this, which is logging a\n> query plan before cleaning up when statement timeout occurs.\n\nI remember this is what you wanted do.[1]\n\n> The\n> re-entrancy issue becomes more interesting then, I think, since we\n> would then have automated calling of the logging code. BTW: I'd\n> thought that would make a nice follow-up patch for this, but if you'd\n> prefer I could add it as another patch in the series here.\n> \n> What do you think about resetting the flag versus just not having it?\n\nIf I understand you correctly, adding the flag is not necessary for this\nproposal.\nTo keep the patch simple, I prefer not having it.\n\n\n[1] \nhttps://www.postgresql.org/message-id/flat/CA%2BTgmoYW_rSOW4JMQ9_0Df9PKQ%3DsQDOKUGA4Gc9D8w4wui8fSA%40mail.gmail.com#b57432077f8045be8588049269f7a8dd\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION\n\n\n","msg_date":"Wed, 14 Jun 2023 00:22:14 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Tue, Jun 13, 2023 at 11:22 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-06-13 00:52, James Coleman wrote:\n> >>\n> >> > I've attached v27. The important change here in 0001 is that it\n> >> > guarantees the interrupt handler is re-entrant, since that was a bug\n> >> > exposed by my testing. I've also included 0002 which is only meant for\n> >> > testing (it attempts to log in the plan in every\n> >> > CHECK_FOR_INTERRUPTS() call).\n> >>\n> >> When SIGINT is sent during ProcessLogQueryPlanInterrupt(),\n> >> ProcessLogQueryPlanInterruptActive can remain true.\n> >> After that, pg_log_query_plan() does nothing but just returns.\n> >>\n> >> AFAIU, v26 logs plan for each pg_log_query_plan() even when another\n> >> pg_log_query_plan() is running, but it doesn't cause errors or\n> >> critical\n> >> problem.\n> >>\n> >> Considering the problem solved and introduced by v27, I think v26\n> >> might\n> >> be fine.\n> >> How do you think?\n> >\n> > The testing I did with calling this during every CFI is what uncovered\n> > the re-entrancy problem. IIRC (without running that test again) the\n> > problem was a stack overflow. Now, to be sure this is a particularly\n> > degenerate case because in real-world usage it'd be impossible in\n> > practice, I think, to trigger that many calls to this function (and by\n> > extension the interrupt handler).\n>\n> Yeah.In addition, currently only superusers are allowed to execute\n> pg_log_query_plan(), I think we don't need to think about cases\n> where users are malicious.\n>\n> > If SIGINT is the only concern we could reset\n> > ProcessLogQueryPlanInterruptActive in error handling code. I admit\n> > that part of my thought process here is thinking ahead to an\n> > additional patch I'd like to see on top of this, which is logging a\n> > query plan before cleaning up when statement timeout occurs.\n>\n> I remember this is what you wanted do.[1]\n>\n> > The\n> > re-entrancy issue becomes more interesting then, I think, since we\n> > would then have automated calling of the logging code. BTW: I'd\n> > thought that would make a nice follow-up patch for this, but if you'd\n> > prefer I could add it as another patch in the series here.\n> >\n> > What do you think about resetting the flag versus just not having it?\n>\n> If I understand you correctly, adding the flag is not necessary for this\n> proposal.\n> To keep the patch simple, I prefer not having it.\n>\n\nI'm going to re-run tests with my patch version + resetting the flag\non SIGINT (and any other error condition) to be certain that the issue\nyou uncovered (where backends get stuck after a SIGINT not responding\nto the requested plan logging) wasn't masking any other issues.\n\nAs long as that run is clean also then I believe the patch is safe\nas-is even without the re-entrancy guard.\n\nI'll report back with the results of that testing.\n\nRegards,\nJames Coleman\n\n\n","msg_date":"Tue, 13 Jun 2023 11:53:39 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Tue, Jun 13, 2023 at 11:53 AM James Coleman <jtc331@gmail.com> wrote:\n>\n> ...\n> I'm going to re-run tests with my patch version + resetting the flag\n> on SIGINT (and any other error condition) to be certain that the issue\n> you uncovered (where backends get stuck after a SIGINT not responding\n> to the requested plan logging) wasn't masking any other issues.\n>\n> As long as that run is clean also then I believe the patch is safe\n> as-is even without the re-entrancy guard.\n>\n> I'll report back with the results of that testing.\n\nThe tests have been running since last night, but have been apparently\nhung now for many hours. I haven't been able to fully look into it,\nbut so far I know the hung (100% CPU) backend last logged this:\n\n2023-06-14 02:00:30.045 UTC client backend[84461]\npg_regress/updatable_views LOG: query plan running on backend with\nPID 84461 is:\n Query Text: SELECT table_name, column_name, is_updatable\n FROM information_schema.columns\n WHERE table_name LIKE E'r_\\\\_view%'\n ORDER BY table_name, ordinal_position;\n\nThe last output from the regression test harness was:\n\n# parallel group (5 tests): index_including create_view\nindex_including_gist create_index create_index_spgist\nok 66 + create_index 36508 ms\nok 67 + create_index_spgist 38588 ms\nok 68 + create_view 1394 ms\nok 69 + index_including 654 ms\nok 70 + index_including_gist 1701 ms\n# parallel group (16 tests): errors create_cast drop_if_exists\ncreate_aggregate roleattributes constraints hash_func typed_table\ninfinite_recurse\n\nAttaching gdb to the hung backend shows this:\n\n#0 0x00005601ab1f9529 in ProcLockWakeup\n(lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n<default_lockmethod>, lock=lock@entry=0x7f5325c913f0) at proc.c:1655\n#1 0x00005601ab1e99dc in CleanUpLock (lock=lock@entry=0x7f5325c913f0,\nproclock=proclock@entry=0x7f5325d40d60,\nlockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n<default_lockmethod>,\n hashcode=hashcode@entry=573498161, wakeupNeeded=<optimized out>)\nat lock.c:1673\n#2 0x00005601ab1e9e21 in LockRefindAndRelease\n(lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n<default_lockmethod>, proc=<optimized out>,\nlocktag=locktag@entry=0x5601ac3d7998, lockmode=lockmode@entry=1,\n decrement_strong_lock_count=decrement_strong_lock_count@entry=false)\nat lock.c:3150\n#3 0x00005601ab1edb27 in LockReleaseAll\n(lockmethodid=lockmethodid@entry=1, allLocks=false) at lock.c:2295\n#4 0x00005601ab1f8599 in ProcReleaseLocks\n(isCommit=isCommit@entry=true) at proc.c:781\n#5 0x00005601ab37f1f4 in ResourceOwnerReleaseInternal\n(owner=<optimized out>, phase=phase@entry=RESOURCE_RELEASE_LOCKS,\nisCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\nresowner.c:618\n#6 0x00005601ab37f7b7 in ResourceOwnerRelease (owner=<optimized out>,\nphase=phase@entry=RESOURCE_RELEASE_LOCKS,\nisCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\nresowner.c:494\n#7 0x00005601aaec1d84 in CommitTransaction () at xact.c:2334\n#8 0x00005601aaec2b22 in CommitTransactionCommand () at xact.c:3067\n#9 0x00005601ab200a66 in finish_xact_command () at postgres.c:2783\n#10 0x00005601ab20338f in exec_simple_query (\n query_string=query_string@entry=0x5601ac3b0858 \"SELECT table_name,\ncolumn_name, is_updatable\\n FROM information_schema.columns\\n WHERE\ntable_name LIKE E'r_\\\\\\\\_view%'\\n ORDER BY table_name,\nordinal_position;\") at postgres.c:1300\n\nI am unable to connect to the regression test Postgres instance --\npsql just hangs, so the lock seems to have affected the postmaster\nalso.\n\nI'm wondering if this might represent a bug in the current patch.\n\nRegards,\nJames Coleman\n\n\n","msg_date":"Wed, 14 Jun 2023 12:48:52 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-06-15 01:48, James Coleman wrote:\n> On Tue, Jun 13, 2023 at 11:53 AM James Coleman <jtc331@gmail.com> \n> wrote:\n>> \n>> ...\n>> I'm going to re-run tests with my patch version + resetting the flag\n>> on SIGINT (and any other error condition) to be certain that the issue\n>> you uncovered (where backends get stuck after a SIGINT not responding\n>> to the requested plan logging) wasn't masking any other issues.\n>> \n>> As long as that run is clean also then I believe the patch is safe\n>> as-is even without the re-entrancy guard.\n>> \n>> I'll report back with the results of that testing.\n> \n> The tests have been running since last night, but have been apparently\n> hung now for many hours. I haven't been able to fully look into it,\n> but so far I know the hung (100% CPU) backend last logged this:\n> \n> 2023-06-14 02:00:30.045 UTC client backend[84461]\n> pg_regress/updatable_views LOG: query plan running on backend with\n> PID 84461 is:\n> Query Text: SELECT table_name, column_name, is_updatable\n> FROM information_schema.columns\n> WHERE table_name LIKE E'r_\\\\_view%'\n> ORDER BY table_name, ordinal_position;\n> \n> The last output from the regression test harness was:\n> \n> # parallel group (5 tests): index_including create_view\n> index_including_gist create_index create_index_spgist\n> ok 66 + create_index 36508 ms\n> ok 67 + create_index_spgist 38588 ms\n> ok 68 + create_view 1394 ms\n> ok 69 + index_including 654 ms\n> ok 70 + index_including_gist 1701 ms\n> # parallel group (16 tests): errors create_cast drop_if_exists\n> create_aggregate roleattributes constraints hash_func typed_table\n> infinite_recurse\n> \n> Attaching gdb to the hung backend shows this:\n> \n> #0 0x00005601ab1f9529 in ProcLockWakeup\n> (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n> <default_lockmethod>, lock=lock@entry=0x7f5325c913f0) at proc.c:1655\n> #1 0x00005601ab1e99dc in CleanUpLock (lock=lock@entry=0x7f5325c913f0,\n> proclock=proclock@entry=0x7f5325d40d60,\n> lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n> <default_lockmethod>,\n> hashcode=hashcode@entry=573498161, wakeupNeeded=<optimized out>)\n> at lock.c:1673\n> #2 0x00005601ab1e9e21 in LockRefindAndRelease\n> (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n> <default_lockmethod>, proc=<optimized out>,\n> locktag=locktag@entry=0x5601ac3d7998, lockmode=lockmode@entry=1,\n> \n> decrement_strong_lock_count=decrement_strong_lock_count@entry=false)\n> at lock.c:3150\n> #3 0x00005601ab1edb27 in LockReleaseAll\n> (lockmethodid=lockmethodid@entry=1, allLocks=false) at lock.c:2295\n> #4 0x00005601ab1f8599 in ProcReleaseLocks\n> (isCommit=isCommit@entry=true) at proc.c:781\n> #5 0x00005601ab37f1f4 in ResourceOwnerReleaseInternal\n> (owner=<optimized out>, phase=phase@entry=RESOURCE_RELEASE_LOCKS,\n> isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\n> resowner.c:618\n> #6 0x00005601ab37f7b7 in ResourceOwnerRelease (owner=<optimized out>,\n> phase=phase@entry=RESOURCE_RELEASE_LOCKS,\n> isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\n> resowner.c:494\n> #7 0x00005601aaec1d84 in CommitTransaction () at xact.c:2334\n> #8 0x00005601aaec2b22 in CommitTransactionCommand () at xact.c:3067\n> #9 0x00005601ab200a66 in finish_xact_command () at postgres.c:2783\n> #10 0x00005601ab20338f in exec_simple_query (\n> query_string=query_string@entry=0x5601ac3b0858 \"SELECT table_name,\n> column_name, is_updatable\\n FROM information_schema.columns\\n WHERE\n> table_name LIKE E'r_\\\\\\\\_view%'\\n ORDER BY table_name,\n> ordinal_position;\") at postgres.c:1300\n> \n> I am unable to connect to the regression test Postgres instance --\n> psql just hangs, so the lock seems to have affected the postmaster\n> also.\n> \n> I'm wondering if this might represent a bug in the current patch.\n\nThanks for running and analyzing the test!\n\nCould you share me how you are running the test?\n\nI imagined something like below, but currently couldn't reproduce it.\n- apply both v26-0001 and v27-0002 and build\n- run PostgreSQL with default GUCssaaa\n- make installcheck-world\n- run 'SELECT pg_log_query_plan(pid) FROM pg_stat_activity \\watch 0.1' \nduring make installcheck-world\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION\n\n\n","msg_date":"Thu, 15 Jun 2023 22:00:46 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Thu, Jun 15, 2023 at 9:00 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-06-15 01:48, James Coleman wrote:\n> > On Tue, Jun 13, 2023 at 11:53 AM James Coleman <jtc331@gmail.com>\n> > wrote:\n> >>\n> >> ...\n> >> I'm going to re-run tests with my patch version + resetting the flag\n> >> on SIGINT (and any other error condition) to be certain that the issue\n> >> you uncovered (where backends get stuck after a SIGINT not responding\n> >> to the requested plan logging) wasn't masking any other issues.\n> >>\n> >> As long as that run is clean also then I believe the patch is safe\n> >> as-is even without the re-entrancy guard.\n> >>\n> >> I'll report back with the results of that testing.\n> >\n> > The tests have been running since last night, but have been apparently\n> > hung now for many hours. I haven't been able to fully look into it,\n> > but so far I know the hung (100% CPU) backend last logged this:\n> >\n> > 2023-06-14 02:00:30.045 UTC client backend[84461]\n> > pg_regress/updatable_views LOG: query plan running on backend with\n> > PID 84461 is:\n> > Query Text: SELECT table_name, column_name, is_updatable\n> > FROM information_schema.columns\n> > WHERE table_name LIKE E'r_\\\\_view%'\n> > ORDER BY table_name, ordinal_position;\n> >\n> > The last output from the regression test harness was:\n> >\n> > # parallel group (5 tests): index_including create_view\n> > index_including_gist create_index create_index_spgist\n> > ok 66 + create_index 36508 ms\n> > ok 67 + create_index_spgist 38588 ms\n> > ok 68 + create_view 1394 ms\n> > ok 69 + index_including 654 ms\n> > ok 70 + index_including_gist 1701 ms\n> > # parallel group (16 tests): errors create_cast drop_if_exists\n> > create_aggregate roleattributes constraints hash_func typed_table\n> > infinite_recurse\n> >\n> > Attaching gdb to the hung backend shows this:\n> >\n> > #0 0x00005601ab1f9529 in ProcLockWakeup\n> > (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n> > <default_lockmethod>, lock=lock@entry=0x7f5325c913f0) at proc.c:1655\n> > #1 0x00005601ab1e99dc in CleanUpLock (lock=lock@entry=0x7f5325c913f0,\n> > proclock=proclock@entry=0x7f5325d40d60,\n> > lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n> > <default_lockmethod>,\n> > hashcode=hashcode@entry=573498161, wakeupNeeded=<optimized out>)\n> > at lock.c:1673\n> > #2 0x00005601ab1e9e21 in LockRefindAndRelease\n> > (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n> > <default_lockmethod>, proc=<optimized out>,\n> > locktag=locktag@entry=0x5601ac3d7998, lockmode=lockmode@entry=1,\n> >\n> > decrement_strong_lock_count=decrement_strong_lock_count@entry=false)\n> > at lock.c:3150\n> > #3 0x00005601ab1edb27 in LockReleaseAll\n> > (lockmethodid=lockmethodid@entry=1, allLocks=false) at lock.c:2295\n> > #4 0x00005601ab1f8599 in ProcReleaseLocks\n> > (isCommit=isCommit@entry=true) at proc.c:781\n> > #5 0x00005601ab37f1f4 in ResourceOwnerReleaseInternal\n> > (owner=<optimized out>, phase=phase@entry=RESOURCE_RELEASE_LOCKS,\n> > isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\n> > resowner.c:618\n> > #6 0x00005601ab37f7b7 in ResourceOwnerRelease (owner=<optimized out>,\n> > phase=phase@entry=RESOURCE_RELEASE_LOCKS,\n> > isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\n> > resowner.c:494\n> > #7 0x00005601aaec1d84 in CommitTransaction () at xact.c:2334\n> > #8 0x00005601aaec2b22 in CommitTransactionCommand () at xact.c:3067\n> > #9 0x00005601ab200a66 in finish_xact_command () at postgres.c:2783\n> > #10 0x00005601ab20338f in exec_simple_query (\n> > query_string=query_string@entry=0x5601ac3b0858 \"SELECT table_name,\n> > column_name, is_updatable\\n FROM information_schema.columns\\n WHERE\n> > table_name LIKE E'r_\\\\\\\\_view%'\\n ORDER BY table_name,\n> > ordinal_position;\") at postgres.c:1300\n> >\n> > I am unable to connect to the regression test Postgres instance --\n> > psql just hangs, so the lock seems to have affected the postmaster\n> > also.\n> >\n> > I'm wondering if this might represent a bug in the current patch.\n>\n> Thanks for running and analyzing the test!\n\nSure thing!\n\n> Could you share me how you are running the test?\n>\n> I imagined something like below, but currently couldn't reproduce it.\n> - apply both v26-0001 and v27-0002 and build\n> - run PostgreSQL with default GUCssaaa\n> - make installcheck-world\n> - run 'SELECT pg_log_query_plan(pid) FROM pg_stat_activity \\watch 0.1'\n> during make installcheck-world\n\nApologies, I should have attached my updated patch (with the fix for\nthe bug you'd reporting with the re-entrancy guard). Attached is v28\nwhich sets ProcessLogQueryPlanInterruptActive to false in errfinish\nwhen necessary. Once built with those two patches I'm simply running\n`make check`.\n\nRegards,\nJames Coleman","msg_date":"Thu, 15 Jun 2023 12:34:23 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-06-16 01:34, James Coleman wrote:\n> On Thu, Jun 15, 2023 at 9:00 AM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> \n>> On 2023-06-15 01:48, James Coleman wrote:\n>> > On Tue, Jun 13, 2023 at 11:53 AM James Coleman <jtc331@gmail.com>\n>> > wrote:\n>> >>\n>> >> ...\n>> >> I'm going to re-run tests with my patch version + resetting the flag\n>> >> on SIGINT (and any other error condition) to be certain that the issue\n>> >> you uncovered (where backends get stuck after a SIGINT not responding\n>> >> to the requested plan logging) wasn't masking any other issues.\n>> >>\n>> >> As long as that run is clean also then I believe the patch is safe\n>> >> as-is even without the re-entrancy guard.\n>> >>\n>> >> I'll report back with the results of that testing.\n>> >\n>> > The tests have been running since last night, but have been apparently\n>> > hung now for many hours. I haven't been able to fully look into it,\n>> > but so far I know the hung (100% CPU) backend last logged this:\n>> >\n>> > 2023-06-14 02:00:30.045 UTC client backend[84461]\n>> > pg_regress/updatable_views LOG: query plan running on backend with\n>> > PID 84461 is:\n>> > Query Text: SELECT table_name, column_name, is_updatable\n>> > FROM information_schema.columns\n>> > WHERE table_name LIKE E'r_\\\\_view%'\n>> > ORDER BY table_name, ordinal_position;\n>> >\n>> > The last output from the regression test harness was:\n>> >\n>> > # parallel group (5 tests): index_including create_view\n>> > index_including_gist create_index create_index_spgist\n>> > ok 66 + create_index 36508 ms\n>> > ok 67 + create_index_spgist 38588 ms\n>> > ok 68 + create_view 1394 ms\n>> > ok 69 + index_including 654 ms\n>> > ok 70 + index_including_gist 1701 ms\n>> > # parallel group (16 tests): errors create_cast drop_if_exists\n>> > create_aggregate roleattributes constraints hash_func typed_table\n>> > infinite_recurse\n>> >\n>> > Attaching gdb to the hung backend shows this:\n>> >\n>> > #0 0x00005601ab1f9529 in ProcLockWakeup\n>> > (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n>> > <default_lockmethod>, lock=lock@entry=0x7f5325c913f0) at proc.c:1655\n>> > #1 0x00005601ab1e99dc in CleanUpLock (lock=lock@entry=0x7f5325c913f0,\n>> > proclock=proclock@entry=0x7f5325d40d60,\n>> > lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n>> > <default_lockmethod>,\n>> > hashcode=hashcode@entry=573498161, wakeupNeeded=<optimized out>)\n>> > at lock.c:1673\n>> > #2 0x00005601ab1e9e21 in LockRefindAndRelease\n>> > (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\n>> > <default_lockmethod>, proc=<optimized out>,\n>> > locktag=locktag@entry=0x5601ac3d7998, lockmode=lockmode@entry=1,\n>> >\n>> > decrement_strong_lock_count=decrement_strong_lock_count@entry=false)\n>> > at lock.c:3150\n>> > #3 0x00005601ab1edb27 in LockReleaseAll\n>> > (lockmethodid=lockmethodid@entry=1, allLocks=false) at lock.c:2295\n>> > #4 0x00005601ab1f8599 in ProcReleaseLocks\n>> > (isCommit=isCommit@entry=true) at proc.c:781\n>> > #5 0x00005601ab37f1f4 in ResourceOwnerReleaseInternal\n>> > (owner=<optimized out>, phase=phase@entry=RESOURCE_RELEASE_LOCKS,\n>> > isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\n>> > resowner.c:618\n>> > #6 0x00005601ab37f7b7 in ResourceOwnerRelease (owner=<optimized out>,\n>> > phase=phase@entry=RESOURCE_RELEASE_LOCKS,\n>> > isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\n>> > resowner.c:494\n>> > #7 0x00005601aaec1d84 in CommitTransaction () at xact.c:2334\n>> > #8 0x00005601aaec2b22 in CommitTransactionCommand () at xact.c:3067\n>> > #9 0x00005601ab200a66 in finish_xact_command () at postgres.c:2783\n>> > #10 0x00005601ab20338f in exec_simple_query (\n>> > query_string=query_string@entry=0x5601ac3b0858 \"SELECT table_name,\n>> > column_name, is_updatable\\n FROM information_schema.columns\\n WHERE\n>> > table_name LIKE E'r_\\\\\\\\_view%'\\n ORDER BY table_name,\n>> > ordinal_position;\") at postgres.c:1300\n>> >\n>> > I am unable to connect to the regression test Postgres instance --\n>> > psql just hangs, so the lock seems to have affected the postmaster\n>> > also.\n>> >\n>> > I'm wondering if this might represent a bug in the current patch.\n>> \n>> Thanks for running and analyzing the test!\n> \n> Sure thing!\n> \n>> Could you share me how you are running the test?\n>> \n>> I imagined something like below, but currently couldn't reproduce it.\n>> - apply both v26-0001 and v27-0002 and build\n>> - run PostgreSQL with default GUCssaaa\n>> - make installcheck-world\n>> - run 'SELECT pg_log_query_plan(pid) FROM pg_stat_activity \\watch 0.1'\n>> during make installcheck-world\n> \n> Apologies, I should have attached my updated patch (with the fix for\n> the bug you'd reporting with the re-entrancy guard). Attached is v28\n> which sets ProcessLogQueryPlanInterruptActive to false in errfinish\n> when necessary. Once built with those two patches I'm simply running\n> `make check`.\n\nThanks!\n\nHowever, I haven't succeeded in reproducing the problem as below:\n\n> The tests have been running since last night, but have been apparently\n> hung now for many hours. I haven't been able to fully look into it,\n> but so far I know the hung (100% CPU) backend last logged this:\n\nDid you do something like this?\n\n $ ./configure --prefix=/home/ubuntu/pgsql/master --enable-cassert\n $ git apply \n../patch/v28-0001-Add-function-to-log-the-plan-of-the-query.patch \n../patch/v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch\n $ make\n $ make check\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION\n\n\n","msg_date":"Tue, 27 Jun 2023 19:03:21 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-06-16 01:34, James Coleman wrote:\n> Attached is v28\n> which sets ProcessLogQueryPlanInterruptActive to false in errfinish\n> when necessary. Once built with those two patches I'm simply running\n> `make check`.\n\nWith v28-0001 and v28-0002 patch, I confirmed backend processes consume \nhuge\namount of memory and under some environments they were terminated by OOM\nkiller.\n\nThis was because memory was allocated from existing memory contexts and \nthey\nwere not freed after ProcessLogQueryPlanInterrupt().\nUpdated the patch to use dedicated memory context for\nProcessLogQueryPlanInterrupt().\n\nApplying attached patch and v28-0002 patch, `make check` successfully\ncompleted after 20min and 50GB of logs on my environment.\n\n>>> On 2023-06-15 01:48, James Coleman wrote:\n>>> > The tests have been running since last night, but have been apparently\n>>> > hung now for many hours.\n\nI don't know if this has anything to do with the hung you faced, but I \nthought\nit might be possible that the large amount of memory usage resulted in\nswapping, which caused a significant delay in processing.\n\nIf possible, I would be very grateful if you could try to reproduce this \nwith\nthe v29 patch.\n\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA CORPORATION","msg_date":"Thu, 17 Aug 2023 23:02:22 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Thu, Aug 17, 2023 at 10:02 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-06-16 01:34, James Coleman wrote:\n> > Attached is v28\n> > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\n> > when necessary. Once built with those two patches I'm simply running\n> > `make check`.\n>\n> With v28-0001 and v28-0002 patch, I confirmed backend processes consume\n> huge\n> amount of memory and under some environments they were terminated by OOM\n> killer.\n>\n> This was because memory was allocated from existing memory contexts and\n> they\n> were not freed after ProcessLogQueryPlanInterrupt().\n> Updated the patch to use dedicated memory context for\n> ProcessLogQueryPlanInterrupt().\n>\n> Applying attached patch and v28-0002 patch, `make check` successfully\n> completed after 20min and 50GB of logs on my environment.\n>\n> >>> On 2023-06-15 01:48, James Coleman wrote:\n> >>> > The tests have been running since last night, but have been apparently\n> >>> > hung now for many hours.\n>\n> I don't know if this has anything to do with the hung you faced, but I\n> thought\n> it might be possible that the large amount of memory usage resulted in\n> swapping, which caused a significant delay in processing.\n\nAh, yes, I think that could be a possible explanation. I was delaying\non this thread because I wasn't comfortable with having caused an\nissue once (even if I couldn't easily reproduce) without at least some\ntheory as to the cause (and a fix).\n\n> If possible, I would be very grateful if you could try to reproduce this\n> with\n> the v29 patch.\n\nI'll kick off some testing.\n\nThanks,\nJames Coleman\n\n\n","msg_date":"Fri, 25 Aug 2023 07:43:26 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Fri, Aug 25, 2023 at 7:43 AM James Coleman <jtc331@gmail.com> wrote:\n>\n> On Thu, Aug 17, 2023 at 10:02 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n> >\n> > On 2023-06-16 01:34, James Coleman wrote:\n> > > Attached is v28\n> > > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\n> > > when necessary. Once built with those two patches I'm simply running\n> > > `make check`.\n> >\n> > With v28-0001 and v28-0002 patch, I confirmed backend processes consume\n> > huge\n> > amount of memory and under some environments they were terminated by OOM\n> > killer.\n> >\n> > This was because memory was allocated from existing memory contexts and\n> > they\n> > were not freed after ProcessLogQueryPlanInterrupt().\n> > Updated the patch to use dedicated memory context for\n> > ProcessLogQueryPlanInterrupt().\n> >\n> > Applying attached patch and v28-0002 patch, `make check` successfully\n> > completed after 20min and 50GB of logs on my environment.\n> >\n> > >>> On 2023-06-15 01:48, James Coleman wrote:\n> > >>> > The tests have been running since last night, but have been apparently\n> > >>> > hung now for many hours.\n> >\n> > I don't know if this has anything to do with the hung you faced, but I\n> > thought\n> > it might be possible that the large amount of memory usage resulted in\n> > swapping, which caused a significant delay in processing.\n>\n> Ah, yes, I think that could be a possible explanation. I was delaying\n> on this thread because I wasn't comfortable with having caused an\n> issue once (even if I couldn't easily reproduce) without at least some\n> theory as to the cause (and a fix).\n>\n> > If possible, I would be very grateful if you could try to reproduce this\n> > with\n> > the v29 patch.\n>\n> I'll kick off some testing.\n>\n\nI don't have time to investigate what's happening here, but 24 hours\nlater the first \"make check\" is still running, and at first glance it\nseems to have the same behavior I'd seen that first time. The test\noutput is to this point:\n\n# parallel group (5 tests): index_including create_view\nindex_including_gist create_index create_index_spgist\nok 66 + create_index 26365 ms\nok 67 + create_index_spgist 27675 ms\nok 68 + create_view 1235 ms\nok 69 + index_including 1102 ms\nok 70 + index_including_gist 1633 ms\n# parallel group (16 tests): create_aggregate create_cast errors\nroleattributes drop_if_exists hash_func typed_table create_am\ninfinite_recurse\n\nand it hasn't progressed past that point since at least ~16 hours ago\n(the first several hours of the run I wasn't monitoring it).\n\nI haven't connected up gdb yet, and won't be able to until maybe\ntomorrow, but here's the ps output for postgres processes that are\nrunning:\n\nadmin 3213249 0.0 0.0 196824 20552 ? Ss Aug25 0:00\n/home/admin/postgresql-test/bin/postgres -D\n/home/admin/postgresql-test-data\nadmin 3213250 0.0 0.0 196964 7284 ? Ss Aug25 0:00\npostgres: checkpointer\nadmin 3213251 0.0 0.0 196956 4276 ? Ss Aug25 0:00\npostgres: background writer\nadmin 3213253 0.0 0.0 196956 8600 ? Ss Aug25 0:00\npostgres: walwriter\nadmin 3213254 0.0 0.0 198424 7316 ? Ss Aug25 0:00\npostgres: autovacuum launcher\nadmin 3213255 0.0 0.0 198412 5488 ? Ss Aug25 0:00\npostgres: logical replication launcher\nadmin 3237967 0.0 0.0 2484 516 pts/4 S+ Aug25 0:00\n/bin/sh -c echo \"# +++ regress check in src/test/regress +++\" &&\nPATH=\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/bin:/home/admin/postgres/src/test/regress:$PATH\"\nLD_LIBRARY_PATH=\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/lib\"\nINITDB_TEMPLATE='/home/admin/postgres'/tmp_install/initdb-template\n../../../src/test/regress/pg_regress --temp-instance=./tmp_check\n--inputdir=. --bindir= --dlpath=. --max-concurrent-tests=20\n--schedule=./parallel_schedule\nadmin 3237973 0.0 0.0 197880 20688 pts/4 S+ Aug25 0:00\npostgres -D /home/admin/postgres/src/test/regress/tmp_check/data -F -c\nlisten_addresses= -k /tmp/pg_regress-7mmGUa\nadmin 3237976 0.0 0.1 198332 44608 ? Ss Aug25 0:00\npostgres: checkpointer\nadmin 3237977 0.0 0.0 198032 4640 ? Ss Aug25 0:00\npostgres: background writer\nadmin 3237979 0.0 0.0 197880 8580 ? Ss Aug25 0:00\npostgres: walwriter\nadmin 3237980 0.0 0.0 199484 7608 ? Ss Aug25 0:00\npostgres: autovacuum launcher\nadmin 3237981 0.0 0.0 199460 5488 ? Ss Aug25 0:00\npostgres: logical replication launcher\nadmin 3243644 0.0 0.2 252400 74656 ? Ss Aug25 0:01\npostgres: admin regression [local] SELECT waiting\nadmin 3243645 0.0 0.1 205480 33992 ? Ss Aug25 0:00\npostgres: admin regression [local] SELECT waiting\nadmin 3243654 99.9 0.0 203156 31504 ? Rs Aug25 1437:49\npostgres: admin regression [local] VACUUM\nadmin 3243655 0.0 0.1 212036 38504 ? Ss Aug25 0:00\npostgres: admin regression [local] SELECT waiting\nadmin 3243656 0.0 0.0 206024 30892 ? Ss Aug25 0:00\npostgres: admin regression [local] DELETE waiting\nadmin 3243657 0.0 0.1 205568 32232 ? Ss Aug25 0:00\npostgres: admin regression [local] ALTER TABLE waiting\nadmin 3243658 0.0 0.0 203740 21532 ? Ss Aug25 0:00\npostgres: admin regression [local] ANALYZE waiting\nadmin 3243798 0.0 0.0 199884 8464 ? Ss Aug25 0:00\npostgres: autovacuum worker\nadmin 3244733 0.0 0.0 199492 5956 ? Ss Aug25 0:00\npostgres: autovacuum worker\nadmin 3245652 0.0 0.0 199884 8468 ? Ss Aug25 0:00\npostgres: autovacuum worker\n\nAs you can see there are a bunch of backends presumably waiting, and\nalso the VACUUM process has been pegging a single CPU core for at\nleast since that ~16 hour ago mark.\n\nI hope to be able to do more investigation later, but I wanted to at\nleast give you this information now.\n\nJames Coleman\n\n\n","msg_date":"Sat, 26 Aug 2023 08:03:26 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-08-26 21:03, James Coleman wrote:\n> On Fri, Aug 25, 2023 at 7:43 AM James Coleman <jtc331@gmail.com> wrote:\n>> \n>> On Thu, Aug 17, 2023 at 10:02 AM torikoshia \n>> <torikoshia@oss.nttdata.com> wrote:\n>> >\n>> > On 2023-06-16 01:34, James Coleman wrote:\n>> > > Attached is v28\n>> > > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\n>> > > when necessary. Once built with those two patches I'm simply running\n>> > > `make check`.\n>> >\n>> > With v28-0001 and v28-0002 patch, I confirmed backend processes consume\n>> > huge\n>> > amount of memory and under some environments they were terminated by OOM\n>> > killer.\n>> >\n>> > This was because memory was allocated from existing memory contexts and\n>> > they\n>> > were not freed after ProcessLogQueryPlanInterrupt().\n>> > Updated the patch to use dedicated memory context for\n>> > ProcessLogQueryPlanInterrupt().\n>> >\n>> > Applying attached patch and v28-0002 patch, `make check` successfully\n>> > completed after 20min and 50GB of logs on my environment.\n>> >\n>> > >>> On 2023-06-15 01:48, James Coleman wrote:\n>> > >>> > The tests have been running since last night, but have been apparently\n>> > >>> > hung now for many hours.\n>> >\n>> > I don't know if this has anything to do with the hung you faced, but I\n>> > thought\n>> > it might be possible that the large amount of memory usage resulted in\n>> > swapping, which caused a significant delay in processing.\n>> \n>> Ah, yes, I think that could be a possible explanation. I was delaying\n>> on this thread because I wasn't comfortable with having caused an\n>> issue once (even if I couldn't easily reproduce) without at least some\n>> theory as to the cause (and a fix).\n>> \n>> > If possible, I would be very grateful if you could try to reproduce this\n>> > with\n>> > the v29 patch.\n>> \n>> I'll kick off some testing.\n>> \n> \n> I don't have time to investigate what's happening here, but 24 hours\n> later the first \"make check\" is still running, and at first glance it\n> seems to have the same behavior I'd seen that first time. The test\n> output is to this point:\n> \n> # parallel group (5 tests): index_including create_view\n> index_including_gist create_index create_index_spgist\n> ok 66 + create_index 26365 ms\n> ok 67 + create_index_spgist 27675 ms\n> ok 68 + create_view 1235 ms\n> ok 69 + index_including 1102 ms\n> ok 70 + index_including_gist 1633 ms\n> # parallel group (16 tests): create_aggregate create_cast errors\n> roleattributes drop_if_exists hash_func typed_table create_am\n> infinite_recurse\n> \n> and it hasn't progressed past that point since at least ~16 hours ago\n> (the first several hours of the run I wasn't monitoring it).\n> \n> I haven't connected up gdb yet, and won't be able to until maybe\n> tomorrow, but here's the ps output for postgres processes that are\n> running:\n> \n> admin 3213249 0.0 0.0 196824 20552 ? Ss Aug25 0:00\n> /home/admin/postgresql-test/bin/postgres -D\n> /home/admin/postgresql-test-data\n> admin 3213250 0.0 0.0 196964 7284 ? Ss Aug25 0:00\n> postgres: checkpointer\n> admin 3213251 0.0 0.0 196956 4276 ? Ss Aug25 0:00\n> postgres: background writer\n> admin 3213253 0.0 0.0 196956 8600 ? Ss Aug25 0:00\n> postgres: walwriter\n> admin 3213254 0.0 0.0 198424 7316 ? Ss Aug25 0:00\n> postgres: autovacuum launcher\n> admin 3213255 0.0 0.0 198412 5488 ? Ss Aug25 0:00\n> postgres: logical replication launcher\n> admin 3237967 0.0 0.0 2484 516 pts/4 S+ Aug25 0:00\n> /bin/sh -c echo \"# +++ regress check in src/test/regress +++\" &&\n> PATH=\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/bin:/home/admin/postgres/src/test/regress:$PATH\"\n> LD_LIBRARY_PATH=\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/lib\"\n> INITDB_TEMPLATE='/home/admin/postgres'/tmp_install/initdb-template\n> ../../../src/test/regress/pg_regress --temp-instance=./tmp_check\n> --inputdir=. --bindir= --dlpath=. --max-concurrent-tests=20\n> --schedule=./parallel_schedule\n> admin 3237973 0.0 0.0 197880 20688 pts/4 S+ Aug25 0:00\n> postgres -D /home/admin/postgres/src/test/regress/tmp_check/data -F -c\n> listen_addresses= -k /tmp/pg_regress-7mmGUa\n> admin 3237976 0.0 0.1 198332 44608 ? Ss Aug25 0:00\n> postgres: checkpointer\n> admin 3237977 0.0 0.0 198032 4640 ? Ss Aug25 0:00\n> postgres: background writer\n> admin 3237979 0.0 0.0 197880 8580 ? Ss Aug25 0:00\n> postgres: walwriter\n> admin 3237980 0.0 0.0 199484 7608 ? Ss Aug25 0:00\n> postgres: autovacuum launcher\n> admin 3237981 0.0 0.0 199460 5488 ? Ss Aug25 0:00\n> postgres: logical replication launcher\n> admin 3243644 0.0 0.2 252400 74656 ? Ss Aug25 0:01\n> postgres: admin regression [local] SELECT waiting\n> admin 3243645 0.0 0.1 205480 33992 ? Ss Aug25 0:00\n> postgres: admin regression [local] SELECT waiting\n> admin 3243654 99.9 0.0 203156 31504 ? Rs Aug25 1437:49\n> postgres: admin regression [local] VACUUM\n> admin 3243655 0.0 0.1 212036 38504 ? Ss Aug25 0:00\n> postgres: admin regression [local] SELECT waiting\n> admin 3243656 0.0 0.0 206024 30892 ? Ss Aug25 0:00\n> postgres: admin regression [local] DELETE waiting\n> admin 3243657 0.0 0.1 205568 32232 ? Ss Aug25 0:00\n> postgres: admin regression [local] ALTER TABLE waiting\n> admin 3243658 0.0 0.0 203740 21532 ? Ss Aug25 0:00\n> postgres: admin regression [local] ANALYZE waiting\n> admin 3243798 0.0 0.0 199884 8464 ? Ss Aug25 0:00\n> postgres: autovacuum worker\n> admin 3244733 0.0 0.0 199492 5956 ? Ss Aug25 0:00\n> postgres: autovacuum worker\n> admin 3245652 0.0 0.0 199884 8468 ? Ss Aug25 0:00\n> postgres: autovacuum worker\n> \n> As you can see there are a bunch of backends presumably waiting, and\n> also the VACUUM process has been pegging a single CPU core for at\n> least since that ~16 hour ago mark.\n> \n> I hope to be able to do more investigation later, but I wanted to at\n> least give you this information now.\n\nThanks a lot for testing the patch!\nI really appreciate your cooperation.\n\nHmm, I also tested on the current HEAD(165d581f146b09) again on Ubuntu \n22.04 and macOS, but unfortunately(fortunately?) they succeeded as \nbelow:\n\n```\n$ git apply v29-0001-Add-function-to-log-the-plan-of-the-query.patch\n$ git apply v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch\n$ ./configure --enable-debug --enable-cassert\n$ make\n$ make check\n\n...(snip)...\n\n# parallel group (5 tests): index_including index_including_gist \ncreate_view create_index create_index_spgist\nok 66 + create_index 25033 ms\nok 67 + create_index_spgist 26144 ms\nok 68 + create_view 3061 ms\nok 69 + index_including 976 ms\nok 70 + index_including_gist 2998 ms\n# parallel group (16 tests): create_cast errors create_aggregate \nroleattributes drop_if_exists hash_func typed_table\ncreate_am select constraints updatable_views inherit triggers vacuum \ncreate_function_sql infinite_recurse\nok 71 + create_aggregate 225 ms\nok 72 + create_function_sql 18874 ms\nok 73 + create_cast 168 ms\n\n...(snip)...\n\n# All 215 tests passed.\n```\n\nIf you notice any difference, I would be grateful if you could let me \nknow.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Mon, 28 Aug 2023 16:01:53 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Mon, Aug 28, 2023 at 3:01 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-08-26 21:03, James Coleman wrote:\n> > On Fri, Aug 25, 2023 at 7:43 AM James Coleman <jtc331@gmail.com> wrote:\n> >>\n> >> On Thu, Aug 17, 2023 at 10:02 AM torikoshia\n> >> <torikoshia@oss.nttdata.com> wrote:\n> >> >\n> >> > On 2023-06-16 01:34, James Coleman wrote:\n> >> > > Attached is v28\n> >> > > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\n> >> > > when necessary. Once built with those two patches I'm simply running\n> >> > > `make check`.\n> >> >\n> >> > With v28-0001 and v28-0002 patch, I confirmed backend processes consume\n> >> > huge\n> >> > amount of memory and under some environments they were terminated by OOM\n> >> > killer.\n> >> >\n> >> > This was because memory was allocated from existing memory contexts and\n> >> > they\n> >> > were not freed after ProcessLogQueryPlanInterrupt().\n> >> > Updated the patch to use dedicated memory context for\n> >> > ProcessLogQueryPlanInterrupt().\n> >> >\n> >> > Applying attached patch and v28-0002 patch, `make check` successfully\n> >> > completed after 20min and 50GB of logs on my environment.\n> >> >\n> >> > >>> On 2023-06-15 01:48, James Coleman wrote:\n> >> > >>> > The tests have been running since last night, but have been apparently\n> >> > >>> > hung now for many hours.\n> >> >\n> >> > I don't know if this has anything to do with the hung you faced, but I\n> >> > thought\n> >> > it might be possible that the large amount of memory usage resulted in\n> >> > swapping, which caused a significant delay in processing.\n> >>\n> >> Ah, yes, I think that could be a possible explanation. I was delaying\n> >> on this thread because I wasn't comfortable with having caused an\n> >> issue once (even if I couldn't easily reproduce) without at least some\n> >> theory as to the cause (and a fix).\n> >>\n> >> > If possible, I would be very grateful if you could try to reproduce this\n> >> > with\n> >> > the v29 patch.\n> >>\n> >> I'll kick off some testing.\n> >>\n> >\n> > I don't have time to investigate what's happening here, but 24 hours\n> > later the first \"make check\" is still running, and at first glance it\n> > seems to have the same behavior I'd seen that first time. The test\n> > output is to this point:\n> >\n> > # parallel group (5 tests): index_including create_view\n> > index_including_gist create_index create_index_spgist\n> > ok 66 + create_index 26365 ms\n> > ok 67 + create_index_spgist 27675 ms\n> > ok 68 + create_view 1235 ms\n> > ok 69 + index_including 1102 ms\n> > ok 70 + index_including_gist 1633 ms\n> > # parallel group (16 tests): create_aggregate create_cast errors\n> > roleattributes drop_if_exists hash_func typed_table create_am\n> > infinite_recurse\n> >\n> > and it hasn't progressed past that point since at least ~16 hours ago\n> > (the first several hours of the run I wasn't monitoring it).\n> >\n> > I haven't connected up gdb yet, and won't be able to until maybe\n> > tomorrow, but here's the ps output for postgres processes that are\n> > running:\n> >\n> > admin 3213249 0.0 0.0 196824 20552 ? Ss Aug25 0:00\n> > /home/admin/postgresql-test/bin/postgres -D\n> > /home/admin/postgresql-test-data\n> > admin 3213250 0.0 0.0 196964 7284 ? Ss Aug25 0:00\n> > postgres: checkpointer\n> > admin 3213251 0.0 0.0 196956 4276 ? Ss Aug25 0:00\n> > postgres: background writer\n> > admin 3213253 0.0 0.0 196956 8600 ? Ss Aug25 0:00\n> > postgres: walwriter\n> > admin 3213254 0.0 0.0 198424 7316 ? Ss Aug25 0:00\n> > postgres: autovacuum launcher\n> > admin 3213255 0.0 0.0 198412 5488 ? Ss Aug25 0:00\n> > postgres: logical replication launcher\n> > admin 3237967 0.0 0.0 2484 516 pts/4 S+ Aug25 0:00\n> > /bin/sh -c echo \"# +++ regress check in src/test/regress +++\" &&\n> > PATH=\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/bin:/home/admin/postgres/src/test/regress:$PATH\"\n> > LD_LIBRARY_PATH=\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/lib\"\n> > INITDB_TEMPLATE='/home/admin/postgres'/tmp_install/initdb-template\n> > ../../../src/test/regress/pg_regress --temp-instance=./tmp_check\n> > --inputdir=. --bindir= --dlpath=. --max-concurrent-tests=20\n> > --schedule=./parallel_schedule\n> > admin 3237973 0.0 0.0 197880 20688 pts/4 S+ Aug25 0:00\n> > postgres -D /home/admin/postgres/src/test/regress/tmp_check/data -F -c\n> > listen_addresses= -k /tmp/pg_regress-7mmGUa\n> > admin 3237976 0.0 0.1 198332 44608 ? Ss Aug25 0:00\n> > postgres: checkpointer\n> > admin 3237977 0.0 0.0 198032 4640 ? Ss Aug25 0:00\n> > postgres: background writer\n> > admin 3237979 0.0 0.0 197880 8580 ? Ss Aug25 0:00\n> > postgres: walwriter\n> > admin 3237980 0.0 0.0 199484 7608 ? Ss Aug25 0:00\n> > postgres: autovacuum launcher\n> > admin 3237981 0.0 0.0 199460 5488 ? Ss Aug25 0:00\n> > postgres: logical replication launcher\n> > admin 3243644 0.0 0.2 252400 74656 ? Ss Aug25 0:01\n> > postgres: admin regression [local] SELECT waiting\n> > admin 3243645 0.0 0.1 205480 33992 ? Ss Aug25 0:00\n> > postgres: admin regression [local] SELECT waiting\n> > admin 3243654 99.9 0.0 203156 31504 ? Rs Aug25 1437:49\n> > postgres: admin regression [local] VACUUM\n> > admin 3243655 0.0 0.1 212036 38504 ? Ss Aug25 0:00\n> > postgres: admin regression [local] SELECT waiting\n> > admin 3243656 0.0 0.0 206024 30892 ? Ss Aug25 0:00\n> > postgres: admin regression [local] DELETE waiting\n> > admin 3243657 0.0 0.1 205568 32232 ? Ss Aug25 0:00\n> > postgres: admin regression [local] ALTER TABLE waiting\n> > admin 3243658 0.0 0.0 203740 21532 ? Ss Aug25 0:00\n> > postgres: admin regression [local] ANALYZE waiting\n> > admin 3243798 0.0 0.0 199884 8464 ? Ss Aug25 0:00\n> > postgres: autovacuum worker\n> > admin 3244733 0.0 0.0 199492 5956 ? Ss Aug25 0:00\n> > postgres: autovacuum worker\n> > admin 3245652 0.0 0.0 199884 8468 ? Ss Aug25 0:00\n> > postgres: autovacuum worker\n> >\n> > As you can see there are a bunch of backends presumably waiting, and\n> > also the VACUUM process has been pegging a single CPU core for at\n> > least since that ~16 hour ago mark.\n> >\n> > I hope to be able to do more investigation later, but I wanted to at\n> > least give you this information now.\n>\n> Thanks a lot for testing the patch!\n> I really appreciate your cooperation.\n>\n> Hmm, I also tested on the current HEAD(165d581f146b09) again on Ubuntu\n> 22.04 and macOS, but unfortunately(fortunately?) they succeeded as\n> below:\n>\n> ```\n> $ git apply v29-0001-Add-function-to-log-the-plan-of-the-query.patch\n> $ git apply v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch\n> $ ./configure --enable-debug --enable-cassert\n> $ make\n> $ make check\n>\n> ...(snip)...\n>\n> # parallel group (5 tests): index_including index_including_gist\n> create_view create_index create_index_spgist\n> ok 66 + create_index 25033 ms\n> ok 67 + create_index_spgist 26144 ms\n> ok 68 + create_view 3061 ms\n> ok 69 + index_including 976 ms\n> ok 70 + index_including_gist 2998 ms\n> # parallel group (16 tests): create_cast errors create_aggregate\n> roleattributes drop_if_exists hash_func typed_table\n> create_am select constraints updatable_views inherit triggers vacuum\n> create_function_sql infinite_recurse\n> ok 71 + create_aggregate 225 ms\n> ok 72 + create_function_sql 18874 ms\n> ok 73 + create_cast 168 ms\n>\n> ...(snip)...\n>\n> # All 215 tests passed.\n> ```\n>\n> If you notice any difference, I would be grateful if you could let me\n> know.\n\nI've never been able to reproduce it (haven't tested the new version,\nbut v28 at least) on my M1 Mac; where I've reproduced it is on Debian\n(first buster and now bullseye).\n\nI'm attaching several stacktraces in the hope that they provide some\nclues. These all match the ps output I sent earlier, though note in\nthat output there is both the regress instance and my test instance\n(pid 3213249) running (different ports, of course, and they are from\nthe exact same compilation run). I've attached ps output for the\npostgres processes under the make check process to simplify cross\nreferencing.\n\nA few interesting things:\n- There's definitely a lock on a relation that seems to be what's\nblocking the processes.\n- When I try to connect with psql the process forks but then hangs\n(see the ps output with task names stuck in \"authentication\"). I've\nalso included a trace from one of these.\n\nIf you think a core file for any of these processes would be helpful\nfor debugging I'd be happy to try to figure out a way to get that to\nyou.\n\nRegards,\nJames Coleman","msg_date":"Mon, 28 Aug 2023 09:47:19 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-08-28 22:47, James Coleman wrote:\n> On Mon, Aug 28, 2023 at 3:01 AM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> \n>> On 2023-08-26 21:03, James Coleman wrote:\n>> > On Fri, Aug 25, 2023 at 7:43 AM James Coleman <jtc331@gmail.com> wrote:\n>> >>\n>> >> On Thu, Aug 17, 2023 at 10:02 AM torikoshia\n>> >> <torikoshia@oss.nttdata.com> wrote:\n>> >> >\n>> >> > On 2023-06-16 01:34, James Coleman wrote:\n>> >> > > Attached is v28\n>> >> > > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\n>> >> > > when necessary. Once built with those two patches I'm simply running\n>> >> > > `make check`.\n>> >> >\n>> >> > With v28-0001 and v28-0002 patch, I confirmed backend processes consume\n>> >> > huge\n>> >> > amount of memory and under some environments they were terminated by OOM\n>> >> > killer.\n>> >> >\n>> >> > This was because memory was allocated from existing memory contexts and\n>> >> > they\n>> >> > were not freed after ProcessLogQueryPlanInterrupt().\n>> >> > Updated the patch to use dedicated memory context for\n>> >> > ProcessLogQueryPlanInterrupt().\n>> >> >\n>> >> > Applying attached patch and v28-0002 patch, `make check` successfully\n>> >> > completed after 20min and 50GB of logs on my environment.\n>> >> >\n>> >> > >>> On 2023-06-15 01:48, James Coleman wrote:\n>> >> > >>> > The tests have been running since last night, but have been apparently\n>> >> > >>> > hung now for many hours.\n>> >> >\n>> >> > I don't know if this has anything to do with the hung you faced, but I\n>> >> > thought\n>> >> > it might be possible that the large amount of memory usage resulted in\n>> >> > swapping, which caused a significant delay in processing.\n>> >>\n>> >> Ah, yes, I think that could be a possible explanation. I was delaying\n>> >> on this thread because I wasn't comfortable with having caused an\n>> >> issue once (even if I couldn't easily reproduce) without at least some\n>> >> theory as to the cause (and a fix).\n>> >>\n>> >> > If possible, I would be very grateful if you could try to reproduce this\n>> >> > with\n>> >> > the v29 patch.\n>> >>\n>> >> I'll kick off some testing.\n>> >>\n>> >\n>> > I don't have time to investigate what's happening here, but 24 hours\n>> > later the first \"make check\" is still running, and at first glance it\n>> > seems to have the same behavior I'd seen that first time. The test\n>> > output is to this point:\n>> >\n>> > # parallel group (5 tests): index_including create_view\n>> > index_including_gist create_index create_index_spgist\n>> > ok 66 + create_index 26365 ms\n>> > ok 67 + create_index_spgist 27675 ms\n>> > ok 68 + create_view 1235 ms\n>> > ok 69 + index_including 1102 ms\n>> > ok 70 + index_including_gist 1633 ms\n>> > # parallel group (16 tests): create_aggregate create_cast errors\n>> > roleattributes drop_if_exists hash_func typed_table create_am\n>> > infinite_recurse\n>> >\n>> > and it hasn't progressed past that point since at least ~16 hours ago\n>> > (the first several hours of the run I wasn't monitoring it).\n>> >\n>> > I haven't connected up gdb yet, and won't be able to until maybe\n>> > tomorrow, but here's the ps output for postgres processes that are\n>> > running:\n>> >\n>> > admin 3213249 0.0 0.0 196824 20552 ? Ss Aug25 0:00\n>> > /home/admin/postgresql-test/bin/postgres -D\n>> > /home/admin/postgresql-test-data\n>> > admin 3213250 0.0 0.0 196964 7284 ? Ss Aug25 0:00\n>> > postgres: checkpointer\n>> > admin 3213251 0.0 0.0 196956 4276 ? Ss Aug25 0:00\n>> > postgres: background writer\n>> > admin 3213253 0.0 0.0 196956 8600 ? Ss Aug25 0:00\n>> > postgres: walwriter\n>> > admin 3213254 0.0 0.0 198424 7316 ? Ss Aug25 0:00\n>> > postgres: autovacuum launcher\n>> > admin 3213255 0.0 0.0 198412 5488 ? Ss Aug25 0:00\n>> > postgres: logical replication launcher\n>> > admin 3237967 0.0 0.0 2484 516 pts/4 S+ Aug25 0:00\n>> > /bin/sh -c echo \"# +++ regress check in src/test/regress +++\" &&\n>> > PATH=\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/bin:/home/admin/postgres/src/test/regress:$PATH\"\n>> > LD_LIBRARY_PATH=\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/lib\"\n>> > INITDB_TEMPLATE='/home/admin/postgres'/tmp_install/initdb-template\n>> > ../../../src/test/regress/pg_regress --temp-instance=./tmp_check\n>> > --inputdir=. --bindir= --dlpath=. --max-concurrent-tests=20\n>> > --schedule=./parallel_schedule\n>> > admin 3237973 0.0 0.0 197880 20688 pts/4 S+ Aug25 0:00\n>> > postgres -D /home/admin/postgres/src/test/regress/tmp_check/data -F -c\n>> > listen_addresses= -k /tmp/pg_regress-7mmGUa\n>> > admin 3237976 0.0 0.1 198332 44608 ? Ss Aug25 0:00\n>> > postgres: checkpointer\n>> > admin 3237977 0.0 0.0 198032 4640 ? Ss Aug25 0:00\n>> > postgres: background writer\n>> > admin 3237979 0.0 0.0 197880 8580 ? Ss Aug25 0:00\n>> > postgres: walwriter\n>> > admin 3237980 0.0 0.0 199484 7608 ? Ss Aug25 0:00\n>> > postgres: autovacuum launcher\n>> > admin 3237981 0.0 0.0 199460 5488 ? Ss Aug25 0:00\n>> > postgres: logical replication launcher\n>> > admin 3243644 0.0 0.2 252400 74656 ? Ss Aug25 0:01\n>> > postgres: admin regression [local] SELECT waiting\n>> > admin 3243645 0.0 0.1 205480 33992 ? Ss Aug25 0:00\n>> > postgres: admin regression [local] SELECT waiting\n>> > admin 3243654 99.9 0.0 203156 31504 ? Rs Aug25 1437:49\n>> > postgres: admin regression [local] VACUUM\n>> > admin 3243655 0.0 0.1 212036 38504 ? Ss Aug25 0:00\n>> > postgres: admin regression [local] SELECT waiting\n>> > admin 3243656 0.0 0.0 206024 30892 ? Ss Aug25 0:00\n>> > postgres: admin regression [local] DELETE waiting\n>> > admin 3243657 0.0 0.1 205568 32232 ? Ss Aug25 0:00\n>> > postgres: admin regression [local] ALTER TABLE waiting\n>> > admin 3243658 0.0 0.0 203740 21532 ? Ss Aug25 0:00\n>> > postgres: admin regression [local] ANALYZE waiting\n>> > admin 3243798 0.0 0.0 199884 8464 ? Ss Aug25 0:00\n>> > postgres: autovacuum worker\n>> > admin 3244733 0.0 0.0 199492 5956 ? Ss Aug25 0:00\n>> > postgres: autovacuum worker\n>> > admin 3245652 0.0 0.0 199884 8468 ? Ss Aug25 0:00\n>> > postgres: autovacuum worker\n>> >\n>> > As you can see there are a bunch of backends presumably waiting, and\n>> > also the VACUUM process has been pegging a single CPU core for at\n>> > least since that ~16 hour ago mark.\n>> >\n>> > I hope to be able to do more investigation later, but I wanted to at\n>> > least give you this information now.\n>> \n>> Thanks a lot for testing the patch!\n>> I really appreciate your cooperation.\n>> \n>> Hmm, I also tested on the current HEAD(165d581f146b09) again on Ubuntu\n>> 22.04 and macOS, but unfortunately(fortunately?) they succeeded as\n>> below:\n>> \n>> ```\n>> $ git apply v29-0001-Add-function-to-log-the-plan-of-the-query.patch\n>> $ git apply \n>> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch\n>> $ ./configure --enable-debug --enable-cassert\n>> $ make\n>> $ make check\n>> \n>> ...(snip)...\n>> \n>> # parallel group (5 tests): index_including index_including_gist\n>> create_view create_index create_index_spgist\n>> ok 66 + create_index 25033 ms\n>> ok 67 + create_index_spgist 26144 ms\n>> ok 68 + create_view 3061 ms\n>> ok 69 + index_including 976 ms\n>> ok 70 + index_including_gist 2998 ms\n>> # parallel group (16 tests): create_cast errors create_aggregate\n>> roleattributes drop_if_exists hash_func typed_table\n>> create_am select constraints updatable_views inherit triggers vacuum\n>> create_function_sql infinite_recurse\n>> ok 71 + create_aggregate 225 ms\n>> ok 72 + create_function_sql 18874 ms\n>> ok 73 + create_cast 168 ms\n>> \n>> ...(snip)...\n>> \n>> # All 215 tests passed.\n>> ```\n>> \n>> If you notice any difference, I would be grateful if you could let me\n>> know.\n> \n> I've never been able to reproduce it (haven't tested the new version,\n> but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\n> (first buster and now bullseye).\n> \n> I'm attaching several stacktraces in the hope that they provide some\n> clues. These all match the ps output I sent earlier, though note in\n> that output there is both the regress instance and my test instance\n> (pid 3213249) running (different ports, of course, and they are from\n> the exact same compilation run). I've attached ps output for the\n> postgres processes under the make check process to simplify cross\n> referencing.\n> \n> A few interesting things:\n> - There's definitely a lock on a relation that seems to be what's\n> blocking the processes.\n> - When I try to connect with psql the process forks but then hangs\n> (see the ps output with task names stuck in \"authentication\"). I've\n> also included a trace from one of these.\n\nThanks for sharing them!\n\nMany processes are waiting to acquire the LW lock, including the process \ntrying to output the plan(select1.trace).\n\nI suspect that this is due to a lock that was acquired prior to being \ninterrupted by ProcessLogQueryPlanInterrupt(), but have not been able to \nreproduce the same situation..\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Tue, 05 Sep 2023 22:58:59 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Tue, Sep 5, 2023 at 9:59 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-08-28 22:47, James Coleman wrote:\n> > On Mon, Aug 28, 2023 at 3:01 AM torikoshia <torikoshia@oss.nttdata.com>\n> > wrote:\n> >>\n> >> On 2023-08-26 21:03, James Coleman wrote:\n> >> > On Fri, Aug 25, 2023 at 7:43 AM James Coleman <jtc331@gmail.com> wrote:\n> >> >>\n> >> >> On Thu, Aug 17, 2023 at 10:02 AM torikoshia\n> >> >> <torikoshia@oss.nttdata.com> wrote:\n> >> >> >\n> >> >> > On 2023-06-16 01:34, James Coleman wrote:\n> >> >> > > Attached is v28\n> >> >> > > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\n> >> >> > > when necessary. Once built with those two patches I'm simply running\n> >> >> > > `make check`.\n> >> >> >\n> >> >> > With v28-0001 and v28-0002 patch, I confirmed backend processes consume\n> >> >> > huge\n> >> >> > amount of memory and under some environments they were terminated by OOM\n> >> >> > killer.\n> >> >> >\n> >> >> > This was because memory was allocated from existing memory contexts and\n> >> >> > they\n> >> >> > were not freed after ProcessLogQueryPlanInterrupt().\n> >> >> > Updated the patch to use dedicated memory context for\n> >> >> > ProcessLogQueryPlanInterrupt().\n> >> >> >\n> >> >> > Applying attached patch and v28-0002 patch, `make check` successfully\n> >> >> > completed after 20min and 50GB of logs on my environment.\n> >> >> >\n> >> >> > >>> On 2023-06-15 01:48, James Coleman wrote:\n> >> >> > >>> > The tests have been running since last night, but have been apparently\n> >> >> > >>> > hung now for many hours.\n> >> >> >\n> >> >> > I don't know if this has anything to do with the hung you faced, but I\n> >> >> > thought\n> >> >> > it might be possible that the large amount of memory usage resulted in\n> >> >> > swapping, which caused a significant delay in processing.\n> >> >>\n> >> >> Ah, yes, I think that could be a possible explanation. I was delaying\n> >> >> on this thread because I wasn't comfortable with having caused an\n> >> >> issue once (even if I couldn't easily reproduce) without at least some\n> >> >> theory as to the cause (and a fix).\n> >> >>\n> >> >> > If possible, I would be very grateful if you could try to reproduce this\n> >> >> > with\n> >> >> > the v29 patch.\n> >> >>\n> >> >> I'll kick off some testing.\n> >> >>\n> >> >\n> >> > I don't have time to investigate what's happening here, but 24 hours\n> >> > later the first \"make check\" is still running, and at first glance it\n> >> > seems to have the same behavior I'd seen that first time. The test\n> >> > output is to this point:\n> >> >\n> >> > # parallel group (5 tests): index_including create_view\n> >> > index_including_gist create_index create_index_spgist\n> >> > ok 66 + create_index 26365 ms\n> >> > ok 67 + create_index_spgist 27675 ms\n> >> > ok 68 + create_view 1235 ms\n> >> > ok 69 + index_including 1102 ms\n> >> > ok 70 + index_including_gist 1633 ms\n> >> > # parallel group (16 tests): create_aggregate create_cast errors\n> >> > roleattributes drop_if_exists hash_func typed_table create_am\n> >> > infinite_recurse\n> >> >\n> >> > and it hasn't progressed past that point since at least ~16 hours ago\n> >> > (the first several hours of the run I wasn't monitoring it).\n> >> >\n> >> > I haven't connected up gdb yet, and won't be able to until maybe\n> >> > tomorrow, but here's the ps output for postgres processes that are\n> >> > running:\n> >> >\n> >> > admin 3213249 0.0 0.0 196824 20552 ? Ss Aug25 0:00\n> >> > /home/admin/postgresql-test/bin/postgres -D\n> >> > /home/admin/postgresql-test-data\n> >> > admin 3213250 0.0 0.0 196964 7284 ? Ss Aug25 0:00\n> >> > postgres: checkpointer\n> >> > admin 3213251 0.0 0.0 196956 4276 ? Ss Aug25 0:00\n> >> > postgres: background writer\n> >> > admin 3213253 0.0 0.0 196956 8600 ? Ss Aug25 0:00\n> >> > postgres: walwriter\n> >> > admin 3213254 0.0 0.0 198424 7316 ? Ss Aug25 0:00\n> >> > postgres: autovacuum launcher\n> >> > admin 3213255 0.0 0.0 198412 5488 ? Ss Aug25 0:00\n> >> > postgres: logical replication launcher\n> >> > admin 3237967 0.0 0.0 2484 516 pts/4 S+ Aug25 0:00\n> >> > /bin/sh -c echo \"# +++ regress check in src/test/regress +++\" &&\n> >> > PATH=\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/bin:/home/admin/postgres/src/test/regress:$PATH\"\n> >> > LD_LIBRARY_PATH=\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/lib\"\n> >> > INITDB_TEMPLATE='/home/admin/postgres'/tmp_install/initdb-template\n> >> > ../../../src/test/regress/pg_regress --temp-instance=./tmp_check\n> >> > --inputdir=. --bindir= --dlpath=. --max-concurrent-tests=20\n> >> > --schedule=./parallel_schedule\n> >> > admin 3237973 0.0 0.0 197880 20688 pts/4 S+ Aug25 0:00\n> >> > postgres -D /home/admin/postgres/src/test/regress/tmp_check/data -F -c\n> >> > listen_addresses= -k /tmp/pg_regress-7mmGUa\n> >> > admin 3237976 0.0 0.1 198332 44608 ? Ss Aug25 0:00\n> >> > postgres: checkpointer\n> >> > admin 3237977 0.0 0.0 198032 4640 ? Ss Aug25 0:00\n> >> > postgres: background writer\n> >> > admin 3237979 0.0 0.0 197880 8580 ? Ss Aug25 0:00\n> >> > postgres: walwriter\n> >> > admin 3237980 0.0 0.0 199484 7608 ? Ss Aug25 0:00\n> >> > postgres: autovacuum launcher\n> >> > admin 3237981 0.0 0.0 199460 5488 ? Ss Aug25 0:00\n> >> > postgres: logical replication launcher\n> >> > admin 3243644 0.0 0.2 252400 74656 ? Ss Aug25 0:01\n> >> > postgres: admin regression [local] SELECT waiting\n> >> > admin 3243645 0.0 0.1 205480 33992 ? Ss Aug25 0:00\n> >> > postgres: admin regression [local] SELECT waiting\n> >> > admin 3243654 99.9 0.0 203156 31504 ? Rs Aug25 1437:49\n> >> > postgres: admin regression [local] VACUUM\n> >> > admin 3243655 0.0 0.1 212036 38504 ? Ss Aug25 0:00\n> >> > postgres: admin regression [local] SELECT waiting\n> >> > admin 3243656 0.0 0.0 206024 30892 ? Ss Aug25 0:00\n> >> > postgres: admin regression [local] DELETE waiting\n> >> > admin 3243657 0.0 0.1 205568 32232 ? Ss Aug25 0:00\n> >> > postgres: admin regression [local] ALTER TABLE waiting\n> >> > admin 3243658 0.0 0.0 203740 21532 ? Ss Aug25 0:00\n> >> > postgres: admin regression [local] ANALYZE waiting\n> >> > admin 3243798 0.0 0.0 199884 8464 ? Ss Aug25 0:00\n> >> > postgres: autovacuum worker\n> >> > admin 3244733 0.0 0.0 199492 5956 ? Ss Aug25 0:00\n> >> > postgres: autovacuum worker\n> >> > admin 3245652 0.0 0.0 199884 8468 ? Ss Aug25 0:00\n> >> > postgres: autovacuum worker\n> >> >\n> >> > As you can see there are a bunch of backends presumably waiting, and\n> >> > also the VACUUM process has been pegging a single CPU core for at\n> >> > least since that ~16 hour ago mark.\n> >> >\n> >> > I hope to be able to do more investigation later, but I wanted to at\n> >> > least give you this information now.\n> >>\n> >> Thanks a lot for testing the patch!\n> >> I really appreciate your cooperation.\n> >>\n> >> Hmm, I also tested on the current HEAD(165d581f146b09) again on Ubuntu\n> >> 22.04 and macOS, but unfortunately(fortunately?) they succeeded as\n> >> below:\n> >>\n> >> ```\n> >> $ git apply v29-0001-Add-function-to-log-the-plan-of-the-query.patch\n> >> $ git apply\n> >> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch\n> >> $ ./configure --enable-debug --enable-cassert\n> >> $ make\n> >> $ make check\n> >>\n> >> ...(snip)...\n> >>\n> >> # parallel group (5 tests): index_including index_including_gist\n> >> create_view create_index create_index_spgist\n> >> ok 66 + create_index 25033 ms\n> >> ok 67 + create_index_spgist 26144 ms\n> >> ok 68 + create_view 3061 ms\n> >> ok 69 + index_including 976 ms\n> >> ok 70 + index_including_gist 2998 ms\n> >> # parallel group (16 tests): create_cast errors create_aggregate\n> >> roleattributes drop_if_exists hash_func typed_table\n> >> create_am select constraints updatable_views inherit triggers vacuum\n> >> create_function_sql infinite_recurse\n> >> ok 71 + create_aggregate 225 ms\n> >> ok 72 + create_function_sql 18874 ms\n> >> ok 73 + create_cast 168 ms\n> >>\n> >> ...(snip)...\n> >>\n> >> # All 215 tests passed.\n> >> ```\n> >>\n> >> If you notice any difference, I would be grateful if you could let me\n> >> know.\n> >\n> > I've never been able to reproduce it (haven't tested the new version,\n> > but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\n> > (first buster and now bullseye).\n> >\n> > I'm attaching several stacktraces in the hope that they provide some\n> > clues. These all match the ps output I sent earlier, though note in\n> > that output there is both the regress instance and my test instance\n> > (pid 3213249) running (different ports, of course, and they are from\n> > the exact same compilation run). I've attached ps output for the\n> > postgres processes under the make check process to simplify cross\n> > referencing.\n> >\n> > A few interesting things:\n> > - There's definitely a lock on a relation that seems to be what's\n> > blocking the processes.\n> > - When I try to connect with psql the process forks but then hangs\n> > (see the ps output with task names stuck in \"authentication\"). I've\n> > also included a trace from one of these.\n>\n> Thanks for sharing them!\n>\n> Many processes are waiting to acquire the LW lock, including the process\n> trying to output the plan(select1.trace).\n>\n> I suspect that this is due to a lock that was acquired prior to being\n> interrupted by ProcessLogQueryPlanInterrupt(), but have not been able to\n> reproduce the same situation..\n>\n\nI don't have time immediately to dig into this, but perhaps loading up\nthe core dumps would allow us to see what query is running in each\nbackend process (if it hasn't already been discarded by that point)\nand thereby determine what point in each test process led to the error\ncondition?\n\nAlternatively we might be able to apply the same trick to the test\nclient instead...\n\nBTW, for my own easy reference in this thread: relid 1259 is pg_class\nif I'm not mistaken.\n\nThoughts?\nJames Coleman\n\n\n","msg_date":"Tue, 5 Sep 2023 22:17:08 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-09-06 11:17, James Coleman wrote:\n\n>> > I've never been able to reproduce it (haven't tested the new version,\n>> > but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\n>> > (first buster and now bullseye).\n>> >\n>> > I'm attaching several stacktraces in the hope that they provide some\n>> > clues. These all match the ps output I sent earlier, though note in\n>> > that output there is both the regress instance and my test instance\n>> > (pid 3213249) running (different ports, of course, and they are from\n>> > the exact same compilation run). I've attached ps output for the\n>> > postgres processes under the make check process to simplify cross\n>> > referencing.\n>> >\n>> > A few interesting things:\n>> > - There's definitely a lock on a relation that seems to be what's\n>> > blocking the processes.\n>> > - When I try to connect with psql the process forks but then hangs\n>> > (see the ps output with task names stuck in \"authentication\"). I've\n>> > also included a trace from one of these.\n>> \n>> Thanks for sharing them!\n>> \n>> Many processes are waiting to acquire the LW lock, including the \n>> process\n>> trying to output the plan(select1.trace).\n>> \n>> I suspect that this is due to a lock that was acquired prior to being\n>> interrupted by ProcessLogQueryPlanInterrupt(), but have not been able \n>> to\n>> reproduce the same situation..\n>> \n> \n> I don't have time immediately to dig into this, but perhaps loading up\n> the core dumps would allow us to see what query is running in each\n> backend process (if it hasn't already been discarded by that point)\n> and thereby determine what point in each test process led to the error\n> condition?\n\nThanks for the suggestion.\nI am concerned that core dumps may not be readable on different \noperating systems or other environments. (Unfortunately, I do not have \nDebian on hand)\n\nIt seems that we can know what queries were running from the stack \ntraces you shared.\nAs described above, I suspect a lock which was acquired prior to \nProcessLogQueryPlanInterrupt() caused the issue.\nI will try a little more to see if I can devise a way to create the same \nsituation.\n\n> Alternatively we might be able to apply the same trick to the test\n> client instead...\n> \n> BTW, for my own easy reference in this thread: relid 1259 is pg_class\n> if I'm not mistaken.\n\nYeah, and I think it's strange that the lock to 1259 appears twice and \nmust be avoided.\n\n #10 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) at \nlmgr.c:117\n ..\n #49 0x0000559d61b4989d in ProcessLogQueryPlanInterrupt () at \nexplain.c:5158\n ..\n #53 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) at \nlmgr.c:117\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Thu, 07 Sep 2023 15:09:12 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Thu, Sep 7, 2023, at 1:09 PM, torikoshia wrote:\n> On 2023-09-06 11:17, James Coleman wrote:\n> It seems that we can know what queries were running from the stack \n> traces you shared.\n> As described above, I suspect a lock which was acquired prior to \n> ProcessLogQueryPlanInterrupt() caused the issue.\n> I will try a little more to see if I can devise a way to create the same \n> situation.\nHi,\nI have explored this patch and, by and large, agree with the code. But some questions I want to discuss:\n1. Maybe add a hook to give a chance for extensions, like pg_query_state, to do their job?\n2. In this implementation, ProcessInterrupts does a lot of work during the explain creation: memory allocations, pass through the plan, systables locks, syscache access, etc. I guess it can change the semantic meaning of 'safe place' where CHECK_FOR_INTERRUPTS can be called - I can imagine a SELECT query, which would call a stored procedure, which executes some DDL and acquires row exclusive lock at the time of interruption. So, in my mind, it is too unpredictable to make the explain in the place of interruption processing. Maybe it is better to think about some hook at the end of ExecProcNode, where a pending explain could be created?\n\nAlso, I suggest minor code change with the diff in attachment. \n\n-- \nRegards,\nAndrei Lepikhov","msg_date":"Fri, 15 Sep 2023 13:21:27 +0700","msg_from":"\"Lepikhov Andrei\" <a.lepikhov@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-09-15 15:21, Lepikhov Andrei wrote:\n> On Thu, Sep 7, 2023, at 1:09 PM, torikoshia wrote:\n>> On 2023-09-06 11:17, James Coleman wrote:\n>> It seems that we can know what queries were running from the stack\n>> traces you shared.\n>> As described above, I suspect a lock which was acquired prior to\n>> ProcessLogQueryPlanInterrupt() caused the issue.\n>> I will try a little more to see if I can devise a way to create the \n>> same\n>> situation.\n> Hi,\n> I have explored this patch and, by and large, agree with the code. But\n> some questions I want to discuss:\n> 1. Maybe add a hook to give a chance for extensions, like\n> pg_query_state, to do their job?\n\nDo you imagine adding a hook which enables adding custom interrupt codes \nlike below?\n\nhttps://github.com/postgrespro/pg_query_state/blob/master/patches/custom_signals_15.0.patch\n\nIf so, that would be possible, but this patch doesn't require the \nfunctionality and I feel it'd be better doing in independent patch.\n\n\n> 2. In this implementation, ProcessInterrupts does a lot of work during\n> the explain creation: memory allocations, pass through the plan,\n> systables locks, syscache access, etc. I guess it can change the\n> semantic meaning of 'safe place' where CHECK_FOR_INTERRUPTS can be\n> called - I can imagine a SELECT query, which would call a stored\n> procedure, which executes some DDL and acquires row exclusive lock at\n> the time of interruption. So, in my mind, it is too unpredictable to\n> make the explain in the place of interruption processing. Maybe it is\n> better to think about some hook at the end of ExecProcNode, where a\n> pending explain could be created?\n\nYeah, I withdrew this patch once for that reason, but we are resuming \ndevelopment in response to the results of a discussion by James and \nothers at this year's pgcon about the safety of this process in CFI:\n\nhttps://www.postgresql.org/message-id/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw%40mail.gmail.com\n\nBTW it seems pg_query_state also enables users to get running query \nplans using CFI.\nDoes pg_query_state do something for this safety concern?\n\n> Also, I suggest minor code change with the diff in attachment.\n\nThanks!\n\nThis might be biased opinion and objections are welcomed, but I feel the \noriginal patch is easier to read since PG_RETURN_BOOL(true/false) is \nlocated in near place to each cases.\nAlso the existing function pg_log_backend_memory_contexts(), which does \nthe same thing, has the same form as the original patch.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Tue, 19 Sep 2023 22:39:59 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Tue, Sep 19, 2023, at 8:39 PM, torikoshia wrote:\n> On 2023-09-15 15:21, Lepikhov Andrei wrote:\n>> On Thu, Sep 7, 2023, at 1:09 PM, torikoshia wrote:\n>> I have explored this patch and, by and large, agree with the code. But\n>> some questions I want to discuss:\n>> 1. Maybe add a hook to give a chance for extensions, like\n>> pg_query_state, to do their job?\n>\n> Do you imagine adding a hook which enables adding custom interrupt codes \n> like below?\n>\n> https://github.com/postgrespro/pg_query_state/blob/master/patches/custom_signals_15.0.patch\n\nNo, I think around the hook, which would allow us to rewrite the pg_query_state extension without additional patches by using the functionality provided by your patch. I mean, an extension could provide console UI, not only server logging. And obtain from target backend so much information in the explain as the instrumentation level of the current query can give.\nIt may make pg_query_state shorter and more stable.\n\n>> 2. In this implementation, ProcessInterrupts does a lot of work during\n>> the explain creation: memory allocations, pass through the plan,\n>> systables locks, syscache access, etc. I guess it can change the\n>> semantic meaning of 'safe place' where CHECK_FOR_INTERRUPTS can be\n>> called - I can imagine a SELECT query, which would call a stored\n>> procedure, which executes some DDL and acquires row exclusive lock at\n>> the time of interruption. So, in my mind, it is too unpredictable to\n>> make the explain in the place of interruption processing. Maybe it is\n>> better to think about some hook at the end of ExecProcNode, where a\n>> pending explain could be created?\n>\n> Yeah, I withdrew this patch once for that reason, but we are resuming \n> development in response to the results of a discussion by James and \n> others at this year's pgcon about the safety of this process in CFI:\n>\n> https://www.postgresql.org/message-id/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw%40mail.gmail.com\n\nI can't track the logic path of the decision provided at this conference. But my anxiety related to specific place, where ActiveQueryDesc points top-level query and interruption comes during DDL, wrapped up in stored procedure. For example:\nCREATE TABLE test();\nCREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\nBEGIN\n EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\n ...\nEND; $$ LANGUAGE plpgsql VOLATILE;\nSELECT ddl(), ... FROM ...;\n\n> BTW it seems pg_query_state also enables users to get running query \n> plans using CFI.\n> Does pg_query_state do something for this safety concern?\n\nNo, and I'm looking for the solution, which could help to rewrite pg_query_state as a clean extension, without patches.\n\n>> Also, I suggest minor code change with the diff in attachment.\n>\n> Thanks!\n>\n> This might be biased opinion and objections are welcomed, but I feel the \n> original patch is easier to read since PG_RETURN_BOOL(true/false) is \n> located in near place to each cases.\n> Also the existing function pg_log_backend_memory_contexts(), which does \n> the same thing, has the same form as the original patch.\nI got it, thank you.\n\n-- \nRegards,\nAndrei Lepikhov\n\n\n","msg_date":"Wed, 20 Sep 2023 12:39:13 +0700","msg_from":"\"Lepikhov Andrei\" <a.lepikhov@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-09-20 14:39, Lepikhov Andrei wrote:\nThanks for your reply.\n\n> On Tue, Sep 19, 2023, at 8:39 PM, torikoshia wrote:\n>> On 2023-09-15 15:21, Lepikhov Andrei wrote:\n>>> On Thu, Sep 7, 2023, at 1:09 PM, torikoshia wrote:\n>>> I have explored this patch and, by and large, agree with the code. \n>>> But\n>>> some questions I want to discuss:\n>>> 1. Maybe add a hook to give a chance for extensions, like\n>>> pg_query_state, to do their job?\n>> \n>> Do you imagine adding a hook which enables adding custom interrupt \n>> codes\n>> like below?\n>> \n>> https://github.com/postgrespro/pg_query_state/blob/master/patches/custom_signals_15.0.patch\n> \n> No, I think around the hook, which would allow us to rewrite the\n> pg_query_state extension without additional patches by using the\n> functionality provided by your patch. I mean, an extension could\n> provide console UI, not only server logging. And obtain from target\n> backend so much information in the explain as the instrumentation\n> level of the current query can give.\n> It may make pg_query_state shorter and more stable.\n> \n>>> 2. In this implementation, ProcessInterrupts does a lot of work \n>>> during\n>>> the explain creation: memory allocations, pass through the plan,\n>>> systables locks, syscache access, etc. I guess it can change the\n>>> semantic meaning of 'safe place' where CHECK_FOR_INTERRUPTS can be\n>>> called - I can imagine a SELECT query, which would call a stored\n>>> procedure, which executes some DDL and acquires row exclusive lock at\n>>> the time of interruption. So, in my mind, it is too unpredictable to\n>>> make the explain in the place of interruption processing. Maybe it is\n>>> better to think about some hook at the end of ExecProcNode, where a\n>>> pending explain could be created?\n>> \n>> Yeah, I withdrew this patch once for that reason, but we are resuming\n>> development in response to the results of a discussion by James and\n>> others at this year's pgcon about the safety of this process in CFI:\n>> \n>> https://www.postgresql.org/message-id/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw%40mail.gmail.com\n> \n> I can't track the logic path of the decision provided at this\n> conference. But my anxiety related to specific place, where\n> ActiveQueryDesc points top-level query and interruption comes during\n> DDL, wrapped up in stored procedure. For example:\n> CREATE TABLE test();\n> CREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\n> BEGIN\n> EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\n> ...\n> END; $$ LANGUAGE plpgsql VOLATILE;\n> SELECT ddl(), ... FROM ...;\n\nHmm, as a test, I made sure to call ProcessLogQueryPlanInterrupt() on \nall CFI using \nv28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch, and then \nran the following query but did not cause any problems.\n\n```\n=# CREATE TABLE test();\n=# CREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\nBEGIN\n EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\n PERFORM pg_sleep(5);\nEND; $$ LANGUAGE plpgsql VOLATILE;\n=# SELECT ddl();\n```\n\nIs this the case you're worrying about?\n\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Mon, 25 Sep 2023 16:21:31 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 25/9/2023 14:21, torikoshia wrote:\n> On 2023-09-20 14:39, Lepikhov Andrei wrote:\n> Hmm, as a test, I made sure to call ProcessLogQueryPlanInterrupt() on \n> all CFI using \n> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch, and then \n> ran the following query but did not cause any problems.\n> \n> ```\n> =# CREATE TABLE test();\n> =# CREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\n> BEGIN\n> EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\n> PERFORM pg_sleep(5);\n> END; $$ LANGUAGE plpgsql VOLATILE;\n> =# SELECT ddl();\n> ```\n> \n> Is this the case you're worrying about?\n\nI didn't find a problem either. I just feel uncomfortable if, at the \nmoment of interruption, we have a descriptor of another query than the \nquery have been executing and holding resources.\n\n-- \nregards,\nAndrey Lepikhov\nPostgres Professional\n\n\n\n","msg_date":"Mon, 25 Sep 2023 16:49:18 +0700","msg_from":"Andrey Lepikhov <a.lepikhov@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-09-25 18:49, Andrey Lepikhov wrote:\n> On 25/9/2023 14:21, torikoshia wrote:\n>> On 2023-09-20 14:39, Lepikhov Andrei wrote:\n>> Hmm, as a test, I made sure to call ProcessLogQueryPlanInterrupt() on \n>> all CFI using \n>> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch, and then \n>> ran the following query but did not cause any problems.\n>> \n>> ```\n>> =# CREATE TABLE test();\n>> =# CREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\n>> BEGIN\n>> EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\n>> PERFORM pg_sleep(5);\n>> END; $$ LANGUAGE plpgsql VOLATILE;\n>> =# SELECT ddl();\n>> ```\n>> \n>> Is this the case you're worrying about?\n> \n> I didn't find a problem either. I just feel uncomfortable if, at the\n> moment of interruption, we have a descriptor of another query than the\n> query have been executing and holding resources.\n\nI think that \"descriptor\" here refers to ActiveQueryDesc, in which case\nit is updated at the beginning of ExecutorRun(), so I am wondering if\nthe situation you're worried about would not occur.\n\n---------\n@@ -303,10 +306,21 @@ ExecutorRun(QueryDesc *queryDesc,\n ScanDirection direction, uint64 count,\n bool execute_once)\n {\n+ /*\n+ * Update ActiveQueryDesc here to enable retrieval of the currently\n+ * running queryDesc for nested queries.\n+ */\n+ QueryDesc *save_ActiveQueryDesc;\n+\n+ save_ActiveQueryDesc = ActiveQueryDesc;\n+ ActiveQueryDesc = queryDesc;\n---------\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Thu, 28 Sep 2023 11:04:09 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 28/9/2023 09:04, torikoshia wrote:\n> On 2023-09-25 18:49, Andrey Lepikhov wrote:\n>> On 25/9/2023 14:21, torikoshia wrote:\n>>> On 2023-09-20 14:39, Lepikhov Andrei wrote:\n>>> Hmm, as a test, I made sure to call ProcessLogQueryPlanInterrupt() on \n>>> all CFI using \n>>> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch, and \n>>> then ran the following query but did not cause any problems.\n>>>\n>>> ```\n>>> =# CREATE TABLE test();\n>>> =# CREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\n>>> BEGIN\n>>> EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\n>>> PERFORM pg_sleep(5);\n>>> END; $$ LANGUAGE plpgsql VOLATILE;\n>>> =# SELECT ddl();\n>>> ```\n>>>\n>>> Is this the case you're worrying about?\n>>\n>> I didn't find a problem either. I just feel uncomfortable if, at the\n>> moment of interruption, we have a descriptor of another query than the\n>> query have been executing and holding resources.\n> \n> I think that \"descriptor\" here refers to ActiveQueryDesc, in which case\n> it is updated at the beginning of ExecutorRun(), so I am wondering if\n> the situation you're worried about would not occur.\nAs you can see, in my example we have the only DDL and no queries with \nplans. In this case postgres doesn't call ExecutorRun() just because it \ndoesn't have a plan. But locks will be obtained.\n\n-- \nregards,\nAndrey Lepikhov\nPostgres Professional\n\n\n\n","msg_date":"Thu, 28 Sep 2023 09:32:46 +0700","msg_from":"Andrey Lepikhov <a.lepikhov@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Thu, Sep 7, 2023 at 2:09 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-09-06 11:17, James Coleman wrote:\n>\n> >> > I've never been able to reproduce it (haven't tested the new version,\n> >> > but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\n> >> > (first buster and now bullseye).\n> >> >\n> >> > I'm attaching several stacktraces in the hope that they provide some\n> >> > clues. These all match the ps output I sent earlier, though note in\n> >> > that output there is both the regress instance and my test instance\n> >> > (pid 3213249) running (different ports, of course, and they are from\n> >> > the exact same compilation run). I've attached ps output for the\n> >> > postgres processes under the make check process to simplify cross\n> >> > referencing.\n> >> >\n> >> > A few interesting things:\n> >> > - There's definitely a lock on a relation that seems to be what's\n> >> > blocking the processes.\n> >> > - When I try to connect with psql the process forks but then hangs\n> >> > (see the ps output with task names stuck in \"authentication\"). I've\n> >> > also included a trace from one of these.\n> >>\n> >> Thanks for sharing them!\n> >>\n> >> Many processes are waiting to acquire the LW lock, including the\n> >> process\n> >> trying to output the plan(select1.trace).\n> >>\n> >> I suspect that this is due to a lock that was acquired prior to being\n> >> interrupted by ProcessLogQueryPlanInterrupt(), but have not been able\n> >> to\n> >> reproduce the same situation..\n> >>\n> >\n> > I don't have time immediately to dig into this, but perhaps loading up\n> > the core dumps would allow us to see what query is running in each\n> > backend process (if it hasn't already been discarded by that point)\n> > and thereby determine what point in each test process led to the error\n> > condition?\n>\n> Thanks for the suggestion.\n> I am concerned that core dumps may not be readable on different\n> operating systems or other environments. (Unfortunately, I do not have\n> Debian on hand)\n>\n> It seems that we can know what queries were running from the stack\n> traces you shared.\n> As described above, I suspect a lock which was acquired prior to\n> ProcessLogQueryPlanInterrupt() caused the issue.\n> I will try a little more to see if I can devise a way to create the same\n> situation.\n>\n> > Alternatively we might be able to apply the same trick to the test\n> > client instead...\n> >\n> > BTW, for my own easy reference in this thread: relid 1259 is pg_class\n> > if I'm not mistaken.\n>\n> Yeah, and I think it's strange that the lock to 1259 appears twice and\n> must be avoided.\n>\n> #10 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) at\n> lmgr.c:117\n> ..\n> #49 0x0000559d61b4989d in ProcessLogQueryPlanInterrupt () at\n> explain.c:5158\n> ..\n> #53 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) at\n> lmgr.c:117\n\nI chatted with Andres and David about this at PGConf.NYC, and I think\nwhat we need to do is explicitly disallow running this code any time\nwe are inside of lock acquisition code.\n\nRegards,\nJames Coleman\n\n\n","msg_date":"Tue, 3 Oct 2023 14:00:11 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-10-04 03:00, James Coleman wrote:\n> On Thu, Sep 7, 2023 at 2:09 AM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> \n>> On 2023-09-06 11:17, James Coleman wrote:\n>> \n>> >> > I've never been able to reproduce it (haven't tested the new version,\n>> >> > but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\n>> >> > (first buster and now bullseye).\n>> >> >\n>> >> > I'm attaching several stacktraces in the hope that they provide some\n>> >> > clues. These all match the ps output I sent earlier, though note in\n>> >> > that output there is both the regress instance and my test instance\n>> >> > (pid 3213249) running (different ports, of course, and they are from\n>> >> > the exact same compilation run). I've attached ps output for the\n>> >> > postgres processes under the make check process to simplify cross\n>> >> > referencing.\n>> >> >\n>> >> > A few interesting things:\n>> >> > - There's definitely a lock on a relation that seems to be what's\n>> >> > blocking the processes.\n>> >> > - When I try to connect with psql the process forks but then hangs\n>> >> > (see the ps output with task names stuck in \"authentication\"). I've\n>> >> > also included a trace from one of these.\n>> >>\n>> >> Thanks for sharing them!\n>> >>\n>> >> Many processes are waiting to acquire the LW lock, including the\n>> >> process\n>> >> trying to output the plan(select1.trace).\n>> >>\n>> >> I suspect that this is due to a lock that was acquired prior to being\n>> >> interrupted by ProcessLogQueryPlanInterrupt(), but have not been able\n>> >> to\n>> >> reproduce the same situation..\n>> >>\n>> >\n>> > I don't have time immediately to dig into this, but perhaps loading up\n>> > the core dumps would allow us to see what query is running in each\n>> > backend process (if it hasn't already been discarded by that point)\n>> > and thereby determine what point in each test process led to the error\n>> > condition?\n>> \n>> Thanks for the suggestion.\n>> I am concerned that core dumps may not be readable on different\n>> operating systems or other environments. (Unfortunately, I do not have\n>> Debian on hand)\n>> \n>> It seems that we can know what queries were running from the stack\n>> traces you shared.\n>> As described above, I suspect a lock which was acquired prior to\n>> ProcessLogQueryPlanInterrupt() caused the issue.\n>> I will try a little more to see if I can devise a way to create the \n>> same\n>> situation.\n>> \n>> > Alternatively we might be able to apply the same trick to the test\n>> > client instead...\n>> >\n>> > BTW, for my own easy reference in this thread: relid 1259 is pg_class\n>> > if I'm not mistaken.\n>> \n>> Yeah, and I think it's strange that the lock to 1259 appears twice and\n>> must be avoided.\n>> \n>> #10 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) \n>> at\n>> lmgr.c:117\n>> ..\n>> #49 0x0000559d61b4989d in ProcessLogQueryPlanInterrupt () at\n>> explain.c:5158\n>> ..\n>> #53 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) \n>> at\n>> lmgr.c:117\n> \n> I chatted with Andres and David about this at PGConf.NYC,\nThanks again for the discussion with hackers!\n\n> and I think\n> what we need to do is explicitly disallow running this code any time\n> we are inside of lock acquisition code.\n\nYeah, I think it's a straightforward workaround.\nAnd I'm also concerned that the condition of being in the process\nof acquiring some kind of lock is too strict and will make it almost\nimpossible to output a running plan.\n\nAnyway I'm going to implement it and run pg_log_query_plan()\nwhile the regression test is running to see how successful the plan\noutput is.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Fri, 06 Oct 2023 21:58:46 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Fri, Oct 6, 2023 at 8:58 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-10-04 03:00, James Coleman wrote:\n> > On Thu, Sep 7, 2023 at 2:09 AM torikoshia <torikoshia@oss.nttdata.com>\n> > wrote:\n> >>\n> >> On 2023-09-06 11:17, James Coleman wrote:\n> >>\n> >> >> > I've never been able to reproduce it (haven't tested the new version,\n> >> >> > but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\n> >> >> > (first buster and now bullseye).\n> >> >> >\n> >> >> > I'm attaching several stacktraces in the hope that they provide some\n> >> >> > clues. These all match the ps output I sent earlier, though note in\n> >> >> > that output there is both the regress instance and my test instance\n> >> >> > (pid 3213249) running (different ports, of course, and they are from\n> >> >> > the exact same compilation run). I've attached ps output for the\n> >> >> > postgres processes under the make check process to simplify cross\n> >> >> > referencing.\n> >> >> >\n> >> >> > A few interesting things:\n> >> >> > - There's definitely a lock on a relation that seems to be what's\n> >> >> > blocking the processes.\n> >> >> > - When I try to connect with psql the process forks but then hangs\n> >> >> > (see the ps output with task names stuck in \"authentication\"). I've\n> >> >> > also included a trace from one of these.\n> >> >>\n> >> >> Thanks for sharing them!\n> >> >>\n> >> >> Many processes are waiting to acquire the LW lock, including the\n> >> >> process\n> >> >> trying to output the plan(select1.trace).\n> >> >>\n> >> >> I suspect that this is due to a lock that was acquired prior to being\n> >> >> interrupted by ProcessLogQueryPlanInterrupt(), but have not been able\n> >> >> to\n> >> >> reproduce the same situation..\n> >> >>\n> >> >\n> >> > I don't have time immediately to dig into this, but perhaps loading up\n> >> > the core dumps would allow us to see what query is running in each\n> >> > backend process (if it hasn't already been discarded by that point)\n> >> > and thereby determine what point in each test process led to the error\n> >> > condition?\n> >>\n> >> Thanks for the suggestion.\n> >> I am concerned that core dumps may not be readable on different\n> >> operating systems or other environments. (Unfortunately, I do not have\n> >> Debian on hand)\n> >>\n> >> It seems that we can know what queries were running from the stack\n> >> traces you shared.\n> >> As described above, I suspect a lock which was acquired prior to\n> >> ProcessLogQueryPlanInterrupt() caused the issue.\n> >> I will try a little more to see if I can devise a way to create the\n> >> same\n> >> situation.\n> >>\n> >> > Alternatively we might be able to apply the same trick to the test\n> >> > client instead...\n> >> >\n> >> > BTW, for my own easy reference in this thread: relid 1259 is pg_class\n> >> > if I'm not mistaken.\n> >>\n> >> Yeah, and I think it's strange that the lock to 1259 appears twice and\n> >> must be avoided.\n> >>\n> >> #10 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1)\n> >> at\n> >> lmgr.c:117\n> >> ..\n> >> #49 0x0000559d61b4989d in ProcessLogQueryPlanInterrupt () at\n> >> explain.c:5158\n> >> ..\n> >> #53 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1)\n> >> at\n> >> lmgr.c:117\n> >\n> > I chatted with Andres and David about this at PGConf.NYC,\n> Thanks again for the discussion with hackers!\n>\n> > and I think\n> > what we need to do is explicitly disallow running this code any time\n> > we are inside of lock acquisition code.\n>\n> Yeah, I think it's a straightforward workaround.\n> And I'm also concerned that the condition of being in the process\n> of acquiring some kind of lock is too strict and will make it almost\n> impossible to output a running plan.\n\nI was concerned about this too, but I was wondering if we might be\nable to cheat a bit by making such a case not clear the flag but\ninstead just skip it for now.\n\nRegards,\nJames\n\n\n","msg_date":"Fri, 6 Oct 2023 09:10:15 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi,\n\nOn 2023-10-06 21:58:46 +0900, torikoshia wrote:\n> Yeah, I think it's a straightforward workaround.\n> And I'm also concerned that the condition of being in the process\n> of acquiring some kind of lock is too strict and will make it almost\n> impossible to output a running plan.\n\nHow so? We shouldn't commonly acquire relevant locks while executing a query?\nWith a few exceptions, they should instead be acquired t the start of query\nprocessing. We do acquire a lot of lwlocks, obviously, but we don't process\ninterrupts during the acquisition / holding of lwlocks.\n\nAnd presumably the interrupt would just be processed the next time interrupt\nprocessing is happening?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 6 Oct 2023 08:58:50 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-10-04 03:00, James Coleman wrote:\n> and I think\n> what we need to do is explicitly disallow running this code any time\n> we are inside of lock acquisition code.\n\nUpdated patch to check if any locks have already been acquired by \nexamining MyProc->heldLocks.\n\nI'm not sure this change can \"disallow running this code `any time` we \nare inside of lock acquisition code\", but as far as select1.trace, which \nyou shared, I believe it can prevent running explain codes since it must \nhave set MyProc->heldLocks in LockAcquireExtended() before WaitOnLock():\n\n```\n /*\n * Set bitmask of locks this process already holds on this \nobject.\n */\n MyProc->heldLocks = proclock->holdMask;\n\n ..(snip)..\n\n WaitOnLock(locallock, owner);\n```\n\nOn 2023-10-07 00:58, Andres Freund wrote:\n\n> How so? We shouldn't commonly acquire relevant locks while executing a \n> query?\n> With a few exceptions, they should instead be acquired t the start of \n> query\n> processing. We do acquire a lot of lwlocks, obviously, but we don't \n> process\n> interrupts during the acquisition / holding of lwlocks.\n> \n> And presumably the interrupt would just be processed the next time \n> interrupt\n> processing is happening?\n\nThanks for your comments!\n\nI tested v30 patch with \nv28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch which makes \nCFI() call ProcessLogQueryPlanInterrupt() internally, and confirmed that \nvery few logging queries failed with v30 patch.\n\nThis is a result in line with your prediction.\n\n\n```\n $ rg -c'ignored request for logging query plan due to lock confilcts' \npostmaster.log\n 8\n```\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation","msg_date":"Tue, 10 Oct 2023 22:27:08 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi,\n\nOn Tue, Oct 10, 2023 at 7:00 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\n\n> Thanks for your comments!\n>\n> I tested v30 patch with\n> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch which makes\n> CFI() call ProcessLogQueryPlanInterrupt() internally, and confirmed that\n> very few logging queries failed with v30 patch.\n>\n> This is a result in line with your prediction.\n>\n>\n> ```\n> $ rg -c'ignored request for logging query plan due to lock confilcts'\n> postmaster.log\n> 8\n\nLike many others I think this feature is useful to debug a long running query.\n\nSorry for jumping late into this.\n\nI have a few of high level comments\n\nThere is a lot of similarity between what this feature does and what\nauto explain does. I see the code is also duplicated. There is some\nmerit in avoiding this duplication\n1. we will get all the features of auto_explain automatically like\nchoosing a format (this was expressed somebody earlier in this\nthread), setings etc.\n2. avoid bugs. E.g your code switches context after ExplainState has\nbeen allocated. These states may leak depending upon when this\nfunction gets called.\n3. Building features on top as James envisions will be easier.\n\nConsidering the similarity with auto_explain I wondered whether this\nfunction should be part of auto_explain contrib module itself? If we\ndo that users will need to load auto_explain extension and thus\ninstall executor hooks when this function doesn't need those. So may\nnot be such a good idea. I didn't see any discussion on this.\n\nI tried following query to pass PID of a non-client backend to this function.\n#select pg_log_query_plan(pid), application_name, backend_type from\npg_stat_activity where backend_type = 'autovacuum launcher';\n pg_log_query_plan | application_name | backend_type\n-------------------+------------------+---------------------\n t | | autovacuum launcher\n(1 row)\nI see \"LOG: backend with PID 2733631 is not running a query or a\nsubtransaction is aborted\" in server logs. That's ok. But may be we\nshould not send signal to these kinds of backends at all, thus\navoiding some system calls.\n\nI am also wondering whether it's better to report the WARNING as\nstatus column in the output. E.g. instead of\n#select pg_log_query_plan(100);\nWARNING: PID 100 is not a PostgreSQL backend process\n pg_log_query_plan\n-------------------\n f\n(1 row)\nwe output\n#select pg_log_query_plan(100);\n pg_log_query_plan | status\n-------------------+---------------------------------------------\n f | PID 100 is not a PostgreSQL backend process\n(1 row)\n\nThat looks neater and can easily be handled by scripts, applications\nand such. But it will be inconsistent with other functions like\npg_terminate_backend() and pg_log_backend_memory_contexts().\n\nI do share a concern that was discussed earlier. If a query is running\nlonger, there's something problematic with it. A diagnostic\nintervention breaking it further would be unwelcome. James has run\nexperiments to shake this code for any loose breakages. He has not\nfound any. So may be we are good. And we wouldn't know about very rare\ncorner cases so easily without using it in the field. So fine with it.\nIf we could add some safety net that will be great but may not be\nnecessary for the first cut.\n\n--\nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Wed, 11 Oct 2023 12:52:00 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-10-11 16:22, Ashutosh Bapat wrote:\n\n> Like many others I think this feature is useful to debug a long running \n> query.\n> \n> Sorry for jumping late into this.\n> \n> I have a few of high level comments\n\nThanks for your comments!\n\n> There is a lot of similarity between what this feature does and what\n> auto explain does. I see the code is also duplicated. There is some\n> merit in avoiding this duplication\n> 1. we will get all the features of auto_explain automatically like\n> choosing a format (this was expressed somebody earlier in this\n> thread), setings etc.\n> 2. avoid bugs. E.g your code switches context after ExplainState has\n> been allocated. These states may leak depending upon when this\n> function gets called.\n> 3. Building features on top as James envisions will be easier.\n> \n> Considering the similarity with auto_explain I wondered whether this\n> function should be part of auto_explain contrib module itself? If we\n> do that users will need to load auto_explain extension and thus\n> install executor hooks when this function doesn't need those. So may\n> not be such a good idea. I didn't see any discussion on this.\n\nI once thought about adding this to auto_explain, but I left it asis for \nbelow reasons:\n\n- One of the typical use case of pg_log_query_plan() would be analyzing \nslow query on customer environments. On such environments, We cannot \nalways control what extensions to install.\n Of course auto_explain is a major extension and it is quite possible \nthat they installed auto_explain, but but it is also possible they do \nnot.\n- It seems a bit counter-intuitive that pg_log_query_plan() is in an \nextension called auto_explain, since it `manually`` logs plans\n\n> I tried following query to pass PID of a non-client backend to this \n> function.\n> #select pg_log_query_plan(pid), application_name, backend_type from\n> pg_stat_activity where backend_type = 'autovacuum launcher';\n> pg_log_query_plan | application_name | backend_type\n> -------------------+------------------+---------------------\n> t | | autovacuum launcher\n> (1 row)\n> I see \"LOG: backend with PID 2733631 is not running a query or a\n> subtransaction is aborted\" in server logs. That's ok. But may be we\n> should not send signal to these kinds of backends at all, thus\n> avoiding some system calls.\n\nAgreed, it seems better.\nAttached patch checks if the backendType of target process is 'client \nbackend'.\n\n =# select pg_log_query_plan(pid), application_name, backend_type from \npg_stat_activity where backend_type = 'autovacuum launcher';\n WARNING: PID 63323 is not a PostgreSQL client backend process\n pg_log_query_plan | application_name | backend_type\n -------------------+------------------+---------------------\n f | | autovacuum launcher\n\n\n> I am also wondering whether it's better to report the WARNING as\n> status column in the output. E.g. instead of\n> #select pg_log_query_plan(100);\n> WARNING: PID 100 is not a PostgreSQL backend process\n> pg_log_query_plan\n> -------------------\n> f\n> (1 row)\n> we output\n> #select pg_log_query_plan(100);\n> pg_log_query_plan | status\n> -------------------+---------------------------------------------\n> f | PID 100 is not a PostgreSQL backend process\n> (1 row)\n> \n> That looks neater and can easily be handled by scripts, applications\n> and such. But it will be inconsistent with other functions like\n> pg_terminate_backend() and pg_log_backend_memory_contexts().\n\nIt seems neater, but it might be inconvenient because we can no longer \nuse it in select list like the following query as you wrote:\n\n #select pg_log_query_plan(pid), application_name, backend_type from\n pg_stat_activity where backend_type = 'autovacuum launcher';\n\n> I do share a concern that was discussed earlier. If a query is running\n> longer, there's something problematic with it. A diagnostic\n> intervention breaking it further would be unwelcome. James has run\n> experiments to shake this code for any loose breakages. He has not\n> found any. So may be we are good. And we wouldn't know about very rare\n> corner cases so easily without using it in the field. So fine with it.\n> If we could add some safety net that will be great but may not be\n> necessary for the first cut.\n\nIf there are candidates for the safety net, I'm willing to add them.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation","msg_date":"Thu, 12 Oct 2023 22:21:26 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Thu, Oct 12, 2023 at 6:51 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-10-11 16:22, Ashutosh Bapat wrote:\n> >\n> > Considering the similarity with auto_explain I wondered whether this\n> > function should be part of auto_explain contrib module itself? If we\n> > do that users will need to load auto_explain extension and thus\n> > install executor hooks when this function doesn't need those. So may\n> > not be such a good idea. I didn't see any discussion on this.\n>\n> I once thought about adding this to auto_explain, but I left it asis for\n> below reasons:\n>\n> - One of the typical use case of pg_log_query_plan() would be analyzing\n> slow query on customer environments. On such environments, We cannot\n> always control what extensions to install.\n\nThe same argument applies to auto_explain functionality as well. But\nit's not part of the core.\n\n> Of course auto_explain is a major extension and it is quite possible\n> that they installed auto_explain, but but it is also possible they do\n> not.\n> - It seems a bit counter-intuitive that pg_log_query_plan() is in an\n> extension called auto_explain, since it `manually`` logs plans\n>\n\npg_log_query_plan() may not fit auto_explain but\npg_explain_backend_query() does. What we are logging is more than just\nplan of the query, it might expand to be closer to explain output.\nWhile auto in auto_explain would refer to its automatically logging\nexplain outputs, it can provide an additional function which provides\nsimilar functionality by manually triggering it.\n\nBut we can defer this to a committer, if you want.\n\nI am more interested in avoiding the duplication of code, esp. the\nfirst comment in my reply\n>> There is a lot of similarity between what this feature does and what\n>> auto explain does. I see the code is also duplicated. There is some\n>> merit in avoiding this duplication\n>> 1. we will get all the features of auto_explain automatically like\n>> choosing a format (this was expressed somebody earlier in this\n>> thread), setings etc.\n>> 2. avoid bugs. E.g your code switches context after ExplainState has\n>> been allocated. These states may leak depending upon when this\n>> function gets called.\n>> 3. Building features on top as James envisions will be easier.\n\n>\n> =# select pg_log_query_plan(pid), application_name, backend_type from\n> pg_stat_activity where backend_type = 'autovacuum launcher';\n> WARNING: PID 63323 is not a PostgreSQL client backend process\n> pg_log_query_plan | application_name | backend_type\n> -------------------+------------------+---------------------\n> f | | autovacuum launcher\n>\n>\n> > I am also wondering whether it's better to report the WARNING as\n> > status column in the output. E.g. instead of\n> > #select pg_log_query_plan(100);\n> > WARNING: PID 100 is not a PostgreSQL backend process\n> > pg_log_query_plan\n> > -------------------\n> > f\n> > (1 row)\n> > we output\n> > #select pg_log_query_plan(100);\n> > pg_log_query_plan | status\n> > -------------------+---------------------------------------------\n> > f | PID 100 is not a PostgreSQL backend process\n> > (1 row)\n> >\n> > That looks neater and can easily be handled by scripts, applications\n> > and such. But it will be inconsistent with other functions like\n> > pg_terminate_backend() and pg_log_backend_memory_contexts().\n>\n> It seems neater, but it might be inconvenient because we can no longer\n> use it in select list like the following query as you wrote:\n>\n> #select pg_log_query_plan(pid), application_name, backend_type from\n> pg_stat_activity where backend_type = 'autovacuum launcher';\n\nWhy is that?\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Mon, 16 Oct 2023 15:16:28 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-10-16 18:46, Ashutosh Bapat wrote:\n> On Thu, Oct 12, 2023 at 6:51 PM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> \n>> On 2023-10-11 16:22, Ashutosh Bapat wrote:\n>> >\n>> > Considering the similarity with auto_explain I wondered whether this\n>> > function should be part of auto_explain contrib module itself? If we\n>> > do that users will need to load auto_explain extension and thus\n>> > install executor hooks when this function doesn't need those. So may\n>> > not be such a good idea. I didn't see any discussion on this.\n>> \n>> I once thought about adding this to auto_explain, but I left it asis \n>> for\n>> below reasons:\n>> \n>> - One of the typical use case of pg_log_query_plan() would be \n>> analyzing\n>> slow query on customer environments. On such environments, We cannot\n>> always control what extensions to install.\n> \n> The same argument applies to auto_explain functionality as well. But\n> it's not part of the core.\n\nYeah, and when we have a situation where we want to run\npg_log_query_plan(), we can run it in any environment as long as it is\n bundled with the core.\nOn the other hand, if it is built into auto_explain, we need to start by\ninstalling auto_explain if we do not have auto_explain, which is often\ndifficult to do in production environments.\n\n>> Of course auto_explain is a major extension and it is quite \n>> possible\n>> that they installed auto_explain, but but it is also possible they do\n>> not.\n>> - It seems a bit counter-intuitive that pg_log_query_plan() is in an\n>> extension called auto_explain, since it `manually`` logs plans\n>> \n> \n> pg_log_query_plan() may not fit auto_explain but\n> pg_explain_backend_query() does. What we are logging is more than just\n> plan of the query, it might expand to be closer to explain output.\n> While auto in auto_explain would refer to its automatically logging\n> explain outputs, it can provide an additional function which provides\n> similar functionality by manually triggering it.\n> \n> But we can defer this to a committer, if you want.\n> \n> I am more interested in avoiding the duplication of code, esp. the\n> first comment in my reply\n\nIf there are no objections, I will try porting it to auto_explain and\nsee its feasibility.\n\n>>> There is a lot of similarity between what this feature does and what\n>>> auto explain does. I see the code is also duplicated. There is some\n>>> merit in avoiding this duplication\n>>> 1. we will get all the features of auto_explain automatically like\n>>> choosing a format (this was expressed somebody earlier in this\n>>> thread), setings etc.\n>>> 2. avoid bugs. E.g your code switches context after ExplainState has\n>>> been allocated. These states may leak depending upon when this\n>>> function gets called.\n>>> 3. Building features on top as James envisions will be easier.\n> \n>> \n>> =# select pg_log_query_plan(pid), application_name, backend_type \n>> from\n>> pg_stat_activity where backend_type = 'autovacuum launcher';\n>> WARNING: PID 63323 is not a PostgreSQL client backend process\n>> pg_log_query_plan | application_name | backend_type\n>> -------------------+------------------+---------------------\n>> f | | autovacuum launcher\n>> \n>> \n>> > I am also wondering whether it's better to report the WARNING as\n>> > status column in the output. E.g. instead of\n>> > #select pg_log_query_plan(100);\n>> > WARNING: PID 100 is not a PostgreSQL backend process\n>> > pg_log_query_plan\n>> > -------------------\n>> > f\n>> > (1 row)\n>> > we output\n>> > #select pg_log_query_plan(100);\n>> > pg_log_query_plan | status\n>> > -------------------+---------------------------------------------\n>> > f | PID 100 is not a PostgreSQL backend process\n>> > (1 row)\n>> >\n>> > That looks neater and can easily be handled by scripts, applications\n>> > and such. But it will be inconsistent with other functions like\n>> > pg_terminate_backend() and pg_log_backend_memory_contexts().\n>> \n>> It seems neater, but it might be inconvenient because we can no longer\n>> use it in select list like the following query as you wrote:\n>> \n>> #select pg_log_query_plan(pid), application_name, backend_type from\n>> pg_stat_activity where backend_type = 'autovacuum launcher';\n> \n> Why is that?\n\nSorry, I misunderstood and confirmed we can run queries like below:\n\n```\n=# CREATE OR REPLACE FUNCTION pg_log_query_plan_stab(i int) \n RETURNS TABLE( \n \n pg_log_query_plan bool, \n status text \n \n ) AS $$ \n DECLARE \n \n BEGIN \n \n RETURN QUERY SELECT false::bool, 'PID 100 is not a PostgreSQL \nbackend process'::text; END; \n \n $$ LANGUAGE plpgsql;\n\n=# select pg_log_query_plan_stab(pid), application_name, backend_type \nfrom pg_stat_activity where backend_type = 'autovacuum launcher';\n pg_log_query_plan_stab | application_name | \n backend_type\n---------------------------------------------------+------------------+---------------------\n (f,\"PID 100 is not a PostgreSQL backend process\") | | \nautovacuum launcher\n```\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Wed, 18 Oct 2023 15:09:38 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Wed, Oct 18, 2023 at 11:39 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n> >\n> > I am more interested in avoiding the duplication of code, esp. the\n> > first comment in my reply\n>\n> If there are no objections, I will try porting it to auto_explain and\n> see its feasibility.\n\nIf you want it in core, you will need to port relevant parts of\nauto_explain code to core.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Wed, 18 Oct 2023 15:43:01 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Wed, Oct 18, 2023 at 2:09 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-10-16 18:46, Ashutosh Bapat wrote:\n> > On Thu, Oct 12, 2023 at 6:51 PM torikoshia <torikoshia@oss.nttdata.com>\n> > wrote:\n> >>\n> >> On 2023-10-11 16:22, Ashutosh Bapat wrote:\n> >> >\n> >> > Considering the similarity with auto_explain I wondered whether this\n> >> > function should be part of auto_explain contrib module itself? If we\n> >> > do that users will need to load auto_explain extension and thus\n> >> > install executor hooks when this function doesn't need those. So may\n> >> > not be such a good idea. I didn't see any discussion on this.\n> >>\n> >> I once thought about adding this to auto_explain, but I left it asis\n> >> for\n> >> below reasons:\n> >>\n> >> - One of the typical use case of pg_log_query_plan() would be\n> >> analyzing\n> >> slow query on customer environments. On such environments, We cannot\n> >> always control what extensions to install.\n> >\n> > The same argument applies to auto_explain functionality as well. But\n> > it's not part of the core.\n>\n> Yeah, and when we have a situation where we want to run\n> pg_log_query_plan(), we can run it in any environment as long as it is\n> bundled with the core.\n> On the other hand, if it is built into auto_explain, we need to start by\n> installing auto_explain if we do not have auto_explain, which is often\n> difficult to do in production environments.\n>\n> >> Of course auto_explain is a major extension and it is quite\n> >> possible\n> >> that they installed auto_explain, but but it is also possible they do\n> >> not.\n> >> - It seems a bit counter-intuitive that pg_log_query_plan() is in an\n> >> extension called auto_explain, since it `manually`` logs plans\n> >>\n> >\n> > pg_log_query_plan() may not fit auto_explain but\n> > pg_explain_backend_query() does. What we are logging is more than just\n> > plan of the query, it might expand to be closer to explain output.\n> > While auto in auto_explain would refer to its automatically logging\n> > explain outputs, it can provide an additional function which provides\n> > similar functionality by manually triggering it.\n> >\n> > But we can defer this to a committer, if you want.\n> >\n> > I am more interested in avoiding the duplication of code, esp. the\n> > first comment in my reply\n>\n> If there are no objections, I will try porting it to auto_explain and\n> see its feasibility.\n>\n> >>> There is a lot of similarity between what this feature does and what\n> >>> auto explain does. I see the code is also duplicated. There is some\n> >>> merit in avoiding this duplication\n> >>> 1. we will get all the features of auto_explain automatically like\n> >>> choosing a format (this was expressed somebody earlier in this\n> >>> thread), setings etc.\n> >>> 2. avoid bugs. E.g your code switches context after ExplainState has\n> >>> been allocated. These states may leak depending upon when this\n> >>> function gets called.\n> >>> 3. Building features on top as James envisions will be easier.\n\nIn my view the fact that auto_explain is itself not part of core is a\nmistake, and I know there are more prominent hackers than myself who\nhold that view.\n\nWhile that decision as regards auto_explain has long since been made\n(and there would be work to undo it), I don't believe that we should\nrepeat that choice here. I'm -10 on moving this into auto_explain.\n\nHowever perhaps there is still an opportunity for moving some of the\nauto_explain code into core so as to enable deduplicating the code.\n\nRegards,\nJames Coleman\n\n\n","msg_date":"Wed, 18 Oct 2023 12:34:36 -0400","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"> Hi,\n> \n> Yeah, and when we have a situation where we want to run\n> pg_log_query_plan(), we can run it in any environment as long as it\n> is bundled with the core.\n\nIs it possible to get the plan of a backend programmatically without\naccess to the logs?\n\nSomething like pg_get_query_plan(pid, format) which output would be the\nsame as EXPLAIN.\n\nRegards,\nÉtienne BERSAC\nDalibo\n\n\n","msg_date":"Tue, 24 Oct 2023 09:12:46 +0200","msg_from":"=?ISO-8859-1?Q?=C9tienne?= BERSAC <etienne.bersac@dalibo.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-10-24 16:12, Étienne BERSAC wrote:\n>> Hi,\n>> \n>> Yeah, and when we have a situation where we want to run\n>> pg_log_query_plan(), we can run it in any environment as long as it\n>> is bundled with the core.\n> \n> Is it possible to get the plan of a backend programmatically without\n> access to the logs?\n> \n> Something like pg_get_query_plan(pid, format) which output would be the\n> same as EXPLAIN.\n\nDo you imagine a function like below?\n\n =# select pg_get_query_plan(100, 'plain');\n QUERY PLAN\n -------------------------------------------------------------------\n Limit (cost=0.00..0.04 rows=1 width=273)\n -> Seq Scan on pg_class (cost=0.00..17.14 rows=414 width=273)\n\nIf so, we once tried to implement such function for getting memory \ncontexts.\nHowever, this attempt didn't succeed because it could lead dead lock \nsituation[1].\n\nI think the same problem can occur when implementing \npg_get_query_plan().\n\n[1] \nhttps://www.postgresql.org/message-id/9a50371e15e741e295accabc72a41df1%40oss.nttdata.com\n> \n> Regards,\n> Étienne BERSAC\n> Dalibo\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Tue, 24 Oct 2023 21:30:49 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Wed, Oct 18, 2023 at 10:04 PM James Coleman <jtc331@gmail.com> wrote:\n>\n> While that decision as regards auto_explain has long since been made\n> (and there would be work to undo it), I don't believe that we should\n> repeat that choice here. I'm -10 on moving this into auto_explain.\n>\n\nRight.\n\n> However perhaps there is still an opportunity for moving some of the\n> auto_explain code into core so as to enable deduplicating the code.\n>\n\nRight. That's what I also think.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Wed, 25 Oct 2023 09:10:06 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-10-25 12:40, Ashutosh Bapat wrote:\n> On Wed, Oct 18, 2023 at 10:04 PM James Coleman <jtc331@gmail.com> \n> wrote:\n>> \n>> While that decision as regards auto_explain has long since been made\n>> (and there would be work to undo it), I don't believe that we should\n>> repeat that choice here. I'm -10 on moving this into auto_explain.\n>> \n> \n> Right.\n> \n>> However perhaps there is still an opportunity for moving some of the\n>> auto_explain code into core so as to enable deduplicating the code.\n>> \n> \n> Right. That's what I also think.\n\nThanks for your comments.\n\nAttached patch adds a new function which assembles es->str for logging\naccording to specified contents and format.\nThis is called from both auto_explain and pg_log_query_plan().\n\nOn 2023-10-11 16:22, Ashutosh Bapat wrote:\n> I am also wondering whether it's better to report the WARNING as\n> status column in the output.\n\nAttached patch left as it was since the inconsistency with\npg_terminate_backend() and pg_log_backend_memory_contexts() as you\npointed out.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation","msg_date":"Thu, 26 Oct 2023 15:53:22 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"\nHi Torikoshia,\n\n> If so, we once tried to implement such function for getting memory \n> contexts.\n> However, this attempt didn't succeed because it could lead dead lock \n> situation[1].\n\nThanks for the pointer. Why not use client log message to allow client\nto get plan without locking backend memory context and without access\nto server log ? I missed the rationnal for not sending the plan to\nclient.\n\nRegards,\nÉtienne\n\n\n","msg_date":"Fri, 27 Oct 2023 09:06:31 +0200","msg_from":"=?ISO-8859-1?Q?=C9tienne?= BERSAC <etienne.bersac@dalibo.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-10-27 16:06, Étienne BERSAC wrote:\n> Hi Torikoshia,\n> \n>> If so, we once tried to implement such function for getting memory\n>> contexts.\n>> However, this attempt didn't succeed because it could lead dead lock\n>> situation[1].\n> \n> Thanks for the pointer. Why not use client log message to allow client\n> to get plan without locking backend memory context and without access\n> to server log ? I missed the rationnal for not sending the plan to\n> client.\n\nIf we use client log message, that message is shown on the target \nprocess whose pid is specified by the parameter of pg_log_query_plan():\n\n (pid:1000)=# select pg_sleep(60);\n (pid:1001)=# select pg_log_query_plan(1000);\n (pid:1000)=# LOG: query plan running on backend with PID 1000 is:\n Query Text: select pg_sleep(1000);\n Result (cost=0.00..0.01 rows=1 width=4)\n Output: pg_sleep('1000'::double precision)\n\nI think this is not an expected behavior and we set elevel to \nLOG_SERVER_ONLY.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Fri, 27 Oct 2023 21:30:02 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"> Hi,\n> \n\n> If we use client log message, that message is shown on the target \n> process whose pid is specified by the parameter of\n> pg_log_query_plan():\n> \n> (pid:1000)=# select pg_sleep(60);\n> (pid:1001)=# select pg_log_query_plan(1000);\n> (pid:1000)=# LOG: query plan running on backend with PID 1000 is:\n> Query Text: select pg_sleep(1000);\n> Result (cost=0.00..0.01 rows=1 width=4)\n> Output: pg_sleep('1000'::double precision)\n> \n> I think this is not an expected behavior and we set elevel to \n> LOG_SERVER_ONLY.\n\n\nMakes sens. Thanks.\n\n\n","msg_date":"Fri, 27 Oct 2023 14:54:48 +0200","msg_from":"=?ISO-8859-1?Q?=C9tienne?= BERSAC <etienne.bersac@dalibo.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"I have following questions related to the functionality. (Please point\nme to the relevant discussion if this has been already discussed.)\n\n1. When a backend is running nested queries, we will see the plan of\nthe innermost query. That query may not be the actual culprit if the\nuser query is running slowly. E.g a query being run as part of inner\nside nested loop when the nested loop itself is the bottleneck. I\nthink it will be useful to print plans of all the whole query stack.\n\n2. When a query is running in parallel worker do we want to print that\nquery? It may or may not be interesting given the situation. If the\noverall plan itself is faulty, output of the parallel worker query is\nnot useful. If the plan is fine but a given worker's leg is running\nslowly it may be interesting.\n\nAs a side note, you may want to fix the indentation in\nExplainAssembleLogOutput().\n\nOn Fri, Oct 27, 2023 at 6:24 PM Étienne BERSAC\n<etienne.bersac@dalibo.com> wrote:\n>\n> > Hi,\n> >\n>\n> > If we use client log message, that message is shown on the target\n> > process whose pid is specified by the parameter of\n> > pg_log_query_plan():\n> >\n> > (pid:1000)=# select pg_sleep(60);\n> > (pid:1001)=# select pg_log_query_plan(1000);\n> > (pid:1000)=# LOG: query plan running on backend with PID 1000 is:\n> > Query Text: select pg_sleep(1000);\n> > Result (cost=0.00..0.01 rows=1 width=4)\n> > Output: pg_sleep('1000'::double precision)\n> >\n> > I think this is not an expected behavior and we set elevel to\n> > LOG_SERVER_ONLY.\n>\n>\n> Makes sens. Thanks.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Fri, 3 Nov 2023 19:31:14 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Fri, Nov 3, 2023 at 7:31 PM Ashutosh Bapat\n<ashutosh.bapat.oss@gmail.com> wrote:\n>\n> I have following questions related to the functionality. (Please point\n> me to the relevant discussion if this has been already discussed.)\n>\n> 1. When a backend is running nested queries, we will see the plan of\n> the innermost query. That query may not be the actual culprit if the\n> user query is running slowly. E.g a query being run as part of inner\n> side nested loop when the nested loop itself is the bottleneck. I\n> think it will be useful to print plans of all the whole query stack.\n\nTo further explain this point, consider following scenario\n\n-- A completely useless function which executes two SQL statements\nwhich take small amount individually\n#create function multstmt() returns int\nlanguage sql as $$\nselect count(*) from pg_class where reltype = 12345;\nselect count(*) from pg_type limit 10;\n$$;\n\n-- force a suboptimal plan\n#set enable_hashjoin to false;\n#set enable_mergejoin to false;\n\n-- A completely useless but long running query\n#select c.oid, t.oid from pg_class c, pg_type t where multstmt(c.oid)\n= multstmt(t.oid);\n\nThis take a few minutes on my laptop.\n\nIn a separate session query pg_stat_activity. We will see the original query\n#select pid, query, backend_type from pg_stat_activity where pid = 349330;\n pid | query\n | backend_type\n--------+-----------------------------------------------------------------------------------------+----------------\n 349330 | select c.oid, t.oid from pg_class c, pg_type t where\nmultstmt(c.oid) = multstmt(t.oid); | client backend\n(1 row)\n\nRun the plan output function a few times\n#select pg_log_query_plan(349330);\n\nYou will observe different plans based on which of the innermost query\nis runnning\nLOG: query plan running on backend with PID 349330 is:\n Query Text:\n select count(*) from pg_class where reltype = typeid;\n select count(*) from pg_type where oid = typeid;\n\n Query Parameters: $1 = '600'\n Aggregate (cost=18.18..18.19 rows=1 width=8)\n Output: count(*)\n -> Seq Scan on pg_catalog.pg_class (cost=0.00..18.18 rows=2 width=0)\n Output: oid, relname, relnamespace, reltype,\nreloftype, relowner, relam, relfilenode, reltablespace, relpages,\nreltuples, relallvisible, reltoastrelid, relhasindex, relisshared,\nrelpersistence, relkind, relnatts, relchecks, relhasrules,\nrelhastriggers, relhassubclass, relrowsecurity, relforcerowsecurity,\nrelispopulated, relreplident, relispartition, relrewrite,\nrelfrozenxid, relminmxid, relacl, reloptions, relpartbound\n Filter: (pg_class.reltype = $1)\n Settings: enable_hashjoin = 'off', enable_mergejoin = 'off'\n2023-11-06 11:52:25.610 IST [349330] LOG: query plan running on\nbackend with PID 349330 is:\n Query Text:\n select count(*) from pg_class where reltype = typeid;\n select count(*) from pg_type where oid = typeid;\n\n Query Parameters: $1 = '2203'\n Aggregate (cost=4.30..4.31 rows=1 width=4)\n Output: count(*)\n -> Index Only Scan using pg_type_oid_index on\npg_catalog.pg_type (cost=0.28..4.29 rows=1 width=0)\n Output: oid\n Index Cond: (pg_type.oid = $1)\n Settings: enable_hashjoin = 'off', enable_mergejoin = 'off'\n\nIndividual pieces are confusing here since the query run by the\nbackend is not the one being EXPLAINed. A user may not know that the\nqueries being EXPLAINed arise from the function call multstmt(). So\nthey won't be able to stitch the pieces together unless they see plan\nof the outermost query with loops and costs. What might help if we\nexplain each query in the hierarchy.\n\nI think we can start with what auto_explain is doing. Always print the\nplan of the outermost query; the query found in pg_stat_activity. In a\nlater version we might find a way to print plans of all the queries in\nthe stack and do so in a readable manner.\n\nThis makes tracking activeQueryDesc a bit tricky. My guess is that the\noutermost query's descriptor will be available through ActivePortal\nmost of the time. But there are cases when ExecutorRun is called by\npassing a queryDesc directly. So may be there are some cases where\nthat's not true.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Mon, 6 Nov 2023 12:02:12 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-11-06 15:32, Ashutosh Bapat wrote:\n\nThanks for the suggestion and example.\n\n> On Fri, Nov 3, 2023 at 7:31 PM Ashutosh Bapat\n> <ashutosh.bapat.oss@gmail.com> wrote:\n>> \n>> I have following questions related to the functionality. (Please point\n>> me to the relevant discussion if this has been already discussed.)\n>> \n>> 1. When a backend is running nested queries, we will see the plan of\n>> the innermost query. That query may not be the actual culprit if the\n>> user query is running slowly. E.g a query being run as part of inner\n>> side nested loop when the nested loop itself is the bottleneck. I\n>> think it will be useful to print plans of all the whole query stack.\n\nThis was discussed in previous threads[1] and we thought it'd be useful \nbut since it needed some extra works, we stopped widening the scope.\n\n> To further explain this point, consider following scenario\n> \n> -- A completely useless function which executes two SQL statements\n> which take small amount individually\n> #create function multstmt() returns int\n> language sql as $$\n> select count(*) from pg_class where reltype = 12345;\n> select count(*) from pg_type limit 10;\n> $$;\n> \n> -- force a suboptimal plan\n> #set enable_hashjoin to false;\n> #set enable_mergejoin to false;\n> \n> -- A completely useless but long running query\n> #select c.oid, t.oid from pg_class c, pg_type t where multstmt(c.oid)\n> = multstmt(t.oid);\n> \n> This take a few minutes on my laptop.\n> \n> In a separate session query pg_stat_activity. We will see the original \n> query\n> #select pid, query, backend_type from pg_stat_activity where pid = \n> 349330;\n> pid | query\n> | backend_type\n> --------+-----------------------------------------------------------------------------------------+----------------\n> 349330 | select c.oid, t.oid from pg_class c, pg_type t where\n> multstmt(c.oid) = multstmt(t.oid); | client backend\n> (1 row)\n> \n> Run the plan output function a few times\n> #select pg_log_query_plan(349330);\n> \n> You will observe different plans based on which of the innermost query\n> is runnning\n> LOG: query plan running on backend with PID 349330 is:\n> Query Text:\n> select count(*) from pg_class where reltype = typeid;\n> select count(*) from pg_type where oid = typeid;\n> \n> Query Parameters: $1 = '600'\n> Aggregate (cost=18.18..18.19 rows=1 width=8)\n> Output: count(*)\n> -> Seq Scan on pg_catalog.pg_class (cost=0.00..18.18 rows=2 \n> width=0)\n> Output: oid, relname, relnamespace, reltype,\n> reloftype, relowner, relam, relfilenode, reltablespace, relpages,\n> reltuples, relallvisible, reltoastrelid, relhasindex, relisshared,\n> relpersistence, relkind, relnatts, relchecks, relhasrules,\n> relhastriggers, relhassubclass, relrowsecurity, relforcerowsecurity,\n> relispopulated, relreplident, relispartition, relrewrite,\n> relfrozenxid, relminmxid, relacl, reloptions, relpartbound\n> Filter: (pg_class.reltype = $1)\n> Settings: enable_hashjoin = 'off', enable_mergejoin = 'off'\n> 2023-11-06 11:52:25.610 IST [349330] LOG: query plan running on\n> backend with PID 349330 is:\n> Query Text:\n> select count(*) from pg_class where reltype = typeid;\n> select count(*) from pg_type where oid = typeid;\n> \n> Query Parameters: $1 = '2203'\n> Aggregate (cost=4.30..4.31 rows=1 width=4)\n> Output: count(*)\n> -> Index Only Scan using pg_type_oid_index on\n> pg_catalog.pg_type (cost=0.28..4.29 rows=1 width=0)\n> Output: oid\n> Index Cond: (pg_type.oid = $1)\n> Settings: enable_hashjoin = 'off', enable_mergejoin = 'off'\n> \n> Individual pieces are confusing here since the query run by the\n> backend is not the one being EXPLAINed. A user may not know that the\n> queries being EXPLAINed arise from the function call multstmt(). So\n> they won't be able to stitch the pieces together unless they see plan\n> of the outermost query with loops and costs. What might help if we\n> explain each query in the hierarchy.\n> \n> I think we can start with what auto_explain is doing. Always print the\n> plan of the outermost query; the query found in pg_stat_activity. In a\n> later version we might find a way to print plans of all the queries in\n> the stack and do so in a readable manner.\n\nAgreed there are cases printing plan of the outermost query is more \nuseful.\n\n> \n> This makes tracking activeQueryDesc a bit tricky. My guess is that the\n> outermost query's descriptor will be available through ActivePortal\n> most of the time. But there are cases when ExecutorRun is called by\n> passing a queryDesc directly. So may be there are some cases where\n> that's not true.\n\nYeah, actually the original version of the patch got the plan from \nActivePortal, but it failed logging plan when the query was something \nlike this[2]:\n\n DO $$\n BEGIN\n PERFORM pg_sleep(100);\n END$$;\n\n> 2. When a query is running in parallel worker do we want to print that\n> query? It may or may not be interesting given the situation. If the\n> overall plan itself is faulty, output of the parallel worker query is\n> not useful. If the plan is fine but a given worker's leg is running\n> slowly it may be interesting.\n\nI think it can be useful.\nI'm wondering if we can add this after the first patch for this feature \nis committed.\n\n> As a side note, you may want to fix the indentation in\n> ExplainAssembleLogOutput().\n\nThanks, modified it.\nSince the documentation was obsoleted, attached patch also updated it.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation","msg_date":"Thu, 09 Nov 2023 15:33:36 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Thu, Nov 9, 2023 at 12:03 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\n> >>\n> >> 1. When a backend is running nested queries, we will see the plan of\n> >> the innermost query. That query may not be the actual culprit if the\n> >> user query is running slowly. E.g a query being run as part of inner\n> >> side nested loop when the nested loop itself is the bottleneck. I\n> >> think it will be useful to print plans of all the whole query stack.\n>\n> This was discussed in previous threads[1] and we thought it'd be useful\n> but since it needed some extra works, we stopped widening the scope.\n>\n> >\n> > I think we can start with what auto_explain is doing. Always print the\n> > plan of the outermost query; the query found in pg_stat_activity. In a\n> > later version we might find a way to print plans of all the queries in\n> > the stack and do so in a readable manner.\n>\n> Agreed there are cases printing plan of the outermost query is more\n> useful.\n>\n\nI am fine printing the plan of the outermost query. This will help\nmany cases. Printing plans of the whole query stack can be added as an\nadd on later.\n\n> >\n> > This makes tracking activeQueryDesc a bit tricky. My guess is that the\n> > outermost query's descriptor will be available through ActivePortal\n> > most of the time. But there are cases when ExecutorRun is called by\n> > passing a queryDesc directly. So may be there are some cases where\n> > that's not true.\n>\n> Yeah, actually the original version of the patch got the plan from\n> ActivePortal, but it failed logging plan when the query was something\n> like this[2]:\n>\n> DO $$\n> BEGIN\n> PERFORM pg_sleep(100);\n> END$$;\n\nReferences [1] and [2] are not listed in your email.\n\nIs that because there was no ActivePortal created or the ActivePortal\npointed to DO block instead of PERFORM pg_sleep?\n\n>\n> > 2. When a query is running in parallel worker do we want to print that\n> > query? It may or may not be interesting given the situation. If the\n> > overall plan itself is faulty, output of the parallel worker query is\n> > not useful. If the plan is fine but a given worker's leg is running\n> > slowly it may be interesting.\n>\n> I think it can be useful.\n> I'm wondering if we can add this after the first patch for this feature\n> is committed.\n\nWith the current patches, it will print the query from a parallel\nbackend. If that's not desirable we should prohibit that case at\nleast.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Thu, 9 Nov 2023 12:41:49 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-11-09 16:11, Ashutosh Bapat wrote:\n> On Thu, Nov 9, 2023 at 12:03 PM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> >>\n>> >> 1. When a backend is running nested queries, we will see the plan of\n>> >> the innermost query. That query may not be the actual culprit if the\n>> >> user query is running slowly. E.g a query being run as part of inner\n>> >> side nested loop when the nested loop itself is the bottleneck. I\n>> >> think it will be useful to print plans of all the whole query stack.\n>> \n>> This was discussed in previous threads[1] and we thought it'd be \n>> useful\n>> but since it needed some extra works, we stopped widening the scope.\n>> \n>> >\n>> > I think we can start with what auto_explain is doing. Always print the\n>> > plan of the outermost query; the query found in pg_stat_activity. In a\n>> > later version we might find a way to print plans of all the queries in\n>> > the stack and do so in a readable manner.\n>> \n>> Agreed there are cases printing plan of the outermost query is more\n>> useful.\n>> \n> \n> I am fine printing the plan of the outermost query. This will help\n> many cases. Printing plans of the whole query stack can be added as an\n> add on later.\n> \n>> >\n>> > This makes tracking activeQueryDesc a bit tricky. My guess is that the\n>> > outermost query's descriptor will be available through ActivePortal\n>> > most of the time. But there are cases when ExecutorRun is called by\n>> > passing a queryDesc directly. So may be there are some cases where\n>> > that's not true.\n>> \n>> Yeah, actually the original version of the patch got the plan from\n>> ActivePortal, but it failed logging plan when the query was something\n>> like this[2]:\n>> \n>> DO $$\n>> BEGIN\n>> PERFORM pg_sleep(100);\n>> END$$;\n> \n> References [1] and [2] are not listed in your email.\n\nOops, sorry. Here are links:\n\n[1] \nhttps://www.postgresql.org/message-id/64f716c44629e303b66e6c24502147cc%40oss.nttdata.com\n[2] \nhttps://www.postgresql.org/message-id/632e99eb-8090-53e6-1b1a-101601904cbd%40oss.nttdata.com\n\n> Is that because there was no ActivePortal created or the ActivePortal\n> pointed to DO block instead of PERFORM pg_sleep?\n\nActivePortal is created but ActivePortal->queryDesc is null.\n\n>> > 2. When a query is running in parallel worker do we want to print that\n>> > query? It may or may not be interesting given the situation. If the\n>> > overall plan itself is faulty, output of the parallel worker query is\n>> > not useful. If the plan is fine but a given worker's leg is running\n>> > slowly it may be interesting.\n>> \n>> I think it can be useful.\n>> I'm wondering if we can add this after the first patch for this \n>> feature\n>> is committed.\n> \n> With the current patches, it will print the query from a parallel\n> backend. If that's not desirable we should prohibit that case at\n> least.\n\nCurrent patch prohibits printing plan if backend type is parallel worker \nas below:\n\n =# select pg_log_query_plan(pid), backend_type from pg_stat_activity \nwhere backend_type = 'parallel worker';\n\n pg_log_query_plan | backend_type\n -------------------+-----------------\n f | parallel worker\n f | parallel worker\n (2 rows)\n\n WARNING: PID 4618 is not a PostgreSQL client backend process\n WARNING: PID 4617 is not a PostgreSQL client backend process\n\nIs this the behavior you expect?\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Thu, 09 Nov 2023 20:26:27 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Thu, Nov 9, 2023 at 4:56 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On 2023-11-09 16:11, Ashutosh Bapat wrote:\n> > On Thu, Nov 9, 2023 at 12:03 PM torikoshia <torikoshia@oss.nttdata.com>\n> > wrote:\n> >> >>\n> >> >> 1. When a backend is running nested queries, we will see the plan of\n> >> >> the innermost query. That query may not be the actual culprit if the\n> >> >> user query is running slowly. E.g a query being run as part of inner\n> >> >> side nested loop when the nested loop itself is the bottleneck. I\n> >> >> think it will be useful to print plans of all the whole query stack.\n> >>\n> >> This was discussed in previous threads[1] and we thought it'd be\n> >> useful\n> >> but since it needed some extra works, we stopped widening the scope.\n> >>\n> >> >\n> >> > I think we can start with what auto_explain is doing. Always print the\n> >> > plan of the outermost query; the query found in pg_stat_activity. In a\n> >> > later version we might find a way to print plans of all the queries in\n> >> > the stack and do so in a readable manner.\n> >>\n> >> Agreed there are cases printing plan of the outermost query is more\n> >> useful.\n> >>\n> >\n> > I am fine printing the plan of the outermost query. This will help\n> > many cases. Printing plans of the whole query stack can be added as an\n> > add on later.\n> >\n> >> >\n> >> > This makes tracking activeQueryDesc a bit tricky. My guess is that the\n> >> > outermost query's descriptor will be available through ActivePortal\n> >> > most of the time. But there are cases when ExecutorRun is called by\n> >> > passing a queryDesc directly. So may be there are some cases where\n> >> > that's not true.\n> >>\n> >> Yeah, actually the original version of the patch got the plan from\n> >> ActivePortal, but it failed logging plan when the query was something\n> >> like this[2]:\n> >>\n> >> DO $$\n> >> BEGIN\n> >> PERFORM pg_sleep(100);\n> >> END$$;\n> >\n> > References [1] and [2] are not listed in your email.\n>\n> Oops, sorry. Here are links:\n>\n> [1]\n> https://www.postgresql.org/message-id/64f716c44629e303b66e6c24502147cc%40oss.nttdata.com\n> [2]\n> https://www.postgresql.org/message-id/632e99eb-8090-53e6-1b1a-101601904cbd%40oss.nttdata.com\n>\n> > Is that because there was no ActivePortal created or the ActivePortal\n> > pointed to DO block instead of PERFORM pg_sleep?\n>\n> ActivePortal is created but ActivePortal->queryDesc is null.\n\nThanks.\n\n>\n> >> > 2. When a query is running in parallel worker do we want to print that\n> >> > query? It may or may not be interesting given the situation. If the\n> >> > overall plan itself is faulty, output of the parallel worker query is\n> >> > not useful. If the plan is fine but a given worker's leg is running\n> >> > slowly it may be interesting.\n> >>\n> >> I think it can be useful.\n> >> I'm wondering if we can add this after the first patch for this\n> >> feature\n> >> is committed.\n> >\n> > With the current patches, it will print the query from a parallel\n> > backend. If that's not desirable we should prohibit that case at\n> > least.\n>\n> Current patch prohibits printing plan if backend type is parallel worker\n> as below:\n>\n> =# select pg_log_query_plan(pid), backend_type from pg_stat_activity\n> where backend_type = 'parallel worker';\n>\n> pg_log_query_plan | backend_type\n> -------------------+-----------------\n> f | parallel worker\n> f | parallel worker\n> (2 rows)\n>\n> WARNING: PID 4618 is not a PostgreSQL client backend process\n> WARNING: PID 4617 is not a PostgreSQL client backend process\n>\n> Is this the behavior you expect?\n>\n\nI misread then. Thanks for correcting me. We could consider plans from\nparallel workers in v2 of this feature.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Thu, 9 Nov 2023 20:28:48 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hello hackers,\n\nLast Saturday I submitted a patch to the pgsql-hackers list with the title\n\"Proposal: In-flight explain logging\" with a patch proposing a feature very\nsimilar to the one being worked on in this thread. I should have done a\nbetter\nsearch in the commitfest before implementing something from scratch.\n\nSo, as recommended by Ashutosh, I am sending an incremental patch containing\nan additional feature I personally think we should include: logging the plan\nwith instrumentation details if enabled.\n\nWhen targeting a query with instrumentation PG should log the complete\nEXPLAIN ANALYZE plan with current row count and, if enabled, timing for each\nnode. This gives the user not only the ability to see what the plan is\nbut also what was executed so far, which is super useful when\ntroubleshooting queries that never finish.\n\nConsidering that the query is in progress the output will include the\nstatement (never executed) for nodes that weren't touched yet (or may\nnever be). This feature is already present in the current ExplainNode\nimplementation.\n\nI added a new statement (in progress) for nodes currently being executed,\nie,\nInstrStartNode was called and clock is ticking there.\n\nBack-reading this thread I saw the discussion about extending\npg_log_query_plan\nto parallel workers or not. I added it in my incremental patch as the new\ncapability of logging instrumentation makes it useful to be able to see\nwhat parallel workers are currently doing.\n\n# DEMONSTRATION:\n\npostgres=# select pid, backend_type,pg_log_query_plan(pid)\npostgres=# from pg_stat_activity\npostgres=# where (backend_type = 'client backend' and pid !=\npg_backend_pid())\npostgres=# or backend_type = 'parallel worker';\n pid | backend_type | pg_log_query_plan\n-------+-----------------+-------------------\n 33833 | client backend | t\n 47202 | parallel worker | t\n 47203 | parallel worker | t\n(3 rows)\n\n2023-12-06 23:14:41.756 UTC [33833] LOG: query plan running on backend\nwith PID 33833 is:\nQuery Text: explain (analyze, buffers) select *\nfrom t2 a\ninner join t1 b on a.c1=b.c1\ninner join t1 c on a.c1=c.c1\ninner join t1 d on a.c1=d.c1\ninner join t1 e on a.c1=e.c1;\nGather (cost=70894.63..202643.27 rows=1000000 width=20) (never executed)\n Output: a.c1, b.c1, c.c1, d.c1, e.c1\n Workers Planned: 2\n Workers Launched: 2\n -> Parallel Hash Join (cost=69894.63..101643.27 rows=416667 width=20)\n(never executed)\n Output: a.c1, b.c1, c.c1, d.c1, e.c1\n Hash Cond: (a.c1 = e.c1)\n -> Parallel Hash Join (cost=54466.62..77218.65 rows=416667\nwidth=16) (never executed)\n Output: a.c1, b.c1, c.c1, d.c1\n Hash Cond: (a.c1 = c.c1)\n -> Parallel Hash Join (cost=15428.00..29997.42 rows=416667\nwidth=8) (actual time=2500.914..2625.922 rows=250412 loops=1) (in progress)\n Output: a.c1, b.c1\n Inner Unique: true\n Hash Cond: (b.c1 = a.c1)\n Buffers: shared hit=755 read=2175, temp read=1473\nwritten=1860\n -> Parallel Seq Scan on public.t1 b\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.022..20.904\nrows=331492 loops=1)\n Output: b.c1\n Buffers: shared hit=602 read=865\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667\nwidth=4) (actual time=1745.107..1745.107 rows=330638 loops=1)\n Output: a.c1\n Buffers: shared hit=153 read=1310, temp written=868\n -> Parallel Seq Scan on public.t2 a\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.042..27.695\nrows=330638 loops=1)\n Output: a.c1\n Buffers: shared hit=153 read=1310\n -> Parallel Hash (cost=32202.28..32202.28 rows=416667\nwidth=8) (actual time=2450.489..2450.489 rows=407941 loops=1)\n Output: c.c1, d.c1\n Buffers: shared hit=1141 read=1833, temp read=1938\nwritten=2836\n -> Parallel Hash Join (cost=15428.00..32202.28\nrows=416667 width=8) (actual time=1323.422..1575.245 rows=407941 loops=1)\n Output: c.c1, d.c1\n Hash Cond: (c.c1 = d.c1)\n Buffers: shared hit=1141 read=1833, temp read=1938\nwritten=1864\n -> Parallel Seq Scan on public.t1 c\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.026..22.223\nrows=336238 loops=1)\n Output: c.c1\n Buffers: shared hit=590 read=898\n -> Parallel Hash (cost=8591.67..8591.67\nrows=416667 width=4) (actual time=653.306..653.306 rows=335836 loops=1)\n Output: d.c1\n Buffers: shared hit=551 read=935, temp\nwritten=872\n -> Parallel Seq Scan on public.t1 d\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.022..23.127\nrows=335836 loops=1)\n Output: d.c1\n Buffers: shared hit=551 read=935\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4)\n(actual time=590.086..590.086 rows=343696 loops=1)\n Output: e.c1\n Buffers: shared hit=519 read=1002, temp written=896\n -> Parallel Seq Scan on public.t1 e (cost=0.00..8591.67\nrows=416667 width=4) (actual time=0.066..21.797 rows=343696 loops=1)\n Output: e.c1\n Buffers: shared hit=519 read=1002\n2023-12-06 23:14:41.757 UTC [47203] LOG: query plan running on backend\nwith PID 47203 is:\nQuery Text: explain (analyze, buffers) select *\nfrom t2 a\ninner join t1 b on a.c1=b.c1\ninner join t1 c on a.c1=c.c1\ninner join t1 d on a.c1=d.c1\ninner join t1 e on a.c1=e.c1;\nParallel Hash Join (cost=69894.63..101643.27 rows=416667 width=20) (never\nexecuted)\n Output: a.c1, b.c1, c.c1, d.c1, e.c1\n Hash Cond: (a.c1 = e.c1)\n -> Parallel Hash Join (cost=54466.62..77218.65 rows=416667 width=16)\n(never executed)\n Output: a.c1, b.c1, c.c1, d.c1\n Hash Cond: (a.c1 = c.c1)\n -> Parallel Hash Join (cost=15428.00..29997.42 rows=416667\nwidth=8) (actual time=2464.367..2628.476 rows=258293 loops=1)\n Output: a.c1, b.c1\n Inner Unique: true\n Hash Cond: (b.c1 = a.c1)\n Buffers: shared hit=704 read=2213, temp read=1497 written=1860\n -> Parallel Seq Scan on public.t1 b (cost=0.00..8591.67\nrows=416667 width=4) (actual time=0.032..22.247 rows=330412 loops=1)\n Output: b.c1\n Buffers: shared hit=594 read=868\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4)\n(actual time=1745.093..1745.093 rows=328830 loops=1)\n Output: a.c1\n Buffers: shared hit=110 read=1345, temp written=868\n -> Parallel Seq Scan on public.t2 a\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=1.141..30.128\nrows=328830 loops=1)\n Output: a.c1\n Buffers: shared hit=110 read=1345\n -> Parallel Hash (cost=32202.28..32202.28 rows=416667 width=8)\n(actual time=2449.694..2449.694 rows=295462 loops=1)\n Output: c.c1, d.c1\n Buffers: shared hit=1161 read=1844, temp read=1971 written=2872\n -> Parallel Hash Join (cost=15428.00..32202.28 rows=416667\nwidth=8) (actual time=1378.678..1577.182 rows=295462 loops=1)\n Output: c.c1, d.c1\n Hash Cond: (c.c1 = d.c1)\n Buffers: shared hit=1161 read=1844, temp read=1971\nwritten=1856\n -> Parallel Seq Scan on public.t1 c\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.031..22.459\nrows=336288 loops=1)\n Output: c.c1\n Buffers: shared hit=590 read=898\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667\nwidth=4) (actual time=651.048..651.048 rows=328378 loops=1)\n Output: d.c1\n Buffers: shared hit=507 read=946, temp written=860\n -> Parallel Seq Scan on public.t1 d\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.034..24.462\nrows=328378 loops=1)\n Output: d.c1\n Buffers: shared hit=507 read=946\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4) (actual\ntime=575.417..575.417 rows=327926 loops=1)\n Output: e.c1\n Buffers: shared hit=522 read=929, temp written=860\n -> Parallel Seq Scan on public.t1 e (cost=0.00..8591.67\nrows=416667 width=4) (actual time=0.786..23.114 rows=327926 loops=1)\n Output: e.c1\n Buffers: shared hit=522 read=929\n2023-12-06 23:14:41.758 UTC [47202] LOG: query plan running on backend\nwith PID 47202 is:\nQuery Text: explain (analyze, buffers) select *\nfrom t2 a\ninner join t1 b on a.c1=b.c1\ninner join t1 c on a.c1=c.c1\ninner join t1 d on a.c1=d.c1\ninner join t1 e on a.c1=e.c1;\nParallel Hash Join (cost=69894.63..101643.27 rows=416667 width=20) (never\nexecuted)\n Output: a.c1, b.c1, c.c1, d.c1, e.c1\n Hash Cond: (a.c1 = e.c1)\n -> Parallel Hash Join (cost=54466.62..77218.65 rows=416667 width=16)\n(never executed)\n Output: a.c1, b.c1, c.c1, d.c1\n Hash Cond: (a.c1 = c.c1)\n -> Parallel Hash Join (cost=15428.00..29997.42 rows=416667\nwidth=8) (actual time=2500.138..2631.855 rows=254125 loops=1)\n Output: a.c1, b.c1\n Inner Unique: true\n Hash Cond: (b.c1 = a.c1)\n Buffers: shared hit=749 read=2254, temp read=1483 written=1896\n -> Parallel Seq Scan on public.t1 b (cost=0.00..8591.67\nrows=416667 width=4) (actual time=0.027..22.224 rows=338096 loops=1)\n Output: b.c1\n Buffers: shared hit=628 read=868\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4)\n(actual time=1744.616..1744.616 rows=340532 loops=1)\n Output: a.c1\n Buffers: shared hit=121 read=1386, temp written=892\n -> Parallel Seq Scan on public.t2 a\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=1.138..30.003\nrows=340532 loops=1)\n Output: a.c1\n Buffers: shared hit=121 read=1386\n -> Parallel Hash (cost=32202.28..32202.28 rows=416667 width=8)\n(actual time=2450.470..2450.470 rows=296597 loops=1)\n Output: c.c1, d.c1\n Buffers: shared hit=1186 read=1813, temp read=1610 written=2872\n -> Parallel Hash Join (cost=15428.00..32202.28 rows=416667\nwidth=8) (actual time=1380.286..1561.412 rows=296597 loops=1)\n Output: c.c1, d.c1\n Hash Cond: (c.c1 = d.c1)\n Buffers: shared hit=1186 read=1813, temp read=1610\nwritten=1848\n -> Parallel Seq Scan on public.t1 c\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.032..22.170\nrows=327474 loops=1)\n Output: c.c1\n Buffers: shared hit=548 read=901\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667\nwidth=4) (actual time=653.297..653.297 rows=335786 loops=1)\n Output: d.c1\n Buffers: shared hit=574 read=912, temp written=868\n -> Parallel Seq Scan on public.t1 d\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.023..24.210\nrows=335786 loops=1)\n Output: d.c1\n Buffers: shared hit=574 read=912\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4) (actual\ntime=574.553..574.553 rows=328378 loops=1)\n Output: e.c1\n Buffers: shared hit=495 read=958, temp written=864\n -> Parallel Seq Scan on public.t1 e (cost=0.00..8591.67\nrows=416667 width=4) (actual time=0.717..23.270 rows=328378 loops=1)\n Output: e.c1\n Buffers: shared hit=495 read=958\n\nThe thread of my initial patch contains all important implementation\ndetails. Sharing the ones relevant to my incremental patch here:\n\n- Safely printing signaled plans with instrumentation\n\nA plan string is built in function ExplainNode here (\nhttps://github.com/postgres/postgres/blob/REL_16_STABLE/src/backend/commands/explain.c#L1178\n)\nwhich is called at the end of a query execution when EXPLAIN is used.\nThat function performs logic using a PlanState (part of QueryDesc) of\nthe running query and a ExplainState.\n\nThe main challenge there is that ExplainNode calls InstrEndLoop which\nchanges values in Instrumentation. This is ok for a regular EXPLAIN\nwhere the query is already complete but not ok for the new feature with\nsignaled plan logging.\n\nSo the new code has custom logic to clone Instrumentation instance of\nthe current node. The cloned object can be safely written.\n\nFunction InstrEndLoop has a safety rule here (\nhttps://github.com/postgres/postgres/blob/REL_16_STABLE/src/backend/executor/instrument.c#L148\n)\nthat prevents adjusting instrumentation details in a running node. This\nnever happens in the current code logic but with the new feature\nit will happen very often.\n\nI didn't want to remove this safety rule as InstrEndLoop gets called in\nother places too (even in auto_explain) so the solution was to keep\nInstrEndLoop and have a new InstrEndLoopForce for the signaled\nplan logging with instrumentation. Both InstrEndLoop and InstrEndLoopForce\ncall a new internal InstrEndLoopInternal to avoid duplicating the code.\n\n- Memory management\n\nConsidering that pg_log_query_plan() creates its own memory context before\ncalling ExplainAssembleLogOutput, the new logic that allocates memory\nto clone instrumentation doesn't need to free anything.\n\n# FINAL CONSIDERATIONS\n\nLet me know what you think about this incremental patch. If you think it\nis useful but needs adjustments I will be happy to change as needed.\n\nKind Regards,\n\nRafael Castro.","msg_date":"Wed, 6 Dec 2023 20:33:56 -0300","msg_from":"Rafael Thofehrn Castro <rafaelthca@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2023-12-07 08:33, Rafael Thofehrn Castro wrote:\n> Hello hackers,\n> \n> Last Saturday I submitted a patch to the pgsql-hackers list with the\n> title\n> \"Proposal: In-flight explain logging\" with a patch proposing a feature\n> very\n> similar to the one being worked on in this thread. I should have done\n> a better\n> search in the commitfest before implementing something from scratch.\n> \n> So, as recommended by Ashutosh, I am sending an incremental patch\n> containing\n> an additional feature I personally think we should include: logging\n> the plan\n> with instrumentation details if enabled.\n\nThanks for the proposal and making the patch!\n\n> When targeting a query with instrumentation PG should log the complete\n> EXPLAIN ANALYZE plan with current row count and, if enabled, timing\n> for each\n> node. This gives the user not only the ability to see what the plan is\n> but also what was executed so far, which is super useful when\n> troubleshooting queries that never finish.\n\n\nI think showing the progress of the query execution would be useful.\n\nOTOH it seems to at least need some modifications around Instrumentation \nas your patch.\nAs a first step, I think it would better to minimize the scope and focus \non the fundamental function.\nFor the same reason, getting queries for parallel workers is also \nprohibited in the current patch as discussed here[1].\n\n[1] \nhttps://www.postgresql.org/message-id/c25ae6015be96a1964eddd964657660b%40oss.nttdata.com\n\nSo I think below steps would be better than pushing all the \nfunctionalities to the 1st commit.\n\n- First, develop function to enable output of query \nprogress(v34-0001-Add-function-to-log-the-plan-of-the-query.patch).\n- Then enhance the function\n - showing the progress of the query \nexecution(v34-0002-Log-plan-along-with-instrumentation-details.patch), \netc.\n\n> --https://www.postgresql.org/message-id/CAG0ozMp3g3drnkDa6RZxXO_OmnisL2sD9vBrmpu5fOBoYpC-3w%40mail.gmail.com\n> - ExplainState customization\n> \n> A ExplainState is allocated and customized for the in-flight logging.\n> Instrumentation related settings are enabled based on how the target\n> query started, which is usually via EXPLAIN ANALYZE or with \n> auto_explain.\n\nDoes this mean the progress can be got only when the target query was \nrun with EXPLAIN ANALYZE or auto_explain.log_analyze?\n\nIf so, there might be limited situations we can get the progress since I \nimagine EXPLAIN ANALYZE is used when user want to get the plan from the \nbeginning and auto_explain.log_analyze can give negative impact on \nperformance as described in the manual and there may not be many \nenvironments which enable it.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Mon, 11 Dec 2023 14:46:23 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi,\n\nUpdated the patch to fix typos and move \nProcessLogQueryPlanInterruptActive from errfinish() to AbortTransaction.\n\n\nBTW since the thread is getting long, I list the some points of the \ndiscussion so far:\n\n# Safety concern\n## Catalog access inside CFI\n- it seems safe if the CFI call is inside an existing valid \ntransaction/query state[1]\n\n- We did some tests, for example calling ProcessLogQueryPlanInterrupt() \nin every single CHECK_FOR_INTERRUPTS()[2]. This test passed on my env \nbut was stucked on James's env, so I modified to exit \nProcessLogQueryPlanInterrupt() when target process is inside of lock \nacquisition code[3]\n\n## Risk of calling EXPLAIN code in CFI\n- EXPLAIN is not a simple logic code, and there exists risk calling it \nfrom CFI. For example, if there is a bug, we may find ourselves in a \nsituation where we can't cancel the query\n\n- it's a trade-off that's worth making for the introspection benefits \nthis patch would provide?[4]\n\n# Design\n- Although some suggested it should be in auto_explain, current patch \nintroduces this feature to the core[5]\n\n- When the target query is nested, only the most inner query's plan is \nexplained. In future, all the nested queries' plans are expected to \nexplained optionally like auto_explain.log_nested_statements[6]\n\n- When the target process is a parallel worker, the plan is not shown[6]\n\n- When the target query is nested and its subtransaction is aborted, \npg_log_query_plan cannot log the parental query plan after the abort \neven parental query is running[7]\n\n- The output corresponds to EXPLAIN with VERBOSE, COST, SETTINGS and \nFORMAT text. It doesn't do ANALYZE or show the progress of the query \nexecution. Future work proposed by Rafael Thofehrn Castro may realize \nthis[8]\n\n- To prevent assertion error, this patch ensures no page lock is held by \nchecking all the LocalLock entries before running explain code, but \nthere is a discussion that ginInsertCleanup() should be modified[9]\n\n\nIt may be not so difficult to improve some of restrictions in \"Design\", \nbut I'd like to limit the scope of the 1st patch to make it simpler.\n\n\n[1] \nhttps://www.postgresql.org/message-id/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw%40mail.gmail.com\n[2] \nhttps://www.postgresql.org/message-id/CAAaqYe8LXVXQhYy3yT0QOHUymdM%3Duha0dJ0%3DBEPzVAx2nG1gsw%40mail.gmail.com\n[3] \nhttps://www.postgresql.org/message-id/0e0e7ca08dff077a625c27a5e0c2ef0a%40oss.nttdata.com\n[4] \nhttps://www.postgresql.org/message-id/CAAaqYe8LXVXQhYy3yT0QOHUymdM%3Duha0dJ0%3DBEPzVAx2nG1gsw%40mail.gmail.com\n[5] \nhttps://www.postgresql.org/message-id/CAAaqYe_1EuoTudAz8mr8-qtN5SoNtYRm4JM2J9CqeverpE3B2A%40mail.gmail.com\n[6] \nhttps://www.postgresql.org/message-id/CAExHW5sh4ahrJgmMAGfptWVmESt1JLKCNm148XVxTunRr%2B-6gA%40mail.gmail.com\n[7] \nhttps://www.postgresql.org/message-id/3d121ed5f81cef588bac836b43f5d1f9%40oss.nttdata.com\n[8] \nhttps://www.postgresql.org/message-id/c161b5e7e1888eb9c9eb182a7d9dcf89%40oss.nttdata.com\n[9] \nhttps://www.postgresql.org/message-id/20220201.172757.1480996662235658750.horikyota.ntt%40gmail.com\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation","msg_date":"Mon, 29 Jan 2024 22:02:29 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi Atsushi,\n\n\nOn Mon, Jan 29, 2024 at 6:32 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> Hi,\n>\n> Updated the patch to fix typos and move\n> ProcessLogQueryPlanInterruptActive from errfinish() to AbortTransaction.\n>\n>\n> BTW since the thread is getting long, I list the some points of the\n> discussion so far:\n>\n> # Safety concern\n> ## Catalog access inside CFI\n> - it seems safe if the CFI call is inside an existing valid\n> transaction/query state[1]\n>\n> - We did some tests, for example calling ProcessLogQueryPlanInterrupt()\n> in every single CHECK_FOR_INTERRUPTS()[2]. This test passed on my env\n> but was stucked on James's env, so I modified to exit\n> ProcessLogQueryPlanInterrupt() when target process is inside of lock\n> acquisition code[3]\n>\n> ## Risk of calling EXPLAIN code in CFI\n> - EXPLAIN is not a simple logic code, and there exists risk calling it\n> from CFI. For example, if there is a bug, we may find ourselves in a\n> situation where we can't cancel the query\n>\n> - it's a trade-off that's worth making for the introspection benefits\n> this patch would provide?[4]\n>\n> # Design\n> - Although some suggested it should be in auto_explain, current patch\n> introduces this feature to the core[5]\n>\n> - When the target query is nested, only the most inner query's plan is\n> explained. In future, all the nested queries' plans are expected to\n> explained optionally like auto_explain.log_nested_statements[6]\n>\n> - When the target process is a parallel worker, the plan is not shown[6]\n>\n> - When the target query is nested and its subtransaction is aborted,\n> pg_log_query_plan cannot log the parental query plan after the abort\n> even parental query is running[7]\n>\n> - The output corresponds to EXPLAIN with VERBOSE, COST, SETTINGS and\n> FORMAT text. It doesn't do ANALYZE or show the progress of the query\n> execution. Future work proposed by Rafael Thofehrn Castro may realize\n> this[8]\n>\n> - To prevent assertion error, this patch ensures no page lock is held by\n> checking all the LocalLock entries before running explain code, but\n> there is a discussion that ginInsertCleanup() should be modified[9]\n>\n>\n> It may be not so difficult to improve some of restrictions in \"Design\",\n> but I'd like to limit the scope of the 1st patch to make it simpler.\n\nThanks for the summary. It is helpful. I think patch is also getting better.\n\nI have a few questions and suggestions\n1. Prologue of GetLockMethodLocalHash() mentions\n * NOTE: When there are many entries in LockMethodLocalHash, calling this\n * function and looking into all of them can lead to performance problems.\n */\nHow bad this performance could be. Let's assume that a query is taking\ntime and pg_log_query_plan() is invoked to examine the plan of this\nquery. Is it possible that the looping over all the locks itself takes\na lot of time delaying the query execution further?\n\n2. What happens if auto_explain is enabled in the backend and\npg_log_query_plan() is called on the same backend? Will they conflict?\nI think we should add a test for the same.\n\nUsing injection point support we should be able to add tests for\ntesting pg_log_query_plan behaviour when there are page locks held or\nwhen auto_explain (with instrumentation) and pg_log_query_plan() work\non the same query plan. Use injection point to make the backend\nrunning query wait at a suitable point to delay its execution and fire\npg_log_query_plan() from other backend. May be the same test could\nexamine the server log file to see if the plan is indeed output to the\nserver log file.\n\nGiven that the feature will be used when the things have already gone\nwrong, it should not make things more serious. So more testing and\nespecially automated would help.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Tue, 6 Feb 2024 16:21:58 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi Ashutosh,\n\nOn 2024-02-06 19:51, Ashutosh Bapat wrote:\n\n> Thanks for the summary. It is helpful. I think patch is also getting \n> better.\n> \n> I have a few questions and suggestions\n\nThanks for your comments.\n\n> 1. Prologue of GetLockMethodLocalHash() mentions\n> * NOTE: When there are many entries in LockMethodLocalHash, calling \n> this\n> * function and looking into all of them can lead to performance \n> problems.\n> */\n> How bad this performance could be. Let's assume that a query is taking\n> time and pg_log_query_plan() is invoked to examine the plan of this\n> query. Is it possible that the looping over all the locks itself takes\n> a lot of time delaying the query execution further?\n\nI think it depends on the number of local locks, but I've measured cpu \ntime for this page lock check by adding below codes and \nv27-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch[1], which \ncalls ProcessLogQueryPlanInterrupt() in every CFI on my laptop just for \nyour information:\n\n diff --git a/src/backend/commands/explain.c \nb/src/backend/commands/explain.c\n index 5f7d77d567..65b7cb4925 100644\n --- a/src/backend/commands/explain.c\n +++ b/src/backend/commands/explain.c\n @@ -44,6 +44,8 @@\n\n +#include \"time.h\"\n ...\n @@ -5287,6 +5292,7 @@ ProcessLogQueryPlanInterrupt(void)\n * we check all the LocalLock entries and when finding even \none, give up\n * logging the plan.\n */\n + start = clock();\n hash_seq_init(&status, GetLockMethodLocalHash());\n while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != \nNULL)\n {\n if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\n {\n ereport(LOG_SERVER_ONLY,\n errmsg(\"ignored request for logging query plan due \nto page lock conflicts\"),\n errdetail(\"You can try again in a moment.\"));\n hash_seq_term(&status);\n\n ProcessLogQueryPlanInterruptActive = false;\n return;\n }\n }\n + end = clock();\n + cpu_time_used = ((double) (end - start)) / CLOCKS_PER_SEC;\n +\n + ereport(LOG,\n + errmsg(\"all locallock entry search took: %f\", \ncpu_time_used));\n +\n\nThere were about 3 million log lines which recorded the cpu time, and \nthe duration was quite short:\n\n =# -- Extracted cpu_time_used from log and loaded it to cpu_time.d.\n =# select max(d), min(d), avg(d) from cpu_time ;\n max | min | avg\n ----------+-----+-----------------------\n 0.000116 | 0 | 4.706274625332238e-07\n\nI'm not certain that this is valid for actual use cases, but these \nresults seem to suggest that it will not take that long.\n\n\n> 2. What happens if auto_explain is enabled in the backend and\n> pg_log_query_plan() is called on the same backend? Will they conflict?\n> I think we should add a test for the same.\n\nHmm, I think they don't conflict since they just refer QueryDesc and \ndon't modify it and don't use same objects for locking.\n(I imagine 'conflict' here is something like 'hard conflict' in \nreplication[2].)\n\nActually using both auto_explain and pg_log_query_plan() output each \nlogs separately:\n\n (pid:62835)=# select pg_sleep(10);\n (pid:70000)=# select pg_log_query_plan(62835);\n\n (pid:70000)=# \\! cat data/log/postgres.log\n ...\n 2024-02-06 21:44:17.837 JST [62835:4:0] LOG: 00000: query plan \nrunning on backend with PID 62835 is:\n Query Text: select pg_sleep(10);\n Result (cost=0.00..0.01 rows=1 width=4)\n Output: pg_sleep('10'::double precision)\n Query Identifier: 3506829283127886044\n 2024-02-06 21:44:17.837 JST [62835:5:0] LOCATION: \nProcessLogQueryPlanInterrupt, explain.c:5336\n 2024-02-06 21:44:26.974 JST [62835:6:0] LOG: 00000: duration: \n10000.868 ms plan:\n Query Text: select pg_sleep(10);\n Result (cost=0.00..0.01 rows=1 width=4) (actual \ntime=10000.802..10000.804 rows=1 loops=1)\n\n> Using injection point support we should be able to add tests for\n> testing pg_log_query_plan behaviour when there are page locks held or\n> when auto_explain (with instrumentation) and pg_log_query_plan() work\n> on the same query plan. Use injection point to make the backend\n> running query wait at a suitable point to delay its execution and fire\n> pg_log_query_plan() from other backend. May be the same test could\n> examine the server log file to see if the plan is indeed output to the\n> server log file.\n> \n> Given that the feature will be used when the things have already gone\n> wrong, it should not make things more serious. So more testing and\n> especially automated would help.\n\nThanks for the advice, it seems a good idea.\nI'm going to try to add tests using injection point.\n\n\n[1] \nhttps://www.postgresql.org/message-id/CAAaqYe8LXVXQhYy3yT0QOHUymdM%3Duha0dJ0%3DBEPzVAx2nG1gsw%40mail.gmail.com\n[2] \nhttps://www.postgresql.org/docs/devel/hot-standby.html#HOT-STANDBY-CONFLICT\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Wed, 07 Feb 2024 13:08:24 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Wed, Feb 7, 2024 at 9:38 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> Hi Ashutosh,\n>\n> On 2024-02-06 19:51, Ashutosh Bapat wrote:\n>\n> > Thanks for the summary. It is helpful. I think patch is also getting\n> > better.\n> >\n> > I have a few questions and suggestions\n>\n> Thanks for your comments.\n>\n> > 1. Prologue of GetLockMethodLocalHash() mentions\n> > * NOTE: When there are many entries in LockMethodLocalHash, calling\n> > this\n> > * function and looking into all of them can lead to performance\n> > problems.\n> > */\n> > How bad this performance could be. Let's assume that a query is taking\n> > time and pg_log_query_plan() is invoked to examine the plan of this\n> > query. Is it possible that the looping over all the locks itself takes\n> > a lot of time delaying the query execution further?\n>\n> I think it depends on the number of local locks, but I've measured cpu\n> time for this page lock check by adding below codes and\n> v27-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch[1], which\n> calls ProcessLogQueryPlanInterrupt() in every CFI on my laptop just for\n> your information:\n>\n> diff --git a/src/backend/commands/explain.c\n> b/src/backend/commands/explain.c\n> index 5f7d77d567..65b7cb4925 100644\n> --- a/src/backend/commands/explain.c\n> +++ b/src/backend/commands/explain.c\n> @@ -44,6 +44,8 @@\n>\n> +#include \"time.h\"\n> ...\n> @@ -5287,6 +5292,7 @@ ProcessLogQueryPlanInterrupt(void)\n> * we check all the LocalLock entries and when finding even\n> one, give up\n> * logging the plan.\n> */\n> + start = clock();\n> hash_seq_init(&status, GetLockMethodLocalHash());\n> while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) !=\n> NULL)\n> {\n> if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\n> {\n> ereport(LOG_SERVER_ONLY,\n> errmsg(\"ignored request for logging query plan due\n> to page lock conflicts\"),\n> errdetail(\"You can try again in a moment.\"));\n> hash_seq_term(&status);\n>\n> ProcessLogQueryPlanInterruptActive = false;\n> return;\n> }\n> }\n> + end = clock();\n> + cpu_time_used = ((double) (end - start)) / CLOCKS_PER_SEC;\n> +\n> + ereport(LOG,\n> + errmsg(\"all locallock entry search took: %f\",\n> cpu_time_used));\n> +\n>\n> There were about 3 million log lines which recorded the cpu time, and\n> the duration was quite short:\n>\n> =# -- Extracted cpu_time_used from log and loaded it to cpu_time.d.\n> =# select max(d), min(d), avg(d) from cpu_time ;\n> max | min | avg\n> ----------+-----+-----------------------\n> 0.000116 | 0 | 4.706274625332238e-07\n>\n> I'm not certain that this is valid for actual use cases, but these\n> results seem to suggest that it will not take that long.\n\nWhat load did you run? I don't think any query in make check would\ntake say thousands of locks. The prologue refers to a very populated\nlock hash table. I think that will happen if thousands of tables are\nqueried in a single query OR a query runs on a partitioned table with\nthousands of partitions. May be we want to try that scenario.\n\n>\n>\n> > 2. What happens if auto_explain is enabled in the backend and\n> > pg_log_query_plan() is called on the same backend? Will they conflict?\n> > I think we should add a test for the same.\n>\n> Hmm, I think they don't conflict since they just refer QueryDesc and\n> don't modify it and don't use same objects for locking.\n> (I imagine 'conflict' here is something like 'hard conflict' in\n> replication[2].)\n\nBy conflict, I mean the two features behave weird when used together\ne.g give wrong results or crash etc.\n\n>\n> Actually using both auto_explain and pg_log_query_plan() output each\n> logs separately:\n>\n> (pid:62835)=# select pg_sleep(10);\n> (pid:70000)=# select pg_log_query_plan(62835);\n>\n> (pid:70000)=# \\! cat data/log/postgres.log\n> ...\n> 2024-02-06 21:44:17.837 JST [62835:4:0] LOG: 00000: query plan\n> running on backend with PID 62835 is:\n> Query Text: select pg_sleep(10);\n> Result (cost=0.00..0.01 rows=1 width=4)\n> Output: pg_sleep('10'::double precision)\n> Query Identifier: 3506829283127886044\n> 2024-02-06 21:44:17.837 JST [62835:5:0] LOCATION:\n> ProcessLogQueryPlanInterrupt, explain.c:5336\n> 2024-02-06 21:44:26.974 JST [62835:6:0] LOG: 00000: duration:\n> 10000.868 ms plan:\n> Query Text: select pg_sleep(10);\n> Result (cost=0.00..0.01 rows=1 width=4) (actual\n> time=10000.802..10000.804 rows=1 loops=1)\n>\n> > Using injection point support we should be able to add tests for\n> > testing pg_log_query_plan behaviour when there are page locks held or\n> > when auto_explain (with instrumentation) and pg_log_query_plan() work\n> > on the same query plan. Use injection point to make the backend\n> > running query wait at a suitable point to delay its execution and fire\n> > pg_log_query_plan() from other backend. May be the same test could\n> > examine the server log file to see if the plan is indeed output to the\n> > server log file.\n> >\n> > Given that the feature will be used when the things have already gone\n> > wrong, it should not make things more serious. So more testing and\n> > especially automated would help.\n>\n> Thanks for the advice, it seems a good idea.\n> I'm going to try to add tests using injection point.\n\nYour test with pg_sleep() is a good basic test. But more involved\ntesting might need something like injection points.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Wed, 7 Feb 2024 10:28:18 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2024-02-07 13:58, Ashutosh Bapat wrote:\n> On Wed, Feb 7, 2024 at 9:38 AM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> \n>> Hi Ashutosh,\n>> \n>> On 2024-02-06 19:51, Ashutosh Bapat wrote:\n>> \n>> > Thanks for the summary. It is helpful. I think patch is also getting\n>> > better.\n>> >\n>> > I have a few questions and suggestions\n>> \n>> Thanks for your comments.\n>> \n>> > 1. Prologue of GetLockMethodLocalHash() mentions\n>> > * NOTE: When there are many entries in LockMethodLocalHash, calling\n>> > this\n>> > * function and looking into all of them can lead to performance\n>> > problems.\n>> > */\n>> > How bad this performance could be. Let's assume that a query is taking\n>> > time and pg_log_query_plan() is invoked to examine the plan of this\n>> > query. Is it possible that the looping over all the locks itself takes\n>> > a lot of time delaying the query execution further?\n>> \n>> I think it depends on the number of local locks, but I've measured cpu\n>> time for this page lock check by adding below codes and\n>> v27-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch[1], which\n>> calls ProcessLogQueryPlanInterrupt() in every CFI on my laptop just \n>> for\n>> your information:\n>> \n>> diff --git a/src/backend/commands/explain.c\n>> b/src/backend/commands/explain.c\n>> index 5f7d77d567..65b7cb4925 100644\n>> --- a/src/backend/commands/explain.c\n>> +++ b/src/backend/commands/explain.c\n>> @@ -44,6 +44,8 @@\n>> \n>> +#include \"time.h\"\n>> ...\n>> @@ -5287,6 +5292,7 @@ ProcessLogQueryPlanInterrupt(void)\n>> * we check all the LocalLock entries and when finding even\n>> one, give up\n>> * logging the plan.\n>> */\n>> + start = clock();\n>> hash_seq_init(&status, GetLockMethodLocalHash());\n>> while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) \n>> !=\n>> NULL)\n>> {\n>> if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\n>> {\n>> ereport(LOG_SERVER_ONLY,\n>> errmsg(\"ignored request for logging query plan \n>> due\n>> to page lock conflicts\"),\n>> errdetail(\"You can try again in a moment.\"));\n>> hash_seq_term(&status);\n>> \n>> ProcessLogQueryPlanInterruptActive = false;\n>> return;\n>> }\n>> }\n>> + end = clock();\n>> + cpu_time_used = ((double) (end - start)) / CLOCKS_PER_SEC;\n>> +\n>> + ereport(LOG,\n>> + errmsg(\"all locallock entry search took: %f\",\n>> cpu_time_used));\n>> +\n>> \n>> There were about 3 million log lines which recorded the cpu time, and\n>> the duration was quite short:\n>> \n>> =# -- Extracted cpu_time_used from log and loaded it to cpu_time.d.\n>> =# select max(d), min(d), avg(d) from cpu_time ;\n>> max | min | avg\n>> ----------+-----+-----------------------\n>> 0.000116 | 0 | 4.706274625332238e-07\n>> \n>> I'm not certain that this is valid for actual use cases, but these\n>> results seem to suggest that it will not take that long.\n> \n> What load did you run? I don't think any query in make check would\n> take say thousands of locks.\n\nSorry, I forgot to write it but ran make check as you imagined.\n\n> The prologue refers to a very populated\n> lock hash table. I think that will happen if thousands of tables are\n> queried in a single query OR a query runs on a partitioned table with\n> thousands of partitions. May be we want to try that scenario.\n\nOK, I'll try such cases.\n\n>> > 2. What happens if auto_explain is enabled in the backend and\n>> > pg_log_query_plan() is called on the same backend? Will they conflict?\n>> > I think we should add a test for the same.\n>> \n>> Hmm, I think they don't conflict since they just refer QueryDesc and\n>> don't modify it and don't use same objects for locking.\n>> (I imagine 'conflict' here is something like 'hard conflict' in\n>> replication[2].)\n> \n> By conflict, I mean the two features behave weird when used together\n> e.g give wrong results or crash etc.\n> \n>> \n>> Actually using both auto_explain and pg_log_query_plan() output each\n>> logs separately:\n>> \n>> (pid:62835)=# select pg_sleep(10);\n>> (pid:70000)=# select pg_log_query_plan(62835);\n>> \n>> (pid:70000)=# \\! cat data/log/postgres.log\n>> ...\n>> 2024-02-06 21:44:17.837 JST [62835:4:0] LOG: 00000: query plan\n>> running on backend with PID 62835 is:\n>> Query Text: select pg_sleep(10);\n>> Result (cost=0.00..0.01 rows=1 width=4)\n>> Output: pg_sleep('10'::double precision)\n>> Query Identifier: 3506829283127886044\n>> 2024-02-06 21:44:17.837 JST [62835:5:0] LOCATION:\n>> ProcessLogQueryPlanInterrupt, explain.c:5336\n>> 2024-02-06 21:44:26.974 JST [62835:6:0] LOG: 00000: duration:\n>> 10000.868 ms plan:\n>> Query Text: select pg_sleep(10);\n>> Result (cost=0.00..0.01 rows=1 width=4) (actual\n>> time=10000.802..10000.804 rows=1 loops=1)\n>> \n>> > Using injection point support we should be able to add tests for\n>> > testing pg_log_query_plan behaviour when there are page locks held or\n>> > when auto_explain (with instrumentation) and pg_log_query_plan() work\n>> > on the same query plan. Use injection point to make the backend\n>> > running query wait at a suitable point to delay its execution and fire\n>> > pg_log_query_plan() from other backend. May be the same test could\n>> > examine the server log file to see if the plan is indeed output to the\n>> > server log file.\n>> >\n>> > Given that the feature will be used when the things have already gone\n>> > wrong, it should not make things more serious. So more testing and\n>> > especially automated would help.\n>> \n>> Thanks for the advice, it seems a good idea.\n>> I'm going to try to add tests using injection point.\n> \n> Your test with pg_sleep() is a good basic test. But more involved\n> testing might need something like injection points.\n\nIt might be so, I will consider whether there are any subtle timing \nissues, etc.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Wed, 07 Feb 2024 19:14:40 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Wed, Feb 7, 2024 at 12:58 PM Ashutosh Bapat\n<ashutosh.bapat.oss@gmail.com> wrote:\n>\n> >\n> > > */\n> > > How bad this performance could be. Let's assume that a query is taking\n> > > time and pg_log_query_plan() is invoked to examine the plan of this\n> > > query. Is it possible that the looping over all the locks itself takes\n> > > a lot of time delaying the query execution further?\n> >\ncorner case test:\npgbench --initialize --partition-method=range --partitions=20000\nSomehow my setup, the pg_bench didn't populate the data but there are\n20000 partitions there.\n(all my other settings are default)\n\nsome interesting things happened when a query touch so many partitions like:\nselect abalance, aid from public.pgbench_accounts,pg_sleep(4) where aid > 1;\n\nin another session, if you immediate call SELECT pg_log_query_plan(9482);\nthen output be\n`\nLOG: backend with PID 9482 is not running a query or a subtransaction\nis aborted\n`\nhowever if you delay a little bit of time (like 1 second), then\nLOG will emit the plan with lots of text (not sure the plan is right).\n\nI think the reason is that the `InitPlan` within\nstandard_ExecutorStart takes more time to finish\nwhen your query touches a lot of partitions.\n\n\n","msg_date":"Mon, 12 Feb 2024 08:00:00 +0800","msg_from":"jian he <jian.universality@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Mon, Jan 29, 2024 at 9:02 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> Hi,\n>\n> Updated the patch to fix typos and move\n> ProcessLogQueryPlanInterruptActive from errfinish() to AbortTransaction.\n>\n\n+ <row>\n+ <entry role=\"func_table_entry\"><para role=\"func_signature\">\n+ <indexterm>\n+ <primary>pg_log_query_plan</primary>\n+ </indexterm>\n+ <function>pg_log_query_plan</function> (\n<parameter>pid</parameter> <type>integer</type> )\n+ <returnvalue>boolean</returnvalue>\n+ </para>\n+ <para>\n+ Requests to log the plan of the query currently running on the\n+ backend with specified process ID.\n+ It will be logged at <literal>LOG</literal> message level and\n+ will appear in the server log based on the log\n+ configuration set (See <xref linkend=\"runtime-config-logging\"/>\n+ for more information), but will not be sent to the client\n+ regardless of <xref linkend=\"guc-client-min-messages\"/>.\n+ </para></entry>\n+ </row>\nit would be better to explain the meaning of return value TRUE/FALSE?\n\n+# logging plan of the running query on the specified backend\n+{ oid => '8000', descr => 'log plan of the running query on the\nspecified backend',\n+ proname => 'pg_log_query_plan',\n+ provolatile => 'v', prorettype => 'bool',\nyou can add\n`proargnames => '{pid}'`\n\n+ if (proc == NULL)\n+ {\n+ /*\n+ * This is just a warning so a loop-through-resultset will not abort\n+ * if one backend terminated on its own during the run.\n+ */\n+ ereport(WARNING,\n+ (errmsg(\"PID %d is not a PostgreSQL backend process\", pid)));\n+ PG_RETURN_BOOL(false);\n+ }\n+\n+ be_status = pgstat_get_beentry_by_backend_id(proc->backendId);\n+ if (be_status->st_backendType != B_BACKEND)\n+ {\n+ ereport(WARNING,\n+ (errmsg(\"PID %d is not a PostgreSQL client backend process\", pid)));\n+ PG_RETURN_BOOL(false);\n+ }\n+\n+ if (SendProcSignal(pid, PROCSIG_LOG_QUERY_PLAN, proc->backendId) < 0)\n+ {\n+ /* Again, just a warning to allow loops */\n+ ereport(WARNING,\n+ (errmsg(\"could not send signal to process %d: %m\", pid)));\n+ PG_RETURN_BOOL(false);\n+ }\n\nI found out that pg_log_query_plan's comments look like\npg_log_backend_memory_contexts.\npg_log_backend_memory_contexts will iterate through many memory contexts.\nbut pg_log_query_plan for one specific pid will only output one plan?\nso I am a little bit confused by the comments.\n\n+ /*\n+ * Ensure no page lock is held on this process.\n+ *\n+ * If page lock is held at the time of the interrupt, we can't acquire any\n+ * other heavyweight lock, which might be necessary for explaining the plan\n+ * when retrieving column names.\n+ *\n+ * This may be overkill, but since page locks are held for a short duration\n+ * we check all the LocalLock entries and when finding even one, give up\n+ * logging the plan.\n+ */\n+ hash_seq_init(&status, GetLockMethodLocalHash());\n+ while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)\n+ {\n+ if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\nmaybe not that self evident, the above comments still not explained\nwhy we need to ensure only\nPAGE lock was held on this process?\n\nIn the commit message, can you add all the discussion links?\nMy gmail account doesn't have a previous discussion history.\nI am not sure this\n(https://www.postgresql.org/message-id/flat/d68c3ae31672664876b22d2dcbb526d2%40postgrespro.ru)\nis the only discussion link?\n\nI found a bug:\nsrc8=# select *, pg_sleep(10) from tenk1 for update;\n2024-02-11 15:54:17.944 CST [48602] LOG: query plan running on\nbackend with PID 48602 is:\n Query Text: select *, pg_sleep(10) from tenk1 for update;\n LockRows (cost=0.00..570.00 rows=10000 width=254)\n Output: unique1, unique2, two, four, ten, twenty, hundred,\nthousand, twothousand, fivethous, tenthous, odd, even, stringu1,\nstringu2, string4, (pg_sleep('10'::double precision)), ctid\n -> Seq Scan on public.tenk1 (cost=0.00..470.00 rows=10000 width=254)\n Output: unique1, unique2, two, four, ten, twenty,\nhundred, thousand, twothousand, fivethous, tenthous, odd, even,\nstringu1, stringu2, string4, pg_sleep('10'::double precision), ctid\n\nanother session (PID) executes `SELECT pg_log_query_plan(48602);` in\nthe meantime.\npg_log_query_plan returns true successfully, but PID 48602 was stuck.\n\nI have problem using git apply:\nerror: patch failed: src/include/commands/explain.h:94\nerror: src/include/commands/explain.h: patch does not apply\n\n`patch -p1 < /v35-0001-Add-function-to-log-the-plan-of-the-query.patch` works\n\n\n","msg_date":"Mon, 12 Feb 2024 08:00:00 +0800","msg_from":"jian he <jian.universality@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Mon, Feb 12, 2024 at 5:31 AM jian he <jian.universality@gmail.com> wrote:\n>\n> I found a bug:\n> src8=# select *, pg_sleep(10) from tenk1 for update;\n> 2024-02-11 15:54:17.944 CST [48602] LOG: query plan running on\n> backend with PID 48602 is:\n> Query Text: select *, pg_sleep(10) from tenk1 for update;\n> LockRows (cost=0.00..570.00 rows=10000 width=254)\n> Output: unique1, unique2, two, four, ten, twenty, hundred,\n> thousand, twothousand, fivethous, tenthous, odd, even, stringu1,\n> stringu2, string4, (pg_sleep('10'::double precision)), ctid\n> -> Seq Scan on public.tenk1 (cost=0.00..470.00 rows=10000 width=254)\n> Output: unique1, unique2, two, four, ten, twenty,\n> hundred, thousand, twothousand, fivethous, tenthous, odd, even,\n> stringu1, stringu2, string4, pg_sleep('10'::double precision), ctid\n>\n> another session (PID) executes `SELECT pg_log_query_plan(48602);` in\n> the meantime.\n> pg_log_query_plan returns true successfully, but PID 48602 was stuck.\n\nWhat do you mean by PID 48602 was stuck?\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Mon, 12 Feb 2024 10:09:18 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Mon, Feb 12, 2024 at 5:31 AM jian he <jian.universality@gmail.com> wrote:\n>\n> On Wed, Feb 7, 2024 at 12:58 PM Ashutosh Bapat\n> <ashutosh.bapat.oss@gmail.com> wrote:\n> >\n> > >\n> > > > */\n> > > > How bad this performance could be. Let's assume that a query is taking\n> > > > time and pg_log_query_plan() is invoked to examine the plan of this\n> > > > query. Is it possible that the looping over all the locks itself takes\n> > > > a lot of time delaying the query execution further?\n> > >\n> corner case test:\n> pgbench --initialize --partition-method=range --partitions=20000\n> Somehow my setup, the pg_bench didn't populate the data but there are\n> 20000 partitions there.\n> (all my other settings are default)\n>\n> some interesting things happened when a query touch so many partitions like:\n> select abalance, aid from public.pgbench_accounts,pg_sleep(4) where aid > 1;\n>\n> in another session, if you immediate call SELECT pg_log_query_plan(9482);\n> then output be\n> `\n> LOG: backend with PID 9482 is not running a query or a subtransaction\n> is aborted\n> `\n> however if you delay a little bit of time (like 1 second), then\n> LOG will emit the plan with lots of text (not sure the plan is right).\n>\n> I think the reason is that the `InitPlan` within\n> standard_ExecutorStart takes more time to finish\n> when your query touches a lot of partitions.\n\nThat's probably expected unless we make the ActiveQueryDesc available\nbefore ExecutorRun.\n\nHow much time did it took between issuing SELECT\npg_log_query_plan(9482); and plan getting output to the server error\nlogs? How does this time compare with say the same time difference for\na simple query and how much of that time can be attributed to Lock\ntable hash scan, if the difference between time difference is huge.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Mon, 12 Feb 2024 10:12:21 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Mon, Feb 12, 2024 at 12:42 PM Ashutosh Bapat\n<ashutosh.bapat.oss@gmail.com> wrote:\n>\n> On Mon, Feb 12, 2024 at 5:31 AM jian he <jian.universality@gmail.com> wrote:\n> >\n> > On Wed, Feb 7, 2024 at 12:58 PM Ashutosh Bapat\n> > <ashutosh.bapat.oss@gmail.com> wrote:\n> > >\n> > > >\n> > > > > */\n> > > > > How bad this performance could be. Let's assume that a query is taking\n> > > > > time and pg_log_query_plan() is invoked to examine the plan of this\n> > > > > query. Is it possible that the looping over all the locks itself takes\n> > > > > a lot of time delaying the query execution further?\n> > > >\n> > corner case test:\n> > pgbench --initialize --partition-method=range --partitions=20000\n> > Somehow my setup, the pg_bench didn't populate the data but there are\n> > 20000 partitions there.\n> > (all my other settings are default)\n> >\n> > some interesting things happened when a query touch so many partitions like:\n> > select abalance, aid from public.pgbench_accounts,pg_sleep(4) where aid > 1;\n> >\n> > in another session, if you immediate call SELECT pg_log_query_plan(9482);\n> > then output be\n> > `\n> > LOG: backend with PID 9482 is not running a query or a subtransaction\n> > is aborted\n> > `\n> > however if you delay a little bit of time (like 1 second), then\n> > LOG will emit the plan with lots of text (not sure the plan is right).\n> >\n> > I think the reason is that the `InitPlan` within\n> > standard_ExecutorStart takes more time to finish\n> > when your query touches a lot of partitions.\n>\n> That's probably expected unless we make the ActiveQueryDesc available\n> before ExecutorRun.\n\nso the error message should be something like:\nerrmsg(\"backend with PID %d is not running a query or a subtransaction\nis aborted or the plan is not generated\",\n\n>\n> How much time did it took between issuing SELECT\n> pg_log_query_plan(9482); and plan getting output to the server error\n> logs?\n\nit either says errmsg(\"backend with PID %d is not running a query,.....)\nor outputs the plan immediately, if i wait one or two seconds to call\npg_log_query_plan.\n\nbecause of previously mentioned: with lots of partitions, initplan\ntook longer to finish.\nsetup: 2 sessions, one runs the query (select abalance, aid from\npublic.pgbench_accounts,pg_sleep(4) where aid > 1;),\nanother one calls pg_log_query_plan.\n\n\n","msg_date":"Mon, 12 Feb 2024 15:08:00 +0800","msg_from":"jian he <jian.universality@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2024-02-07 19:14, torikoshia wrote:\n> On 2024-02-07 13:58, Ashutosh Bapat wrote:\n\n>> The prologue refers to a very populated\n>> lock hash table. I think that will happen if thousands of tables are\n>> queried in a single query OR a query runs on a partitioned table with\n>> thousands of partitions. May be we want to try that scenario.\n> \n> OK, I'll try such cases.\n\nI measured this using partitioned pgbench_accounts with some \nmodification to v36[1].\nThe results[2] show that CPU time increases in proportion to the number \nof partitions, and the increase is not that large.\n\nHowever I've noticed that these ensuring no page lock logic would not be \nnecessary anymore since cc32ec24fdf3b98 removed the assertion which \ncaused an error[1].\n\n $ git show cc32ec24fdf3b98\n ..\n diff --git a/src/backend/storage/lmgr/lock.c \nb/src/backend/storage/lmgr/lock.c\n index 0a692ee0a6..f595bce31b 100644\n --- a/src/backend/storage/lmgr/lock.c\n +++ b/src/backend/storage/lmgr/lock.c\n @@ -186,18 +186,6 @@ static int FastPathLocalUseCount = 0;\n */\n static bool IsRelationExtensionLockHeld PG_USED_FOR_ASSERTS_ONLY = \nfalse;\n\n - /*\n - * We don't acquire any other heavyweight lock while holding \nthe page lock\n - * except for relation extension.\n - */\n - Assert(!IsPageLockHeld ||\n - (locktag->locktag_type == \nLOCKTAG_RELATION_EXTEND));\n\nI'm going to remove ensuring no page lock logic after some testings.\n\n\n[1]\n$ git diff _submission/log_running_query-v36\n+#include \"time.h\"\n+\n bool ProcessLogQueryPlanInterruptActive = false;\n\n /* Hook for plugins to get control in ExplainOneQuery() */\n@@ -5258,6 +5260,10 @@ ProcessLogQueryPlanInterrupt(void)\n MemoryContext old_cxt;\n LogQueryPlanPending = false;\n\n+ clock_t start, end;\n+ double cpu_time_used;\n+ int num_hash_entry = 0;\n+\n /* Cannot re-enter. */\n if (ProcessLogQueryPlanInterruptActive)\n return;\n@@ -5287,9 +5293,11 @@ ProcessLogQueryPlanInterrupt(void)\n * we check all the LocalLock entries and when finding even one, \ngive up\n * logging the plan.\n */\n+ start = clock();\n hash_seq_init(&status, GetLockMethodLocalHash());\n while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)\n {\n+ num_hash_entry++;\n if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\n {\n ereport(LOG_SERVER_ONLY,\n@@ -5301,6 +5309,12 @@ ProcessLogQueryPlanInterrupt(void)\n return;\n }\n }\n+ end = clock();\n+ cpu_time_used = ((double) (end - start)) / CLOCKS_PER_SEC;\n+\n+ ereport(LOG,\n+ errmsg(\"locallock entry search took: %f for %d entries\", \ncpu_time_used, num_hash_entry));\n\n[2]\n# partition number: 512\nlocallock entry search took: 0.000029 for 1026 entries\nlocallock entry search took: 0.000030 for 1026 entries\nlocallock entry search took: 0.000036 for 1026 entries\n\n# partition number: 1024\nlocallock entry search took: 0.000070 for 2050 entries\nlocallock entry search took: 0.000059 for 2050 entries\nlocallock entry search took: 0.000049 for 2050 entries\n\n# partition number: 2048\nlocallock entry search took: 0.000100 for 4098 entries\nlocallock entry search took: 0.000103 for 4098 entries\nlocallock entry search took: 0.000101 for 4098 entries\n\n# partition number: 4096\nlocallock entry search took: 0.000197 for 8194 entries\nlocallock entry search took: 0.000193 for 8194 entries\nlocallock entry search took: 0.000192 for 8194 entries\n\n[3] \nhttps://www.postgresql.org/message-id/0642712f-1298-960a-a3ba-e256d56040ac%40oss.nttdata.com\n\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Tue, 13 Feb 2024 09:19:03 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2024-02-12 09:00, jian he wrote:\n\nThanks for you comments.\n\n> On Mon, Jan 29, 2024 at 9:02 PM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> \n>> Hi,\n>> \n>> Updated the patch to fix typos and move\n>> ProcessLogQueryPlanInterruptActive from errfinish() to \n>> AbortTransaction.\n>> \n> \n> + <row>\n> + <entry role=\"func_table_entry\"><para role=\"func_signature\">\n> + <indexterm>\n> + <primary>pg_log_query_plan</primary>\n> + </indexterm>\n> + <function>pg_log_query_plan</function> (\n> <parameter>pid</parameter> <type>integer</type> )\n> + <returnvalue>boolean</returnvalue>\n> + </para>\n> + <para>\n> + Requests to log the plan of the query currently running on the\n> + backend with specified process ID.\n> + It will be logged at <literal>LOG</literal> message level and\n> + will appear in the server log based on the log\n> + configuration set (See <xref \n> linkend=\"runtime-config-logging\"/>\n> + for more information), but will not be sent to the client\n> + regardless of <xref linkend=\"guc-client-min-messages\"/>.\n> + </para></entry>\n> + </row>\n> it would be better to explain the meaning of return value TRUE/FALSE?\n\nYeah, but I've noticed that this should be located in 'Table Server \nSignaling Functions' not 'Table Control Data Functions'.\nSince 'Table Server Signaling Functions' describes the return code as \nbelow, just relocation seems fine.\n\n Each of these functions returns true if the signal was successfully \nsent and false if sending the signal failed.\n\n> +# logging plan of the running query on the specified backend\n> +{ oid => '8000', descr => 'log plan of the running query on the\n> specified backend',\n> + proname => 'pg_log_query_plan',\n> + provolatile => 'v', prorettype => 'bool',\n> you can add\n> `proargnames => '{pid}'`\n\nHmm, pg_log_query_plan() can take one argument, I'm not sure how much \nsense it makes.\nOther functions which take one argument such as pg_cancel_backend() does \nnot have proargnames.\n\n\n> + if (proc == NULL)\n> + {\n> + /*\n> + * This is just a warning so a loop-through-resultset will not abort\n> + * if one backend terminated on its own during the run.\n> + */\n> + ereport(WARNING,\n> + (errmsg(\"PID %d is not a PostgreSQL backend process\", pid)));\n> + PG_RETURN_BOOL(false);\n> + }\n> +\n> + be_status = pgstat_get_beentry_by_backend_id(proc->backendId);\n> + if (be_status->st_backendType != B_BACKEND)\n> + {\n> + ereport(WARNING,\n> + (errmsg(\"PID %d is not a PostgreSQL client backend process\", pid)));\n> + PG_RETURN_BOOL(false);\n> + }\n> +\n> + if (SendProcSignal(pid, PROCSIG_LOG_QUERY_PLAN, proc->backendId) < 0)\n> + {\n> + /* Again, just a warning to allow loops */\n> + ereport(WARNING,\n> + (errmsg(\"could not send signal to process %d: %m\", pid)));\n> + PG_RETURN_BOOL(false);\n> + }\n> \n> I found out that pg_log_query_plan's comments look like\n> pg_log_backend_memory_contexts.\n> pg_log_backend_memory_contexts will iterate through many memory \n> contexts.\n> but pg_log_query_plan for one specific pid will only output one plan?\n> so I am a little bit confused by the comments.\n\nThese \"loop\" mean backend can run pg_log_query_plan() repeatedly even \nwhen failing sending signals.\npg_signal_backend() also have such comments.\n\n> + /*\n> + * Ensure no page lock is held on this process.\n> + *\n> + * If page lock is held at the time of the interrupt, we can't acquire \n> any\n> + * other heavyweight lock, which might be necessary for explaining the \n> plan\n> + * when retrieving column names.\n> + *\n> + * This may be overkill, but since page locks are held for a short \n> duration\n> + * we check all the LocalLock entries and when finding even one, give \n> up\n> + * logging the plan.\n> + */\n> + hash_seq_init(&status, GetLockMethodLocalHash());\n> + while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)\n> + {\n> + if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\n> maybe not that self evident, the above comments still not explained\n> why we need to ensure only\n> PAGE lock was held on this process?\n\nThis is for preventing assertion error and it seems not necessary \nanymore as described in [1].\nI'm going remove them.\n\n> \n> In the commit message, can you add all the discussion links?\n> My gmail account doesn't have a previous discussion history.\n\nSure.\n\n> I am not sure this\n> (https://www.postgresql.org/message-id/flat/d68c3ae31672664876b22d2dcbb526d2%40postgrespro.ru)\n> is the only discussion link?\n\nThis is the original one:\nhttps://www.postgresql.org/message-id/cf8501bcd95ba4d727cbba886ba9eea8%40oss.nttdata.com\n\n> I found a bug:\n> src8=# select *, pg_sleep(10) from tenk1 for update;\n> 2024-02-11 15:54:17.944 CST [48602] LOG: query plan running on\n> backend with PID 48602 is:\n> Query Text: select *, pg_sleep(10) from tenk1 for update;\n> LockRows (cost=0.00..570.00 rows=10000 width=254)\n> Output: unique1, unique2, two, four, ten, twenty, hundred,\n> thousand, twothousand, fivethous, tenthous, odd, even, stringu1,\n> stringu2, string4, (pg_sleep('10'::double precision)), ctid\n> -> Seq Scan on public.tenk1 (cost=0.00..470.00 rows=10000 \n> width=254)\n> Output: unique1, unique2, two, four, ten, twenty,\n> hundred, thousand, twothousand, fivethous, tenthous, odd, even,\n> stringu1, stringu2, string4, pg_sleep('10'::double precision), ctid\n> \n> another session (PID) executes `SELECT pg_log_query_plan(48602);` in\n> the meantime.\n> pg_log_query_plan returns true successfully, but PID 48602 was stuck.\n\nHmm, it's not simply sleeping, is it?\nI'm concerned a bit this because estimated rows of tenk1 is 10000.\nIf so, the query will take 10000 * 10 seconds.\n\n> I have problem using git apply:\n> error: patch failed: src/include/commands/explain.h:94\n> error: src/include/commands/explain.h: patch does not apply\n> \n> `patch -p1 < /v35-0001-Add-function-to-log-the-plan-of-the-query.patch` \n> works\n\nI'll update the patch including other points such as removing ensuring \nno page lock code.\n\n\n[1] \nhttps://www.postgresql.org/message-id/1b2b247530f3ff3afab4ddc2df222e8b%40oss.nttdata.com\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Tue, 13 Feb 2024 11:30:53 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2024-02-13 11:30, torikoshia wrote:\n\n> I'll update the patch including other points such as removing ensuring\n> no page lock code.\n\nUpdated the patch.\n\n> Using injection point support we should be able to add tests for\n> testing pg_log_query_plan behaviour when there are page locks held or\n> when auto_explain (with instrumentation) and pg_log_query_plan() work\n> on the same query plan. Use injection point to make the backend\n> running query wait at a suitable point to delay its execution and fire\n> pg_log_query_plan() from other backend. May be the same test could\n> examine the server log file to see if the plan is indeed output to the\n> server log file.\n\nAttached patch uses injection point as below:\n\n- There may be more points to inject, but added an injection point at \nExecutorRun(), which seems to be the first interruption point where \nplans can be reliably displayed.\n- At injection point, it'd be possible to wait for some duration and \nfire pg_log_plan_query() as you suggested. However, I'm not sure how \nlong duration is appropriate considering the variety of testing \nenvironments. Instead, attached patch calls \nHandleLogQueryPlanInterrupt() directly and set InterruptPending.\n- Tests both pg_log_plan_query() and auto_explain logs for their output, \nand the logged plans are the same.\n\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation","msg_date":"Wed, 14 Feb 2024 15:29:35 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi,\n\nI've just been catching up on this thread.\n\n+ if (MyProc->heldLocks)\n+ {\n+ ereport(LOG_SERVER_ONLY,\n+ errmsg(\"ignored request for logging query plan due to lock conflicts\"),\n+ errdetail(\"You can try again in a moment.\"));\n+ return;\n+ }\n\nI don't like this for several reasons.\n\nFirst, I think it's not nice to have a request just get ignored. A\nuser will expect that if we cannot immediately respond to some\nrequest, we'll respond later at the first time that it's safe to do\nso, rather than just ignoring it and telling them to retry.\n\nSecond, I don't think that the error message is very good. It talks\nabout lock conflicts, but we don't know that there is any real\nproblem. We know that, if we enter this block, the server is in the\nmiddle of trying to acquire some lock, and we also know that we could\nattempt to acquire a lock as part of generating the EXPLAIN output,\nand maybe that's an issue. But that's not a lock conflict. That's a\nre-entrancy problem. I don't know that we want to talk about\nre-entrancy problems in an error message, and I don't really think\nthis error message should exist in the first place, but if we're going\nto error out in this case surely we shouldn't do so with an error\nmessage that describes a problem we don't have.\n\nThird, I think that the re-entrancy problems with this patch may\nextend well beyond lock acquisition. This is one example of how it can\nbe unsafe to do something as complicated as EXPLAIN at any arbitrary\nCHECK_FOR_INTERRUPTS(). It is not correct to say, as\nhttp://postgr.es/m/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw@mail.gmail.com\ndoes, that the problems with running EXPLAIN at an arbitrary point are\nspecific to running this code in an aborted transaction. The existence\nof this code shows that there is at least one hazard even if the\ncurrent transaction is not aborted, and I see no analysis on this\nthread indicating whether there are any more such hazards, or how we\ncould go about finding them all.\n\nI think the issue is very general. We have lots of subsystems that\nboth (a) use global variables and (b) contain CHECK_FOR_INTERRUPTS().\nIf we process an interrupt while that code is in the middle of\nmanipulating its global variables and then again re-enter that code,\nbad things might happen. We could try to rule that out by analyzing\nall such subsystems and all places where CHECK_FOR_INTERRUPTS() may\nappear, but that's very difficult. Suppose we took the alternative\napproach recommended by Andrey Lepikhov in\nhttp://postgr.es/m/b1b110ae-61f6-4fd9-9b94-f967db9b53d4@app.fastmail.com\nand instead set a flag that gets handled in some suitable place in the\nexecutor code, like ExecProcNode(). If we did that, then we would know\nthat we're not in the middle of a syscache lookup, a catcache lookup,\na heavyweight lock acquisition, an ereport, or any other low-level\nsubsystem call that might create problems for the EXPLAIN code.\n\nIn that design, the hack shown above would go away, and we could be\nmuch more certain that we don't need any other similar hacks for other\nsubsystems. The only downside is that we might experience a slightly\nlonger delay before a requested EXPLAIN plan actually shows up, but\nthat seems like a pretty small price to pay for being able to reason\nabout the behavior of the system. I don't *think* there are any cases\nwhere we run in the executor for a particularly long time without a\nnew call to ExecProcNode().\n\nI think this case is a bit like vacuum_delay_point(). You might think\nthat vacuum_delay_point() could be moved inside of\nCHECK_FOR_INTERRUPTS(), but we've made the opposite decision: every\nvacuum_delay_point() calls CHECK_FOR_INTERRUPTS() but not every\nCHECK_FOR_INTERRUPTS() calls vacuum_delay_point(). That means that we\ncan allow vacuum_delay_point() only at cases where we know it's safe,\nrather than at every CHECK_FOR_INTERRUPTS(). I think that's a pretty\nsmart decision, even for vacuum_delay_point(), and AFAICS the code\nwe're proposing to run here does things that are substantially more\ncomplicated than what vacuum_delay_point() does. That code just does a\nbit of reading of shared memory, reports a wait event, and sleeps.\nThat's really simple compared to code that could try to do relcache\nbuilds, which can trigger syscache lookups, which can both trigger\nheavyweight lock acquisition, lightweight lock acquisition, bringing\npages into shared_buffers possibly through physical I/O, processing of\ninvalidation messages, and a bunch of other stuff.\n\nIt's really hard for me to accept that the heavyweight lock problem\nfor which the patch contains a workaround is the only one that exists.\nI can't see any reason why that should be true.\n\n...Robert\n\n\n","msg_date":"Thu, 15 Feb 2024 14:42:11 +0530","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi,\n\nOn 2024-02-15 14:42:11 +0530, Robert Haas wrote:\n> I think the issue is very general. We have lots of subsystems that\n> both (a) use global variables and (b) contain CHECK_FOR_INTERRUPTS().\n> If we process an interrupt while that code is in the middle of\n> manipulating its global variables and then again re-enter that code,\n> bad things might happen. We could try to rule that out by analyzing\n> all such subsystems and all places where CHECK_FOR_INTERRUPTS() may\n> appear, but that's very difficult. Suppose we took the alternative\n> approach recommended by Andrey Lepikhov in\n> http://postgr.es/m/b1b110ae-61f6-4fd9-9b94-f967db9b53d4@app.fastmail.com\n> and instead set a flag that gets handled in some suitable place in the\n> executor code, like ExecProcNode(). If we did that, then we would know\n> that we're not in the middle of a syscache lookup, a catcache lookup,\n> a heavyweight lock acquisition, an ereport, or any other low-level\n> subsystem call that might create problems for the EXPLAIN code.\n> \n> In that design, the hack shown above would go away, and we could be\n> much more certain that we don't need any other similar hacks for other\n> subsystems. The only downside is that we might experience a slightly\n> longer delay before a requested EXPLAIN plan actually shows up, but\n> that seems like a pretty small price to pay for being able to reason\n> about the behavior of the system.\n\nI am very wary of adding overhead to ExecProcNode() - I'm quite sure that\nadding code there would trigger visible overhead for query times.\n\nIf we went with something like tht approach, I think we'd have to do something\nlike redirecting node->ExecProcNode to a wrapper, presumably from within a\nCFI. That wrapper could then implement the explain support, without slowing\ndown the normal execution path.\n\n\n> I don't *think* there are any cases where we run in the executor for a\n> particularly long time without a new call to ExecProcNode().\n\nI guess it depends on what you call a long time. A large sort, for example,\ncould spend a fair amount of time inside tuplesort, similarly, a gather node\nmight need to wait for a worker for a while etc.\n\n\n> It's really hard for me to accept that the heavyweight lock problem\n> for which the patch contains a workaround is the only one that exists.\n> I can't see any reason why that should be true.\n\nI suspect you're right.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 15 Feb 2024 10:59:11 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Thu, Feb 15, 2024 at 6:12 PM Robert Haas <robertmhaas@gmail.com> \nwrote:\n> Hi,\n> \n> I've just been catching up on this thread.\n> \n> + if (MyProc->heldLocks)\n> + {\n> + ereport(LOG_SERVER_ONLY,\n> + errmsg(\"ignored request for logging query plan due to lock \n> conflicts\"),\n> + errdetail(\"You can try again in a moment.\"));\n> + return;\n> + }\n> \n> I don't like this for several reasons.\n> \n> First, I think it's not nice to have a request just get ignored. A\n> user will expect that if we cannot immediately respond to some\n> request, we'll respond later at the first time that it's safe to do\n> so, rather than just ignoring it and telling them to retry.\n> \n> Second, I don't think that the error message is very good. It talks\n> about lock conflicts, but we don't know that there is any real\n> problem. We know that, if we enter this block, the server is in the\n> middle of trying to acquire some lock, and we also know that we could\n> attempt to acquire a lock as part of generating the EXPLAIN output,\n> and maybe that's an issue. But that's not a lock conflict. That's a\n> re-entrancy problem. I don't know that we want to talk about\n> re-entrancy problems in an error message, and I don't really think\n> this error message should exist in the first place, but if we're going\n> to error out in this case surely we shouldn't do so with an error\n> message that describes a problem we don't have.\n> \n> Third, I think that the re-entrancy problems with this patch may\n> extend well beyond lock acquisition. This is one example of how it can\n> be unsafe to do something as complicated as EXPLAIN at any arbitrary\n> CHECK_FOR_INTERRUPTS(). It is not correct to say, as\n> http://postgr.es/m/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw@mail.gmail.com\n> does, that the problems with running EXPLAIN at an arbitrary point are\n> specific to running this code in an aborted transaction. The existence\n> of this code shows that there is at least one hazard even if the\n> current transaction is not aborted, and I see no analysis on this\n> thread indicating whether there are any more such hazards, or how we\n> could go about finding them all.\n> \n> I think the issue is very general. We have lots of subsystems that\n> both (a) use global variables and (b) contain CHECK_FOR_INTERRUPTS().\n> If we process an interrupt while that code is in the middle of\n> manipulating its global variables and then again re-enter that code,\n> bad things might happen. We could try to rule that out by analyzing\n> all such subsystems and all places where CHECK_FOR_INTERRUPTS() may\n> appear, but that's very difficult. Suppose we took the alternative\n> approach recommended by Andrey Lepikhov in\n> http://postgr.es/m/b1b110ae-61f6-4fd9-9b94-f967db9b53d4@app.fastmail.com\n> and instead set a flag that gets handled in some suitable place in the\n> executor code, like ExecProcNode(). If we did that, then we would know\n> that we're not in the middle of a syscache lookup, a catcache lookup,\n> a heavyweight lock acquisition, an ereport, or any other low-level\n> subsystem call that might create problems for the EXPLAIN code.\n> \n> In that design, the hack shown above would go away, and we could be\n> much more certain that we don't need any other similar hacks for other\n> subsystems. The only downside is that we might experience a slightly\n> longer delay before a requested EXPLAIN plan actually shows up, but\n> that seems like a pretty small price to pay for being able to reason\n> about the behavior of the system. I don't *think* there are any cases\n> where we run in the executor for a particularly long time without a\n> new call to ExecProcNode().\n> \n> I think this case is a bit like vacuum_delay_point(). You might think\n> that vacuum_delay_point() could be moved inside of\n> CHECK_FOR_INTERRUPTS(), but we've made the opposite decision: every\n> vacuum_delay_point() calls CHECK_FOR_INTERRUPTS() but not every\n> CHECK_FOR_INTERRUPTS() calls vacuum_delay_point(). That means that we\n> can allow vacuum_delay_point() only at cases where we know it's safe,\n> rather than at every CHECK_FOR_INTERRUPTS(). I think that's a pretty\n> smart decision, even for vacuum_delay_point(), and AFAICS the code\n> we're proposing to run here does things that are substantially more\n> complicated than what vacuum_delay_point() does. That code just does a\n> bit of reading of shared memory, reports a wait event, and sleeps.\n> That's really simple compared to code that could try to do relcache\n> builds, which can trigger syscache lookups, which can both trigger\n> heavyweight lock acquisition, lightweight lock acquisition, bringing\n> pages into shared_buffers possibly through physical I/O, processing of\n> invalidation messages, and a bunch of other stuff.\n> \n> It's really hard for me to accept that the heavyweight lock problem\n> for which the patch contains a workaround is the only one that exists.\n> I can't see any reason why that should be true.\n\nThanks for the review and the very detailed explanation!\n\nI'm convinced that it's unsafe to execute EXPLAIN codes during \nCHECK_FOR_INTERRUPTS() and we need to execute it in other safe place, as \nwell as the first and second point.\n\n\nOn 2024-02-16 03:59, Andres Freund wrote:\n> Hi,\n> \n> On 2024-02-15 14:42:11 +0530, Robert Haas wrote:\n>> I think the issue is very general. We have lots of subsystems that\n>> both (a) use global variables and (b) contain CHECK_FOR_INTERRUPTS().\n>> If we process an interrupt while that code is in the middle of\n>> manipulating its global variables and then again re-enter that code,\n>> bad things might happen. We could try to rule that out by analyzing\n>> all such subsystems and all places where CHECK_FOR_INTERRUPTS() may\n>> appear, but that's very difficult. Suppose we took the alternative\n>> approach recommended by Andrey Lepikhov in\n>> http://postgr.es/m/b1b110ae-61f6-4fd9-9b94-f967db9b53d4@app.fastmail.com\n>> and instead set a flag that gets handled in some suitable place in the\n>> executor code, like ExecProcNode(). If we did that, then we would know\n>> that we're not in the middle of a syscache lookup, a catcache lookup,\n>> a heavyweight lock acquisition, an ereport, or any other low-level\n>> subsystem call that might create problems for the EXPLAIN code.\n>> \n>> In that design, the hack shown above would go away, and we could be\n>> much more certain that we don't need any other similar hacks for other\n>> subsystems. The only downside is that we might experience a slightly\n>> longer delay before a requested EXPLAIN plan actually shows up, but\n>> that seems like a pretty small price to pay for being able to reason\n>> about the behavior of the system.\n> \n> I am very wary of adding overhead to ExecProcNode() - I'm quite sure \n> that\n> adding code there would trigger visible overhead for query times.\n> \n> If we went with something like tht approach, I think we'd have to do \n> something\n> like redirecting node->ExecProcNode to a wrapper, presumably from \n> within a\n> CFI. That wrapper could then implement the explain support, without \n> slowing\n> down the normal execution path.\n\nThanks for the idea!\nI'm not so sure about the implementation now, i.e. finding the next node \nto be executed from the planstate tree, but I'm going to try this \napproach.\n\n>> I don't *think* there are any cases where we run in the executor for a\n>> particularly long time without a new call to ExecProcNode().\n> \n> I guess it depends on what you call a long time. A large sort, for \n> example,\n> could spend a fair amount of time inside tuplesort, similarly, a gather \n> node\n> might need to wait for a worker for a while etc.\n> \n> \n>> It's really hard for me to accept that the heavyweight lock problem\n>> for which the patch contains a workaround is the only one that exists.\n>> I can't see any reason why that should be true.\n> \n> I suspect you're right.\n> \n> Greetings,\n> \n> Andres Freund\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Fri, 16 Feb 2024 23:42:36 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Fri, Feb 16, 2024 at 12:29 AM Andres Freund <andres@anarazel.de> wrote:\n> If we went with something like tht approach, I think we'd have to do something\n> like redirecting node->ExecProcNode to a wrapper, presumably from within a\n> CFI. That wrapper could then implement the explain support, without slowing\n> down the normal execution path.\n\nThat's an annoying complication; maybe there's some better way to\nhandle this. But I think we need to do something different than what\nthe patch does currently because...\n\n> > It's really hard for me to accept that the heavyweight lock problem\n> > for which the patch contains a workaround is the only one that exists.\n> > I can't see any reason why that should be true.\n>\n> I suspect you're right.\n\n...I think the current approach is just plain dead, because of this\nissue. We can't take an approach that creates an unbounded number of\nunclear reentrancy issues and then insert hacks one by one to cure\nthem (or hack around them, more to the point) as they're discovered.\n\nThe premise has to be that we only allow logging the query plan at\npoints where we know it's safe, rather than, as at present, allowing\nit in places that are unsafe and then trying to compensate with code\nelsewhere. That's not likely to ever be as stable as we want\nPostgreSQL to be.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 20 Feb 2024 10:23:04 +0530","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Mon, Feb 19, 2024 at 11:53 PM Robert Haas <robertmhaas@gmail.com> wrote:\n>\n> On Fri, Feb 16, 2024 at 12:29 AM Andres Freund <andres@anarazel.de> wrote:\n> > If we went with something like tht approach, I think we'd have to do something\n> > like redirecting node->ExecProcNode to a wrapper, presumably from within a\n> > CFI. That wrapper could then implement the explain support, without slowing\n> > down the normal execution path.\n>\n> That's an annoying complication; maybe there's some better way to\n> handle this. But I think we need to do something different than what\n> the patch does currently because...\n>\n> > > It's really hard for me to accept that the heavyweight lock problem\n> > > for which the patch contains a workaround is the only one that exists.\n> > > I can't see any reason why that should be true.\n> >\n> > I suspect you're right.\n>\n> ...I think the current approach is just plain dead, because of this\n> issue. We can't take an approach that creates an unbounded number of\n> unclear reentrancy issues and then insert hacks one by one to cure\n> them (or hack around them, more to the point) as they're discovered.\n>\n> The premise has to be that we only allow logging the query plan at\n> points where we know it's safe, rather than, as at present, allowing\n> it in places that are unsafe and then trying to compensate with code\n> elsewhere. That's not likely to ever be as stable as we want\n> PostgreSQL to be.\n\nThis is potentially a bit of a wild idea, but I wonder if having some\nkind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\n\"normal\" as opposed to \"critical\" (using that word differently than\nthe existing critical sections) would be worth it.\n\nRegards,\nJames Coleman\n\n\n","msg_date":"Wed, 21 Feb 2024 19:54:57 -0500","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\n> >\n> > ...I think the current approach is just plain dead, because of this\n> > issue. We can't take an approach that creates an unbounded number of\n> > unclear reentrancy issues and then insert hacks one by one to cure\n> > them (or hack around them, more to the point) as they're discovered.\n> >\n> > The premise has to be that we only allow logging the query plan at\n> > points where we know it's safe, rather than, as at present, allowing\n> > it in places that are unsafe and then trying to compensate with code\n> > elsewhere. That's not likely to ever be as stable as we want\n> > PostgreSQL to be.\n>\n> This is potentially a bit of a wild idea, but I wonder if having some\n> kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\n> \"normal\" as opposed to \"critical\" (using that word differently than\n> the existing critical sections) would be worth it.\n\nMy hunch is this will end up being a maintenance burden since every\ncaller has to decide (carefully) whether the call is under normal\ncondition or not. Developers will tend to take a safe approach and\nflag calls as critical. But importantly, what's normal for one\ninterrupt action may be critical for another and vice versa. Approach\nwould be useful depending upon how easy it is to comprehend the\ndefinition of \"normal\".\n\nIf a query executes for longer than a user defined threashold (session\nlevel GUC? or same value as auto_explain parameter), the executor\nproactively prepares an EXPLAIN output and keeps it handy in case\nasked for. It can do so at a \"known\" safe place and time rather than\nat any random time and location. Extra time spent in creating EXPLAIN\noutput may not be noticeable in a long running query. The EXPLAIN\noutput could be saved in pg_stat_activity or similar place. This will\navoid signaling the backend.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Thu, 22 Feb 2024 09:36:06 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\n> This is potentially a bit of a wild idea, but I wonder if having some\n> kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\n> \"normal\" as opposed to \"critical\" (using that word differently than\n> the existing critical sections) would be worth it.\n\nIt's worth considering, but the definition of \"normal\" vs. \"critical\"\nmight be hard to pin down. Or, we might end up with a definition that\nis specific to this particular case and not generalizable to others.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 23 Feb 2024 10:22:32 +0530","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi,\n\nOn Fri, Feb 23, 2024 at 10:22:32AM +0530, Robert Haas wrote:\n> On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\n> > This is potentially a bit of a wild idea, but I wonder if having some\n> > kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\n> > \"normal\" as opposed to \"critical\" (using that word differently than\n> > the existing critical sections) would be worth it.\n>\n> It's worth considering, but the definition of \"normal\" vs. \"critical\"\n> might be hard to pin down. Or, we might end up with a definition that\n> is specific to this particular case and not generalizable to others.\n\nBut it doesn't have to be all or nothing right? I mean each call could say\nwhat the situation is like in their context, like\nCHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\nslowly tag calls as needed, similarly to how we add already CFI based on users\nreport.\n\n\n","msg_date":"Fri, 23 Feb 2024 22:20:51 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> On Fri, Feb 23, 2024 at 10:22:32AM +0530, Robert Haas wrote:\n> > On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\n> > > This is potentially a bit of a wild idea, but I wonder if having some\n> > > kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\n> > > \"normal\" as opposed to \"critical\" (using that word differently than\n> > > the existing critical sections) would be worth it.\n> >\n> > It's worth considering, but the definition of \"normal\" vs. \"critical\"\n> > might be hard to pin down. Or, we might end up with a definition that\n> > is specific to this particular case and not generalizable to others.\n>\n> But it doesn't have to be all or nothing right? I mean each call could say\n> what the situation is like in their context, like\n> CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\n> slowly tag calls as needed, similarly to how we add already CFI based on users\n> report.\n\nAbsolutely. My gut feeling is that it's going to be simpler to pick a\nsmall number of places that are safe and sufficient for this\nparticular feature and add an extra call there, similar to how we do\nvacuum_delay_point(). The reason I think that's likely to be better is\nthat it will likely require changing only a relatively small number of\nplaces. If we instead start annotating CFIs, well, we've got hundreds\nof those. That's a lot more to change, and it also inconveniences\nthird-party extension authors and people doing back-patching. I'm not\nhere to say it can't work; I just think it's likely not the easiest\npath.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 23 Feb 2024 20:53:22 +0530","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Fri, Feb 23, 2024 at 10:23 AM Robert Haas <robertmhaas@gmail.com> wrote:\n>\n> On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> > On Fri, Feb 23, 2024 at 10:22:32AM +0530, Robert Haas wrote:\n> > > On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\n> > > > This is potentially a bit of a wild idea, but I wonder if having some\n> > > > kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\n> > > > \"normal\" as opposed to \"critical\" (using that word differently than\n> > > > the existing critical sections) would be worth it.\n> > >\n> > > It's worth considering, but the definition of \"normal\" vs. \"critical\"\n> > > might be hard to pin down. Or, we might end up with a definition that\n> > > is specific to this particular case and not generalizable to others.\n> >\n> > But it doesn't have to be all or nothing right? I mean each call could say\n> > what the situation is like in their context, like\n> > CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\n> > slowly tag calls as needed, similarly to how we add already CFI based on users\n> > report.\n>\n> Absolutely. My gut feeling is that it's going to be simpler to pick a\n> small number of places that are safe and sufficient for this\n> particular feature and add an extra call there, similar to how we do\n> vacuum_delay_point(). The reason I think that's likely to be better is\n> that it will likely require changing only a relatively small number of\n> places. If we instead start annotating CFIs, well, we've got hundreds\n> of those. That's a lot more to change, and it also inconveniences\n> third-party extension authors and people doing back-patching. I'm not\n> here to say it can't work; I just think it's likely not the easiest\n> path.\n\nYes, I suspect it's not the easiest path. I have a small hunch it\nmight end up paying more dividends in the future: there isn't just one\nof these things that is regularly a thorny discussion for the same\nreasons each time (basically \"no way to trigger this safely from\nanother backend interrupting another one at an arbitrary point\"), and\nif we were able to generalize a solution we may have multiple wins (a\nvery obvious one in my mind is the inability of auto explain to run an\nexplain at the precise time it's most useful: when statement timeout\nfires).\n\nBut it's also possible there are simply ways that get us more than\nthis scenario also, so I might be wrong; it's merely a hunch.\n\nRegards,\nJames Coleman\n\n\n","msg_date":"Sat, 24 Feb 2024 08:56:41 -0500","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Sat, Feb 24, 2024 at 08:56:41AM -0500, James Coleman wrote:\n> On Fri, Feb 23, 2024 at 10:23 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> >\n> > On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> > >\n> > > But it doesn't have to be all or nothing right? I mean each call could say\n> > > what the situation is like in their context, like\n> > > CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\n> > > slowly tag calls as needed, similarly to how we add already CFI based on users\n> > > report.\n> >\n> > Absolutely. My gut feeling is that it's going to be simpler to pick a\n> > small number of places that are safe and sufficient for this\n> > particular feature and add an extra call there, similar to how we do\n> > vacuum_delay_point(). The reason I think that's likely to be better is\n> > that it will likely require changing only a relatively small number of\n> > places. If we instead start annotating CFIs, well, we've got hundreds\n> > of those. That's a lot more to change, and it also inconveniences\n> > third-party extension authors and people doing back-patching. I'm not\n> > here to say it can't work; I just think it's likely not the easiest\n> > path.\n>\n> Yes, I suspect it's not the easiest path. I have a small hunch it\n> might end up paying more dividends in the future: there isn't just one\n> of these things that is regularly a thorny discussion for the same\n> reasons each time (basically \"no way to trigger this safely from\n> another backend interrupting another one at an arbitrary point\"), and\n> if we were able to generalize a solution we may have multiple wins (a\n> very obvious one in my mind is the inability of auto explain to run an\n> explain at the precise time it's most useful: when statement timeout\n> fires).\n\nYeah, trying to find a generalized solution seems like worth investing some\ntime to avoid having a bunch of CHECK_FOR_XXX() calls scattered in the code a\nfew years down the road.\n\nI might be missing something, but since we already have a ton of macro hacks,\nwhy not get another to allow CFI() overloading rather than modifying every\nsingle call? Something like that should do the trick (and CFIFlagHandler() is\njust a naive example with a function call to avoid multiple evaluation, should\nbe replaced with anything that required more than 10s thinking):\n\n#define CHECK_FOR_INTERRUPTS_0() \\\ndo { \\\n\tif (INTERRUPTS_PENDING_CONDITION()) \\\n\t\tProcessInterrupts(); \\\n} while(0)\n\n#define CHECK_FOR_INTERRUPTS_1(f) \\\ndo { \\\n\tif (INTERRUPTS_PENDING_CONDITION()) \\\n\t\tProcessInterrupts(); \\\n\t\\\n\tCFIFlagHandler(f); \\\n} while(0)\n\n#define CHECK_FOR_INTERRUPTS_X(x, f, CFI_IMPL, ...) CFI_IMPL\n\n#define CHECK_FOR_INTERRUPTS(...) \\\n\tCHECK_FOR_INTERRUPTS_X(, ##__VA_ARGS__, \\\n\t\t\t\t\t\t\tCHECK_FOR_INTERRUPTS_1(__VA_ARGS__), \\\n\t\t\t\t\t\t\tCHECK_FOR_INTERRUPTS_0(__VA_ARGS__) \\\n\t\t\t\t\t\t)\n\nWe would have to duplicate the current CFI body, but it should never really\nchange, and we shouldn't end up with more than 2 overloads anyway so I don't\nsee it being much of a problem.\n\n\n","msg_date":"Sun, 25 Feb 2024 19:30:37 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Sun, Feb 25, 2024 at 5:00 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n>\n> On Sat, Feb 24, 2024 at 08:56:41AM -0500, James Coleman wrote:\n> > On Fri, Feb 23, 2024 at 10:23 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> > >\n> > > On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> > > >\n> > > > But it doesn't have to be all or nothing right? I mean each call could say\n> > > > what the situation is like in their context, like\n> > > > CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\n> > > > slowly tag calls as needed, similarly to how we add already CFI based on users\n> > > > report.\n\nThat has some potential ...\n\n>\n> Yeah, trying to find a generalized solution seems like worth investing some\n> time to avoid having a bunch of CHECK_FOR_XXX() calls scattered in the code a\n> few years down the road.\n>\n> I might be missing something, but since we already have a ton of macro hacks,\n> why not get another to allow CFI() overloading rather than modifying every\n> single call? Something like that should do the trick (and CFIFlagHandler() is\n> just a naive example with a function call to avoid multiple evaluation, should\n> be replaced with anything that required more than 10s thinking):\n>\n> #define CHECK_FOR_INTERRUPTS_0() \\\n> do { \\\n> if (INTERRUPTS_PENDING_CONDITION()) \\\n> ProcessInterrupts(); \\\n> } while(0)\n>\n> #define CHECK_FOR_INTERRUPTS_1(f) \\\n> do { \\\n> if (INTERRUPTS_PENDING_CONDITION()) \\\n> ProcessInterrupts(); \\\n> \\\n> CFIFlagHandler(f); \\\n> } while(0)\n\n From your earlier description I thought you are talking about flags\nthat can be ORed. We need only two macros above. Why are we calling\nCFIFLagHandler() after ProcessInterrupts()? Shouldn't we pass flags to\nProcessInterrupts() itself.\n\n>\n> #define CHECK_FOR_INTERRUPTS_X(x, f, CFI_IMPL, ...) CFI_IMPL\n>\n> #define CHECK_FOR_INTERRUPTS(...) \\\n> CHECK_FOR_INTERRUPTS_X(, ##__VA_ARGS__, \\\n> CHECK_FOR_INTERRUPTS_1(__VA_ARGS__), \\\n> CHECK_FOR_INTERRUPTS_0(__VA_ARGS__) \\\n> )\n>\n> We would have to duplicate the current CFI body, but it should never really\n> change, and we shouldn't end up with more than 2 overloads anyway so I don't\n> see it being much of a problem.\n\nWhy do we need this complex macro?\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Mon, 26 Feb 2024 12:19:42 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Mon, Feb 26, 2024 at 12:19:42PM +0530, Ashutosh Bapat wrote:\n> On Sun, Feb 25, 2024 at 5:00 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> >\n> > > > On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> > > > >\n> > > > > But it doesn't have to be all or nothing right? I mean each call could say\n> > > > > what the situation is like in their context, like\n> > > > > CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\n> > > > > slowly tag calls as needed, similarly to how we add already CFI based on users\n> > > > > report.\n>\n> That has some potential ...\n>\n> >\n> > I might be missing something, but since we already have a ton of macro hacks,\n> > why not get another to allow CFI() overloading rather than modifying every\n> > single call? Something like that should do the trick (and CFIFlagHandler() is\n> > just a naive example with a function call to avoid multiple evaluation, should\n> > be replaced with anything that required more than 10s thinking):\n> >\n> > #define CHECK_FOR_INTERRUPTS_0() \\\n> > do { \\\n> > if (INTERRUPTS_PENDING_CONDITION()) \\\n> > ProcessInterrupts(); \\\n> > } while(0)\n> >\n> > #define CHECK_FOR_INTERRUPTS_1(f) \\\n> > do { \\\n> > if (INTERRUPTS_PENDING_CONDITION()) \\\n> > ProcessInterrupts(); \\\n> > \\\n> > CFIFlagHandler(f); \\\n> > } while(0)\n>\n> From your earlier description I thought you are talking about flags\n> that can be ORed. We need only two macros above. Why are we calling\n> CFIFLagHandler() after ProcessInterrupts()? Shouldn't we pass flags to\n> ProcessInterrupts() itself.\n\nYes, I'm still talking about ORed flags passed to CFI(). That CFIFlagHandler\ncall is just an example for a generalized function that would act those flags,\nrather than having it coded inside the macro.\n>\n> >\n> > #define CHECK_FOR_INTERRUPTS_X(x, f, CFI_IMPL, ...) CFI_IMPL\n> >\n> > #define CHECK_FOR_INTERRUPTS(...) \\\n> > CHECK_FOR_INTERRUPTS_X(, ##__VA_ARGS__, \\\n> > CHECK_FOR_INTERRUPTS_1(__VA_ARGS__), \\\n> > CHECK_FOR_INTERRUPTS_0(__VA_ARGS__) \\\n> > )\n> >\n> > We would have to duplicate the current CFI body, but it should never really\n> > change, and we shouldn't end up with more than 2 overloads anyway so I don't\n> > see it being much of a problem.\n>\n> Why do we need this complex macro?\n\nSo that client code can use either CHECK_FOR_INTERRUPTS() or\nCHECK_FOR_INTERRUPTS(flag) rather that transforming every single\nCHECK_FOR_INTERRUPTS() to CHECK_FOR_INTERRUPTS(0), which was Robert's\ncomplaint about this approach.\n\n\n","msg_date":"Mon, 26 Feb 2024 15:55:04 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Mon, Feb 26, 2024 at 1:25 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> >\n> > >\n> > > #define CHECK_FOR_INTERRUPTS_X(x, f, CFI_IMPL, ...) CFI_IMPL\n> > >\n> > > #define CHECK_FOR_INTERRUPTS(...) \\\n> > > CHECK_FOR_INTERRUPTS_X(, ##__VA_ARGS__, \\\n> > > CHECK_FOR_INTERRUPTS_1(__VA_ARGS__), \\\n> > > CHECK_FOR_INTERRUPTS_0(__VA_ARGS__) \\\n> > > )\n> > >\n> > > We would have to duplicate the current CFI body, but it should never really\n> > > change, and we shouldn't end up with more than 2 overloads anyway so I don't\n> > > see it being much of a problem.\n> >\n> > Why do we need this complex macro?\n>\n> So that client code can use either CHECK_FOR_INTERRUPTS() or\n> CHECK_FOR_INTERRUPTS(flag) rather that transforming every single\n> CHECK_FOR_INTERRUPTS() to CHECK_FOR_INTERRUPTS(0), which was Robert's\n> complaint about this approach.\n\nIt might be better to just create two marcos (with names like\nCHECK_FOR_INTERRUPTS() and CHECK_FOR_INTERRUPTS_SAFE()) to call\nProcessInterrupt() directly and modify ProcessInterrupts() to accept\nthe flag (and if required call CFIFlagHandler() additionally). Last\nmacro is hard to understand.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Mon, 26 Feb 2024 13:45:08 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Sun, Feb 25, 2024 at 5:00 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> Yeah, trying to find a generalized solution seems like worth investing some\n> time to avoid having a bunch of CHECK_FOR_XXX() calls scattered in the code a\n> few years down the road.\n\nI just don't really see how to do it. I suspect that every task that\nwants to run at some CFIs but not others is going to have slightly\ndifferent requirements, and we probably can't foresee what all of\nthose requirements are.\n\nSaid another way, if in the future we want to call\nDoSomethingOrOther() from the CFI handler, well then we need to know\nthat we're not already in the middle of using any subsystem that\nDoSomethingOrOther() might also try to use ... and we also need to\nknow that we're not in the middle of doing anything that's more\ncritical than DoSomethingOrOther(). But both of these are likely to\nvary in each case.\n\nEXPLAIN might be one member of a general class of things that require\ncatalog access (and thus might take locks or lwlocks, access the\ncatalogs, trigger invalidations, etc.) but it's not clear how general\nthat class really is. Also, I think if we try to do too many different\nkinds of things at CFIs, the whole thing is going to fall apart.\nYou'll end up failing to foresee some interactions, or the stack will\nget too deep, or ... something.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 26 Feb 2024 13:56:44 +0530","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2024-02-24 00:23, Robert Haas wrote:\n> On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> \n> wrote:\n>> On Fri, Feb 23, 2024 at 10:22:32AM +0530, Robert Haas wrote:\n>> > On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\n>> > > This is potentially a bit of a wild idea, but I wonder if having some\n>> > > kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\n>> > > \"normal\" as opposed to \"critical\" (using that word differently than\n>> > > the existing critical sections) would be worth it.\n>> >\n>> > It's worth considering, but the definition of \"normal\" vs. \"critical\"\n>> > might be hard to pin down. Or, we might end up with a definition that\n>> > is specific to this particular case and not generalizable to others.\n>> \n>> But it doesn't have to be all or nothing right? I mean each call \n>> could say\n>> what the situation is like in their context, like\n>> CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | \n>> GUARANTEE_WHATEVER), and\n>> slowly tag calls as needed, similarly to how we add already CFI based \n>> on users\n>> report.\n> \n> Absolutely. My gut feeling is that it's going to be simpler to pick a\n> small number of places that are safe and sufficient for this\n> particular feature and add an extra call there\n\nHmm, whether extending CHECK_FOR_INTERRUPTS() or adding extras call \ndirectly, currently I'm not sure where are the good 'places', which \ndon't give performance impact.\n\nAs attached PoC patch, I experimentally added extra calls on \nExecScanFetch() which would be less called than ExecProcNode()[1].\nWhen running sequential scan on pgbench_accounts which is on the memory, \nthere seems a performance degradation.\n\n - Executed \"select * from pgbench_accounts\" for 20 times\n - Compared the elapsed time between the patch applied and not applied \non 874d817baa160ca7e68\n - there were no heap_blks_read during the query\n - pgbench_accounts has 3000000 rows\n\n patch NOT applied:\n - average: 335.88 ms\n - max: 367.313 ms\n - min: 309.609 ms\n\n patch applied:\n - average: 342.57 ms\n - max: 380.099 ms\n - min: 324.270 ms\n\nIt would be nice if there was a place accessed once every few seconds or \nso..\n\n[1] \nhttps://www.postgresql.org/message-id/20240215185911.v4o6fo444md6a3w7%40awork3.anarazel.de\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation","msg_date":"Mon, 26 Feb 2024 21:01:45 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Mon, Feb 26, 2024 at 01:56:44PM +0530, Robert Haas wrote:\n> On Sun, Feb 25, 2024 at 5:00 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> > Yeah, trying to find a generalized solution seems like worth investing some\n> > time to avoid having a bunch of CHECK_FOR_XXX() calls scattered in the code a\n> > few years down the road.\n>\n> I just don't really see how to do it. I suspect that every task that\n> wants to run at some CFIs but not others is going to have slightly\n> different requirements, and we probably can't foresee what all of\n> those requirements are.\n>\n> Said another way, if in the future we want to call\n> DoSomethingOrOther() from the CFI handler, well then we need to know\n> that we're not already in the middle of using any subsystem that\n> DoSomethingOrOther() might also try to use ... and we also need to\n> know that we're not in the middle of doing anything that's more\n> critical than DoSomethingOrOther(). But both of these are likely to\n> vary in each case.\n>\n> EXPLAIN might be one member of a general class of things that require\n> catalog access (and thus might take locks or lwlocks, access the\n> catalogs, trigger invalidations, etc.) but it's not clear how general\n> that class really is. Also, I think if we try to do too many different\n> kinds of things at CFIs, the whole thing is going to fall apart.\n> You'll end up failing to foresee some interactions, or the stack will\n> get too deep, or ... something.\n\nI still fail to understand your point. So you say we might want to check for\nsafe condition to run explain or DoSomethingOrOther or even DoMeh, with all\ndifferent requirements. IIUC what you're saying is that we should have\nCHECK_FOR_EXPLAIN(), CHECK_FOR_DOSOMETHINGOROTHER() and CHECK_FOR_MEH()?\n\nAnd so in some code path A we could have\n\nCHECK_FOR_INTERRUPTS();\nCHECK_FOR_EXPLAIN();\n\nIn another\n\nCHECK_FOR_INTERRUPTS();\nCHECK_FOR_DOSOMETHINGOROTHER();\n\nand in one happy path\n\nCHECK_FOR_INTERRUPTS();\nCHECK_FOR_EXPLAIN();\nCHECK_FOR_DOSOMETHINGOROTHER();\nCHECK_FOR_MEH();\n\nOr that we should still have all of those but they shouldn't be anywhere close\nto a CHECK_FOR_INTERRUPTS(), or something totally different?\n\nIn the first case, I'm not sure why having CHECK_FOR_INTERRUPTS(EXPLAIN|...),\nwith a combination of flags and with the flag handling being done after\nProcessIterrupts(), would be any different apart from having less boilerplate\nlines. Similarly for the 2nd case, but relying on a single more general\nCHECK_FOR_CONDITION(EXPLAIN | ...) rather than N CHECK_FOR_XXX?\n\nIf you just want something totally different then sure.\n\n\n","msg_date":"Mon, 26 Feb 2024 21:54:41 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Mon, Feb 26, 2024 at 5:31 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\n> It would be nice if there was a place accessed once every few seconds or\n> so..\n\nI think this comment earlier from Andres deserves close attention:\n\n# If we went with something like tht approach, I think we'd have to do something\n# like redirecting node->ExecProcNode to a wrapper, presumably from within a\n# CFI. That wrapper could then implement the explain support, without slowing\n# down the normal execution path.\n\nIf this is correctly implemented, the overhead in the case where the\nfeature isn't used should be essentially zero, I believe.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 28 Feb 2024 11:48:32 +0530","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Wed, Feb 28, 2024 at 1:18 AM Robert Haas <robertmhaas@gmail.com> wrote:\n>\n> On Mon, Feb 26, 2024 at 5:31 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\n> > It would be nice if there was a place accessed once every few seconds or\n> > so..\n>\n> I think this comment earlier from Andres deserves close attention:\n>\n> # If we went with something like tht approach, I think we'd have to do something\n> # like redirecting node->ExecProcNode to a wrapper, presumably from within a\n> # CFI. That wrapper could then implement the explain support, without slowing\n> # down the normal execution path.\n>\n> If this is correctly implemented, the overhead in the case where the\n> feature isn't used should be essentially zero, I believe.\n\nIf I can rephrase this idea: it's basically \"delay this interrupt\nuntil inline to the next ExecProcNode execution\".\n\nThat seems pretty promising to me as well.\n\nRegards,\nJames Coleman\n\n\n","msg_date":"Sat, 2 Mar 2024 10:46:36 -0500","msg_from":"James Coleman <jtc331@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Sat, Mar 2, 2024 at 10:46 AM James Coleman <jtc331@gmail.com> wrote:\n> If I can rephrase this idea: it's basically \"delay this interrupt\n> until inline to the next ExecProcNode execution\".\n\nYes, but it's not just that. It also means that the code which would\nhandle the interrupt doesn't need to be called at every ExecProcNode.\nOnly when the interrupt actually arrives do we enable the code that\nhandles it.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 4 Mar 2024 13:19:45 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Fri, Feb 16, 2024 at 11:42 PM torikoshia <torikoshia@oss.nttdata.com> \nwrote:\n> I'm not so sure about the implementation now, i.e. finding the next \n> node\n> to be executed from the planstate tree, but I'm going to try this\n> approach.\n\nAttached a patch which takes this approach.\n\n- I saw no way to find the next node to be executed from the planstate \ntree, so the patch wraps all the ExecProcNode of the planstate tree at \nCHECK_FOR_INTERRUPTS().\n- To prevent overhead of this wrapped function call, unwrap it at the \nend of EXPLAIN code execution.\n- I first tried to use ExecSetExecProcNode() for wrapping, but it \n'changes' ExecProcNodeMtd of nodes, not 'adds' some process to \nExecProcNodeMtd. I'm not sure this is the right approach, but attached \npatch adds new member ExecProcNodeOriginal to PlanState to preserve \noriginal ExecProcNodeMtd.\n\nAny comments are welcomed.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation","msg_date":"Wed, 13 Mar 2024 14:28:42 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Wed, Mar 13, 2024 at 1:28 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\n> - I saw no way to find the next node to be executed from the planstate\n> tree, so the patch wraps all the ExecProcNode of the planstate tree at\n> CHECK_FOR_INTERRUPTS().\n\nI don't think it does this correctly, because some node types have\nchildren other than the left and right node. See /* special child\nplans */ in ExplainNode().\n\nBut also ... having to wrap the entire plan tree like this seems\npretty awful. I don't really like the idea of a large-scan plan\nmodification like this in the middle of the query. I also wonder\nwhether it interacts properly with JIT. But at the same time, I wonder\nhow you're supposed to avoid it.\n\nAndres, did you have some clever idea for this feature that would\navoid the need to do this?\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 13 Mar 2024 15:33:02 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On Wed, Mar 13, 2024 at 1:28 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\n>\n> On Fri, Feb 16, 2024 at 11:42 PM torikoshia <torikoshia@oss.nttdata.com>\n> wrote:\n> > I'm not so sure about the implementation now, i.e. finding the next\n> > node\n> > to be executed from the planstate tree, but I'm going to try this\n> > approach.\n>\n> Attached a patch which takes this approach.\n>\n\none minor issue.\nI understand the regress test, compare the expected outcome with\ntestrun outcome,\nbut can you enlighten me about how you check if the change you made in\ncontrib/auto_explain/t/001_auto_explain.pl is correct.\n(i am not familiar with perl).\n\ndiff --git a/src/include/commands/explain.h b/src/include/commands/explain.h\nindex cf195f1359..2d06bf297e 100644\n--- a/src/include/commands/explain.h\n+++ b/src/include/commands/explain.h\n@@ -17,6 +17,8 @@\n #include \"lib/stringinfo.h\"\n #include \"parser/parse_node.h\"\n+extern PGDLLIMPORT bool ProcessLogQueryPlanInterruptActive;\n\ndiff --git a/src/include/utils/elog.h b/src/include/utils/elog.h\nindex 054dd2bf62..3c6bd1ea7c 100644\n--- a/src/include/utils/elog.h\n+++ b/src/include/utils/elog.h\n@@ -167,6 +167,7 @@ struct Node;\n extern bool message_level_is_interesting(int elevel);\n+extern PGDLLIMPORT bool ProcessLogQueryPlanInterruptActive;\n\nutils/elog.h is already included in src/include/postgres.h.\nyou don't need to declare ProcessLogQueryPlanInterruptActive at\ninclude/commands/explain.h?\ngenerally a variable should only declare once?\n\n\n","msg_date":"Thu, 14 Mar 2024 09:02:47 +0800","msg_from":"jian he <jian.universality@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"On 2024-03-14 04:33, Robert Haas wrote:\nThanks for your review!\n\n> On Wed, Mar 13, 2024 at 1:28 AM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> - I saw no way to find the next node to be executed from the planstate\n>> tree, so the patch wraps all the ExecProcNode of the planstate tree at\n>> CHECK_FOR_INTERRUPTS().\n> \n> I don't think it does this correctly, because some node types have\n> children other than the left and right node. See /* special child\n> plans */ in ExplainNode().\n\nAgreed.\n\n> But also ... having to wrap the entire plan tree like this seems\n> pretty awful. I don't really like the idea of a large-scan plan\n> modification like this in the middle of the query.\n\nYeah, but I haven't come up with other ideas to select the appropriate \nnode to wrap.\n\n> Andres, did you have some clever idea for this feature that would\n> avoid the need to do this?\n\n\n\nOn 2024-03-14 10:02, jian he wrote:\n> On Wed, Mar 13, 2024 at 1:28 PM torikoshia <torikoshia@oss.nttdata.com> \n> wrote:\n>> \n>> On Fri, Feb 16, 2024 at 11:42 PM torikoshia \n>> <torikoshia@oss.nttdata.com>\n>> wrote:\n>> > I'm not so sure about the implementation now, i.e. finding the next\n>> > node\n>> > to be executed from the planstate tree, but I'm going to try this\n>> > approach.\n>> \n>> Attached a patch which takes this approach.\n>> \n> \n> one minor issue.\n> I understand the regress test, compare the expected outcome with\n> testrun outcome,\n> but can you enlighten me about how you check if the change you made in\n> contrib/auto_explain/t/001_auto_explain.pl is correct.\n> (i am not familiar with perl).\n\n$pg_log_plan_query_output saves the output plan of pg_log_query_plan() \nand $auto_explain_output saves the output plan of auto_explain.\nThe test checks the both are the same using cmp_ok().\n\nDid I answer your question?\n\n> diff --git a/src/include/commands/explain.h \n> b/src/include/commands/explain.h\n> index cf195f1359..2d06bf297e 100644\n> --- a/src/include/commands/explain.h\n> +++ b/src/include/commands/explain.h\n> @@ -17,6 +17,8 @@\n> #include \"lib/stringinfo.h\"\n> #include \"parser/parse_node.h\"\n> +extern PGDLLIMPORT bool ProcessLogQueryPlanInterruptActive;\n> \n> diff --git a/src/include/utils/elog.h b/src/include/utils/elog.h\n> index 054dd2bf62..3c6bd1ea7c 100644\n> --- a/src/include/utils/elog.h\n> +++ b/src/include/utils/elog.h\n> @@ -167,6 +167,7 @@ struct Node;\n> extern bool message_level_is_interesting(int elevel);\n> +extern PGDLLIMPORT bool ProcessLogQueryPlanInterruptActive;\n> \n> utils/elog.h is already included in src/include/postgres.h.\n> you don't need to declare ProcessLogQueryPlanInterruptActive at\n> include/commands/explain.h?\n> generally a variable should only declare once?\n\nYeah, this declaration is not necessary and we should add include \ncommands/explain.h to src/backend/access/transam/xact.c.\n\n-- \nRegards,\n\n--\nAtsushi Torikoshi\nNTT DATA Group Corporation\n\n\n","msg_date":"Mon, 18 Mar 2024 14:39:48 +0900","msg_from":"torikoshia <torikoshia@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"},{"msg_contents":"Hi,\n\nOn 2024-03-13 15:33:02 -0400, Robert Haas wrote:\n> But also ... having to wrap the entire plan tree like this seems\n> pretty awful. I don't really like the idea of a large-scan plan\n> modification like this in the middle of the query.\n\nIt's not great. But I also don't really see an alternative with this approach.\n\nI guess we could invent a new CFI version that gets the current PlanState and\nuse that in all of src/backend/executor/node* and pass the PlanState to that -\nbut then we could just as well just directly process the interrupt there.\n\n\n> I also wonder whether it interacts properly with JIT.\n\nI don't think there's a problem unless somebody invests a lot of time in\nJITing much more of the query. Which will require a lot more work, basically\nredesigning the executor...\n\n\n\n> Andres, did you have some clever idea for this feature that would\n> avoid the need to do this?\n\nNo. I think it's acceptable though.\n\nHowever it might be worth inventing an executor tree walker in a preliminary\nstep. We have already quite a few switches over all plan nodes, which we could\nlargely replace with a helper.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 25 Mar 2024 19:34:47 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: RFC: Logging plan of the running query"}],"string":"[\n {\n \"msg_contents\": \"Hi!\\n\\nI'm here to answer your questions about contrib/pg_query_state.\\n>> I only took a quick look at pg_query_state, I have some questions.\\n\\n>> pg_query_state seems using shm_mq to expose the plan information, but\\n>> there was a discussion that this kind of architecture would be tricky\\n>> to do properly [1].\\n>> Does pg_query_state handle difficulties listed on the discussion?\\n> [1] \\n> https://www.postgresql.org/message-id/9a50371e15e741e295accabc72a41df1%40oss.nttdata.com\\n\\nI doubt that it was the right link.\\nBut on the topic I will say that extension really use shared memory, \\ninteraction is implemented by sending / receiving messages. This \\narchitecture provides the required reliability and convenience.\\n\\n>> It seems the caller of the pg_query_state() has to wait until the\\n>> target process pushes the plan information into shared memory, can it\\n>> lead to deadlock situations?\\n>> I came up with this question because when trying to make a view for\\n>> memory contexts of other backends, we encountered deadlock situations.\\n>> After all, we gave up view design and adopted sending signal and\\n>> logging.\\n> \\n> Discussion at the following URL.\\n> https://www.postgresql.org/message-id/9a50371e15e741e295accabc72a41df1%40oss.nttdata.com\\n\\nBefore extracting information about side process we check its state. \\nInformation will only be retrieved for a process willing to provide it. \\nOtherwise, we will receive an error message about impossibility of \\ngetting query execution statistics + process status. Also checking fact \\nof extracting your own status exists. This is even verified in tests.\\n\\nThanks for your attention.\\nJust in case, I am ready to discuss this topic in more detail.\\n\\nAbout overhead:\\n> I haven't measured it yet, but I believe that the overhead for backends\\n> which are not called pg_log_current_plan() would be slight since the\\n> patch just adds the logic for saving QueryDesc on ExecutorRun().\\n> The overhead for backends which is called pg_log_current_plan() might\\n> not slight, but since the target process are assumed dealing with\\n> long-running query and the user want to know its plan, its overhead\\n> would be worth the cost.\\nI think it would be useful for us to have couple of examples with a \\ndifferent number of rows compared to using without this functionality.\\n\\nHope this helps.\\n\\n-- \\nEkaterina Sokolova\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\\n\\n\\n\",\n \"msg_date\": \"Tue, 02 Nov 2021 14:32:59 +0300\",\n \"msg_from\": \"Ekaterina Sokolova <e.sokolova@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"This patch no longer applies on top of HEAD, please submit a rebased version.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 3 Nov 2021 13:48:51 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2021-11-02 20:32, Ekaterina Sokolova wrote:\\nThanks for your response!\\n\\n> Hi!\\n> \\n> I'm here to answer your questions about contrib/pg_query_state.\\n>>> I only took a quick look at pg_query_state, I have some questions.\\n> \\n>>> pg_query_state seems using shm_mq to expose the plan information, but\\n>>> there was a discussion that this kind of architecture would be tricky\\n>>> to do properly [1].\\n>>> Does pg_query_state handle difficulties listed on the discussion?\\n>> [1] \\n>> https://www.postgresql.org/message-id/9a50371e15e741e295accabc72a41df1%40oss.nttdata.com\\n> \\n> I doubt that it was the right link.\\n\\nSorry for make you confused, here is the link.\\n\\n \\nhttps://www.postgresql.org/message-id/CA%2BTgmobkpFV0UB67kzXuD36--OFHwz1bs%3DL_6PZbD4nxKqUQMw%40mail.gmail.com\\n\\n> But on the topic I will say that extension really use shared memory,\\n> interaction is implemented by sending / receiving messages. This\\n> architecture provides the required reliability and convenience.\\n\\nAs described in the link, using shared memory for this kind of work \\nwould need DSM and It'd be also necessary to exchange information \\nbetween requestor and responder.\\n\\nFor example, when I looked at a little bit of pg_query_state code, it \\nlooks like the size of the queue is fixed at QUEUE_SIZE, and I wonder \\nhow plans that exceed QUEUE_SIZE are handled.\\n\\n>>> It seems the caller of the pg_query_state() has to wait until the\\n>>> target process pushes the plan information into shared memory, can it\\n>>> lead to deadlock situations?\\n>>> I came up with this question because when trying to make a view for\\n>>> memory contexts of other backends, we encountered deadlock \\n>>> situations.\\n>>> After all, we gave up view design and adopted sending signal and\\n>>> logging.\\n>> \\n>> Discussion at the following URL.\\n>> https://www.postgresql.org/message-id/9a50371e15e741e295accabc72a41df1%40oss.nttdata.com\\n> \\n> Before extracting information about side process we check its state.\\n> Information will only be retrieved for a process willing to provide\\n> it. Otherwise, we will receive an error message about impossibility of\\n> getting query execution statistics + process status. Also checking\\n> fact of extracting your own status exists. This is even verified in\\n> tests.\\n> \\n> Thanks for your attention.\\n> Just in case, I am ready to discuss this topic in more detail.\\n\\nI imagined the following procedure.\\nDoes it cause dead lock in pg_query_state?\\n\\n- session1\\nBEGIN; TRUNCATE t;\\n\\n- session2\\nBEGIN; TRUNCATE t; -- wait\\n\\n- session1\\nSELECT * FROM pg_query_state(<pid of session>); -- wait and dead locked?\\n\\n> About overhead:\\n>> I haven't measured it yet, but I believe that the overhead for \\n>> backends\\n>> which are not called pg_log_current_plan() would be slight since the\\n>> patch just adds the logic for saving QueryDesc on ExecutorRun().\\n>> The overhead for backends which is called pg_log_current_plan() might\\n>> not slight, but since the target process are assumed dealing with\\n>> long-running query and the user want to know its plan, its overhead\\n>> would be worth the cost.\\n> I think it would be useful for us to have couple of examples with a\\n> different number of rows compared to using without this functionality.\\n\\nDo you have any expectaion that the number of rows would affect the \\nperformance of this functionality?\\nThis patch adds some codes to ExecutorRun(), but I thought the number of \\nrows would not give impact on the performance.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Thu, 04 Nov 2021 21:49:21 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi!\\n\\nYou forgot my last fix to build correctly on Mac. I have added it.\\n\\nAbout our discussion of pg_query_state:\\n\\ntorikoshia писал 2021-11-04 15:49:\\n>> I doubt that it was the right link.\\n> Sorry for make you confused, here is the link.\\n> https://www.postgresql.org/message-id/CA%2BTgmobkpFV0UB67kzXuD36--OFHwz1bs%3DL_6PZbD4nxKqUQMw%40mail.gmail.com\\n\\nThank you. I'll see it soon.\\n\\n> I imagined the following procedure.\\n> Does it cause dead lock in pg_query_state?\\n> \\n> - session1\\n> BEGIN; TRUNCATE t;\\n> \\n> - session2\\n> BEGIN; TRUNCATE t; -- wait\\n> \\n> - session1\\n> SELECT * FROM pg_query_state(<pid of session>); -- wait and dead \\n> locked?\\n\\nAs I know, pg_query_state use non-blocking read and write. I have wrote \\nfew tests trying to deadlock it (on 14 version), but all finished \\ncorrectly.\\n\\nHave a nice day. Please feel free to contact me if you need any further \\ninformation.\\n\\n-- \\nEkaterina Sokolova\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\",\n \"msg_date\": \"Wed, 17 Nov 2021 16:44:51 +0300\",\n \"msg_from\": \"Ekaterina Sokolova <e.sokolova@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2021-11-17 22:44, Ekaterina Sokolova wrote:\\n> Hi!\\n> \\n> You forgot my last fix to build correctly on Mac. I have added it.\\n\\nThanks for the notification!\\nSince the patch could not be applied to the HEAD anymore, I also updated \\nit.\\n\\n> \\n> About our discussion of pg_query_state:\\n> \\n> torikoshia писал 2021-11-04 15:49:\\n>>> I doubt that it was the right link.\\n>> Sorry for make you confused, here is the link.\\n>> https://www.postgresql.org/message-id/CA%2BTgmobkpFV0UB67kzXuD36--OFHwz1bs%3DL_6PZbD4nxKqUQMw%40mail.gmail.com\\n> \\n> Thank you. I'll see it soon.\\n> \\n>> I imagined the following procedure.\\n>> Does it cause dead lock in pg_query_state?\\n>> \\n>> - session1\\n>> BEGIN; TRUNCATE t;\\n>> \\n>> - session2\\n>> BEGIN; TRUNCATE t; -- wait\\n>> \\n>> - session1\\n>> SELECT * FROM pg_query_state(<pid of session>); -- wait and dead \\n>> locked?\\n> \\n> As I know, pg_query_state use non-blocking read and write. I have\\n> wrote few tests trying to deadlock it (on 14 version), but all\\n> finished correctly.\\n> \\n> Have a nice day. Please feel free to contact me if you need any\\n> further information.\\n\\nThanks for your information and help!\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Fri, 26 Nov 2021 12:39:48 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2021-11-26 12:39, torikoshia wrote:\\n> Since the patch could not be applied to the HEAD anymore, I also \\n> updated it.\\n\\nUpdated the patch for fixing compiler warning about the format on \\nwindows.\\n\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Fri, 07 Jan 2022 14:30:37 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2022-01-07 14:30, torikoshia wrote:\\n\\n> Updated the patch for fixing compiler warning about the format on \\n> windows.\\n\\nI got another compiler warning, updated the patch again.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Fri, 07 Jan 2022 20:58:19 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"\\r\\n\\r\\nOn 2022/01/07 20:58, torikoshia wrote:\\r\\n> On 2022-01-07 14:30, torikoshia wrote:\\r\\n> \\r\\n>> Updated the patch for fixing compiler warning about the format on windows.\\r\\n> \\r\\n> I got another compiler warning, updated the patch again.\\r\\n\\r\\nThanks for updating the patch!\\r\\n\\r\\nI ran the following query every 0.1s by using \\\\watch psql command from three different sessions while make installcheck test was running.\\r\\n\\r\\n SELECT pg_log_query_plan(pid) FROM pg_stat_activity;\\r\\n\\r\\n\\r\\nAnd then I got the segmentation fault as follows.\\r\\n\\r\\n2022-01-07 21:40:32 JST [postmaster] LOG: server process (PID 51017) was terminated by signal 11: Segmentation fault: 11\\r\\n2022-01-07 21:40:32 JST [postmaster] DETAIL: Failed process was running: select description, (test_conv(inbytes, 'utf8', 'utf8')).* from utf8_verification_inputs;\\r\\n2022-01-07 21:40:32 JST [postmaster] LOG: terminating any other active server processes\\r\\n2022-01-07 21:40:32 JST [postmaster] LOG: all server processes terminated; reinitializing\\r\\n\\r\\n\\r\\nThe backtrace I got from the core file was:\\r\\n\\r\\n(lldb) target create --core \\\"/cores/core.51017\\\"\\r\\nCore file '/cores/core.51017' (x86_64) was loaded.\\r\\n(lldb) bt\\r\\n* thread #1, stop reason = signal SIGSTOP\\r\\n * frame #0: 0x00007fff20484552 libsystem_platform.dylib`_platform_strlen + 18\\r\\n frame #1: 0x000000011076e36c postgres`dopr(target=0x00007ffedfd1b1d8, format=\\\"\\\\n\\\", args=0x00007ffedfd1b450) at snprintf.c:444:20\\r\\n frame #2: 0x000000011076e133 postgres`pg_vsnprintf(str=\\\"Query Text: \\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\"..., count=1024, fmt=\\\"%s: %s\\\\n\\\", args=0x00007ffedfd1b450) at snprintf.c:195:2\\r\\n frame #3: 0x000000011075e51d postgres`pvsnprintf(buf=\\\"Query Text: \\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\"..., len=1024, fmt=\\\"%s: %s\\\\n\\\", args=0x00007ffedfd1b450) at psprintf.c:110:13\\r\\n frame #4: 0x000000011076052d postgres`appendStringInfoVA(str=0x00007fe56804bec0, fmt=\\\"%s: %s\\\\n\\\", args=0x00007ffedfd1b450) at stringinfo.c:149:13\\r\\n frame #5: 0x0000000110760449 postgres`appendStringInfo(str=0x00007fe56804bec0, fmt=\\\"%s: %s\\\\n\\\") at stringinfo.c:103:12\\r\\n frame #6: 0x000000011016f6cf postgres`ExplainProperty(qlabel=\\\"Query Text\\\", unit=0x0000000000000000, value=\\\"\\\", numeric=false, es=0x00007fe56804be28) at explain.c:4390:5\\r\\n frame #7: 0x000000011016eace postgres`ExplainPropertyText(qlabel=\\\"Query Text\\\", value=\\\"\\\", es=0x00007fe56804be28) at explain.c:4435:2\\r\\n frame #8: 0x000000011016ea97 postgres`ExplainQueryText(es=0x00007fe56804be28, queryDesc=0x00007fe568022320) at explain.c:969:3\\r\\n frame #9: 0x000000011016faeb postgres`ProcessLogCurrentPlanInterrupt at explain.c:5009:2\\r\\n frame #10: 0x00000001104dda84 postgres`ProcessInterrupts at postgres.c:3373:3\\r\\n frame #11: 0x0000000110c81c57 plpgsql.so`exec_stmts(estate=0x00007ffedfd1bab8, stmts=0x00007fe568050e00) at pl_exec.c:1994:3\\r\\n frame #12: 0x0000000110c80a78 plpgsql.so`exec_stmt_block(estate=0x00007ffedfd1bab8, block=0x00007fe56801c570) at pl_exec.c:1777:9\\r\\n frame #13: 0x0000000110c81c91 plpgsql.so`exec_stmts(estate=0x00007ffedfd1bab8, stmts=0x00007fe56801c5c8) at pl_exec.c:1999:10\\r\\n frame #14: 0x0000000110c80f21 plpgsql.so`exec_stmt_block(estate=0x00007ffedfd1bab8, block=0x00007fe56801c678) at pl_exec.c:1926:8\\r\\n frame #15: 0x0000000110c7ed92 plpgsql.so`exec_toplevel_block(estate=0x00007ffedfd1bab8, block=0x00007fe56801c678) at pl_exec.c:1617:7\\r\\n frame #16: 0x0000000110c7d3e4 plpgsql.so`plpgsql_exec_function(func=0x00007fe56a00aa20, fcinfo=0x00007fe56803c888, simple_eval_estate=0x0000000000000000, simple_eval_resowner=0x0000000000000000, procedure_resowner=0x0000000000000000, atomic=true) at pl_exec.c:611:7\\r\\n frame #17: 0x0000000110c9bf12 plpgsql.so`plpgsql_call_handler(fcinfo=0x00007fe56803c888) at pl_handler.c:277:13\\r\\n frame #18: 0x0000000110228b06 postgres`ExecInterpExpr(state=0x00007fe56803bf88, econtext=0x00007fe56803ba70, isnull=0x00007ffedfd1bf7f) at execExprInterp.c:725:8\\r\\n frame #19: 0x000000011024858b postgres`ExecEvalExprSwitchContext(state=0x00007fe56803bf88, econtext=0x00007fe56803ba70, isNull=0x00007ffedfd1bf7f) at executor.h:339:13\\r\\n frame #20: 0x00000001102482aa postgres`ExecProject(projInfo=0x00007fe56803bf80) at executor.h:373:9\\r\\n frame #21: 0x0000000110247e2f postgres`ExecScan(node=0x00007fe56803b958, accessMtd=(postgres`SeqNext at nodeSeqscan.c:51), recheckMtd=(postgres`SeqRecheck at nodeSeqscan.c:90)) at execScan.c:238:12\\r\\n frame #22: 0x000000011028c575 postgres`ExecSeqScan(pstate=0x00007fe56803b958) at nodeSeqscan.c:112:9\\r\\n frame #23: 0x000000011023bbe2 postgres`ExecProcNode(node=0x00007fe56803b958) at executor.h:257:9\\r\\n frame #24: 0x0000000110237471 postgres`ExecutePlan(estate=0x00007fe56803b720, planstate=0x00007fe56803b958, use_parallel_mode=false, operation=CMD_SELECT, sendTuples=true, numberTuples=0, direction=ForwardScanDirection, dest=0x00007fe568046ac8, execute_once=true) at execMain.c:1561:10\\r\\n frame #25: 0x0000000110237320 postgres`standard_ExecutorRun(queryDesc=0x00007fe568037320, direction=ForwardScanDirection, count=0, execute_once=true) at execMain.c:371:3\\r\\n frame #26: 0x00000001102370b8 postgres`ExecutorRun(queryDesc=0x00007fe568037320, direction=ForwardScanDirection, count=0, execute_once=true) at execMain.c:313:3\\r\\n frame #27: 0x00000001104e6bdb postgres`PortalRunSelect(portal=0x00007fe567825920, forward=true, count=0, dest=0x00007fe568046ac8) at pquery.c:921:4\\r\\n frame #28: 0x00000001104e65ba postgres`PortalRun(portal=0x00007fe567825920, count=9223372036854775807, isTopLevel=true, run_once=true, dest=0x00007fe568046ac8, altdest=0x00007fe568046ac8, qc=0x00007ffedfd1c350) at pquery.c:765:18\\r\\n frame #29: 0x00000001104e1861 postgres`exec_simple_query(query_string=\\\"select description, (test_conv(inbytes, 'utf8', 'utf8')).* from utf8_verification_inputs;\\\") at postgres.c:1216:10\\r\\n frame #30: 0x00000001104e099c postgres`PostgresMain(dbname=\\\"regression\\\", username=\\\"postgres\\\") at postgres.c:4505:7\\r\\n frame #31: 0x00000001103ee2a2 postgres`BackendRun(port=0x00007fe567704080) at postmaster.c:4594:2\\r\\n frame #32: 0x00000001103ed868 postgres`BackendStartup(port=0x00007fe567704080) at postmaster.c:4322:3\\r\\n frame #33: 0x00000001103ec64c postgres`ServerLoop at postmaster.c:1802:7\\r\\n frame #34: 0x00000001103e9e29 postgres`PostmasterMain(argc=3, argv=0x00007fe567405e70) at postmaster.c:1474:11\\r\\n frame #35: 0x00000001102c8b69 postgres`main(argc=3, argv=0x00007fe567405e70) at main.c:198:3\\r\\n frame #36: 0x00007fff2045cf3d libdyld.dylib`start + 1\\r\\n\\r\\nRegards,\\r\\n\\r\\n-- \\r\\nFujii Masao\\r\\nAdvanced Computing Technology Center\\r\\nResearch and Development Headquarters\\r\\nNTT DATA CORPORATION\\r\\n\",\n \"msg_date\": \"Fri, 7 Jan 2022 21:54:31 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Fri, Jan 07, 2022 at 09:54:31PM +0900, Fujii Masao wrote:\\n> \\n> On 2022/01/07 20:58, torikoshia wrote:\\n> > On 2022-01-07 14:30, torikoshia wrote:\\n> > \\n> > > Updated the patch for fixing compiler warning about the format on windows.\\n> > \\n> > I got another compiler warning, updated the patch again.\\n> \\n> Thanks for updating the patch!\\n> \\n> I ran the following query every 0.1s by using \\\\watch psql command from three different sessions while make installcheck test was running.\\n> \\n> SELECT pg_log_query_plan(pid) FROM pg_stat_activity;\\n> \\n> \\n> And then I got the segmentation fault as follows.\\n> \\n> 2022-01-07 21:40:32 JST [postmaster] LOG: server process (PID 51017) was terminated by signal 11: Segmentation fault: 11\\n> 2022-01-07 21:40:32 JST [postmaster] DETAIL: Failed process was running: select description, (test_conv(inbytes, 'utf8', 'utf8')).* from utf8_verification_inputs;\\n> 2022-01-07 21:40:32 JST [postmaster] LOG: terminating any other active server processes\\n> 2022-01-07 21:40:32 JST [postmaster] LOG: all server processes terminated; reinitializing\\n> \\n> \\n> The backtrace I got from the core file was:\\n> \\n> (lldb) target create --core \\\"/cores/core.51017\\\"\\n> Core file '/cores/core.51017' (x86_64) was loaded.\\n> (lldb) bt\\n> * thread #1, stop reason = signal SIGSTOP\\n> * frame #0: 0x00007fff20484552 libsystem_platform.dylib`_platform_strlen + 18\\n> frame #1: 0x000000011076e36c postgres`dopr(target=0x00007ffedfd1b1d8, format=\\\"\\\\n\\\", args=0x00007ffedfd1b450) at snprintf.c:444:20\\n> frame #2: 0x000000011076e133 postgres`pg_vsnprintf(str=\\\"Query Text: \\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\"..., count=1024, fmt=\\\"%s: %s\\\\n\\\", args=0x00007ffedfd1b450) at snprintf.c:195:2\\n\\nOn top of the visibly use-after-pfree memory issue, the patch doesn't apply\\nanymore:\\n\\nhttp://cfbot.cputube.org/patch_36_3142.log\\n=== Applying patches on top of PostgreSQL commit ID a18b6d2dc288dfa6e7905ede1d4462edd6a8af47 ===\\n=== applying patch ./v16-0001-log-running-query-plan.patch\\n[...]\\npatching file src/test/regress/expected/misc_functions.out\\nHunk #1 FAILED at 134.\\nHunk #2 succeeded at 163 (offset 7 lines).\\nHunk #3 succeeded at 180 (offset 7 lines).\\n1 out of 3 hunks FAILED -- saving rejects to file src/test/regress/expected/misc_functions.out.rej\\npatching file src/test/regress/sql/misc_functions.sql\\nHunk #1 FAILED at 31.\\n1 out of 1 hunk FAILED -- saving rejects to file src/test/regress/sql/misc_functions.sql.rej\\n\\nI'm switching the patch to Waiting on Author.\\n\\n\\n\",\n \"msg_date\": \"Fri, 14 Jan 2022 14:38:46 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2022-01-14 15:38, Julien Rouhaud wrote:\\n> Hi,\\n> \\n> On Fri, Jan 07, 2022 at 09:54:31PM +0900, Fujii Masao wrote:\\n\\n>> I ran the following query every 0.1s by using \\\\watch psql command from \\n>> three different sessions while make installcheck test was running.\\n>> \\n>> SELECT pg_log_query_plan(pid) FROM pg_stat_activity;\\n>> \\n>> \\n>> And then I got the segmentation fault as follows.\\n>> \\n>> 2022-01-07 21:40:32 JST [postmaster] LOG: server process (PID 51017) \\n>> was terminated by signal 11: Segmentation fault: 11\\n>> 2022-01-07 21:40:32 JST [postmaster] DETAIL: Failed process was \\n>> running: select description, (test_conv(inbytes, 'utf8', 'utf8')).* \\n>> from utf8_verification_inputs;\\n>> 2022-01-07 21:40:32 JST [postmaster] LOG: terminating any other \\n>> active server processes\\n>> 2022-01-07 21:40:32 JST [postmaster] LOG: all server processes \\n>> terminated; reinitializing\\n>> \\n>> \\n>> The backtrace I got from the core file was:\\n>> \\n>> (lldb) target create --core \\\"/cores/core.51017\\\"\\n>> Core file '/cores/core.51017' (x86_64) was loaded.\\n>> (lldb) bt\\n>> * thread #1, stop reason = signal SIGSTOP\\n>> * frame #0: 0x00007fff20484552 \\n>> libsystem_platform.dylib`_platform_strlen + 18\\n>> frame #1: 0x000000011076e36c \\n>> postgres`dopr(target=0x00007ffedfd1b1d8, format=\\\"\\\\n\\\", \\n>> args=0x00007ffedfd1b450) at snprintf.c:444:20\\n>> frame #2: 0x000000011076e133 postgres`pg_vsnprintf(str=\\\"Query \\n>> Text: \\n>>\\n\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x\\n7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\n\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x\\n7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\n\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\\x7f\\\"..., \\n>> count=1024, fmt=\\\"%s: %s\\\\n\\\", args=0x00007ffedfd1b450) at \\n>> snprintf.c:195:2\\n\\nThanks for testing!\\nI could reproduce the situation and confirmed almost the same backtrace.\\n\\nThe cause seems that the previous patch didn't do cleanup when the \\ntransactions were aborted. It leaded segmentation faults when \\nreferencing ActiveQueryDesc-> after the transaction were aborted.\\n\\nThe attached patch added cleanups when transactions and subtransactions \\nare aborted.\\n\\nBTW, the latest patch for pg_log_backtrace() added a new function for \\ndeduplicating codes for pg_log_*()[1].\\nIt seems better to use it for pg_log_query_plan() after the patch is \\nmerged.\\n\\n\\n[1] \\nhttps://www.postgresql.org/message-id/flat/20211115194251.GP17618%40telsasoft.com#4e681d04c78867b2cce44b3cb1791bf5\\n duplicate code to a new function\\nCheckPostgresProcessId which can be used by pg_log_backtrace,\\n\\n> On top of the visibly use-after-pfree memory issue, the patch doesn't \\n> apply\\n\\nThanks for letting me know!\\nAs mentioned above, I updated the patch.\\n\\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Mon, 24 Jan 2022 14:33:05 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2022/01/24 14:33, torikoshia wrote:\\n> As mentioned above, I updated the patch.\\n\\nThanks for updating the patch!\\n\\nHere are another review comments:\\n\\n+LOG: plan of the query running on backend with PID 17793 is:\\n\\nThis seems not the same as what actually logged.\\n\\n+\\t\\tereport(WARNING,\\n+\\t\\t\\t\\t(errmsg(\\\"PID %d is not a PostgreSQL server process\\\", pid)));\\n\\nLike commit 7fa945b857 changed, this warning message should be \\\"PID %d is not a PostgreSQL backend process\\\"?\\n\\n+\\tif (SendProcSignal(pid, PROCSIG_LOG_CURRENT_PLAN, InvalidBackendId) < 0)\\n\\nproc->backendId should be specified instead of InvalidBackendId, to speed up the processing in SendProcSignal()?\\n\\n+\\tPROCSIG_LOG_CURRENT_PLAN, /* ask backend to log plan of the current query */\\n+volatile sig_atomic_t LogCurrentPlanPending = false;\\n+extern void HandleLogCurrentPlanInterrupt(void);\\n+extern void ProcessLogCurrentPlanInterrupt(void);\\n\\nIsn't it better to use the names that are more consistent with the function name, i.e., pg_log_query_plan? For example, PROCSIG_LOG_QUERY_PLAN instead of PROCSIG_LOG_CURRENT_PLAN?\\n\\n+\\t\\tereport(LOG_SERVER_ONLY,\\n+\\t\\t\\t\\terrmsg(\\\"backend with PID %d is not running a query\\\",\\n+\\t\\t\\t\\t\\tMyProcPid));\\n\\nerrhidestmt(true) and errhidecontext(true) need to be added, don't they? Otherwise, for example, if pg_log_query_plan() is executed after debug_query_string is set but before ActiveQueryDesc is set, STATEMENT message would be output even though the message saying \\\"not running a query\\\" is output. Which seems confusing.\\n\\n+\\thash_seq_init(&status, GetLockMethodLocalHash());\\n+\\twhile ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)\\n+\\t{\\n+\\t\\tif (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\\n\\nWhy did you use the search for local lock hash instead of IsPageLockHeld flag variable, to check whether a page lock is held or not? Because there is the corner case where the interrupt is processed after the local lock is registered into the hash but before IsPageLockHeld is enabled?\\n\\nThere is the case where the request to log a query plan is skipped even while the target backend is running a query. If this happens, users can just retry pg_log_query_plan(). These things should be documented?\\n\\n+\\t\\t\\tereport(LOG_SERVER_ONLY,\\n+\\t\\t\\t\\terrmsg(\\\"backend with PID %d is holding a page lock. Try again\\\",\\n+\\t\\t\\t\\t\\tMyProcPid));\\n\\nIt seems more proper to output this message in DETAIL or HINT, instead. So how about something like the following messages?\\n\\nLOG: could not log the query plan\\nDETAIL: query plan cannot be logged while page level lock is being held\\nHINT: Try pg_log_query_plan() after a few ....\\n\\nRegards,\\n\\n-- \\nFujii Masao\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Thu, 27 Jan 2022 20:18:37 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2022-01-27 20:18, Fujii Masao wrote:\\n> Here are another review comments:\\n\\nThanks for reviewing!\\n\\n> +LOG: plan of the query running on backend with PID 17793 is:\\n> \\n> This seems not the same as what actually logged.\\n\\nModified.\\n\\n> +\\t\\tereport(WARNING,\\n> +\\t\\t\\t\\t(errmsg(\\\"PID %d is not a PostgreSQL server process\\\", pid)));\\n> \\n> Like commit 7fa945b857 changed, this warning message should be \\\"PID %d\\n> is not a PostgreSQL backend process\\\"?\\n\\nModified.\\n\\n> +\\tif (SendProcSignal(pid, PROCSIG_LOG_CURRENT_PLAN, InvalidBackendId) < \\n> 0)\\n> \\n> proc->backendId should be specified instead of InvalidBackendId, to\\n> speed up the processing in SendProcSignal()?\\n\\nAgreed. Modified.\\n\\n> +\\tPROCSIG_LOG_CURRENT_PLAN, /* ask backend to log plan of the current \\n> query */\\n> +volatile sig_atomic_t LogCurrentPlanPending = false;\\n> +extern void HandleLogCurrentPlanInterrupt(void);\\n> +extern void ProcessLogCurrentPlanInterrupt(void);\\n> \\n> Isn't it better to use the names that are more consistent with the\\n> function name, i.e., pg_log_query_plan? For example,\\n> PROCSIG_LOG_QUERY_PLAN instead of PROCSIG_LOG_CURRENT_PLAN?\\n\\nAgreed.\\nI removed 'current' from the variable and function names and used \\n'query' instead.\\n\\n> +\\t\\tereport(LOG_SERVER_ONLY,\\n> +\\t\\t\\t\\terrmsg(\\\"backend with PID %d is not running a query\\\",\\n> +\\t\\t\\t\\t\\tMyProcPid));\\n> \\n> errhidestmt(true) and errhidecontext(true) need to be added, don't\\n> they? Otherwise, for example, if pg_log_query_plan() is executed after\\n> debug_query_string is set but before ActiveQueryDesc is set, STATEMENT\\n> message would be output even though the message saying \\\"not running a\\n> query\\\" is output. Which seems confusing.\\n\\nAgreed. Added errhidestmt(true) and errhidecontext(true).\\n\\n> +\\thash_seq_init(&status, GetLockMethodLocalHash());\\n> +\\twhile ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)\\n> +\\t{\\n> +\\t\\tif (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\\n> \\n> Why did you use the search for local lock hash instead of\\n> IsPageLockHeld flag variable, to check whether a page lock is held or\\n> not? Because there is the corner case where the interrupt is processed\\n> after the local lock is registered into the hash but before\\n> IsPageLockHeld is enabled?\\n\\nAs far as I read CheckAndSetLockHeld(), IsPageLockHeld can be used only \\nwhen USE_ASSERT_CHECKING is enabled.\\nSince removing USE_ASSERT_CHECKING from CheckAndSetLockHeld() would give \\nperformance impact on every granting/removing local lock, I used the \\nsearch for local local hash.\\n\\n> There is the case where the request to log a query plan is skipped\\n> even while the target backend is running a query. If this happens,\\n> users can just retry pg_log_query_plan(). These things should be\\n> documented?\\n\\nAgreed.\\nAdded following:\\n\\n + Note that there is the case where the request to log a query\\n + plan is skipped even while the target backend is running a\\n + query due to lock conflict avoidance.\\n + If this happens, users can just retry pg_log_query_plan().\\n\\n|\\n\\n> +\\t\\t\\tereport(LOG_SERVER_ONLY,\\n> +\\t\\t\\t\\terrmsg(\\\"backend with PID %d is holding a page lock. Try again\\\",\\n> +\\t\\t\\t\\t\\tMyProcPid));\\n> \\n> It seems more proper to output this message in DETAIL or HINT,\\n> instead. So how about something like the following messages?\\n> \\n> LOG: could not log the query plan\\n> DETAIL: query plan cannot be logged while page level lock is being held\\n> HINT: Try pg_log_query_plan() after a few ....\\n\\nAgreed.\\nI felt the HINT message 'after a few ...' is difficult to describe, and \\nwrote as below.\\n\\n| HINT: Retrying pg_log_query_plan() might succeed since the lock \\nduration of page level locks are usually short\\n\\nHow do you think?\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Fri, 28 Jan 2022 17:45:49 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2022/01/28 17:45, torikoshia wrote:\\n>> There is the case where the request to log a query plan is skipped\\n>> even while the target backend is running a query. If this happens,\\n>> users can just retry pg_log_query_plan(). These things should be\\n>> documented?\\n> \\n> Agreed.\\n> Added following:\\n> \\n> + Note that there is the case where the request to log a query\\n> + plan is skipped even while the target backend is running a\\n> + query due to lock conflict avoidance.\\n> + If this happens, users can just retry pg_log_query_plan().\\n\\nThis may cause users to misunderstand that pg_log_query_plan() fails while the target backend is holding *any* locks? Isn't it better to mention \\\"page-level locks\\\", instead? So how about the following?\\n\\n--------------------------\\nNote that the request to log the query plan will be ignored if it's received during a short period while the target backend is holding a page-level lock.\\n--------------------------\\n\\n\\n>> + ereport(LOG_SERVER_ONLY,\\n>> + errmsg(\\\"backend with PID %d is holding a page lock. Try again\\\",\\n>> + MyProcPid));\\n>>\\n>> It seems more proper to output this message in DETAIL or HINT,\\n>> instead. So how about something like the following messages?\\n>>\\n>> LOG: could not log the query plan\\n>> DETAIL: query plan cannot be logged while page level lock is being held\\n>> HINT: Try pg_log_query_plan() after a few ....\\n> \\n> Agreed.\\n> I felt the HINT message 'after a few ...' is difficult to describe, and wrote as below.\\n> \\n> | HINT: Retrying pg_log_query_plan() might succeed since the lock duration of page level locks are usually short\\n> \\n> How do you think?\\n\\nOr we don't need HINT message?\\n\\n\\n+\\t\\t\\t\\terrmsg(\\\"could not log the query plan\\\"),\\n+\\t\\t\\t\\terrdetail(\\\"query plan cannot be logged while page level lock is being held\\\"),\\n\\nIn detail message, the first word of sentences should be capitalized. How about \\\"Cannot log the query plan while holding page-level lock.\\\", instead?\\n\\n\\nThanks for updating the patch! Here are some review comments.\\n\\n+ <row>\\n+ <entry role=\\\"func_table_entry\\\"><para role=\\\"func_signature\\\">\\n+ <indexterm>\\n+ <primary>pg_log_query_plan</primary>\\n+ </indexterm>\\n\\nThis entry is placed before one for pg_log_backend_memory_contexts(). But it should be *after* that since those entries seem to be placed in alphabetical order in the table?\\n\\n\\n+ Requests to log the plan of the query currently running on the\\n+ backend with specified process ID along with the untruncated\\n+ query string.\\n\\nOther descriptions about logging of query string seem not to mention something like \\\"untruncated query string\\\". For example, auto_explain, log_statement, etc. Why do we need to mention \\\"along with the untruncated query string\\\" specially for pg_log_query_plan()?\\n\\n\\n+ Note that nested statements (statements executed inside a function) are not\\n+ considered for logging. Only the plan of the most deeply nested query is logged.\\n\\nNow the plan of even nested statement can be logged. So this description needs to be updated?\\n\\n\\n@@ -440,6 +450,7 @@ standard_ExecutorFinish(QueryDesc *queryDesc)\\n \\n \\tMemoryContextSwitchTo(oldcontext);\\n \\n+\\tActiveQueryDesc = NULL;\\n\\nActiveQueryDesc seems unnecessary. Why does ActiveQueryDesc need to be reset to NULL in standard_ExecutorFinish()?\\n\\n\\nCurrently even during ProcessLogQueryPlanInterrupt(), pg_log_query_plan() can be call and another ProcessLogQueryPlanInterrupt() can be executed. So repeatable re-entrances to ProcessLogQueryPlanInterrupt() could cause \\\"stack depth limit exceeded\\\" error. To avoid this, shouldn't we make ProcessLogQueryPlanInterrupt() do nothing and return immediately, if it's called during another ProcessLogQueryPlanInterrupt()?\\n\\npg_log_backend_memory_contexts() also might have the same issue.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Feb 2022 01:51:11 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"At Tue, 1 Feb 2022 01:51:11 +0900, Fujii Masao <masao.fujii@oss.nttdata.com> wrote in \\n> \\n> \\n> On 2022/01/28 17:45, torikoshia wrote:\\n> >> There is the case where the request to log a query plan is skipped\\n> >> even while the target backend is running a query. If this happens,\\n> >> users can just retry pg_log_query_plan(). These things should be\\n> >> documented?\\n> > Agreed.\\n> > Added following:\\n> > + Note that there is the case where the request to log a\\n> > query\\n> > + plan is skipped even while the target backend is running a\\n> > + query due to lock conflict avoidance.\\n> > + If this happens, users can just retry pg_log_query_plan().\\n> \\n> This may cause users to misunderstand that pg_log_query_plan() fails\\n> while the target backend is holding *any* locks? Isn't it better to\\n> mention \\\"page-level locks\\\", instead? So how about the following?\\n> \\n> --------------------------\\n> Note that the request to log the query plan will be ignored if it's\\n> received during a short period while the target backend is holding a\\n> page-level lock.\\n> --------------------------\\n> \\n> \\n> >> + ereport(LOG_SERVER_ONLY,\\n> >> + errmsg(\\\"backend with PID %d is holding a page\\n> >> lock. Try again\\\",\\n> >> + MyProcPid));\\n> >>\\n> >> It seems more proper to output this message in DETAIL or HINT,\\n> >> instead. So how about something like the following messages?\\n> >>\\n> >> LOG: could not log the query plan\\n> >> DETAIL: query plan cannot be logged while page level lock is being\\n> >> held\\n> >> HINT: Try pg_log_query_plan() after a few ....\\n> > Agreed.\\n> > I felt the HINT message 'after a few ...' is difficult to describe,\\n> > and wrote as below.\\n> > | HINT: Retrying pg_log_query_plan() might succeed since the lock\\n> > | duration of page level locks are usually short\\n> > How do you think?\\n> \\n> Or we don't need HINT message?\\n> \\n> \\n> +\\t\\t\\t\\terrmsg(\\\"could not log the query plan\\\"),\\n> + errdetail(\\\"query plan cannot be logged while page level lock is\\n> being held\\\"),\\n> \\n> In detail message, the first word of sentences should be\\n> capitalized. How about \\\"Cannot log the query plan while holding\\n> page-level lock.\\\", instead?\\n> \\n> \\n> Thanks for updating the patch! Here are some review comments.\\n> \\n> + <row>\\n> + <entry role=\\\"func_table_entry\\\"><para role=\\\"func_signature\\\">\\n> + <indexterm>\\n> + <primary>pg_log_query_plan</primary>\\n> + </indexterm>\\n> \\n> This entry is placed before one for\\n> pg_log_backend_memory_contexts(). But it should be *after* that since\\n> those entries seem to be placed in alphabetical order in the table?\\n> \\n> \\n> + Requests to log the plan of the query currently running on the\\n> + backend with specified process ID along with the untruncated\\n> + query string.\\n> \\n> Other descriptions about logging of query string seem not to mention\\n> something like \\\"untruncated query string\\\". For example, auto_explain,\\n> log_statement, etc. Why do we need to mention \\\"along with the\\n> untruncated query string\\\" specially for pg_log_query_plan()?\\n> \\n> \\n> + Note that nested statements (statements executed inside a function)\\n> are not\\n> + considered for logging. Only the plan of the most deeply nested\\n> query is logged.\\n> \\n> Now the plan of even nested statement can be logged. So this\\n> description needs to be updated?\\n> \\n> \\n> @@ -440,6 +450,7 @@ standard_ExecutorFinish(QueryDesc *queryDesc)\\n> \\tMemoryContextSwitchTo(oldcontext);\\n> +\\tActiveQueryDesc = NULL;\\n> \\n> ActiveQueryDesc seems unnecessary. Why does ActiveQueryDesc need to be\\n> reset to NULL in standard_ExecutorFinish()?\\n> \\n> \\n> Currently even during ProcessLogQueryPlanInterrupt(),\\n> pg_log_query_plan() can be call and another\\n> ProcessLogQueryPlanInterrupt() can be executed. So repeatable\\n> re-entrances to ProcessLogQueryPlanInterrupt() could cause \\\"stack\\n> depth limit exceeded\\\" error. To avoid this, shouldn't we make\\n> ProcessLogQueryPlanInterrupt() do nothing and return immediately, if\\n> it's called during another ProcessLogQueryPlanInterrupt()?\\n> \\n> pg_log_backend_memory_contexts() also might have the same issue.\\n\\nGood catch.\\n\\nBy the way, I'm anxious about the following part and I'd like to\\nremove it.\\n\\n+\\t * Ensure no page lock is held on this process.\\n\\nIt seems to me what is wrong is ginInsertCleanup(), not this feature.\\nAs I read the comment for the assertion, I don't believe we want to\\nallow CFI while holding a page lock. And AFAICS the function is the\\nonly point where doing that. (It is the alone user of LockPage()...)\\n\\nThis is the assertion.\\n\\nlock.c: 902\\n> \\t/*\\n> \\t * We don't acquire any other heavyweight lock while holding the page lock\\n> \\t * except for relation extension.\\n> \\t */\\n> \\tAssert(!IsPageLockHeld ||\\n> \\t\\t (locktag->locktag_type == LOCKTAG_RELATION_EXTEND));\\n\\nIt is added by a recent commit 72e78d831ab5550c39f2dcc7cc5d44c406ec3dc2.\\n\\n> * Similar to relation extension, page locks are also held for a short\\n> * duration, so imposing such a restriction won't hurt.\\n\\nI don't believe a path involving vacuum_delay_point() calls is\\nshort-duration'ed.\\n\\nhttps://www.postgresql.org/message-id/CAH2-WznzCPUKnOV%2Bre30_rHLCkqQWm2JTSVjTCAei9LySTc2pw%40mail.gmail.com\\n\\n> One thing that really bothers me about commit e2c79e14 is that\\n> LockPage() is called, not LockBuffer(). GIN had no LockPage() calls\\n> before that commit, and is now the only code in the entire system that\\n> calls LockPage()/ConditionalLockPage() (the hash am no longer uses\\n> page heavyweight locks following recent work there).\\n\\nI agree to the discussion. Can't we use other mechanism here to get\\nrid of the Lockpage()?\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Tue, 01 Feb 2022 17:27:57 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2022/02/01 17:27, Kyotaro Horiguchi wrote:\\n>> * Similar to relation extension, page locks are also held for a short\\n>> * duration, so imposing such a restriction won't hurt.\\n> \\n> I don't believe a path involving vacuum_delay_point() calls is\\n> short-duration'ed.\\n\\nYes.\\n\\n\\n>> One thing that really bothers me about commit e2c79e14 is that\\n>> LockPage() is called, not LockBuffer(). GIN had no LockPage() calls\\n>> before that commit, and is now the only code in the entire system that\\n>> calls LockPage()/ConditionalLockPage() (the hash am no longer uses\\n>> page heavyweight locks following recent work there).\\n> \\n> I agree to the discussion. Can't we use other mechanism here to get\\n> rid of the Lockpage()?\\n\\nI have no good idea for that yet, but I agree it's better to get rid of page level lock.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Feb 2022 23:11:03 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"At Tue, 1 Feb 2022 23:11:03 +0900, Fujii Masao <masao.fujii@oss.nttdata.com> wrote in \\n> >> One thing that really bothers me about commit e2c79e14 is that\\n> >> LockPage() is called, not LockBuffer(). GIN had no LockPage() calls\\n> >> before that commit, and is now the only code in the entire system that\\n> >> calls LockPage()/ConditionalLockPage() (the hash am no longer uses\\n> >> page heavyweight locks following recent work there).\\n> > I agree to the discussion. Can't we use other mechanism here to get\\n> > rid of the Lockpage()?\\n> \\n> I have no good idea for that yet, but I agree it's better to get rid\\n> of page level lock.\\n\\nIt's my turn?\\n\\nThe page lock is used to hold-off simultaneous cleanups on the same\\nindex. ShareUpdateExclusive lock on the index relation works that\\nway. In that path it seems like we are always holding a RowExclusive\\nlock, so it seems to me we can use ShareUpdateExclusive for our\\npurpose.\\n\\nThere might be a false blocking case when another backend is holding a\\nconflicting lock on the index. They are, Share, ShareRowExclusive,\\nExclusive and AccessExclusive. The last three cases don't seem worth\\ndiscussion. I'm not sure about Share and Share Row cases. AFAICS\\nShare lock is taken on an index in ATExecReplicaIdentity, and there no\\nuse of ShareRowExclusive lock on an index. It's no use discussing about\\nexplicit locking.\\n\\nSo aren't we able to use ShareUpdateExclusive lock for that?\\n\\nIn the attached patch, ginInsertCleanup has an extra check for such\\nstronger locks not being held. At least \\\"make check\\\" doesn't cause\\nthe extra assertion to fire.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\ndiff --git a/src/backend/access/gin/ginfast.c b/src/backend/access/gin/ginfast.c\\nindex 7409fdc165..1af9a69abb 100644\\n--- a/src/backend/access/gin/ginfast.c\\n+++ b/src/backend/access/gin/ginfast.c\\n@@ -791,20 +791,29 @@ ginInsertCleanup(GinState *ginstate, bool full_clean,\\n \\tbool\\t\\tfsm_vac = false;\\n \\tSize\\t\\tworkMemory;\\n \\n-\\t/*\\n+\\t/*r\\n \\t * We would like to prevent concurrent cleanup process. For that we will\\n \\t * lock metapage in exclusive mode using LockPage() call. Nobody other\\n \\t * will use that lock for metapage, so we keep possibility of concurrent\\n \\t * insertion into pending list\\n \\t */\\n \\n+\\t/*\\n+\\t * we use ShareUpdateExclusive lock on this relation to hold-off concurrent\\n+\\t * cleanup\\n+\\t */\\n+\\tAssert(!CheckRelationLockedByMe(index, ShareUpdateExclusiveLock, false));\\n+\\n+\\t/* tentative debug-purpose assertion for stronger locks */\\n+\\tAssert(!CheckRelationLockedByMe(index, ShareLock, true));\\n+\\t\\n \\tif (forceCleanup)\\n \\t{\\n \\t\\t/*\\n \\t\\t * We are called from [auto]vacuum/analyze or gin_clean_pending_list()\\n \\t\\t * and we would like to wait concurrent cleanup to finish.\\n \\t\\t */\\n-\\t\\tLockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock);\\n+\\t\\tLockRelation(index, ShareUpdateExclusiveLock);\\n \\t\\tworkMemory =\\n \\t\\t\\t(IsAutoVacuumWorkerProcess() && autovacuum_work_mem != -1) ?\\n \\t\\t\\tautovacuum_work_mem : maintenance_work_mem;\\n@@ -816,7 +825,7 @@ ginInsertCleanup(GinState *ginstate, bool full_clean,\\n \\t\\t * just exit in hope that concurrent process will clean up pending\\n \\t\\t * list.\\n \\t\\t */\\n-\\t\\tif (!ConditionalLockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock))\\n+\\t\\tif (!ConditionalLockRelation(index, ShareUpdateExclusiveLock))\\n \\t\\t\\treturn;\\n \\t\\tworkMemory = work_mem;\\n \\t}\\n@@ -830,7 +839,7 @@ ginInsertCleanup(GinState *ginstate, bool full_clean,\\n \\t{\\n \\t\\t/* Nothing to do */\\n \\t\\tUnlockReleaseBuffer(metabuffer);\\n-\\t\\tUnlockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock);\\n+\\t\\tUnlockRelation(index, ShareUpdateExclusiveLock);\\n \\t\\treturn;\\n \\t}\\n \\n@@ -1002,7 +1011,7 @@ ginInsertCleanup(GinState *ginstate, bool full_clean,\\n \\t\\tpage = BufferGetPage(buffer);\\n \\t}\\n \\n-\\tUnlockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock);\\n+\\tUnlockRelation(index, ShareUpdateExclusiveLock);\\n \\tReleaseBuffer(metabuffer);\\n \\n \\t/*\",\n \"msg_date\": \"Wed, 02 Feb 2022 16:49:57 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"At Wed, 02 Feb 2022 16:49:57 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in \\n> So aren't we able to use ShareUpdateExclusive lock for that?\\n> \\n> In the attached patch, ginInsertCleanup has an extra check for such\\n> stronger locks not being held. At least \\\"make check\\\" doesn't cause\\n> the extra assertion to fire.\\n\\nActually, the discussion is a bit dubious. What we need really to\\ncheck is wheter such locks are not held on an index *elsewhere*.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Wed, 02 Feb 2022 17:04:21 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"2022-02-01 01:51, Fujii Masao wrote:\\nThanks for reviewing and suggestions!\\n\\n>> + Note that there is the case where the request to log a \\n>> query\\n>> + plan is skipped even while the target backend is running a\\n>> + query due to lock conflict avoidance.\\n>> + If this happens, users can just retry pg_log_query_plan().\\n> \\n> This may cause users to misunderstand that pg_log_query_plan() fails\\n> while the target backend is holding *any* locks? Isn't it better to\\n> mention \\\"page-level locks\\\", instead? So how about the following?\\n> \\n> --------------------------\\n> Note that the request to log the query plan will be ignored if it's\\n> received during a short period while the target backend is holding a\\n> page-level lock.\\n> --------------------------\\n\\nAgreed.\\n\\n>>> + ereport(LOG_SERVER_ONLY,\\n>>> + errmsg(\\\"backend with PID %d is holding a page lock. \\n>>> Try again\\\",\\n>>> + MyProcPid));\\n>>> \\n>>> It seems more proper to output this message in DETAIL or HINT,\\n>>> instead. So how about something like the following messages?\\n>>> \\n>>> LOG: could not log the query plan\\n>>> DETAIL: query plan cannot be logged while page level lock is being \\n>>> held\\n>>> HINT: Try pg_log_query_plan() after a few ....\\n>> \\n>> Agreed.\\n>> I felt the HINT message 'after a few ...' is difficult to describe, \\n>> and wrote as below.\\n>> \\n>> | HINT: Retrying pg_log_query_plan() might succeed since the lock \\n>> duration of page level locks are usually short\\n>> \\n>> How do you think?\\n> \\n> Or we don't need HINT message?\\n\\nRemoved the HINT message.\\n\\n> + errmsg(\\\"could not log the query plan\\\"),\\n> + errdetail(\\\"query plan cannot be logged \\n> while page level lock is\\n> being held\\\"),\\n> \\n> In detail message, the first word of sentences should be capitalized.\\n> How about \\\"Cannot log the query plan while holding page-level lock.\\\",\\n> instead?\\n\\nAgreed.\\n\\n> Thanks for updating the patch! Here are some review comments.\\n> \\n> + <row>\\n> + <entry role=\\\"func_table_entry\\\"><para role=\\\"func_signature\\\">\\n> + <indexterm>\\n> + <primary>pg_log_query_plan</primary>\\n> + </indexterm>\\n> \\n> This entry is placed before one for pg_log_backend_memory_contexts().\\n> But it should be *after* that since those entries seem to be placed in\\n> alphabetical order in the table?\\n\\nModified it.\\n\\n> + Requests to log the plan of the query currently running on the\\n> + backend with specified process ID along with the untruncated\\n> + query string.\\n> \\n> Other descriptions about logging of query string seem not to mention\\n> something like \\\"untruncated query string\\\". For example, auto_explain,\\n> log_statement, etc. Why do we need to mention \\\"along with the\\n> untruncated query string\\\" specially for pg_log_query_plan()?\\n\\nModified it as below:\\n\\n Requests to log the plan of the query currently running on \\nthe\\n - backend with specified process ID along with the \\nuntruncated\\n - query string.\\n - They will be logged at <literal>LOG</literal> message level \\nand\\n + backend with specified process ID.\\n + It will be logged at <literal>LOG</literal> message level \\nand\\n\\n> + Note that nested statements (statements executed inside a \\n> function) are not\\n> + considered for logging. Only the plan of the most deeply nested\\n> query is logged.\\n> \\n> Now the plan of even nested statement can be logged. So this\\n> description needs to be updated?\\n\\nModified it as below:\\n\\n - Note that nested statements (statements executed inside a \\nfunction) are not\\n - considered for logging. Only the plan of the most deeply nested \\nquery is logged.\\n + Note that when the statements are executed inside a function, \\nonly the\\n + plan of the most deeply nested query is logged.\\n\\n> @@ -440,6 +450,7 @@ standard_ExecutorFinish(QueryDesc *queryDesc)\\n> MemoryContextSwitchTo(oldcontext);\\n> + ActiveQueryDesc = NULL;\\n> \\n> ActiveQueryDesc seems unnecessary. Why does ActiveQueryDesc need to be\\n> reset to NULL in standard_ExecutorFinish()?\\n\\nActiveQueryDesc should not be reset in standard_ExecutorFinish().\\nRemoved it.\\n\\n> Currently even during ProcessLogQueryPlanInterrupt(),\\n> pg_log_query_plan() can be call and another\\n> ProcessLogQueryPlanInterrupt() can be executed. So repeatable\\n> re-entrances to ProcessLogQueryPlanInterrupt() could cause \\\"stack\\n> depth limit exceeded\\\" error. To avoid this, shouldn't we make\\n> ProcessLogQueryPlanInterrupt() do nothing and return immediately, if\\n> it's called during another ProcessLogQueryPlanInterrupt()?\\n> \\n> pg_log_backend_memory_contexts() also might have the same issue.\\n\\nAs you pointed out offlist, the issue could occur even when both \\npg_log_backend_memory_contexts and pg_log_query_plan are called and it \\nmay be better to make another patch.\\n\\n\\nYou also pointed out offlist that it would be necessary to call \\nLockErrorCleanup() if ProcessLogQueryPlanInterrupt() can incur ERROR.\\nAFAICS it can call ereport(ERROR), i.e., palloc0() in NewExplainState(), \\nso I added PG_TRY/CATCH and make it call LockErrorCleanup() when ERROR \\noccurs.\\n\\n\\n\\nOn 2022-02-01 17:27, Kyotaro Horiguchi wrote:\\n\\nThanks for reviewing Horiguchi-san!\\n\\n> By the way, I'm anxious about the following part and I'd like to\\n> remove it.\\n\\nI also think it would be nice if it's possible.\\n> \\n> + * Ensure no page lock is held on this process.\\n> \\n> It seems to me what is wrong is ginInsertCleanup(), not this feature.\\n\\n> Actually, the discussion is a bit dubious. What we need really to\\ncheck is wheter such locks are not held on an index *elsewhere*.\\n\\nSince I'm not sure how long it will take to discuss this point, the \\nattached patch is based on the current HEAD at this time.\\nI also think it may be better to discuss it on another thread.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Wed, 02 Feb 2022 21:59:12 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 2, 2022 at 7:59 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> 2022-02-01 01:51, Fujii Masao wrote:\\n<snip>\\n> > + Note that nested statements (statements executed inside a\\n> > function) are not\\n> > + considered for logging. Only the plan of the most deeply nested\\n> > query is logged.\\n> >\\n> > Now the plan of even nested statement can be logged. So this\\n> > description needs to be updated?\\n>\\n> Modified it as below:\\n>\\n> - Note that nested statements (statements executed inside a\\n> function) are not\\n> - considered for logging. Only the plan of the most deeply nested\\n> query is logged.\\n> + Note that when the statements are executed inside a function,\\n> only the\\n> + plan of the most deeply nested query is logged.\\n>\\n\\nMinor nit, but I think the \\\"the\\\" is superfluous.. ie.\\n\\n\\\"Note that when statements are executed inside a function,\\nonly the plan of the most deeply nested query is logged\\\"\\n\\n<snip>\\n> On 2022-02-01 17:27, Kyotaro Horiguchi wrote:\\n>\\n> Thanks for reviewing Horiguchi-san!\\n>\\n> > By the way, I'm anxious about the following part and I'd like to\\n> > remove it.\\n>\\n> I also think it would be nice if it's possible.\\n> >\\n> > + * Ensure no page lock is held on this process.\\n> >\\n> > It seems to me what is wrong is ginInsertCleanup(), not this feature.\\n>\\n> > Actually, the discussion is a bit dubious. What we need really to\\n> check is wheter such locks are not held on an index *elsewhere*.\\n>\\n> Since I'm not sure how long it will take to discuss this point, the\\n> attached patch is based on the current HEAD at this time.\\n> I also think it may be better to discuss it on another thread.\\n>\\n\\nWhile I agree on the above points, IMHO I don't believe it should be a\\nshow-stopper for adding this functionality to v15, but we have a few\\nmore commitments before we get to that point.\\n\\n\\nRobert Treat\\nhttps://xzilla.net\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Feb 2022 03:09:07 -0500\",\n \"msg_from\": \"Robert Treat <rob@xzilla.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2022/02/02 21:59, torikoshia wrote:\\n>> This may cause users to misunderstand that pg_log_query_plan() fails\\n>> while the target backend is holding *any* locks? Isn't it better to\\n>> mention \\\"page-level locks\\\", instead? So how about the following?\\n>>\\n>> --------------------------\\n>> Note that the request to log the query plan will be ignored if it's\\n>> received during a short period while the target backend is holding a\\n>> page-level lock.\\n>> --------------------------\\n> \\n> Agreed.\\n\\nOn second thought, this note is confusing rather than helpful? Because the users don't know when and what operation needs page-level lock. So now I'm thinking it's better to remove this note.\\n\\n\\n> As you pointed out offlist, the issue could occur even when both pg_log_backend_memory_contexts and pg_log_query_plan are called and it may be better to make another patch.\\n\\nOK.\\n\\n\\n> You also pointed out offlist that it would be necessary to call LockErrorCleanup() if ProcessLogQueryPlanInterrupt() can incur ERROR.\\n> AFAICS it can call ereport(ERROR), i.e., palloc0() in NewExplainState(), so I added PG_TRY/CATCH and make it call LockErrorCleanup() when ERROR occurs.\\n\\nAs we discussed off-list, this treatment might not be necessary. Because when ERROR or FATAL error happens, AbortTransaction() is called and LockErrorCleanup() is also called inside there.\\n\\n\\n> Since I'm not sure how long it will take to discuss this point, the attached patch is based on the current HEAD at this time.\\n\\nThanks for updating the patch!\\n\\n@@ -5048,6 +5055,12 @@ AbortSubTransaction(void)\\n \\t */\\n \\tPG_SETMASK(&UnBlockSig);\\n \\n+\\t/*\\n+\\t * When ActiveQueryDesc is referenced after abort, some of its elements\\n+\\t * are freed. To avoid accessing them, reset ActiveQueryDesc here.\\n+\\t */\\n+\\tActiveQueryDesc = NULL;\\n\\nAbortSubTransaction() should reset ActiveQueryDesc to save_ActiveQueryDesc that ExecutorRun() set, instead of NULL? Otherwise ActiveQueryDesc of top-level statement will be unavailable after subtransaction is aborted in the nested statements.\\n\\nFor example, no plan is logged while the following \\\"select pg_sleep(test())\\\" is running, because the exception inside test() function aborts the subtransaction and resets ActiveQueryDesc to NULL.\\n\\ncreate or replace function test () returns int as $$\\nbegin\\n perform 1/0;\\nexception when others then\\n return 30;\\nend;\\n$$ language plpgsql;\\n\\nselect pg_sleep(test());\\n\\n\\n-CREATE ROLE regress_log_memory;\\n+CREATE ROLE regress_log;\\n\\nIsn't this name too generic? How about regress_log_function, for example?\\n\\nRegards,\\n\\n-- \\nFujii Masao\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Tue, 8 Feb 2022 01:13:44 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"At Tue, 8 Feb 2022 01:13:44 +0900, Fujii Masao <masao.fujii@oss.nttdata.com> wrote in \\n> \\n> \\n> On 2022/02/02 21:59, torikoshia wrote:\\n> >> This may cause users to misunderstand that pg_log_query_plan() fails\\n> >> while the target backend is holding *any* locks? Isn't it better to\\n> >> mention \\\"page-level locks\\\", instead? So how about the following?\\n> >>\\n> >> --------------------------\\n> >> Note that the request to log the query plan will be ignored if it's\\n> >> received during a short period while the target backend is holding a\\n> >> page-level lock.\\n> >> --------------------------\\n> > Agreed.\\n> \\n> On second thought, this note is confusing rather than helpful? Because\\n> the users don't know when and what operation needs page-level lock. So\\n> now I'm thinking it's better to remove this note.\\n\\n*I* agree to removing the note. And the following error message looks\\nas mysterious as the note is, and the DETAIL doesn't help..\\n\\n\\t\\t\\tereport(LOG_SERVER_ONLY,\\n+\\t\\t\\t\\terrmsg(\\\"could not log the query plan\\\"),\\n+\\t\\t\\t\\terrdetail(\\\"Cannot log the query plan while holding page-level lock.\\\"));\\n+\\t\\t\\thash_seq_term(&status);\\n\\nWe should tell the command can be retried soon, like this?\\n\\n\\\"LOG: ignored request for logging query plan due to lock confilcts\\\"\\n\\\"HINT: You can try again in a moment.\\\"\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Tue, 08 Feb 2022 17:18:52 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2022-02-03 17:09, Robert Treat wrote:\\n> On Wed, Feb 2, 2022 at 7:59 AM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> \\n>> 2022-02-01 01:51, Fujii Masao wrote:\\n> <snip>\\n>> > + Note that nested statements (statements executed inside a\\n>> > function) are not\\n>> > + considered for logging. Only the plan of the most deeply nested\\n>> > query is logged.\\n>> >\\n>> > Now the plan of even nested statement can be logged. So this\\n>> > description needs to be updated?\\n>> \\n>> Modified it as below:\\n>> \\n>> - Note that nested statements (statements executed inside a\\n>> function) are not\\n>> - considered for logging. Only the plan of the most deeply \\n>> nested\\n>> query is logged.\\n>> + Note that when the statements are executed inside a \\n>> function,\\n>> only the\\n>> + plan of the most deeply nested query is logged.\\n>> \\n> \\n> Minor nit, but I think the \\\"the\\\" is superfluous.. ie.\\n> \\n> \\\"Note that when statements are executed inside a function,\\n> only the plan of the most deeply nested query is logged\\\"\\n\\nThanks! Modified it.\\n\\n\\nOn 2022-02-08 01:13, Fujii Masao wrote:\\nThanks for the comments!\\n\\n> On 2022/02/02 21:59, torikoshia wrote:\\n>>> This may cause users to misunderstand that pg_log_query_plan() fails\\n>>> while the target backend is holding *any* locks? Isn't it better to\\n>>> mention \\\"page-level locks\\\", instead? So how about the following?\\n>>> \\n>>> --------------------------\\n>>> Note that the request to log the query plan will be ignored if it's\\n>>> received during a short period while the target backend is holding a\\n>>> page-level lock.\\n>>> --------------------------\\n>> \\n>> Agreed.\\n> \\n> On second thought, this note is confusing rather than helpful? Because\\n> the users don't know when and what operation needs page-level lock. So\\n> now I'm thinking it's better to remove this note.\\n\\nRemoved it.\\n> \\n> \\n>> As you pointed out offlist, the issue could occur even when both \\n>> pg_log_backend_memory_contexts and pg_log_query_plan are called and it \\n>> may be better to make another patch.\\n> \\n> OK.\\n> \\n> \\n>> You also pointed out offlist that it would be necessary to call \\n>> LockErrorCleanup() if ProcessLogQueryPlanInterrupt() can incur ERROR.\\n>> AFAICS it can call ereport(ERROR), i.e., palloc0() in \\n>> NewExplainState(), so I added PG_TRY/CATCH and make it call \\n>> LockErrorCleanup() when ERROR occurs.\\n> \\n> As we discussed off-list, this treatment might not be necessary.\\n> Because when ERROR or FATAL error happens, AbortTransaction() is\\n> called and LockErrorCleanup() is also called inside there.\\n\\nAgreed.\\n\\n>> Since I'm not sure how long it will take to discuss this point, the \\n>> attached patch is based on the current HEAD at this time.\\n> \\n> Thanks for updating the patch!\\n> \\n> @@ -5048,6 +5055,12 @@ AbortSubTransaction(void)\\n> \\t */\\n> \\tPG_SETMASK(&UnBlockSig);\\n> +\\t/*\\n> +\\t * When ActiveQueryDesc is referenced after abort, some of its \\n> elements\\n> +\\t * are freed. To avoid accessing them, reset ActiveQueryDesc here.\\n> +\\t */\\n> +\\tActiveQueryDesc = NULL;\\n> \\n> AbortSubTransaction() should reset ActiveQueryDesc to\\n> save_ActiveQueryDesc that ExecutorRun() set, instead of NULL?\\n> Otherwise ActiveQueryDesc of top-level statement will be unavailable\\n> after subtransaction is aborted in the nested statements.\\n> \\n> For example, no plan is logged while the following \\\"select\\n> pg_sleep(test())\\\" is running, because the exception inside test()\\n> function aborts the subtransaction and resets ActiveQueryDesc to NULL.\\n> \\n> create or replace function test () returns int as $$\\n> begin\\n> perform 1/0;\\n> exception when others then\\n> return 30;\\n> end;\\n> $$ language plpgsql;\\n> \\n> select pg_sleep(test());\\n\\nAgreed.\\n\\nBTW, since the above example results in calling ExecutorRun() only once, \\nthe output didn't differ even after ActiveQueryDesc is reset to \\nsave_ActiveQueryDesc.\\n\\nThe below definition of test() worked as expected.\\n\\n create or replace function test () returns int as $$\\n begin\\n perform 1;\\n perform 1/0;\\n exception when others then\\n return 30;\\n end;\\n $$ language plpgsql;\\n\\n> \\n> \\n> -CREATE ROLE regress_log_memory;\\n> +CREATE ROLE regress_log;\\n> \\n> Isn't this name too generic? How about regress_log_function, for \\n> example?\\n\\nAgreed.\\n\\n\\nOn 2022-02-08 17:18, Kyotaro Horiguchi wrote:\\n> At Tue, 8 Feb 2022 01:13:44 +0900, Fujii Masao\\n> <masao.fujii@oss.nttdata.com> wrote in\\n>> \\n>> \\n>> On 2022/02/02 21:59, torikoshia wrote:\\n>> >> This may cause users to misunderstand that pg_log_query_plan() fails\\n>> >> while the target backend is holding *any* locks? Isn't it better to\\n>> >> mention \\\"page-level locks\\\", instead? So how about the following?\\n>> >>\\n>> >> --------------------------\\n>> >> Note that the request to log the query plan will be ignored if it's\\n>> >> received during a short period while the target backend is holding a\\n>> >> page-level lock.\\n>> >> --------------------------\\n>> > Agreed.\\n>> \\n>> On second thought, this note is confusing rather than helpful? Because\\n>> the users don't know when and what operation needs page-level lock. So\\n>> now I'm thinking it's better to remove this note.\\n> \\n> *I* agree to removing the note. And the following error message looks\\n> as mysterious as the note is, and the DETAIL doesn't help..\\n> \\n> \\t\\t\\tereport(LOG_SERVER_ONLY,\\n> +\\t\\t\\t\\terrmsg(\\\"could not log the query plan\\\"),\\n> +\\t\\t\\t\\terrdetail(\\\"Cannot log the query plan while holding page-level \\n> lock.\\\"));\\n> +\\t\\t\\thash_seq_term(&status);\\n> \\n> We should tell the command can be retried soon, like this?\\n> \\n> \\\"LOG: ignored request for logging query plan due to lock confilcts\\\"\\n> \\\"HINT: You can try again in a moment.\\\"\\n\\nThanks, it seems better.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Wed, 09 Feb 2022 00:12:50 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2022/02/09 0:12, torikoshia wrote:\\n> BTW, since the above example results in calling ExecutorRun() only once, the output didn't differ even after ActiveQueryDesc is reset to save_ActiveQueryDesc.\\n> \\n> The below definition of test() worked as expected.\\n> \\n> create or replace function test () returns int as $$\\n> begin\\n> perform 1;\\n> perform 1/0;\\n> exception when others then\\n> return 30;\\n> end;\\n> $$ language plpgsql;\\nSo in this case ActiveQueryDesc set by ExecutorRun() called only once is still valid even when AbortSubTransaction() is called. That is, that ActiveQueryDesc should NOT be reset to save_ActiveQueryDesc in this case, should it?\\n\\nOTOH, in your example, ActiveQueryDesc set by the second call to ExecutorRun() should be reset in AbortSubTransaction(). Then ActiveQueryDesc set by the first call should be valid.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Tue, 22 Feb 2022 02:39:16 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2022-02-08 01:13, Fujii Masao wrote:\\n> AbortSubTransaction() should reset ActiveQueryDesc to\\n> save_ActiveQueryDesc that ExecutorRun() set, instead of NULL?\\n> Otherwise ActiveQueryDesc of top-level statement will be unavailable\\n> after subtransaction is aborted in the nested statements.\\n\\nI once agreed above suggestion and made v20 patch making \\nsave_ActiveQueryDesc a global variable, but it caused segfault when \\ncalling pg_log_query_plan() after FreeQueryDesc().\\n\\nOTOH, doing some kind of reset of ActiveQueryDesc seems necessary since \\nit also caused segfault when running pg_log_query_plan() during \\ninstallcheck.\\n\\nThere may be a better way, but resetting ActiveQueryDesc to NULL seems \\nsafe and simple.\\nOf course it makes pg_log_query_plan() useless after a subtransaction is \\naborted.\\nHowever, if it does not often happen that people want to know the \\nrunning query's plan whose subtransaction is aborted, resetting \\nActiveQueryDesc to NULL would be acceptable.\\n\\nAttached is a patch that sets ActiveQueryDesc to NULL when a \\nsubtransaction is aborted.\\n\\nHow do you think?\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Wed, 09 Mar 2022 19:04:49 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2022-03-09 19:04, torikoshia wrote:\\n> On 2022-02-08 01:13, Fujii Masao wrote:\\n>> AbortSubTransaction() should reset ActiveQueryDesc to\\n>> save_ActiveQueryDesc that ExecutorRun() set, instead of NULL?\\n>> Otherwise ActiveQueryDesc of top-level statement will be unavailable\\n>> after subtransaction is aborted in the nested statements.\\n> \\n> I once agreed above suggestion and made v20 patch making\\n> save_ActiveQueryDesc a global variable, but it caused segfault when\\n> calling pg_log_query_plan() after FreeQueryDesc().\\n> \\n> OTOH, doing some kind of reset of ActiveQueryDesc seems necessary\\n> since it also caused segfault when running pg_log_query_plan() during\\n> installcheck.\\n> \\n> There may be a better way, but resetting ActiveQueryDesc to NULL seems\\n> safe and simple.\\n> Of course it makes pg_log_query_plan() useless after a subtransaction\\n> is aborted.\\n> However, if it does not often happen that people want to know the\\n> running query's plan whose subtransaction is aborted, resetting\\n> ActiveQueryDesc to NULL would be acceptable.\\n> \\n> Attached is a patch that sets ActiveQueryDesc to NULL when a\\n> subtransaction is aborted.\\n> \\n> How do you think?\\n\\nAttached new patch to fix patch apply failures.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Mon, 16 May 2022 17:02:40 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2022-05-16 17:02, torikoshia wrote:\\n> On 2022-03-09 19:04, torikoshia wrote:\\n>> On 2022-02-08 01:13, Fujii Masao wrote:\\n>>> AbortSubTransaction() should reset ActiveQueryDesc to\\n>>> save_ActiveQueryDesc that ExecutorRun() set, instead of NULL?\\n>>> Otherwise ActiveQueryDesc of top-level statement will be unavailable\\n>>> after subtransaction is aborted in the nested statements.\\n>> \\n>> I once agreed above suggestion and made v20 patch making\\n>> save_ActiveQueryDesc a global variable, but it caused segfault when\\n>> calling pg_log_query_plan() after FreeQueryDesc().\\n>> \\n>> OTOH, doing some kind of reset of ActiveQueryDesc seems necessary\\n>> since it also caused segfault when running pg_log_query_plan() during\\n>> installcheck.\\n>> \\n>> There may be a better way, but resetting ActiveQueryDesc to NULL seems\\n>> safe and simple.\\n>> Of course it makes pg_log_query_plan() useless after a subtransaction\\n>> is aborted.\\n>> However, if it does not often happen that people want to know the\\n>> running query's plan whose subtransaction is aborted, resetting\\n>> ActiveQueryDesc to NULL would be acceptable.\\n>> \\n>> Attached is a patch that sets ActiveQueryDesc to NULL when a\\n>> subtransaction is aborted.\\n>> \\n>> How do you think?\\nAttached new patch to fix patch apply failures again.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Thu, 08 Sep 2022 19:19:09 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi,I'm sorry,if this message is duplicated previous this one, but I'm not sure that the previous message is sent correctly. I sent it from email address a.rybakina@postgrespro.ru and I couldn't send this one email from those address.I like idea to create patch for logging query plan. After reviewing this code and notice some moments and I'd rather ask you some questions.Firstly, I suggest some editing in the comment of commit. I think, it is turned out the more laconic and the same clear. I wrote it below since I can't think of any other way to add it.```Currently, we have to wait for finishing of the query execution to check its plan.This is not so convenient in investigation long-running queries on productionenvironments where we cannot use debuggers.To improve this situation there is proposed the patch containing the pg_log_query_plan()function which request to log plan of the specified backend process.By default, only superusers are allowed to request log of the plan otherwiseallowing any users to issue this request could create cause lots of log messagesand it can lead to denial of service.At the next requesting CHECK_FOR_INTERRUPTS(), the target backend logs its plan atLOG_SERVER_ONLY level and therefore this plan will appear in the server log only,not to be sent to the client.```Secondly, I have question about deleting USE_ASSERT_CHECKING in lock.h?It supposed to have been checked in another placed of the code by matching values. I am worry about skipping errors due to untesting with assert option in the places where it (GetLockMethodLocalHash) participates and we won't able to get core file in segfault cases. I might not understand something, then can you please explain to me?Thirdly, I have incomprehension of the point why save_ActiveQueryDesc is declared in the pquery.h? I am seemed to save_ActiveQueryDesc to be used in an once time in the ExecutorRun function and its declaration superfluous. I added it in the attached patch.Fourthly, it seems to me there are not enough explanatory comments in the code. I also added them in the attached patch.Lastly, I have incomprehension about handling signals since have been unused it before. Could another signal disabled calling this signal to log query plan? I noticed this signal to be checked the latest in procsignal_sigusr1_handler function. -- Regards,Alena RybakinaPostgres Professional 19.09.2022, 11:01, \\\"torikoshia\\\" <torikoshia@oss.nttdata.com>:On 2022-05-16 17:02, torikoshia wrote: On 2022-03-09 19:04, torikoshia wrote: On 2022-02-08 01:13, Fujii Masao wrote: AbortSubTransaction() should reset ActiveQueryDesc to save_ActiveQueryDesc that ExecutorRun() set, instead of NULL? Otherwise ActiveQueryDesc of top-level statement will be unavailable after subtransaction is aborted in the nested statements. I once agreed above suggestion and made v20 patch making save_ActiveQueryDesc a global variable, but it caused segfault when calling pg_log_query_plan() after FreeQueryDesc(). OTOH, doing some kind of reset of ActiveQueryDesc seems necessary since it also caused segfault when running pg_log_query_plan() during installcheck. There may be a better way, but resetting ActiveQueryDesc to NULL seems safe and simple. Of course it makes pg_log_query_plan() useless after a subtransaction is aborted. However, if it does not often happen that people want to know the running query's plan whose subtransaction is aborted, resetting ActiveQueryDesc to NULL would be acceptable. Attached is a patch that sets ActiveQueryDesc to NULL when a subtransaction is aborted. How do you think?Attached new patch to fix patch apply failures again. --Regards,--Atsushi TorikoshiNTT DATA CORPORATION\",\n \"msg_date\": \"Mon, 19 Sep 2022 11:47:21 +0300\",\n \"msg_from\": \"=?utf-8?B?0JDQu9C10L3QsCDQoNGL0LHQsNC60LjQvdCw?=\\n <lena.ribackina@yandex.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2022-09-19 17:47, Алена Рыбакина wrote:\\nThanks for your review and comments!\\n\\n> Hi,\\n> \\n> I'm sorry,if this message is duplicated previous this one, but I'm not\\n> sure that the previous message is sent correctly. I sent it from email\\n> address a.rybakina@postgrespro.ru and I couldn't send this one email\\n> from those address.\\n\\nI've successfully received your mail from both a.rybakina@postgrespro.ru \\nand lena.ribackina@yandex.ru.\\n\\n> I like idea to create patch for logging query plan. After reviewing\\n> this code and notice some moments and I'd rather ask you some\\n> questions.\\n> \\n> Firstly, I suggest some editing in the comment of commit. I think, it \\n> is\\n> turned out the more laconic and the same clear. I wrote it below since \\n> I\\n> can't think of any other way to add it.\\n\\n> ```\\n> Currently, we have to wait for finishing of the query execution to \\n> check\\n> its plan.\\n> This is not so convenient in investigation long-running queries on\\n> production\\n> environments where we cannot use debuggers.\\n\\n> To improve this situation there is proposed the patch containing the\\n> pg_log_query_plan()\\n> function which request to log plan of the specified backend process.\\n\\n> By default, only superusers are allowed to request log of the plan\\n> otherwise\\n> allowing any users to issue this request could create cause lots of log\\n> messages\\n> and it can lead to denial of service.\\n> \\n> At the next requesting CHECK_FOR_INTERRUPTS(), the target backend logs\\n> its plan at\\n> LOG_SERVER_ONLY level and therefore this plan will appear in the server\\n> log only,\\n> not to be sent to the client.\\nThanks, I have incorporated your comments.\\nSince the latter part of the original message comes from the commit \\nmessage of pg_log_backend_memory_contexts(43620e328617c), so I left it \\nas it was for consistency.\\n\\n\\n> Secondly, I have question about deleting USE_ASSERT_CHECKING in lock.h?\\n> It supposed to have been checked in another placed of the code by\\n> matching values. I am worry about skipping errors due to untesting with\\n> assert option in the places where it (GetLockMethodLocalHash)\\n> participates and we won't able to get core file in segfault cases. I\\n> might not understand something, then can you please explain to me?\\nSince GetLockMethodLocalHash() is only used for assertions, this is only \\ndefined when USE_ASSERT_CHECKING is enabled.\\nThis patch uses GetLockMethodLocalHash() not only for the assertion \\npurpose, so I removed \\\"ifdef USE_ASSERT_CHECKING\\\" for this function.\\nI belive it does not lead to skip errors.\\n\\n> Thirdly, I have incomprehension of the point why save_ActiveQueryDesc \\n> is\\n> declared in the pquery.h? I am seemed to save_ActiveQueryDesc to be \\n> used\\n> in an once time in the ExecutorRun function and its declaration\\n> superfluous. I added it in the attached patch.\\nExactly.\\n\\n> Fourthly, it seems to me there are not enough explanatory comments in\\n> the code. I also added them in the attached patch.\\nThanks!\\n\\n| + /*\\n| + * Save value of ActiveQueryDesc before having called\\n| + * ExecutorRun_hook function due to having reset by\\n| + * AbortSubTransaction.\\n| + */\\n| +\\n| save_ActiveQueryDesc = ActiveQueryDesc;\\n| ActiveQueryDesc = queryDesc;\\n|\\n| @@ -314,6 +320,7 @@ ExecutorRun(QueryDesc *queryDesc,\\n| else\\n| standard_ExecutorRun(queryDesc, direction, count, \\nexecute_once);\\n|\\n| + /* We set the actual value of ActiveQueryDesc */\\n| ActiveQueryDesc = save_ActiveQueryDesc;\\n\\nSince these processes are needed for nested queries, not only for\\nAbortSubTransaction[1], added comments on it.\\n\\n| +/* Function to handle the signal to output the query plan. */\\n| extern void HandleLogQueryPlanInterrupt(void);\\n\\nI feel this comment is unnecessary since the explanation of \\nHandleLogQueryPlanInterrupt() is written in explain.c and no functions \\nin explain.h have comments in it.\\n\\n> Lastly, I have incomprehension about handling signals since have been\\n> unused it before. Could another signal disabled calling this signal to\\n> log query plan? I noticed this signal to be checked the latest in\\n> procsignal_sigusr1_handler function.\\nAre you concerned that one signal will not be processed when multiple \\nsignals are sent in succession?\\nAFAIU both of them are processed since SendProcSignal flags \\nps_signalFlags for each signal.\\n\\n```\\n SendProcSignal(pid_t pid, ProcSignalReason reason, BackendId backendId)\\n {\\n volatile ProcSignalSlot *slot;\\n...(snip)...\\n278 if (slot->pss_pid == pid)\\n279 {\\n280 /* Atomically set the proper flag */\\n281 slot->pss_signalFlags[reason] = true;\\n282 /* Send signal */\\n283 return kill(pid, SIGUSR1);\\n```\\n\\nComments of ProcSignalReason also say 'We can cope with concurrent \\nsignals for different reasons'.\\n\\n```C\\n /*\\n * Reasons for signaling a Postgres child process (a backend or an \\nauxiliary\\n * process, like checkpointer). We can cope with concurrent signals \\nfor different\\n * reasons. However, if the same reason is signaled multiple times in \\nquick\\n * succession, the process is likely to observe only one notification \\nof it.\\n * This is okay for the present uses.\\n ...\\n typedef enum\\n {\\n PROCSIG_CATCHUP_INTERRUPT, /* sinval catchup interrupt */\\n PROCSIG_NOTIFY_INTERRUPT, /* listen/notify interrupt */\\n PROCSIG_PARALLEL_MESSAGE, /* message from cooperating parallel \\nbackend */\\n PROCSIG_WALSND_INIT_STOPPING, /* ask walsenders to prepare for \\nshutdown */\\n PROCSIG_BARRIER, /* global barrier interrupt */\\n PROCSIG_LOG_MEMORY_CONTEXT, /* ask backend to log the memory \\ncontexts */\\n PROCSIG_LOG_QUERY_PLAN, /* ask backend to log plan of the current \\nquery */\\n...\\n } ProcSignalReason;\\n```\\n\\n\\n[1] \\nhttps://www.postgresql.org/message-id/8b53b32f-26cc-0531-4ac0-27310e0bef4b%40oss.nttdata.com\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Tue, 20 Sep 2022 23:41:57 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Ok, I get it.\\n\\n> Since GetLockMethodLocalHash() is only used for assertions, this is \\n> only defined when USE_ASSERT_CHECKING is enabled. This patch uses \\n> GetLockMethodLocalHash() not only for the assertion purpose, so I \\n> removed \\\"ifdef USE_ASSERT_CHECKING\\\" for this function. I belive it \\n> does not lead to skip errors. \\nAgree.\\n> Since these processes are needed for nested queries, not only for \\n> AbortSubTransaction[1], added comments on it.\\n\\nI also noticed it. However I also discovered that above function \\ndeclarations to be aplied for explain command and used to be printed \\ndetails of the executed query.\\n\\nWe have a similar task to print the plan of an interrupted process \\nmaking a request for a specific pid.\\n\\nIn short, I think, this task is different and for separating these parts \\nI added this comment.\\n\\n> I feel this comment is unnecessary since the explanation of \\n> HandleLogQueryPlanInterrupt() is written in explain.c and no functions \\n> in explain.h have comments in it. \\n\\nYes, I was worried about it. I understood it, thank for explaining.\\n\\n>\\n> AFAIU both of them are processed since SendProcSignal flags \\n> ps_signalFlags for each signal.\\n>\\n> ```\\n> SendProcSignal(pid_t pid, ProcSignalReason reason, BackendId backendId)\\n> {\\n> volatile ProcSignalSlot *slot;\\n> ...(snip)...\\n> 278 if (slot->pss_pid == pid)\\n> 279 {\\n> 280 /* Atomically set the proper flag */\\n> 281 slot->pss_signalFlags[reason] = true;\\n> 282 /* Send signal */\\n> 283 return kill(pid, SIGUSR1);\\n> ```\\n>\\n> Comments of ProcSignalReason also say 'We can cope with concurrent \\n> signals for different reasons'.\\n>\\n> ```C\\n> /*\\n> * Reasons for signaling a Postgres child process (a backend or an \\n> auxiliary\\n> * process, like checkpointer). We can cope with concurrent signals for \\n> different\\n> * reasons. However, if the same reason is signaled multiple times in \\n> quick\\n> * succession, the process is likely to observe only one notification \\n> of it.\\n> * This is okay for the present uses.\\n> ...\\n> typedef enum\\n> {\\n> PROCSIG_CATCHUP_INTERRUPT, /* sinval catchup interrupt */\\n> PROCSIG_NOTIFY_INTERRUPT, /* listen/notify interrupt */\\n> PROCSIG_PARALLEL_MESSAGE, /* message from cooperating parallel backend */\\n> PROCSIG_WALSND_INIT_STOPPING, /* ask walsenders to prepare for \\n> shutdown */\\n> PROCSIG_BARRIER, /* global barrier interrupt */\\n> PROCSIG_LOG_MEMORY_CONTEXT, /* ask backend to log the memory contexts */\\n> PROCSIG_LOG_QUERY_PLAN, /* ask backend to log plan of the current \\n> query */\\n> ...\\n> } ProcSignalReason;\\n> ```\\n>\\n>\\n> [1] \\n> https://www.postgresql.org/message-id/8b53b32f-26cc-0531-4ac0-27310e0bef4b%40oss.nttdata.com\\n\\n\\n\\n\\n\\n Ok, I get it.\\n\\n Since GetLockMethodLocalHash() is only used for assertions, this\\n is only defined when USE_ASSERT_CHECKING is enabled. This patch\\n uses GetLockMethodLocalHash() not only for the assertion purpose,\\n so I removed \\\"ifdef USE_ASSERT_CHECKING\\\" for this function. I\\n belive it does not lead to skip errors. \\nAgree.\\n \\n\\n Since these processes are needed for nested queries, not only for\\n AbortSubTransaction[1], added comments on it.\\nI also noticed it. However I also discovered that\\n above function declarations to be aplied for explain command and\\n used to be printed details of the executed query.\\nWe have a similar task to print the plan of an\\n interrupted process making a request for a specific pid.\\nIn short, I think, this task is different and for\\n separating these parts I added this comment.\\n\\n I feel this comment is unnecessary since the explanation of\\n HandleLogQueryPlanInterrupt() is written in explain.c and no\\n functions\\n in explain.h have comments in it.\\n \\n\\nYes, I was worried about it. I understood it, thank\\n for explaining.\\n\\n\\n AFAIU both of them are processed since SendProcSignal flags\\n ps_signalFlags for each signal.\\n \\n\\n ```\\n \\n SendProcSignal(pid_t pid, ProcSignalReason reason, BackendId\\n backendId)\\n \\n {\\n \\n volatile ProcSignalSlot *slot;\\n \\n ...(snip)...\\n \\n 278 if (slot->pss_pid == pid)\\n \\n 279 {\\n \\n 280 /* Atomically set the proper flag */\\n \\n 281 slot->pss_signalFlags[reason] = true;\\n \\n 282 /* Send signal */\\n \\n 283 return kill(pid, SIGUSR1);\\n \\n ```\\n \\n\\n Comments of ProcSignalReason also say 'We can cope with concurrent\\n signals for different reasons'.\\n \\n\\n ```C\\n \\n /*\\n \\n * Reasons for signaling a Postgres child process (a backend or an\\n auxiliary\\n \\n * process, like checkpointer). We can cope with concurrent signals\\n for different\\n \\n * reasons. However, if the same reason is signaled multiple times\\n in\\n quick\\n \\n * succession, the process is likely to observe only one\\n notification\\n of it.\\n \\n * This is okay for the present uses.\\n \\n ...\\n \\n typedef enum\\n \\n {\\n \\n PROCSIG_CATCHUP_INTERRUPT, /* sinval catchup interrupt */\\n \\n PROCSIG_NOTIFY_INTERRUPT, /* listen/notify interrupt */\\n \\n PROCSIG_PARALLEL_MESSAGE, /* message from cooperating parallel\\n backend */\\n \\n PROCSIG_WALSND_INIT_STOPPING, /* ask walsenders to prepare for\\n shutdown */\\n \\n PROCSIG_BARRIER, /* global barrier interrupt */\\n \\n PROCSIG_LOG_MEMORY_CONTEXT, /* ask backend to log the memory\\n contexts */\\n \\n PROCSIG_LOG_QUERY_PLAN, /* ask backend to log plan of the current\\n query */\\n \\n ...\\n \\n } ProcSignalReason;\\n \\n ```\\n \\n\\n\\n [1]\\nhttps://www.postgresql.org/message-id/8b53b32f-26cc-0531-4ac0-27310e0bef4b%40oss.nttdata.com\",\n \"msg_date\": \"Wed, 21 Sep 2022 11:30:25 +0300\",\n \"msg_from\": \"Alena Rybakina <lena.ribackina@yandex.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2022-09-21 17:30, Alena Rybakina wrote:\\n\\nThanks for your reply!\\n\\n> I also noticed it. However I also discovered that above function\\n> declarations to be aplied for explain command and used to be printed\\n> details of the executed query.\\n> \\n> We have a similar task to print the plan of an interrupted process\\n> making a request for a specific pid.\\n> \\n> In short, I think, this task is different and for separating these\\n> parts I added this comment.\\n\\nI'm not sure I understand your comment correctly, do you mean \\nHandleLogQueryPlanInterrupt() should not be placed in explain.c?\\n\\nIt may be so.\\nHowever, given that ProcesLogMemoryContextInterrupt(), which similarly \\nhandles interrupts for pg_log_backend_memory_contexts(), is located in \\nmcxt.c, I also think current location might be acceptable.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Thu, 22 Sep 2022 23:12:16 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Sorry, I wrote confusingly at that time.\\n\\nNo, I suggested adding comment about the explanation of \\nHandleLogQueryPlanInterrupt() only in the explain.h and not removing \\nfrom the explain.c.\\n\\nI seemed to be necessary separating declaration function for 'explaining \\nfeature' of executed query from our logging plan of the running query \\ninterrupts function. But now, I doubt it.\\n\\n> I'm not sure I understand your comment correctly, do you mean \\n> HandleLogQueryPlanInterrupt() should not be placed in explain.c? \\n\\n\\nThank you for having reminded about this function and I looked at \\nProcessLogMemoryContextInterrupt() declaration. I'm noticed comments in \\nthe memutils.h are missed tooю\\n\\nDescription of this function is written only in mcxt.c.\\n> However, given that ProcesLogMemoryContextInterrupt(), which similarly \\n> handles interrupts for pg_log_backend_memory_contexts(), is located in \\n> mcxt.c, I also think current location might be acceptable. \\n\\n\\nSo I think you are right and the comment about the explanation of \\nHandleLogQueryPlanInterrupt() written in explain.h is redundant.\\n\\n> I feel this comment is unnecessary since the explanation of \\n> HandleLogQueryPlanInterrupt() is written in explain.c and no functions \\n> in explain.h have comments in it. \\n\\nRegards,\\n\\n-- \\nAlena Rybakina\\nPostgres Professional\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Sep 2022 12:43:19 +0300\",\n \"msg_from\": \"Alena Rybakina <lena.ribackina@yandex.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nThis patch does not currently build, due to a conflicting oid:\\n\\nhttps://cirrus-ci.com/task/4638460594618368?logs=build#L108\\n[17:26:44.602] /usr/bin/perl ../src/include/catalog/../../backend/catalog/genbki.pl --include-path=../src/include --set-version=16 --output=src/include/catalog ../src/include/catalog/pg_proc.h ../src/include/catalog/pg_type.h ../src/include/catalog/pg_attribute.h ../src/include/catalog/pg_class.h ../src/include/catalog/pg_attrdef.h ../src/include/catalog/pg_constraint.h ../src/include/catalog/pg_inherits.h ../src/include/catalog/pg_index.h ../src/include/catalog/pg_operator.h ../src/include/catalog/pg_opfamily.h ../src/include/catalog/pg_opclass.h ../src/include/catalog/pg_am.h ../src/include/catalog/pg_amop.h ../src/include/catalog/pg_amproc.h ../src/include/catalog/pg_language.h ../src/include/catalog/pg_largeobject_metadata.h ../src/include/catalog/pg_largeobject.h ../src/include/catalog/pg_aggregate.h ../src/include/catalog/pg_statistic.h ../src/include/catalog/pg_statistic_ext.h ../src/include/catalog/pg_statistic_ext_data.h ../src/include/catalog/pg_rewrite.h ../src/include/catalog/pg_trigger.h ../src/include/catalog/pg_event_trigger.h ../src/include/catalog/pg_description.h ../src/include/catalog/pg_cast.h ../src/include/catalog/pg_enum.h ../src/include/catalog/pg_namespace.h ../src/include/catalog/pg_conversion.h ../src/include/catalog/pg_depend.h ../src/include/catalog/pg_database.h ../src/include/catalog/pg_db_role_setting.h ../src/include/catalog/pg_tablespace.h ../src/include/catalog/pg_authid.h ../src/include/catalog/pg_auth_members.h ../src/include/catalog/pg_shdepend.h ../src/include/catalog/pg_shdescription.h ../src/include/catalog/pg_ts_config.h ../src/include/catalog/pg_ts_config_map.h ../src/include/catalog/pg_ts_dict.h ../src/include/catalog/pg_ts_parser.h ../src/include/catalog/pg_ts_template.h ../src/include/catalog/pg_extension.h ../src/include/catalog/pg_foreign_data_wrapper.h ../src/include/catalog/pg_foreign_server.h ../src/include/catalog/pg_user_mapping.h ../src/include/catalog/pg_foreign_table.h ../src/include/catalog/pg_policy.h ../src/include/catalog/pg_replication_origin.h ../src/include/catalog/pg_default_acl.h ../src/include/catalog/pg_init_privs.h ../src/include/catalog/pg_seclabel.h ../src/include/catalog/pg_shseclabel.h ../src/include/catalog/pg_collation.h ../src/include/catalog/pg_parameter_acl.h ../src/include/catalog/pg_partitioned_table.h ../src/include/catalog/pg_range.h ../src/include/catalog/pg_transform.h ../src/include/catalog/pg_sequence.h ../src/include/catalog/pg_publication.h ../src/include/catalog/pg_publication_namespace.h ../src/include/catalog/pg_publication_rel.h ../src/include/catalog/pg_subscription.h ../src/include/catalog/pg_subscription_rel.h\\n[17:26:44.602] Duplicate OIDs detected:\\n[17:26:44.602] 4550\\n[17:26:44.602] found 1 duplicate OID(s) in catalog data\\n\\nI suggest you choose a random oid out of the \\\"development purposes\\\" range:\\n\\n *\\t\\tOIDs 1-9999 are reserved for manual assignment (see .dat files in\\n *\\t\\tsrc/include/catalog/). Of these, 8000-9999 are reserved for\\n *\\t\\tdevelopment purposes (such as in-progress patches and forks);\\n *\\t\\tthey should not appear in released versions.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 6 Dec 2022 10:41:01 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2022-12-07 03:41, Andres Freund wrote:\\n> Hi,\\n> \\n> This patch does not currently build, due to a conflicting oid:\\n> \\n> I suggest you choose a random oid out of the \\\"development purposes\\\" \\n> range:\\n\\nThanks for your advice!\\nAttached updated patch.\\n\\n\\nBTW, since this patch depends on ProcessInterrupts() and EXPLAIN codes \\nwhich is used in auto_explain, I'm feeling that the following discussion \\nalso applies to this patch.\\n\\n> -- \\n> https://www.postgresql.org/message-id/CA%2BTgmoYW_rSOW4JMQ9_0Df9PKQ%3DsQDOKUGA4Gc9D8w4wui8fSA%40mail.gmail.com\\n> \\n> explaining a query is a pretty\\n> complicated operation that involves catalog lookups and lots of\\n> complicated stuff, and I don't think that it would be safe to do all\\n> of that at any arbitrary point in the code where ProcessInterrupts()\\n> happened to fire.\\n\\nIf I can't come up with some workaround during the next Commitfest, I'm \\ngoing to withdraw this proposal.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Thu, 08 Dec 2022 14:10:32 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hello,\\n\\nThanks for working on this patch!\\n\\nOn Thu, Dec 8, 2022 at 12:10 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> BTW, since this patch depends on ProcessInterrupts() and EXPLAIN codes\\n> which is used in auto_explain, I'm feeling that the following discussion\\n> also applies to this patch.\\n>\\n> > --\\n> > https://www.postgresql.org/message-id/CA%2BTgmoYW_rSOW4JMQ9_0Df9PKQ%3DsQDOKUGA4Gc9D8w4wui8fSA%40mail.gmail.com\\n> >\\n> > explaining a query is a pretty\\n> > complicated operation that involves catalog lookups and lots of\\n> > complicated stuff, and I don't think that it would be safe to do all\\n> > of that at any arbitrary point in the code where ProcessInterrupts()\\n> > happened to fire.\\n>\\n> If I can't come up with some workaround during the next Commitfest, I'm\\n> going to withdraw this proposal.\\n\\nWhile at PGCon this week I'd brought up this idea with a few people\\nwithout realizing you'd already worked on it previously, and then\\nafter I discovered this thread several of us (Greg, Ronan, David,\\nHeikki, and myself -- all cc'd) discussed the safety concerns over\\ndinner last night.\\n\\nOur conclusion was that all of the concerns we could come up with (for\\nexample, walking though the code for ExplainTargetRel() and discussing\\nthe relevant catalog and syscache accesses) all applied specifically\\nto Robert's concerns about running explain inside an aborted\\ntransaction. After all, I'd originally started that thread to ask\\nabout running auto-explain after a query timeout.\\n\\nTo put it positively: we believe that, for example, catalog accesses\\ninside CHECK_FOR_INTERRUPTS() -- assuming that the CFI call is inside\\nan existing valid transaction/query state, as it would be for this\\npatch -- are safe. If there were problems, then those problems are\\nlikely bugs we already have in other CFI cases.\\n\\nAnother concern Robert raised may apply here: what if a person tries\\nto cancel a query when we're explaining? I believe that's a reasonable\\nquestion to ask, but I believe it's a trade-off that's worth making\\nfor the (significant) introspection benefits this patch would provide.\\n\\nOn to the patch itself: overall I think it looks like it's in pretty\\ngood shape. I also noticed we don't have any tests (I assume it'd have\\nto be TAP tests) of the actual output happening, and I think it would\\nbe worth adding that.\\n\\nAre you interested in re-opening this patch? I'd be happy to provide\\nfurther review and help to try to push this along.\\n\\nI've rebased the patch and attached as v26.\\n\\nThanks,\\nJames Coleman\",\n \"msg_date\": \"Fri, 2 Jun 2023 13:51:22 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-06-03 02:51, James Coleman wrote:\\n> Hello,\\n> \\n> Thanks for working on this patch!\\n> \\n> On Thu, Dec 8, 2022 at 12:10 AM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> \\n>> BTW, since this patch depends on ProcessInterrupts() and EXPLAIN codes\\n>> which is used in auto_explain, I'm feeling that the following \\n>> discussion\\n>> also applies to this patch.\\n>> \\n>> > --\\n>> > https://www.postgresql.org/message-id/CA%2BTgmoYW_rSOW4JMQ9_0Df9PKQ%3DsQDOKUGA4Gc9D8w4wui8fSA%40mail.gmail.com\\n>> >\\n>> > explaining a query is a pretty\\n>> > complicated operation that involves catalog lookups and lots of\\n>> > complicated stuff, and I don't think that it would be safe to do all\\n>> > of that at any arbitrary point in the code where ProcessInterrupts()\\n>> > happened to fire.\\n>> \\n>> If I can't come up with some workaround during the next Commitfest, \\n>> I'm\\n>> going to withdraw this proposal.\\n> \\n> While at PGCon this week I'd brought up this idea with a few people\\n> without realizing you'd already worked on it previously, and then\\n> after I discovered this thread several of us (Greg, Ronan, David,\\n> Heikki, and myself -- all cc'd) discussed the safety concerns over\\n> dinner last night.\\n> \\n> Our conclusion was that all of the concerns we could come up with (for\\n> example, walking though the code for ExplainTargetRel() and discussing\\n> the relevant catalog and syscache accesses) all applied specifically\\n> to Robert's concerns about running explain inside an aborted\\n> transaction. After all, I'd originally started that thread to ask\\n> about running auto-explain after a query timeout.\\n> \\n> To put it positively: we believe that, for example, catalog accesses\\n> inside CHECK_FOR_INTERRUPTS() -- assuming that the CFI call is inside\\n> an existing valid transaction/query state, as it would be for this\\n> patch -- are safe. If there were problems, then those problems are\\n> likely bugs we already have in other CFI cases.\\n\\nThanks a lot for the discussion and sharing it!\\nI really appreciate it.\\n\\nBTW I'm not sure whether all the CFI are called in valid transaction,\\ndo you think we should check each of them?\\n\\n> Another concern Robert raised may apply here: what if a person tries\\n> to cancel a query when we're explaining? I believe that's a reasonable\\n> question to ask, but I believe it's a trade-off that's worth making\\n> for the (significant) introspection benefits this patch would provide.\\n\\nIs the concern here limited to the case where explain code goes crazy\\nas Robert pointed out?\\nIf so, this may be a trade-off worth doing.\\nI am a little concerned about whether there will be cases where the\\nexplain code is not problematic but just takes long time.\\n\\n> On to the patch itself: overall I think it looks like it's in pretty\\n> good shape. I also noticed we don't have any tests (I assume it'd have\\n> to be TAP tests) of the actual output happening, and I think it would\\n> be worth adding that.\\n\\nChecking the log output may be better, but I didn't add it since there\\nis only a little content that can be checked, and similar function\\npg_log_backend_memory_contexts() does not do such type of tests.\\n\\n> Are you interested in re-opening this patch? I'd be happy to provide\\n> further review and help to try to push this along.\\nSure, I'm going to re-open this.\\n\\n> I've rebased the patch and attached as v26.\\nThanks again for your work!\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Mon, 05 Jun 2023 17:30:18 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Mon, Jun 5, 2023 at 4:30 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-06-03 02:51, James Coleman wrote:\\n> > Hello,\\n> >\\n> > Thanks for working on this patch!\\n\\nSure thing! I'm *very interested* in seeing this available, and I\\nthink it paves the way for some additional features later on...\\n\\n> > On Thu, Dec 8, 2022 at 12:10 AM torikoshia <torikoshia@oss.nttdata.com>\\n> ...\\n> > To put it positively: we believe that, for example, catalog accesses\\n> > inside CHECK_FOR_INTERRUPTS() -- assuming that the CFI call is inside\\n> > an existing valid transaction/query state, as it would be for this\\n> > patch -- are safe. If there were problems, then those problems are\\n> > likely bugs we already have in other CFI cases.\\n>\\n> Thanks a lot for the discussion and sharing it!\\n> I really appreciate it.\\n>\\n> BTW I'm not sure whether all the CFI are called in valid transaction,\\n> do you think we should check each of them?\\n\\nI kicked off the regressions tests with a call to\\nProcessLogQueryPlanInterrupt() in every single CHECK_FOR_INTERRUPTS()\\ncall. Several hours and 52 GB of logs later I have confirmed that\\n(with the attached revision) at the very least the regression test\\nsuite can't trigger any kind of failures regardless of when we trigger\\nthis. The existing code in the patch for only running the explain when\\nthere's an active query handling that.\\n\\n> > Another concern Robert raised may apply here: what if a person tries\\n> > to cancel a query when we're explaining? I believe that's a reasonable\\n> > question to ask, but I believe it's a trade-off that's worth making\\n> > for the (significant) introspection benefits this patch would provide.\\n>\\n> Is the concern here limited to the case where explain code goes crazy\\n> as Robert pointed out?\\n> If so, this may be a trade-off worth doing.\\n> I am a little concerned about whether there will be cases where the\\n> explain code is not problematic but just takes long time.\\n\\nThe explain code could take a long time in some rare cases (e.g., we\\ndiscovered a bug a few years back with the planning code that actually\\ndescends an index to find the max value), but I think the trade-off is\\nworth it.\\n\\n> > On to the patch itself: overall I think it looks like it's in pretty\\n> > good shape. I also noticed we don't have any tests (I assume it'd have\\n> > to be TAP tests) of the actual output happening, and I think it would\\n> > be worth adding that.\\n>\\n> Checking the log output may be better, but I didn't add it since there\\n> is only a little content that can be checked, and similar function\\n> pg_log_backend_memory_contexts() does not do such type of tests.\\n\\nFair enough. I still think that would be an improvement here, but\\nothers could also weigh in.\\n\\n> > Are you interested in re-opening this patch? I'd be happy to provide\\n> > further review and help to try to push this along.\\n> Sure, I'm going to re-open this.\\n\\nI've attached v27. The important change here in 0001 is that it\\nguarantees the interrupt handler is re-entrant, since that was a bug\\nexposed by my testing. I've also included 0002 which is only meant for\\ntesting (it attempts to log in the plan in every\\nCHECK_FOR_INTERRUPTS() call).\\n\\nRegards,\\nJames\",\n \"msg_date\": \"Mon, 5 Jun 2023 14:26:25 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-06-06 03:26, James Coleman wrote:\\n> On Mon, Jun 5, 2023 at 4:30 AM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> \\n>> On 2023-06-03 02:51, James Coleman wrote:\\n>> > Hello,\\n>> >\\n>> > Thanks for working on this patch!\\n> \\n> Sure thing! I'm *very interested* in seeing this available, and I\\n> think it paves the way for some additional features later on...\\n> \\n>> > On Thu, Dec 8, 2022 at 12:10 AM torikoshia <torikoshia@oss.nttdata.com>\\n>> ...\\n>> > To put it positively: we believe that, for example, catalog accesses\\n>> > inside CHECK_FOR_INTERRUPTS() -- assuming that the CFI call is inside\\n>> > an existing valid transaction/query state, as it would be for this\\n>> > patch -- are safe. If there were problems, then those problems are\\n>> > likely bugs we already have in other CFI cases.\\n>> \\n>> Thanks a lot for the discussion and sharing it!\\n>> I really appreciate it.\\n>> \\n>> BTW I'm not sure whether all the CFI are called in valid transaction,\\n>> do you think we should check each of them?\\n> \\n> I kicked off the regressions tests with a call to\\n> ProcessLogQueryPlanInterrupt() in every single CHECK_FOR_INTERRUPTS()\\n> call. Several hours and 52 GB of logs later I have confirmed that\\n> (with the attached revision) at the very least the regression test\\n> suite can't trigger any kind of failures regardless of when we trigger\\n> this. The existing code in the patch for only running the explain when\\n> there's an active query handling that.\\n\\nThanks for the testing!\\n\\n> I've attached v27. The important change here in 0001 is that it\\n> guarantees the interrupt handler is re-entrant, since that was a bug\\n> exposed by my testing. I've also included 0002 which is only meant for\\n> testing (it attempts to log in the plan in every\\n> CHECK_FOR_INTERRUPTS() call).\\n\\nWhen SIGINT is sent during ProcessLogQueryPlanInterrupt(),\\nProcessLogQueryPlanInterruptActive can remain true.\\nAfter that, pg_log_query_plan() does nothing but just returns.\\n\\nAFAIU, v26 logs plan for each pg_log_query_plan() even when another\\npg_log_query_plan() is running, but it doesn't cause errors or critical\\nproblem.\\n\\nConsidering the problem solved and introduced by v27, I think v26 might\\nbe fine.\\nHow do you think?\\n\\n> \\n> Regards,\\n> James\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Jun 2023 13:07:34 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Sun, Jun 11, 2023 at 11:07 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-06-06 03:26, James Coleman wrote:\\n> > On Mon, Jun 5, 2023 at 4:30 AM torikoshia <torikoshia@oss.nttdata.com>\\n> > wrote:\\n> >>\\n> >> On 2023-06-03 02:51, James Coleman wrote:\\n> >> > Hello,\\n> >> >\\n> >> > Thanks for working on this patch!\\n> >\\n> > Sure thing! I'm *very interested* in seeing this available, and I\\n> > think it paves the way for some additional features later on...\\n> >\\n> >> > On Thu, Dec 8, 2022 at 12:10 AM torikoshia <torikoshia@oss.nttdata.com>\\n> >> ...\\n> >> > To put it positively: we believe that, for example, catalog accesses\\n> >> > inside CHECK_FOR_INTERRUPTS() -- assuming that the CFI call is inside\\n> >> > an existing valid transaction/query state, as it would be for this\\n> >> > patch -- are safe. If there were problems, then those problems are\\n> >> > likely bugs we already have in other CFI cases.\\n> >>\\n> >> Thanks a lot for the discussion and sharing it!\\n> >> I really appreciate it.\\n> >>\\n> >> BTW I'm not sure whether all the CFI are called in valid transaction,\\n> >> do you think we should check each of them?\\n> >\\n> > I kicked off the regressions tests with a call to\\n> > ProcessLogQueryPlanInterrupt() in every single CHECK_FOR_INTERRUPTS()\\n> > call. Several hours and 52 GB of logs later I have confirmed that\\n> > (with the attached revision) at the very least the regression test\\n> > suite can't trigger any kind of failures regardless of when we trigger\\n> > this. The existing code in the patch for only running the explain when\\n> > there's an active query handling that.\\n>\\n> Thanks for the testing!\\n>\\n> > I've attached v27. The important change here in 0001 is that it\\n> > guarantees the interrupt handler is re-entrant, since that was a bug\\n> > exposed by my testing. I've also included 0002 which is only meant for\\n> > testing (it attempts to log in the plan in every\\n> > CHECK_FOR_INTERRUPTS() call).\\n>\\n> When SIGINT is sent during ProcessLogQueryPlanInterrupt(),\\n> ProcessLogQueryPlanInterruptActive can remain true.\\n> After that, pg_log_query_plan() does nothing but just returns.\\n>\\n> AFAIU, v26 logs plan for each pg_log_query_plan() even when another\\n> pg_log_query_plan() is running, but it doesn't cause errors or critical\\n> problem.\\n>\\n> Considering the problem solved and introduced by v27, I think v26 might\\n> be fine.\\n> How do you think?\\n\\nThe testing I did with calling this during every CFI is what uncovered\\nthe re-entrancy problem. IIRC (without running that test again) the\\nproblem was a stack overflow. Now, to be sure this is a particularly\\ndegenerate case because in real-world usage it'd be impossible in\\npractice, I think, to trigger that many calls to this function (and by\\nextension the interrupt handler).\\n\\nIf SIGINT is the only concern we could reset\\nProcessLogQueryPlanInterruptActive in error handling code. I admit\\nthat part of my thought process here is thinking ahead to an\\nadditional patch I'd like to see on top of this, which is logging a\\nquery plan before cleaning up when statement timeout occurs. The\\nre-entrancy issue becomes more interesting then, I think, since we\\nwould then have automated calling of the logging code. BTW: I'd\\nthought that would make a nice follow-up patch for this, but if you'd\\nprefer I could add it as another patch in the series here.\\n\\nWhat do you think about resetting the flag versus just not having it?\\n\\nRegards,\\nJames Coleman\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Jun 2023 10:52:01 -0500\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-06-13 00:52, James Coleman wrote:\\n>> \\n>> > I've attached v27. The important change here in 0001 is that it\\n>> > guarantees the interrupt handler is re-entrant, since that was a bug\\n>> > exposed by my testing. I've also included 0002 which is only meant for\\n>> > testing (it attempts to log in the plan in every\\n>> > CHECK_FOR_INTERRUPTS() call).\\n>> \\n>> When SIGINT is sent during ProcessLogQueryPlanInterrupt(),\\n>> ProcessLogQueryPlanInterruptActive can remain true.\\n>> After that, pg_log_query_plan() does nothing but just returns.\\n>> \\n>> AFAIU, v26 logs plan for each pg_log_query_plan() even when another\\n>> pg_log_query_plan() is running, but it doesn't cause errors or \\n>> critical\\n>> problem.\\n>> \\n>> Considering the problem solved and introduced by v27, I think v26 \\n>> might\\n>> be fine.\\n>> How do you think?\\n> \\n> The testing I did with calling this during every CFI is what uncovered\\n> the re-entrancy problem. IIRC (without running that test again) the\\n> problem was a stack overflow. Now, to be sure this is a particularly\\n> degenerate case because in real-world usage it'd be impossible in\\n> practice, I think, to trigger that many calls to this function (and by\\n> extension the interrupt handler).\\n\\nYeah.In addition, currently only superusers are allowed to execute\\npg_log_query_plan(), I think we don't need to think about cases\\nwhere users are malicious.\\n\\n> If SIGINT is the only concern we could reset\\n> ProcessLogQueryPlanInterruptActive in error handling code. I admit\\n> that part of my thought process here is thinking ahead to an\\n> additional patch I'd like to see on top of this, which is logging a\\n> query plan before cleaning up when statement timeout occurs.\\n\\nI remember this is what you wanted do.[1]\\n\\n> The\\n> re-entrancy issue becomes more interesting then, I think, since we\\n> would then have automated calling of the logging code. BTW: I'd\\n> thought that would make a nice follow-up patch for this, but if you'd\\n> prefer I could add it as another patch in the series here.\\n> \\n> What do you think about resetting the flag versus just not having it?\\n\\nIf I understand you correctly, adding the flag is not necessary for this\\nproposal.\\nTo keep the patch simple, I prefer not having it.\\n\\n\\n[1] \\nhttps://www.postgresql.org/message-id/flat/CA%2BTgmoYW_rSOW4JMQ9_0Df9PKQ%3DsQDOKUGA4Gc9D8w4wui8fSA%40mail.gmail.com#b57432077f8045be8588049269f7a8dd\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Jun 2023 00:22:14 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Tue, Jun 13, 2023 at 11:22 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-06-13 00:52, James Coleman wrote:\\n> >>\\n> >> > I've attached v27. The important change here in 0001 is that it\\n> >> > guarantees the interrupt handler is re-entrant, since that was a bug\\n> >> > exposed by my testing. I've also included 0002 which is only meant for\\n> >> > testing (it attempts to log in the plan in every\\n> >> > CHECK_FOR_INTERRUPTS() call).\\n> >>\\n> >> When SIGINT is sent during ProcessLogQueryPlanInterrupt(),\\n> >> ProcessLogQueryPlanInterruptActive can remain true.\\n> >> After that, pg_log_query_plan() does nothing but just returns.\\n> >>\\n> >> AFAIU, v26 logs plan for each pg_log_query_plan() even when another\\n> >> pg_log_query_plan() is running, but it doesn't cause errors or\\n> >> critical\\n> >> problem.\\n> >>\\n> >> Considering the problem solved and introduced by v27, I think v26\\n> >> might\\n> >> be fine.\\n> >> How do you think?\\n> >\\n> > The testing I did with calling this during every CFI is what uncovered\\n> > the re-entrancy problem. IIRC (without running that test again) the\\n> > problem was a stack overflow. Now, to be sure this is a particularly\\n> > degenerate case because in real-world usage it'd be impossible in\\n> > practice, I think, to trigger that many calls to this function (and by\\n> > extension the interrupt handler).\\n>\\n> Yeah.In addition, currently only superusers are allowed to execute\\n> pg_log_query_plan(), I think we don't need to think about cases\\n> where users are malicious.\\n>\\n> > If SIGINT is the only concern we could reset\\n> > ProcessLogQueryPlanInterruptActive in error handling code. I admit\\n> > that part of my thought process here is thinking ahead to an\\n> > additional patch I'd like to see on top of this, which is logging a\\n> > query plan before cleaning up when statement timeout occurs.\\n>\\n> I remember this is what you wanted do.[1]\\n>\\n> > The\\n> > re-entrancy issue becomes more interesting then, I think, since we\\n> > would then have automated calling of the logging code. BTW: I'd\\n> > thought that would make a nice follow-up patch for this, but if you'd\\n> > prefer I could add it as another patch in the series here.\\n> >\\n> > What do you think about resetting the flag versus just not having it?\\n>\\n> If I understand you correctly, adding the flag is not necessary for this\\n> proposal.\\n> To keep the patch simple, I prefer not having it.\\n>\\n\\nI'm going to re-run tests with my patch version + resetting the flag\\non SIGINT (and any other error condition) to be certain that the issue\\nyou uncovered (where backends get stuck after a SIGINT not responding\\nto the requested plan logging) wasn't masking any other issues.\\n\\nAs long as that run is clean also then I believe the patch is safe\\nas-is even without the re-entrancy guard.\\n\\nI'll report back with the results of that testing.\\n\\nRegards,\\nJames Coleman\\n\\n\\n\",\n \"msg_date\": \"Tue, 13 Jun 2023 11:53:39 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Tue, Jun 13, 2023 at 11:53 AM James Coleman <jtc331@gmail.com> wrote:\\n>\\n> ...\\n> I'm going to re-run tests with my patch version + resetting the flag\\n> on SIGINT (and any other error condition) to be certain that the issue\\n> you uncovered (where backends get stuck after a SIGINT not responding\\n> to the requested plan logging) wasn't masking any other issues.\\n>\\n> As long as that run is clean also then I believe the patch is safe\\n> as-is even without the re-entrancy guard.\\n>\\n> I'll report back with the results of that testing.\\n\\nThe tests have been running since last night, but have been apparently\\nhung now for many hours. I haven't been able to fully look into it,\\nbut so far I know the hung (100% CPU) backend last logged this:\\n\\n2023-06-14 02:00:30.045 UTC client backend[84461]\\npg_regress/updatable_views LOG: query plan running on backend with\\nPID 84461 is:\\n Query Text: SELECT table_name, column_name, is_updatable\\n FROM information_schema.columns\\n WHERE table_name LIKE E'r_\\\\\\\\_view%'\\n ORDER BY table_name, ordinal_position;\\n\\nThe last output from the regression test harness was:\\n\\n# parallel group (5 tests): index_including create_view\\nindex_including_gist create_index create_index_spgist\\nok 66 + create_index 36508 ms\\nok 67 + create_index_spgist 38588 ms\\nok 68 + create_view 1394 ms\\nok 69 + index_including 654 ms\\nok 70 + index_including_gist 1701 ms\\n# parallel group (16 tests): errors create_cast drop_if_exists\\ncreate_aggregate roleattributes constraints hash_func typed_table\\ninfinite_recurse\\n\\nAttaching gdb to the hung backend shows this:\\n\\n#0 0x00005601ab1f9529 in ProcLockWakeup\\n(lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n<default_lockmethod>, lock=lock@entry=0x7f5325c913f0) at proc.c:1655\\n#1 0x00005601ab1e99dc in CleanUpLock (lock=lock@entry=0x7f5325c913f0,\\nproclock=proclock@entry=0x7f5325d40d60,\\nlockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n<default_lockmethod>,\\n hashcode=hashcode@entry=573498161, wakeupNeeded=<optimized out>)\\nat lock.c:1673\\n#2 0x00005601ab1e9e21 in LockRefindAndRelease\\n(lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n<default_lockmethod>, proc=<optimized out>,\\nlocktag=locktag@entry=0x5601ac3d7998, lockmode=lockmode@entry=1,\\n decrement_strong_lock_count=decrement_strong_lock_count@entry=false)\\nat lock.c:3150\\n#3 0x00005601ab1edb27 in LockReleaseAll\\n(lockmethodid=lockmethodid@entry=1, allLocks=false) at lock.c:2295\\n#4 0x00005601ab1f8599 in ProcReleaseLocks\\n(isCommit=isCommit@entry=true) at proc.c:781\\n#5 0x00005601ab37f1f4 in ResourceOwnerReleaseInternal\\n(owner=<optimized out>, phase=phase@entry=RESOURCE_RELEASE_LOCKS,\\nisCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\\nresowner.c:618\\n#6 0x00005601ab37f7b7 in ResourceOwnerRelease (owner=<optimized out>,\\nphase=phase@entry=RESOURCE_RELEASE_LOCKS,\\nisCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\\nresowner.c:494\\n#7 0x00005601aaec1d84 in CommitTransaction () at xact.c:2334\\n#8 0x00005601aaec2b22 in CommitTransactionCommand () at xact.c:3067\\n#9 0x00005601ab200a66 in finish_xact_command () at postgres.c:2783\\n#10 0x00005601ab20338f in exec_simple_query (\\n query_string=query_string@entry=0x5601ac3b0858 \\\"SELECT table_name,\\ncolumn_name, is_updatable\\\\n FROM information_schema.columns\\\\n WHERE\\ntable_name LIKE E'r_\\\\\\\\\\\\\\\\_view%'\\\\n ORDER BY table_name,\\nordinal_position;\\\") at postgres.c:1300\\n\\nI am unable to connect to the regression test Postgres instance --\\npsql just hangs, so the lock seems to have affected the postmaster\\nalso.\\n\\nI'm wondering if this might represent a bug in the current patch.\\n\\nRegards,\\nJames Coleman\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Jun 2023 12:48:52 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-06-15 01:48, James Coleman wrote:\\n> On Tue, Jun 13, 2023 at 11:53 AM James Coleman <jtc331@gmail.com> \\n> wrote:\\n>> \\n>> ...\\n>> I'm going to re-run tests with my patch version + resetting the flag\\n>> on SIGINT (and any other error condition) to be certain that the issue\\n>> you uncovered (where backends get stuck after a SIGINT not responding\\n>> to the requested plan logging) wasn't masking any other issues.\\n>> \\n>> As long as that run is clean also then I believe the patch is safe\\n>> as-is even without the re-entrancy guard.\\n>> \\n>> I'll report back with the results of that testing.\\n> \\n> The tests have been running since last night, but have been apparently\\n> hung now for many hours. I haven't been able to fully look into it,\\n> but so far I know the hung (100% CPU) backend last logged this:\\n> \\n> 2023-06-14 02:00:30.045 UTC client backend[84461]\\n> pg_regress/updatable_views LOG: query plan running on backend with\\n> PID 84461 is:\\n> Query Text: SELECT table_name, column_name, is_updatable\\n> FROM information_schema.columns\\n> WHERE table_name LIKE E'r_\\\\\\\\_view%'\\n> ORDER BY table_name, ordinal_position;\\n> \\n> The last output from the regression test harness was:\\n> \\n> # parallel group (5 tests): index_including create_view\\n> index_including_gist create_index create_index_spgist\\n> ok 66 + create_index 36508 ms\\n> ok 67 + create_index_spgist 38588 ms\\n> ok 68 + create_view 1394 ms\\n> ok 69 + index_including 654 ms\\n> ok 70 + index_including_gist 1701 ms\\n> # parallel group (16 tests): errors create_cast drop_if_exists\\n> create_aggregate roleattributes constraints hash_func typed_table\\n> infinite_recurse\\n> \\n> Attaching gdb to the hung backend shows this:\\n> \\n> #0 0x00005601ab1f9529 in ProcLockWakeup\\n> (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n> <default_lockmethod>, lock=lock@entry=0x7f5325c913f0) at proc.c:1655\\n> #1 0x00005601ab1e99dc in CleanUpLock (lock=lock@entry=0x7f5325c913f0,\\n> proclock=proclock@entry=0x7f5325d40d60,\\n> lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n> <default_lockmethod>,\\n> hashcode=hashcode@entry=573498161, wakeupNeeded=<optimized out>)\\n> at lock.c:1673\\n> #2 0x00005601ab1e9e21 in LockRefindAndRelease\\n> (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n> <default_lockmethod>, proc=<optimized out>,\\n> locktag=locktag@entry=0x5601ac3d7998, lockmode=lockmode@entry=1,\\n> \\n> decrement_strong_lock_count=decrement_strong_lock_count@entry=false)\\n> at lock.c:3150\\n> #3 0x00005601ab1edb27 in LockReleaseAll\\n> (lockmethodid=lockmethodid@entry=1, allLocks=false) at lock.c:2295\\n> #4 0x00005601ab1f8599 in ProcReleaseLocks\\n> (isCommit=isCommit@entry=true) at proc.c:781\\n> #5 0x00005601ab37f1f4 in ResourceOwnerReleaseInternal\\n> (owner=<optimized out>, phase=phase@entry=RESOURCE_RELEASE_LOCKS,\\n> isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\\n> resowner.c:618\\n> #6 0x00005601ab37f7b7 in ResourceOwnerRelease (owner=<optimized out>,\\n> phase=phase@entry=RESOURCE_RELEASE_LOCKS,\\n> isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\\n> resowner.c:494\\n> #7 0x00005601aaec1d84 in CommitTransaction () at xact.c:2334\\n> #8 0x00005601aaec2b22 in CommitTransactionCommand () at xact.c:3067\\n> #9 0x00005601ab200a66 in finish_xact_command () at postgres.c:2783\\n> #10 0x00005601ab20338f in exec_simple_query (\\n> query_string=query_string@entry=0x5601ac3b0858 \\\"SELECT table_name,\\n> column_name, is_updatable\\\\n FROM information_schema.columns\\\\n WHERE\\n> table_name LIKE E'r_\\\\\\\\\\\\\\\\_view%'\\\\n ORDER BY table_name,\\n> ordinal_position;\\\") at postgres.c:1300\\n> \\n> I am unable to connect to the regression test Postgres instance --\\n> psql just hangs, so the lock seems to have affected the postmaster\\n> also.\\n> \\n> I'm wondering if this might represent a bug in the current patch.\\n\\nThanks for running and analyzing the test!\\n\\nCould you share me how you are running the test?\\n\\nI imagined something like below, but currently couldn't reproduce it.\\n- apply both v26-0001 and v27-0002 and build\\n- run PostgreSQL with default GUCssaaa\\n- make installcheck-world\\n- run 'SELECT pg_log_query_plan(pid) FROM pg_stat_activity \\\\watch 0.1' \\nduring make installcheck-world\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Thu, 15 Jun 2023 22:00:46 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 15, 2023 at 9:00 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-06-15 01:48, James Coleman wrote:\\n> > On Tue, Jun 13, 2023 at 11:53 AM James Coleman <jtc331@gmail.com>\\n> > wrote:\\n> >>\\n> >> ...\\n> >> I'm going to re-run tests with my patch version + resetting the flag\\n> >> on SIGINT (and any other error condition) to be certain that the issue\\n> >> you uncovered (where backends get stuck after a SIGINT not responding\\n> >> to the requested plan logging) wasn't masking any other issues.\\n> >>\\n> >> As long as that run is clean also then I believe the patch is safe\\n> >> as-is even without the re-entrancy guard.\\n> >>\\n> >> I'll report back with the results of that testing.\\n> >\\n> > The tests have been running since last night, but have been apparently\\n> > hung now for many hours. I haven't been able to fully look into it,\\n> > but so far I know the hung (100% CPU) backend last logged this:\\n> >\\n> > 2023-06-14 02:00:30.045 UTC client backend[84461]\\n> > pg_regress/updatable_views LOG: query plan running on backend with\\n> > PID 84461 is:\\n> > Query Text: SELECT table_name, column_name, is_updatable\\n> > FROM information_schema.columns\\n> > WHERE table_name LIKE E'r_\\\\\\\\_view%'\\n> > ORDER BY table_name, ordinal_position;\\n> >\\n> > The last output from the regression test harness was:\\n> >\\n> > # parallel group (5 tests): index_including create_view\\n> > index_including_gist create_index create_index_spgist\\n> > ok 66 + create_index 36508 ms\\n> > ok 67 + create_index_spgist 38588 ms\\n> > ok 68 + create_view 1394 ms\\n> > ok 69 + index_including 654 ms\\n> > ok 70 + index_including_gist 1701 ms\\n> > # parallel group (16 tests): errors create_cast drop_if_exists\\n> > create_aggregate roleattributes constraints hash_func typed_table\\n> > infinite_recurse\\n> >\\n> > Attaching gdb to the hung backend shows this:\\n> >\\n> > #0 0x00005601ab1f9529 in ProcLockWakeup\\n> > (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n> > <default_lockmethod>, lock=lock@entry=0x7f5325c913f0) at proc.c:1655\\n> > #1 0x00005601ab1e99dc in CleanUpLock (lock=lock@entry=0x7f5325c913f0,\\n> > proclock=proclock@entry=0x7f5325d40d60,\\n> > lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n> > <default_lockmethod>,\\n> > hashcode=hashcode@entry=573498161, wakeupNeeded=<optimized out>)\\n> > at lock.c:1673\\n> > #2 0x00005601ab1e9e21 in LockRefindAndRelease\\n> > (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n> > <default_lockmethod>, proc=<optimized out>,\\n> > locktag=locktag@entry=0x5601ac3d7998, lockmode=lockmode@entry=1,\\n> >\\n> > decrement_strong_lock_count=decrement_strong_lock_count@entry=false)\\n> > at lock.c:3150\\n> > #3 0x00005601ab1edb27 in LockReleaseAll\\n> > (lockmethodid=lockmethodid@entry=1, allLocks=false) at lock.c:2295\\n> > #4 0x00005601ab1f8599 in ProcReleaseLocks\\n> > (isCommit=isCommit@entry=true) at proc.c:781\\n> > #5 0x00005601ab37f1f4 in ResourceOwnerReleaseInternal\\n> > (owner=<optimized out>, phase=phase@entry=RESOURCE_RELEASE_LOCKS,\\n> > isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\\n> > resowner.c:618\\n> > #6 0x00005601ab37f7b7 in ResourceOwnerRelease (owner=<optimized out>,\\n> > phase=phase@entry=RESOURCE_RELEASE_LOCKS,\\n> > isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\\n> > resowner.c:494\\n> > #7 0x00005601aaec1d84 in CommitTransaction () at xact.c:2334\\n> > #8 0x00005601aaec2b22 in CommitTransactionCommand () at xact.c:3067\\n> > #9 0x00005601ab200a66 in finish_xact_command () at postgres.c:2783\\n> > #10 0x00005601ab20338f in exec_simple_query (\\n> > query_string=query_string@entry=0x5601ac3b0858 \\\"SELECT table_name,\\n> > column_name, is_updatable\\\\n FROM information_schema.columns\\\\n WHERE\\n> > table_name LIKE E'r_\\\\\\\\\\\\\\\\_view%'\\\\n ORDER BY table_name,\\n> > ordinal_position;\\\") at postgres.c:1300\\n> >\\n> > I am unable to connect to the regression test Postgres instance --\\n> > psql just hangs, so the lock seems to have affected the postmaster\\n> > also.\\n> >\\n> > I'm wondering if this might represent a bug in the current patch.\\n>\\n> Thanks for running and analyzing the test!\\n\\nSure thing!\\n\\n> Could you share me how you are running the test?\\n>\\n> I imagined something like below, but currently couldn't reproduce it.\\n> - apply both v26-0001 and v27-0002 and build\\n> - run PostgreSQL with default GUCssaaa\\n> - make installcheck-world\\n> - run 'SELECT pg_log_query_plan(pid) FROM pg_stat_activity \\\\watch 0.1'\\n> during make installcheck-world\\n\\nApologies, I should have attached my updated patch (with the fix for\\nthe bug you'd reporting with the re-entrancy guard). Attached is v28\\nwhich sets ProcessLogQueryPlanInterruptActive to false in errfinish\\nwhen necessary. Once built with those two patches I'm simply running\\n`make check`.\\n\\nRegards,\\nJames Coleman\",\n \"msg_date\": \"Thu, 15 Jun 2023 12:34:23 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-06-16 01:34, James Coleman wrote:\\n> On Thu, Jun 15, 2023 at 9:00 AM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> \\n>> On 2023-06-15 01:48, James Coleman wrote:\\n>> > On Tue, Jun 13, 2023 at 11:53 AM James Coleman <jtc331@gmail.com>\\n>> > wrote:\\n>> >>\\n>> >> ...\\n>> >> I'm going to re-run tests with my patch version + resetting the flag\\n>> >> on SIGINT (and any other error condition) to be certain that the issue\\n>> >> you uncovered (where backends get stuck after a SIGINT not responding\\n>> >> to the requested plan logging) wasn't masking any other issues.\\n>> >>\\n>> >> As long as that run is clean also then I believe the patch is safe\\n>> >> as-is even without the re-entrancy guard.\\n>> >>\\n>> >> I'll report back with the results of that testing.\\n>> >\\n>> > The tests have been running since last night, but have been apparently\\n>> > hung now for many hours. I haven't been able to fully look into it,\\n>> > but so far I know the hung (100% CPU) backend last logged this:\\n>> >\\n>> > 2023-06-14 02:00:30.045 UTC client backend[84461]\\n>> > pg_regress/updatable_views LOG: query plan running on backend with\\n>> > PID 84461 is:\\n>> > Query Text: SELECT table_name, column_name, is_updatable\\n>> > FROM information_schema.columns\\n>> > WHERE table_name LIKE E'r_\\\\\\\\_view%'\\n>> > ORDER BY table_name, ordinal_position;\\n>> >\\n>> > The last output from the regression test harness was:\\n>> >\\n>> > # parallel group (5 tests): index_including create_view\\n>> > index_including_gist create_index create_index_spgist\\n>> > ok 66 + create_index 36508 ms\\n>> > ok 67 + create_index_spgist 38588 ms\\n>> > ok 68 + create_view 1394 ms\\n>> > ok 69 + index_including 654 ms\\n>> > ok 70 + index_including_gist 1701 ms\\n>> > # parallel group (16 tests): errors create_cast drop_if_exists\\n>> > create_aggregate roleattributes constraints hash_func typed_table\\n>> > infinite_recurse\\n>> >\\n>> > Attaching gdb to the hung backend shows this:\\n>> >\\n>> > #0 0x00005601ab1f9529 in ProcLockWakeup\\n>> > (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n>> > <default_lockmethod>, lock=lock@entry=0x7f5325c913f0) at proc.c:1655\\n>> > #1 0x00005601ab1e99dc in CleanUpLock (lock=lock@entry=0x7f5325c913f0,\\n>> > proclock=proclock@entry=0x7f5325d40d60,\\n>> > lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n>> > <default_lockmethod>,\\n>> > hashcode=hashcode@entry=573498161, wakeupNeeded=<optimized out>)\\n>> > at lock.c:1673\\n>> > #2 0x00005601ab1e9e21 in LockRefindAndRelease\\n>> > (lockMethodTable=lockMethodTable@entry=0x5601ab6484e0\\n>> > <default_lockmethod>, proc=<optimized out>,\\n>> > locktag=locktag@entry=0x5601ac3d7998, lockmode=lockmode@entry=1,\\n>> >\\n>> > decrement_strong_lock_count=decrement_strong_lock_count@entry=false)\\n>> > at lock.c:3150\\n>> > #3 0x00005601ab1edb27 in LockReleaseAll\\n>> > (lockmethodid=lockmethodid@entry=1, allLocks=false) at lock.c:2295\\n>> > #4 0x00005601ab1f8599 in ProcReleaseLocks\\n>> > (isCommit=isCommit@entry=true) at proc.c:781\\n>> > #5 0x00005601ab37f1f4 in ResourceOwnerReleaseInternal\\n>> > (owner=<optimized out>, phase=phase@entry=RESOURCE_RELEASE_LOCKS,\\n>> > isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\\n>> > resowner.c:618\\n>> > #6 0x00005601ab37f7b7 in ResourceOwnerRelease (owner=<optimized out>,\\n>> > phase=phase@entry=RESOURCE_RELEASE_LOCKS,\\n>> > isCommit=isCommit@entry=true, isTopLevel=isTopLevel@entry=true) at\\n>> > resowner.c:494\\n>> > #7 0x00005601aaec1d84 in CommitTransaction () at xact.c:2334\\n>> > #8 0x00005601aaec2b22 in CommitTransactionCommand () at xact.c:3067\\n>> > #9 0x00005601ab200a66 in finish_xact_command () at postgres.c:2783\\n>> > #10 0x00005601ab20338f in exec_simple_query (\\n>> > query_string=query_string@entry=0x5601ac3b0858 \\\"SELECT table_name,\\n>> > column_name, is_updatable\\\\n FROM information_schema.columns\\\\n WHERE\\n>> > table_name LIKE E'r_\\\\\\\\\\\\\\\\_view%'\\\\n ORDER BY table_name,\\n>> > ordinal_position;\\\") at postgres.c:1300\\n>> >\\n>> > I am unable to connect to the regression test Postgres instance --\\n>> > psql just hangs, so the lock seems to have affected the postmaster\\n>> > also.\\n>> >\\n>> > I'm wondering if this might represent a bug in the current patch.\\n>> \\n>> Thanks for running and analyzing the test!\\n> \\n> Sure thing!\\n> \\n>> Could you share me how you are running the test?\\n>> \\n>> I imagined something like below, but currently couldn't reproduce it.\\n>> - apply both v26-0001 and v27-0002 and build\\n>> - run PostgreSQL with default GUCssaaa\\n>> - make installcheck-world\\n>> - run 'SELECT pg_log_query_plan(pid) FROM pg_stat_activity \\\\watch 0.1'\\n>> during make installcheck-world\\n> \\n> Apologies, I should have attached my updated patch (with the fix for\\n> the bug you'd reporting with the re-entrancy guard). Attached is v28\\n> which sets ProcessLogQueryPlanInterruptActive to false in errfinish\\n> when necessary. Once built with those two patches I'm simply running\\n> `make check`.\\n\\nThanks!\\n\\nHowever, I haven't succeeded in reproducing the problem as below:\\n\\n> The tests have been running since last night, but have been apparently\\n> hung now for many hours. I haven't been able to fully look into it,\\n> but so far I know the hung (100% CPU) backend last logged this:\\n\\nDid you do something like this?\\n\\n $ ./configure --prefix=/home/ubuntu/pgsql/master --enable-cassert\\n $ git apply \\n../patch/v28-0001-Add-function-to-log-the-plan-of-the-query.patch \\n../patch/v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch\\n $ make\\n $ make check\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Tue, 27 Jun 2023 19:03:21 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-06-16 01:34, James Coleman wrote:\\n> Attached is v28\\n> which sets ProcessLogQueryPlanInterruptActive to false in errfinish\\n> when necessary. Once built with those two patches I'm simply running\\n> `make check`.\\n\\nWith v28-0001 and v28-0002 patch, I confirmed backend processes consume \\nhuge\\namount of memory and under some environments they were terminated by OOM\\nkiller.\\n\\nThis was because memory was allocated from existing memory contexts and \\nthey\\nwere not freed after ProcessLogQueryPlanInterrupt().\\nUpdated the patch to use dedicated memory context for\\nProcessLogQueryPlanInterrupt().\\n\\nApplying attached patch and v28-0002 patch, `make check` successfully\\ncompleted after 20min and 50GB of logs on my environment.\\n\\n>>> On 2023-06-15 01:48, James Coleman wrote:\\n>>> > The tests have been running since last night, but have been apparently\\n>>> > hung now for many hours.\\n\\nI don't know if this has anything to do with the hung you faced, but I \\nthought\\nit might be possible that the large amount of memory usage resulted in\\nswapping, which caused a significant delay in processing.\\n\\nIf possible, I would be very grateful if you could try to reproduce this \\nwith\\nthe v29 patch.\\n\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Thu, 17 Aug 2023 23:02:22 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 17, 2023 at 10:02 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-06-16 01:34, James Coleman wrote:\\n> > Attached is v28\\n> > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\\n> > when necessary. Once built with those two patches I'm simply running\\n> > `make check`.\\n>\\n> With v28-0001 and v28-0002 patch, I confirmed backend processes consume\\n> huge\\n> amount of memory and under some environments they were terminated by OOM\\n> killer.\\n>\\n> This was because memory was allocated from existing memory contexts and\\n> they\\n> were not freed after ProcessLogQueryPlanInterrupt().\\n> Updated the patch to use dedicated memory context for\\n> ProcessLogQueryPlanInterrupt().\\n>\\n> Applying attached patch and v28-0002 patch, `make check` successfully\\n> completed after 20min and 50GB of logs on my environment.\\n>\\n> >>> On 2023-06-15 01:48, James Coleman wrote:\\n> >>> > The tests have been running since last night, but have been apparently\\n> >>> > hung now for many hours.\\n>\\n> I don't know if this has anything to do with the hung you faced, but I\\n> thought\\n> it might be possible that the large amount of memory usage resulted in\\n> swapping, which caused a significant delay in processing.\\n\\nAh, yes, I think that could be a possible explanation. I was delaying\\non this thread because I wasn't comfortable with having caused an\\nissue once (even if I couldn't easily reproduce) without at least some\\ntheory as to the cause (and a fix).\\n\\n> If possible, I would be very grateful if you could try to reproduce this\\n> with\\n> the v29 patch.\\n\\nI'll kick off some testing.\\n\\nThanks,\\nJames Coleman\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Aug 2023 07:43:26 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 25, 2023 at 7:43 AM James Coleman <jtc331@gmail.com> wrote:\\n>\\n> On Thu, Aug 17, 2023 at 10:02 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n> >\\n> > On 2023-06-16 01:34, James Coleman wrote:\\n> > > Attached is v28\\n> > > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\\n> > > when necessary. Once built with those two patches I'm simply running\\n> > > `make check`.\\n> >\\n> > With v28-0001 and v28-0002 patch, I confirmed backend processes consume\\n> > huge\\n> > amount of memory and under some environments they were terminated by OOM\\n> > killer.\\n> >\\n> > This was because memory was allocated from existing memory contexts and\\n> > they\\n> > were not freed after ProcessLogQueryPlanInterrupt().\\n> > Updated the patch to use dedicated memory context for\\n> > ProcessLogQueryPlanInterrupt().\\n> >\\n> > Applying attached patch and v28-0002 patch, `make check` successfully\\n> > completed after 20min and 50GB of logs on my environment.\\n> >\\n> > >>> On 2023-06-15 01:48, James Coleman wrote:\\n> > >>> > The tests have been running since last night, but have been apparently\\n> > >>> > hung now for many hours.\\n> >\\n> > I don't know if this has anything to do with the hung you faced, but I\\n> > thought\\n> > it might be possible that the large amount of memory usage resulted in\\n> > swapping, which caused a significant delay in processing.\\n>\\n> Ah, yes, I think that could be a possible explanation. I was delaying\\n> on this thread because I wasn't comfortable with having caused an\\n> issue once (even if I couldn't easily reproduce) without at least some\\n> theory as to the cause (and a fix).\\n>\\n> > If possible, I would be very grateful if you could try to reproduce this\\n> > with\\n> > the v29 patch.\\n>\\n> I'll kick off some testing.\\n>\\n\\nI don't have time to investigate what's happening here, but 24 hours\\nlater the first \\\"make check\\\" is still running, and at first glance it\\nseems to have the same behavior I'd seen that first time. The test\\noutput is to this point:\\n\\n# parallel group (5 tests): index_including create_view\\nindex_including_gist create_index create_index_spgist\\nok 66 + create_index 26365 ms\\nok 67 + create_index_spgist 27675 ms\\nok 68 + create_view 1235 ms\\nok 69 + index_including 1102 ms\\nok 70 + index_including_gist 1633 ms\\n# parallel group (16 tests): create_aggregate create_cast errors\\nroleattributes drop_if_exists hash_func typed_table create_am\\ninfinite_recurse\\n\\nand it hasn't progressed past that point since at least ~16 hours ago\\n(the first several hours of the run I wasn't monitoring it).\\n\\nI haven't connected up gdb yet, and won't be able to until maybe\\ntomorrow, but here's the ps output for postgres processes that are\\nrunning:\\n\\nadmin 3213249 0.0 0.0 196824 20552 ? Ss Aug25 0:00\\n/home/admin/postgresql-test/bin/postgres -D\\n/home/admin/postgresql-test-data\\nadmin 3213250 0.0 0.0 196964 7284 ? Ss Aug25 0:00\\npostgres: checkpointer\\nadmin 3213251 0.0 0.0 196956 4276 ? Ss Aug25 0:00\\npostgres: background writer\\nadmin 3213253 0.0 0.0 196956 8600 ? Ss Aug25 0:00\\npostgres: walwriter\\nadmin 3213254 0.0 0.0 198424 7316 ? Ss Aug25 0:00\\npostgres: autovacuum launcher\\nadmin 3213255 0.0 0.0 198412 5488 ? Ss Aug25 0:00\\npostgres: logical replication launcher\\nadmin 3237967 0.0 0.0 2484 516 pts/4 S+ Aug25 0:00\\n/bin/sh -c echo \\\"# +++ regress check in src/test/regress +++\\\" &&\\nPATH=\\\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/bin:/home/admin/postgres/src/test/regress:$PATH\\\"\\nLD_LIBRARY_PATH=\\\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/lib\\\"\\nINITDB_TEMPLATE='/home/admin/postgres'/tmp_install/initdb-template\\n../../../src/test/regress/pg_regress --temp-instance=./tmp_check\\n--inputdir=. --bindir= --dlpath=. --max-concurrent-tests=20\\n--schedule=./parallel_schedule\\nadmin 3237973 0.0 0.0 197880 20688 pts/4 S+ Aug25 0:00\\npostgres -D /home/admin/postgres/src/test/regress/tmp_check/data -F -c\\nlisten_addresses= -k /tmp/pg_regress-7mmGUa\\nadmin 3237976 0.0 0.1 198332 44608 ? Ss Aug25 0:00\\npostgres: checkpointer\\nadmin 3237977 0.0 0.0 198032 4640 ? Ss Aug25 0:00\\npostgres: background writer\\nadmin 3237979 0.0 0.0 197880 8580 ? Ss Aug25 0:00\\npostgres: walwriter\\nadmin 3237980 0.0 0.0 199484 7608 ? Ss Aug25 0:00\\npostgres: autovacuum launcher\\nadmin 3237981 0.0 0.0 199460 5488 ? Ss Aug25 0:00\\npostgres: logical replication launcher\\nadmin 3243644 0.0 0.2 252400 74656 ? Ss Aug25 0:01\\npostgres: admin regression [local] SELECT waiting\\nadmin 3243645 0.0 0.1 205480 33992 ? Ss Aug25 0:00\\npostgres: admin regression [local] SELECT waiting\\nadmin 3243654 99.9 0.0 203156 31504 ? Rs Aug25 1437:49\\npostgres: admin regression [local] VACUUM\\nadmin 3243655 0.0 0.1 212036 38504 ? Ss Aug25 0:00\\npostgres: admin regression [local] SELECT waiting\\nadmin 3243656 0.0 0.0 206024 30892 ? Ss Aug25 0:00\\npostgres: admin regression [local] DELETE waiting\\nadmin 3243657 0.0 0.1 205568 32232 ? Ss Aug25 0:00\\npostgres: admin regression [local] ALTER TABLE waiting\\nadmin 3243658 0.0 0.0 203740 21532 ? Ss Aug25 0:00\\npostgres: admin regression [local] ANALYZE waiting\\nadmin 3243798 0.0 0.0 199884 8464 ? Ss Aug25 0:00\\npostgres: autovacuum worker\\nadmin 3244733 0.0 0.0 199492 5956 ? Ss Aug25 0:00\\npostgres: autovacuum worker\\nadmin 3245652 0.0 0.0 199884 8468 ? Ss Aug25 0:00\\npostgres: autovacuum worker\\n\\nAs you can see there are a bunch of backends presumably waiting, and\\nalso the VACUUM process has been pegging a single CPU core for at\\nleast since that ~16 hour ago mark.\\n\\nI hope to be able to do more investigation later, but I wanted to at\\nleast give you this information now.\\n\\nJames Coleman\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Aug 2023 08:03:26 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-08-26 21:03, James Coleman wrote:\\n> On Fri, Aug 25, 2023 at 7:43 AM James Coleman <jtc331@gmail.com> wrote:\\n>> \\n>> On Thu, Aug 17, 2023 at 10:02 AM torikoshia \\n>> <torikoshia@oss.nttdata.com> wrote:\\n>> >\\n>> > On 2023-06-16 01:34, James Coleman wrote:\\n>> > > Attached is v28\\n>> > > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\\n>> > > when necessary. Once built with those two patches I'm simply running\\n>> > > `make check`.\\n>> >\\n>> > With v28-0001 and v28-0002 patch, I confirmed backend processes consume\\n>> > huge\\n>> > amount of memory and under some environments they were terminated by OOM\\n>> > killer.\\n>> >\\n>> > This was because memory was allocated from existing memory contexts and\\n>> > they\\n>> > were not freed after ProcessLogQueryPlanInterrupt().\\n>> > Updated the patch to use dedicated memory context for\\n>> > ProcessLogQueryPlanInterrupt().\\n>> >\\n>> > Applying attached patch and v28-0002 patch, `make check` successfully\\n>> > completed after 20min and 50GB of logs on my environment.\\n>> >\\n>> > >>> On 2023-06-15 01:48, James Coleman wrote:\\n>> > >>> > The tests have been running since last night, but have been apparently\\n>> > >>> > hung now for many hours.\\n>> >\\n>> > I don't know if this has anything to do with the hung you faced, but I\\n>> > thought\\n>> > it might be possible that the large amount of memory usage resulted in\\n>> > swapping, which caused a significant delay in processing.\\n>> \\n>> Ah, yes, I think that could be a possible explanation. I was delaying\\n>> on this thread because I wasn't comfortable with having caused an\\n>> issue once (even if I couldn't easily reproduce) without at least some\\n>> theory as to the cause (and a fix).\\n>> \\n>> > If possible, I would be very grateful if you could try to reproduce this\\n>> > with\\n>> > the v29 patch.\\n>> \\n>> I'll kick off some testing.\\n>> \\n> \\n> I don't have time to investigate what's happening here, but 24 hours\\n> later the first \\\"make check\\\" is still running, and at first glance it\\n> seems to have the same behavior I'd seen that first time. The test\\n> output is to this point:\\n> \\n> # parallel group (5 tests): index_including create_view\\n> index_including_gist create_index create_index_spgist\\n> ok 66 + create_index 26365 ms\\n> ok 67 + create_index_spgist 27675 ms\\n> ok 68 + create_view 1235 ms\\n> ok 69 + index_including 1102 ms\\n> ok 70 + index_including_gist 1633 ms\\n> # parallel group (16 tests): create_aggregate create_cast errors\\n> roleattributes drop_if_exists hash_func typed_table create_am\\n> infinite_recurse\\n> \\n> and it hasn't progressed past that point since at least ~16 hours ago\\n> (the first several hours of the run I wasn't monitoring it).\\n> \\n> I haven't connected up gdb yet, and won't be able to until maybe\\n> tomorrow, but here's the ps output for postgres processes that are\\n> running:\\n> \\n> admin 3213249 0.0 0.0 196824 20552 ? Ss Aug25 0:00\\n> /home/admin/postgresql-test/bin/postgres -D\\n> /home/admin/postgresql-test-data\\n> admin 3213250 0.0 0.0 196964 7284 ? Ss Aug25 0:00\\n> postgres: checkpointer\\n> admin 3213251 0.0 0.0 196956 4276 ? Ss Aug25 0:00\\n> postgres: background writer\\n> admin 3213253 0.0 0.0 196956 8600 ? Ss Aug25 0:00\\n> postgres: walwriter\\n> admin 3213254 0.0 0.0 198424 7316 ? Ss Aug25 0:00\\n> postgres: autovacuum launcher\\n> admin 3213255 0.0 0.0 198412 5488 ? Ss Aug25 0:00\\n> postgres: logical replication launcher\\n> admin 3237967 0.0 0.0 2484 516 pts/4 S+ Aug25 0:00\\n> /bin/sh -c echo \\\"# +++ regress check in src/test/regress +++\\\" &&\\n> PATH=\\\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/bin:/home/admin/postgres/src/test/regress:$PATH\\\"\\n> LD_LIBRARY_PATH=\\\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/lib\\\"\\n> INITDB_TEMPLATE='/home/admin/postgres'/tmp_install/initdb-template\\n> ../../../src/test/regress/pg_regress --temp-instance=./tmp_check\\n> --inputdir=. --bindir= --dlpath=. --max-concurrent-tests=20\\n> --schedule=./parallel_schedule\\n> admin 3237973 0.0 0.0 197880 20688 pts/4 S+ Aug25 0:00\\n> postgres -D /home/admin/postgres/src/test/regress/tmp_check/data -F -c\\n> listen_addresses= -k /tmp/pg_regress-7mmGUa\\n> admin 3237976 0.0 0.1 198332 44608 ? Ss Aug25 0:00\\n> postgres: checkpointer\\n> admin 3237977 0.0 0.0 198032 4640 ? Ss Aug25 0:00\\n> postgres: background writer\\n> admin 3237979 0.0 0.0 197880 8580 ? Ss Aug25 0:00\\n> postgres: walwriter\\n> admin 3237980 0.0 0.0 199484 7608 ? Ss Aug25 0:00\\n> postgres: autovacuum launcher\\n> admin 3237981 0.0 0.0 199460 5488 ? Ss Aug25 0:00\\n> postgres: logical replication launcher\\n> admin 3243644 0.0 0.2 252400 74656 ? Ss Aug25 0:01\\n> postgres: admin regression [local] SELECT waiting\\n> admin 3243645 0.0 0.1 205480 33992 ? Ss Aug25 0:00\\n> postgres: admin regression [local] SELECT waiting\\n> admin 3243654 99.9 0.0 203156 31504 ? Rs Aug25 1437:49\\n> postgres: admin regression [local] VACUUM\\n> admin 3243655 0.0 0.1 212036 38504 ? Ss Aug25 0:00\\n> postgres: admin regression [local] SELECT waiting\\n> admin 3243656 0.0 0.0 206024 30892 ? Ss Aug25 0:00\\n> postgres: admin regression [local] DELETE waiting\\n> admin 3243657 0.0 0.1 205568 32232 ? Ss Aug25 0:00\\n> postgres: admin regression [local] ALTER TABLE waiting\\n> admin 3243658 0.0 0.0 203740 21532 ? Ss Aug25 0:00\\n> postgres: admin regression [local] ANALYZE waiting\\n> admin 3243798 0.0 0.0 199884 8464 ? Ss Aug25 0:00\\n> postgres: autovacuum worker\\n> admin 3244733 0.0 0.0 199492 5956 ? Ss Aug25 0:00\\n> postgres: autovacuum worker\\n> admin 3245652 0.0 0.0 199884 8468 ? Ss Aug25 0:00\\n> postgres: autovacuum worker\\n> \\n> As you can see there are a bunch of backends presumably waiting, and\\n> also the VACUUM process has been pegging a single CPU core for at\\n> least since that ~16 hour ago mark.\\n> \\n> I hope to be able to do more investigation later, but I wanted to at\\n> least give you this information now.\\n\\nThanks a lot for testing the patch!\\nI really appreciate your cooperation.\\n\\nHmm, I also tested on the current HEAD(165d581f146b09) again on Ubuntu \\n22.04 and macOS, but unfortunately(fortunately?) they succeeded as \\nbelow:\\n\\n```\\n$ git apply v29-0001-Add-function-to-log-the-plan-of-the-query.patch\\n$ git apply v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch\\n$ ./configure --enable-debug --enable-cassert\\n$ make\\n$ make check\\n\\n...(snip)...\\n\\n# parallel group (5 tests): index_including index_including_gist \\ncreate_view create_index create_index_spgist\\nok 66 + create_index 25033 ms\\nok 67 + create_index_spgist 26144 ms\\nok 68 + create_view 3061 ms\\nok 69 + index_including 976 ms\\nok 70 + index_including_gist 2998 ms\\n# parallel group (16 tests): create_cast errors create_aggregate \\nroleattributes drop_if_exists hash_func typed_table\\ncreate_am select constraints updatable_views inherit triggers vacuum \\ncreate_function_sql infinite_recurse\\nok 71 + create_aggregate 225 ms\\nok 72 + create_function_sql 18874 ms\\nok 73 + create_cast 168 ms\\n\\n...(snip)...\\n\\n# All 215 tests passed.\\n```\\n\\nIf you notice any difference, I would be grateful if you could let me \\nknow.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Aug 2023 16:01:53 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 28, 2023 at 3:01 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-08-26 21:03, James Coleman wrote:\\n> > On Fri, Aug 25, 2023 at 7:43 AM James Coleman <jtc331@gmail.com> wrote:\\n> >>\\n> >> On Thu, Aug 17, 2023 at 10:02 AM torikoshia\\n> >> <torikoshia@oss.nttdata.com> wrote:\\n> >> >\\n> >> > On 2023-06-16 01:34, James Coleman wrote:\\n> >> > > Attached is v28\\n> >> > > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\\n> >> > > when necessary. Once built with those two patches I'm simply running\\n> >> > > `make check`.\\n> >> >\\n> >> > With v28-0001 and v28-0002 patch, I confirmed backend processes consume\\n> >> > huge\\n> >> > amount of memory and under some environments they were terminated by OOM\\n> >> > killer.\\n> >> >\\n> >> > This was because memory was allocated from existing memory contexts and\\n> >> > they\\n> >> > were not freed after ProcessLogQueryPlanInterrupt().\\n> >> > Updated the patch to use dedicated memory context for\\n> >> > ProcessLogQueryPlanInterrupt().\\n> >> >\\n> >> > Applying attached patch and v28-0002 patch, `make check` successfully\\n> >> > completed after 20min and 50GB of logs on my environment.\\n> >> >\\n> >> > >>> On 2023-06-15 01:48, James Coleman wrote:\\n> >> > >>> > The tests have been running since last night, but have been apparently\\n> >> > >>> > hung now for many hours.\\n> >> >\\n> >> > I don't know if this has anything to do with the hung you faced, but I\\n> >> > thought\\n> >> > it might be possible that the large amount of memory usage resulted in\\n> >> > swapping, which caused a significant delay in processing.\\n> >>\\n> >> Ah, yes, I think that could be a possible explanation. I was delaying\\n> >> on this thread because I wasn't comfortable with having caused an\\n> >> issue once (even if I couldn't easily reproduce) without at least some\\n> >> theory as to the cause (and a fix).\\n> >>\\n> >> > If possible, I would be very grateful if you could try to reproduce this\\n> >> > with\\n> >> > the v29 patch.\\n> >>\\n> >> I'll kick off some testing.\\n> >>\\n> >\\n> > I don't have time to investigate what's happening here, but 24 hours\\n> > later the first \\\"make check\\\" is still running, and at first glance it\\n> > seems to have the same behavior I'd seen that first time. The test\\n> > output is to this point:\\n> >\\n> > # parallel group (5 tests): index_including create_view\\n> > index_including_gist create_index create_index_spgist\\n> > ok 66 + create_index 26365 ms\\n> > ok 67 + create_index_spgist 27675 ms\\n> > ok 68 + create_view 1235 ms\\n> > ok 69 + index_including 1102 ms\\n> > ok 70 + index_including_gist 1633 ms\\n> > # parallel group (16 tests): create_aggregate create_cast errors\\n> > roleattributes drop_if_exists hash_func typed_table create_am\\n> > infinite_recurse\\n> >\\n> > and it hasn't progressed past that point since at least ~16 hours ago\\n> > (the first several hours of the run I wasn't monitoring it).\\n> >\\n> > I haven't connected up gdb yet, and won't be able to until maybe\\n> > tomorrow, but here's the ps output for postgres processes that are\\n> > running:\\n> >\\n> > admin 3213249 0.0 0.0 196824 20552 ? Ss Aug25 0:00\\n> > /home/admin/postgresql-test/bin/postgres -D\\n> > /home/admin/postgresql-test-data\\n> > admin 3213250 0.0 0.0 196964 7284 ? Ss Aug25 0:00\\n> > postgres: checkpointer\\n> > admin 3213251 0.0 0.0 196956 4276 ? Ss Aug25 0:00\\n> > postgres: background writer\\n> > admin 3213253 0.0 0.0 196956 8600 ? Ss Aug25 0:00\\n> > postgres: walwriter\\n> > admin 3213254 0.0 0.0 198424 7316 ? Ss Aug25 0:00\\n> > postgres: autovacuum launcher\\n> > admin 3213255 0.0 0.0 198412 5488 ? Ss Aug25 0:00\\n> > postgres: logical replication launcher\\n> > admin 3237967 0.0 0.0 2484 516 pts/4 S+ Aug25 0:00\\n> > /bin/sh -c echo \\\"# +++ regress check in src/test/regress +++\\\" &&\\n> > PATH=\\\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/bin:/home/admin/postgres/src/test/regress:$PATH\\\"\\n> > LD_LIBRARY_PATH=\\\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/lib\\\"\\n> > INITDB_TEMPLATE='/home/admin/postgres'/tmp_install/initdb-template\\n> > ../../../src/test/regress/pg_regress --temp-instance=./tmp_check\\n> > --inputdir=. --bindir= --dlpath=. --max-concurrent-tests=20\\n> > --schedule=./parallel_schedule\\n> > admin 3237973 0.0 0.0 197880 20688 pts/4 S+ Aug25 0:00\\n> > postgres -D /home/admin/postgres/src/test/regress/tmp_check/data -F -c\\n> > listen_addresses= -k /tmp/pg_regress-7mmGUa\\n> > admin 3237976 0.0 0.1 198332 44608 ? Ss Aug25 0:00\\n> > postgres: checkpointer\\n> > admin 3237977 0.0 0.0 198032 4640 ? Ss Aug25 0:00\\n> > postgres: background writer\\n> > admin 3237979 0.0 0.0 197880 8580 ? Ss Aug25 0:00\\n> > postgres: walwriter\\n> > admin 3237980 0.0 0.0 199484 7608 ? Ss Aug25 0:00\\n> > postgres: autovacuum launcher\\n> > admin 3237981 0.0 0.0 199460 5488 ? Ss Aug25 0:00\\n> > postgres: logical replication launcher\\n> > admin 3243644 0.0 0.2 252400 74656 ? Ss Aug25 0:01\\n> > postgres: admin regression [local] SELECT waiting\\n> > admin 3243645 0.0 0.1 205480 33992 ? Ss Aug25 0:00\\n> > postgres: admin regression [local] SELECT waiting\\n> > admin 3243654 99.9 0.0 203156 31504 ? Rs Aug25 1437:49\\n> > postgres: admin regression [local] VACUUM\\n> > admin 3243655 0.0 0.1 212036 38504 ? Ss Aug25 0:00\\n> > postgres: admin regression [local] SELECT waiting\\n> > admin 3243656 0.0 0.0 206024 30892 ? Ss Aug25 0:00\\n> > postgres: admin regression [local] DELETE waiting\\n> > admin 3243657 0.0 0.1 205568 32232 ? Ss Aug25 0:00\\n> > postgres: admin regression [local] ALTER TABLE waiting\\n> > admin 3243658 0.0 0.0 203740 21532 ? Ss Aug25 0:00\\n> > postgres: admin regression [local] ANALYZE waiting\\n> > admin 3243798 0.0 0.0 199884 8464 ? Ss Aug25 0:00\\n> > postgres: autovacuum worker\\n> > admin 3244733 0.0 0.0 199492 5956 ? Ss Aug25 0:00\\n> > postgres: autovacuum worker\\n> > admin 3245652 0.0 0.0 199884 8468 ? Ss Aug25 0:00\\n> > postgres: autovacuum worker\\n> >\\n> > As you can see there are a bunch of backends presumably waiting, and\\n> > also the VACUUM process has been pegging a single CPU core for at\\n> > least since that ~16 hour ago mark.\\n> >\\n> > I hope to be able to do more investigation later, but I wanted to at\\n> > least give you this information now.\\n>\\n> Thanks a lot for testing the patch!\\n> I really appreciate your cooperation.\\n>\\n> Hmm, I also tested on the current HEAD(165d581f146b09) again on Ubuntu\\n> 22.04 and macOS, but unfortunately(fortunately?) they succeeded as\\n> below:\\n>\\n> ```\\n> $ git apply v29-0001-Add-function-to-log-the-plan-of-the-query.patch\\n> $ git apply v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch\\n> $ ./configure --enable-debug --enable-cassert\\n> $ make\\n> $ make check\\n>\\n> ...(snip)...\\n>\\n> # parallel group (5 tests): index_including index_including_gist\\n> create_view create_index create_index_spgist\\n> ok 66 + create_index 25033 ms\\n> ok 67 + create_index_spgist 26144 ms\\n> ok 68 + create_view 3061 ms\\n> ok 69 + index_including 976 ms\\n> ok 70 + index_including_gist 2998 ms\\n> # parallel group (16 tests): create_cast errors create_aggregate\\n> roleattributes drop_if_exists hash_func typed_table\\n> create_am select constraints updatable_views inherit triggers vacuum\\n> create_function_sql infinite_recurse\\n> ok 71 + create_aggregate 225 ms\\n> ok 72 + create_function_sql 18874 ms\\n> ok 73 + create_cast 168 ms\\n>\\n> ...(snip)...\\n>\\n> # All 215 tests passed.\\n> ```\\n>\\n> If you notice any difference, I would be grateful if you could let me\\n> know.\\n\\nI've never been able to reproduce it (haven't tested the new version,\\nbut v28 at least) on my M1 Mac; where I've reproduced it is on Debian\\n(first buster and now bullseye).\\n\\nI'm attaching several stacktraces in the hope that they provide some\\nclues. These all match the ps output I sent earlier, though note in\\nthat output there is both the regress instance and my test instance\\n(pid 3213249) running (different ports, of course, and they are from\\nthe exact same compilation run). I've attached ps output for the\\npostgres processes under the make check process to simplify cross\\nreferencing.\\n\\nA few interesting things:\\n- There's definitely a lock on a relation that seems to be what's\\nblocking the processes.\\n- When I try to connect with psql the process forks but then hangs\\n(see the ps output with task names stuck in \\\"authentication\\\"). I've\\nalso included a trace from one of these.\\n\\nIf you think a core file for any of these processes would be helpful\\nfor debugging I'd be happy to try to figure out a way to get that to\\nyou.\\n\\nRegards,\\nJames Coleman\",\n \"msg_date\": \"Mon, 28 Aug 2023 09:47:19 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-08-28 22:47, James Coleman wrote:\\n> On Mon, Aug 28, 2023 at 3:01 AM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> \\n>> On 2023-08-26 21:03, James Coleman wrote:\\n>> > On Fri, Aug 25, 2023 at 7:43 AM James Coleman <jtc331@gmail.com> wrote:\\n>> >>\\n>> >> On Thu, Aug 17, 2023 at 10:02 AM torikoshia\\n>> >> <torikoshia@oss.nttdata.com> wrote:\\n>> >> >\\n>> >> > On 2023-06-16 01:34, James Coleman wrote:\\n>> >> > > Attached is v28\\n>> >> > > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\\n>> >> > > when necessary. Once built with those two patches I'm simply running\\n>> >> > > `make check`.\\n>> >> >\\n>> >> > With v28-0001 and v28-0002 patch, I confirmed backend processes consume\\n>> >> > huge\\n>> >> > amount of memory and under some environments they were terminated by OOM\\n>> >> > killer.\\n>> >> >\\n>> >> > This was because memory was allocated from existing memory contexts and\\n>> >> > they\\n>> >> > were not freed after ProcessLogQueryPlanInterrupt().\\n>> >> > Updated the patch to use dedicated memory context for\\n>> >> > ProcessLogQueryPlanInterrupt().\\n>> >> >\\n>> >> > Applying attached patch and v28-0002 patch, `make check` successfully\\n>> >> > completed after 20min and 50GB of logs on my environment.\\n>> >> >\\n>> >> > >>> On 2023-06-15 01:48, James Coleman wrote:\\n>> >> > >>> > The tests have been running since last night, but have been apparently\\n>> >> > >>> > hung now for many hours.\\n>> >> >\\n>> >> > I don't know if this has anything to do with the hung you faced, but I\\n>> >> > thought\\n>> >> > it might be possible that the large amount of memory usage resulted in\\n>> >> > swapping, which caused a significant delay in processing.\\n>> >>\\n>> >> Ah, yes, I think that could be a possible explanation. I was delaying\\n>> >> on this thread because I wasn't comfortable with having caused an\\n>> >> issue once (even if I couldn't easily reproduce) without at least some\\n>> >> theory as to the cause (and a fix).\\n>> >>\\n>> >> > If possible, I would be very grateful if you could try to reproduce this\\n>> >> > with\\n>> >> > the v29 patch.\\n>> >>\\n>> >> I'll kick off some testing.\\n>> >>\\n>> >\\n>> > I don't have time to investigate what's happening here, but 24 hours\\n>> > later the first \\\"make check\\\" is still running, and at first glance it\\n>> > seems to have the same behavior I'd seen that first time. The test\\n>> > output is to this point:\\n>> >\\n>> > # parallel group (5 tests): index_including create_view\\n>> > index_including_gist create_index create_index_spgist\\n>> > ok 66 + create_index 26365 ms\\n>> > ok 67 + create_index_spgist 27675 ms\\n>> > ok 68 + create_view 1235 ms\\n>> > ok 69 + index_including 1102 ms\\n>> > ok 70 + index_including_gist 1633 ms\\n>> > # parallel group (16 tests): create_aggregate create_cast errors\\n>> > roleattributes drop_if_exists hash_func typed_table create_am\\n>> > infinite_recurse\\n>> >\\n>> > and it hasn't progressed past that point since at least ~16 hours ago\\n>> > (the first several hours of the run I wasn't monitoring it).\\n>> >\\n>> > I haven't connected up gdb yet, and won't be able to until maybe\\n>> > tomorrow, but here's the ps output for postgres processes that are\\n>> > running:\\n>> >\\n>> > admin 3213249 0.0 0.0 196824 20552 ? Ss Aug25 0:00\\n>> > /home/admin/postgresql-test/bin/postgres -D\\n>> > /home/admin/postgresql-test-data\\n>> > admin 3213250 0.0 0.0 196964 7284 ? Ss Aug25 0:00\\n>> > postgres: checkpointer\\n>> > admin 3213251 0.0 0.0 196956 4276 ? Ss Aug25 0:00\\n>> > postgres: background writer\\n>> > admin 3213253 0.0 0.0 196956 8600 ? Ss Aug25 0:00\\n>> > postgres: walwriter\\n>> > admin 3213254 0.0 0.0 198424 7316 ? Ss Aug25 0:00\\n>> > postgres: autovacuum launcher\\n>> > admin 3213255 0.0 0.0 198412 5488 ? Ss Aug25 0:00\\n>> > postgres: logical replication launcher\\n>> > admin 3237967 0.0 0.0 2484 516 pts/4 S+ Aug25 0:00\\n>> > /bin/sh -c echo \\\"# +++ regress check in src/test/regress +++\\\" &&\\n>> > PATH=\\\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/bin:/home/admin/postgres/src/test/regress:$PATH\\\"\\n>> > LD_LIBRARY_PATH=\\\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/lib\\\"\\n>> > INITDB_TEMPLATE='/home/admin/postgres'/tmp_install/initdb-template\\n>> > ../../../src/test/regress/pg_regress --temp-instance=./tmp_check\\n>> > --inputdir=. --bindir= --dlpath=. --max-concurrent-tests=20\\n>> > --schedule=./parallel_schedule\\n>> > admin 3237973 0.0 0.0 197880 20688 pts/4 S+ Aug25 0:00\\n>> > postgres -D /home/admin/postgres/src/test/regress/tmp_check/data -F -c\\n>> > listen_addresses= -k /tmp/pg_regress-7mmGUa\\n>> > admin 3237976 0.0 0.1 198332 44608 ? Ss Aug25 0:00\\n>> > postgres: checkpointer\\n>> > admin 3237977 0.0 0.0 198032 4640 ? Ss Aug25 0:00\\n>> > postgres: background writer\\n>> > admin 3237979 0.0 0.0 197880 8580 ? Ss Aug25 0:00\\n>> > postgres: walwriter\\n>> > admin 3237980 0.0 0.0 199484 7608 ? Ss Aug25 0:00\\n>> > postgres: autovacuum launcher\\n>> > admin 3237981 0.0 0.0 199460 5488 ? Ss Aug25 0:00\\n>> > postgres: logical replication launcher\\n>> > admin 3243644 0.0 0.2 252400 74656 ? Ss Aug25 0:01\\n>> > postgres: admin regression [local] SELECT waiting\\n>> > admin 3243645 0.0 0.1 205480 33992 ? Ss Aug25 0:00\\n>> > postgres: admin regression [local] SELECT waiting\\n>> > admin 3243654 99.9 0.0 203156 31504 ? Rs Aug25 1437:49\\n>> > postgres: admin regression [local] VACUUM\\n>> > admin 3243655 0.0 0.1 212036 38504 ? Ss Aug25 0:00\\n>> > postgres: admin regression [local] SELECT waiting\\n>> > admin 3243656 0.0 0.0 206024 30892 ? Ss Aug25 0:00\\n>> > postgres: admin regression [local] DELETE waiting\\n>> > admin 3243657 0.0 0.1 205568 32232 ? Ss Aug25 0:00\\n>> > postgres: admin regression [local] ALTER TABLE waiting\\n>> > admin 3243658 0.0 0.0 203740 21532 ? Ss Aug25 0:00\\n>> > postgres: admin regression [local] ANALYZE waiting\\n>> > admin 3243798 0.0 0.0 199884 8464 ? Ss Aug25 0:00\\n>> > postgres: autovacuum worker\\n>> > admin 3244733 0.0 0.0 199492 5956 ? Ss Aug25 0:00\\n>> > postgres: autovacuum worker\\n>> > admin 3245652 0.0 0.0 199884 8468 ? Ss Aug25 0:00\\n>> > postgres: autovacuum worker\\n>> >\\n>> > As you can see there are a bunch of backends presumably waiting, and\\n>> > also the VACUUM process has been pegging a single CPU core for at\\n>> > least since that ~16 hour ago mark.\\n>> >\\n>> > I hope to be able to do more investigation later, but I wanted to at\\n>> > least give you this information now.\\n>> \\n>> Thanks a lot for testing the patch!\\n>> I really appreciate your cooperation.\\n>> \\n>> Hmm, I also tested on the current HEAD(165d581f146b09) again on Ubuntu\\n>> 22.04 and macOS, but unfortunately(fortunately?) they succeeded as\\n>> below:\\n>> \\n>> ```\\n>> $ git apply v29-0001-Add-function-to-log-the-plan-of-the-query.patch\\n>> $ git apply \\n>> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch\\n>> $ ./configure --enable-debug --enable-cassert\\n>> $ make\\n>> $ make check\\n>> \\n>> ...(snip)...\\n>> \\n>> # parallel group (5 tests): index_including index_including_gist\\n>> create_view create_index create_index_spgist\\n>> ok 66 + create_index 25033 ms\\n>> ok 67 + create_index_spgist 26144 ms\\n>> ok 68 + create_view 3061 ms\\n>> ok 69 + index_including 976 ms\\n>> ok 70 + index_including_gist 2998 ms\\n>> # parallel group (16 tests): create_cast errors create_aggregate\\n>> roleattributes drop_if_exists hash_func typed_table\\n>> create_am select constraints updatable_views inherit triggers vacuum\\n>> create_function_sql infinite_recurse\\n>> ok 71 + create_aggregate 225 ms\\n>> ok 72 + create_function_sql 18874 ms\\n>> ok 73 + create_cast 168 ms\\n>> \\n>> ...(snip)...\\n>> \\n>> # All 215 tests passed.\\n>> ```\\n>> \\n>> If you notice any difference, I would be grateful if you could let me\\n>> know.\\n> \\n> I've never been able to reproduce it (haven't tested the new version,\\n> but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\\n> (first buster and now bullseye).\\n> \\n> I'm attaching several stacktraces in the hope that they provide some\\n> clues. These all match the ps output I sent earlier, though note in\\n> that output there is both the regress instance and my test instance\\n> (pid 3213249) running (different ports, of course, and they are from\\n> the exact same compilation run). I've attached ps output for the\\n> postgres processes under the make check process to simplify cross\\n> referencing.\\n> \\n> A few interesting things:\\n> - There's definitely a lock on a relation that seems to be what's\\n> blocking the processes.\\n> - When I try to connect with psql the process forks but then hangs\\n> (see the ps output with task names stuck in \\\"authentication\\\"). I've\\n> also included a trace from one of these.\\n\\nThanks for sharing them!\\n\\nMany processes are waiting to acquire the LW lock, including the process \\ntrying to output the plan(select1.trace).\\n\\nI suspect that this is due to a lock that was acquired prior to being \\ninterrupted by ProcessLogQueryPlanInterrupt(), but have not been able to \\nreproduce the same situation..\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Tue, 05 Sep 2023 22:58:59 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Tue, Sep 5, 2023 at 9:59 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-08-28 22:47, James Coleman wrote:\\n> > On Mon, Aug 28, 2023 at 3:01 AM torikoshia <torikoshia@oss.nttdata.com>\\n> > wrote:\\n> >>\\n> >> On 2023-08-26 21:03, James Coleman wrote:\\n> >> > On Fri, Aug 25, 2023 at 7:43 AM James Coleman <jtc331@gmail.com> wrote:\\n> >> >>\\n> >> >> On Thu, Aug 17, 2023 at 10:02 AM torikoshia\\n> >> >> <torikoshia@oss.nttdata.com> wrote:\\n> >> >> >\\n> >> >> > On 2023-06-16 01:34, James Coleman wrote:\\n> >> >> > > Attached is v28\\n> >> >> > > which sets ProcessLogQueryPlanInterruptActive to false in errfinish\\n> >> >> > > when necessary. Once built with those two patches I'm simply running\\n> >> >> > > `make check`.\\n> >> >> >\\n> >> >> > With v28-0001 and v28-0002 patch, I confirmed backend processes consume\\n> >> >> > huge\\n> >> >> > amount of memory and under some environments they were terminated by OOM\\n> >> >> > killer.\\n> >> >> >\\n> >> >> > This was because memory was allocated from existing memory contexts and\\n> >> >> > they\\n> >> >> > were not freed after ProcessLogQueryPlanInterrupt().\\n> >> >> > Updated the patch to use dedicated memory context for\\n> >> >> > ProcessLogQueryPlanInterrupt().\\n> >> >> >\\n> >> >> > Applying attached patch and v28-0002 patch, `make check` successfully\\n> >> >> > completed after 20min and 50GB of logs on my environment.\\n> >> >> >\\n> >> >> > >>> On 2023-06-15 01:48, James Coleman wrote:\\n> >> >> > >>> > The tests have been running since last night, but have been apparently\\n> >> >> > >>> > hung now for many hours.\\n> >> >> >\\n> >> >> > I don't know if this has anything to do with the hung you faced, but I\\n> >> >> > thought\\n> >> >> > it might be possible that the large amount of memory usage resulted in\\n> >> >> > swapping, which caused a significant delay in processing.\\n> >> >>\\n> >> >> Ah, yes, I think that could be a possible explanation. I was delaying\\n> >> >> on this thread because I wasn't comfortable with having caused an\\n> >> >> issue once (even if I couldn't easily reproduce) without at least some\\n> >> >> theory as to the cause (and a fix).\\n> >> >>\\n> >> >> > If possible, I would be very grateful if you could try to reproduce this\\n> >> >> > with\\n> >> >> > the v29 patch.\\n> >> >>\\n> >> >> I'll kick off some testing.\\n> >> >>\\n> >> >\\n> >> > I don't have time to investigate what's happening here, but 24 hours\\n> >> > later the first \\\"make check\\\" is still running, and at first glance it\\n> >> > seems to have the same behavior I'd seen that first time. The test\\n> >> > output is to this point:\\n> >> >\\n> >> > # parallel group (5 tests): index_including create_view\\n> >> > index_including_gist create_index create_index_spgist\\n> >> > ok 66 + create_index 26365 ms\\n> >> > ok 67 + create_index_spgist 27675 ms\\n> >> > ok 68 + create_view 1235 ms\\n> >> > ok 69 + index_including 1102 ms\\n> >> > ok 70 + index_including_gist 1633 ms\\n> >> > # parallel group (16 tests): create_aggregate create_cast errors\\n> >> > roleattributes drop_if_exists hash_func typed_table create_am\\n> >> > infinite_recurse\\n> >> >\\n> >> > and it hasn't progressed past that point since at least ~16 hours ago\\n> >> > (the first several hours of the run I wasn't monitoring it).\\n> >> >\\n> >> > I haven't connected up gdb yet, and won't be able to until maybe\\n> >> > tomorrow, but here's the ps output for postgres processes that are\\n> >> > running:\\n> >> >\\n> >> > admin 3213249 0.0 0.0 196824 20552 ? Ss Aug25 0:00\\n> >> > /home/admin/postgresql-test/bin/postgres -D\\n> >> > /home/admin/postgresql-test-data\\n> >> > admin 3213250 0.0 0.0 196964 7284 ? Ss Aug25 0:00\\n> >> > postgres: checkpointer\\n> >> > admin 3213251 0.0 0.0 196956 4276 ? Ss Aug25 0:00\\n> >> > postgres: background writer\\n> >> > admin 3213253 0.0 0.0 196956 8600 ? Ss Aug25 0:00\\n> >> > postgres: walwriter\\n> >> > admin 3213254 0.0 0.0 198424 7316 ? Ss Aug25 0:00\\n> >> > postgres: autovacuum launcher\\n> >> > admin 3213255 0.0 0.0 198412 5488 ? Ss Aug25 0:00\\n> >> > postgres: logical replication launcher\\n> >> > admin 3237967 0.0 0.0 2484 516 pts/4 S+ Aug25 0:00\\n> >> > /bin/sh -c echo \\\"# +++ regress check in src/test/regress +++\\\" &&\\n> >> > PATH=\\\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/bin:/home/admin/postgres/src/test/regress:$PATH\\\"\\n> >> > LD_LIBRARY_PATH=\\\"/home/admin/postgres/tmp_install/home/admin/postgresql-test/lib\\\"\\n> >> > INITDB_TEMPLATE='/home/admin/postgres'/tmp_install/initdb-template\\n> >> > ../../../src/test/regress/pg_regress --temp-instance=./tmp_check\\n> >> > --inputdir=. --bindir= --dlpath=. --max-concurrent-tests=20\\n> >> > --schedule=./parallel_schedule\\n> >> > admin 3237973 0.0 0.0 197880 20688 pts/4 S+ Aug25 0:00\\n> >> > postgres -D /home/admin/postgres/src/test/regress/tmp_check/data -F -c\\n> >> > listen_addresses= -k /tmp/pg_regress-7mmGUa\\n> >> > admin 3237976 0.0 0.1 198332 44608 ? Ss Aug25 0:00\\n> >> > postgres: checkpointer\\n> >> > admin 3237977 0.0 0.0 198032 4640 ? Ss Aug25 0:00\\n> >> > postgres: background writer\\n> >> > admin 3237979 0.0 0.0 197880 8580 ? Ss Aug25 0:00\\n> >> > postgres: walwriter\\n> >> > admin 3237980 0.0 0.0 199484 7608 ? Ss Aug25 0:00\\n> >> > postgres: autovacuum launcher\\n> >> > admin 3237981 0.0 0.0 199460 5488 ? Ss Aug25 0:00\\n> >> > postgres: logical replication launcher\\n> >> > admin 3243644 0.0 0.2 252400 74656 ? Ss Aug25 0:01\\n> >> > postgres: admin regression [local] SELECT waiting\\n> >> > admin 3243645 0.0 0.1 205480 33992 ? Ss Aug25 0:00\\n> >> > postgres: admin regression [local] SELECT waiting\\n> >> > admin 3243654 99.9 0.0 203156 31504 ? Rs Aug25 1437:49\\n> >> > postgres: admin regression [local] VACUUM\\n> >> > admin 3243655 0.0 0.1 212036 38504 ? Ss Aug25 0:00\\n> >> > postgres: admin regression [local] SELECT waiting\\n> >> > admin 3243656 0.0 0.0 206024 30892 ? Ss Aug25 0:00\\n> >> > postgres: admin regression [local] DELETE waiting\\n> >> > admin 3243657 0.0 0.1 205568 32232 ? Ss Aug25 0:00\\n> >> > postgres: admin regression [local] ALTER TABLE waiting\\n> >> > admin 3243658 0.0 0.0 203740 21532 ? Ss Aug25 0:00\\n> >> > postgres: admin regression [local] ANALYZE waiting\\n> >> > admin 3243798 0.0 0.0 199884 8464 ? Ss Aug25 0:00\\n> >> > postgres: autovacuum worker\\n> >> > admin 3244733 0.0 0.0 199492 5956 ? Ss Aug25 0:00\\n> >> > postgres: autovacuum worker\\n> >> > admin 3245652 0.0 0.0 199884 8468 ? Ss Aug25 0:00\\n> >> > postgres: autovacuum worker\\n> >> >\\n> >> > As you can see there are a bunch of backends presumably waiting, and\\n> >> > also the VACUUM process has been pegging a single CPU core for at\\n> >> > least since that ~16 hour ago mark.\\n> >> >\\n> >> > I hope to be able to do more investigation later, but I wanted to at\\n> >> > least give you this information now.\\n> >>\\n> >> Thanks a lot for testing the patch!\\n> >> I really appreciate your cooperation.\\n> >>\\n> >> Hmm, I also tested on the current HEAD(165d581f146b09) again on Ubuntu\\n> >> 22.04 and macOS, but unfortunately(fortunately?) they succeeded as\\n> >> below:\\n> >>\\n> >> ```\\n> >> $ git apply v29-0001-Add-function-to-log-the-plan-of-the-query.patch\\n> >> $ git apply\\n> >> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch\\n> >> $ ./configure --enable-debug --enable-cassert\\n> >> $ make\\n> >> $ make check\\n> >>\\n> >> ...(snip)...\\n> >>\\n> >> # parallel group (5 tests): index_including index_including_gist\\n> >> create_view create_index create_index_spgist\\n> >> ok 66 + create_index 25033 ms\\n> >> ok 67 + create_index_spgist 26144 ms\\n> >> ok 68 + create_view 3061 ms\\n> >> ok 69 + index_including 976 ms\\n> >> ok 70 + index_including_gist 2998 ms\\n> >> # parallel group (16 tests): create_cast errors create_aggregate\\n> >> roleattributes drop_if_exists hash_func typed_table\\n> >> create_am select constraints updatable_views inherit triggers vacuum\\n> >> create_function_sql infinite_recurse\\n> >> ok 71 + create_aggregate 225 ms\\n> >> ok 72 + create_function_sql 18874 ms\\n> >> ok 73 + create_cast 168 ms\\n> >>\\n> >> ...(snip)...\\n> >>\\n> >> # All 215 tests passed.\\n> >> ```\\n> >>\\n> >> If you notice any difference, I would be grateful if you could let me\\n> >> know.\\n> >\\n> > I've never been able to reproduce it (haven't tested the new version,\\n> > but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\\n> > (first buster and now bullseye).\\n> >\\n> > I'm attaching several stacktraces in the hope that they provide some\\n> > clues. These all match the ps output I sent earlier, though note in\\n> > that output there is both the regress instance and my test instance\\n> > (pid 3213249) running (different ports, of course, and they are from\\n> > the exact same compilation run). I've attached ps output for the\\n> > postgres processes under the make check process to simplify cross\\n> > referencing.\\n> >\\n> > A few interesting things:\\n> > - There's definitely a lock on a relation that seems to be what's\\n> > blocking the processes.\\n> > - When I try to connect with psql the process forks but then hangs\\n> > (see the ps output with task names stuck in \\\"authentication\\\"). I've\\n> > also included a trace from one of these.\\n>\\n> Thanks for sharing them!\\n>\\n> Many processes are waiting to acquire the LW lock, including the process\\n> trying to output the plan(select1.trace).\\n>\\n> I suspect that this is due to a lock that was acquired prior to being\\n> interrupted by ProcessLogQueryPlanInterrupt(), but have not been able to\\n> reproduce the same situation..\\n>\\n\\nI don't have time immediately to dig into this, but perhaps loading up\\nthe core dumps would allow us to see what query is running in each\\nbackend process (if it hasn't already been discarded by that point)\\nand thereby determine what point in each test process led to the error\\ncondition?\\n\\nAlternatively we might be able to apply the same trick to the test\\nclient instead...\\n\\nBTW, for my own easy reference in this thread: relid 1259 is pg_class\\nif I'm not mistaken.\\n\\nThoughts?\\nJames Coleman\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Sep 2023 22:17:08 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-09-06 11:17, James Coleman wrote:\\n\\n>> > I've never been able to reproduce it (haven't tested the new version,\\n>> > but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\\n>> > (first buster and now bullseye).\\n>> >\\n>> > I'm attaching several stacktraces in the hope that they provide some\\n>> > clues. These all match the ps output I sent earlier, though note in\\n>> > that output there is both the regress instance and my test instance\\n>> > (pid 3213249) running (different ports, of course, and they are from\\n>> > the exact same compilation run). I've attached ps output for the\\n>> > postgres processes under the make check process to simplify cross\\n>> > referencing.\\n>> >\\n>> > A few interesting things:\\n>> > - There's definitely a lock on a relation that seems to be what's\\n>> > blocking the processes.\\n>> > - When I try to connect with psql the process forks but then hangs\\n>> > (see the ps output with task names stuck in \\\"authentication\\\"). I've\\n>> > also included a trace from one of these.\\n>> \\n>> Thanks for sharing them!\\n>> \\n>> Many processes are waiting to acquire the LW lock, including the \\n>> process\\n>> trying to output the plan(select1.trace).\\n>> \\n>> I suspect that this is due to a lock that was acquired prior to being\\n>> interrupted by ProcessLogQueryPlanInterrupt(), but have not been able \\n>> to\\n>> reproduce the same situation..\\n>> \\n> \\n> I don't have time immediately to dig into this, but perhaps loading up\\n> the core dumps would allow us to see what query is running in each\\n> backend process (if it hasn't already been discarded by that point)\\n> and thereby determine what point in each test process led to the error\\n> condition?\\n\\nThanks for the suggestion.\\nI am concerned that core dumps may not be readable on different \\noperating systems or other environments. (Unfortunately, I do not have \\nDebian on hand)\\n\\nIt seems that we can know what queries were running from the stack \\ntraces you shared.\\nAs described above, I suspect a lock which was acquired prior to \\nProcessLogQueryPlanInterrupt() caused the issue.\\nI will try a little more to see if I can devise a way to create the same \\nsituation.\\n\\n> Alternatively we might be able to apply the same trick to the test\\n> client instead...\\n> \\n> BTW, for my own easy reference in this thread: relid 1259 is pg_class\\n> if I'm not mistaken.\\n\\nYeah, and I think it's strange that the lock to 1259 appears twice and \\nmust be avoided.\\n\\n #10 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) at \\nlmgr.c:117\\n ..\\n #49 0x0000559d61b4989d in ProcessLogQueryPlanInterrupt () at \\nexplain.c:5158\\n ..\\n #53 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) at \\nlmgr.c:117\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Sep 2023 15:09:12 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Thu, Sep 7, 2023, at 1:09 PM, torikoshia wrote:\\n> On 2023-09-06 11:17, James Coleman wrote:\\n> It seems that we can know what queries were running from the stack \\n> traces you shared.\\n> As described above, I suspect a lock which was acquired prior to \\n> ProcessLogQueryPlanInterrupt() caused the issue.\\n> I will try a little more to see if I can devise a way to create the same \\n> situation.\\nHi,\\nI have explored this patch and, by and large, agree with the code. But some questions I want to discuss:\\n1. Maybe add a hook to give a chance for extensions, like pg_query_state, to do their job?\\n2. In this implementation, ProcessInterrupts does a lot of work during the explain creation: memory allocations, pass through the plan, systables locks, syscache access, etc. I guess it can change the semantic meaning of 'safe place' where CHECK_FOR_INTERRUPTS can be called - I can imagine a SELECT query, which would call a stored procedure, which executes some DDL and acquires row exclusive lock at the time of interruption. So, in my mind, it is too unpredictable to make the explain in the place of interruption processing. Maybe it is better to think about some hook at the end of ExecProcNode, where a pending explain could be created?\\n\\nAlso, I suggest minor code change with the diff in attachment. \\n\\n-- \\nRegards,\\nAndrei Lepikhov\",\n \"msg_date\": \"Fri, 15 Sep 2023 13:21:27 +0700\",\n \"msg_from\": \"\\\"Lepikhov Andrei\\\" <a.lepikhov@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-09-15 15:21, Lepikhov Andrei wrote:\\n> On Thu, Sep 7, 2023, at 1:09 PM, torikoshia wrote:\\n>> On 2023-09-06 11:17, James Coleman wrote:\\n>> It seems that we can know what queries were running from the stack\\n>> traces you shared.\\n>> As described above, I suspect a lock which was acquired prior to\\n>> ProcessLogQueryPlanInterrupt() caused the issue.\\n>> I will try a little more to see if I can devise a way to create the \\n>> same\\n>> situation.\\n> Hi,\\n> I have explored this patch and, by and large, agree with the code. But\\n> some questions I want to discuss:\\n> 1. Maybe add a hook to give a chance for extensions, like\\n> pg_query_state, to do their job?\\n\\nDo you imagine adding a hook which enables adding custom interrupt codes \\nlike below?\\n\\nhttps://github.com/postgrespro/pg_query_state/blob/master/patches/custom_signals_15.0.patch\\n\\nIf so, that would be possible, but this patch doesn't require the \\nfunctionality and I feel it'd be better doing in independent patch.\\n\\n\\n> 2. In this implementation, ProcessInterrupts does a lot of work during\\n> the explain creation: memory allocations, pass through the plan,\\n> systables locks, syscache access, etc. I guess it can change the\\n> semantic meaning of 'safe place' where CHECK_FOR_INTERRUPTS can be\\n> called - I can imagine a SELECT query, which would call a stored\\n> procedure, which executes some DDL and acquires row exclusive lock at\\n> the time of interruption. So, in my mind, it is too unpredictable to\\n> make the explain in the place of interruption processing. Maybe it is\\n> better to think about some hook at the end of ExecProcNode, where a\\n> pending explain could be created?\\n\\nYeah, I withdrew this patch once for that reason, but we are resuming \\ndevelopment in response to the results of a discussion by James and \\nothers at this year's pgcon about the safety of this process in CFI:\\n\\nhttps://www.postgresql.org/message-id/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw%40mail.gmail.com\\n\\nBTW it seems pg_query_state also enables users to get running query \\nplans using CFI.\\nDoes pg_query_state do something for this safety concern?\\n\\n> Also, I suggest minor code change with the diff in attachment.\\n\\nThanks!\\n\\nThis might be biased opinion and objections are welcomed, but I feel the \\noriginal patch is easier to read since PG_RETURN_BOOL(true/false) is \\nlocated in near place to each cases.\\nAlso the existing function pg_log_backend_memory_contexts(), which does \\nthe same thing, has the same form as the original patch.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Tue, 19 Sep 2023 22:39:59 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Tue, Sep 19, 2023, at 8:39 PM, torikoshia wrote:\\n> On 2023-09-15 15:21, Lepikhov Andrei wrote:\\n>> On Thu, Sep 7, 2023, at 1:09 PM, torikoshia wrote:\\n>> I have explored this patch and, by and large, agree with the code. But\\n>> some questions I want to discuss:\\n>> 1. Maybe add a hook to give a chance for extensions, like\\n>> pg_query_state, to do their job?\\n>\\n> Do you imagine adding a hook which enables adding custom interrupt codes \\n> like below?\\n>\\n> https://github.com/postgrespro/pg_query_state/blob/master/patches/custom_signals_15.0.patch\\n\\nNo, I think around the hook, which would allow us to rewrite the pg_query_state extension without additional patches by using the functionality provided by your patch. I mean, an extension could provide console UI, not only server logging. And obtain from target backend so much information in the explain as the instrumentation level of the current query can give.\\nIt may make pg_query_state shorter and more stable.\\n\\n>> 2. In this implementation, ProcessInterrupts does a lot of work during\\n>> the explain creation: memory allocations, pass through the plan,\\n>> systables locks, syscache access, etc. I guess it can change the\\n>> semantic meaning of 'safe place' where CHECK_FOR_INTERRUPTS can be\\n>> called - I can imagine a SELECT query, which would call a stored\\n>> procedure, which executes some DDL and acquires row exclusive lock at\\n>> the time of interruption. So, in my mind, it is too unpredictable to\\n>> make the explain in the place of interruption processing. Maybe it is\\n>> better to think about some hook at the end of ExecProcNode, where a\\n>> pending explain could be created?\\n>\\n> Yeah, I withdrew this patch once for that reason, but we are resuming \\n> development in response to the results of a discussion by James and \\n> others at this year's pgcon about the safety of this process in CFI:\\n>\\n> https://www.postgresql.org/message-id/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw%40mail.gmail.com\\n\\nI can't track the logic path of the decision provided at this conference. But my anxiety related to specific place, where ActiveQueryDesc points top-level query and interruption comes during DDL, wrapped up in stored procedure. For example:\\nCREATE TABLE test();\\nCREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\\nBEGIN\\n EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\\n ...\\nEND; $$ LANGUAGE plpgsql VOLATILE;\\nSELECT ddl(), ... FROM ...;\\n\\n> BTW it seems pg_query_state also enables users to get running query \\n> plans using CFI.\\n> Does pg_query_state do something for this safety concern?\\n\\nNo, and I'm looking for the solution, which could help to rewrite pg_query_state as a clean extension, without patches.\\n\\n>> Also, I suggest minor code change with the diff in attachment.\\n>\\n> Thanks!\\n>\\n> This might be biased opinion and objections are welcomed, but I feel the \\n> original patch is easier to read since PG_RETURN_BOOL(true/false) is \\n> located in near place to each cases.\\n> Also the existing function pg_log_backend_memory_contexts(), which does \\n> the same thing, has the same form as the original patch.\\nI got it, thank you.\\n\\n-- \\nRegards,\\nAndrei Lepikhov\\n\\n\\n\",\n \"msg_date\": \"Wed, 20 Sep 2023 12:39:13 +0700\",\n \"msg_from\": \"\\\"Lepikhov Andrei\\\" <a.lepikhov@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-09-20 14:39, Lepikhov Andrei wrote:\\nThanks for your reply.\\n\\n> On Tue, Sep 19, 2023, at 8:39 PM, torikoshia wrote:\\n>> On 2023-09-15 15:21, Lepikhov Andrei wrote:\\n>>> On Thu, Sep 7, 2023, at 1:09 PM, torikoshia wrote:\\n>>> I have explored this patch and, by and large, agree with the code. \\n>>> But\\n>>> some questions I want to discuss:\\n>>> 1. Maybe add a hook to give a chance for extensions, like\\n>>> pg_query_state, to do their job?\\n>> \\n>> Do you imagine adding a hook which enables adding custom interrupt \\n>> codes\\n>> like below?\\n>> \\n>> https://github.com/postgrespro/pg_query_state/blob/master/patches/custom_signals_15.0.patch\\n> \\n> No, I think around the hook, which would allow us to rewrite the\\n> pg_query_state extension without additional patches by using the\\n> functionality provided by your patch. I mean, an extension could\\n> provide console UI, not only server logging. And obtain from target\\n> backend so much information in the explain as the instrumentation\\n> level of the current query can give.\\n> It may make pg_query_state shorter and more stable.\\n> \\n>>> 2. In this implementation, ProcessInterrupts does a lot of work \\n>>> during\\n>>> the explain creation: memory allocations, pass through the plan,\\n>>> systables locks, syscache access, etc. I guess it can change the\\n>>> semantic meaning of 'safe place' where CHECK_FOR_INTERRUPTS can be\\n>>> called - I can imagine a SELECT query, which would call a stored\\n>>> procedure, which executes some DDL and acquires row exclusive lock at\\n>>> the time of interruption. So, in my mind, it is too unpredictable to\\n>>> make the explain in the place of interruption processing. Maybe it is\\n>>> better to think about some hook at the end of ExecProcNode, where a\\n>>> pending explain could be created?\\n>> \\n>> Yeah, I withdrew this patch once for that reason, but we are resuming\\n>> development in response to the results of a discussion by James and\\n>> others at this year's pgcon about the safety of this process in CFI:\\n>> \\n>> https://www.postgresql.org/message-id/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw%40mail.gmail.com\\n> \\n> I can't track the logic path of the decision provided at this\\n> conference. But my anxiety related to specific place, where\\n> ActiveQueryDesc points top-level query and interruption comes during\\n> DDL, wrapped up in stored procedure. For example:\\n> CREATE TABLE test();\\n> CREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\\n> BEGIN\\n> EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\\n> ...\\n> END; $$ LANGUAGE plpgsql VOLATILE;\\n> SELECT ddl(), ... FROM ...;\\n\\nHmm, as a test, I made sure to call ProcessLogQueryPlanInterrupt() on \\nall CFI using \\nv28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch, and then \\nran the following query but did not cause any problems.\\n\\n```\\n=# CREATE TABLE test();\\n=# CREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\\nBEGIN\\n EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\\n PERFORM pg_sleep(5);\\nEND; $$ LANGUAGE plpgsql VOLATILE;\\n=# SELECT ddl();\\n```\\n\\nIs this the case you're worrying about?\\n\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Sep 2023 16:21:31 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 25/9/2023 14:21, torikoshia wrote:\\n> On 2023-09-20 14:39, Lepikhov Andrei wrote:\\n> Hmm, as a test, I made sure to call ProcessLogQueryPlanInterrupt() on \\n> all CFI using \\n> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch, and then \\n> ran the following query but did not cause any problems.\\n> \\n> ```\\n> =# CREATE TABLE test();\\n> =# CREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\\n> BEGIN\\n> EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\\n> PERFORM pg_sleep(5);\\n> END; $$ LANGUAGE plpgsql VOLATILE;\\n> =# SELECT ddl();\\n> ```\\n> \\n> Is this the case you're worrying about?\\n\\nI didn't find a problem either. I just feel uncomfortable if, at the \\nmoment of interruption, we have a descriptor of another query than the \\nquery have been executing and holding resources.\\n\\n-- \\nregards,\\nAndrey Lepikhov\\nPostgres Professional\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Sep 2023 16:49:18 +0700\",\n \"msg_from\": \"Andrey Lepikhov <a.lepikhov@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-09-25 18:49, Andrey Lepikhov wrote:\\n> On 25/9/2023 14:21, torikoshia wrote:\\n>> On 2023-09-20 14:39, Lepikhov Andrei wrote:\\n>> Hmm, as a test, I made sure to call ProcessLogQueryPlanInterrupt() on \\n>> all CFI using \\n>> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch, and then \\n>> ran the following query but did not cause any problems.\\n>> \\n>> ```\\n>> =# CREATE TABLE test();\\n>> =# CREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\\n>> BEGIN\\n>> EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\\n>> PERFORM pg_sleep(5);\\n>> END; $$ LANGUAGE plpgsql VOLATILE;\\n>> =# SELECT ddl();\\n>> ```\\n>> \\n>> Is this the case you're worrying about?\\n> \\n> I didn't find a problem either. I just feel uncomfortable if, at the\\n> moment of interruption, we have a descriptor of another query than the\\n> query have been executing and holding resources.\\n\\nI think that \\\"descriptor\\\" here refers to ActiveQueryDesc, in which case\\nit is updated at the beginning of ExecutorRun(), so I am wondering if\\nthe situation you're worried about would not occur.\\n\\n---------\\n@@ -303,10 +306,21 @@ ExecutorRun(QueryDesc *queryDesc,\\n ScanDirection direction, uint64 count,\\n bool execute_once)\\n {\\n+ /*\\n+ * Update ActiveQueryDesc here to enable retrieval of the currently\\n+ * running queryDesc for nested queries.\\n+ */\\n+ QueryDesc *save_ActiveQueryDesc;\\n+\\n+ save_ActiveQueryDesc = ActiveQueryDesc;\\n+ ActiveQueryDesc = queryDesc;\\n---------\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Thu, 28 Sep 2023 11:04:09 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 28/9/2023 09:04, torikoshia wrote:\\n> On 2023-09-25 18:49, Andrey Lepikhov wrote:\\n>> On 25/9/2023 14:21, torikoshia wrote:\\n>>> On 2023-09-20 14:39, Lepikhov Andrei wrote:\\n>>> Hmm, as a test, I made sure to call ProcessLogQueryPlanInterrupt() on \\n>>> all CFI using \\n>>> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch, and \\n>>> then ran the following query but did not cause any problems.\\n>>>\\n>>> ```\\n>>> =# CREATE TABLE test();\\n>>> =# CREATE OR REPLACE FUNCTION ddl() RETURNS void AS $$\\n>>> BEGIN\\n>>> EXECUTE format('ALTER TABLE test ADD COLUMN x integer;');\\n>>> PERFORM pg_sleep(5);\\n>>> END; $$ LANGUAGE plpgsql VOLATILE;\\n>>> =# SELECT ddl();\\n>>> ```\\n>>>\\n>>> Is this the case you're worrying about?\\n>>\\n>> I didn't find a problem either. I just feel uncomfortable if, at the\\n>> moment of interruption, we have a descriptor of another query than the\\n>> query have been executing and holding resources.\\n> \\n> I think that \\\"descriptor\\\" here refers to ActiveQueryDesc, in which case\\n> it is updated at the beginning of ExecutorRun(), so I am wondering if\\n> the situation you're worried about would not occur.\\nAs you can see, in my example we have the only DDL and no queries with \\nplans. In this case postgres doesn't call ExecutorRun() just because it \\ndoesn't have a plan. But locks will be obtained.\\n\\n-- \\nregards,\\nAndrey Lepikhov\\nPostgres Professional\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 28 Sep 2023 09:32:46 +0700\",\n \"msg_from\": \"Andrey Lepikhov <a.lepikhov@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Thu, Sep 7, 2023 at 2:09 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-09-06 11:17, James Coleman wrote:\\n>\\n> >> > I've never been able to reproduce it (haven't tested the new version,\\n> >> > but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\\n> >> > (first buster and now bullseye).\\n> >> >\\n> >> > I'm attaching several stacktraces in the hope that they provide some\\n> >> > clues. These all match the ps output I sent earlier, though note in\\n> >> > that output there is both the regress instance and my test instance\\n> >> > (pid 3213249) running (different ports, of course, and they are from\\n> >> > the exact same compilation run). I've attached ps output for the\\n> >> > postgres processes under the make check process to simplify cross\\n> >> > referencing.\\n> >> >\\n> >> > A few interesting things:\\n> >> > - There's definitely a lock on a relation that seems to be what's\\n> >> > blocking the processes.\\n> >> > - When I try to connect with psql the process forks but then hangs\\n> >> > (see the ps output with task names stuck in \\\"authentication\\\"). I've\\n> >> > also included a trace from one of these.\\n> >>\\n> >> Thanks for sharing them!\\n> >>\\n> >> Many processes are waiting to acquire the LW lock, including the\\n> >> process\\n> >> trying to output the plan(select1.trace).\\n> >>\\n> >> I suspect that this is due to a lock that was acquired prior to being\\n> >> interrupted by ProcessLogQueryPlanInterrupt(), but have not been able\\n> >> to\\n> >> reproduce the same situation..\\n> >>\\n> >\\n> > I don't have time immediately to dig into this, but perhaps loading up\\n> > the core dumps would allow us to see what query is running in each\\n> > backend process (if it hasn't already been discarded by that point)\\n> > and thereby determine what point in each test process led to the error\\n> > condition?\\n>\\n> Thanks for the suggestion.\\n> I am concerned that core dumps may not be readable on different\\n> operating systems or other environments. (Unfortunately, I do not have\\n> Debian on hand)\\n>\\n> It seems that we can know what queries were running from the stack\\n> traces you shared.\\n> As described above, I suspect a lock which was acquired prior to\\n> ProcessLogQueryPlanInterrupt() caused the issue.\\n> I will try a little more to see if I can devise a way to create the same\\n> situation.\\n>\\n> > Alternatively we might be able to apply the same trick to the test\\n> > client instead...\\n> >\\n> > BTW, for my own easy reference in this thread: relid 1259 is pg_class\\n> > if I'm not mistaken.\\n>\\n> Yeah, and I think it's strange that the lock to 1259 appears twice and\\n> must be avoided.\\n>\\n> #10 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) at\\n> lmgr.c:117\\n> ..\\n> #49 0x0000559d61b4989d in ProcessLogQueryPlanInterrupt () at\\n> explain.c:5158\\n> ..\\n> #53 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) at\\n> lmgr.c:117\\n\\nI chatted with Andres and David about this at PGConf.NYC, and I think\\nwhat we need to do is explicitly disallow running this code any time\\nwe are inside of lock acquisition code.\\n\\nRegards,\\nJames Coleman\\n\\n\\n\",\n \"msg_date\": \"Tue, 3 Oct 2023 14:00:11 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-10-04 03:00, James Coleman wrote:\\n> On Thu, Sep 7, 2023 at 2:09 AM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> \\n>> On 2023-09-06 11:17, James Coleman wrote:\\n>> \\n>> >> > I've never been able to reproduce it (haven't tested the new version,\\n>> >> > but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\\n>> >> > (first buster and now bullseye).\\n>> >> >\\n>> >> > I'm attaching several stacktraces in the hope that they provide some\\n>> >> > clues. These all match the ps output I sent earlier, though note in\\n>> >> > that output there is both the regress instance and my test instance\\n>> >> > (pid 3213249) running (different ports, of course, and they are from\\n>> >> > the exact same compilation run). I've attached ps output for the\\n>> >> > postgres processes under the make check process to simplify cross\\n>> >> > referencing.\\n>> >> >\\n>> >> > A few interesting things:\\n>> >> > - There's definitely a lock on a relation that seems to be what's\\n>> >> > blocking the processes.\\n>> >> > - When I try to connect with psql the process forks but then hangs\\n>> >> > (see the ps output with task names stuck in \\\"authentication\\\"). I've\\n>> >> > also included a trace from one of these.\\n>> >>\\n>> >> Thanks for sharing them!\\n>> >>\\n>> >> Many processes are waiting to acquire the LW lock, including the\\n>> >> process\\n>> >> trying to output the plan(select1.trace).\\n>> >>\\n>> >> I suspect that this is due to a lock that was acquired prior to being\\n>> >> interrupted by ProcessLogQueryPlanInterrupt(), but have not been able\\n>> >> to\\n>> >> reproduce the same situation..\\n>> >>\\n>> >\\n>> > I don't have time immediately to dig into this, but perhaps loading up\\n>> > the core dumps would allow us to see what query is running in each\\n>> > backend process (if it hasn't already been discarded by that point)\\n>> > and thereby determine what point in each test process led to the error\\n>> > condition?\\n>> \\n>> Thanks for the suggestion.\\n>> I am concerned that core dumps may not be readable on different\\n>> operating systems or other environments. (Unfortunately, I do not have\\n>> Debian on hand)\\n>> \\n>> It seems that we can know what queries were running from the stack\\n>> traces you shared.\\n>> As described above, I suspect a lock which was acquired prior to\\n>> ProcessLogQueryPlanInterrupt() caused the issue.\\n>> I will try a little more to see if I can devise a way to create the \\n>> same\\n>> situation.\\n>> \\n>> > Alternatively we might be able to apply the same trick to the test\\n>> > client instead...\\n>> >\\n>> > BTW, for my own easy reference in this thread: relid 1259 is pg_class\\n>> > if I'm not mistaken.\\n>> \\n>> Yeah, and I think it's strange that the lock to 1259 appears twice and\\n>> must be avoided.\\n>> \\n>> #10 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) \\n>> at\\n>> lmgr.c:117\\n>> ..\\n>> #49 0x0000559d61b4989d in ProcessLogQueryPlanInterrupt () at\\n>> explain.c:5158\\n>> ..\\n>> #53 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1) \\n>> at\\n>> lmgr.c:117\\n> \\n> I chatted with Andres and David about this at PGConf.NYC,\\nThanks again for the discussion with hackers!\\n\\n> and I think\\n> what we need to do is explicitly disallow running this code any time\\n> we are inside of lock acquisition code.\\n\\nYeah, I think it's a straightforward workaround.\\nAnd I'm also concerned that the condition of being in the process\\nof acquiring some kind of lock is too strict and will make it almost\\nimpossible to output a running plan.\\n\\nAnyway I'm going to implement it and run pg_log_query_plan()\\nwhile the regression test is running to see how successful the plan\\noutput is.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Fri, 06 Oct 2023 21:58:46 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 6, 2023 at 8:58 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-10-04 03:00, James Coleman wrote:\\n> > On Thu, Sep 7, 2023 at 2:09 AM torikoshia <torikoshia@oss.nttdata.com>\\n> > wrote:\\n> >>\\n> >> On 2023-09-06 11:17, James Coleman wrote:\\n> >>\\n> >> >> > I've never been able to reproduce it (haven't tested the new version,\\n> >> >> > but v28 at least) on my M1 Mac; where I've reproduced it is on Debian\\n> >> >> > (first buster and now bullseye).\\n> >> >> >\\n> >> >> > I'm attaching several stacktraces in the hope that they provide some\\n> >> >> > clues. These all match the ps output I sent earlier, though note in\\n> >> >> > that output there is both the regress instance and my test instance\\n> >> >> > (pid 3213249) running (different ports, of course, and they are from\\n> >> >> > the exact same compilation run). I've attached ps output for the\\n> >> >> > postgres processes under the make check process to simplify cross\\n> >> >> > referencing.\\n> >> >> >\\n> >> >> > A few interesting things:\\n> >> >> > - There's definitely a lock on a relation that seems to be what's\\n> >> >> > blocking the processes.\\n> >> >> > - When I try to connect with psql the process forks but then hangs\\n> >> >> > (see the ps output with task names stuck in \\\"authentication\\\"). I've\\n> >> >> > also included a trace from one of these.\\n> >> >>\\n> >> >> Thanks for sharing them!\\n> >> >>\\n> >> >> Many processes are waiting to acquire the LW lock, including the\\n> >> >> process\\n> >> >> trying to output the plan(select1.trace).\\n> >> >>\\n> >> >> I suspect that this is due to a lock that was acquired prior to being\\n> >> >> interrupted by ProcessLogQueryPlanInterrupt(), but have not been able\\n> >> >> to\\n> >> >> reproduce the same situation..\\n> >> >>\\n> >> >\\n> >> > I don't have time immediately to dig into this, but perhaps loading up\\n> >> > the core dumps would allow us to see what query is running in each\\n> >> > backend process (if it hasn't already been discarded by that point)\\n> >> > and thereby determine what point in each test process led to the error\\n> >> > condition?\\n> >>\\n> >> Thanks for the suggestion.\\n> >> I am concerned that core dumps may not be readable on different\\n> >> operating systems or other environments. (Unfortunately, I do not have\\n> >> Debian on hand)\\n> >>\\n> >> It seems that we can know what queries were running from the stack\\n> >> traces you shared.\\n> >> As described above, I suspect a lock which was acquired prior to\\n> >> ProcessLogQueryPlanInterrupt() caused the issue.\\n> >> I will try a little more to see if I can devise a way to create the\\n> >> same\\n> >> situation.\\n> >>\\n> >> > Alternatively we might be able to apply the same trick to the test\\n> >> > client instead...\\n> >> >\\n> >> > BTW, for my own easy reference in this thread: relid 1259 is pg_class\\n> >> > if I'm not mistaken.\\n> >>\\n> >> Yeah, and I think it's strange that the lock to 1259 appears twice and\\n> >> must be avoided.\\n> >>\\n> >> #10 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1)\\n> >> at\\n> >> lmgr.c:117\\n> >> ..\\n> >> #49 0x0000559d61b4989d in ProcessLogQueryPlanInterrupt () at\\n> >> explain.c:5158\\n> >> ..\\n> >> #53 0x0000559d61d8ee6e in LockRelationOid (relid=1259, lockmode=1)\\n> >> at\\n> >> lmgr.c:117\\n> >\\n> > I chatted with Andres and David about this at PGConf.NYC,\\n> Thanks again for the discussion with hackers!\\n>\\n> > and I think\\n> > what we need to do is explicitly disallow running this code any time\\n> > we are inside of lock acquisition code.\\n>\\n> Yeah, I think it's a straightforward workaround.\\n> And I'm also concerned that the condition of being in the process\\n> of acquiring some kind of lock is too strict and will make it almost\\n> impossible to output a running plan.\\n\\nI was concerned about this too, but I was wondering if we might be\\nable to cheat a bit by making such a case not clear the flag but\\ninstead just skip it for now.\\n\\nRegards,\\nJames\\n\\n\\n\",\n \"msg_date\": \"Fri, 6 Oct 2023 09:10:15 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2023-10-06 21:58:46 +0900, torikoshia wrote:\\n> Yeah, I think it's a straightforward workaround.\\n> And I'm also concerned that the condition of being in the process\\n> of acquiring some kind of lock is too strict and will make it almost\\n> impossible to output a running plan.\\n\\nHow so? We shouldn't commonly acquire relevant locks while executing a query?\\nWith a few exceptions, they should instead be acquired t the start of query\\nprocessing. We do acquire a lot of lwlocks, obviously, but we don't process\\ninterrupts during the acquisition / holding of lwlocks.\\n\\nAnd presumably the interrupt would just be processed the next time interrupt\\nprocessing is happening?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 6 Oct 2023 08:58:50 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-10-04 03:00, James Coleman wrote:\\n> and I think\\n> what we need to do is explicitly disallow running this code any time\\n> we are inside of lock acquisition code.\\n\\nUpdated patch to check if any locks have already been acquired by \\nexamining MyProc->heldLocks.\\n\\nI'm not sure this change can \\\"disallow running this code `any time` we \\nare inside of lock acquisition code\\\", but as far as select1.trace, which \\nyou shared, I believe it can prevent running explain codes since it must \\nhave set MyProc->heldLocks in LockAcquireExtended() before WaitOnLock():\\n\\n```\\n /*\\n * Set bitmask of locks this process already holds on this \\nobject.\\n */\\n MyProc->heldLocks = proclock->holdMask;\\n\\n ..(snip)..\\n\\n WaitOnLock(locallock, owner);\\n```\\n\\nOn 2023-10-07 00:58, Andres Freund wrote:\\n\\n> How so? We shouldn't commonly acquire relevant locks while executing a \\n> query?\\n> With a few exceptions, they should instead be acquired t the start of \\n> query\\n> processing. We do acquire a lot of lwlocks, obviously, but we don't \\n> process\\n> interrupts during the acquisition / holding of lwlocks.\\n> \\n> And presumably the interrupt would just be processed the next time \\n> interrupt\\n> processing is happening?\\n\\nThanks for your comments!\\n\\nI tested v30 patch with \\nv28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch which makes \\nCFI() call ProcessLogQueryPlanInterrupt() internally, and confirmed that \\nvery few logging queries failed with v30 patch.\\n\\nThis is a result in line with your prediction.\\n\\n\\n```\\n $ rg -c'ignored request for logging query plan due to lock confilcts' \\npostmaster.log\\n 8\\n```\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\",\n \"msg_date\": \"Tue, 10 Oct 2023 22:27:08 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Tue, Oct 10, 2023 at 7:00 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n\\n> Thanks for your comments!\\n>\\n> I tested v30 patch with\\n> v28-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch which makes\\n> CFI() call ProcessLogQueryPlanInterrupt() internally, and confirmed that\\n> very few logging queries failed with v30 patch.\\n>\\n> This is a result in line with your prediction.\\n>\\n>\\n> ```\\n> $ rg -c'ignored request for logging query plan due to lock confilcts'\\n> postmaster.log\\n> 8\\n\\nLike many others I think this feature is useful to debug a long running query.\\n\\nSorry for jumping late into this.\\n\\nI have a few of high level comments\\n\\nThere is a lot of similarity between what this feature does and what\\nauto explain does. I see the code is also duplicated. There is some\\nmerit in avoiding this duplication\\n1. we will get all the features of auto_explain automatically like\\nchoosing a format (this was expressed somebody earlier in this\\nthread), setings etc.\\n2. avoid bugs. E.g your code switches context after ExplainState has\\nbeen allocated. These states may leak depending upon when this\\nfunction gets called.\\n3. Building features on top as James envisions will be easier.\\n\\nConsidering the similarity with auto_explain I wondered whether this\\nfunction should be part of auto_explain contrib module itself? If we\\ndo that users will need to load auto_explain extension and thus\\ninstall executor hooks when this function doesn't need those. So may\\nnot be such a good idea. I didn't see any discussion on this.\\n\\nI tried following query to pass PID of a non-client backend to this function.\\n#select pg_log_query_plan(pid), application_name, backend_type from\\npg_stat_activity where backend_type = 'autovacuum launcher';\\n pg_log_query_plan | application_name | backend_type\\n-------------------+------------------+---------------------\\n t | | autovacuum launcher\\n(1 row)\\nI see \\\"LOG: backend with PID 2733631 is not running a query or a\\nsubtransaction is aborted\\\" in server logs. That's ok. But may be we\\nshould not send signal to these kinds of backends at all, thus\\navoiding some system calls.\\n\\nI am also wondering whether it's better to report the WARNING as\\nstatus column in the output. E.g. instead of\\n#select pg_log_query_plan(100);\\nWARNING: PID 100 is not a PostgreSQL backend process\\n pg_log_query_plan\\n-------------------\\n f\\n(1 row)\\nwe output\\n#select pg_log_query_plan(100);\\n pg_log_query_plan | status\\n-------------------+---------------------------------------------\\n f | PID 100 is not a PostgreSQL backend process\\n(1 row)\\n\\nThat looks neater and can easily be handled by scripts, applications\\nand such. But it will be inconsistent with other functions like\\npg_terminate_backend() and pg_log_backend_memory_contexts().\\n\\nI do share a concern that was discussed earlier. If a query is running\\nlonger, there's something problematic with it. A diagnostic\\nintervention breaking it further would be unwelcome. James has run\\nexperiments to shake this code for any loose breakages. He has not\\nfound any. So may be we are good. And we wouldn't know about very rare\\ncorner cases so easily without using it in the field. So fine with it.\\nIf we could add some safety net that will be great but may not be\\nnecessary for the first cut.\\n\\n--\\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Wed, 11 Oct 2023 12:52:00 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-10-11 16:22, Ashutosh Bapat wrote:\\n\\n> Like many others I think this feature is useful to debug a long running \\n> query.\\n> \\n> Sorry for jumping late into this.\\n> \\n> I have a few of high level comments\\n\\nThanks for your comments!\\n\\n> There is a lot of similarity between what this feature does and what\\n> auto explain does. I see the code is also duplicated. There is some\\n> merit in avoiding this duplication\\n> 1. we will get all the features of auto_explain automatically like\\n> choosing a format (this was expressed somebody earlier in this\\n> thread), setings etc.\\n> 2. avoid bugs. E.g your code switches context after ExplainState has\\n> been allocated. These states may leak depending upon when this\\n> function gets called.\\n> 3. Building features on top as James envisions will be easier.\\n> \\n> Considering the similarity with auto_explain I wondered whether this\\n> function should be part of auto_explain contrib module itself? If we\\n> do that users will need to load auto_explain extension and thus\\n> install executor hooks when this function doesn't need those. So may\\n> not be such a good idea. I didn't see any discussion on this.\\n\\nI once thought about adding this to auto_explain, but I left it asis for \\nbelow reasons:\\n\\n- One of the typical use case of pg_log_query_plan() would be analyzing \\nslow query on customer environments. On such environments, We cannot \\nalways control what extensions to install.\\n Of course auto_explain is a major extension and it is quite possible \\nthat they installed auto_explain, but but it is also possible they do \\nnot.\\n- It seems a bit counter-intuitive that pg_log_query_plan() is in an \\nextension called auto_explain, since it `manually`` logs plans\\n\\n> I tried following query to pass PID of a non-client backend to this \\n> function.\\n> #select pg_log_query_plan(pid), application_name, backend_type from\\n> pg_stat_activity where backend_type = 'autovacuum launcher';\\n> pg_log_query_plan | application_name | backend_type\\n> -------------------+------------------+---------------------\\n> t | | autovacuum launcher\\n> (1 row)\\n> I see \\\"LOG: backend with PID 2733631 is not running a query or a\\n> subtransaction is aborted\\\" in server logs. That's ok. But may be we\\n> should not send signal to these kinds of backends at all, thus\\n> avoiding some system calls.\\n\\nAgreed, it seems better.\\nAttached patch checks if the backendType of target process is 'client \\nbackend'.\\n\\n =# select pg_log_query_plan(pid), application_name, backend_type from \\npg_stat_activity where backend_type = 'autovacuum launcher';\\n WARNING: PID 63323 is not a PostgreSQL client backend process\\n pg_log_query_plan | application_name | backend_type\\n -------------------+------------------+---------------------\\n f | | autovacuum launcher\\n\\n\\n> I am also wondering whether it's better to report the WARNING as\\n> status column in the output. E.g. instead of\\n> #select pg_log_query_plan(100);\\n> WARNING: PID 100 is not a PostgreSQL backend process\\n> pg_log_query_plan\\n> -------------------\\n> f\\n> (1 row)\\n> we output\\n> #select pg_log_query_plan(100);\\n> pg_log_query_plan | status\\n> -------------------+---------------------------------------------\\n> f | PID 100 is not a PostgreSQL backend process\\n> (1 row)\\n> \\n> That looks neater and can easily be handled by scripts, applications\\n> and such. But it will be inconsistent with other functions like\\n> pg_terminate_backend() and pg_log_backend_memory_contexts().\\n\\nIt seems neater, but it might be inconvenient because we can no longer \\nuse it in select list like the following query as you wrote:\\n\\n #select pg_log_query_plan(pid), application_name, backend_type from\\n pg_stat_activity where backend_type = 'autovacuum launcher';\\n\\n> I do share a concern that was discussed earlier. If a query is running\\n> longer, there's something problematic with it. A diagnostic\\n> intervention breaking it further would be unwelcome. James has run\\n> experiments to shake this code for any loose breakages. He has not\\n> found any. So may be we are good. And we wouldn't know about very rare\\n> corner cases so easily without using it in the field. So fine with it.\\n> If we could add some safety net that will be great but may not be\\n> necessary for the first cut.\\n\\nIf there are candidates for the safety net, I'm willing to add them.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\",\n \"msg_date\": \"Thu, 12 Oct 2023 22:21:26 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 12, 2023 at 6:51 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-10-11 16:22, Ashutosh Bapat wrote:\\n> >\\n> > Considering the similarity with auto_explain I wondered whether this\\n> > function should be part of auto_explain contrib module itself? If we\\n> > do that users will need to load auto_explain extension and thus\\n> > install executor hooks when this function doesn't need those. So may\\n> > not be such a good idea. I didn't see any discussion on this.\\n>\\n> I once thought about adding this to auto_explain, but I left it asis for\\n> below reasons:\\n>\\n> - One of the typical use case of pg_log_query_plan() would be analyzing\\n> slow query on customer environments. On such environments, We cannot\\n> always control what extensions to install.\\n\\nThe same argument applies to auto_explain functionality as well. But\\nit's not part of the core.\\n\\n> Of course auto_explain is a major extension and it is quite possible\\n> that they installed auto_explain, but but it is also possible they do\\n> not.\\n> - It seems a bit counter-intuitive that pg_log_query_plan() is in an\\n> extension called auto_explain, since it `manually`` logs plans\\n>\\n\\npg_log_query_plan() may not fit auto_explain but\\npg_explain_backend_query() does. What we are logging is more than just\\nplan of the query, it might expand to be closer to explain output.\\nWhile auto in auto_explain would refer to its automatically logging\\nexplain outputs, it can provide an additional function which provides\\nsimilar functionality by manually triggering it.\\n\\nBut we can defer this to a committer, if you want.\\n\\nI am more interested in avoiding the duplication of code, esp. the\\nfirst comment in my reply\\n>> There is a lot of similarity between what this feature does and what\\n>> auto explain does. I see the code is also duplicated. There is some\\n>> merit in avoiding this duplication\\n>> 1. we will get all the features of auto_explain automatically like\\n>> choosing a format (this was expressed somebody earlier in this\\n>> thread), setings etc.\\n>> 2. avoid bugs. E.g your code switches context after ExplainState has\\n>> been allocated. These states may leak depending upon when this\\n>> function gets called.\\n>> 3. Building features on top as James envisions will be easier.\\n\\n>\\n> =# select pg_log_query_plan(pid), application_name, backend_type from\\n> pg_stat_activity where backend_type = 'autovacuum launcher';\\n> WARNING: PID 63323 is not a PostgreSQL client backend process\\n> pg_log_query_plan | application_name | backend_type\\n> -------------------+------------------+---------------------\\n> f | | autovacuum launcher\\n>\\n>\\n> > I am also wondering whether it's better to report the WARNING as\\n> > status column in the output. E.g. instead of\\n> > #select pg_log_query_plan(100);\\n> > WARNING: PID 100 is not a PostgreSQL backend process\\n> > pg_log_query_plan\\n> > -------------------\\n> > f\\n> > (1 row)\\n> > we output\\n> > #select pg_log_query_plan(100);\\n> > pg_log_query_plan | status\\n> > -------------------+---------------------------------------------\\n> > f | PID 100 is not a PostgreSQL backend process\\n> > (1 row)\\n> >\\n> > That looks neater and can easily be handled by scripts, applications\\n> > and such. But it will be inconsistent with other functions like\\n> > pg_terminate_backend() and pg_log_backend_memory_contexts().\\n>\\n> It seems neater, but it might be inconvenient because we can no longer\\n> use it in select list like the following query as you wrote:\\n>\\n> #select pg_log_query_plan(pid), application_name, backend_type from\\n> pg_stat_activity where backend_type = 'autovacuum launcher';\\n\\nWhy is that?\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Mon, 16 Oct 2023 15:16:28 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-10-16 18:46, Ashutosh Bapat wrote:\\n> On Thu, Oct 12, 2023 at 6:51 PM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> \\n>> On 2023-10-11 16:22, Ashutosh Bapat wrote:\\n>> >\\n>> > Considering the similarity with auto_explain I wondered whether this\\n>> > function should be part of auto_explain contrib module itself? If we\\n>> > do that users will need to load auto_explain extension and thus\\n>> > install executor hooks when this function doesn't need those. So may\\n>> > not be such a good idea. I didn't see any discussion on this.\\n>> \\n>> I once thought about adding this to auto_explain, but I left it asis \\n>> for\\n>> below reasons:\\n>> \\n>> - One of the typical use case of pg_log_query_plan() would be \\n>> analyzing\\n>> slow query on customer environments. On such environments, We cannot\\n>> always control what extensions to install.\\n> \\n> The same argument applies to auto_explain functionality as well. But\\n> it's not part of the core.\\n\\nYeah, and when we have a situation where we want to run\\npg_log_query_plan(), we can run it in any environment as long as it is\\n bundled with the core.\\nOn the other hand, if it is built into auto_explain, we need to start by\\ninstalling auto_explain if we do not have auto_explain, which is often\\ndifficult to do in production environments.\\n\\n>> Of course auto_explain is a major extension and it is quite \\n>> possible\\n>> that they installed auto_explain, but but it is also possible they do\\n>> not.\\n>> - It seems a bit counter-intuitive that pg_log_query_plan() is in an\\n>> extension called auto_explain, since it `manually`` logs plans\\n>> \\n> \\n> pg_log_query_plan() may not fit auto_explain but\\n> pg_explain_backend_query() does. What we are logging is more than just\\n> plan of the query, it might expand to be closer to explain output.\\n> While auto in auto_explain would refer to its automatically logging\\n> explain outputs, it can provide an additional function which provides\\n> similar functionality by manually triggering it.\\n> \\n> But we can defer this to a committer, if you want.\\n> \\n> I am more interested in avoiding the duplication of code, esp. the\\n> first comment in my reply\\n\\nIf there are no objections, I will try porting it to auto_explain and\\nsee its feasibility.\\n\\n>>> There is a lot of similarity between what this feature does and what\\n>>> auto explain does. I see the code is also duplicated. There is some\\n>>> merit in avoiding this duplication\\n>>> 1. we will get all the features of auto_explain automatically like\\n>>> choosing a format (this was expressed somebody earlier in this\\n>>> thread), setings etc.\\n>>> 2. avoid bugs. E.g your code switches context after ExplainState has\\n>>> been allocated. These states may leak depending upon when this\\n>>> function gets called.\\n>>> 3. Building features on top as James envisions will be easier.\\n> \\n>> \\n>> =# select pg_log_query_plan(pid), application_name, backend_type \\n>> from\\n>> pg_stat_activity where backend_type = 'autovacuum launcher';\\n>> WARNING: PID 63323 is not a PostgreSQL client backend process\\n>> pg_log_query_plan | application_name | backend_type\\n>> -------------------+------------------+---------------------\\n>> f | | autovacuum launcher\\n>> \\n>> \\n>> > I am also wondering whether it's better to report the WARNING as\\n>> > status column in the output. E.g. instead of\\n>> > #select pg_log_query_plan(100);\\n>> > WARNING: PID 100 is not a PostgreSQL backend process\\n>> > pg_log_query_plan\\n>> > -------------------\\n>> > f\\n>> > (1 row)\\n>> > we output\\n>> > #select pg_log_query_plan(100);\\n>> > pg_log_query_plan | status\\n>> > -------------------+---------------------------------------------\\n>> > f | PID 100 is not a PostgreSQL backend process\\n>> > (1 row)\\n>> >\\n>> > That looks neater and can easily be handled by scripts, applications\\n>> > and such. But it will be inconsistent with other functions like\\n>> > pg_terminate_backend() and pg_log_backend_memory_contexts().\\n>> \\n>> It seems neater, but it might be inconvenient because we can no longer\\n>> use it in select list like the following query as you wrote:\\n>> \\n>> #select pg_log_query_plan(pid), application_name, backend_type from\\n>> pg_stat_activity where backend_type = 'autovacuum launcher';\\n> \\n> Why is that?\\n\\nSorry, I misunderstood and confirmed we can run queries like below:\\n\\n```\\n=# CREATE OR REPLACE FUNCTION pg_log_query_plan_stab(i int) \\n RETURNS TABLE( \\n \\n pg_log_query_plan bool, \\n status text \\n \\n ) AS $$ \\n DECLARE \\n \\n BEGIN \\n \\n RETURN QUERY SELECT false::bool, 'PID 100 is not a PostgreSQL \\nbackend process'::text; END; \\n \\n $$ LANGUAGE plpgsql;\\n\\n=# select pg_log_query_plan_stab(pid), application_name, backend_type \\nfrom pg_stat_activity where backend_type = 'autovacuum launcher';\\n pg_log_query_plan_stab | application_name | \\n backend_type\\n---------------------------------------------------+------------------+---------------------\\n (f,\\\"PID 100 is not a PostgreSQL backend process\\\") | | \\nautovacuum launcher\\n```\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Wed, 18 Oct 2023 15:09:38 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 18, 2023 at 11:39 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n> >\\n> > I am more interested in avoiding the duplication of code, esp. the\\n> > first comment in my reply\\n>\\n> If there are no objections, I will try porting it to auto_explain and\\n> see its feasibility.\\n\\nIf you want it in core, you will need to port relevant parts of\\nauto_explain code to core.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Wed, 18 Oct 2023 15:43:01 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 18, 2023 at 2:09 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-10-16 18:46, Ashutosh Bapat wrote:\\n> > On Thu, Oct 12, 2023 at 6:51 PM torikoshia <torikoshia@oss.nttdata.com>\\n> > wrote:\\n> >>\\n> >> On 2023-10-11 16:22, Ashutosh Bapat wrote:\\n> >> >\\n> >> > Considering the similarity with auto_explain I wondered whether this\\n> >> > function should be part of auto_explain contrib module itself? If we\\n> >> > do that users will need to load auto_explain extension and thus\\n> >> > install executor hooks when this function doesn't need those. So may\\n> >> > not be such a good idea. I didn't see any discussion on this.\\n> >>\\n> >> I once thought about adding this to auto_explain, but I left it asis\\n> >> for\\n> >> below reasons:\\n> >>\\n> >> - One of the typical use case of pg_log_query_plan() would be\\n> >> analyzing\\n> >> slow query on customer environments. On such environments, We cannot\\n> >> always control what extensions to install.\\n> >\\n> > The same argument applies to auto_explain functionality as well. But\\n> > it's not part of the core.\\n>\\n> Yeah, and when we have a situation where we want to run\\n> pg_log_query_plan(), we can run it in any environment as long as it is\\n> bundled with the core.\\n> On the other hand, if it is built into auto_explain, we need to start by\\n> installing auto_explain if we do not have auto_explain, which is often\\n> difficult to do in production environments.\\n>\\n> >> Of course auto_explain is a major extension and it is quite\\n> >> possible\\n> >> that they installed auto_explain, but but it is also possible they do\\n> >> not.\\n> >> - It seems a bit counter-intuitive that pg_log_query_plan() is in an\\n> >> extension called auto_explain, since it `manually`` logs plans\\n> >>\\n> >\\n> > pg_log_query_plan() may not fit auto_explain but\\n> > pg_explain_backend_query() does. What we are logging is more than just\\n> > plan of the query, it might expand to be closer to explain output.\\n> > While auto in auto_explain would refer to its automatically logging\\n> > explain outputs, it can provide an additional function which provides\\n> > similar functionality by manually triggering it.\\n> >\\n> > But we can defer this to a committer, if you want.\\n> >\\n> > I am more interested in avoiding the duplication of code, esp. the\\n> > first comment in my reply\\n>\\n> If there are no objections, I will try porting it to auto_explain and\\n> see its feasibility.\\n>\\n> >>> There is a lot of similarity between what this feature does and what\\n> >>> auto explain does. I see the code is also duplicated. There is some\\n> >>> merit in avoiding this duplication\\n> >>> 1. we will get all the features of auto_explain automatically like\\n> >>> choosing a format (this was expressed somebody earlier in this\\n> >>> thread), setings etc.\\n> >>> 2. avoid bugs. E.g your code switches context after ExplainState has\\n> >>> been allocated. These states may leak depending upon when this\\n> >>> function gets called.\\n> >>> 3. Building features on top as James envisions will be easier.\\n\\nIn my view the fact that auto_explain is itself not part of core is a\\nmistake, and I know there are more prominent hackers than myself who\\nhold that view.\\n\\nWhile that decision as regards auto_explain has long since been made\\n(and there would be work to undo it), I don't believe that we should\\nrepeat that choice here. I'm -10 on moving this into auto_explain.\\n\\nHowever perhaps there is still an opportunity for moving some of the\\nauto_explain code into core so as to enable deduplicating the code.\\n\\nRegards,\\nJames Coleman\\n\\n\\n\",\n \"msg_date\": \"Wed, 18 Oct 2023 12:34:36 -0400\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"> Hi,\\n> \\n> Yeah, and when we have a situation where we want to run\\n> pg_log_query_plan(), we can run it in any environment as long as it\\n> is bundled with the core.\\n\\nIs it possible to get the plan of a backend programmatically without\\naccess to the logs?\\n\\nSomething like pg_get_query_plan(pid, format) which output would be the\\nsame as EXPLAIN.\\n\\nRegards,\\nÉtienne BERSAC\\nDalibo\\n\\n\\n\",\n \"msg_date\": \"Tue, 24 Oct 2023 09:12:46 +0200\",\n \"msg_from\": \"=?ISO-8859-1?Q?=C9tienne?= BERSAC <etienne.bersac@dalibo.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-10-24 16:12, Étienne BERSAC wrote:\\n>> Hi,\\n>> \\n>> Yeah, and when we have a situation where we want to run\\n>> pg_log_query_plan(), we can run it in any environment as long as it\\n>> is bundled with the core.\\n> \\n> Is it possible to get the plan of a backend programmatically without\\n> access to the logs?\\n> \\n> Something like pg_get_query_plan(pid, format) which output would be the\\n> same as EXPLAIN.\\n\\nDo you imagine a function like below?\\n\\n =# select pg_get_query_plan(100, 'plain');\\n QUERY PLAN\\n -------------------------------------------------------------------\\n Limit (cost=0.00..0.04 rows=1 width=273)\\n -> Seq Scan on pg_class (cost=0.00..17.14 rows=414 width=273)\\n\\nIf so, we once tried to implement such function for getting memory \\ncontexts.\\nHowever, this attempt didn't succeed because it could lead dead lock \\nsituation[1].\\n\\nI think the same problem can occur when implementing \\npg_get_query_plan().\\n\\n[1] \\nhttps://www.postgresql.org/message-id/9a50371e15e741e295accabc72a41df1%40oss.nttdata.com\\n> \\n> Regards,\\n> Étienne BERSAC\\n> Dalibo\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Tue, 24 Oct 2023 21:30:49 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 18, 2023 at 10:04 PM James Coleman <jtc331@gmail.com> wrote:\\n>\\n> While that decision as regards auto_explain has long since been made\\n> (and there would be work to undo it), I don't believe that we should\\n> repeat that choice here. I'm -10 on moving this into auto_explain.\\n>\\n\\nRight.\\n\\n> However perhaps there is still an opportunity for moving some of the\\n> auto_explain code into core so as to enable deduplicating the code.\\n>\\n\\nRight. That's what I also think.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Wed, 25 Oct 2023 09:10:06 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-10-25 12:40, Ashutosh Bapat wrote:\\n> On Wed, Oct 18, 2023 at 10:04 PM James Coleman <jtc331@gmail.com> \\n> wrote:\\n>> \\n>> While that decision as regards auto_explain has long since been made\\n>> (and there would be work to undo it), I don't believe that we should\\n>> repeat that choice here. I'm -10 on moving this into auto_explain.\\n>> \\n> \\n> Right.\\n> \\n>> However perhaps there is still an opportunity for moving some of the\\n>> auto_explain code into core so as to enable deduplicating the code.\\n>> \\n> \\n> Right. That's what I also think.\\n\\nThanks for your comments.\\n\\nAttached patch adds a new function which assembles es->str for logging\\naccording to specified contents and format.\\nThis is called from both auto_explain and pg_log_query_plan().\\n\\nOn 2023-10-11 16:22, Ashutosh Bapat wrote:\\n> I am also wondering whether it's better to report the WARNING as\\n> status column in the output.\\n\\nAttached patch left as it was since the inconsistency with\\npg_terminate_backend() and pg_log_backend_memory_contexts() as you\\npointed out.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\",\n \"msg_date\": \"Thu, 26 Oct 2023 15:53:22 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"\\nHi Torikoshia,\\n\\n> If so, we once tried to implement such function for getting memory \\n> contexts.\\n> However, this attempt didn't succeed because it could lead dead lock \\n> situation[1].\\n\\nThanks for the pointer. Why not use client log message to allow client\\nto get plan without locking backend memory context and without access\\nto server log ? I missed the rationnal for not sending the plan to\\nclient.\\n\\nRegards,\\nÉtienne\\n\\n\\n\",\n \"msg_date\": \"Fri, 27 Oct 2023 09:06:31 +0200\",\n \"msg_from\": \"=?ISO-8859-1?Q?=C9tienne?= BERSAC <etienne.bersac@dalibo.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-10-27 16:06, Étienne BERSAC wrote:\\n> Hi Torikoshia,\\n> \\n>> If so, we once tried to implement such function for getting memory\\n>> contexts.\\n>> However, this attempt didn't succeed because it could lead dead lock\\n>> situation[1].\\n> \\n> Thanks for the pointer. Why not use client log message to allow client\\n> to get plan without locking backend memory context and without access\\n> to server log ? I missed the rationnal for not sending the plan to\\n> client.\\n\\nIf we use client log message, that message is shown on the target \\nprocess whose pid is specified by the parameter of pg_log_query_plan():\\n\\n (pid:1000)=# select pg_sleep(60);\\n (pid:1001)=# select pg_log_query_plan(1000);\\n (pid:1000)=# LOG: query plan running on backend with PID 1000 is:\\n Query Text: select pg_sleep(1000);\\n Result (cost=0.00..0.01 rows=1 width=4)\\n Output: pg_sleep('1000'::double precision)\\n\\nI think this is not an expected behavior and we set elevel to \\nLOG_SERVER_ONLY.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Fri, 27 Oct 2023 21:30:02 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"> Hi,\\n> \\n\\n> If we use client log message, that message is shown on the target \\n> process whose pid is specified by the parameter of\\n> pg_log_query_plan():\\n> \\n> (pid:1000)=# select pg_sleep(60);\\n> (pid:1001)=# select pg_log_query_plan(1000);\\n> (pid:1000)=# LOG: query plan running on backend with PID 1000 is:\\n> Query Text: select pg_sleep(1000);\\n> Result (cost=0.00..0.01 rows=1 width=4)\\n> Output: pg_sleep('1000'::double precision)\\n> \\n> I think this is not an expected behavior and we set elevel to \\n> LOG_SERVER_ONLY.\\n\\n\\nMakes sens. Thanks.\\n\\n\\n\",\n \"msg_date\": \"Fri, 27 Oct 2023 14:54:48 +0200\",\n \"msg_from\": \"=?ISO-8859-1?Q?=C9tienne?= BERSAC <etienne.bersac@dalibo.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"I have following questions related to the functionality. (Please point\\nme to the relevant discussion if this has been already discussed.)\\n\\n1. When a backend is running nested queries, we will see the plan of\\nthe innermost query. That query may not be the actual culprit if the\\nuser query is running slowly. E.g a query being run as part of inner\\nside nested loop when the nested loop itself is the bottleneck. I\\nthink it will be useful to print plans of all the whole query stack.\\n\\n2. When a query is running in parallel worker do we want to print that\\nquery? It may or may not be interesting given the situation. If the\\noverall plan itself is faulty, output of the parallel worker query is\\nnot useful. If the plan is fine but a given worker's leg is running\\nslowly it may be interesting.\\n\\nAs a side note, you may want to fix the indentation in\\nExplainAssembleLogOutput().\\n\\nOn Fri, Oct 27, 2023 at 6:24 PM Étienne BERSAC\\n<etienne.bersac@dalibo.com> wrote:\\n>\\n> > Hi,\\n> >\\n>\\n> > If we use client log message, that message is shown on the target\\n> > process whose pid is specified by the parameter of\\n> > pg_log_query_plan():\\n> >\\n> > (pid:1000)=# select pg_sleep(60);\\n> > (pid:1001)=# select pg_log_query_plan(1000);\\n> > (pid:1000)=# LOG: query plan running on backend with PID 1000 is:\\n> > Query Text: select pg_sleep(1000);\\n> > Result (cost=0.00..0.01 rows=1 width=4)\\n> > Output: pg_sleep('1000'::double precision)\\n> >\\n> > I think this is not an expected behavior and we set elevel to\\n> > LOG_SERVER_ONLY.\\n>\\n>\\n> Makes sens. Thanks.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Fri, 3 Nov 2023 19:31:14 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 3, 2023 at 7:31 PM Ashutosh Bapat\\n<ashutosh.bapat.oss@gmail.com> wrote:\\n>\\n> I have following questions related to the functionality. (Please point\\n> me to the relevant discussion if this has been already discussed.)\\n>\\n> 1. When a backend is running nested queries, we will see the plan of\\n> the innermost query. That query may not be the actual culprit if the\\n> user query is running slowly. E.g a query being run as part of inner\\n> side nested loop when the nested loop itself is the bottleneck. I\\n> think it will be useful to print plans of all the whole query stack.\\n\\nTo further explain this point, consider following scenario\\n\\n-- A completely useless function which executes two SQL statements\\nwhich take small amount individually\\n#create function multstmt() returns int\\nlanguage sql as $$\\nselect count(*) from pg_class where reltype = 12345;\\nselect count(*) from pg_type limit 10;\\n$$;\\n\\n-- force a suboptimal plan\\n#set enable_hashjoin to false;\\n#set enable_mergejoin to false;\\n\\n-- A completely useless but long running query\\n#select c.oid, t.oid from pg_class c, pg_type t where multstmt(c.oid)\\n= multstmt(t.oid);\\n\\nThis take a few minutes on my laptop.\\n\\nIn a separate session query pg_stat_activity. We will see the original query\\n#select pid, query, backend_type from pg_stat_activity where pid = 349330;\\n pid | query\\n | backend_type\\n--------+-----------------------------------------------------------------------------------------+----------------\\n 349330 | select c.oid, t.oid from pg_class c, pg_type t where\\nmultstmt(c.oid) = multstmt(t.oid); | client backend\\n(1 row)\\n\\nRun the plan output function a few times\\n#select pg_log_query_plan(349330);\\n\\nYou will observe different plans based on which of the innermost query\\nis runnning\\nLOG: query plan running on backend with PID 349330 is:\\n Query Text:\\n select count(*) from pg_class where reltype = typeid;\\n select count(*) from pg_type where oid = typeid;\\n\\n Query Parameters: $1 = '600'\\n Aggregate (cost=18.18..18.19 rows=1 width=8)\\n Output: count(*)\\n -> Seq Scan on pg_catalog.pg_class (cost=0.00..18.18 rows=2 width=0)\\n Output: oid, relname, relnamespace, reltype,\\nreloftype, relowner, relam, relfilenode, reltablespace, relpages,\\nreltuples, relallvisible, reltoastrelid, relhasindex, relisshared,\\nrelpersistence, relkind, relnatts, relchecks, relhasrules,\\nrelhastriggers, relhassubclass, relrowsecurity, relforcerowsecurity,\\nrelispopulated, relreplident, relispartition, relrewrite,\\nrelfrozenxid, relminmxid, relacl, reloptions, relpartbound\\n Filter: (pg_class.reltype = $1)\\n Settings: enable_hashjoin = 'off', enable_mergejoin = 'off'\\n2023-11-06 11:52:25.610 IST [349330] LOG: query plan running on\\nbackend with PID 349330 is:\\n Query Text:\\n select count(*) from pg_class where reltype = typeid;\\n select count(*) from pg_type where oid = typeid;\\n\\n Query Parameters: $1 = '2203'\\n Aggregate (cost=4.30..4.31 rows=1 width=4)\\n Output: count(*)\\n -> Index Only Scan using pg_type_oid_index on\\npg_catalog.pg_type (cost=0.28..4.29 rows=1 width=0)\\n Output: oid\\n Index Cond: (pg_type.oid = $1)\\n Settings: enable_hashjoin = 'off', enable_mergejoin = 'off'\\n\\nIndividual pieces are confusing here since the query run by the\\nbackend is not the one being EXPLAINed. A user may not know that the\\nqueries being EXPLAINed arise from the function call multstmt(). So\\nthey won't be able to stitch the pieces together unless they see plan\\nof the outermost query with loops and costs. What might help if we\\nexplain each query in the hierarchy.\\n\\nI think we can start with what auto_explain is doing. Always print the\\nplan of the outermost query; the query found in pg_stat_activity. In a\\nlater version we might find a way to print plans of all the queries in\\nthe stack and do so in a readable manner.\\n\\nThis makes tracking activeQueryDesc a bit tricky. My guess is that the\\noutermost query's descriptor will be available through ActivePortal\\nmost of the time. But there are cases when ExecutorRun is called by\\npassing a queryDesc directly. So may be there are some cases where\\nthat's not true.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 Nov 2023 12:02:12 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-11-06 15:32, Ashutosh Bapat wrote:\\n\\nThanks for the suggestion and example.\\n\\n> On Fri, Nov 3, 2023 at 7:31 PM Ashutosh Bapat\\n> <ashutosh.bapat.oss@gmail.com> wrote:\\n>> \\n>> I have following questions related to the functionality. (Please point\\n>> me to the relevant discussion if this has been already discussed.)\\n>> \\n>> 1. When a backend is running nested queries, we will see the plan of\\n>> the innermost query. That query may not be the actual culprit if the\\n>> user query is running slowly. E.g a query being run as part of inner\\n>> side nested loop when the nested loop itself is the bottleneck. I\\n>> think it will be useful to print plans of all the whole query stack.\\n\\nThis was discussed in previous threads[1] and we thought it'd be useful \\nbut since it needed some extra works, we stopped widening the scope.\\n\\n> To further explain this point, consider following scenario\\n> \\n> -- A completely useless function which executes two SQL statements\\n> which take small amount individually\\n> #create function multstmt() returns int\\n> language sql as $$\\n> select count(*) from pg_class where reltype = 12345;\\n> select count(*) from pg_type limit 10;\\n> $$;\\n> \\n> -- force a suboptimal plan\\n> #set enable_hashjoin to false;\\n> #set enable_mergejoin to false;\\n> \\n> -- A completely useless but long running query\\n> #select c.oid, t.oid from pg_class c, pg_type t where multstmt(c.oid)\\n> = multstmt(t.oid);\\n> \\n> This take a few minutes on my laptop.\\n> \\n> In a separate session query pg_stat_activity. We will see the original \\n> query\\n> #select pid, query, backend_type from pg_stat_activity where pid = \\n> 349330;\\n> pid | query\\n> | backend_type\\n> --------+-----------------------------------------------------------------------------------------+----------------\\n> 349330 | select c.oid, t.oid from pg_class c, pg_type t where\\n> multstmt(c.oid) = multstmt(t.oid); | client backend\\n> (1 row)\\n> \\n> Run the plan output function a few times\\n> #select pg_log_query_plan(349330);\\n> \\n> You will observe different plans based on which of the innermost query\\n> is runnning\\n> LOG: query plan running on backend with PID 349330 is:\\n> Query Text:\\n> select count(*) from pg_class where reltype = typeid;\\n> select count(*) from pg_type where oid = typeid;\\n> \\n> Query Parameters: $1 = '600'\\n> Aggregate (cost=18.18..18.19 rows=1 width=8)\\n> Output: count(*)\\n> -> Seq Scan on pg_catalog.pg_class (cost=0.00..18.18 rows=2 \\n> width=0)\\n> Output: oid, relname, relnamespace, reltype,\\n> reloftype, relowner, relam, relfilenode, reltablespace, relpages,\\n> reltuples, relallvisible, reltoastrelid, relhasindex, relisshared,\\n> relpersistence, relkind, relnatts, relchecks, relhasrules,\\n> relhastriggers, relhassubclass, relrowsecurity, relforcerowsecurity,\\n> relispopulated, relreplident, relispartition, relrewrite,\\n> relfrozenxid, relminmxid, relacl, reloptions, relpartbound\\n> Filter: (pg_class.reltype = $1)\\n> Settings: enable_hashjoin = 'off', enable_mergejoin = 'off'\\n> 2023-11-06 11:52:25.610 IST [349330] LOG: query plan running on\\n> backend with PID 349330 is:\\n> Query Text:\\n> select count(*) from pg_class where reltype = typeid;\\n> select count(*) from pg_type where oid = typeid;\\n> \\n> Query Parameters: $1 = '2203'\\n> Aggregate (cost=4.30..4.31 rows=1 width=4)\\n> Output: count(*)\\n> -> Index Only Scan using pg_type_oid_index on\\n> pg_catalog.pg_type (cost=0.28..4.29 rows=1 width=0)\\n> Output: oid\\n> Index Cond: (pg_type.oid = $1)\\n> Settings: enable_hashjoin = 'off', enable_mergejoin = 'off'\\n> \\n> Individual pieces are confusing here since the query run by the\\n> backend is not the one being EXPLAINed. A user may not know that the\\n> queries being EXPLAINed arise from the function call multstmt(). So\\n> they won't be able to stitch the pieces together unless they see plan\\n> of the outermost query with loops and costs. What might help if we\\n> explain each query in the hierarchy.\\n> \\n> I think we can start with what auto_explain is doing. Always print the\\n> plan of the outermost query; the query found in pg_stat_activity. In a\\n> later version we might find a way to print plans of all the queries in\\n> the stack and do so in a readable manner.\\n\\nAgreed there are cases printing plan of the outermost query is more \\nuseful.\\n\\n> \\n> This makes tracking activeQueryDesc a bit tricky. My guess is that the\\n> outermost query's descriptor will be available through ActivePortal\\n> most of the time. But there are cases when ExecutorRun is called by\\n> passing a queryDesc directly. So may be there are some cases where\\n> that's not true.\\n\\nYeah, actually the original version of the patch got the plan from \\nActivePortal, but it failed logging plan when the query was something \\nlike this[2]:\\n\\n DO $$\\n BEGIN\\n PERFORM pg_sleep(100);\\n END$$;\\n\\n> 2. When a query is running in parallel worker do we want to print that\\n> query? It may or may not be interesting given the situation. If the\\n> overall plan itself is faulty, output of the parallel worker query is\\n> not useful. If the plan is fine but a given worker's leg is running\\n> slowly it may be interesting.\\n\\nI think it can be useful.\\nI'm wondering if we can add this after the first patch for this feature \\nis committed.\\n\\n> As a side note, you may want to fix the indentation in\\n> ExplainAssembleLogOutput().\\n\\nThanks, modified it.\\nSince the documentation was obsoleted, attached patch also updated it.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\",\n \"msg_date\": \"Thu, 09 Nov 2023 15:33:36 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 9, 2023 at 12:03 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n> >>\\n> >> 1. When a backend is running nested queries, we will see the plan of\\n> >> the innermost query. That query may not be the actual culprit if the\\n> >> user query is running slowly. E.g a query being run as part of inner\\n> >> side nested loop when the nested loop itself is the bottleneck. I\\n> >> think it will be useful to print plans of all the whole query stack.\\n>\\n> This was discussed in previous threads[1] and we thought it'd be useful\\n> but since it needed some extra works, we stopped widening the scope.\\n>\\n> >\\n> > I think we can start with what auto_explain is doing. Always print the\\n> > plan of the outermost query; the query found in pg_stat_activity. In a\\n> > later version we might find a way to print plans of all the queries in\\n> > the stack and do so in a readable manner.\\n>\\n> Agreed there are cases printing plan of the outermost query is more\\n> useful.\\n>\\n\\nI am fine printing the plan of the outermost query. This will help\\nmany cases. Printing plans of the whole query stack can be added as an\\nadd on later.\\n\\n> >\\n> > This makes tracking activeQueryDesc a bit tricky. My guess is that the\\n> > outermost query's descriptor will be available through ActivePortal\\n> > most of the time. But there are cases when ExecutorRun is called by\\n> > passing a queryDesc directly. So may be there are some cases where\\n> > that's not true.\\n>\\n> Yeah, actually the original version of the patch got the plan from\\n> ActivePortal, but it failed logging plan when the query was something\\n> like this[2]:\\n>\\n> DO $$\\n> BEGIN\\n> PERFORM pg_sleep(100);\\n> END$$;\\n\\nReferences [1] and [2] are not listed in your email.\\n\\nIs that because there was no ActivePortal created or the ActivePortal\\npointed to DO block instead of PERFORM pg_sleep?\\n\\n>\\n> > 2. When a query is running in parallel worker do we want to print that\\n> > query? It may or may not be interesting given the situation. If the\\n> > overall plan itself is faulty, output of the parallel worker query is\\n> > not useful. If the plan is fine but a given worker's leg is running\\n> > slowly it may be interesting.\\n>\\n> I think it can be useful.\\n> I'm wondering if we can add this after the first patch for this feature\\n> is committed.\\n\\nWith the current patches, it will print the query from a parallel\\nbackend. If that's not desirable we should prohibit that case at\\nleast.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Nov 2023 12:41:49 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-11-09 16:11, Ashutosh Bapat wrote:\\n> On Thu, Nov 9, 2023 at 12:03 PM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> >>\\n>> >> 1. When a backend is running nested queries, we will see the plan of\\n>> >> the innermost query. That query may not be the actual culprit if the\\n>> >> user query is running slowly. E.g a query being run as part of inner\\n>> >> side nested loop when the nested loop itself is the bottleneck. I\\n>> >> think it will be useful to print plans of all the whole query stack.\\n>> \\n>> This was discussed in previous threads[1] and we thought it'd be \\n>> useful\\n>> but since it needed some extra works, we stopped widening the scope.\\n>> \\n>> >\\n>> > I think we can start with what auto_explain is doing. Always print the\\n>> > plan of the outermost query; the query found in pg_stat_activity. In a\\n>> > later version we might find a way to print plans of all the queries in\\n>> > the stack and do so in a readable manner.\\n>> \\n>> Agreed there are cases printing plan of the outermost query is more\\n>> useful.\\n>> \\n> \\n> I am fine printing the plan of the outermost query. This will help\\n> many cases. Printing plans of the whole query stack can be added as an\\n> add on later.\\n> \\n>> >\\n>> > This makes tracking activeQueryDesc a bit tricky. My guess is that the\\n>> > outermost query's descriptor will be available through ActivePortal\\n>> > most of the time. But there are cases when ExecutorRun is called by\\n>> > passing a queryDesc directly. So may be there are some cases where\\n>> > that's not true.\\n>> \\n>> Yeah, actually the original version of the patch got the plan from\\n>> ActivePortal, but it failed logging plan when the query was something\\n>> like this[2]:\\n>> \\n>> DO $$\\n>> BEGIN\\n>> PERFORM pg_sleep(100);\\n>> END$$;\\n> \\n> References [1] and [2] are not listed in your email.\\n\\nOops, sorry. Here are links:\\n\\n[1] \\nhttps://www.postgresql.org/message-id/64f716c44629e303b66e6c24502147cc%40oss.nttdata.com\\n[2] \\nhttps://www.postgresql.org/message-id/632e99eb-8090-53e6-1b1a-101601904cbd%40oss.nttdata.com\\n\\n> Is that because there was no ActivePortal created or the ActivePortal\\n> pointed to DO block instead of PERFORM pg_sleep?\\n\\nActivePortal is created but ActivePortal->queryDesc is null.\\n\\n>> > 2. When a query is running in parallel worker do we want to print that\\n>> > query? It may or may not be interesting given the situation. If the\\n>> > overall plan itself is faulty, output of the parallel worker query is\\n>> > not useful. If the plan is fine but a given worker's leg is running\\n>> > slowly it may be interesting.\\n>> \\n>> I think it can be useful.\\n>> I'm wondering if we can add this after the first patch for this \\n>> feature\\n>> is committed.\\n> \\n> With the current patches, it will print the query from a parallel\\n> backend. If that's not desirable we should prohibit that case at\\n> least.\\n\\nCurrent patch prohibits printing plan if backend type is parallel worker \\nas below:\\n\\n =# select pg_log_query_plan(pid), backend_type from pg_stat_activity \\nwhere backend_type = 'parallel worker';\\n\\n pg_log_query_plan | backend_type\\n -------------------+-----------------\\n f | parallel worker\\n f | parallel worker\\n (2 rows)\\n\\n WARNING: PID 4618 is not a PostgreSQL client backend process\\n WARNING: PID 4617 is not a PostgreSQL client backend process\\n\\nIs this the behavior you expect?\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Thu, 09 Nov 2023 20:26:27 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 9, 2023 at 4:56 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On 2023-11-09 16:11, Ashutosh Bapat wrote:\\n> > On Thu, Nov 9, 2023 at 12:03 PM torikoshia <torikoshia@oss.nttdata.com>\\n> > wrote:\\n> >> >>\\n> >> >> 1. When a backend is running nested queries, we will see the plan of\\n> >> >> the innermost query. That query may not be the actual culprit if the\\n> >> >> user query is running slowly. E.g a query being run as part of inner\\n> >> >> side nested loop when the nested loop itself is the bottleneck. I\\n> >> >> think it will be useful to print plans of all the whole query stack.\\n> >>\\n> >> This was discussed in previous threads[1] and we thought it'd be\\n> >> useful\\n> >> but since it needed some extra works, we stopped widening the scope.\\n> >>\\n> >> >\\n> >> > I think we can start with what auto_explain is doing. Always print the\\n> >> > plan of the outermost query; the query found in pg_stat_activity. In a\\n> >> > later version we might find a way to print plans of all the queries in\\n> >> > the stack and do so in a readable manner.\\n> >>\\n> >> Agreed there are cases printing plan of the outermost query is more\\n> >> useful.\\n> >>\\n> >\\n> > I am fine printing the plan of the outermost query. This will help\\n> > many cases. Printing plans of the whole query stack can be added as an\\n> > add on later.\\n> >\\n> >> >\\n> >> > This makes tracking activeQueryDesc a bit tricky. My guess is that the\\n> >> > outermost query's descriptor will be available through ActivePortal\\n> >> > most of the time. But there are cases when ExecutorRun is called by\\n> >> > passing a queryDesc directly. So may be there are some cases where\\n> >> > that's not true.\\n> >>\\n> >> Yeah, actually the original version of the patch got the plan from\\n> >> ActivePortal, but it failed logging plan when the query was something\\n> >> like this[2]:\\n> >>\\n> >> DO $$\\n> >> BEGIN\\n> >> PERFORM pg_sleep(100);\\n> >> END$$;\\n> >\\n> > References [1] and [2] are not listed in your email.\\n>\\n> Oops, sorry. Here are links:\\n>\\n> [1]\\n> https://www.postgresql.org/message-id/64f716c44629e303b66e6c24502147cc%40oss.nttdata.com\\n> [2]\\n> https://www.postgresql.org/message-id/632e99eb-8090-53e6-1b1a-101601904cbd%40oss.nttdata.com\\n>\\n> > Is that because there was no ActivePortal created or the ActivePortal\\n> > pointed to DO block instead of PERFORM pg_sleep?\\n>\\n> ActivePortal is created but ActivePortal->queryDesc is null.\\n\\nThanks.\\n\\n>\\n> >> > 2. When a query is running in parallel worker do we want to print that\\n> >> > query? It may or may not be interesting given the situation. If the\\n> >> > overall plan itself is faulty, output of the parallel worker query is\\n> >> > not useful. If the plan is fine but a given worker's leg is running\\n> >> > slowly it may be interesting.\\n> >>\\n> >> I think it can be useful.\\n> >> I'm wondering if we can add this after the first patch for this\\n> >> feature\\n> >> is committed.\\n> >\\n> > With the current patches, it will print the query from a parallel\\n> > backend. If that's not desirable we should prohibit that case at\\n> > least.\\n>\\n> Current patch prohibits printing plan if backend type is parallel worker\\n> as below:\\n>\\n> =# select pg_log_query_plan(pid), backend_type from pg_stat_activity\\n> where backend_type = 'parallel worker';\\n>\\n> pg_log_query_plan | backend_type\\n> -------------------+-----------------\\n> f | parallel worker\\n> f | parallel worker\\n> (2 rows)\\n>\\n> WARNING: PID 4618 is not a PostgreSQL client backend process\\n> WARNING: PID 4617 is not a PostgreSQL client backend process\\n>\\n> Is this the behavior you expect?\\n>\\n\\nI misread then. Thanks for correcting me. We could consider plans from\\nparallel workers in v2 of this feature.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Nov 2023 20:28:48 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hello hackers,\\n\\nLast Saturday I submitted a patch to the pgsql-hackers list with the title\\n\\\"Proposal: In-flight explain logging\\\" with a patch proposing a feature very\\nsimilar to the one being worked on in this thread. I should have done a\\nbetter\\nsearch in the commitfest before implementing something from scratch.\\n\\nSo, as recommended by Ashutosh, I am sending an incremental patch containing\\nan additional feature I personally think we should include: logging the plan\\nwith instrumentation details if enabled.\\n\\nWhen targeting a query with instrumentation PG should log the complete\\nEXPLAIN ANALYZE plan with current row count and, if enabled, timing for each\\nnode. This gives the user not only the ability to see what the plan is\\nbut also what was executed so far, which is super useful when\\ntroubleshooting queries that never finish.\\n\\nConsidering that the query is in progress the output will include the\\nstatement (never executed) for nodes that weren't touched yet (or may\\nnever be). This feature is already present in the current ExplainNode\\nimplementation.\\n\\nI added a new statement (in progress) for nodes currently being executed,\\nie,\\nInstrStartNode was called and clock is ticking there.\\n\\nBack-reading this thread I saw the discussion about extending\\npg_log_query_plan\\nto parallel workers or not. I added it in my incremental patch as the new\\ncapability of logging instrumentation makes it useful to be able to see\\nwhat parallel workers are currently doing.\\n\\n# DEMONSTRATION:\\n\\npostgres=# select pid, backend_type,pg_log_query_plan(pid)\\npostgres=# from pg_stat_activity\\npostgres=# where (backend_type = 'client backend' and pid !=\\npg_backend_pid())\\npostgres=# or backend_type = 'parallel worker';\\n pid | backend_type | pg_log_query_plan\\n-------+-----------------+-------------------\\n 33833 | client backend | t\\n 47202 | parallel worker | t\\n 47203 | parallel worker | t\\n(3 rows)\\n\\n2023-12-06 23:14:41.756 UTC [33833] LOG: query plan running on backend\\nwith PID 33833 is:\\nQuery Text: explain (analyze, buffers) select *\\nfrom t2 a\\ninner join t1 b on a.c1=b.c1\\ninner join t1 c on a.c1=c.c1\\ninner join t1 d on a.c1=d.c1\\ninner join t1 e on a.c1=e.c1;\\nGather (cost=70894.63..202643.27 rows=1000000 width=20) (never executed)\\n Output: a.c1, b.c1, c.c1, d.c1, e.c1\\n Workers Planned: 2\\n Workers Launched: 2\\n -> Parallel Hash Join (cost=69894.63..101643.27 rows=416667 width=20)\\n(never executed)\\n Output: a.c1, b.c1, c.c1, d.c1, e.c1\\n Hash Cond: (a.c1 = e.c1)\\n -> Parallel Hash Join (cost=54466.62..77218.65 rows=416667\\nwidth=16) (never executed)\\n Output: a.c1, b.c1, c.c1, d.c1\\n Hash Cond: (a.c1 = c.c1)\\n -> Parallel Hash Join (cost=15428.00..29997.42 rows=416667\\nwidth=8) (actual time=2500.914..2625.922 rows=250412 loops=1) (in progress)\\n Output: a.c1, b.c1\\n Inner Unique: true\\n Hash Cond: (b.c1 = a.c1)\\n Buffers: shared hit=755 read=2175, temp read=1473\\nwritten=1860\\n -> Parallel Seq Scan on public.t1 b\\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.022..20.904\\nrows=331492 loops=1)\\n Output: b.c1\\n Buffers: shared hit=602 read=865\\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667\\nwidth=4) (actual time=1745.107..1745.107 rows=330638 loops=1)\\n Output: a.c1\\n Buffers: shared hit=153 read=1310, temp written=868\\n -> Parallel Seq Scan on public.t2 a\\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.042..27.695\\nrows=330638 loops=1)\\n Output: a.c1\\n Buffers: shared hit=153 read=1310\\n -> Parallel Hash (cost=32202.28..32202.28 rows=416667\\nwidth=8) (actual time=2450.489..2450.489 rows=407941 loops=1)\\n Output: c.c1, d.c1\\n Buffers: shared hit=1141 read=1833, temp read=1938\\nwritten=2836\\n -> Parallel Hash Join (cost=15428.00..32202.28\\nrows=416667 width=8) (actual time=1323.422..1575.245 rows=407941 loops=1)\\n Output: c.c1, d.c1\\n Hash Cond: (c.c1 = d.c1)\\n Buffers: shared hit=1141 read=1833, temp read=1938\\nwritten=1864\\n -> Parallel Seq Scan on public.t1 c\\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.026..22.223\\nrows=336238 loops=1)\\n Output: c.c1\\n Buffers: shared hit=590 read=898\\n -> Parallel Hash (cost=8591.67..8591.67\\nrows=416667 width=4) (actual time=653.306..653.306 rows=335836 loops=1)\\n Output: d.c1\\n Buffers: shared hit=551 read=935, temp\\nwritten=872\\n -> Parallel Seq Scan on public.t1 d\\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.022..23.127\\nrows=335836 loops=1)\\n Output: d.c1\\n Buffers: shared hit=551 read=935\\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4)\\n(actual time=590.086..590.086 rows=343696 loops=1)\\n Output: e.c1\\n Buffers: shared hit=519 read=1002, temp written=896\\n -> Parallel Seq Scan on public.t1 e (cost=0.00..8591.67\\nrows=416667 width=4) (actual time=0.066..21.797 rows=343696 loops=1)\\n Output: e.c1\\n Buffers: shared hit=519 read=1002\\n2023-12-06 23:14:41.757 UTC [47203] LOG: query plan running on backend\\nwith PID 47203 is:\\nQuery Text: explain (analyze, buffers) select *\\nfrom t2 a\\ninner join t1 b on a.c1=b.c1\\ninner join t1 c on a.c1=c.c1\\ninner join t1 d on a.c1=d.c1\\ninner join t1 e on a.c1=e.c1;\\nParallel Hash Join (cost=69894.63..101643.27 rows=416667 width=20) (never\\nexecuted)\\n Output: a.c1, b.c1, c.c1, d.c1, e.c1\\n Hash Cond: (a.c1 = e.c1)\\n -> Parallel Hash Join (cost=54466.62..77218.65 rows=416667 width=16)\\n(never executed)\\n Output: a.c1, b.c1, c.c1, d.c1\\n Hash Cond: (a.c1 = c.c1)\\n -> Parallel Hash Join (cost=15428.00..29997.42 rows=416667\\nwidth=8) (actual time=2464.367..2628.476 rows=258293 loops=1)\\n Output: a.c1, b.c1\\n Inner Unique: true\\n Hash Cond: (b.c1 = a.c1)\\n Buffers: shared hit=704 read=2213, temp read=1497 written=1860\\n -> Parallel Seq Scan on public.t1 b (cost=0.00..8591.67\\nrows=416667 width=4) (actual time=0.032..22.247 rows=330412 loops=1)\\n Output: b.c1\\n Buffers: shared hit=594 read=868\\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4)\\n(actual time=1745.093..1745.093 rows=328830 loops=1)\\n Output: a.c1\\n Buffers: shared hit=110 read=1345, temp written=868\\n -> Parallel Seq Scan on public.t2 a\\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=1.141..30.128\\nrows=328830 loops=1)\\n Output: a.c1\\n Buffers: shared hit=110 read=1345\\n -> Parallel Hash (cost=32202.28..32202.28 rows=416667 width=8)\\n(actual time=2449.694..2449.694 rows=295462 loops=1)\\n Output: c.c1, d.c1\\n Buffers: shared hit=1161 read=1844, temp read=1971 written=2872\\n -> Parallel Hash Join (cost=15428.00..32202.28 rows=416667\\nwidth=8) (actual time=1378.678..1577.182 rows=295462 loops=1)\\n Output: c.c1, d.c1\\n Hash Cond: (c.c1 = d.c1)\\n Buffers: shared hit=1161 read=1844, temp read=1971\\nwritten=1856\\n -> Parallel Seq Scan on public.t1 c\\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.031..22.459\\nrows=336288 loops=1)\\n Output: c.c1\\n Buffers: shared hit=590 read=898\\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667\\nwidth=4) (actual time=651.048..651.048 rows=328378 loops=1)\\n Output: d.c1\\n Buffers: shared hit=507 read=946, temp written=860\\n -> Parallel Seq Scan on public.t1 d\\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.034..24.462\\nrows=328378 loops=1)\\n Output: d.c1\\n Buffers: shared hit=507 read=946\\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4) (actual\\ntime=575.417..575.417 rows=327926 loops=1)\\n Output: e.c1\\n Buffers: shared hit=522 read=929, temp written=860\\n -> Parallel Seq Scan on public.t1 e (cost=0.00..8591.67\\nrows=416667 width=4) (actual time=0.786..23.114 rows=327926 loops=1)\\n Output: e.c1\\n Buffers: shared hit=522 read=929\\n2023-12-06 23:14:41.758 UTC [47202] LOG: query plan running on backend\\nwith PID 47202 is:\\nQuery Text: explain (analyze, buffers) select *\\nfrom t2 a\\ninner join t1 b on a.c1=b.c1\\ninner join t1 c on a.c1=c.c1\\ninner join t1 d on a.c1=d.c1\\ninner join t1 e on a.c1=e.c1;\\nParallel Hash Join (cost=69894.63..101643.27 rows=416667 width=20) (never\\nexecuted)\\n Output: a.c1, b.c1, c.c1, d.c1, e.c1\\n Hash Cond: (a.c1 = e.c1)\\n -> Parallel Hash Join (cost=54466.62..77218.65 rows=416667 width=16)\\n(never executed)\\n Output: a.c1, b.c1, c.c1, d.c1\\n Hash Cond: (a.c1 = c.c1)\\n -> Parallel Hash Join (cost=15428.00..29997.42 rows=416667\\nwidth=8) (actual time=2500.138..2631.855 rows=254125 loops=1)\\n Output: a.c1, b.c1\\n Inner Unique: true\\n Hash Cond: (b.c1 = a.c1)\\n Buffers: shared hit=749 read=2254, temp read=1483 written=1896\\n -> Parallel Seq Scan on public.t1 b (cost=0.00..8591.67\\nrows=416667 width=4) (actual time=0.027..22.224 rows=338096 loops=1)\\n Output: b.c1\\n Buffers: shared hit=628 read=868\\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4)\\n(actual time=1744.616..1744.616 rows=340532 loops=1)\\n Output: a.c1\\n Buffers: shared hit=121 read=1386, temp written=892\\n -> Parallel Seq Scan on public.t2 a\\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=1.138..30.003\\nrows=340532 loops=1)\\n Output: a.c1\\n Buffers: shared hit=121 read=1386\\n -> Parallel Hash (cost=32202.28..32202.28 rows=416667 width=8)\\n(actual time=2450.470..2450.470 rows=296597 loops=1)\\n Output: c.c1, d.c1\\n Buffers: shared hit=1186 read=1813, temp read=1610 written=2872\\n -> Parallel Hash Join (cost=15428.00..32202.28 rows=416667\\nwidth=8) (actual time=1380.286..1561.412 rows=296597 loops=1)\\n Output: c.c1, d.c1\\n Hash Cond: (c.c1 = d.c1)\\n Buffers: shared hit=1186 read=1813, temp read=1610\\nwritten=1848\\n -> Parallel Seq Scan on public.t1 c\\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.032..22.170\\nrows=327474 loops=1)\\n Output: c.c1\\n Buffers: shared hit=548 read=901\\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667\\nwidth=4) (actual time=653.297..653.297 rows=335786 loops=1)\\n Output: d.c1\\n Buffers: shared hit=574 read=912, temp written=868\\n -> Parallel Seq Scan on public.t1 d\\n (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.023..24.210\\nrows=335786 loops=1)\\n Output: d.c1\\n Buffers: shared hit=574 read=912\\n -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4) (actual\\ntime=574.553..574.553 rows=328378 loops=1)\\n Output: e.c1\\n Buffers: shared hit=495 read=958, temp written=864\\n -> Parallel Seq Scan on public.t1 e (cost=0.00..8591.67\\nrows=416667 width=4) (actual time=0.717..23.270 rows=328378 loops=1)\\n Output: e.c1\\n Buffers: shared hit=495 read=958\\n\\nThe thread of my initial patch contains all important implementation\\ndetails. Sharing the ones relevant to my incremental patch here:\\n\\n- Safely printing signaled plans with instrumentation\\n\\nA plan string is built in function ExplainNode here (\\nhttps://github.com/postgres/postgres/blob/REL_16_STABLE/src/backend/commands/explain.c#L1178\\n)\\nwhich is called at the end of a query execution when EXPLAIN is used.\\nThat function performs logic using a PlanState (part of QueryDesc) of\\nthe running query and a ExplainState.\\n\\nThe main challenge there is that ExplainNode calls InstrEndLoop which\\nchanges values in Instrumentation. This is ok for a regular EXPLAIN\\nwhere the query is already complete but not ok for the new feature with\\nsignaled plan logging.\\n\\nSo the new code has custom logic to clone Instrumentation instance of\\nthe current node. The cloned object can be safely written.\\n\\nFunction InstrEndLoop has a safety rule here (\\nhttps://github.com/postgres/postgres/blob/REL_16_STABLE/src/backend/executor/instrument.c#L148\\n)\\nthat prevents adjusting instrumentation details in a running node. This\\nnever happens in the current code logic but with the new feature\\nit will happen very often.\\n\\nI didn't want to remove this safety rule as InstrEndLoop gets called in\\nother places too (even in auto_explain) so the solution was to keep\\nInstrEndLoop and have a new InstrEndLoopForce for the signaled\\nplan logging with instrumentation. Both InstrEndLoop and InstrEndLoopForce\\ncall a new internal InstrEndLoopInternal to avoid duplicating the code.\\n\\n- Memory management\\n\\nConsidering that pg_log_query_plan() creates its own memory context before\\ncalling ExplainAssembleLogOutput, the new logic that allocates memory\\nto clone instrumentation doesn't need to free anything.\\n\\n# FINAL CONSIDERATIONS\\n\\nLet me know what you think about this incremental patch. If you think it\\nis useful but needs adjustments I will be happy to change as needed.\\n\\nKind Regards,\\n\\nRafael Castro.\",\n \"msg_date\": \"Wed, 6 Dec 2023 20:33:56 -0300\",\n \"msg_from\": \"Rafael Thofehrn Castro <rafaelthca@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2023-12-07 08:33, Rafael Thofehrn Castro wrote:\\n> Hello hackers,\\n> \\n> Last Saturday I submitted a patch to the pgsql-hackers list with the\\n> title\\n> \\\"Proposal: In-flight explain logging\\\" with a patch proposing a feature\\n> very\\n> similar to the one being worked on in this thread. I should have done\\n> a better\\n> search in the commitfest before implementing something from scratch.\\n> \\n> So, as recommended by Ashutosh, I am sending an incremental patch\\n> containing\\n> an additional feature I personally think we should include: logging\\n> the plan\\n> with instrumentation details if enabled.\\n\\nThanks for the proposal and making the patch!\\n\\n> When targeting a query with instrumentation PG should log the complete\\n> EXPLAIN ANALYZE plan with current row count and, if enabled, timing\\n> for each\\n> node. This gives the user not only the ability to see what the plan is\\n> but also what was executed so far, which is super useful when\\n> troubleshooting queries that never finish.\\n\\n\\nI think showing the progress of the query execution would be useful.\\n\\nOTOH it seems to at least need some modifications around Instrumentation \\nas your patch.\\nAs a first step, I think it would better to minimize the scope and focus \\non the fundamental function.\\nFor the same reason, getting queries for parallel workers is also \\nprohibited in the current patch as discussed here[1].\\n\\n[1] \\nhttps://www.postgresql.org/message-id/c25ae6015be96a1964eddd964657660b%40oss.nttdata.com\\n\\nSo I think below steps would be better than pushing all the \\nfunctionalities to the 1st commit.\\n\\n- First, develop function to enable output of query \\nprogress(v34-0001-Add-function-to-log-the-plan-of-the-query.patch).\\n- Then enhance the function\\n - showing the progress of the query \\nexecution(v34-0002-Log-plan-along-with-instrumentation-details.patch), \\netc.\\n\\n> --https://www.postgresql.org/message-id/CAG0ozMp3g3drnkDa6RZxXO_OmnisL2sD9vBrmpu5fOBoYpC-3w%40mail.gmail.com\\n> - ExplainState customization\\n> \\n> A ExplainState is allocated and customized for the in-flight logging.\\n> Instrumentation related settings are enabled based on how the target\\n> query started, which is usually via EXPLAIN ANALYZE or with \\n> auto_explain.\\n\\nDoes this mean the progress can be got only when the target query was \\nrun with EXPLAIN ANALYZE or auto_explain.log_analyze?\\n\\nIf so, there might be limited situations we can get the progress since I \\nimagine EXPLAIN ANALYZE is used when user want to get the plan from the \\nbeginning and auto_explain.log_analyze can give negative impact on \\nperformance as described in the manual and there may not be many \\nenvironments which enable it.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Dec 2023 14:46:23 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nUpdated the patch to fix typos and move \\nProcessLogQueryPlanInterruptActive from errfinish() to AbortTransaction.\\n\\n\\nBTW since the thread is getting long, I list the some points of the \\ndiscussion so far:\\n\\n# Safety concern\\n## Catalog access inside CFI\\n- it seems safe if the CFI call is inside an existing valid \\ntransaction/query state[1]\\n\\n- We did some tests, for example calling ProcessLogQueryPlanInterrupt() \\nin every single CHECK_FOR_INTERRUPTS()[2]. This test passed on my env \\nbut was stucked on James's env, so I modified to exit \\nProcessLogQueryPlanInterrupt() when target process is inside of lock \\nacquisition code[3]\\n\\n## Risk of calling EXPLAIN code in CFI\\n- EXPLAIN is not a simple logic code, and there exists risk calling it \\nfrom CFI. For example, if there is a bug, we may find ourselves in a \\nsituation where we can't cancel the query\\n\\n- it's a trade-off that's worth making for the introspection benefits \\nthis patch would provide?[4]\\n\\n# Design\\n- Although some suggested it should be in auto_explain, current patch \\nintroduces this feature to the core[5]\\n\\n- When the target query is nested, only the most inner query's plan is \\nexplained. In future, all the nested queries' plans are expected to \\nexplained optionally like auto_explain.log_nested_statements[6]\\n\\n- When the target process is a parallel worker, the plan is not shown[6]\\n\\n- When the target query is nested and its subtransaction is aborted, \\npg_log_query_plan cannot log the parental query plan after the abort \\neven parental query is running[7]\\n\\n- The output corresponds to EXPLAIN with VERBOSE, COST, SETTINGS and \\nFORMAT text. It doesn't do ANALYZE or show the progress of the query \\nexecution. Future work proposed by Rafael Thofehrn Castro may realize \\nthis[8]\\n\\n- To prevent assertion error, this patch ensures no page lock is held by \\nchecking all the LocalLock entries before running explain code, but \\nthere is a discussion that ginInsertCleanup() should be modified[9]\\n\\n\\nIt may be not so difficult to improve some of restrictions in \\\"Design\\\", \\nbut I'd like to limit the scope of the 1st patch to make it simpler.\\n\\n\\n[1] \\nhttps://www.postgresql.org/message-id/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw%40mail.gmail.com\\n[2] \\nhttps://www.postgresql.org/message-id/CAAaqYe8LXVXQhYy3yT0QOHUymdM%3Duha0dJ0%3DBEPzVAx2nG1gsw%40mail.gmail.com\\n[3] \\nhttps://www.postgresql.org/message-id/0e0e7ca08dff077a625c27a5e0c2ef0a%40oss.nttdata.com\\n[4] \\nhttps://www.postgresql.org/message-id/CAAaqYe8LXVXQhYy3yT0QOHUymdM%3Duha0dJ0%3DBEPzVAx2nG1gsw%40mail.gmail.com\\n[5] \\nhttps://www.postgresql.org/message-id/CAAaqYe_1EuoTudAz8mr8-qtN5SoNtYRm4JM2J9CqeverpE3B2A%40mail.gmail.com\\n[6] \\nhttps://www.postgresql.org/message-id/CAExHW5sh4ahrJgmMAGfptWVmESt1JLKCNm148XVxTunRr%2B-6gA%40mail.gmail.com\\n[7] \\nhttps://www.postgresql.org/message-id/3d121ed5f81cef588bac836b43f5d1f9%40oss.nttdata.com\\n[8] \\nhttps://www.postgresql.org/message-id/c161b5e7e1888eb9c9eb182a7d9dcf89%40oss.nttdata.com\\n[9] \\nhttps://www.postgresql.org/message-id/20220201.172757.1480996662235658750.horikyota.ntt%40gmail.com\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\",\n \"msg_date\": \"Mon, 29 Jan 2024 22:02:29 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi Atsushi,\\n\\n\\nOn Mon, Jan 29, 2024 at 6:32 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> Hi,\\n>\\n> Updated the patch to fix typos and move\\n> ProcessLogQueryPlanInterruptActive from errfinish() to AbortTransaction.\\n>\\n>\\n> BTW since the thread is getting long, I list the some points of the\\n> discussion so far:\\n>\\n> # Safety concern\\n> ## Catalog access inside CFI\\n> - it seems safe if the CFI call is inside an existing valid\\n> transaction/query state[1]\\n>\\n> - We did some tests, for example calling ProcessLogQueryPlanInterrupt()\\n> in every single CHECK_FOR_INTERRUPTS()[2]. This test passed on my env\\n> but was stucked on James's env, so I modified to exit\\n> ProcessLogQueryPlanInterrupt() when target process is inside of lock\\n> acquisition code[3]\\n>\\n> ## Risk of calling EXPLAIN code in CFI\\n> - EXPLAIN is not a simple logic code, and there exists risk calling it\\n> from CFI. For example, if there is a bug, we may find ourselves in a\\n> situation where we can't cancel the query\\n>\\n> - it's a trade-off that's worth making for the introspection benefits\\n> this patch would provide?[4]\\n>\\n> # Design\\n> - Although some suggested it should be in auto_explain, current patch\\n> introduces this feature to the core[5]\\n>\\n> - When the target query is nested, only the most inner query's plan is\\n> explained. In future, all the nested queries' plans are expected to\\n> explained optionally like auto_explain.log_nested_statements[6]\\n>\\n> - When the target process is a parallel worker, the plan is not shown[6]\\n>\\n> - When the target query is nested and its subtransaction is aborted,\\n> pg_log_query_plan cannot log the parental query plan after the abort\\n> even parental query is running[7]\\n>\\n> - The output corresponds to EXPLAIN with VERBOSE, COST, SETTINGS and\\n> FORMAT text. It doesn't do ANALYZE or show the progress of the query\\n> execution. Future work proposed by Rafael Thofehrn Castro may realize\\n> this[8]\\n>\\n> - To prevent assertion error, this patch ensures no page lock is held by\\n> checking all the LocalLock entries before running explain code, but\\n> there is a discussion that ginInsertCleanup() should be modified[9]\\n>\\n>\\n> It may be not so difficult to improve some of restrictions in \\\"Design\\\",\\n> but I'd like to limit the scope of the 1st patch to make it simpler.\\n\\nThanks for the summary. It is helpful. I think patch is also getting better.\\n\\nI have a few questions and suggestions\\n1. Prologue of GetLockMethodLocalHash() mentions\\n * NOTE: When there are many entries in LockMethodLocalHash, calling this\\n * function and looking into all of them can lead to performance problems.\\n */\\nHow bad this performance could be. Let's assume that a query is taking\\ntime and pg_log_query_plan() is invoked to examine the plan of this\\nquery. Is it possible that the looping over all the locks itself takes\\na lot of time delaying the query execution further?\\n\\n2. What happens if auto_explain is enabled in the backend and\\npg_log_query_plan() is called on the same backend? Will they conflict?\\nI think we should add a test for the same.\\n\\nUsing injection point support we should be able to add tests for\\ntesting pg_log_query_plan behaviour when there are page locks held or\\nwhen auto_explain (with instrumentation) and pg_log_query_plan() work\\non the same query plan. Use injection point to make the backend\\nrunning query wait at a suitable point to delay its execution and fire\\npg_log_query_plan() from other backend. May be the same test could\\nexamine the server log file to see if the plan is indeed output to the\\nserver log file.\\n\\nGiven that the feature will be used when the things have already gone\\nwrong, it should not make things more serious. So more testing and\\nespecially automated would help.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Tue, 6 Feb 2024 16:21:58 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi Ashutosh,\\n\\nOn 2024-02-06 19:51, Ashutosh Bapat wrote:\\n\\n> Thanks for the summary. It is helpful. I think patch is also getting \\n> better.\\n> \\n> I have a few questions and suggestions\\n\\nThanks for your comments.\\n\\n> 1. Prologue of GetLockMethodLocalHash() mentions\\n> * NOTE: When there are many entries in LockMethodLocalHash, calling \\n> this\\n> * function and looking into all of them can lead to performance \\n> problems.\\n> */\\n> How bad this performance could be. Let's assume that a query is taking\\n> time and pg_log_query_plan() is invoked to examine the plan of this\\n> query. Is it possible that the looping over all the locks itself takes\\n> a lot of time delaying the query execution further?\\n\\nI think it depends on the number of local locks, but I've measured cpu \\ntime for this page lock check by adding below codes and \\nv27-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch[1], which \\ncalls ProcessLogQueryPlanInterrupt() in every CFI on my laptop just for \\nyour information:\\n\\n diff --git a/src/backend/commands/explain.c \\nb/src/backend/commands/explain.c\\n index 5f7d77d567..65b7cb4925 100644\\n --- a/src/backend/commands/explain.c\\n +++ b/src/backend/commands/explain.c\\n @@ -44,6 +44,8 @@\\n\\n +#include \\\"time.h\\\"\\n ...\\n @@ -5287,6 +5292,7 @@ ProcessLogQueryPlanInterrupt(void)\\n * we check all the LocalLock entries and when finding even \\none, give up\\n * logging the plan.\\n */\\n + start = clock();\\n hash_seq_init(&status, GetLockMethodLocalHash());\\n while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != \\nNULL)\\n {\\n if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\\n {\\n ereport(LOG_SERVER_ONLY,\\n errmsg(\\\"ignored request for logging query plan due \\nto page lock conflicts\\\"),\\n errdetail(\\\"You can try again in a moment.\\\"));\\n hash_seq_term(&status);\\n\\n ProcessLogQueryPlanInterruptActive = false;\\n return;\\n }\\n }\\n + end = clock();\\n + cpu_time_used = ((double) (end - start)) / CLOCKS_PER_SEC;\\n +\\n + ereport(LOG,\\n + errmsg(\\\"all locallock entry search took: %f\\\", \\ncpu_time_used));\\n +\\n\\nThere were about 3 million log lines which recorded the cpu time, and \\nthe duration was quite short:\\n\\n =# -- Extracted cpu_time_used from log and loaded it to cpu_time.d.\\n =# select max(d), min(d), avg(d) from cpu_time ;\\n max | min | avg\\n ----------+-----+-----------------------\\n 0.000116 | 0 | 4.706274625332238e-07\\n\\nI'm not certain that this is valid for actual use cases, but these \\nresults seem to suggest that it will not take that long.\\n\\n\\n> 2. What happens if auto_explain is enabled in the backend and\\n> pg_log_query_plan() is called on the same backend? Will they conflict?\\n> I think we should add a test for the same.\\n\\nHmm, I think they don't conflict since they just refer QueryDesc and \\ndon't modify it and don't use same objects for locking.\\n(I imagine 'conflict' here is something like 'hard conflict' in \\nreplication[2].)\\n\\nActually using both auto_explain and pg_log_query_plan() output each \\nlogs separately:\\n\\n (pid:62835)=# select pg_sleep(10);\\n (pid:70000)=# select pg_log_query_plan(62835);\\n\\n (pid:70000)=# \\\\! cat data/log/postgres.log\\n ...\\n 2024-02-06 21:44:17.837 JST [62835:4:0] LOG: 00000: query plan \\nrunning on backend with PID 62835 is:\\n Query Text: select pg_sleep(10);\\n Result (cost=0.00..0.01 rows=1 width=4)\\n Output: pg_sleep('10'::double precision)\\n Query Identifier: 3506829283127886044\\n 2024-02-06 21:44:17.837 JST [62835:5:0] LOCATION: \\nProcessLogQueryPlanInterrupt, explain.c:5336\\n 2024-02-06 21:44:26.974 JST [62835:6:0] LOG: 00000: duration: \\n10000.868 ms plan:\\n Query Text: select pg_sleep(10);\\n Result (cost=0.00..0.01 rows=1 width=4) (actual \\ntime=10000.802..10000.804 rows=1 loops=1)\\n\\n> Using injection point support we should be able to add tests for\\n> testing pg_log_query_plan behaviour when there are page locks held or\\n> when auto_explain (with instrumentation) and pg_log_query_plan() work\\n> on the same query plan. Use injection point to make the backend\\n> running query wait at a suitable point to delay its execution and fire\\n> pg_log_query_plan() from other backend. May be the same test could\\n> examine the server log file to see if the plan is indeed output to the\\n> server log file.\\n> \\n> Given that the feature will be used when the things have already gone\\n> wrong, it should not make things more serious. So more testing and\\n> especially automated would help.\\n\\nThanks for the advice, it seems a good idea.\\nI'm going to try to add tests using injection point.\\n\\n\\n[1] \\nhttps://www.postgresql.org/message-id/CAAaqYe8LXVXQhYy3yT0QOHUymdM%3Duha0dJ0%3DBEPzVAx2nG1gsw%40mail.gmail.com\\n[2] \\nhttps://www.postgresql.org/docs/devel/hot-standby.html#HOT-STANDBY-CONFLICT\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Wed, 07 Feb 2024 13:08:24 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 7, 2024 at 9:38 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> Hi Ashutosh,\\n>\\n> On 2024-02-06 19:51, Ashutosh Bapat wrote:\\n>\\n> > Thanks for the summary. It is helpful. I think patch is also getting\\n> > better.\\n> >\\n> > I have a few questions and suggestions\\n>\\n> Thanks for your comments.\\n>\\n> > 1. Prologue of GetLockMethodLocalHash() mentions\\n> > * NOTE: When there are many entries in LockMethodLocalHash, calling\\n> > this\\n> > * function and looking into all of them can lead to performance\\n> > problems.\\n> > */\\n> > How bad this performance could be. Let's assume that a query is taking\\n> > time and pg_log_query_plan() is invoked to examine the plan of this\\n> > query. Is it possible that the looping over all the locks itself takes\\n> > a lot of time delaying the query execution further?\\n>\\n> I think it depends on the number of local locks, but I've measured cpu\\n> time for this page lock check by adding below codes and\\n> v27-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch[1], which\\n> calls ProcessLogQueryPlanInterrupt() in every CFI on my laptop just for\\n> your information:\\n>\\n> diff --git a/src/backend/commands/explain.c\\n> b/src/backend/commands/explain.c\\n> index 5f7d77d567..65b7cb4925 100644\\n> --- a/src/backend/commands/explain.c\\n> +++ b/src/backend/commands/explain.c\\n> @@ -44,6 +44,8 @@\\n>\\n> +#include \\\"time.h\\\"\\n> ...\\n> @@ -5287,6 +5292,7 @@ ProcessLogQueryPlanInterrupt(void)\\n> * we check all the LocalLock entries and when finding even\\n> one, give up\\n> * logging the plan.\\n> */\\n> + start = clock();\\n> hash_seq_init(&status, GetLockMethodLocalHash());\\n> while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) !=\\n> NULL)\\n> {\\n> if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\\n> {\\n> ereport(LOG_SERVER_ONLY,\\n> errmsg(\\\"ignored request for logging query plan due\\n> to page lock conflicts\\\"),\\n> errdetail(\\\"You can try again in a moment.\\\"));\\n> hash_seq_term(&status);\\n>\\n> ProcessLogQueryPlanInterruptActive = false;\\n> return;\\n> }\\n> }\\n> + end = clock();\\n> + cpu_time_used = ((double) (end - start)) / CLOCKS_PER_SEC;\\n> +\\n> + ereport(LOG,\\n> + errmsg(\\\"all locallock entry search took: %f\\\",\\n> cpu_time_used));\\n> +\\n>\\n> There were about 3 million log lines which recorded the cpu time, and\\n> the duration was quite short:\\n>\\n> =# -- Extracted cpu_time_used from log and loaded it to cpu_time.d.\\n> =# select max(d), min(d), avg(d) from cpu_time ;\\n> max | min | avg\\n> ----------+-----+-----------------------\\n> 0.000116 | 0 | 4.706274625332238e-07\\n>\\n> I'm not certain that this is valid for actual use cases, but these\\n> results seem to suggest that it will not take that long.\\n\\nWhat load did you run? I don't think any query in make check would\\ntake say thousands of locks. The prologue refers to a very populated\\nlock hash table. I think that will happen if thousands of tables are\\nqueried in a single query OR a query runs on a partitioned table with\\nthousands of partitions. May be we want to try that scenario.\\n\\n>\\n>\\n> > 2. What happens if auto_explain is enabled in the backend and\\n> > pg_log_query_plan() is called on the same backend? Will they conflict?\\n> > I think we should add a test for the same.\\n>\\n> Hmm, I think they don't conflict since they just refer QueryDesc and\\n> don't modify it and don't use same objects for locking.\\n> (I imagine 'conflict' here is something like 'hard conflict' in\\n> replication[2].)\\n\\nBy conflict, I mean the two features behave weird when used together\\ne.g give wrong results or crash etc.\\n\\n>\\n> Actually using both auto_explain and pg_log_query_plan() output each\\n> logs separately:\\n>\\n> (pid:62835)=# select pg_sleep(10);\\n> (pid:70000)=# select pg_log_query_plan(62835);\\n>\\n> (pid:70000)=# \\\\! cat data/log/postgres.log\\n> ...\\n> 2024-02-06 21:44:17.837 JST [62835:4:0] LOG: 00000: query plan\\n> running on backend with PID 62835 is:\\n> Query Text: select pg_sleep(10);\\n> Result (cost=0.00..0.01 rows=1 width=4)\\n> Output: pg_sleep('10'::double precision)\\n> Query Identifier: 3506829283127886044\\n> 2024-02-06 21:44:17.837 JST [62835:5:0] LOCATION:\\n> ProcessLogQueryPlanInterrupt, explain.c:5336\\n> 2024-02-06 21:44:26.974 JST [62835:6:0] LOG: 00000: duration:\\n> 10000.868 ms plan:\\n> Query Text: select pg_sleep(10);\\n> Result (cost=0.00..0.01 rows=1 width=4) (actual\\n> time=10000.802..10000.804 rows=1 loops=1)\\n>\\n> > Using injection point support we should be able to add tests for\\n> > testing pg_log_query_plan behaviour when there are page locks held or\\n> > when auto_explain (with instrumentation) and pg_log_query_plan() work\\n> > on the same query plan. Use injection point to make the backend\\n> > running query wait at a suitable point to delay its execution and fire\\n> > pg_log_query_plan() from other backend. May be the same test could\\n> > examine the server log file to see if the plan is indeed output to the\\n> > server log file.\\n> >\\n> > Given that the feature will be used when the things have already gone\\n> > wrong, it should not make things more serious. So more testing and\\n> > especially automated would help.\\n>\\n> Thanks for the advice, it seems a good idea.\\n> I'm going to try to add tests using injection point.\\n\\nYour test with pg_sleep() is a good basic test. But more involved\\ntesting might need something like injection points.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Wed, 7 Feb 2024 10:28:18 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2024-02-07 13:58, Ashutosh Bapat wrote:\\n> On Wed, Feb 7, 2024 at 9:38 AM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> \\n>> Hi Ashutosh,\\n>> \\n>> On 2024-02-06 19:51, Ashutosh Bapat wrote:\\n>> \\n>> > Thanks for the summary. It is helpful. I think patch is also getting\\n>> > better.\\n>> >\\n>> > I have a few questions and suggestions\\n>> \\n>> Thanks for your comments.\\n>> \\n>> > 1. Prologue of GetLockMethodLocalHash() mentions\\n>> > * NOTE: When there are many entries in LockMethodLocalHash, calling\\n>> > this\\n>> > * function and looking into all of them can lead to performance\\n>> > problems.\\n>> > */\\n>> > How bad this performance could be. Let's assume that a query is taking\\n>> > time and pg_log_query_plan() is invoked to examine the plan of this\\n>> > query. Is it possible that the looping over all the locks itself takes\\n>> > a lot of time delaying the query execution further?\\n>> \\n>> I think it depends on the number of local locks, but I've measured cpu\\n>> time for this page lock check by adding below codes and\\n>> v27-0002-Testing-attempt-logging-plan-on-ever-CFI-call.patch[1], which\\n>> calls ProcessLogQueryPlanInterrupt() in every CFI on my laptop just \\n>> for\\n>> your information:\\n>> \\n>> diff --git a/src/backend/commands/explain.c\\n>> b/src/backend/commands/explain.c\\n>> index 5f7d77d567..65b7cb4925 100644\\n>> --- a/src/backend/commands/explain.c\\n>> +++ b/src/backend/commands/explain.c\\n>> @@ -44,6 +44,8 @@\\n>> \\n>> +#include \\\"time.h\\\"\\n>> ...\\n>> @@ -5287,6 +5292,7 @@ ProcessLogQueryPlanInterrupt(void)\\n>> * we check all the LocalLock entries and when finding even\\n>> one, give up\\n>> * logging the plan.\\n>> */\\n>> + start = clock();\\n>> hash_seq_init(&status, GetLockMethodLocalHash());\\n>> while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) \\n>> !=\\n>> NULL)\\n>> {\\n>> if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\\n>> {\\n>> ereport(LOG_SERVER_ONLY,\\n>> errmsg(\\\"ignored request for logging query plan \\n>> due\\n>> to page lock conflicts\\\"),\\n>> errdetail(\\\"You can try again in a moment.\\\"));\\n>> hash_seq_term(&status);\\n>> \\n>> ProcessLogQueryPlanInterruptActive = false;\\n>> return;\\n>> }\\n>> }\\n>> + end = clock();\\n>> + cpu_time_used = ((double) (end - start)) / CLOCKS_PER_SEC;\\n>> +\\n>> + ereport(LOG,\\n>> + errmsg(\\\"all locallock entry search took: %f\\\",\\n>> cpu_time_used));\\n>> +\\n>> \\n>> There were about 3 million log lines which recorded the cpu time, and\\n>> the duration was quite short:\\n>> \\n>> =# -- Extracted cpu_time_used from log and loaded it to cpu_time.d.\\n>> =# select max(d), min(d), avg(d) from cpu_time ;\\n>> max | min | avg\\n>> ----------+-----+-----------------------\\n>> 0.000116 | 0 | 4.706274625332238e-07\\n>> \\n>> I'm not certain that this is valid for actual use cases, but these\\n>> results seem to suggest that it will not take that long.\\n> \\n> What load did you run? I don't think any query in make check would\\n> take say thousands of locks.\\n\\nSorry, I forgot to write it but ran make check as you imagined.\\n\\n> The prologue refers to a very populated\\n> lock hash table. I think that will happen if thousands of tables are\\n> queried in a single query OR a query runs on a partitioned table with\\n> thousands of partitions. May be we want to try that scenario.\\n\\nOK, I'll try such cases.\\n\\n>> > 2. What happens if auto_explain is enabled in the backend and\\n>> > pg_log_query_plan() is called on the same backend? Will they conflict?\\n>> > I think we should add a test for the same.\\n>> \\n>> Hmm, I think they don't conflict since they just refer QueryDesc and\\n>> don't modify it and don't use same objects for locking.\\n>> (I imagine 'conflict' here is something like 'hard conflict' in\\n>> replication[2].)\\n> \\n> By conflict, I mean the two features behave weird when used together\\n> e.g give wrong results or crash etc.\\n> \\n>> \\n>> Actually using both auto_explain and pg_log_query_plan() output each\\n>> logs separately:\\n>> \\n>> (pid:62835)=# select pg_sleep(10);\\n>> (pid:70000)=# select pg_log_query_plan(62835);\\n>> \\n>> (pid:70000)=# \\\\! cat data/log/postgres.log\\n>> ...\\n>> 2024-02-06 21:44:17.837 JST [62835:4:0] LOG: 00000: query plan\\n>> running on backend with PID 62835 is:\\n>> Query Text: select pg_sleep(10);\\n>> Result (cost=0.00..0.01 rows=1 width=4)\\n>> Output: pg_sleep('10'::double precision)\\n>> Query Identifier: 3506829283127886044\\n>> 2024-02-06 21:44:17.837 JST [62835:5:0] LOCATION:\\n>> ProcessLogQueryPlanInterrupt, explain.c:5336\\n>> 2024-02-06 21:44:26.974 JST [62835:6:0] LOG: 00000: duration:\\n>> 10000.868 ms plan:\\n>> Query Text: select pg_sleep(10);\\n>> Result (cost=0.00..0.01 rows=1 width=4) (actual\\n>> time=10000.802..10000.804 rows=1 loops=1)\\n>> \\n>> > Using injection point support we should be able to add tests for\\n>> > testing pg_log_query_plan behaviour when there are page locks held or\\n>> > when auto_explain (with instrumentation) and pg_log_query_plan() work\\n>> > on the same query plan. Use injection point to make the backend\\n>> > running query wait at a suitable point to delay its execution and fire\\n>> > pg_log_query_plan() from other backend. May be the same test could\\n>> > examine the server log file to see if the plan is indeed output to the\\n>> > server log file.\\n>> >\\n>> > Given that the feature will be used when the things have already gone\\n>> > wrong, it should not make things more serious. So more testing and\\n>> > especially automated would help.\\n>> \\n>> Thanks for the advice, it seems a good idea.\\n>> I'm going to try to add tests using injection point.\\n> \\n> Your test with pg_sleep() is a good basic test. But more involved\\n> testing might need something like injection points.\\n\\nIt might be so, I will consider whether there are any subtle timing \\nissues, etc.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Wed, 07 Feb 2024 19:14:40 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 7, 2024 at 12:58 PM Ashutosh Bapat\\n<ashutosh.bapat.oss@gmail.com> wrote:\\n>\\n> >\\n> > > */\\n> > > How bad this performance could be. Let's assume that a query is taking\\n> > > time and pg_log_query_plan() is invoked to examine the plan of this\\n> > > query. Is it possible that the looping over all the locks itself takes\\n> > > a lot of time delaying the query execution further?\\n> >\\ncorner case test:\\npgbench --initialize --partition-method=range --partitions=20000\\nSomehow my setup, the pg_bench didn't populate the data but there are\\n20000 partitions there.\\n(all my other settings are default)\\n\\nsome interesting things happened when a query touch so many partitions like:\\nselect abalance, aid from public.pgbench_accounts,pg_sleep(4) where aid > 1;\\n\\nin another session, if you immediate call SELECT pg_log_query_plan(9482);\\nthen output be\\n`\\nLOG: backend with PID 9482 is not running a query or a subtransaction\\nis aborted\\n`\\nhowever if you delay a little bit of time (like 1 second), then\\nLOG will emit the plan with lots of text (not sure the plan is right).\\n\\nI think the reason is that the `InitPlan` within\\nstandard_ExecutorStart takes more time to finish\\nwhen your query touches a lot of partitions.\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Feb 2024 08:00:00 +0800\",\n \"msg_from\": \"jian he <jian.universality@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 29, 2024 at 9:02 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> Hi,\\n>\\n> Updated the patch to fix typos and move\\n> ProcessLogQueryPlanInterruptActive from errfinish() to AbortTransaction.\\n>\\n\\n+ <row>\\n+ <entry role=\\\"func_table_entry\\\"><para role=\\\"func_signature\\\">\\n+ <indexterm>\\n+ <primary>pg_log_query_plan</primary>\\n+ </indexterm>\\n+ <function>pg_log_query_plan</function> (\\n<parameter>pid</parameter> <type>integer</type> )\\n+ <returnvalue>boolean</returnvalue>\\n+ </para>\\n+ <para>\\n+ Requests to log the plan of the query currently running on the\\n+ backend with specified process ID.\\n+ It will be logged at <literal>LOG</literal> message level and\\n+ will appear in the server log based on the log\\n+ configuration set (See <xref linkend=\\\"runtime-config-logging\\\"/>\\n+ for more information), but will not be sent to the client\\n+ regardless of <xref linkend=\\\"guc-client-min-messages\\\"/>.\\n+ </para></entry>\\n+ </row>\\nit would be better to explain the meaning of return value TRUE/FALSE?\\n\\n+# logging plan of the running query on the specified backend\\n+{ oid => '8000', descr => 'log plan of the running query on the\\nspecified backend',\\n+ proname => 'pg_log_query_plan',\\n+ provolatile => 'v', prorettype => 'bool',\\nyou can add\\n`proargnames => '{pid}'`\\n\\n+ if (proc == NULL)\\n+ {\\n+ /*\\n+ * This is just a warning so a loop-through-resultset will not abort\\n+ * if one backend terminated on its own during the run.\\n+ */\\n+ ereport(WARNING,\\n+ (errmsg(\\\"PID %d is not a PostgreSQL backend process\\\", pid)));\\n+ PG_RETURN_BOOL(false);\\n+ }\\n+\\n+ be_status = pgstat_get_beentry_by_backend_id(proc->backendId);\\n+ if (be_status->st_backendType != B_BACKEND)\\n+ {\\n+ ereport(WARNING,\\n+ (errmsg(\\\"PID %d is not a PostgreSQL client backend process\\\", pid)));\\n+ PG_RETURN_BOOL(false);\\n+ }\\n+\\n+ if (SendProcSignal(pid, PROCSIG_LOG_QUERY_PLAN, proc->backendId) < 0)\\n+ {\\n+ /* Again, just a warning to allow loops */\\n+ ereport(WARNING,\\n+ (errmsg(\\\"could not send signal to process %d: %m\\\", pid)));\\n+ PG_RETURN_BOOL(false);\\n+ }\\n\\nI found out that pg_log_query_plan's comments look like\\npg_log_backend_memory_contexts.\\npg_log_backend_memory_contexts will iterate through many memory contexts.\\nbut pg_log_query_plan for one specific pid will only output one plan?\\nso I am a little bit confused by the comments.\\n\\n+ /*\\n+ * Ensure no page lock is held on this process.\\n+ *\\n+ * If page lock is held at the time of the interrupt, we can't acquire any\\n+ * other heavyweight lock, which might be necessary for explaining the plan\\n+ * when retrieving column names.\\n+ *\\n+ * This may be overkill, but since page locks are held for a short duration\\n+ * we check all the LocalLock entries and when finding even one, give up\\n+ * logging the plan.\\n+ */\\n+ hash_seq_init(&status, GetLockMethodLocalHash());\\n+ while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)\\n+ {\\n+ if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\\nmaybe not that self evident, the above comments still not explained\\nwhy we need to ensure only\\nPAGE lock was held on this process?\\n\\nIn the commit message, can you add all the discussion links?\\nMy gmail account doesn't have a previous discussion history.\\nI am not sure this\\n(https://www.postgresql.org/message-id/flat/d68c3ae31672664876b22d2dcbb526d2%40postgrespro.ru)\\nis the only discussion link?\\n\\nI found a bug:\\nsrc8=# select *, pg_sleep(10) from tenk1 for update;\\n2024-02-11 15:54:17.944 CST [48602] LOG: query plan running on\\nbackend with PID 48602 is:\\n Query Text: select *, pg_sleep(10) from tenk1 for update;\\n LockRows (cost=0.00..570.00 rows=10000 width=254)\\n Output: unique1, unique2, two, four, ten, twenty, hundred,\\nthousand, twothousand, fivethous, tenthous, odd, even, stringu1,\\nstringu2, string4, (pg_sleep('10'::double precision)), ctid\\n -> Seq Scan on public.tenk1 (cost=0.00..470.00 rows=10000 width=254)\\n Output: unique1, unique2, two, four, ten, twenty,\\nhundred, thousand, twothousand, fivethous, tenthous, odd, even,\\nstringu1, stringu2, string4, pg_sleep('10'::double precision), ctid\\n\\nanother session (PID) executes `SELECT pg_log_query_plan(48602);` in\\nthe meantime.\\npg_log_query_plan returns true successfully, but PID 48602 was stuck.\\n\\nI have problem using git apply:\\nerror: patch failed: src/include/commands/explain.h:94\\nerror: src/include/commands/explain.h: patch does not apply\\n\\n`patch -p1 < /v35-0001-Add-function-to-log-the-plan-of-the-query.patch` works\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Feb 2024 08:00:00 +0800\",\n \"msg_from\": \"jian he <jian.universality@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 12, 2024 at 5:31 AM jian he <jian.universality@gmail.com> wrote:\\n>\\n> I found a bug:\\n> src8=# select *, pg_sleep(10) from tenk1 for update;\\n> 2024-02-11 15:54:17.944 CST [48602] LOG: query plan running on\\n> backend with PID 48602 is:\\n> Query Text: select *, pg_sleep(10) from tenk1 for update;\\n> LockRows (cost=0.00..570.00 rows=10000 width=254)\\n> Output: unique1, unique2, two, four, ten, twenty, hundred,\\n> thousand, twothousand, fivethous, tenthous, odd, even, stringu1,\\n> stringu2, string4, (pg_sleep('10'::double precision)), ctid\\n> -> Seq Scan on public.tenk1 (cost=0.00..470.00 rows=10000 width=254)\\n> Output: unique1, unique2, two, four, ten, twenty,\\n> hundred, thousand, twothousand, fivethous, tenthous, odd, even,\\n> stringu1, stringu2, string4, pg_sleep('10'::double precision), ctid\\n>\\n> another session (PID) executes `SELECT pg_log_query_plan(48602);` in\\n> the meantime.\\n> pg_log_query_plan returns true successfully, but PID 48602 was stuck.\\n\\nWhat do you mean by PID 48602 was stuck?\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Feb 2024 10:09:18 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 12, 2024 at 5:31 AM jian he <jian.universality@gmail.com> wrote:\\n>\\n> On Wed, Feb 7, 2024 at 12:58 PM Ashutosh Bapat\\n> <ashutosh.bapat.oss@gmail.com> wrote:\\n> >\\n> > >\\n> > > > */\\n> > > > How bad this performance could be. Let's assume that a query is taking\\n> > > > time and pg_log_query_plan() is invoked to examine the plan of this\\n> > > > query. Is it possible that the looping over all the locks itself takes\\n> > > > a lot of time delaying the query execution further?\\n> > >\\n> corner case test:\\n> pgbench --initialize --partition-method=range --partitions=20000\\n> Somehow my setup, the pg_bench didn't populate the data but there are\\n> 20000 partitions there.\\n> (all my other settings are default)\\n>\\n> some interesting things happened when a query touch so many partitions like:\\n> select abalance, aid from public.pgbench_accounts,pg_sleep(4) where aid > 1;\\n>\\n> in another session, if you immediate call SELECT pg_log_query_plan(9482);\\n> then output be\\n> `\\n> LOG: backend with PID 9482 is not running a query or a subtransaction\\n> is aborted\\n> `\\n> however if you delay a little bit of time (like 1 second), then\\n> LOG will emit the plan with lots of text (not sure the plan is right).\\n>\\n> I think the reason is that the `InitPlan` within\\n> standard_ExecutorStart takes more time to finish\\n> when your query touches a lot of partitions.\\n\\nThat's probably expected unless we make the ActiveQueryDesc available\\nbefore ExecutorRun.\\n\\nHow much time did it took between issuing SELECT\\npg_log_query_plan(9482); and plan getting output to the server error\\nlogs? How does this time compare with say the same time difference for\\na simple query and how much of that time can be attributed to Lock\\ntable hash scan, if the difference between time difference is huge.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Feb 2024 10:12:21 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 12, 2024 at 12:42 PM Ashutosh Bapat\\n<ashutosh.bapat.oss@gmail.com> wrote:\\n>\\n> On Mon, Feb 12, 2024 at 5:31 AM jian he <jian.universality@gmail.com> wrote:\\n> >\\n> > On Wed, Feb 7, 2024 at 12:58 PM Ashutosh Bapat\\n> > <ashutosh.bapat.oss@gmail.com> wrote:\\n> > >\\n> > > >\\n> > > > > */\\n> > > > > How bad this performance could be. Let's assume that a query is taking\\n> > > > > time and pg_log_query_plan() is invoked to examine the plan of this\\n> > > > > query. Is it possible that the looping over all the locks itself takes\\n> > > > > a lot of time delaying the query execution further?\\n> > > >\\n> > corner case test:\\n> > pgbench --initialize --partition-method=range --partitions=20000\\n> > Somehow my setup, the pg_bench didn't populate the data but there are\\n> > 20000 partitions there.\\n> > (all my other settings are default)\\n> >\\n> > some interesting things happened when a query touch so many partitions like:\\n> > select abalance, aid from public.pgbench_accounts,pg_sleep(4) where aid > 1;\\n> >\\n> > in another session, if you immediate call SELECT pg_log_query_plan(9482);\\n> > then output be\\n> > `\\n> > LOG: backend with PID 9482 is not running a query or a subtransaction\\n> > is aborted\\n> > `\\n> > however if you delay a little bit of time (like 1 second), then\\n> > LOG will emit the plan with lots of text (not sure the plan is right).\\n> >\\n> > I think the reason is that the `InitPlan` within\\n> > standard_ExecutorStart takes more time to finish\\n> > when your query touches a lot of partitions.\\n>\\n> That's probably expected unless we make the ActiveQueryDesc available\\n> before ExecutorRun.\\n\\nso the error message should be something like:\\nerrmsg(\\\"backend with PID %d is not running a query or a subtransaction\\nis aborted or the plan is not generated\\\",\\n\\n>\\n> How much time did it took between issuing SELECT\\n> pg_log_query_plan(9482); and plan getting output to the server error\\n> logs?\\n\\nit either says errmsg(\\\"backend with PID %d is not running a query,.....)\\nor outputs the plan immediately, if i wait one or two seconds to call\\npg_log_query_plan.\\n\\nbecause of previously mentioned: with lots of partitions, initplan\\ntook longer to finish.\\nsetup: 2 sessions, one runs the query (select abalance, aid from\\npublic.pgbench_accounts,pg_sleep(4) where aid > 1;),\\nanother one calls pg_log_query_plan.\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Feb 2024 15:08:00 +0800\",\n \"msg_from\": \"jian he <jian.universality@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2024-02-07 19:14, torikoshia wrote:\\n> On 2024-02-07 13:58, Ashutosh Bapat wrote:\\n\\n>> The prologue refers to a very populated\\n>> lock hash table. I think that will happen if thousands of tables are\\n>> queried in a single query OR a query runs on a partitioned table with\\n>> thousands of partitions. May be we want to try that scenario.\\n> \\n> OK, I'll try such cases.\\n\\nI measured this using partitioned pgbench_accounts with some \\nmodification to v36[1].\\nThe results[2] show that CPU time increases in proportion to the number \\nof partitions, and the increase is not that large.\\n\\nHowever I've noticed that these ensuring no page lock logic would not be \\nnecessary anymore since cc32ec24fdf3b98 removed the assertion which \\ncaused an error[1].\\n\\n $ git show cc32ec24fdf3b98\\n ..\\n diff --git a/src/backend/storage/lmgr/lock.c \\nb/src/backend/storage/lmgr/lock.c\\n index 0a692ee0a6..f595bce31b 100644\\n --- a/src/backend/storage/lmgr/lock.c\\n +++ b/src/backend/storage/lmgr/lock.c\\n @@ -186,18 +186,6 @@ static int FastPathLocalUseCount = 0;\\n */\\n static bool IsRelationExtensionLockHeld PG_USED_FOR_ASSERTS_ONLY = \\nfalse;\\n\\n - /*\\n - * We don't acquire any other heavyweight lock while holding \\nthe page lock\\n - * except for relation extension.\\n - */\\n - Assert(!IsPageLockHeld ||\\n - (locktag->locktag_type == \\nLOCKTAG_RELATION_EXTEND));\\n\\nI'm going to remove ensuring no page lock logic after some testings.\\n\\n\\n[1]\\n$ git diff _submission/log_running_query-v36\\n+#include \\\"time.h\\\"\\n+\\n bool ProcessLogQueryPlanInterruptActive = false;\\n\\n /* Hook for plugins to get control in ExplainOneQuery() */\\n@@ -5258,6 +5260,10 @@ ProcessLogQueryPlanInterrupt(void)\\n MemoryContext old_cxt;\\n LogQueryPlanPending = false;\\n\\n+ clock_t start, end;\\n+ double cpu_time_used;\\n+ int num_hash_entry = 0;\\n+\\n /* Cannot re-enter. */\\n if (ProcessLogQueryPlanInterruptActive)\\n return;\\n@@ -5287,9 +5293,11 @@ ProcessLogQueryPlanInterrupt(void)\\n * we check all the LocalLock entries and when finding even one, \\ngive up\\n * logging the plan.\\n */\\n+ start = clock();\\n hash_seq_init(&status, GetLockMethodLocalHash());\\n while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)\\n {\\n+ num_hash_entry++;\\n if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\\n {\\n ereport(LOG_SERVER_ONLY,\\n@@ -5301,6 +5309,12 @@ ProcessLogQueryPlanInterrupt(void)\\n return;\\n }\\n }\\n+ end = clock();\\n+ cpu_time_used = ((double) (end - start)) / CLOCKS_PER_SEC;\\n+\\n+ ereport(LOG,\\n+ errmsg(\\\"locallock entry search took: %f for %d entries\\\", \\ncpu_time_used, num_hash_entry));\\n\\n[2]\\n# partition number: 512\\nlocallock entry search took: 0.000029 for 1026 entries\\nlocallock entry search took: 0.000030 for 1026 entries\\nlocallock entry search took: 0.000036 for 1026 entries\\n\\n# partition number: 1024\\nlocallock entry search took: 0.000070 for 2050 entries\\nlocallock entry search took: 0.000059 for 2050 entries\\nlocallock entry search took: 0.000049 for 2050 entries\\n\\n# partition number: 2048\\nlocallock entry search took: 0.000100 for 4098 entries\\nlocallock entry search took: 0.000103 for 4098 entries\\nlocallock entry search took: 0.000101 for 4098 entries\\n\\n# partition number: 4096\\nlocallock entry search took: 0.000197 for 8194 entries\\nlocallock entry search took: 0.000193 for 8194 entries\\nlocallock entry search took: 0.000192 for 8194 entries\\n\\n[3] \\nhttps://www.postgresql.org/message-id/0642712f-1298-960a-a3ba-e256d56040ac%40oss.nttdata.com\\n\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Tue, 13 Feb 2024 09:19:03 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2024-02-12 09:00, jian he wrote:\\n\\nThanks for you comments.\\n\\n> On Mon, Jan 29, 2024 at 9:02 PM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> \\n>> Hi,\\n>> \\n>> Updated the patch to fix typos and move\\n>> ProcessLogQueryPlanInterruptActive from errfinish() to \\n>> AbortTransaction.\\n>> \\n> \\n> + <row>\\n> + <entry role=\\\"func_table_entry\\\"><para role=\\\"func_signature\\\">\\n> + <indexterm>\\n> + <primary>pg_log_query_plan</primary>\\n> + </indexterm>\\n> + <function>pg_log_query_plan</function> (\\n> <parameter>pid</parameter> <type>integer</type> )\\n> + <returnvalue>boolean</returnvalue>\\n> + </para>\\n> + <para>\\n> + Requests to log the plan of the query currently running on the\\n> + backend with specified process ID.\\n> + It will be logged at <literal>LOG</literal> message level and\\n> + will appear in the server log based on the log\\n> + configuration set (See <xref \\n> linkend=\\\"runtime-config-logging\\\"/>\\n> + for more information), but will not be sent to the client\\n> + regardless of <xref linkend=\\\"guc-client-min-messages\\\"/>.\\n> + </para></entry>\\n> + </row>\\n> it would be better to explain the meaning of return value TRUE/FALSE?\\n\\nYeah, but I've noticed that this should be located in 'Table Server \\nSignaling Functions' not 'Table Control Data Functions'.\\nSince 'Table Server Signaling Functions' describes the return code as \\nbelow, just relocation seems fine.\\n\\n Each of these functions returns true if the signal was successfully \\nsent and false if sending the signal failed.\\n\\n> +# logging plan of the running query on the specified backend\\n> +{ oid => '8000', descr => 'log plan of the running query on the\\n> specified backend',\\n> + proname => 'pg_log_query_plan',\\n> + provolatile => 'v', prorettype => 'bool',\\n> you can add\\n> `proargnames => '{pid}'`\\n\\nHmm, pg_log_query_plan() can take one argument, I'm not sure how much \\nsense it makes.\\nOther functions which take one argument such as pg_cancel_backend() does \\nnot have proargnames.\\n\\n\\n> + if (proc == NULL)\\n> + {\\n> + /*\\n> + * This is just a warning so a loop-through-resultset will not abort\\n> + * if one backend terminated on its own during the run.\\n> + */\\n> + ereport(WARNING,\\n> + (errmsg(\\\"PID %d is not a PostgreSQL backend process\\\", pid)));\\n> + PG_RETURN_BOOL(false);\\n> + }\\n> +\\n> + be_status = pgstat_get_beentry_by_backend_id(proc->backendId);\\n> + if (be_status->st_backendType != B_BACKEND)\\n> + {\\n> + ereport(WARNING,\\n> + (errmsg(\\\"PID %d is not a PostgreSQL client backend process\\\", pid)));\\n> + PG_RETURN_BOOL(false);\\n> + }\\n> +\\n> + if (SendProcSignal(pid, PROCSIG_LOG_QUERY_PLAN, proc->backendId) < 0)\\n> + {\\n> + /* Again, just a warning to allow loops */\\n> + ereport(WARNING,\\n> + (errmsg(\\\"could not send signal to process %d: %m\\\", pid)));\\n> + PG_RETURN_BOOL(false);\\n> + }\\n> \\n> I found out that pg_log_query_plan's comments look like\\n> pg_log_backend_memory_contexts.\\n> pg_log_backend_memory_contexts will iterate through many memory \\n> contexts.\\n> but pg_log_query_plan for one specific pid will only output one plan?\\n> so I am a little bit confused by the comments.\\n\\nThese \\\"loop\\\" mean backend can run pg_log_query_plan() repeatedly even \\nwhen failing sending signals.\\npg_signal_backend() also have such comments.\\n\\n> + /*\\n> + * Ensure no page lock is held on this process.\\n> + *\\n> + * If page lock is held at the time of the interrupt, we can't acquire \\n> any\\n> + * other heavyweight lock, which might be necessary for explaining the \\n> plan\\n> + * when retrieving column names.\\n> + *\\n> + * This may be overkill, but since page locks are held for a short \\n> duration\\n> + * we check all the LocalLock entries and when finding even one, give \\n> up\\n> + * logging the plan.\\n> + */\\n> + hash_seq_init(&status, GetLockMethodLocalHash());\\n> + while ((locallock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)\\n> + {\\n> + if (LOCALLOCK_LOCKTAG(*locallock) == LOCKTAG_PAGE)\\n> maybe not that self evident, the above comments still not explained\\n> why we need to ensure only\\n> PAGE lock was held on this process?\\n\\nThis is for preventing assertion error and it seems not necessary \\nanymore as described in [1].\\nI'm going remove them.\\n\\n> \\n> In the commit message, can you add all the discussion links?\\n> My gmail account doesn't have a previous discussion history.\\n\\nSure.\\n\\n> I am not sure this\\n> (https://www.postgresql.org/message-id/flat/d68c3ae31672664876b22d2dcbb526d2%40postgrespro.ru)\\n> is the only discussion link?\\n\\nThis is the original one:\\nhttps://www.postgresql.org/message-id/cf8501bcd95ba4d727cbba886ba9eea8%40oss.nttdata.com\\n\\n> I found a bug:\\n> src8=# select *, pg_sleep(10) from tenk1 for update;\\n> 2024-02-11 15:54:17.944 CST [48602] LOG: query plan running on\\n> backend with PID 48602 is:\\n> Query Text: select *, pg_sleep(10) from tenk1 for update;\\n> LockRows (cost=0.00..570.00 rows=10000 width=254)\\n> Output: unique1, unique2, two, four, ten, twenty, hundred,\\n> thousand, twothousand, fivethous, tenthous, odd, even, stringu1,\\n> stringu2, string4, (pg_sleep('10'::double precision)), ctid\\n> -> Seq Scan on public.tenk1 (cost=0.00..470.00 rows=10000 \\n> width=254)\\n> Output: unique1, unique2, two, four, ten, twenty,\\n> hundred, thousand, twothousand, fivethous, tenthous, odd, even,\\n> stringu1, stringu2, string4, pg_sleep('10'::double precision), ctid\\n> \\n> another session (PID) executes `SELECT pg_log_query_plan(48602);` in\\n> the meantime.\\n> pg_log_query_plan returns true successfully, but PID 48602 was stuck.\\n\\nHmm, it's not simply sleeping, is it?\\nI'm concerned a bit this because estimated rows of tenk1 is 10000.\\nIf so, the query will take 10000 * 10 seconds.\\n\\n> I have problem using git apply:\\n> error: patch failed: src/include/commands/explain.h:94\\n> error: src/include/commands/explain.h: patch does not apply\\n> \\n> `patch -p1 < /v35-0001-Add-function-to-log-the-plan-of-the-query.patch` \\n> works\\n\\nI'll update the patch including other points such as removing ensuring \\nno page lock code.\\n\\n\\n[1] \\nhttps://www.postgresql.org/message-id/1b2b247530f3ff3afab4ddc2df222e8b%40oss.nttdata.com\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Tue, 13 Feb 2024 11:30:53 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2024-02-13 11:30, torikoshia wrote:\\n\\n> I'll update the patch including other points such as removing ensuring\\n> no page lock code.\\n\\nUpdated the patch.\\n\\n> Using injection point support we should be able to add tests for\\n> testing pg_log_query_plan behaviour when there are page locks held or\\n> when auto_explain (with instrumentation) and pg_log_query_plan() work\\n> on the same query plan. Use injection point to make the backend\\n> running query wait at a suitable point to delay its execution and fire\\n> pg_log_query_plan() from other backend. May be the same test could\\n> examine the server log file to see if the plan is indeed output to the\\n> server log file.\\n\\nAttached patch uses injection point as below:\\n\\n- There may be more points to inject, but added an injection point at \\nExecutorRun(), which seems to be the first interruption point where \\nplans can be reliably displayed.\\n- At injection point, it'd be possible to wait for some duration and \\nfire pg_log_plan_query() as you suggested. However, I'm not sure how \\nlong duration is appropriate considering the variety of testing \\nenvironments. Instead, attached patch calls \\nHandleLogQueryPlanInterrupt() directly and set InterruptPending.\\n- Tests both pg_log_plan_query() and auto_explain logs for their output, \\nand the logged plans are the same.\\n\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\",\n \"msg_date\": \"Wed, 14 Feb 2024 15:29:35 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nI've just been catching up on this thread.\\n\\n+ if (MyProc->heldLocks)\\n+ {\\n+ ereport(LOG_SERVER_ONLY,\\n+ errmsg(\\\"ignored request for logging query plan due to lock conflicts\\\"),\\n+ errdetail(\\\"You can try again in a moment.\\\"));\\n+ return;\\n+ }\\n\\nI don't like this for several reasons.\\n\\nFirst, I think it's not nice to have a request just get ignored. A\\nuser will expect that if we cannot immediately respond to some\\nrequest, we'll respond later at the first time that it's safe to do\\nso, rather than just ignoring it and telling them to retry.\\n\\nSecond, I don't think that the error message is very good. It talks\\nabout lock conflicts, but we don't know that there is any real\\nproblem. We know that, if we enter this block, the server is in the\\nmiddle of trying to acquire some lock, and we also know that we could\\nattempt to acquire a lock as part of generating the EXPLAIN output,\\nand maybe that's an issue. But that's not a lock conflict. That's a\\nre-entrancy problem. I don't know that we want to talk about\\nre-entrancy problems in an error message, and I don't really think\\nthis error message should exist in the first place, but if we're going\\nto error out in this case surely we shouldn't do so with an error\\nmessage that describes a problem we don't have.\\n\\nThird, I think that the re-entrancy problems with this patch may\\nextend well beyond lock acquisition. This is one example of how it can\\nbe unsafe to do something as complicated as EXPLAIN at any arbitrary\\nCHECK_FOR_INTERRUPTS(). It is not correct to say, as\\nhttp://postgr.es/m/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw@mail.gmail.com\\ndoes, that the problems with running EXPLAIN at an arbitrary point are\\nspecific to running this code in an aborted transaction. The existence\\nof this code shows that there is at least one hazard even if the\\ncurrent transaction is not aborted, and I see no analysis on this\\nthread indicating whether there are any more such hazards, or how we\\ncould go about finding them all.\\n\\nI think the issue is very general. We have lots of subsystems that\\nboth (a) use global variables and (b) contain CHECK_FOR_INTERRUPTS().\\nIf we process an interrupt while that code is in the middle of\\nmanipulating its global variables and then again re-enter that code,\\nbad things might happen. We could try to rule that out by analyzing\\nall such subsystems and all places where CHECK_FOR_INTERRUPTS() may\\nappear, but that's very difficult. Suppose we took the alternative\\napproach recommended by Andrey Lepikhov in\\nhttp://postgr.es/m/b1b110ae-61f6-4fd9-9b94-f967db9b53d4@app.fastmail.com\\nand instead set a flag that gets handled in some suitable place in the\\nexecutor code, like ExecProcNode(). If we did that, then we would know\\nthat we're not in the middle of a syscache lookup, a catcache lookup,\\na heavyweight lock acquisition, an ereport, or any other low-level\\nsubsystem call that might create problems for the EXPLAIN code.\\n\\nIn that design, the hack shown above would go away, and we could be\\nmuch more certain that we don't need any other similar hacks for other\\nsubsystems. The only downside is that we might experience a slightly\\nlonger delay before a requested EXPLAIN plan actually shows up, but\\nthat seems like a pretty small price to pay for being able to reason\\nabout the behavior of the system. I don't *think* there are any cases\\nwhere we run in the executor for a particularly long time without a\\nnew call to ExecProcNode().\\n\\nI think this case is a bit like vacuum_delay_point(). You might think\\nthat vacuum_delay_point() could be moved inside of\\nCHECK_FOR_INTERRUPTS(), but we've made the opposite decision: every\\nvacuum_delay_point() calls CHECK_FOR_INTERRUPTS() but not every\\nCHECK_FOR_INTERRUPTS() calls vacuum_delay_point(). That means that we\\ncan allow vacuum_delay_point() only at cases where we know it's safe,\\nrather than at every CHECK_FOR_INTERRUPTS(). I think that's a pretty\\nsmart decision, even for vacuum_delay_point(), and AFAICS the code\\nwe're proposing to run here does things that are substantially more\\ncomplicated than what vacuum_delay_point() does. That code just does a\\nbit of reading of shared memory, reports a wait event, and sleeps.\\nThat's really simple compared to code that could try to do relcache\\nbuilds, which can trigger syscache lookups, which can both trigger\\nheavyweight lock acquisition, lightweight lock acquisition, bringing\\npages into shared_buffers possibly through physical I/O, processing of\\ninvalidation messages, and a bunch of other stuff.\\n\\nIt's really hard for me to accept that the heavyweight lock problem\\nfor which the patch contains a workaround is the only one that exists.\\nI can't see any reason why that should be true.\\n\\n...Robert\\n\\n\\n\",\n \"msg_date\": \"Thu, 15 Feb 2024 14:42:11 +0530\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2024-02-15 14:42:11 +0530, Robert Haas wrote:\\n> I think the issue is very general. We have lots of subsystems that\\n> both (a) use global variables and (b) contain CHECK_FOR_INTERRUPTS().\\n> If we process an interrupt while that code is in the middle of\\n> manipulating its global variables and then again re-enter that code,\\n> bad things might happen. We could try to rule that out by analyzing\\n> all such subsystems and all places where CHECK_FOR_INTERRUPTS() may\\n> appear, but that's very difficult. Suppose we took the alternative\\n> approach recommended by Andrey Lepikhov in\\n> http://postgr.es/m/b1b110ae-61f6-4fd9-9b94-f967db9b53d4@app.fastmail.com\\n> and instead set a flag that gets handled in some suitable place in the\\n> executor code, like ExecProcNode(). If we did that, then we would know\\n> that we're not in the middle of a syscache lookup, a catcache lookup,\\n> a heavyweight lock acquisition, an ereport, or any other low-level\\n> subsystem call that might create problems for the EXPLAIN code.\\n> \\n> In that design, the hack shown above would go away, and we could be\\n> much more certain that we don't need any other similar hacks for other\\n> subsystems. The only downside is that we might experience a slightly\\n> longer delay before a requested EXPLAIN plan actually shows up, but\\n> that seems like a pretty small price to pay for being able to reason\\n> about the behavior of the system.\\n\\nI am very wary of adding overhead to ExecProcNode() - I'm quite sure that\\nadding code there would trigger visible overhead for query times.\\n\\nIf we went with something like tht approach, I think we'd have to do something\\nlike redirecting node->ExecProcNode to a wrapper, presumably from within a\\nCFI. That wrapper could then implement the explain support, without slowing\\ndown the normal execution path.\\n\\n\\n> I don't *think* there are any cases where we run in the executor for a\\n> particularly long time without a new call to ExecProcNode().\\n\\nI guess it depends on what you call a long time. A large sort, for example,\\ncould spend a fair amount of time inside tuplesort, similarly, a gather node\\nmight need to wait for a worker for a while etc.\\n\\n\\n> It's really hard for me to accept that the heavyweight lock problem\\n> for which the patch contains a workaround is the only one that exists.\\n> I can't see any reason why that should be true.\\n\\nI suspect you're right.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 15 Feb 2024 10:59:11 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 15, 2024 at 6:12 PM Robert Haas <robertmhaas@gmail.com> \\nwrote:\\n> Hi,\\n> \\n> I've just been catching up on this thread.\\n> \\n> + if (MyProc->heldLocks)\\n> + {\\n> + ereport(LOG_SERVER_ONLY,\\n> + errmsg(\\\"ignored request for logging query plan due to lock \\n> conflicts\\\"),\\n> + errdetail(\\\"You can try again in a moment.\\\"));\\n> + return;\\n> + }\\n> \\n> I don't like this for several reasons.\\n> \\n> First, I think it's not nice to have a request just get ignored. A\\n> user will expect that if we cannot immediately respond to some\\n> request, we'll respond later at the first time that it's safe to do\\n> so, rather than just ignoring it and telling them to retry.\\n> \\n> Second, I don't think that the error message is very good. It talks\\n> about lock conflicts, but we don't know that there is any real\\n> problem. We know that, if we enter this block, the server is in the\\n> middle of trying to acquire some lock, and we also know that we could\\n> attempt to acquire a lock as part of generating the EXPLAIN output,\\n> and maybe that's an issue. But that's not a lock conflict. That's a\\n> re-entrancy problem. I don't know that we want to talk about\\n> re-entrancy problems in an error message, and I don't really think\\n> this error message should exist in the first place, but if we're going\\n> to error out in this case surely we shouldn't do so with an error\\n> message that describes a problem we don't have.\\n> \\n> Third, I think that the re-entrancy problems with this patch may\\n> extend well beyond lock acquisition. This is one example of how it can\\n> be unsafe to do something as complicated as EXPLAIN at any arbitrary\\n> CHECK_FOR_INTERRUPTS(). It is not correct to say, as\\n> http://postgr.es/m/CAAaqYe9euUZD8bkjXTVcD9e4n5c7kzHzcvuCJXt-xds9X4c7Fw@mail.gmail.com\\n> does, that the problems with running EXPLAIN at an arbitrary point are\\n> specific to running this code in an aborted transaction. The existence\\n> of this code shows that there is at least one hazard even if the\\n> current transaction is not aborted, and I see no analysis on this\\n> thread indicating whether there are any more such hazards, or how we\\n> could go about finding them all.\\n> \\n> I think the issue is very general. We have lots of subsystems that\\n> both (a) use global variables and (b) contain CHECK_FOR_INTERRUPTS().\\n> If we process an interrupt while that code is in the middle of\\n> manipulating its global variables and then again re-enter that code,\\n> bad things might happen. We could try to rule that out by analyzing\\n> all such subsystems and all places where CHECK_FOR_INTERRUPTS() may\\n> appear, but that's very difficult. Suppose we took the alternative\\n> approach recommended by Andrey Lepikhov in\\n> http://postgr.es/m/b1b110ae-61f6-4fd9-9b94-f967db9b53d4@app.fastmail.com\\n> and instead set a flag that gets handled in some suitable place in the\\n> executor code, like ExecProcNode(). If we did that, then we would know\\n> that we're not in the middle of a syscache lookup, a catcache lookup,\\n> a heavyweight lock acquisition, an ereport, or any other low-level\\n> subsystem call that might create problems for the EXPLAIN code.\\n> \\n> In that design, the hack shown above would go away, and we could be\\n> much more certain that we don't need any other similar hacks for other\\n> subsystems. The only downside is that we might experience a slightly\\n> longer delay before a requested EXPLAIN plan actually shows up, but\\n> that seems like a pretty small price to pay for being able to reason\\n> about the behavior of the system. I don't *think* there are any cases\\n> where we run in the executor for a particularly long time without a\\n> new call to ExecProcNode().\\n> \\n> I think this case is a bit like vacuum_delay_point(). You might think\\n> that vacuum_delay_point() could be moved inside of\\n> CHECK_FOR_INTERRUPTS(), but we've made the opposite decision: every\\n> vacuum_delay_point() calls CHECK_FOR_INTERRUPTS() but not every\\n> CHECK_FOR_INTERRUPTS() calls vacuum_delay_point(). That means that we\\n> can allow vacuum_delay_point() only at cases where we know it's safe,\\n> rather than at every CHECK_FOR_INTERRUPTS(). I think that's a pretty\\n> smart decision, even for vacuum_delay_point(), and AFAICS the code\\n> we're proposing to run here does things that are substantially more\\n> complicated than what vacuum_delay_point() does. That code just does a\\n> bit of reading of shared memory, reports a wait event, and sleeps.\\n> That's really simple compared to code that could try to do relcache\\n> builds, which can trigger syscache lookups, which can both trigger\\n> heavyweight lock acquisition, lightweight lock acquisition, bringing\\n> pages into shared_buffers possibly through physical I/O, processing of\\n> invalidation messages, and a bunch of other stuff.\\n> \\n> It's really hard for me to accept that the heavyweight lock problem\\n> for which the patch contains a workaround is the only one that exists.\\n> I can't see any reason why that should be true.\\n\\nThanks for the review and the very detailed explanation!\\n\\nI'm convinced that it's unsafe to execute EXPLAIN codes during \\nCHECK_FOR_INTERRUPTS() and we need to execute it in other safe place, as \\nwell as the first and second point.\\n\\n\\nOn 2024-02-16 03:59, Andres Freund wrote:\\n> Hi,\\n> \\n> On 2024-02-15 14:42:11 +0530, Robert Haas wrote:\\n>> I think the issue is very general. We have lots of subsystems that\\n>> both (a) use global variables and (b) contain CHECK_FOR_INTERRUPTS().\\n>> If we process an interrupt while that code is in the middle of\\n>> manipulating its global variables and then again re-enter that code,\\n>> bad things might happen. We could try to rule that out by analyzing\\n>> all such subsystems and all places where CHECK_FOR_INTERRUPTS() may\\n>> appear, but that's very difficult. Suppose we took the alternative\\n>> approach recommended by Andrey Lepikhov in\\n>> http://postgr.es/m/b1b110ae-61f6-4fd9-9b94-f967db9b53d4@app.fastmail.com\\n>> and instead set a flag that gets handled in some suitable place in the\\n>> executor code, like ExecProcNode(). If we did that, then we would know\\n>> that we're not in the middle of a syscache lookup, a catcache lookup,\\n>> a heavyweight lock acquisition, an ereport, or any other low-level\\n>> subsystem call that might create problems for the EXPLAIN code.\\n>> \\n>> In that design, the hack shown above would go away, and we could be\\n>> much more certain that we don't need any other similar hacks for other\\n>> subsystems. The only downside is that we might experience a slightly\\n>> longer delay before a requested EXPLAIN plan actually shows up, but\\n>> that seems like a pretty small price to pay for being able to reason\\n>> about the behavior of the system.\\n> \\n> I am very wary of adding overhead to ExecProcNode() - I'm quite sure \\n> that\\n> adding code there would trigger visible overhead for query times.\\n> \\n> If we went with something like tht approach, I think we'd have to do \\n> something\\n> like redirecting node->ExecProcNode to a wrapper, presumably from \\n> within a\\n> CFI. That wrapper could then implement the explain support, without \\n> slowing\\n> down the normal execution path.\\n\\nThanks for the idea!\\nI'm not so sure about the implementation now, i.e. finding the next node \\nto be executed from the planstate tree, but I'm going to try this \\napproach.\\n\\n>> I don't *think* there are any cases where we run in the executor for a\\n>> particularly long time without a new call to ExecProcNode().\\n> \\n> I guess it depends on what you call a long time. A large sort, for \\n> example,\\n> could spend a fair amount of time inside tuplesort, similarly, a gather \\n> node\\n> might need to wait for a worker for a while etc.\\n> \\n> \\n>> It's really hard for me to accept that the heavyweight lock problem\\n>> for which the patch contains a workaround is the only one that exists.\\n>> I can't see any reason why that should be true.\\n> \\n> I suspect you're right.\\n> \\n> Greetings,\\n> \\n> Andres Freund\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Fri, 16 Feb 2024 23:42:36 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 16, 2024 at 12:29 AM Andres Freund <andres@anarazel.de> wrote:\\n> If we went with something like tht approach, I think we'd have to do something\\n> like redirecting node->ExecProcNode to a wrapper, presumably from within a\\n> CFI. That wrapper could then implement the explain support, without slowing\\n> down the normal execution path.\\n\\nThat's an annoying complication; maybe there's some better way to\\nhandle this. But I think we need to do something different than what\\nthe patch does currently because...\\n\\n> > It's really hard for me to accept that the heavyweight lock problem\\n> > for which the patch contains a workaround is the only one that exists.\\n> > I can't see any reason why that should be true.\\n>\\n> I suspect you're right.\\n\\n...I think the current approach is just plain dead, because of this\\nissue. We can't take an approach that creates an unbounded number of\\nunclear reentrancy issues and then insert hacks one by one to cure\\nthem (or hack around them, more to the point) as they're discovered.\\n\\nThe premise has to be that we only allow logging the query plan at\\npoints where we know it's safe, rather than, as at present, allowing\\nit in places that are unsafe and then trying to compensate with code\\nelsewhere. That's not likely to ever be as stable as we want\\nPostgreSQL to be.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 20 Feb 2024 10:23:04 +0530\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 19, 2024 at 11:53 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> On Fri, Feb 16, 2024 at 12:29 AM Andres Freund <andres@anarazel.de> wrote:\\n> > If we went with something like tht approach, I think we'd have to do something\\n> > like redirecting node->ExecProcNode to a wrapper, presumably from within a\\n> > CFI. That wrapper could then implement the explain support, without slowing\\n> > down the normal execution path.\\n>\\n> That's an annoying complication; maybe there's some better way to\\n> handle this. But I think we need to do something different than what\\n> the patch does currently because...\\n>\\n> > > It's really hard for me to accept that the heavyweight lock problem\\n> > > for which the patch contains a workaround is the only one that exists.\\n> > > I can't see any reason why that should be true.\\n> >\\n> > I suspect you're right.\\n>\\n> ...I think the current approach is just plain dead, because of this\\n> issue. We can't take an approach that creates an unbounded number of\\n> unclear reentrancy issues and then insert hacks one by one to cure\\n> them (or hack around them, more to the point) as they're discovered.\\n>\\n> The premise has to be that we only allow logging the query plan at\\n> points where we know it's safe, rather than, as at present, allowing\\n> it in places that are unsafe and then trying to compensate with code\\n> elsewhere. That's not likely to ever be as stable as we want\\n> PostgreSQL to be.\\n\\nThis is potentially a bit of a wild idea, but I wonder if having some\\nkind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\\n\\\"normal\\\" as opposed to \\\"critical\\\" (using that word differently than\\nthe existing critical sections) would be worth it.\\n\\nRegards,\\nJames Coleman\\n\\n\\n\",\n \"msg_date\": \"Wed, 21 Feb 2024 19:54:57 -0500\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\\n> >\\n> > ...I think the current approach is just plain dead, because of this\\n> > issue. We can't take an approach that creates an unbounded number of\\n> > unclear reentrancy issues and then insert hacks one by one to cure\\n> > them (or hack around them, more to the point) as they're discovered.\\n> >\\n> > The premise has to be that we only allow logging the query plan at\\n> > points where we know it's safe, rather than, as at present, allowing\\n> > it in places that are unsafe and then trying to compensate with code\\n> > elsewhere. That's not likely to ever be as stable as we want\\n> > PostgreSQL to be.\\n>\\n> This is potentially a bit of a wild idea, but I wonder if having some\\n> kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\\n> \\\"normal\\\" as opposed to \\\"critical\\\" (using that word differently than\\n> the existing critical sections) would be worth it.\\n\\nMy hunch is this will end up being a maintenance burden since every\\ncaller has to decide (carefully) whether the call is under normal\\ncondition or not. Developers will tend to take a safe approach and\\nflag calls as critical. But importantly, what's normal for one\\ninterrupt action may be critical for another and vice versa. Approach\\nwould be useful depending upon how easy it is to comprehend the\\ndefinition of \\\"normal\\\".\\n\\nIf a query executes for longer than a user defined threashold (session\\nlevel GUC? or same value as auto_explain parameter), the executor\\nproactively prepares an EXPLAIN output and keeps it handy in case\\nasked for. It can do so at a \\\"known\\\" safe place and time rather than\\nat any random time and location. Extra time spent in creating EXPLAIN\\noutput may not be noticeable in a long running query. The EXPLAIN\\noutput could be saved in pg_stat_activity or similar place. This will\\navoid signaling the backend.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Thu, 22 Feb 2024 09:36:06 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\\n> This is potentially a bit of a wild idea, but I wonder if having some\\n> kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\\n> \\\"normal\\\" as opposed to \\\"critical\\\" (using that word differently than\\n> the existing critical sections) would be worth it.\\n\\nIt's worth considering, but the definition of \\\"normal\\\" vs. \\\"critical\\\"\\nmight be hard to pin down. Or, we might end up with a definition that\\nis specific to this particular case and not generalizable to others.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Feb 2024 10:22:32 +0530\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Fri, Feb 23, 2024 at 10:22:32AM +0530, Robert Haas wrote:\\n> On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\\n> > This is potentially a bit of a wild idea, but I wonder if having some\\n> > kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\\n> > \\\"normal\\\" as opposed to \\\"critical\\\" (using that word differently than\\n> > the existing critical sections) would be worth it.\\n>\\n> It's worth considering, but the definition of \\\"normal\\\" vs. \\\"critical\\\"\\n> might be hard to pin down. Or, we might end up with a definition that\\n> is specific to this particular case and not generalizable to others.\\n\\nBut it doesn't have to be all or nothing right? I mean each call could say\\nwhat the situation is like in their context, like\\nCHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\\nslowly tag calls as needed, similarly to how we add already CFI based on users\\nreport.\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Feb 2024 22:20:51 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> On Fri, Feb 23, 2024 at 10:22:32AM +0530, Robert Haas wrote:\\n> > On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\\n> > > This is potentially a bit of a wild idea, but I wonder if having some\\n> > > kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\\n> > > \\\"normal\\\" as opposed to \\\"critical\\\" (using that word differently than\\n> > > the existing critical sections) would be worth it.\\n> >\\n> > It's worth considering, but the definition of \\\"normal\\\" vs. \\\"critical\\\"\\n> > might be hard to pin down. Or, we might end up with a definition that\\n> > is specific to this particular case and not generalizable to others.\\n>\\n> But it doesn't have to be all or nothing right? I mean each call could say\\n> what the situation is like in their context, like\\n> CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\\n> slowly tag calls as needed, similarly to how we add already CFI based on users\\n> report.\\n\\nAbsolutely. My gut feeling is that it's going to be simpler to pick a\\nsmall number of places that are safe and sufficient for this\\nparticular feature and add an extra call there, similar to how we do\\nvacuum_delay_point(). The reason I think that's likely to be better is\\nthat it will likely require changing only a relatively small number of\\nplaces. If we instead start annotating CFIs, well, we've got hundreds\\nof those. That's a lot more to change, and it also inconveniences\\nthird-party extension authors and people doing back-patching. I'm not\\nhere to say it can't work; I just think it's likely not the easiest\\npath.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Feb 2024 20:53:22 +0530\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 23, 2024 at 10:23 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> > On Fri, Feb 23, 2024 at 10:22:32AM +0530, Robert Haas wrote:\\n> > > On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\\n> > > > This is potentially a bit of a wild idea, but I wonder if having some\\n> > > > kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\\n> > > > \\\"normal\\\" as opposed to \\\"critical\\\" (using that word differently than\\n> > > > the existing critical sections) would be worth it.\\n> > >\\n> > > It's worth considering, but the definition of \\\"normal\\\" vs. \\\"critical\\\"\\n> > > might be hard to pin down. Or, we might end up with a definition that\\n> > > is specific to this particular case and not generalizable to others.\\n> >\\n> > But it doesn't have to be all or nothing right? I mean each call could say\\n> > what the situation is like in their context, like\\n> > CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\\n> > slowly tag calls as needed, similarly to how we add already CFI based on users\\n> > report.\\n>\\n> Absolutely. My gut feeling is that it's going to be simpler to pick a\\n> small number of places that are safe and sufficient for this\\n> particular feature and add an extra call there, similar to how we do\\n> vacuum_delay_point(). The reason I think that's likely to be better is\\n> that it will likely require changing only a relatively small number of\\n> places. If we instead start annotating CFIs, well, we've got hundreds\\n> of those. That's a lot more to change, and it also inconveniences\\n> third-party extension authors and people doing back-patching. I'm not\\n> here to say it can't work; I just think it's likely not the easiest\\n> path.\\n\\nYes, I suspect it's not the easiest path. I have a small hunch it\\nmight end up paying more dividends in the future: there isn't just one\\nof these things that is regularly a thorny discussion for the same\\nreasons each time (basically \\\"no way to trigger this safely from\\nanother backend interrupting another one at an arbitrary point\\\"), and\\nif we were able to generalize a solution we may have multiple wins (a\\nvery obvious one in my mind is the inability of auto explain to run an\\nexplain at the precise time it's most useful: when statement timeout\\nfires).\\n\\nBut it's also possible there are simply ways that get us more than\\nthis scenario also, so I might be wrong; it's merely a hunch.\\n\\nRegards,\\nJames Coleman\\n\\n\\n\",\n \"msg_date\": \"Sat, 24 Feb 2024 08:56:41 -0500\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 24, 2024 at 08:56:41AM -0500, James Coleman wrote:\\n> On Fri, Feb 23, 2024 at 10:23 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> >\\n> > On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> > >\\n> > > But it doesn't have to be all or nothing right? I mean each call could say\\n> > > what the situation is like in their context, like\\n> > > CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\\n> > > slowly tag calls as needed, similarly to how we add already CFI based on users\\n> > > report.\\n> >\\n> > Absolutely. My gut feeling is that it's going to be simpler to pick a\\n> > small number of places that are safe and sufficient for this\\n> > particular feature and add an extra call there, similar to how we do\\n> > vacuum_delay_point(). The reason I think that's likely to be better is\\n> > that it will likely require changing only a relatively small number of\\n> > places. If we instead start annotating CFIs, well, we've got hundreds\\n> > of those. That's a lot more to change, and it also inconveniences\\n> > third-party extension authors and people doing back-patching. I'm not\\n> > here to say it can't work; I just think it's likely not the easiest\\n> > path.\\n>\\n> Yes, I suspect it's not the easiest path. I have a small hunch it\\n> might end up paying more dividends in the future: there isn't just one\\n> of these things that is regularly a thorny discussion for the same\\n> reasons each time (basically \\\"no way to trigger this safely from\\n> another backend interrupting another one at an arbitrary point\\\"), and\\n> if we were able to generalize a solution we may have multiple wins (a\\n> very obvious one in my mind is the inability of auto explain to run an\\n> explain at the precise time it's most useful: when statement timeout\\n> fires).\\n\\nYeah, trying to find a generalized solution seems like worth investing some\\ntime to avoid having a bunch of CHECK_FOR_XXX() calls scattered in the code a\\nfew years down the road.\\n\\nI might be missing something, but since we already have a ton of macro hacks,\\nwhy not get another to allow CFI() overloading rather than modifying every\\nsingle call? Something like that should do the trick (and CFIFlagHandler() is\\njust a naive example with a function call to avoid multiple evaluation, should\\nbe replaced with anything that required more than 10s thinking):\\n\\n#define CHECK_FOR_INTERRUPTS_0() \\\\\\ndo { \\\\\\n\\tif (INTERRUPTS_PENDING_CONDITION()) \\\\\\n\\t\\tProcessInterrupts(); \\\\\\n} while(0)\\n\\n#define CHECK_FOR_INTERRUPTS_1(f) \\\\\\ndo { \\\\\\n\\tif (INTERRUPTS_PENDING_CONDITION()) \\\\\\n\\t\\tProcessInterrupts(); \\\\\\n\\t\\\\\\n\\tCFIFlagHandler(f); \\\\\\n} while(0)\\n\\n#define CHECK_FOR_INTERRUPTS_X(x, f, CFI_IMPL, ...) CFI_IMPL\\n\\n#define CHECK_FOR_INTERRUPTS(...) \\\\\\n\\tCHECK_FOR_INTERRUPTS_X(, ##__VA_ARGS__, \\\\\\n\\t\\t\\t\\t\\t\\t\\tCHECK_FOR_INTERRUPTS_1(__VA_ARGS__), \\\\\\n\\t\\t\\t\\t\\t\\t\\tCHECK_FOR_INTERRUPTS_0(__VA_ARGS__) \\\\\\n\\t\\t\\t\\t\\t\\t)\\n\\nWe would have to duplicate the current CFI body, but it should never really\\nchange, and we shouldn't end up with more than 2 overloads anyway so I don't\\nsee it being much of a problem.\\n\\n\\n\",\n \"msg_date\": \"Sun, 25 Feb 2024 19:30:37 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 25, 2024 at 5:00 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n>\\n> On Sat, Feb 24, 2024 at 08:56:41AM -0500, James Coleman wrote:\\n> > On Fri, Feb 23, 2024 at 10:23 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> > >\\n> > > On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> > > >\\n> > > > But it doesn't have to be all or nothing right? I mean each call could say\\n> > > > what the situation is like in their context, like\\n> > > > CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\\n> > > > slowly tag calls as needed, similarly to how we add already CFI based on users\\n> > > > report.\\n\\nThat has some potential ...\\n\\n>\\n> Yeah, trying to find a generalized solution seems like worth investing some\\n> time to avoid having a bunch of CHECK_FOR_XXX() calls scattered in the code a\\n> few years down the road.\\n>\\n> I might be missing something, but since we already have a ton of macro hacks,\\n> why not get another to allow CFI() overloading rather than modifying every\\n> single call? Something like that should do the trick (and CFIFlagHandler() is\\n> just a naive example with a function call to avoid multiple evaluation, should\\n> be replaced with anything that required more than 10s thinking):\\n>\\n> #define CHECK_FOR_INTERRUPTS_0() \\\\\\n> do { \\\\\\n> if (INTERRUPTS_PENDING_CONDITION()) \\\\\\n> ProcessInterrupts(); \\\\\\n> } while(0)\\n>\\n> #define CHECK_FOR_INTERRUPTS_1(f) \\\\\\n> do { \\\\\\n> if (INTERRUPTS_PENDING_CONDITION()) \\\\\\n> ProcessInterrupts(); \\\\\\n> \\\\\\n> CFIFlagHandler(f); \\\\\\n> } while(0)\\n\\n From your earlier description I thought you are talking about flags\\nthat can be ORed. We need only two macros above. Why are we calling\\nCFIFLagHandler() after ProcessInterrupts()? Shouldn't we pass flags to\\nProcessInterrupts() itself.\\n\\n>\\n> #define CHECK_FOR_INTERRUPTS_X(x, f, CFI_IMPL, ...) CFI_IMPL\\n>\\n> #define CHECK_FOR_INTERRUPTS(...) \\\\\\n> CHECK_FOR_INTERRUPTS_X(, ##__VA_ARGS__, \\\\\\n> CHECK_FOR_INTERRUPTS_1(__VA_ARGS__), \\\\\\n> CHECK_FOR_INTERRUPTS_0(__VA_ARGS__) \\\\\\n> )\\n>\\n> We would have to duplicate the current CFI body, but it should never really\\n> change, and we shouldn't end up with more than 2 overloads anyway so I don't\\n> see it being much of a problem.\\n\\nWhy do we need this complex macro?\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Mon, 26 Feb 2024 12:19:42 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 26, 2024 at 12:19:42PM +0530, Ashutosh Bapat wrote:\\n> On Sun, Feb 25, 2024 at 5:00 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> >\\n> > > > On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> > > > >\\n> > > > > But it doesn't have to be all or nothing right? I mean each call could say\\n> > > > > what the situation is like in their context, like\\n> > > > > CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | GUARANTEE_WHATEVER), and\\n> > > > > slowly tag calls as needed, similarly to how we add already CFI based on users\\n> > > > > report.\\n>\\n> That has some potential ...\\n>\\n> >\\n> > I might be missing something, but since we already have a ton of macro hacks,\\n> > why not get another to allow CFI() overloading rather than modifying every\\n> > single call? Something like that should do the trick (and CFIFlagHandler() is\\n> > just a naive example with a function call to avoid multiple evaluation, should\\n> > be replaced with anything that required more than 10s thinking):\\n> >\\n> > #define CHECK_FOR_INTERRUPTS_0() \\\\\\n> > do { \\\\\\n> > if (INTERRUPTS_PENDING_CONDITION()) \\\\\\n> > ProcessInterrupts(); \\\\\\n> > } while(0)\\n> >\\n> > #define CHECK_FOR_INTERRUPTS_1(f) \\\\\\n> > do { \\\\\\n> > if (INTERRUPTS_PENDING_CONDITION()) \\\\\\n> > ProcessInterrupts(); \\\\\\n> > \\\\\\n> > CFIFlagHandler(f); \\\\\\n> > } while(0)\\n>\\n> From your earlier description I thought you are talking about flags\\n> that can be ORed. We need only two macros above. Why are we calling\\n> CFIFLagHandler() after ProcessInterrupts()? Shouldn't we pass flags to\\n> ProcessInterrupts() itself.\\n\\nYes, I'm still talking about ORed flags passed to CFI(). That CFIFlagHandler\\ncall is just an example for a generalized function that would act those flags,\\nrather than having it coded inside the macro.\\n>\\n> >\\n> > #define CHECK_FOR_INTERRUPTS_X(x, f, CFI_IMPL, ...) CFI_IMPL\\n> >\\n> > #define CHECK_FOR_INTERRUPTS(...) \\\\\\n> > CHECK_FOR_INTERRUPTS_X(, ##__VA_ARGS__, \\\\\\n> > CHECK_FOR_INTERRUPTS_1(__VA_ARGS__), \\\\\\n> > CHECK_FOR_INTERRUPTS_0(__VA_ARGS__) \\\\\\n> > )\\n> >\\n> > We would have to duplicate the current CFI body, but it should never really\\n> > change, and we shouldn't end up with more than 2 overloads anyway so I don't\\n> > see it being much of a problem.\\n>\\n> Why do we need this complex macro?\\n\\nSo that client code can use either CHECK_FOR_INTERRUPTS() or\\nCHECK_FOR_INTERRUPTS(flag) rather that transforming every single\\nCHECK_FOR_INTERRUPTS() to CHECK_FOR_INTERRUPTS(0), which was Robert's\\ncomplaint about this approach.\\n\\n\\n\",\n \"msg_date\": \"Mon, 26 Feb 2024 15:55:04 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 26, 2024 at 1:25 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> >\\n> > >\\n> > > #define CHECK_FOR_INTERRUPTS_X(x, f, CFI_IMPL, ...) CFI_IMPL\\n> > >\\n> > > #define CHECK_FOR_INTERRUPTS(...) \\\\\\n> > > CHECK_FOR_INTERRUPTS_X(, ##__VA_ARGS__, \\\\\\n> > > CHECK_FOR_INTERRUPTS_1(__VA_ARGS__), \\\\\\n> > > CHECK_FOR_INTERRUPTS_0(__VA_ARGS__) \\\\\\n> > > )\\n> > >\\n> > > We would have to duplicate the current CFI body, but it should never really\\n> > > change, and we shouldn't end up with more than 2 overloads anyway so I don't\\n> > > see it being much of a problem.\\n> >\\n> > Why do we need this complex macro?\\n>\\n> So that client code can use either CHECK_FOR_INTERRUPTS() or\\n> CHECK_FOR_INTERRUPTS(flag) rather that transforming every single\\n> CHECK_FOR_INTERRUPTS() to CHECK_FOR_INTERRUPTS(0), which was Robert's\\n> complaint about this approach.\\n\\nIt might be better to just create two marcos (with names like\\nCHECK_FOR_INTERRUPTS() and CHECK_FOR_INTERRUPTS_SAFE()) to call\\nProcessInterrupt() directly and modify ProcessInterrupts() to accept\\nthe flag (and if required call CFIFlagHandler() additionally). Last\\nmacro is hard to understand.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Mon, 26 Feb 2024 13:45:08 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 25, 2024 at 5:00 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> Yeah, trying to find a generalized solution seems like worth investing some\\n> time to avoid having a bunch of CHECK_FOR_XXX() calls scattered in the code a\\n> few years down the road.\\n\\nI just don't really see how to do it. I suspect that every task that\\nwants to run at some CFIs but not others is going to have slightly\\ndifferent requirements, and we probably can't foresee what all of\\nthose requirements are.\\n\\nSaid another way, if in the future we want to call\\nDoSomethingOrOther() from the CFI handler, well then we need to know\\nthat we're not already in the middle of using any subsystem that\\nDoSomethingOrOther() might also try to use ... and we also need to\\nknow that we're not in the middle of doing anything that's more\\ncritical than DoSomethingOrOther(). But both of these are likely to\\nvary in each case.\\n\\nEXPLAIN might be one member of a general class of things that require\\ncatalog access (and thus might take locks or lwlocks, access the\\ncatalogs, trigger invalidations, etc.) but it's not clear how general\\nthat class really is. Also, I think if we try to do too many different\\nkinds of things at CFIs, the whole thing is going to fall apart.\\nYou'll end up failing to foresee some interactions, or the stack will\\nget too deep, or ... something.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 26 Feb 2024 13:56:44 +0530\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2024-02-24 00:23, Robert Haas wrote:\\n> On Fri, Feb 23, 2024 at 7:50 PM Julien Rouhaud <rjuju123@gmail.com> \\n> wrote:\\n>> On Fri, Feb 23, 2024 at 10:22:32AM +0530, Robert Haas wrote:\\n>> > On Thu, Feb 22, 2024 at 6:25 AM James Coleman <jtc331@gmail.com> wrote:\\n>> > > This is potentially a bit of a wild idea, but I wonder if having some\\n>> > > kind of argument to CHECK_FOR_INTERRUPTS() signifying we're in\\n>> > > \\\"normal\\\" as opposed to \\\"critical\\\" (using that word differently than\\n>> > > the existing critical sections) would be worth it.\\n>> >\\n>> > It's worth considering, but the definition of \\\"normal\\\" vs. \\\"critical\\\"\\n>> > might be hard to pin down. Or, we might end up with a definition that\\n>> > is specific to this particular case and not generalizable to others.\\n>> \\n>> But it doesn't have to be all or nothing right? I mean each call \\n>> could say\\n>> what the situation is like in their context, like\\n>> CHECK_FOR_INTERRUPTS(GUARANTEE_NO_HEAVYWEIGHT_LOCK | \\n>> GUARANTEE_WHATEVER), and\\n>> slowly tag calls as needed, similarly to how we add already CFI based \\n>> on users\\n>> report.\\n> \\n> Absolutely. My gut feeling is that it's going to be simpler to pick a\\n> small number of places that are safe and sufficient for this\\n> particular feature and add an extra call there\\n\\nHmm, whether extending CHECK_FOR_INTERRUPTS() or adding extras call \\ndirectly, currently I'm not sure where are the good 'places', which \\ndon't give performance impact.\\n\\nAs attached PoC patch, I experimentally added extra calls on \\nExecScanFetch() which would be less called than ExecProcNode()[1].\\nWhen running sequential scan on pgbench_accounts which is on the memory, \\nthere seems a performance degradation.\\n\\n - Executed \\\"select * from pgbench_accounts\\\" for 20 times\\n - Compared the elapsed time between the patch applied and not applied \\non 874d817baa160ca7e68\\n - there were no heap_blks_read during the query\\n - pgbench_accounts has 3000000 rows\\n\\n patch NOT applied:\\n - average: 335.88 ms\\n - max: 367.313 ms\\n - min: 309.609 ms\\n\\n patch applied:\\n - average: 342.57 ms\\n - max: 380.099 ms\\n - min: 324.270 ms\\n\\nIt would be nice if there was a place accessed once every few seconds or \\nso..\\n\\n[1] \\nhttps://www.postgresql.org/message-id/20240215185911.v4o6fo444md6a3w7%40awork3.anarazel.de\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\",\n \"msg_date\": \"Mon, 26 Feb 2024 21:01:45 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 26, 2024 at 01:56:44PM +0530, Robert Haas wrote:\\n> On Sun, Feb 25, 2024 at 5:00 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> > Yeah, trying to find a generalized solution seems like worth investing some\\n> > time to avoid having a bunch of CHECK_FOR_XXX() calls scattered in the code a\\n> > few years down the road.\\n>\\n> I just don't really see how to do it. I suspect that every task that\\n> wants to run at some CFIs but not others is going to have slightly\\n> different requirements, and we probably can't foresee what all of\\n> those requirements are.\\n>\\n> Said another way, if in the future we want to call\\n> DoSomethingOrOther() from the CFI handler, well then we need to know\\n> that we're not already in the middle of using any subsystem that\\n> DoSomethingOrOther() might also try to use ... and we also need to\\n> know that we're not in the middle of doing anything that's more\\n> critical than DoSomethingOrOther(). But both of these are likely to\\n> vary in each case.\\n>\\n> EXPLAIN might be one member of a general class of things that require\\n> catalog access (and thus might take locks or lwlocks, access the\\n> catalogs, trigger invalidations, etc.) but it's not clear how general\\n> that class really is. Also, I think if we try to do too many different\\n> kinds of things at CFIs, the whole thing is going to fall apart.\\n> You'll end up failing to foresee some interactions, or the stack will\\n> get too deep, or ... something.\\n\\nI still fail to understand your point. So you say we might want to check for\\nsafe condition to run explain or DoSomethingOrOther or even DoMeh, with all\\ndifferent requirements. IIUC what you're saying is that we should have\\nCHECK_FOR_EXPLAIN(), CHECK_FOR_DOSOMETHINGOROTHER() and CHECK_FOR_MEH()?\\n\\nAnd so in some code path A we could have\\n\\nCHECK_FOR_INTERRUPTS();\\nCHECK_FOR_EXPLAIN();\\n\\nIn another\\n\\nCHECK_FOR_INTERRUPTS();\\nCHECK_FOR_DOSOMETHINGOROTHER();\\n\\nand in one happy path\\n\\nCHECK_FOR_INTERRUPTS();\\nCHECK_FOR_EXPLAIN();\\nCHECK_FOR_DOSOMETHINGOROTHER();\\nCHECK_FOR_MEH();\\n\\nOr that we should still have all of those but they shouldn't be anywhere close\\nto a CHECK_FOR_INTERRUPTS(), or something totally different?\\n\\nIn the first case, I'm not sure why having CHECK_FOR_INTERRUPTS(EXPLAIN|...),\\nwith a combination of flags and with the flag handling being done after\\nProcessIterrupts(), would be any different apart from having less boilerplate\\nlines. Similarly for the 2nd case, but relying on a single more general\\nCHECK_FOR_CONDITION(EXPLAIN | ...) rather than N CHECK_FOR_XXX?\\n\\nIf you just want something totally different then sure.\\n\\n\\n\",\n \"msg_date\": \"Mon, 26 Feb 2024 21:54:41 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 26, 2024 at 5:31 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n> It would be nice if there was a place accessed once every few seconds or\\n> so..\\n\\nI think this comment earlier from Andres deserves close attention:\\n\\n# If we went with something like tht approach, I think we'd have to do something\\n# like redirecting node->ExecProcNode to a wrapper, presumably from within a\\n# CFI. That wrapper could then implement the explain support, without slowing\\n# down the normal execution path.\\n\\nIf this is correctly implemented, the overhead in the case where the\\nfeature isn't used should be essentially zero, I believe.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 28 Feb 2024 11:48:32 +0530\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 28, 2024 at 1:18 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> On Mon, Feb 26, 2024 at 5:31 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n> > It would be nice if there was a place accessed once every few seconds or\\n> > so..\\n>\\n> I think this comment earlier from Andres deserves close attention:\\n>\\n> # If we went with something like tht approach, I think we'd have to do something\\n> # like redirecting node->ExecProcNode to a wrapper, presumably from within a\\n> # CFI. That wrapper could then implement the explain support, without slowing\\n> # down the normal execution path.\\n>\\n> If this is correctly implemented, the overhead in the case where the\\n> feature isn't used should be essentially zero, I believe.\\n\\nIf I can rephrase this idea: it's basically \\\"delay this interrupt\\nuntil inline to the next ExecProcNode execution\\\".\\n\\nThat seems pretty promising to me as well.\\n\\nRegards,\\nJames Coleman\\n\\n\\n\",\n \"msg_date\": \"Sat, 2 Mar 2024 10:46:36 -0500\",\n \"msg_from\": \"James Coleman <jtc331@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Sat, Mar 2, 2024 at 10:46 AM James Coleman <jtc331@gmail.com> wrote:\\n> If I can rephrase this idea: it's basically \\\"delay this interrupt\\n> until inline to the next ExecProcNode execution\\\".\\n\\nYes, but it's not just that. It also means that the code which would\\nhandle the interrupt doesn't need to be called at every ExecProcNode.\\nOnly when the interrupt actually arrives do we enable the code that\\nhandles it.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Mar 2024 13:19:45 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 16, 2024 at 11:42 PM torikoshia <torikoshia@oss.nttdata.com> \\nwrote:\\n> I'm not so sure about the implementation now, i.e. finding the next \\n> node\\n> to be executed from the planstate tree, but I'm going to try this\\n> approach.\\n\\nAttached a patch which takes this approach.\\n\\n- I saw no way to find the next node to be executed from the planstate \\ntree, so the patch wraps all the ExecProcNode of the planstate tree at \\nCHECK_FOR_INTERRUPTS().\\n- To prevent overhead of this wrapped function call, unwrap it at the \\nend of EXPLAIN code execution.\\n- I first tried to use ExecSetExecProcNode() for wrapping, but it \\n'changes' ExecProcNodeMtd of nodes, not 'adds' some process to \\nExecProcNodeMtd. I'm not sure this is the right approach, but attached \\npatch adds new member ExecProcNodeOriginal to PlanState to preserve \\noriginal ExecProcNodeMtd.\\n\\nAny comments are welcomed.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\",\n \"msg_date\": \"Wed, 13 Mar 2024 14:28:42 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 13, 2024 at 1:28 AM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n> - I saw no way to find the next node to be executed from the planstate\\n> tree, so the patch wraps all the ExecProcNode of the planstate tree at\\n> CHECK_FOR_INTERRUPTS().\\n\\nI don't think it does this correctly, because some node types have\\nchildren other than the left and right node. See /* special child\\nplans */ in ExplainNode().\\n\\nBut also ... having to wrap the entire plan tree like this seems\\npretty awful. I don't really like the idea of a large-scan plan\\nmodification like this in the middle of the query. I also wonder\\nwhether it interacts properly with JIT. But at the same time, I wonder\\nhow you're supposed to avoid it.\\n\\nAndres, did you have some clever idea for this feature that would\\navoid the need to do this?\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Mar 2024 15:33:02 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 13, 2024 at 1:28 PM torikoshia <torikoshia@oss.nttdata.com> wrote:\\n>\\n> On Fri, Feb 16, 2024 at 11:42 PM torikoshia <torikoshia@oss.nttdata.com>\\n> wrote:\\n> > I'm not so sure about the implementation now, i.e. finding the next\\n> > node\\n> > to be executed from the planstate tree, but I'm going to try this\\n> > approach.\\n>\\n> Attached a patch which takes this approach.\\n>\\n\\none minor issue.\\nI understand the regress test, compare the expected outcome with\\ntestrun outcome,\\nbut can you enlighten me about how you check if the change you made in\\ncontrib/auto_explain/t/001_auto_explain.pl is correct.\\n(i am not familiar with perl).\\n\\ndiff --git a/src/include/commands/explain.h b/src/include/commands/explain.h\\nindex cf195f1359..2d06bf297e 100644\\n--- a/src/include/commands/explain.h\\n+++ b/src/include/commands/explain.h\\n@@ -17,6 +17,8 @@\\n #include \\\"lib/stringinfo.h\\\"\\n #include \\\"parser/parse_node.h\\\"\\n+extern PGDLLIMPORT bool ProcessLogQueryPlanInterruptActive;\\n\\ndiff --git a/src/include/utils/elog.h b/src/include/utils/elog.h\\nindex 054dd2bf62..3c6bd1ea7c 100644\\n--- a/src/include/utils/elog.h\\n+++ b/src/include/utils/elog.h\\n@@ -167,6 +167,7 @@ struct Node;\\n extern bool message_level_is_interesting(int elevel);\\n+extern PGDLLIMPORT bool ProcessLogQueryPlanInterruptActive;\\n\\nutils/elog.h is already included in src/include/postgres.h.\\nyou don't need to declare ProcessLogQueryPlanInterruptActive at\\ninclude/commands/explain.h?\\ngenerally a variable should only declare once?\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Mar 2024 09:02:47 +0800\",\n \"msg_from\": \"jian he <jian.universality@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"On 2024-03-14 04:33, Robert Haas wrote:\\nThanks for your review!\\n\\n> On Wed, Mar 13, 2024 at 1:28 AM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> - I saw no way to find the next node to be executed from the planstate\\n>> tree, so the patch wraps all the ExecProcNode of the planstate tree at\\n>> CHECK_FOR_INTERRUPTS().\\n> \\n> I don't think it does this correctly, because some node types have\\n> children other than the left and right node. See /* special child\\n> plans */ in ExplainNode().\\n\\nAgreed.\\n\\n> But also ... having to wrap the entire plan tree like this seems\\n> pretty awful. I don't really like the idea of a large-scan plan\\n> modification like this in the middle of the query.\\n\\nYeah, but I haven't come up with other ideas to select the appropriate \\nnode to wrap.\\n\\n> Andres, did you have some clever idea for this feature that would\\n> avoid the need to do this?\\n\\n\\n\\nOn 2024-03-14 10:02, jian he wrote:\\n> On Wed, Mar 13, 2024 at 1:28 PM torikoshia <torikoshia@oss.nttdata.com> \\n> wrote:\\n>> \\n>> On Fri, Feb 16, 2024 at 11:42 PM torikoshia \\n>> <torikoshia@oss.nttdata.com>\\n>> wrote:\\n>> > I'm not so sure about the implementation now, i.e. finding the next\\n>> > node\\n>> > to be executed from the planstate tree, but I'm going to try this\\n>> > approach.\\n>> \\n>> Attached a patch which takes this approach.\\n>> \\n> \\n> one minor issue.\\n> I understand the regress test, compare the expected outcome with\\n> testrun outcome,\\n> but can you enlighten me about how you check if the change you made in\\n> contrib/auto_explain/t/001_auto_explain.pl is correct.\\n> (i am not familiar with perl).\\n\\n$pg_log_plan_query_output saves the output plan of pg_log_query_plan() \\nand $auto_explain_output saves the output plan of auto_explain.\\nThe test checks the both are the same using cmp_ok().\\n\\nDid I answer your question?\\n\\n> diff --git a/src/include/commands/explain.h \\n> b/src/include/commands/explain.h\\n> index cf195f1359..2d06bf297e 100644\\n> --- a/src/include/commands/explain.h\\n> +++ b/src/include/commands/explain.h\\n> @@ -17,6 +17,8 @@\\n> #include \\\"lib/stringinfo.h\\\"\\n> #include \\\"parser/parse_node.h\\\"\\n> +extern PGDLLIMPORT bool ProcessLogQueryPlanInterruptActive;\\n> \\n> diff --git a/src/include/utils/elog.h b/src/include/utils/elog.h\\n> index 054dd2bf62..3c6bd1ea7c 100644\\n> --- a/src/include/utils/elog.h\\n> +++ b/src/include/utils/elog.h\\n> @@ -167,6 +167,7 @@ struct Node;\\n> extern bool message_level_is_interesting(int elevel);\\n> +extern PGDLLIMPORT bool ProcessLogQueryPlanInterruptActive;\\n> \\n> utils/elog.h is already included in src/include/postgres.h.\\n> you don't need to declare ProcessLogQueryPlanInterruptActive at\\n> include/commands/explain.h?\\n> generally a variable should only declare once?\\n\\nYeah, this declaration is not necessary and we should add include \\ncommands/explain.h to src/backend/access/transam/xact.c.\\n\\n-- \\nRegards,\\n\\n--\\nAtsushi Torikoshi\\nNTT DATA Group Corporation\\n\\n\\n\",\n \"msg_date\": \"Mon, 18 Mar 2024 14:39:48 +0900\",\n \"msg_from\": \"torikoshia <torikoshia@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2024-03-13 15:33:02 -0400, Robert Haas wrote:\\n> But also ... having to wrap the entire plan tree like this seems\\n> pretty awful. I don't really like the idea of a large-scan plan\\n> modification like this in the middle of the query.\\n\\nIt's not great. But I also don't really see an alternative with this approach.\\n\\nI guess we could invent a new CFI version that gets the current PlanState and\\nuse that in all of src/backend/executor/node* and pass the PlanState to that -\\nbut then we could just as well just directly process the interrupt there.\\n\\n\\n> I also wonder whether it interacts properly with JIT.\\n\\nI don't think there's a problem unless somebody invests a lot of time in\\nJITing much more of the query. Which will require a lot more work, basically\\nredesigning the executor...\\n\\n\\n\\n> Andres, did you have some clever idea for this feature that would\\n> avoid the need to do this?\\n\\nNo. I think it's acceptable though.\\n\\nHowever it might be worth inventing an executor tree walker in a preliminary\\nstep. We have already quite a few switches over all plan nodes, which we could\\nlargely replace with a helper.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Mar 2024 19:34:47 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: Logging plan of the running query\"\n }\n]","isConversational":false}}},{"rowIdx":53040,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\n\nSince we have the regexp_like operator I have found that there is two \nreferences in the documentation about PostgreSQL lacking of LIKE_REGEX \nimplementation. Here is a patch to fix the documentation. I simply \nremove the reference to non exist of LIKE_REGEX in PostgreSQL in chapter \n\"9.7.3.8 Differences from XQuery\" And try to modify chapter \"9.16.2.3. \nSQL/JSON Regular Expressions\" to mention the REGEXP_LIKE operator. For \nthe second fix there should be better wording.\n\n\nBest regards,\n\n-- \nGilles Darold","msg_date":"Tue, 2 Nov 2021 15:58:49 +0100","msg_from":"Gilles Darold <gilles@migops.com>","msg_from_op":true,"msg_subject":"[PATCH] fix references to like_regex"},{"msg_contents":"Gilles Darold <gilles@migops.com> writes:\n> Since we have the regexp_like operator I have found that there is two \n> references in the documentation about PostgreSQL lacking of LIKE_REGEX \n> implementation. Here is a patch to fix the documentation. I simply \n> remove the reference to non exist of LIKE_REGEX in PostgreSQL in chapter \n> \"9.7.3.8 Differences from XQuery\" And try to modify chapter \"9.16.2.3. \n> SQL/JSON Regular Expressions\" to mention the REGEXP_LIKE operator. For \n> the second fix there should be better wording.\n\nI don't think we should change these (yet). regexp_like() is *not*\nLIKE_REGEX, precisely because it's using a slightly different\nregular-expression language than what the spec calls for.\nAt some point we may provide a skin for the regexp engine that\nduplicates the spec's definition, and then we can implement\nLIKE_REGEX for real.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 02 Nov 2021 11:50:47 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] fix references to like_regex"},{"msg_contents":"Le 02/11/2021 à 16:50, Tom Lane a écrit :\n> Gilles Darold <gilles@migops.com> writes:\n>> Since we have the regexp_like operator I have found that there is two\n>> references in the documentation about PostgreSQL lacking of LIKE_REGEX\n>> implementation. Here is a patch to fix the documentation. I simply\n>> remove the reference to non exist of LIKE_REGEX in PostgreSQL in chapter\n>> \"9.7.3.8 Differences from XQuery\" And try to modify chapter \"9.16.2.3.\n>> SQL/JSON Regular Expressions\" to mention the REGEXP_LIKE operator. For\n>> the second fix there should be better wording.\n> I don't think we should change these (yet). regexp_like() is *not*\n> LIKE_REGEX, precisely because it's using a slightly different\n> regular-expression language than what the spec calls for.\n> At some point we may provide a skin for the regexp engine that\n> duplicates the spec's definition, and then we can implement\n> LIKE_REGEX for real.\n\n\nThanks for clarifying, I thought it was an oversight.\n\n\nRegards\n\n-- \nGilles Darold\n\n\n\n\n\n\n\nLe 02/11/2021 à 16:50, Tom Lane a\n écrit :\n\n\nGilles Darold <gilles@migops.com> writes:\n\n\nSince we have the regexp_like operator I have found that there is two \nreferences in the documentation about PostgreSQL lacking of LIKE_REGEX \nimplementation. Here is a patch to fix the documentation. I simply \nremove the reference to non exist of LIKE_REGEX in PostgreSQL in chapter \n\"9.7.3.8 Differences from XQuery\" And try to modify chapter \"9.16.2.3. \nSQL/JSON Regular Expressions\" to mention the REGEXP_LIKE operator. For \nthe second fix there should be better wording.\n\n\n\nI don't think we should change these (yet). regexp_like() is *not*\nLIKE_REGEX, precisely because it's using a slightly different\nregular-expression language than what the spec calls for.\nAt some point we may provide a skin for the regexp engine that\nduplicates the spec's definition, and then we can implement\nLIKE_REGEX for real.\n\n\n\n\nThanks for clarifying, I thought it was an\n oversight.\n\n\nRegards\n\n-- \nGilles Darold","msg_date":"Tue, 2 Nov 2021 16:56:10 +0100","msg_from":"Gilles Darold <gilles@migops.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] fix references to like_regex"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\n\\nSince we have the regexp_like operator I have found that there is two \\nreferences in the documentation about PostgreSQL lacking of LIKE_REGEX \\nimplementation. Here is a patch to fix the documentation. I simply \\nremove the reference to non exist of LIKE_REGEX in PostgreSQL in chapter \\n\\\"9.7.3.8 Differences from XQuery\\\" And try to modify chapter \\\"9.16.2.3. \\nSQL/JSON Regular Expressions\\\" to mention the REGEXP_LIKE operator. For \\nthe second fix there should be better wording.\\n\\n\\nBest regards,\\n\\n-- \\nGilles Darold\",\n \"msg_date\": \"Tue, 2 Nov 2021 15:58:49 +0100\",\n \"msg_from\": \"Gilles Darold <gilles@migops.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[PATCH] fix references to like_regex\"\n },\n {\n \"msg_contents\": \"Gilles Darold <gilles@migops.com> writes:\\n> Since we have the regexp_like operator I have found that there is two \\n> references in the documentation about PostgreSQL lacking of LIKE_REGEX \\n> implementation. Here is a patch to fix the documentation. I simply \\n> remove the reference to non exist of LIKE_REGEX in PostgreSQL in chapter \\n> \\\"9.7.3.8 Differences from XQuery\\\" And try to modify chapter \\\"9.16.2.3. \\n> SQL/JSON Regular Expressions\\\" to mention the REGEXP_LIKE operator. For \\n> the second fix there should be better wording.\\n\\nI don't think we should change these (yet). regexp_like() is *not*\\nLIKE_REGEX, precisely because it's using a slightly different\\nregular-expression language than what the spec calls for.\\nAt some point we may provide a skin for the regexp engine that\\nduplicates the spec's definition, and then we can implement\\nLIKE_REGEX for real.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 02 Nov 2021 11:50:47 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] fix references to like_regex\"\n },\n {\n \"msg_contents\": \"Le 02/11/2021 à 16:50, Tom Lane a écrit :\\n> Gilles Darold <gilles@migops.com> writes:\\n>> Since we have the regexp_like operator I have found that there is two\\n>> references in the documentation about PostgreSQL lacking of LIKE_REGEX\\n>> implementation. Here is a patch to fix the documentation. I simply\\n>> remove the reference to non exist of LIKE_REGEX in PostgreSQL in chapter\\n>> \\\"9.7.3.8 Differences from XQuery\\\" And try to modify chapter \\\"9.16.2.3.\\n>> SQL/JSON Regular Expressions\\\" to mention the REGEXP_LIKE operator. For\\n>> the second fix there should be better wording.\\n> I don't think we should change these (yet). regexp_like() is *not*\\n> LIKE_REGEX, precisely because it's using a slightly different\\n> regular-expression language than what the spec calls for.\\n> At some point we may provide a skin for the regexp engine that\\n> duplicates the spec's definition, and then we can implement\\n> LIKE_REGEX for real.\\n\\n\\nThanks for clarifying, I thought it was an oversight.\\n\\n\\nRegards\\n\\n-- \\nGilles Darold\\n\\n\\n\\n\\n\\n\\n\\nLe 02/11/2021 à 16:50, Tom Lane a\\n écrit :\\n\\n\\nGilles Darold <gilles@migops.com> writes:\\n\\n\\nSince we have the regexp_like operator I have found that there is two \\nreferences in the documentation about PostgreSQL lacking of LIKE_REGEX \\nimplementation. Here is a patch to fix the documentation. I simply \\nremove the reference to non exist of LIKE_REGEX in PostgreSQL in chapter \\n\\\"9.7.3.8 Differences from XQuery\\\" And try to modify chapter \\\"9.16.2.3. \\nSQL/JSON Regular Expressions\\\" to mention the REGEXP_LIKE operator. For \\nthe second fix there should be better wording.\\n\\n\\n\\nI don't think we should change these (yet). regexp_like() is *not*\\nLIKE_REGEX, precisely because it's using a slightly different\\nregular-expression language than what the spec calls for.\\nAt some point we may provide a skin for the regexp engine that\\nduplicates the spec's definition, and then we can implement\\nLIKE_REGEX for real.\\n\\n\\n\\n\\nThanks for clarifying, I thought it was an\\n oversight.\\n\\n\\nRegards\\n\\n-- \\nGilles Darold\",\n \"msg_date\": \"Tue, 2 Nov 2021 16:56:10 +0100\",\n \"msg_from\": \"Gilles Darold <gilles@migops.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] fix references to like_regex\"\n }\n]","isConversational":false}}},{"rowIdx":53041,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Attached is a patch to create a helper function which creates a\ntuplestore to be used by FUNCAPI-callable functions.\n\nIt was suggested in [1] that I start a separate thread for this patch.\n\nA few notes:\n\nThere are a few places with very similar code to the new helper\n(MakeTuplestore()) but which, for example, don't expect the return type\nto be a row (pg_ls_dir(), pg_tablespace_databases(), text_to_table()). I\nwasn't sure if it was worth modifying it for their benefit.\n\nI wasn't sure if the elog() check for call result type should be changed\nto an ereport().\n\nAll callers of MakeTuplestore() pass work_mem as the third parameter, so\nI could just omit it and hard-code it in, but I wasn't sure that felt\nright in a utility/helper function.\n\nI inlined deflist_to_tuplestore() in foreign.c since it was billed as a\nhelper function but wasn't used elsewhere and it made it harder to use\nMakeTuplestore() in this location.\n\n-- melanie\n\n[1] https://www.postgresql.org/message-id/flat/20200124195226.lth52iydq2n2uilq%40alap3.anarazel.de","msg_date":"Tue, 2 Nov 2021 12:34:30 -0400","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"make tuplestore helper function"},{"msg_contents":"On 2021-Nov-02, Melanie Plageman wrote:\n\n> Attached is a patch to create a helper function which creates a\n> tuplestore to be used by FUNCAPI-callable functions.\n\nLooks good, and given the amount of code being removed, it seems a\nno-brainer that some helper(s) is/are needed.\n\nI think the name is a bit too generic. How about\nMakeFuncResultTuplestore()?\n\nI think the function should not modify rsinfo -- having that as\n(undocumented) side effect is weird. I would suggest to just return the\ntuplestore, and have the caller stash it in rsinfo->setResult and modify\nthe other struct members alongside that. It's a couple more lines per\ncaller, but the code is clearer that way.\n\n> There are a few places with very similar code to the new helper\n> (MakeTuplestore()) but which, for example, don't expect the return type\n> to be a row (pg_ls_dir(), pg_tablespace_databases(), text_to_table()). I\n> wasn't sure if it was worth modifying it for their benefit.\n\nIs this just because of the tupdesc? If that's the case, then maybe the\ntupdesc can be optionally passed in by caller, and when it's not, the\nhelper does get_call_result_type() and the tupdesc is used as output\nargument.\n\n> I wasn't sure if the elog() check for call result type should be changed\n> to an ereport().\n\nI don't see a reason for that: it's still an internal (not user-facing) error.\n\n> All callers of MakeTuplestore() pass work_mem as the third parameter, so\n> I could just omit it and hard-code it in, but I wasn't sure that felt\n> right in a utility/helper function.\n\nNo opinion. I would have used the global in the function instead of\npassing it in.\n\n> I inlined deflist_to_tuplestore() in foreign.c since it was billed as a\n> helper function but wasn't used elsewhere and it made it harder to use\n> MakeTuplestore() in this location.\n\nThis is a bit strange. Does it work to pass rsinfo->expectedDesc?\n\n-- \nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\n\"No renuncies a nada. No te aferres a nada.\"\n\n\n","msg_date":"Tue, 2 Nov 2021 15:17:05 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"Several places have a conditional value for the first argument (randomAccess),\nbut your patch changes the behavior to a constant \"true\". I didn't review the\npatch beyond that.\n\n> @@ -740,18 +724,14 @@ pg_prepared_statement(PG_FUNCTION_ARGS)\n> -\ttupstore =\n> -\t\ttuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\n> -\t\t\t\t\t\t\t false, work_mem);\n\n> @@ -2701,42 +2701,13 @@ pg_hba_file_rules(PG_FUNCTION_ARGS)\n> -\ttuple_store =\n> -\t\ttuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\n> -\t\t\t\t\t\t\t false, work_mem);\n\n> @@ -4799,31 +4797,8 @@ pg_timezone_names(PG_FUNCTION_ARGS)\n> -\trandomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\n> -\ttupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\n\n> @@ -575,38 +575,12 @@ pg_ls_dir_1arg(PG_FUNCTION_ARGS)\n> -\trandomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\n> -\ttupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\n\n> @@ -1170,17 +1154,12 @@ pg_cursor(PG_FUNCTION_ARGS)\n> -\ttupstore =\n> -\t\ttuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\n> -\t\t\t\t\t\t\t false, work_mem);\n\n> +++ b/src/backend/utils/fmgr/funcapi.c\n> +\ttupstore = tuplestore_begin_heap(true, false, maxKBytes);\n\n-- \nJustin\n\n\n","msg_date":"Tue, 2 Nov 2021 15:23:53 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"The attached patch contains a helper now named\nMakeFuncResultTuplestore() which aims to eliminate a few lines of\nredundant code that all FUNCAPI functions making tuplestores repeated.\n\nWhile investigating all of these call sites and making the changes\nsuggested by Álvaro, I noticed that the functions making tuplestores\nbroke out into seven groups. For some of them, using my helper changes\nwhat memory context the tuple descriptor is made in.\n\nExplanation of how/why memory context tuple descriptor is in has\nchanged:\n\nAn earlier version of this patch changed into the per query memory\ncontext inside of the helper function. This version does not do so but\ninstead asserts that the caller has changed into the appropriate memory\ncontext.\n\nFor callers which previously did not call get_call_result_type() and\ninstead used a function like CreateTupleDescCopy() to create the tuple\ndescriptor which was used for the tuples in the tuplestore, most of them\nmade this copy in the per query memory context. For these callers it\nseemed error-prone to switch to the per query memory context to create\nthe copied tuple descriptor and then switch back before calling the\nhelper function. Instead, the helper function will assert that the\ncaller has already switched into the per query memory context so that\nthe tuplestore is made in the per query context.\n\nThis means that for the group of functions which, prior to this patch,\ncalled get_call_result_type() in a shorter-lived memory context but now\nrely on the helper function to call get_call_result_type(), the tuple\ndescriptor will now be in the per query memory context.\n\nI think that this makes sense since most of these are set as the\nResultSetInfo->setDesc (the tuple descriptor for returned tuples).\n\nExplanation of the different groups of functions using the\nMakeFuncResultTuplestore() helper:\n\nThe first group are functions which called get_call_result_type(),\nallocating memory for the tuple descriptor in the short-lived memory\ncontext, then switched into the per query memory context before\ncreating the tuplestore, then switched back to the short-lived memory\ncontext directly after making the tuplestore.\n\nNow that they use the helper function and switch into the per query\nmemory context before calling it, the tuple descriptor is allocated in\nthe per query memory context.\n\nThe first group:\n\n pg_stop_backup_v2()\n pg_event_trigger_dropped_objects()\n pg_event_trigger_ddl_commands()\n pg_available_extensions()\n pg_available_extension_versions()\n pg_extension_update_paths()\n pg_hba_file_rules()\n pg_stat_get_subscription()\n pg_logical_slot_get_changes_guts()\n pg_show_replication_origin_status()\n pg_get_replication_slots()\n pg_stat_get_wal_senders()\n pg_get_shmem_allocations()\n pg_timezone_names()\n pg_ls_dir_files()\n pg_get_backend_memory_contexts()\n pg_stat_get_progress_info()\n pg_stat_get_activity()\n pg_stat_get_slru()\n\nThe second group are functions which, instead of calling\nget_call_result_type(), call CreateTemplateTupleDesc() and\nTupleDescInitEntry() to allocate and initialize the tuple descriptor in\nthe per query memory context.\nTheir memory usage is unchanged by this patch.\n\nThe second group:\n\n pg_prepared_statement()\n pg_ls_dir()\n pg_tablespace_databases()\n show_all_file_settings()\n pg_cursor()\n\nThe third and fourth groups are functions that use CreateTupleDescCopy()\nto make a tuple descriptor copy from ResultSetInfo->expectedDesc.\nThese functions made a copy of ResultSetInfo->expectedDesc but did not\ncheck if it was present before doing so (and do not seem to call a\nfunction which would call internal_get_result_type() -- which does this\ncheck), so I have added that check.\nTheir memory usage is unchanged by this patch.\n\nThe third group:\n\n text_to_table()\n\nThe fourth group differs from the third in that the values for the\ntuples are actually created in the per query memory context -- seemingly\nunnecessarily since tuplestore_putvalues() will eventually end up\ncopying everything when it is forming tuples. This behavior is not\naltered at all by the MakeFuncResultTuplestore() helper function patch,\nhowever I wanted to mention that I noticed that perhaps more was being\nallocated in the per query memory context than is necessary.\n\nThe fourth group:\n\n pg_options_to_table()\n pg_config()\n\nThe fifth group is a function that uses CreateTupleDescCoy() to make a\ncopy of a tuple descriptor (NOT ResultSetInfo->expectedDesc) in the\nfunction info memory context (fcinfo->flinfo->fn_mcxt) so that it can be\naccessed later.\nTheir memory usage is unchanged by this patch.\n\nThe fifth group:\n\n populate_recordset_worker()\n\nThe sixth group of functions use CreateTupleDescCopy() to make a copy of\nthe tuple descriptor that they set up using get_call_result_type().\n\nThey called get_call_result_type() in a short-lived memory context, then\nswitched to the per query memory context and made a copy of it right\naway. It is unclear why they couldn't just call get_call_result_type()\nin the per query context and not make a copy. Perhaps they needed some\nspecific behavior of the copy (like that it doesn't copy constraints and\ndefaults?).\n\nWith the attached patch, the initial tuple descriptor is now made in the\nper query memory context (as well as the copy). I wonder if the initial\ntuple descriptor being made in the per query memory context negated the\nneed for a copy, however I left the copy intact since I wasn't sure.\n\nThese functions also checked for the presence of\nResultSetInfo->expectedDesc which is not applicable for all callers of\nMakeFuncResultTuplestore(), so I separated it into a separate check.\nget_call_result_type() will call internal_get_result_type() which will\ncheck for expectedDesc if the result type is TYPEFUNC_RECORD, so I'm not\nsure if the expectedDesc check is required or not in the body of these\nfunctions.\n\nThe sixth group:\n\n each_worker()\n each_worker_jsonb()\n\nThe seventh group of functions use CreateTupleDescCopy() to make a copy of\nResultSetInfo->expectedDesc in the per query memory context.\nThis patch does not change any of that. The only difference is that now\nthe check for expectedDesc is separate.\nTheir memory usage is unchanged by this patch.\n\nThe seventh group:\n\n elements_worker_jsonb()\n elements_worker()\n\nNote that for all of the functions in which I added or changed the check for\nResultSetInfo->expectedDesc, it is a little bit different because I've changed\nthe error code.\nI've used error code: ERRCODE_E_R_I_E_SRF_PROTOCOL_VIOLATED where they\npreviously would have been part of an ereport with error code\nERRCODE_FEATURE_NOT_SUPPORTED.\nI'm not sure if the new one is correct. In fact, I'm not sure if an error\nrelated to expectedDesc being set should be user-facing at all (looking at how\nexpectedDesc is used and set, I wasn't sure). Perhaps it should be an elog() or\nan assert?\n\nFeedback is addressed inline below.\n\nOn Tue, Nov 2, 2021 at 2:17 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>\n> On 2021-Nov-02, Melanie Plageman wrote:\n>\n> > Attached is a patch to create a helper function which creates a\n> > tuplestore to be used by FUNCAPI-callable functions.\n>\n> Looks good, and given the amount of code being removed, it seems a\n> no-brainer that some helper(s) is/are needed.\n>\n> I think the name is a bit too generic. How about\n> MakeFuncResultTuplestore()?\n\nI've done this rename.\n\n> I think the function should not modify rsinfo -- having that as\n> (undocumented) side effect is weird. I would suggest to just return the\n> tuplestore, and have the caller stash it in rsinfo->setResult and modify\n> the other struct members alongside that. It's a couple more lines per\n> caller, but the code is clearer that way.\n\nI've changed the helper not to modify rsinfo.\n\n> > There are a few places with very similar code to the new helper\n> > (MakeTuplestore()) but which, for example, don't expect the return type\n> > to be a row (pg_ls_dir(), pg_tablespace_databases(), text_to_table()). I\n> > wasn't sure if it was worth modifying it for their benefit.\n>\n> Is this just because of the tupdesc? If that's the case, then maybe the\n> tupdesc can be optionally passed in by caller, and when it's not, the\n> helper does get_call_result_type() and the tupdesc is used as output\n> argument.\n\nI've parameterized the helper in this way and now am able to use it for\nall FUNCAPI-callable functions making tuplestores except fmgr_sql (for\nSQL functions) in functions.c because it didn't really seem worth it\ngiven the way the code is written.\n\n> > All callers of MakeTuplestore() pass work_mem as the third parameter, so\n> > I could just omit it and hard-code it in, but I wasn't sure that felt\n> > right in a utility/helper function.\n>\n> No opinion. I would have used the global in the function instead of\n> passing it in.\n\nI've changed this to use the global.\n\n> > I inlined deflist_to_tuplestore() in foreign.c since it was billed as a\n> > helper function but wasn't used elsewhere and it made it harder to use\n> > MakeTuplestore() in this location.\n>\n> This is a bit strange. Does it work to pass rsinfo->expectedDesc?\n\nThis is addressed by changing MakeFuncResultTuplestore() call\nget_call_result_type() optionally and having pg_options_to_table() call\nit with NULL as tuple descriptor. Now the behavior is the same -- it\nmakes a copy of expectedDesc in the per query context and sets that as\nsetDesc in the ReturnSetInfo.\n\n- Melanie","msg_date":"Mon, 8 Nov 2021 14:44:57 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Tue, Nov 2, 2021 at 4:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n>\n> Several places have a conditional value for the first argument (randomAccess),\n> but your patch changes the behavior to a constant \"true\". I didn't review the\n> patch beyond that.\n>\n> > @@ -740,18 +724,14 @@ pg_prepared_statement(PG_FUNCTION_ARGS)\n> > - tupstore =\n> > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\n> > - false, work_mem);\n>\n> > @@ -2701,42 +2701,13 @@ pg_hba_file_rules(PG_FUNCTION_ARGS)\n> > - tuple_store =\n> > - tuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\n> > - false, work_mem);\n>\n> > @@ -4799,31 +4797,8 @@ pg_timezone_names(PG_FUNCTION_ARGS)\n> > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\n> > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\n>\n> > @@ -575,38 +575,12 @@ pg_ls_dir_1arg(PG_FUNCTION_ARGS)\n> > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\n> > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\n>\n> > @@ -1170,17 +1154,12 @@ pg_cursor(PG_FUNCTION_ARGS)\n> > - tupstore =\n> > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\n> > - false, work_mem);\n>\n> > +++ b/src/backend/utils/fmgr/funcapi.c\n> > + tupstore = tuplestore_begin_heap(true, false, maxKBytes);\n>\n\nI believe the patch has preserved the same behavior. All of the callers\nfor which I replaced tuplestore_begin_heap() which passed a variable for\nthe randomAccess parameter had set that variable to something which was\neffectively the same as passing true -- SFRM_Materialize_Random.\n\n- Melanie\n\n\n","msg_date":"Mon, 8 Nov 2021 14:52:28 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Mon, Nov 08, 2021 at 02:52:28PM -0500, Melanie Plageman wrote:\n> On Tue, Nov 2, 2021 at 4:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> >\n> > Several places have a conditional value for the first argument (randomAccess),\n> > but your patch changes the behavior to a constant \"true\". I didn't review the\n> > patch beyond that.\n> >\n> > > @@ -740,18 +724,14 @@ pg_prepared_statement(PG_FUNCTION_ARGS)\n> > > - tupstore =\n> > > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\n> > > - false, work_mem);\n> >\n> > > @@ -2701,42 +2701,13 @@ pg_hba_file_rules(PG_FUNCTION_ARGS)\n> > > - tuple_store =\n> > > - tuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\n> > > - false, work_mem);\n> >\n> > > @@ -4799,31 +4797,8 @@ pg_timezone_names(PG_FUNCTION_ARGS)\n> > > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\n> > > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\n> >\n> > > @@ -575,38 +575,12 @@ pg_ls_dir_1arg(PG_FUNCTION_ARGS)\n> > > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\n> > > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\n> >\n> > > @@ -1170,17 +1154,12 @@ pg_cursor(PG_FUNCTION_ARGS)\n> > > - tupstore =\n> > > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\n> > > - false, work_mem);\n> >\n> > > +++ b/src/backend/utils/fmgr/funcapi.c\n> > > + tupstore = tuplestore_begin_heap(true, false, maxKBytes);\n> \n> I believe the patch has preserved the same behavior. All of the callers\n> for which I replaced tuplestore_begin_heap() which passed a variable for\n> the randomAccess parameter had set that variable to something which was\n> effectively the same as passing true -- SFRM_Materialize_Random.\n\nI don't think so ?\n\nThey callers aren't passing SFRM_Materialize_Random, but rather\n(allowedModes & SFRM_Materialize_Random) != 0\n\nWhere allowedModes is determined EXEC_FLAG_BACKWARD.\n\nsrc/include/executor/executor.h:extern Tuplestorestate *ExecMakeTableFunctionResult(SetExprState *setexpr,\nsrc/include/executor/executor.h- ExprContext *econtext,\nsrc/include/executor/executor.h- MemoryContext argContext,\nsrc/include/executor/executor.h- TupleDesc expectedDesc,\nsrc/include/executor/executor.h- bool randomAccess);\n\nsrc/backend/executor/nodeFunctionscan.c=FunctionNext(FunctionScanState *node)\nsrc/backend/executor/nodeFunctionscan.c: ExecMakeTableFunctionResult(node->funcstates[0].setexpr,\nsrc/backend/executor/nodeFunctionscan.c- node->ss.ps.ps_ExprContext,\nsrc/backend/executor/nodeFunctionscan.c- node->argcontext,\nsrc/backend/executor/nodeFunctionscan.c- node->funcstates[0].tupdesc,\nsrc/backend/executor/nodeFunctionscan.c- node->eflags & EXEC_FLAG_BACKWARD);\n\n-- \nJustin\n\n\n","msg_date":"Mon, 8 Nov 2021 14:13:38 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Mon, Nov 8, 2021 at 3:13 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n>\n> On Mon, Nov 08, 2021 at 02:52:28PM -0500, Melanie Plageman wrote:\n> > On Tue, Nov 2, 2021 at 4:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > >\n> > > Several places have a conditional value for the first argument (randomAccess),\n> > > but your patch changes the behavior to a constant \"true\". I didn't review the\n> > > patch beyond that.\n> > >\n> > > > @@ -740,18 +724,14 @@ pg_prepared_statement(PG_FUNCTION_ARGS)\n> > > > - tupstore =\n> > > > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\n> > > > - false, work_mem);\n> > >\n> > > > @@ -2701,42 +2701,13 @@ pg_hba_file_rules(PG_FUNCTION_ARGS)\n> > > > - tuple_store =\n> > > > - tuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\n> > > > - false, work_mem);\n> > >\n> > > > @@ -4799,31 +4797,8 @@ pg_timezone_names(PG_FUNCTION_ARGS)\n> > > > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\n> > > > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\n> > >\n> > > > @@ -575,38 +575,12 @@ pg_ls_dir_1arg(PG_FUNCTION_ARGS)\n> > > > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\n> > > > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\n> > >\n> > > > @@ -1170,17 +1154,12 @@ pg_cursor(PG_FUNCTION_ARGS)\n> > > > - tupstore =\n> > > > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\n> > > > - false, work_mem);\n> > >\n> > > > +++ b/src/backend/utils/fmgr/funcapi.c\n> > > > + tupstore = tuplestore_begin_heap(true, false, maxKBytes);\n> >\n> > I believe the patch has preserved the same behavior. All of the callers\n> > for which I replaced tuplestore_begin_heap() which passed a variable for\n> > the randomAccess parameter had set that variable to something which was\n> > effectively the same as passing true -- SFRM_Materialize_Random.\n>\n> I don't think so ?\n>\n> They callers aren't passing SFRM_Materialize_Random, but rather\n> (allowedModes & SFRM_Materialize_Random) != 0\n>\n> Where allowedModes is determined EXEC_FLAG_BACKWARD.\n>\n> src/include/executor/executor.h:extern Tuplestorestate *ExecMakeTableFunctionResult(SetExprState *setexpr,\n> src/include/executor/executor.h- ExprContext *econtext,\n> src/include/executor/executor.h- MemoryContext argContext,\n> src/include/executor/executor.h- TupleDesc expectedDesc,\n> src/include/executor/executor.h- bool randomAccess);\n>\n> src/backend/executor/nodeFunctionscan.c=FunctionNext(FunctionScanState *node)\n> src/backend/executor/nodeFunctionscan.c: ExecMakeTableFunctionResult(node->funcstates[0].setexpr,\n> src/backend/executor/nodeFunctionscan.c- node->ss.ps.ps_ExprContext,\n> src/backend/executor/nodeFunctionscan.c- node->argcontext,\n> src/backend/executor/nodeFunctionscan.c- node->funcstates[0].tupdesc,\n> src/backend/executor/nodeFunctionscan.c- node->eflags & EXEC_FLAG_BACKWARD);\n>\n\nYou are right. I misread it.\n\nSo, I've attached a patch where randomAccess is now an additional\nparameter (and registered for the next fest).\n\nI was thinking about how to add a test that would have broken when I\npassed true for randomAccess to tuplestore_begin_heap() when false was\nrequired. But, I don't fully understand the problem. If backward\naccesses to a tuplestore are not allowed but randomAccess is mistakenly\npassed as true, would the potential result be potentially wrong results\nfrom accessing the tuplestore results backwards?\n\n- Melanie","msg_date":"Thu, 18 Nov 2021 12:59:03 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Thu, Nov 18, 2021 at 12:59:03PM -0500, Melanie Plageman wrote:\n> On Mon, Nov 8, 2021 at 3:13 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > On Mon, Nov 08, 2021 at 02:52:28PM -0500, Melanie Plageman wrote:\n> > > On Tue, Nov 2, 2021 at 4:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > > >\n> > > > Several places have a conditional value for the first argument (randomAccess),\n> > > > but your patch changes the behavior to a constant \"true\". I didn't review the\n> > > > patch beyond that.\n> > > >\n...\n> > > I believe the patch has preserved the same behavior. All of the callers\n> > > for which I replaced tuplestore_begin_heap() which passed a variable for\n> > > the randomAccess parameter had set that variable to something which was\n> > > effectively the same as passing true -- SFRM_Materialize_Random.\n> >\n> > I don't think so ?\n> >\n> > They callers aren't passing SFRM_Materialize_Random, but rather\n> > (allowedModes & SFRM_Materialize_Random) != 0\n> >\n> > Where allowedModes is determined EXEC_FLAG_BACKWARD.\n...\n> You are right. I misread it.\n> \n> So, I've attached a patch where randomAccess is now an additional\n> parameter (and registered for the next fest).\n> \n> I was thinking about how to add a test that would have broken when I\n> passed true for randomAccess to tuplestore_begin_heap() when false was\n> required. But, I don't fully understand the problem. If backward\n> accesses to a tuplestore are not allowed but randomAccess is mistakenly\n> passed as true, would the potential result be potentially wrong results\n> from accessing the tuplestore results backwards?\n\nNo, see here\n\nsrc/backend/utils/fmgr/README-The Tuplestore must be created with randomAccess = true if\nsrc/backend/utils/fmgr/README:SFRM_Materialize_Random is set in allowedModes, but it can (and preferably\nsrc/backend/utils/fmgr/README-should) be created with randomAccess = false if not. Callers that can support\nsrc/backend/utils/fmgr/README-both ValuePerCall and Materialize mode will set SFRM_Materialize_Preferred,\nsrc/backend/utils/fmgr/README-or not, depending on which mode they prefer.\n\nIf you use \"randomAccess=true\" when it's not needed, the result might be less\nefficient.\n\nSome callers specify \"true\" when they don't need to. I ran into that here.\nhttps://www.postgresql.org/message-id/21724.1583955158@sss.pgh.pa.us\n\nMaybe you'd want to add an 0002 patch which changes those to conditional ?\n\nBTW, there should be a newline before MakeFuncResultTuplestore().\n\n-- \nJustin\n\n\n","msg_date":"Thu, 18 Nov 2021 12:24:30 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Thu, Nov 18, 2021 at 1:24 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n>\n> On Thu, Nov 18, 2021 at 12:59:03PM -0500, Melanie Plageman wrote:\n> > On Mon, Nov 8, 2021 at 3:13 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > > On Mon, Nov 08, 2021 at 02:52:28PM -0500, Melanie Plageman wrote:\n> > > > On Tue, Nov 2, 2021 at 4:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > > > >\n> > > > > Several places have a conditional value for the first argument (randomAccess),\n> > > > > but your patch changes the behavior to a constant \"true\". I didn't review the\n> > > > > patch beyond that.\n> > > > >\n> ...\n> > > > I believe the patch has preserved the same behavior. All of the callers\n> > > > for which I replaced tuplestore_begin_heap() which passed a variable for\n> > > > the randomAccess parameter had set that variable to something which was\n> > > > effectively the same as passing true -- SFRM_Materialize_Random.\n> > >\n> > > I don't think so ?\n> > >\n> > > They callers aren't passing SFRM_Materialize_Random, but rather\n> > > (allowedModes & SFRM_Materialize_Random) != 0\n> > >\n> > > Where allowedModes is determined EXEC_FLAG_BACKWARD.\n> ...\n> > You are right. I misread it.\n> >\n> > So, I've attached a patch where randomAccess is now an additional\n> > parameter (and registered for the next fest).\n> >\n> > I was thinking about how to add a test that would have broken when I\n> > passed true for randomAccess to tuplestore_begin_heap() when false was\n> > required. But, I don't fully understand the problem. If backward\n> > accesses to a tuplestore are not allowed but randomAccess is mistakenly\n> > passed as true, would the potential result be potentially wrong results\n> > from accessing the tuplestore results backwards?\n>\n> No, see here\n>\n> src/backend/utils/fmgr/README-The Tuplestore must be created with randomAccess = true if\n> src/backend/utils/fmgr/README:SFRM_Materialize_Random is set in allowedModes, but it can (and preferably\n> src/backend/utils/fmgr/README-should) be created with randomAccess = false if not. Callers that can support\n> src/backend/utils/fmgr/README-both ValuePerCall and Materialize mode will set SFRM_Materialize_Preferred,\n> src/backend/utils/fmgr/README-or not, depending on which mode they prefer.\n>\n> If you use \"randomAccess=true\" when it's not needed, the result might be less\n> efficient.\n>\n> Some callers specify \"true\" when they don't need to. I ran into that here.\n> https://www.postgresql.org/message-id/21724.1583955158@sss.pgh.pa.us\n>\n> Maybe you'd want to add an 0002 patch which changes those to conditional ?\n\nDone in attached v4\n\n> BTW, there should be a newline before MakeFuncResultTuplestore().\n\nFixed\n\n- Melanie","msg_date":"Mon, 13 Dec 2021 17:27:52 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Mon, Dec 13, 2021 at 05:27:52PM -0500, Melanie Plageman wrote:\n> Done in attached v4\n\nThanks.\n\nI think these comments can be removed from the callers:\n/* it's a simple type, so don't use get_call_result_type */\n\nYou removed one call to tuplestore_donestoring(), but not the others.\nI guess you could remove them all (optionally as a separate patch).\n\nThe rest of these are questions more than comments, and I'm not necessarily\nsuggesting to change the patch:\n\nThis isn't new in your patch, but for me, it's more clear if the callers have a\nseparate boolean for randomAccess, rather than having the long expression in\nthe function call:\n\n+ tupstore = MakeFuncResultTuplestore(fcinfo, NULL,\n+ rsinfo->allowedModes & SFRM_Materialize_Random);\n\nvs\n\n randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\n- tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\n+ tupstore = MakeFuncResultTuplestore(fcinfo, NULL, randomAccess);\n\nOne could also argue for passing rsinfo->allowedModes, instead of\n(rsinfo->allowedModes & SFRM_Materialize_Random).\n\nThere's a couples places that you're checking expectedDesc where it wasn't\nbeing checked before. Is that some kind of live bug ?\npg_config() text_to_table()\n\nIt'd be nice if the \"expected tuple format\" error didn't need to be duplicated\nfor each SRFs. I suppose the helper function could just take a boolean\ndetermining whether or not to throw an error (passing \"expectedDesc==NULL\").\nYou'd have to call the helper before CreateTupleDescCopy() - which you're\nalready doing in a couple places for similar reasons.\n\nI noticed that tuplestore.h isn't included most places. I suppose it's\nincluded via execnodes.h. Your patch doesn't change that, arguably it\nshould've always been included.\n\n-- \nJustin\n\n\n","msg_date":"Fri, 17 Dec 2021 14:04:19 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"Thanks for the detailed review!\n\nOn Fri, Dec 17, 2021 at 3:04 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n>\n> On Mon, Dec 13, 2021 at 05:27:52PM -0500, Melanie Plageman wrote:\n> > Done in attached v4\n>\n> Thanks.\n>\n> I think these comments can be removed from the callers:\n> /* it's a simple type, so don't use get_call_result_type */\n\nDone in attached v5\n\n> You removed one call to tuplestore_donestoring(), but not the others.\n> I guess you could remove them all (optionally as a separate patch).\n\nI removed it in that one location because I wanted to get rid of the\nlocal variable it used. I am fine with removing the other occurrences,\nbut I thought there might be some reason why instead of removing it,\nthey made it into a no-op.\n\n> The rest of these are questions more than comments, and I'm not necessarily\n> suggesting to change the patch:\n>\n> This isn't new in your patch, but for me, it's more clear if the callers have a\n> separate boolean for randomAccess, rather than having the long expression in\n> the function call:\n>\n> + tupstore = MakeFuncResultTuplestore(fcinfo, NULL,\n> + rsinfo->allowedModes & SFRM_Materialize_Random);\n>\n> vs\n>\n> randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\n> - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\n> + tupstore = MakeFuncResultTuplestore(fcinfo, NULL, randomAccess);\n>\n> One could also argue for passing rsinfo->allowedModes, instead of\n> (rsinfo->allowedModes & SFRM_Materialize_Random).\n\nI've changed the patch to have MakeFuncResultTuplestore() check\nrsinfo->allowedModes and pass in the resulting boolean to\ntuplestore_begin_heap(). Because I modified the\nMakeFuncResultTuplestore() function signature to accommodate this, I had\nto collapse the two patches into one, as I couldn't maintain the call\nsites which passed in a constant.\n\n> There's a couples places that you're checking expectedDesc where it wasn't\n> being checked before. Is that some kind of live bug ?\n> pg_config() text_to_table()\n\nYes, expectedDesc was accessed in these locations without checking that\nit is not NULL. Should that be a separate patch?\n\n> It'd be nice if the \"expected tuple format\" error didn't need to be duplicated\n> for each SRFs. I suppose the helper function could just take a boolean\n> determining whether or not to throw an error (passing \"expectedDesc==NULL\").\n> You'd have to call the helper before CreateTupleDescCopy() - which you're\n> already doing in a couple places for similar reasons.\n\nSo, I don't think it makes sense to move that error message into\nMakeFuncResultTuplestore() for the cases in which NULL is passed in for\nthe tuple descriptor. expectedDesc is not accessed at all in these cases\ninside the function (since get_call_result_type()) is not called. For\nthe cases when a tuple descriptor is passed in, only two callers\nactually check for expectedDesc -- each_worker_jsonb and each_worker.\nTherefore, I don't think it makes sense to add another parameter just to\nmove that check inside for two callers.\n\n> I noticed that tuplestore.h isn't included most places. I suppose it's\n> included via execnodes.h. Your patch doesn't change that, arguably it\n> should've always been included.\n\nI've now included it in funcapi.c.\n\n- Melanie","msg_date":"Wed, 5 Jan 2022 12:09:16 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Wed, Jan 05, 2022 at 12:09:16PM -0500, Melanie Plageman wrote:\n> On Fri, Dec 17, 2021 at 3:04 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > The rest of these are questions more than comments, and I'm not necessarily\n> > suggesting to change the patch:\n> >\n> > This isn't new in your patch, but for me, it's more clear if the callers have a\n> > separate boolean for randomAccess, rather than having the long expression in\n> > the function call:\n> >\n> > + tupstore = MakeFuncResultTuplestore(fcinfo, NULL,\n> > + rsinfo->allowedModes & SFRM_Materialize_Random);\n> >\n> > vs\n> >\n> > randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\n> > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\n> > + tupstore = MakeFuncResultTuplestore(fcinfo, NULL, randomAccess);\n> >\n> > One could also argue for passing rsinfo->allowedModes, instead of\n> > (rsinfo->allowedModes & SFRM_Materialize_Random).\n> \n> I've changed the patch to have MakeFuncResultTuplestore() check\n> rsinfo->allowedModes and pass in the resulting boolean to\n> tuplestore_begin_heap(). Because I modified the\n> MakeFuncResultTuplestore() function signature to accommodate this, I had\n> to collapse the two patches into one, as I couldn't maintain the call\n> sites which passed in a constant.\n\n> > There's a couples places that you're checking expectedDesc where it wasn't\n> > being checked before. Is that some kind of live bug ?\n> > pg_config() text_to_table()\n> \n> Yes, expectedDesc was accessed in these locations without checking that\n> it is not NULL. Should that be a separate patch?\n\nI don't know. The patch is already easy to review, since it's mostly limited\nto removing code and fixing inconsistencies (NULL check) and possible\ninefficiencies (randomAccess).\n\nIf the expectedDesc NULL check were an 0001 patch, then 0002 (the main patch)\nwould be even easier to review. Only foreign.c is different.\n\n> > You removed one call to tuplestore_donestoring(), but not the others.\n> > I guess you could remove them all (optionally as a separate patch).\n> \n> I removed it in that one location because I wanted to get rid of the\n> local variable it used.\n\nWhat local variable ? I see now that logicalfuncs.c never had a local variable\ncalled tupstore. I'm sure it was intended since 2014 (b89e15105) to say\ntupestore_donestoring(p->tupstore). But donestoring(typo) causes no error,\nsince the define is a NOP.\n\nsrc/include/utils/tuplestore.h:#define tuplestore_donestoring(state) ((void) 0)\n\n> I am fine with removing the other occurrences,\n> but I thought there might be some reason why instead of removing it,\n> they made it into a no-op.\n\nI assume Tom left it (actually, added it back in dd04e958c) to avoid breaking\nextensions for no reason. And maybe to preserve the possbility of at some\npoint in the future doing something again during the \"done\" step.\n\nI'd leave it for input from a committer about those:\n\n - remove tuplestore_donestoring() calls ?\n - split expectedDesc NULL check to an 0001 patch ?\n - anything other opened questions ?\n\nI'm marking this as RFC, with the caveat that the newline before\nMakeFuncResultTuplestore went missing again :)\n\n-- \nJustin\n\n\n","msg_date":"Wed, 5 Jan 2022 18:57:48 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Wed, Jan 5, 2022 at 7:57 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n>\n> On Wed, Jan 05, 2022 at 12:09:16PM -0500, Melanie Plageman wrote:\n> > On Fri, Dec 17, 2021 at 3:04 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > > There's a couples places that you're checking expectedDesc where it wasn't\n> > > being checked before. Is that some kind of live bug ?\n> > > pg_config() text_to_table()\n> >\n> > Yes, expectedDesc was accessed in these locations without checking that\n> > it is not NULL. Should that be a separate patch?\n>\n> I don't know. The patch is already easy to review, since it's mostly limited\n> to removing code and fixing inconsistencies (NULL check) and possible\n> inefficiencies (randomAccess).\n>\n> If the expectedDesc NULL check were an 0001 patch, then 0002 (the main patch)\n> would be even easier to review. Only foreign.c is different.\n\nI'll wait to do that if preferred by committer.\nAre you imagining that patch 0001 would only add the check for\nexpectedDesc that is missing from pg_config() and text_to_table()?\n\n> > > You removed one call to tuplestore_donestoring(), but not the others.\n> > > I guess you could remove them all (optionally as a separate patch).\n> >\n> > I removed it in that one location because I wanted to get rid of the\n> > local variable it used.\n>\n> What local variable ? I see now that logicalfuncs.c never had a local variable\n> called tupstore. I'm sure it was intended since 2014 (b89e15105) to say\n> tupestore_donestoring(p->tupstore). But donestoring(typo) causes no error,\n> since the define is a NOP.\n>\n> src/include/utils/tuplestore.h:#define tuplestore_donestoring(state) ((void) 0)\n\nYes, I mean the local variable, tupstore.\n\n> > I am fine with removing the other occurrences,\n> > but I thought there might be some reason why instead of removing it,\n> > they made it into a no-op.\n>\n> I assume Tom left it (actually, added it back in dd04e958c) to avoid breaking\n> extensions for no reason. And maybe to preserve the possbility of at some\n> point in the future doing something again during the \"done\" step.\n>\n> I'd leave it for input from a committer about those:\n>\n> - remove tuplestore_donestoring() calls ?\n> - split expectedDesc NULL check to an 0001 patch ?\n> - anything other opened questions ?\n>\n> I'm marking this as RFC, with the caveat that the newline before\n> MakeFuncResultTuplestore went missing again :)\n\noops. I've attached v6 with the newline.\n\n- Melanie","msg_date":"Tue, 11 Jan 2022 10:19:33 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Tue, Jan 11, 2022 at 10:19:33AM -0500, Melanie Plageman wrote:\n> On Wed, Jan 5, 2022 at 7:57 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n>> I'd leave it for input from a committer about those:\n>>\n>> - remove tuplestore_donestoring() calls ?\n\nThis part has been raised on a different thread, as of:\nhttps://www.postgresql.org/message-id/CAP0=ZVJeeYfAeRfmzqAF2Lumdiv4S4FewyBnZd4DPTrsSQKJKw@mail.gmail.com\n\nAnd it looks sensible from here to just remove any traces of this\nstuff from the C code, while leaving the compatibility macro around\nto avoid any breakages with any external modules.\n\n>> - split expectedDesc NULL check to an 0001 patch ?\n\nI am not sure that this is worth a split, as the same areas are\ntouched.\n\n>> - anything other opened questions ?\n>>\n>> I'm marking this as RFC, with the caveat that the newline before\n>> MakeFuncResultTuplestore went missing again :)\n> \n> oops. I've attached v6 with the newline.\n\nSaying that, the wrapper you are proposing here looks like a good\nidea. I have wanted a centralized wrapper more than once, and this is\na nice cut overall.\n--\nMichael","msg_date":"Wed, 16 Feb 2022 14:43:08 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Tue, Jan 11, 2022 at 10:19:33AM -0500, Melanie Plageman wrote:\n> > If the expectedDesc NULL check were an 0001 patch, then 0002 (the main patch)\n> > would be even easier to review. Only foreign.c is different.\n> \n> I'll wait to do that if preferred by committer.\n> Are you imagining that patch 0001 would only add the check for\n> expectedDesc that is missing from pg_config() and text_to_table()?\n\nYes, that's what I meant, but I haven't been able to determine if a NULL check\nshould be added at all.\n\nI found this commit message, which indicates that it shouldn't be used in every\ncase. I'm not clear when it should/not be used, though.\n\ncommit 60651e4cddbb77e8f1a0c7fc0be6a7e7bf626fe0\nAuthor: Tom Lane <tgl@sss.pgh.pa.us>\nDate: Sat Oct 28 14:02:21 2017 -0400\n\n Support domains over composite types in PL/Perl.\n \n In passing, don't insist on rsi->expectedDesc being set unless we\n actually need it; this allows succeeding in a couple of cases where\n PL/Perl functions returning setof composite would have failed before,\n and makes the error message more apropos in other cases.\n\n\n","msg_date":"Tue, 15 Feb 2022 23:57:56 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Tue, Feb 15, 2022 at 11:57:56PM -0600, Justin Pryzby wrote:\n> On Tue, Jan 11, 2022 at 10:19:33AM -0500, Melanie Plageman wrote:\n>> I'll wait to do that if preferred by committer.\n>> Are you imagining that patch 0001 would only add the check for\n>> expectedDesc that is missing from pg_config() and text_to_table()?\n\nIt took me some time to follow up on this point, but we clearly don't\nexpect a NULL expectedDesc in the context of these two code paths or\nwe finish with a NULL pointer dereference in CreateTupleDescCopy()\nwith a crash, so they had better be added. It also makes sense to me\nto do that first and independently of the rest of the patch: it is a\nseparate issue.\n\n> Yes, that's what I meant, but I haven't been able to determine if a NULL check\n> should be added at all.\n\n+ (errcode(ERRCODE_E_R_I_E_SRF_PROTOCOL_VIOLATED),\n+ errmsg(\"expected tuple format not specified as required for \"\n+ \"set-returning function.\")))\nerrmsgs should be in a single line, to ease the grepping of similar\npatterns.\n\nThere is something else that I think could be cleaned up, while we are\nworking on this area. Some of the callers of\nMakeFuncResultTuplestore() (prepare.c, portalmem.c, one in genfile.c,\nguc.c, misc.c) pass down a NULL tupdesc while building their TupleDesc\nusing CreateTemplateTupleDesc() and a set of TupleDescInitEntry() for\neach attribute. This data exists already in pg_proc.dat, duplicating\nthe effort. Except if I am missing something, wouldn't it be better\nto let get_call_result_type() do this work for us, passing a TupleDesc\npointer to the new MakeFuncResultTuplestore() rather than NULL?\n\nOnce you do that, there is something else that stands out: all the\ncallers of MakeFuncResultTuplestore() passing NULL for tupledesc *have\nto* set expectedDesc, as far as I can see from the remaining callers.\nThis brings an interesting effect, couldn't we move the check on\nReturnSetInfo->expectedDesc being not NULL within\nMakeFuncResultTuplestore(), checking after it only when tupdesc ==\nNULL?\n\nOne would say that this changes the error ordering in code paths like\npg_config(), but, actually, in this case, we don't have any need for\nthe part \"Check to make sure we have a reasonable tuple descriptor\" as\nof pg_proc.dat defining the tuple description this function uses, no?\nAnd we could switch to a tuplestore in this case, as well, reducing by\none the numbers of callers with tupdesc=NULL for the new routine.\nThis could be a separate cleanup, separate from the rest, done first.\n\neach_worker_jsonb() does not really need its check on expectedDesc\nbeing NULL, does it? HEAD grabs the tuple descriptor with\nget_call_result_type(), the patch uses MakeFuncResultTuplestore(), so\nwe make nowhere use of the expected TupleDesc saved by the SRF\nexecutor.\n\nAsserting that we are in the correct memory context in when calling\nMakeFuncResultTuplestore() sounds rather sensible from here as per the\nmagics done in the various json functions. Still, it really feels\nlike we could do a more centralized consolidation of the whole.\n--\nMichael","msg_date":"Thu, 17 Feb 2022 16:10:01 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Thu, Feb 17, 2022 at 04:10:01PM +0900, Michael Paquier wrote:\n> Asserting that we are in the correct memory context in when calling\n> MakeFuncResultTuplestore() sounds rather sensible from here as per the\n> magics done in the various json functions. Still, it really feels\n> like we could do a more centralized consolidation of the whole.\n\nSo, I got my hands on this area, and found myself applying 07daca5 as\na first piece of the puzzle. Anyway, after more review today, I have\nbumped into more pieces that could be consolidated, and finished with\nthe following patch set:\n- 0001 changes a couple of SRFs that I found worth simplifying. These\nrefer mostly to the second and fourth group mentioned upthread by\nMelanie, with two exceptions: pg_tablespace_databases() where it is\nnot worth changing to keep it backward-compatible and pg_ls_dir() as\nper its one-arg version. That's a nice first cut in itself.\n- 0002 changes a couple of places to unify some existing SRF checks,\nthat I bumped into on the way. The value here is in using the same\nerror messages everywhere, reducing the translation effort and\ngenerating the same errors for all cases based on the same conditions.\nThis eases the review of the next patch.\n- 0003 is the actual refactoring meat, where I have been able to move\nthe check on expectedDesc into MakeFuncResultTuplestore(), shaving\nmore code than previously. If you discard the cases of patch 0001, it\nshould actually be possible to set setResult, setDesc and returnMode\nwithin the new function, which would feel natural as that's the place\nwhere we create the function's tuplestore and we want to materialize\nits contents. The cases with the JSON code are also a bit hairy and\nneed more thoughts, but we could also cut this part of the code from\nthe initial refactoring effort.\n\nSo, for now, 0001 and 0002 look like rather committable pieces. 0003\nneeds a bit more thoughts about all the corner cases we need to\nconsider, mostly what I am mentioning above. Perhaps the pieces of\n0003 related to pg_options_to_table() should be moved to 0001 as a\nmatter of clarity.\n--\nMichael","msg_date":"Mon, 21 Feb 2022 16:41:17 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Mon, Feb 21, 2022 at 04:41:17PM +0900, Michael Paquier wrote:\n> So, I got my hands on this area, and found myself applying 07daca5 as\n> a first piece of the puzzle. Anyway, after more review today, I have\n> bumped into more pieces that could be consolidated, and finished with\n> the following patch set:\n> - 0001 changes a couple of SRFs that I found worth simplifying. These\n> refer mostly to the second and fourth group mentioned upthread by\n> Melanie, with two exceptions: pg_tablespace_databases() where it is\n> not worth changing to keep it backward-compatible and pg_ls_dir() as\n> per its one-arg version. That's a nice first cut in itself.\n> - 0002 changes a couple of places to unify some existing SRF checks,\n> that I bumped into on the way. The value here is in using the same\n> error messages everywhere, reducing the translation effort and\n> generating the same errors for all cases based on the same conditions.\n> This eases the review of the next patch.\n\nThese two have been now applied, with some comment improvements and\nthe cleanup of pg_options_to_table() done in 0001, and a slight change\nin 0002 for pageinspect where a check was not necessary for a BRIN\ncode path.\n\n> - 0003 is the actual refactoring meat, where I have been able to move\n> the check on expectedDesc into MakeFuncResultTuplestore(), shaving\n> more code than previously. If you discard the cases of patch 0001, it\n> should actually be possible to set setResult, setDesc and returnMode\n> within the new function, which would feel natural as that's the place\n> where we create the function's tuplestore and we want to materialize\n> its contents. The cases with the JSON code are also a bit hairy and\n> need more thoughts, but we could also cut this part of the code from\n> the initial refactoring effort.\n\nThis is the remaining piece, as attached, that I have not been able to\npoke much at yet.\n--\nMichael","msg_date":"Thu, 24 Feb 2022 20:25:06 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Thu, Feb 24, 2022 at 08:25:06PM +0900, Michael Paquier wrote:\n> This is the remaining piece, as attached, that I have not been able to\n> poke much at yet.\n\nSo, I have finally poked at this last part of the patch set, and I\nfound that we can be more aggressive with the refactoring, by moving\ninto MakeFuncResultTuplestore() the parts where we save the tuplestore\nand the tupledesc in the per-query memory context. There are two\npieces that matter once things are reshaped:\n- The tuple descriptor may need some extra validation via\nBlessTupleDesc() when it comes from a transient record datatype,\nsomething that happens for most of the subroutines related to the JSON\nfunctions.\n- expectedDesc is sometimes required by the caller, though most of the\ntime just needs to be built with the more expensive\nget_call_result_type().\n\nIn order to keep things pluggable at will, MakeFuncResultTuplestore()\nhas been changed to access a set of bits32 flags, able to control the\ntwo options above. With this facility in place, I have been able to\ncut much more code than the initial patch, roughly twice as of:\n 24 files changed, 157 insertions(+), 893 deletions(-)\n\nThis seems in rather good shape to me, the changes are\nstraight-forward and the code cut is really good, so I'd like to move\non with that. 0001 is the initial patch, and 0002 is the extra\nrefactoring I have been working on. The plan would be to merge both,\nbut I am sending a split to ease any checks on what I have changed.\n\nComments or objections?\n--\nMichael","msg_date":"Mon, 28 Feb 2022 16:49:41 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Mon, Feb 28, 2022 at 04:49:41PM +0900, Michael Paquier wrote:\n> In order to keep things pluggable at will, MakeFuncResultTuplestore()\n> has been changed to access a set of bits32 flags, able to control the\n> two options above. With this facility in place, I have been able to\n> cut much more code than the initial patch, roughly twice as of:\n> 24 files changed, 157 insertions(+), 893 deletions(-)\n\nSo, I have been thinking about this patch once again, and after\npondering more on it, MakeFuncResultTuplestore() is actually a wrong\nname now that it does much more than just creating a tuplestore, by\nassigning the TupleDesc and the TupleStore into the function's\nReturnSetInfo.\n\nThis is actually setting up a function in the context of a single call\nwhere we fill the tuplestore with all its values, so instead I have\nsettled down to name that SetSingleFuncCall(), to make a parallel with\nthe existing MultiFuncCall*(). funcapi.c is the right place for\nthat, and I have added more documentation in the fmgr's README as well\nas funcapi.h.\n--\nMichael","msg_date":"Wed, 2 Mar 2022 15:43:17 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Wed, Mar 02, 2022 at 03:43:17PM +0900, Michael Paquier wrote:\n> This is actually setting up a function in the context of a single call\n> where we fill the tuplestore with all its values, so instead I have\n> settled down to name that SetSingleFuncCall(), to make a parallel with\n> the existing MultiFuncCall*(). funcapi.c is the right place for\n> that, and I have added more documentation in the fmgr's README as well\n> as funcapi.h.\n\nI have tortured all those code paths for the last couple of days, and\nthe new function name, as well as its options, still seemed fine to\nme, so I have applied the patch. The buildfarm is cool with it. It\nis worth noting that there are more code paths in contrib/ that could\nbe simplified with this new routine.\n--\nMichael","msg_date":"Mon, 7 Mar 2022 11:29:03 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Sun, Mar 6, 2022 at 9:29 PM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Wed, Mar 02, 2022 at 03:43:17PM +0900, Michael Paquier wrote:\n> > This is actually setting up a function in the context of a single call\n> > where we fill the tuplestore with all its values, so instead I have\n> > settled down to name that SetSingleFuncCall(), to make a parallel with\n> > the existing MultiFuncCall*(). funcapi.c is the right place for\n> > that, and I have added more documentation in the fmgr's README as well\n> > as funcapi.h.\n>\n> I have tortured all those code paths for the last couple of days, and\n> the new function name, as well as its options, still seemed fine to\n> me, so I have applied the patch. The buildfarm is cool with it. It\n> is worth noting that there are more code paths in contrib/ that could\n> be simplified with this new routine.\n\nWow! Thanks so much for taking the time to review, refine, and commit\nthis work.\n\nI've attached a patch using the helper in most locations in contrib\nmodules that seemed useful.\n\nThe following I don't think we can use the helper or it is not worth it:\n\n- pg_logdir_ls() in contrib/adminpack has return type TYPEFUNC_RECORD\n and expectedDesc is not already created, so the helper can't be used.\n\n But basically, since it doesn't specify OUT argument names, it has to\n do TupleDescInitEntry() itself anyway, I think.\n\n- contrib/tablefunc.c was also not simple to refactor. the various parts\n of SetSingleFuncCall are spread throughout different functions in the\n file.\n\n- contrib/dblink has one function which returns a tuplestore that was\n simple to change (dblink_get_notify()) and I've done so.\n\n However, most of the other creation of tuplestore and tupledescriptors\n is in helper functions (like materializeResult()) which separate the\n tuplestore creation from the tuple descriptor initialization in a way\n that made it hard to do a drop-in replacement with the helper function.\n\n- Melanie","msg_date":"Mon, 7 Mar 2022 17:09:19 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"Re: make tuplestore helper function"},{"msg_contents":"On Mon, Mar 07, 2022 at 05:09:19PM -0500, Melanie Plageman wrote:\n> I've attached a patch using the helper in most locations in contrib\n> modules that seemed useful.\n\nThanks for the patch. I was also looking at that yesterday, and this\npretty much maps to what I have finished with, except for dblink and\nxml2 where it was not looking that obvious to me at quick glance.\n\nIn xml2, my eyes hurt a bit on the meaning of doing the \"Switch out of\nSPI context\" in xpath_table() after doing the two SPI calls, but I\nagree that this extra switch back to the old context should not be\nnecessary. For dblink, I did not notice that the change for\ndblink_get_notify() would be this straight-forward, good catch. It\nwould be nice to see prepTuplestoreResult() gone at the end, though I\nam not sure how invasive/worth it would be for dblink/.\n\n> - pg_logdir_ls() in contrib/adminpack has return type TYPEFUNC_RECORD\n> and expectedDesc is not already created, so the helper can't be used.\n> But basically, since it doesn't specify OUT argument names, it has to\n> do TupleDescInitEntry() itself anyway, I think.\n\nHere, actually, I thought first that a simplification should be\npossible, noticing that the regression tests passed with the function\ncalled until I saw what it was testing :)\n\nI am fine to not poke at the beast, and it looks like we may run into\ntrouble if we wish to maintain compatibility down to adminpack 1.0 at\nruntime. This function has a very limited usage, anyway.\n\n> - contrib/tablefunc.c was also not simple to refactor. the various parts\n> of SetSingleFuncCall are spread throughout different functions in the\n> file.\n> \n> - contrib/dblink has one function which returns a tuplestore that was\n> simple to change (dblink_get_notify()) and I've done so.\n\nThis matches my impression, so applied.\n--\nMichael","msg_date":"Tue, 8 Mar 2022 10:17:03 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: make tuplestore helper function"}],"string":"[\n {\n \"msg_contents\": \"Attached is a patch to create a helper function which creates a\\ntuplestore to be used by FUNCAPI-callable functions.\\n\\nIt was suggested in [1] that I start a separate thread for this patch.\\n\\nA few notes:\\n\\nThere are a few places with very similar code to the new helper\\n(MakeTuplestore()) but which, for example, don't expect the return type\\nto be a row (pg_ls_dir(), pg_tablespace_databases(), text_to_table()). I\\nwasn't sure if it was worth modifying it for their benefit.\\n\\nI wasn't sure if the elog() check for call result type should be changed\\nto an ereport().\\n\\nAll callers of MakeTuplestore() pass work_mem as the third parameter, so\\nI could just omit it and hard-code it in, but I wasn't sure that felt\\nright in a utility/helper function.\\n\\nI inlined deflist_to_tuplestore() in foreign.c since it was billed as a\\nhelper function but wasn't used elsewhere and it made it harder to use\\nMakeTuplestore() in this location.\\n\\n-- melanie\\n\\n[1] https://www.postgresql.org/message-id/flat/20200124195226.lth52iydq2n2uilq%40alap3.anarazel.de\",\n \"msg_date\": \"Tue, 2 Nov 2021 12:34:30 -0400\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-02, Melanie Plageman wrote:\\n\\n> Attached is a patch to create a helper function which creates a\\n> tuplestore to be used by FUNCAPI-callable functions.\\n\\nLooks good, and given the amount of code being removed, it seems a\\nno-brainer that some helper(s) is/are needed.\\n\\nI think the name is a bit too generic. How about\\nMakeFuncResultTuplestore()?\\n\\nI think the function should not modify rsinfo -- having that as\\n(undocumented) side effect is weird. I would suggest to just return the\\ntuplestore, and have the caller stash it in rsinfo->setResult and modify\\nthe other struct members alongside that. It's a couple more lines per\\ncaller, but the code is clearer that way.\\n\\n> There are a few places with very similar code to the new helper\\n> (MakeTuplestore()) but which, for example, don't expect the return type\\n> to be a row (pg_ls_dir(), pg_tablespace_databases(), text_to_table()). I\\n> wasn't sure if it was worth modifying it for their benefit.\\n\\nIs this just because of the tupdesc? If that's the case, then maybe the\\ntupdesc can be optionally passed in by caller, and when it's not, the\\nhelper does get_call_result_type() and the tupdesc is used as output\\nargument.\\n\\n> I wasn't sure if the elog() check for call result type should be changed\\n> to an ereport().\\n\\nI don't see a reason for that: it's still an internal (not user-facing) error.\\n\\n> All callers of MakeTuplestore() pass work_mem as the third parameter, so\\n> I could just omit it and hard-code it in, but I wasn't sure that felt\\n> right in a utility/helper function.\\n\\nNo opinion. I would have used the global in the function instead of\\npassing it in.\\n\\n> I inlined deflist_to_tuplestore() in foreign.c since it was billed as a\\n> helper function but wasn't used elsewhere and it made it harder to use\\n> MakeTuplestore() in this location.\\n\\nThis is a bit strange. Does it work to pass rsinfo->expectedDesc?\\n\\n-- \\nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\\n\\\"No renuncies a nada. No te aferres a nada.\\\"\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Nov 2021 15:17:05 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"Several places have a conditional value for the first argument (randomAccess),\\nbut your patch changes the behavior to a constant \\\"true\\\". I didn't review the\\npatch beyond that.\\n\\n> @@ -740,18 +724,14 @@ pg_prepared_statement(PG_FUNCTION_ARGS)\\n> -\\ttupstore =\\n> -\\t\\ttuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\\n> -\\t\\t\\t\\t\\t\\t\\t false, work_mem);\\n\\n> @@ -2701,42 +2701,13 @@ pg_hba_file_rules(PG_FUNCTION_ARGS)\\n> -\\ttuple_store =\\n> -\\t\\ttuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\\n> -\\t\\t\\t\\t\\t\\t\\t false, work_mem);\\n\\n> @@ -4799,31 +4797,8 @@ pg_timezone_names(PG_FUNCTION_ARGS)\\n> -\\trandomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\\n> -\\ttupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\\n\\n> @@ -575,38 +575,12 @@ pg_ls_dir_1arg(PG_FUNCTION_ARGS)\\n> -\\trandomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\\n> -\\ttupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\\n\\n> @@ -1170,17 +1154,12 @@ pg_cursor(PG_FUNCTION_ARGS)\\n> -\\ttupstore =\\n> -\\t\\ttuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\\n> -\\t\\t\\t\\t\\t\\t\\t false, work_mem);\\n\\n> +++ b/src/backend/utils/fmgr/funcapi.c\\n> +\\ttupstore = tuplestore_begin_heap(true, false, maxKBytes);\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Nov 2021 15:23:53 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"The attached patch contains a helper now named\\nMakeFuncResultTuplestore() which aims to eliminate a few lines of\\nredundant code that all FUNCAPI functions making tuplestores repeated.\\n\\nWhile investigating all of these call sites and making the changes\\nsuggested by Álvaro, I noticed that the functions making tuplestores\\nbroke out into seven groups. For some of them, using my helper changes\\nwhat memory context the tuple descriptor is made in.\\n\\nExplanation of how/why memory context tuple descriptor is in has\\nchanged:\\n\\nAn earlier version of this patch changed into the per query memory\\ncontext inside of the helper function. This version does not do so but\\ninstead asserts that the caller has changed into the appropriate memory\\ncontext.\\n\\nFor callers which previously did not call get_call_result_type() and\\ninstead used a function like CreateTupleDescCopy() to create the tuple\\ndescriptor which was used for the tuples in the tuplestore, most of them\\nmade this copy in the per query memory context. For these callers it\\nseemed error-prone to switch to the per query memory context to create\\nthe copied tuple descriptor and then switch back before calling the\\nhelper function. Instead, the helper function will assert that the\\ncaller has already switched into the per query memory context so that\\nthe tuplestore is made in the per query context.\\n\\nThis means that for the group of functions which, prior to this patch,\\ncalled get_call_result_type() in a shorter-lived memory context but now\\nrely on the helper function to call get_call_result_type(), the tuple\\ndescriptor will now be in the per query memory context.\\n\\nI think that this makes sense since most of these are set as the\\nResultSetInfo->setDesc (the tuple descriptor for returned tuples).\\n\\nExplanation of the different groups of functions using the\\nMakeFuncResultTuplestore() helper:\\n\\nThe first group are functions which called get_call_result_type(),\\nallocating memory for the tuple descriptor in the short-lived memory\\ncontext, then switched into the per query memory context before\\ncreating the tuplestore, then switched back to the short-lived memory\\ncontext directly after making the tuplestore.\\n\\nNow that they use the helper function and switch into the per query\\nmemory context before calling it, the tuple descriptor is allocated in\\nthe per query memory context.\\n\\nThe first group:\\n\\n pg_stop_backup_v2()\\n pg_event_trigger_dropped_objects()\\n pg_event_trigger_ddl_commands()\\n pg_available_extensions()\\n pg_available_extension_versions()\\n pg_extension_update_paths()\\n pg_hba_file_rules()\\n pg_stat_get_subscription()\\n pg_logical_slot_get_changes_guts()\\n pg_show_replication_origin_status()\\n pg_get_replication_slots()\\n pg_stat_get_wal_senders()\\n pg_get_shmem_allocations()\\n pg_timezone_names()\\n pg_ls_dir_files()\\n pg_get_backend_memory_contexts()\\n pg_stat_get_progress_info()\\n pg_stat_get_activity()\\n pg_stat_get_slru()\\n\\nThe second group are functions which, instead of calling\\nget_call_result_type(), call CreateTemplateTupleDesc() and\\nTupleDescInitEntry() to allocate and initialize the tuple descriptor in\\nthe per query memory context.\\nTheir memory usage is unchanged by this patch.\\n\\nThe second group:\\n\\n pg_prepared_statement()\\n pg_ls_dir()\\n pg_tablespace_databases()\\n show_all_file_settings()\\n pg_cursor()\\n\\nThe third and fourth groups are functions that use CreateTupleDescCopy()\\nto make a tuple descriptor copy from ResultSetInfo->expectedDesc.\\nThese functions made a copy of ResultSetInfo->expectedDesc but did not\\ncheck if it was present before doing so (and do not seem to call a\\nfunction which would call internal_get_result_type() -- which does this\\ncheck), so I have added that check.\\nTheir memory usage is unchanged by this patch.\\n\\nThe third group:\\n\\n text_to_table()\\n\\nThe fourth group differs from the third in that the values for the\\ntuples are actually created in the per query memory context -- seemingly\\nunnecessarily since tuplestore_putvalues() will eventually end up\\ncopying everything when it is forming tuples. This behavior is not\\naltered at all by the MakeFuncResultTuplestore() helper function patch,\\nhowever I wanted to mention that I noticed that perhaps more was being\\nallocated in the per query memory context than is necessary.\\n\\nThe fourth group:\\n\\n pg_options_to_table()\\n pg_config()\\n\\nThe fifth group is a function that uses CreateTupleDescCoy() to make a\\ncopy of a tuple descriptor (NOT ResultSetInfo->expectedDesc) in the\\nfunction info memory context (fcinfo->flinfo->fn_mcxt) so that it can be\\naccessed later.\\nTheir memory usage is unchanged by this patch.\\n\\nThe fifth group:\\n\\n populate_recordset_worker()\\n\\nThe sixth group of functions use CreateTupleDescCopy() to make a copy of\\nthe tuple descriptor that they set up using get_call_result_type().\\n\\nThey called get_call_result_type() in a short-lived memory context, then\\nswitched to the per query memory context and made a copy of it right\\naway. It is unclear why they couldn't just call get_call_result_type()\\nin the per query context and not make a copy. Perhaps they needed some\\nspecific behavior of the copy (like that it doesn't copy constraints and\\ndefaults?).\\n\\nWith the attached patch, the initial tuple descriptor is now made in the\\nper query memory context (as well as the copy). I wonder if the initial\\ntuple descriptor being made in the per query memory context negated the\\nneed for a copy, however I left the copy intact since I wasn't sure.\\n\\nThese functions also checked for the presence of\\nResultSetInfo->expectedDesc which is not applicable for all callers of\\nMakeFuncResultTuplestore(), so I separated it into a separate check.\\nget_call_result_type() will call internal_get_result_type() which will\\ncheck for expectedDesc if the result type is TYPEFUNC_RECORD, so I'm not\\nsure if the expectedDesc check is required or not in the body of these\\nfunctions.\\n\\nThe sixth group:\\n\\n each_worker()\\n each_worker_jsonb()\\n\\nThe seventh group of functions use CreateTupleDescCopy() to make a copy of\\nResultSetInfo->expectedDesc in the per query memory context.\\nThis patch does not change any of that. The only difference is that now\\nthe check for expectedDesc is separate.\\nTheir memory usage is unchanged by this patch.\\n\\nThe seventh group:\\n\\n elements_worker_jsonb()\\n elements_worker()\\n\\nNote that for all of the functions in which I added or changed the check for\\nResultSetInfo->expectedDesc, it is a little bit different because I've changed\\nthe error code.\\nI've used error code: ERRCODE_E_R_I_E_SRF_PROTOCOL_VIOLATED where they\\npreviously would have been part of an ereport with error code\\nERRCODE_FEATURE_NOT_SUPPORTED.\\nI'm not sure if the new one is correct. In fact, I'm not sure if an error\\nrelated to expectedDesc being set should be user-facing at all (looking at how\\nexpectedDesc is used and set, I wasn't sure). Perhaps it should be an elog() or\\nan assert?\\n\\nFeedback is addressed inline below.\\n\\nOn Tue, Nov 2, 2021 at 2:17 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>\\n> On 2021-Nov-02, Melanie Plageman wrote:\\n>\\n> > Attached is a patch to create a helper function which creates a\\n> > tuplestore to be used by FUNCAPI-callable functions.\\n>\\n> Looks good, and given the amount of code being removed, it seems a\\n> no-brainer that some helper(s) is/are needed.\\n>\\n> I think the name is a bit too generic. How about\\n> MakeFuncResultTuplestore()?\\n\\nI've done this rename.\\n\\n> I think the function should not modify rsinfo -- having that as\\n> (undocumented) side effect is weird. I would suggest to just return the\\n> tuplestore, and have the caller stash it in rsinfo->setResult and modify\\n> the other struct members alongside that. It's a couple more lines per\\n> caller, but the code is clearer that way.\\n\\nI've changed the helper not to modify rsinfo.\\n\\n> > There are a few places with very similar code to the new helper\\n> > (MakeTuplestore()) but which, for example, don't expect the return type\\n> > to be a row (pg_ls_dir(), pg_tablespace_databases(), text_to_table()). I\\n> > wasn't sure if it was worth modifying it for their benefit.\\n>\\n> Is this just because of the tupdesc? If that's the case, then maybe the\\n> tupdesc can be optionally passed in by caller, and when it's not, the\\n> helper does get_call_result_type() and the tupdesc is used as output\\n> argument.\\n\\nI've parameterized the helper in this way and now am able to use it for\\nall FUNCAPI-callable functions making tuplestores except fmgr_sql (for\\nSQL functions) in functions.c because it didn't really seem worth it\\ngiven the way the code is written.\\n\\n> > All callers of MakeTuplestore() pass work_mem as the third parameter, so\\n> > I could just omit it and hard-code it in, but I wasn't sure that felt\\n> > right in a utility/helper function.\\n>\\n> No opinion. I would have used the global in the function instead of\\n> passing it in.\\n\\nI've changed this to use the global.\\n\\n> > I inlined deflist_to_tuplestore() in foreign.c since it was billed as a\\n> > helper function but wasn't used elsewhere and it made it harder to use\\n> > MakeTuplestore() in this location.\\n>\\n> This is a bit strange. Does it work to pass rsinfo->expectedDesc?\\n\\nThis is addressed by changing MakeFuncResultTuplestore() call\\nget_call_result_type() optionally and having pg_options_to_table() call\\nit with NULL as tuple descriptor. Now the behavior is the same -- it\\nmakes a copy of expectedDesc in the per query context and sets that as\\nsetDesc in the ReturnSetInfo.\\n\\n- Melanie\",\n \"msg_date\": \"Mon, 8 Nov 2021 14:44:57 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 2, 2021 at 4:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>\\n> Several places have a conditional value for the first argument (randomAccess),\\n> but your patch changes the behavior to a constant \\\"true\\\". I didn't review the\\n> patch beyond that.\\n>\\n> > @@ -740,18 +724,14 @@ pg_prepared_statement(PG_FUNCTION_ARGS)\\n> > - tupstore =\\n> > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\\n> > - false, work_mem);\\n>\\n> > @@ -2701,42 +2701,13 @@ pg_hba_file_rules(PG_FUNCTION_ARGS)\\n> > - tuple_store =\\n> > - tuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\\n> > - false, work_mem);\\n>\\n> > @@ -4799,31 +4797,8 @@ pg_timezone_names(PG_FUNCTION_ARGS)\\n> > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\\n> > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\\n>\\n> > @@ -575,38 +575,12 @@ pg_ls_dir_1arg(PG_FUNCTION_ARGS)\\n> > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\\n> > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\\n>\\n> > @@ -1170,17 +1154,12 @@ pg_cursor(PG_FUNCTION_ARGS)\\n> > - tupstore =\\n> > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\\n> > - false, work_mem);\\n>\\n> > +++ b/src/backend/utils/fmgr/funcapi.c\\n> > + tupstore = tuplestore_begin_heap(true, false, maxKBytes);\\n>\\n\\nI believe the patch has preserved the same behavior. All of the callers\\nfor which I replaced tuplestore_begin_heap() which passed a variable for\\nthe randomAccess parameter had set that variable to something which was\\neffectively the same as passing true -- SFRM_Materialize_Random.\\n\\n- Melanie\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 Nov 2021 14:52:28 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 08, 2021 at 02:52:28PM -0500, Melanie Plageman wrote:\\n> On Tue, Nov 2, 2021 at 4:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> >\\n> > Several places have a conditional value for the first argument (randomAccess),\\n> > but your patch changes the behavior to a constant \\\"true\\\". I didn't review the\\n> > patch beyond that.\\n> >\\n> > > @@ -740,18 +724,14 @@ pg_prepared_statement(PG_FUNCTION_ARGS)\\n> > > - tupstore =\\n> > > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\\n> > > - false, work_mem);\\n> >\\n> > > @@ -2701,42 +2701,13 @@ pg_hba_file_rules(PG_FUNCTION_ARGS)\\n> > > - tuple_store =\\n> > > - tuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\\n> > > - false, work_mem);\\n> >\\n> > > @@ -4799,31 +4797,8 @@ pg_timezone_names(PG_FUNCTION_ARGS)\\n> > > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\\n> > > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\\n> >\\n> > > @@ -575,38 +575,12 @@ pg_ls_dir_1arg(PG_FUNCTION_ARGS)\\n> > > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\\n> > > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\\n> >\\n> > > @@ -1170,17 +1154,12 @@ pg_cursor(PG_FUNCTION_ARGS)\\n> > > - tupstore =\\n> > > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\\n> > > - false, work_mem);\\n> >\\n> > > +++ b/src/backend/utils/fmgr/funcapi.c\\n> > > + tupstore = tuplestore_begin_heap(true, false, maxKBytes);\\n> \\n> I believe the patch has preserved the same behavior. All of the callers\\n> for which I replaced tuplestore_begin_heap() which passed a variable for\\n> the randomAccess parameter had set that variable to something which was\\n> effectively the same as passing true -- SFRM_Materialize_Random.\\n\\nI don't think so ?\\n\\nThey callers aren't passing SFRM_Materialize_Random, but rather\\n(allowedModes & SFRM_Materialize_Random) != 0\\n\\nWhere allowedModes is determined EXEC_FLAG_BACKWARD.\\n\\nsrc/include/executor/executor.h:extern Tuplestorestate *ExecMakeTableFunctionResult(SetExprState *setexpr,\\nsrc/include/executor/executor.h- ExprContext *econtext,\\nsrc/include/executor/executor.h- MemoryContext argContext,\\nsrc/include/executor/executor.h- TupleDesc expectedDesc,\\nsrc/include/executor/executor.h- bool randomAccess);\\n\\nsrc/backend/executor/nodeFunctionscan.c=FunctionNext(FunctionScanState *node)\\nsrc/backend/executor/nodeFunctionscan.c: ExecMakeTableFunctionResult(node->funcstates[0].setexpr,\\nsrc/backend/executor/nodeFunctionscan.c- node->ss.ps.ps_ExprContext,\\nsrc/backend/executor/nodeFunctionscan.c- node->argcontext,\\nsrc/backend/executor/nodeFunctionscan.c- node->funcstates[0].tupdesc,\\nsrc/backend/executor/nodeFunctionscan.c- node->eflags & EXEC_FLAG_BACKWARD);\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 Nov 2021 14:13:38 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 8, 2021 at 3:13 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>\\n> On Mon, Nov 08, 2021 at 02:52:28PM -0500, Melanie Plageman wrote:\\n> > On Tue, Nov 2, 2021 at 4:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > >\\n> > > Several places have a conditional value for the first argument (randomAccess),\\n> > > but your patch changes the behavior to a constant \\\"true\\\". I didn't review the\\n> > > patch beyond that.\\n> > >\\n> > > > @@ -740,18 +724,14 @@ pg_prepared_statement(PG_FUNCTION_ARGS)\\n> > > > - tupstore =\\n> > > > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\\n> > > > - false, work_mem);\\n> > >\\n> > > > @@ -2701,42 +2701,13 @@ pg_hba_file_rules(PG_FUNCTION_ARGS)\\n> > > > - tuple_store =\\n> > > > - tuplestore_begin_heap(rsi->allowedModes & SFRM_Materialize_Random,\\n> > > > - false, work_mem);\\n> > >\\n> > > > @@ -4799,31 +4797,8 @@ pg_timezone_names(PG_FUNCTION_ARGS)\\n> > > > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\\n> > > > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\\n> > >\\n> > > > @@ -575,38 +575,12 @@ pg_ls_dir_1arg(PG_FUNCTION_ARGS)\\n> > > > - randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\\n> > > > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\\n> > >\\n> > > > @@ -1170,17 +1154,12 @@ pg_cursor(PG_FUNCTION_ARGS)\\n> > > > - tupstore =\\n> > > > - tuplestore_begin_heap(rsinfo->allowedModes & SFRM_Materialize_Random,\\n> > > > - false, work_mem);\\n> > >\\n> > > > +++ b/src/backend/utils/fmgr/funcapi.c\\n> > > > + tupstore = tuplestore_begin_heap(true, false, maxKBytes);\\n> >\\n> > I believe the patch has preserved the same behavior. All of the callers\\n> > for which I replaced tuplestore_begin_heap() which passed a variable for\\n> > the randomAccess parameter had set that variable to something which was\\n> > effectively the same as passing true -- SFRM_Materialize_Random.\\n>\\n> I don't think so ?\\n>\\n> They callers aren't passing SFRM_Materialize_Random, but rather\\n> (allowedModes & SFRM_Materialize_Random) != 0\\n>\\n> Where allowedModes is determined EXEC_FLAG_BACKWARD.\\n>\\n> src/include/executor/executor.h:extern Tuplestorestate *ExecMakeTableFunctionResult(SetExprState *setexpr,\\n> src/include/executor/executor.h- ExprContext *econtext,\\n> src/include/executor/executor.h- MemoryContext argContext,\\n> src/include/executor/executor.h- TupleDesc expectedDesc,\\n> src/include/executor/executor.h- bool randomAccess);\\n>\\n> src/backend/executor/nodeFunctionscan.c=FunctionNext(FunctionScanState *node)\\n> src/backend/executor/nodeFunctionscan.c: ExecMakeTableFunctionResult(node->funcstates[0].setexpr,\\n> src/backend/executor/nodeFunctionscan.c- node->ss.ps.ps_ExprContext,\\n> src/backend/executor/nodeFunctionscan.c- node->argcontext,\\n> src/backend/executor/nodeFunctionscan.c- node->funcstates[0].tupdesc,\\n> src/backend/executor/nodeFunctionscan.c- node->eflags & EXEC_FLAG_BACKWARD);\\n>\\n\\nYou are right. I misread it.\\n\\nSo, I've attached a patch where randomAccess is now an additional\\nparameter (and registered for the next fest).\\n\\nI was thinking about how to add a test that would have broken when I\\npassed true for randomAccess to tuplestore_begin_heap() when false was\\nrequired. But, I don't fully understand the problem. If backward\\naccesses to a tuplestore are not allowed but randomAccess is mistakenly\\npassed as true, would the potential result be potentially wrong results\\nfrom accessing the tuplestore results backwards?\\n\\n- Melanie\",\n \"msg_date\": \"Thu, 18 Nov 2021 12:59:03 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 12:59:03PM -0500, Melanie Plageman wrote:\\n> On Mon, Nov 8, 2021 at 3:13 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > On Mon, Nov 08, 2021 at 02:52:28PM -0500, Melanie Plageman wrote:\\n> > > On Tue, Nov 2, 2021 at 4:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > > >\\n> > > > Several places have a conditional value for the first argument (randomAccess),\\n> > > > but your patch changes the behavior to a constant \\\"true\\\". I didn't review the\\n> > > > patch beyond that.\\n> > > >\\n...\\n> > > I believe the patch has preserved the same behavior. All of the callers\\n> > > for which I replaced tuplestore_begin_heap() which passed a variable for\\n> > > the randomAccess parameter had set that variable to something which was\\n> > > effectively the same as passing true -- SFRM_Materialize_Random.\\n> >\\n> > I don't think so ?\\n> >\\n> > They callers aren't passing SFRM_Materialize_Random, but rather\\n> > (allowedModes & SFRM_Materialize_Random) != 0\\n> >\\n> > Where allowedModes is determined EXEC_FLAG_BACKWARD.\\n...\\n> You are right. I misread it.\\n> \\n> So, I've attached a patch where randomAccess is now an additional\\n> parameter (and registered for the next fest).\\n> \\n> I was thinking about how to add a test that would have broken when I\\n> passed true for randomAccess to tuplestore_begin_heap() when false was\\n> required. But, I don't fully understand the problem. If backward\\n> accesses to a tuplestore are not allowed but randomAccess is mistakenly\\n> passed as true, would the potential result be potentially wrong results\\n> from accessing the tuplestore results backwards?\\n\\nNo, see here\\n\\nsrc/backend/utils/fmgr/README-The Tuplestore must be created with randomAccess = true if\\nsrc/backend/utils/fmgr/README:SFRM_Materialize_Random is set in allowedModes, but it can (and preferably\\nsrc/backend/utils/fmgr/README-should) be created with randomAccess = false if not. Callers that can support\\nsrc/backend/utils/fmgr/README-both ValuePerCall and Materialize mode will set SFRM_Materialize_Preferred,\\nsrc/backend/utils/fmgr/README-or not, depending on which mode they prefer.\\n\\nIf you use \\\"randomAccess=true\\\" when it's not needed, the result might be less\\nefficient.\\n\\nSome callers specify \\\"true\\\" when they don't need to. I ran into that here.\\nhttps://www.postgresql.org/message-id/21724.1583955158@sss.pgh.pa.us\\n\\nMaybe you'd want to add an 0002 patch which changes those to conditional ?\\n\\nBTW, there should be a newline before MakeFuncResultTuplestore().\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 12:24:30 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 1:24 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>\\n> On Thu, Nov 18, 2021 at 12:59:03PM -0500, Melanie Plageman wrote:\\n> > On Mon, Nov 8, 2021 at 3:13 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > > On Mon, Nov 08, 2021 at 02:52:28PM -0500, Melanie Plageman wrote:\\n> > > > On Tue, Nov 2, 2021 at 4:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > > > >\\n> > > > > Several places have a conditional value for the first argument (randomAccess),\\n> > > > > but your patch changes the behavior to a constant \\\"true\\\". I didn't review the\\n> > > > > patch beyond that.\\n> > > > >\\n> ...\\n> > > > I believe the patch has preserved the same behavior. All of the callers\\n> > > > for which I replaced tuplestore_begin_heap() which passed a variable for\\n> > > > the randomAccess parameter had set that variable to something which was\\n> > > > effectively the same as passing true -- SFRM_Materialize_Random.\\n> > >\\n> > > I don't think so ?\\n> > >\\n> > > They callers aren't passing SFRM_Materialize_Random, but rather\\n> > > (allowedModes & SFRM_Materialize_Random) != 0\\n> > >\\n> > > Where allowedModes is determined EXEC_FLAG_BACKWARD.\\n> ...\\n> > You are right. I misread it.\\n> >\\n> > So, I've attached a patch where randomAccess is now an additional\\n> > parameter (and registered for the next fest).\\n> >\\n> > I was thinking about how to add a test that would have broken when I\\n> > passed true for randomAccess to tuplestore_begin_heap() when false was\\n> > required. But, I don't fully understand the problem. If backward\\n> > accesses to a tuplestore are not allowed but randomAccess is mistakenly\\n> > passed as true, would the potential result be potentially wrong results\\n> > from accessing the tuplestore results backwards?\\n>\\n> No, see here\\n>\\n> src/backend/utils/fmgr/README-The Tuplestore must be created with randomAccess = true if\\n> src/backend/utils/fmgr/README:SFRM_Materialize_Random is set in allowedModes, but it can (and preferably\\n> src/backend/utils/fmgr/README-should) be created with randomAccess = false if not. Callers that can support\\n> src/backend/utils/fmgr/README-both ValuePerCall and Materialize mode will set SFRM_Materialize_Preferred,\\n> src/backend/utils/fmgr/README-or not, depending on which mode they prefer.\\n>\\n> If you use \\\"randomAccess=true\\\" when it's not needed, the result might be less\\n> efficient.\\n>\\n> Some callers specify \\\"true\\\" when they don't need to. I ran into that here.\\n> https://www.postgresql.org/message-id/21724.1583955158@sss.pgh.pa.us\\n>\\n> Maybe you'd want to add an 0002 patch which changes those to conditional ?\\n\\nDone in attached v4\\n\\n> BTW, there should be a newline before MakeFuncResultTuplestore().\\n\\nFixed\\n\\n- Melanie\",\n \"msg_date\": \"Mon, 13 Dec 2021 17:27:52 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 13, 2021 at 05:27:52PM -0500, Melanie Plageman wrote:\\n> Done in attached v4\\n\\nThanks.\\n\\nI think these comments can be removed from the callers:\\n/* it's a simple type, so don't use get_call_result_type */\\n\\nYou removed one call to tuplestore_donestoring(), but not the others.\\nI guess you could remove them all (optionally as a separate patch).\\n\\nThe rest of these are questions more than comments, and I'm not necessarily\\nsuggesting to change the patch:\\n\\nThis isn't new in your patch, but for me, it's more clear if the callers have a\\nseparate boolean for randomAccess, rather than having the long expression in\\nthe function call:\\n\\n+ tupstore = MakeFuncResultTuplestore(fcinfo, NULL,\\n+ rsinfo->allowedModes & SFRM_Materialize_Random);\\n\\nvs\\n\\n randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\\n- tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\\n+ tupstore = MakeFuncResultTuplestore(fcinfo, NULL, randomAccess);\\n\\nOne could also argue for passing rsinfo->allowedModes, instead of\\n(rsinfo->allowedModes & SFRM_Materialize_Random).\\n\\nThere's a couples places that you're checking expectedDesc where it wasn't\\nbeing checked before. Is that some kind of live bug ?\\npg_config() text_to_table()\\n\\nIt'd be nice if the \\\"expected tuple format\\\" error didn't need to be duplicated\\nfor each SRFs. I suppose the helper function could just take a boolean\\ndetermining whether or not to throw an error (passing \\\"expectedDesc==NULL\\\").\\nYou'd have to call the helper before CreateTupleDescCopy() - which you're\\nalready doing in a couple places for similar reasons.\\n\\nI noticed that tuplestore.h isn't included most places. I suppose it's\\nincluded via execnodes.h. Your patch doesn't change that, arguably it\\nshould've always been included.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Fri, 17 Dec 2021 14:04:19 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"Thanks for the detailed review!\\n\\nOn Fri, Dec 17, 2021 at 3:04 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>\\n> On Mon, Dec 13, 2021 at 05:27:52PM -0500, Melanie Plageman wrote:\\n> > Done in attached v4\\n>\\n> Thanks.\\n>\\n> I think these comments can be removed from the callers:\\n> /* it's a simple type, so don't use get_call_result_type */\\n\\nDone in attached v5\\n\\n> You removed one call to tuplestore_donestoring(), but not the others.\\n> I guess you could remove them all (optionally as a separate patch).\\n\\nI removed it in that one location because I wanted to get rid of the\\nlocal variable it used. I am fine with removing the other occurrences,\\nbut I thought there might be some reason why instead of removing it,\\nthey made it into a no-op.\\n\\n> The rest of these are questions more than comments, and I'm not necessarily\\n> suggesting to change the patch:\\n>\\n> This isn't new in your patch, but for me, it's more clear if the callers have a\\n> separate boolean for randomAccess, rather than having the long expression in\\n> the function call:\\n>\\n> + tupstore = MakeFuncResultTuplestore(fcinfo, NULL,\\n> + rsinfo->allowedModes & SFRM_Materialize_Random);\\n>\\n> vs\\n>\\n> randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\\n> - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\\n> + tupstore = MakeFuncResultTuplestore(fcinfo, NULL, randomAccess);\\n>\\n> One could also argue for passing rsinfo->allowedModes, instead of\\n> (rsinfo->allowedModes & SFRM_Materialize_Random).\\n\\nI've changed the patch to have MakeFuncResultTuplestore() check\\nrsinfo->allowedModes and pass in the resulting boolean to\\ntuplestore_begin_heap(). Because I modified the\\nMakeFuncResultTuplestore() function signature to accommodate this, I had\\nto collapse the two patches into one, as I couldn't maintain the call\\nsites which passed in a constant.\\n\\n> There's a couples places that you're checking expectedDesc where it wasn't\\n> being checked before. Is that some kind of live bug ?\\n> pg_config() text_to_table()\\n\\nYes, expectedDesc was accessed in these locations without checking that\\nit is not NULL. Should that be a separate patch?\\n\\n> It'd be nice if the \\\"expected tuple format\\\" error didn't need to be duplicated\\n> for each SRFs. I suppose the helper function could just take a boolean\\n> determining whether or not to throw an error (passing \\\"expectedDesc==NULL\\\").\\n> You'd have to call the helper before CreateTupleDescCopy() - which you're\\n> already doing in a couple places for similar reasons.\\n\\nSo, I don't think it makes sense to move that error message into\\nMakeFuncResultTuplestore() for the cases in which NULL is passed in for\\nthe tuple descriptor. expectedDesc is not accessed at all in these cases\\ninside the function (since get_call_result_type()) is not called. For\\nthe cases when a tuple descriptor is passed in, only two callers\\nactually check for expectedDesc -- each_worker_jsonb and each_worker.\\nTherefore, I don't think it makes sense to add another parameter just to\\nmove that check inside for two callers.\\n\\n> I noticed that tuplestore.h isn't included most places. I suppose it's\\n> included via execnodes.h. Your patch doesn't change that, arguably it\\n> should've always been included.\\n\\nI've now included it in funcapi.c.\\n\\n- Melanie\",\n \"msg_date\": \"Wed, 5 Jan 2022 12:09:16 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 05, 2022 at 12:09:16PM -0500, Melanie Plageman wrote:\\n> On Fri, Dec 17, 2021 at 3:04 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > The rest of these are questions more than comments, and I'm not necessarily\\n> > suggesting to change the patch:\\n> >\\n> > This isn't new in your patch, but for me, it's more clear if the callers have a\\n> > separate boolean for randomAccess, rather than having the long expression in\\n> > the function call:\\n> >\\n> > + tupstore = MakeFuncResultTuplestore(fcinfo, NULL,\\n> > + rsinfo->allowedModes & SFRM_Materialize_Random);\\n> >\\n> > vs\\n> >\\n> > randomAccess = (rsinfo->allowedModes & SFRM_Materialize_Random) != 0;\\n> > - tupstore = tuplestore_begin_heap(randomAccess, false, work_mem);\\n> > + tupstore = MakeFuncResultTuplestore(fcinfo, NULL, randomAccess);\\n> >\\n> > One could also argue for passing rsinfo->allowedModes, instead of\\n> > (rsinfo->allowedModes & SFRM_Materialize_Random).\\n> \\n> I've changed the patch to have MakeFuncResultTuplestore() check\\n> rsinfo->allowedModes and pass in the resulting boolean to\\n> tuplestore_begin_heap(). Because I modified the\\n> MakeFuncResultTuplestore() function signature to accommodate this, I had\\n> to collapse the two patches into one, as I couldn't maintain the call\\n> sites which passed in a constant.\\n\\n> > There's a couples places that you're checking expectedDesc where it wasn't\\n> > being checked before. Is that some kind of live bug ?\\n> > pg_config() text_to_table()\\n> \\n> Yes, expectedDesc was accessed in these locations without checking that\\n> it is not NULL. Should that be a separate patch?\\n\\nI don't know. The patch is already easy to review, since it's mostly limited\\nto removing code and fixing inconsistencies (NULL check) and possible\\ninefficiencies (randomAccess).\\n\\nIf the expectedDesc NULL check were an 0001 patch, then 0002 (the main patch)\\nwould be even easier to review. Only foreign.c is different.\\n\\n> > You removed one call to tuplestore_donestoring(), but not the others.\\n> > I guess you could remove them all (optionally as a separate patch).\\n> \\n> I removed it in that one location because I wanted to get rid of the\\n> local variable it used.\\n\\nWhat local variable ? I see now that logicalfuncs.c never had a local variable\\ncalled tupstore. I'm sure it was intended since 2014 (b89e15105) to say\\ntupestore_donestoring(p->tupstore). But donestoring(typo) causes no error,\\nsince the define is a NOP.\\n\\nsrc/include/utils/tuplestore.h:#define tuplestore_donestoring(state) ((void) 0)\\n\\n> I am fine with removing the other occurrences,\\n> but I thought there might be some reason why instead of removing it,\\n> they made it into a no-op.\\n\\nI assume Tom left it (actually, added it back in dd04e958c) to avoid breaking\\nextensions for no reason. And maybe to preserve the possbility of at some\\npoint in the future doing something again during the \\\"done\\\" step.\\n\\nI'd leave it for input from a committer about those:\\n\\n - remove tuplestore_donestoring() calls ?\\n - split expectedDesc NULL check to an 0001 patch ?\\n - anything other opened questions ?\\n\\nI'm marking this as RFC, with the caveat that the newline before\\nMakeFuncResultTuplestore went missing again :)\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Wed, 5 Jan 2022 18:57:48 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 5, 2022 at 7:57 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>\\n> On Wed, Jan 05, 2022 at 12:09:16PM -0500, Melanie Plageman wrote:\\n> > On Fri, Dec 17, 2021 at 3:04 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > > There's a couples places that you're checking expectedDesc where it wasn't\\n> > > being checked before. Is that some kind of live bug ?\\n> > > pg_config() text_to_table()\\n> >\\n> > Yes, expectedDesc was accessed in these locations without checking that\\n> > it is not NULL. Should that be a separate patch?\\n>\\n> I don't know. The patch is already easy to review, since it's mostly limited\\n> to removing code and fixing inconsistencies (NULL check) and possible\\n> inefficiencies (randomAccess).\\n>\\n> If the expectedDesc NULL check were an 0001 patch, then 0002 (the main patch)\\n> would be even easier to review. Only foreign.c is different.\\n\\nI'll wait to do that if preferred by committer.\\nAre you imagining that patch 0001 would only add the check for\\nexpectedDesc that is missing from pg_config() and text_to_table()?\\n\\n> > > You removed one call to tuplestore_donestoring(), but not the others.\\n> > > I guess you could remove them all (optionally as a separate patch).\\n> >\\n> > I removed it in that one location because I wanted to get rid of the\\n> > local variable it used.\\n>\\n> What local variable ? I see now that logicalfuncs.c never had a local variable\\n> called tupstore. I'm sure it was intended since 2014 (b89e15105) to say\\n> tupestore_donestoring(p->tupstore). But donestoring(typo) causes no error,\\n> since the define is a NOP.\\n>\\n> src/include/utils/tuplestore.h:#define tuplestore_donestoring(state) ((void) 0)\\n\\nYes, I mean the local variable, tupstore.\\n\\n> > I am fine with removing the other occurrences,\\n> > but I thought there might be some reason why instead of removing it,\\n> > they made it into a no-op.\\n>\\n> I assume Tom left it (actually, added it back in dd04e958c) to avoid breaking\\n> extensions for no reason. And maybe to preserve the possbility of at some\\n> point in the future doing something again during the \\\"done\\\" step.\\n>\\n> I'd leave it for input from a committer about those:\\n>\\n> - remove tuplestore_donestoring() calls ?\\n> - split expectedDesc NULL check to an 0001 patch ?\\n> - anything other opened questions ?\\n>\\n> I'm marking this as RFC, with the caveat that the newline before\\n> MakeFuncResultTuplestore went missing again :)\\n\\noops. I've attached v6 with the newline.\\n\\n- Melanie\",\n \"msg_date\": \"Tue, 11 Jan 2022 10:19:33 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 11, 2022 at 10:19:33AM -0500, Melanie Plageman wrote:\\n> On Wed, Jan 5, 2022 at 7:57 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>> I'd leave it for input from a committer about those:\\n>>\\n>> - remove tuplestore_donestoring() calls ?\\n\\nThis part has been raised on a different thread, as of:\\nhttps://www.postgresql.org/message-id/CAP0=ZVJeeYfAeRfmzqAF2Lumdiv4S4FewyBnZd4DPTrsSQKJKw@mail.gmail.com\\n\\nAnd it looks sensible from here to just remove any traces of this\\nstuff from the C code, while leaving the compatibility macro around\\nto avoid any breakages with any external modules.\\n\\n>> - split expectedDesc NULL check to an 0001 patch ?\\n\\nI am not sure that this is worth a split, as the same areas are\\ntouched.\\n\\n>> - anything other opened questions ?\\n>>\\n>> I'm marking this as RFC, with the caveat that the newline before\\n>> MakeFuncResultTuplestore went missing again :)\\n> \\n> oops. I've attached v6 with the newline.\\n\\nSaying that, the wrapper you are proposing here looks like a good\\nidea. I have wanted a centralized wrapper more than once, and this is\\na nice cut overall.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 16 Feb 2022 14:43:08 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 11, 2022 at 10:19:33AM -0500, Melanie Plageman wrote:\\n> > If the expectedDesc NULL check were an 0001 patch, then 0002 (the main patch)\\n> > would be even easier to review. Only foreign.c is different.\\n> \\n> I'll wait to do that if preferred by committer.\\n> Are you imagining that patch 0001 would only add the check for\\n> expectedDesc that is missing from pg_config() and text_to_table()?\\n\\nYes, that's what I meant, but I haven't been able to determine if a NULL check\\nshould be added at all.\\n\\nI found this commit message, which indicates that it shouldn't be used in every\\ncase. I'm not clear when it should/not be used, though.\\n\\ncommit 60651e4cddbb77e8f1a0c7fc0be6a7e7bf626fe0\\nAuthor: Tom Lane <tgl@sss.pgh.pa.us>\\nDate: Sat Oct 28 14:02:21 2017 -0400\\n\\n Support domains over composite types in PL/Perl.\\n \\n In passing, don't insist on rsi->expectedDesc being set unless we\\n actually need it; this allows succeeding in a couple of cases where\\n PL/Perl functions returning setof composite would have failed before,\\n and makes the error message more apropos in other cases.\\n\\n\\n\",\n \"msg_date\": \"Tue, 15 Feb 2022 23:57:56 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Tue, Feb 15, 2022 at 11:57:56PM -0600, Justin Pryzby wrote:\\n> On Tue, Jan 11, 2022 at 10:19:33AM -0500, Melanie Plageman wrote:\\n>> I'll wait to do that if preferred by committer.\\n>> Are you imagining that patch 0001 would only add the check for\\n>> expectedDesc that is missing from pg_config() and text_to_table()?\\n\\nIt took me some time to follow up on this point, but we clearly don't\\nexpect a NULL expectedDesc in the context of these two code paths or\\nwe finish with a NULL pointer dereference in CreateTupleDescCopy()\\nwith a crash, so they had better be added. It also makes sense to me\\nto do that first and independently of the rest of the patch: it is a\\nseparate issue.\\n\\n> Yes, that's what I meant, but I haven't been able to determine if a NULL check\\n> should be added at all.\\n\\n+ (errcode(ERRCODE_E_R_I_E_SRF_PROTOCOL_VIOLATED),\\n+ errmsg(\\\"expected tuple format not specified as required for \\\"\\n+ \\\"set-returning function.\\\")))\\nerrmsgs should be in a single line, to ease the grepping of similar\\npatterns.\\n\\nThere is something else that I think could be cleaned up, while we are\\nworking on this area. Some of the callers of\\nMakeFuncResultTuplestore() (prepare.c, portalmem.c, one in genfile.c,\\nguc.c, misc.c) pass down a NULL tupdesc while building their TupleDesc\\nusing CreateTemplateTupleDesc() and a set of TupleDescInitEntry() for\\neach attribute. This data exists already in pg_proc.dat, duplicating\\nthe effort. Except if I am missing something, wouldn't it be better\\nto let get_call_result_type() do this work for us, passing a TupleDesc\\npointer to the new MakeFuncResultTuplestore() rather than NULL?\\n\\nOnce you do that, there is something else that stands out: all the\\ncallers of MakeFuncResultTuplestore() passing NULL for tupledesc *have\\nto* set expectedDesc, as far as I can see from the remaining callers.\\nThis brings an interesting effect, couldn't we move the check on\\nReturnSetInfo->expectedDesc being not NULL within\\nMakeFuncResultTuplestore(), checking after it only when tupdesc ==\\nNULL?\\n\\nOne would say that this changes the error ordering in code paths like\\npg_config(), but, actually, in this case, we don't have any need for\\nthe part \\\"Check to make sure we have a reasonable tuple descriptor\\\" as\\nof pg_proc.dat defining the tuple description this function uses, no?\\nAnd we could switch to a tuplestore in this case, as well, reducing by\\none the numbers of callers with tupdesc=NULL for the new routine.\\nThis could be a separate cleanup, separate from the rest, done first.\\n\\neach_worker_jsonb() does not really need its check on expectedDesc\\nbeing NULL, does it? HEAD grabs the tuple descriptor with\\nget_call_result_type(), the patch uses MakeFuncResultTuplestore(), so\\nwe make nowhere use of the expected TupleDesc saved by the SRF\\nexecutor.\\n\\nAsserting that we are in the correct memory context in when calling\\nMakeFuncResultTuplestore() sounds rather sensible from here as per the\\nmagics done in the various json functions. Still, it really feels\\nlike we could do a more centralized consolidation of the whole.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 17 Feb 2022 16:10:01 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 17, 2022 at 04:10:01PM +0900, Michael Paquier wrote:\\n> Asserting that we are in the correct memory context in when calling\\n> MakeFuncResultTuplestore() sounds rather sensible from here as per the\\n> magics done in the various json functions. Still, it really feels\\n> like we could do a more centralized consolidation of the whole.\\n\\nSo, I got my hands on this area, and found myself applying 07daca5 as\\na first piece of the puzzle. Anyway, after more review today, I have\\nbumped into more pieces that could be consolidated, and finished with\\nthe following patch set:\\n- 0001 changes a couple of SRFs that I found worth simplifying. These\\nrefer mostly to the second and fourth group mentioned upthread by\\nMelanie, with two exceptions: pg_tablespace_databases() where it is\\nnot worth changing to keep it backward-compatible and pg_ls_dir() as\\nper its one-arg version. That's a nice first cut in itself.\\n- 0002 changes a couple of places to unify some existing SRF checks,\\nthat I bumped into on the way. The value here is in using the same\\nerror messages everywhere, reducing the translation effort and\\ngenerating the same errors for all cases based on the same conditions.\\nThis eases the review of the next patch.\\n- 0003 is the actual refactoring meat, where I have been able to move\\nthe check on expectedDesc into MakeFuncResultTuplestore(), shaving\\nmore code than previously. If you discard the cases of patch 0001, it\\nshould actually be possible to set setResult, setDesc and returnMode\\nwithin the new function, which would feel natural as that's the place\\nwhere we create the function's tuplestore and we want to materialize\\nits contents. The cases with the JSON code are also a bit hairy and\\nneed more thoughts, but we could also cut this part of the code from\\nthe initial refactoring effort.\\n\\nSo, for now, 0001 and 0002 look like rather committable pieces. 0003\\nneeds a bit more thoughts about all the corner cases we need to\\nconsider, mostly what I am mentioning above. Perhaps the pieces of\\n0003 related to pg_options_to_table() should be moved to 0001 as a\\nmatter of clarity.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 21 Feb 2022 16:41:17 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 21, 2022 at 04:41:17PM +0900, Michael Paquier wrote:\\n> So, I got my hands on this area, and found myself applying 07daca5 as\\n> a first piece of the puzzle. Anyway, after more review today, I have\\n> bumped into more pieces that could be consolidated, and finished with\\n> the following patch set:\\n> - 0001 changes a couple of SRFs that I found worth simplifying. These\\n> refer mostly to the second and fourth group mentioned upthread by\\n> Melanie, with two exceptions: pg_tablespace_databases() where it is\\n> not worth changing to keep it backward-compatible and pg_ls_dir() as\\n> per its one-arg version. That's a nice first cut in itself.\\n> - 0002 changes a couple of places to unify some existing SRF checks,\\n> that I bumped into on the way. The value here is in using the same\\n> error messages everywhere, reducing the translation effort and\\n> generating the same errors for all cases based on the same conditions.\\n> This eases the review of the next patch.\\n\\nThese two have been now applied, with some comment improvements and\\nthe cleanup of pg_options_to_table() done in 0001, and a slight change\\nin 0002 for pageinspect where a check was not necessary for a BRIN\\ncode path.\\n\\n> - 0003 is the actual refactoring meat, where I have been able to move\\n> the check on expectedDesc into MakeFuncResultTuplestore(), shaving\\n> more code than previously. If you discard the cases of patch 0001, it\\n> should actually be possible to set setResult, setDesc and returnMode\\n> within the new function, which would feel natural as that's the place\\n> where we create the function's tuplestore and we want to materialize\\n> its contents. The cases with the JSON code are also a bit hairy and\\n> need more thoughts, but we could also cut this part of the code from\\n> the initial refactoring effort.\\n\\nThis is the remaining piece, as attached, that I have not been able to\\npoke much at yet.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 24 Feb 2022 20:25:06 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 24, 2022 at 08:25:06PM +0900, Michael Paquier wrote:\\n> This is the remaining piece, as attached, that I have not been able to\\n> poke much at yet.\\n\\nSo, I have finally poked at this last part of the patch set, and I\\nfound that we can be more aggressive with the refactoring, by moving\\ninto MakeFuncResultTuplestore() the parts where we save the tuplestore\\nand the tupledesc in the per-query memory context. There are two\\npieces that matter once things are reshaped:\\n- The tuple descriptor may need some extra validation via\\nBlessTupleDesc() when it comes from a transient record datatype,\\nsomething that happens for most of the subroutines related to the JSON\\nfunctions.\\n- expectedDesc is sometimes required by the caller, though most of the\\ntime just needs to be built with the more expensive\\nget_call_result_type().\\n\\nIn order to keep things pluggable at will, MakeFuncResultTuplestore()\\nhas been changed to access a set of bits32 flags, able to control the\\ntwo options above. With this facility in place, I have been able to\\ncut much more code than the initial patch, roughly twice as of:\\n 24 files changed, 157 insertions(+), 893 deletions(-)\\n\\nThis seems in rather good shape to me, the changes are\\nstraight-forward and the code cut is really good, so I'd like to move\\non with that. 0001 is the initial patch, and 0002 is the extra\\nrefactoring I have been working on. The plan would be to merge both,\\nbut I am sending a split to ease any checks on what I have changed.\\n\\nComments or objections?\\n--\\nMichael\",\n \"msg_date\": \"Mon, 28 Feb 2022 16:49:41 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 04:49:41PM +0900, Michael Paquier wrote:\\n> In order to keep things pluggable at will, MakeFuncResultTuplestore()\\n> has been changed to access a set of bits32 flags, able to control the\\n> two options above. With this facility in place, I have been able to\\n> cut much more code than the initial patch, roughly twice as of:\\n> 24 files changed, 157 insertions(+), 893 deletions(-)\\n\\nSo, I have been thinking about this patch once again, and after\\npondering more on it, MakeFuncResultTuplestore() is actually a wrong\\nname now that it does much more than just creating a tuplestore, by\\nassigning the TupleDesc and the TupleStore into the function's\\nReturnSetInfo.\\n\\nThis is actually setting up a function in the context of a single call\\nwhere we fill the tuplestore with all its values, so instead I have\\nsettled down to name that SetSingleFuncCall(), to make a parallel with\\nthe existing MultiFuncCall*(). funcapi.c is the right place for\\nthat, and I have added more documentation in the fmgr's README as well\\nas funcapi.h.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 2 Mar 2022 15:43:17 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 02, 2022 at 03:43:17PM +0900, Michael Paquier wrote:\\n> This is actually setting up a function in the context of a single call\\n> where we fill the tuplestore with all its values, so instead I have\\n> settled down to name that SetSingleFuncCall(), to make a parallel with\\n> the existing MultiFuncCall*(). funcapi.c is the right place for\\n> that, and I have added more documentation in the fmgr's README as well\\n> as funcapi.h.\\n\\nI have tortured all those code paths for the last couple of days, and\\nthe new function name, as well as its options, still seemed fine to\\nme, so I have applied the patch. The buildfarm is cool with it. It\\nis worth noting that there are more code paths in contrib/ that could\\nbe simplified with this new routine.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 7 Mar 2022 11:29:03 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Sun, Mar 6, 2022 at 9:29 PM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Wed, Mar 02, 2022 at 03:43:17PM +0900, Michael Paquier wrote:\\n> > This is actually setting up a function in the context of a single call\\n> > where we fill the tuplestore with all its values, so instead I have\\n> > settled down to name that SetSingleFuncCall(), to make a parallel with\\n> > the existing MultiFuncCall*(). funcapi.c is the right place for\\n> > that, and I have added more documentation in the fmgr's README as well\\n> > as funcapi.h.\\n>\\n> I have tortured all those code paths for the last couple of days, and\\n> the new function name, as well as its options, still seemed fine to\\n> me, so I have applied the patch. The buildfarm is cool with it. It\\n> is worth noting that there are more code paths in contrib/ that could\\n> be simplified with this new routine.\\n\\nWow! Thanks so much for taking the time to review, refine, and commit\\nthis work.\\n\\nI've attached a patch using the helper in most locations in contrib\\nmodules that seemed useful.\\n\\nThe following I don't think we can use the helper or it is not worth it:\\n\\n- pg_logdir_ls() in contrib/adminpack has return type TYPEFUNC_RECORD\\n and expectedDesc is not already created, so the helper can't be used.\\n\\n But basically, since it doesn't specify OUT argument names, it has to\\n do TupleDescInitEntry() itself anyway, I think.\\n\\n- contrib/tablefunc.c was also not simple to refactor. the various parts\\n of SetSingleFuncCall are spread throughout different functions in the\\n file.\\n\\n- contrib/dblink has one function which returns a tuplestore that was\\n simple to change (dblink_get_notify()) and I've done so.\\n\\n However, most of the other creation of tuplestore and tupledescriptors\\n is in helper functions (like materializeResult()) which separate the\\n tuplestore creation from the tuple descriptor initialization in a way\\n that made it hard to do a drop-in replacement with the helper function.\\n\\n- Melanie\",\n \"msg_date\": \"Mon, 7 Mar 2022 17:09:19 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 07, 2022 at 05:09:19PM -0500, Melanie Plageman wrote:\\n> I've attached a patch using the helper in most locations in contrib\\n> modules that seemed useful.\\n\\nThanks for the patch. I was also looking at that yesterday, and this\\npretty much maps to what I have finished with, except for dblink and\\nxml2 where it was not looking that obvious to me at quick glance.\\n\\nIn xml2, my eyes hurt a bit on the meaning of doing the \\\"Switch out of\\nSPI context\\\" in xpath_table() after doing the two SPI calls, but I\\nagree that this extra switch back to the old context should not be\\nnecessary. For dblink, I did not notice that the change for\\ndblink_get_notify() would be this straight-forward, good catch. It\\nwould be nice to see prepTuplestoreResult() gone at the end, though I\\nam not sure how invasive/worth it would be for dblink/.\\n\\n> - pg_logdir_ls() in contrib/adminpack has return type TYPEFUNC_RECORD\\n> and expectedDesc is not already created, so the helper can't be used.\\n> But basically, since it doesn't specify OUT argument names, it has to\\n> do TupleDescInitEntry() itself anyway, I think.\\n\\nHere, actually, I thought first that a simplification should be\\npossible, noticing that the regression tests passed with the function\\ncalled until I saw what it was testing :)\\n\\nI am fine to not poke at the beast, and it looks like we may run into\\ntrouble if we wish to maintain compatibility down to adminpack 1.0 at\\nruntime. This function has a very limited usage, anyway.\\n\\n> - contrib/tablefunc.c was also not simple to refactor. the various parts\\n> of SetSingleFuncCall are spread throughout different functions in the\\n> file.\\n> \\n> - contrib/dblink has one function which returns a tuplestore that was\\n> simple to change (dblink_get_notify()) and I've done so.\\n\\nThis matches my impression, so applied.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 8 Mar 2022 10:17:03 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: make tuplestore helper function\"\n }\n]","isConversational":false}}},{"rowIdx":53042,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nAndres mentioned in passing that he'd defined\nPG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY only on Itanium, PPC and x86 but\nnot ARM.\n\nI took a look at https://developer.arm.com/documentation/ddi0487/gb/\nunder \"B2.2.1 Requirements for single-copy atomicity\" and it seemed\nlike we should turn this on for __aarch64__. It goes back to the\noriginal ARMv8-A so should cover all 64 bit ARM systems.","msg_date":"Wed, 3 Nov 2021 11:33:40 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":true,"msg_subject":"AArch64 has single-copy 64 bit atomicity"},{"msg_contents":"On Wed, Nov 3, 2021 at 1:34 AM Thomas Munro <thomas.munro@gmail.com> wrote:\n> Andres mentioned in passing that he'd defined\n> PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY only on Itanium, PPC and x86 but\n> not ARM.\n>\n> I took a look at https://developer.arm.com/documentation/ddi0487/gb/\n> under \"B2.2.1 Requirements for single-copy atomicity\" and it seemed\n> like we should turn this on for __aarch64__. It goes back to the\n> original ARMv8-A so should cover all 64 bit ARM systems.\n\nThat should be very good because ARM gains popularity and the effect\nof atomic read is significant\n\n------\nRegards,\nAlexander Korotkov\n\n\n","msg_date":"Wed, 3 Nov 2021 03:49:29 +0300","msg_from":"Alexander Korotkov <aekorotkov@gmail.com>","msg_from_op":false,"msg_subject":"Re: AArch64 has single-copy 64 bit atomicity"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nAndres mentioned in passing that he'd defined\\nPG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY only on Itanium, PPC and x86 but\\nnot ARM.\\n\\nI took a look at https://developer.arm.com/documentation/ddi0487/gb/\\nunder \\\"B2.2.1 Requirements for single-copy atomicity\\\" and it seemed\\nlike we should turn this on for __aarch64__. It goes back to the\\noriginal ARMv8-A so should cover all 64 bit ARM systems.\",\n \"msg_date\": \"Wed, 3 Nov 2021 11:33:40 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"AArch64 has single-copy 64 bit atomicity\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 3, 2021 at 1:34 AM Thomas Munro <thomas.munro@gmail.com> wrote:\\n> Andres mentioned in passing that he'd defined\\n> PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY only on Itanium, PPC and x86 but\\n> not ARM.\\n>\\n> I took a look at https://developer.arm.com/documentation/ddi0487/gb/\\n> under \\\"B2.2.1 Requirements for single-copy atomicity\\\" and it seemed\\n> like we should turn this on for __aarch64__. It goes back to the\\n> original ARMv8-A so should cover all 64 bit ARM systems.\\n\\nThat should be very good because ARM gains popularity and the effect\\nof atomic read is significant\\n\\n------\\nRegards,\\nAlexander Korotkov\\n\\n\\n\",\n \"msg_date\": \"Wed, 3 Nov 2021 03:49:29 +0300\",\n \"msg_from\": \"Alexander Korotkov <aekorotkov@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: AArch64 has single-copy 64 bit atomicity\"\n }\n]","isConversational":false}}},{"rowIdx":53043,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi list,\n\njust compiled PostgreSQL head on RedHat 9 beta. Generally seems to be fine but when building contrib, a lot of these warnings pop up:\n\n/usr/lib64/perl5/CORE/inline.h:2635:5: warning: '}' and ')' tokens terminating statement expression appear in different macro expansion contexts [-Wcompound-token-split-by-macro]\n ENV_UNLOCK;\n ^~~~~~~~~~\n/usr/lib64/perl5/CORE/perl.h:2925:31: note: expanded from macro 'ENV_UNLOCK'\n# define ENV_UNLOCK MUTEX_UNLOCK(&PL_env_mutex)\n ^~~~~~~~~~~~~~~~~~~~~~~~~~~\n/usr/lib64/perl5/CORE/thread.h:227:5: note: expanded from macro 'MUTEX_UNLOCK'\n } STMT_END\n ^\n/usr/lib64/perl5/CORE/inline.h:2635:5: note: ')' token is here\n ENV_UNLOCK;\n ^~~~~~~~~~\n\nFull log attached.\n\nNote: I had to install two different packages compared to RHEL 7 and 8:\nperl-FindBin and perl-Opcode\n\nRegards\nDaniel","msg_date":"Wed, 3 Nov 2021 16:54:52 +0000","msg_from":"\"Daniel Westermann (DWE)\" <daniel.westermann@dbi-services.com>","msg_from_op":true,"msg_subject":"Perl warnings when building contrib on RHEL 9 beta"},{"msg_contents":"\"Daniel Westermann (DWE)\" <daniel.westermann@dbi-services.com> writes:\n> just compiled PostgreSQL head on RedHat 9 beta. Generally seems to be fine but when building contrib, a lot of these warnings pop up:\n\n> /usr/lib64/perl5/CORE/inline.h:2635:5: warning: '}' and ')' tokens terminating statement expression appear in different macro expansion contexts [-Wcompound-token-split-by-macro]\n\nYeah, I've been seeing that on macOS as well, with the latest Xcode;\nand several other buildfarm members are showing it too.\n\nI think this is a new clang warning that the Perl headers haven't\ncaught up with yet. I'm not sure there is anything we can do about\nit other than add -Wno-compound-token-split-by-macro.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 03 Nov 2021 13:27:42 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Perl warnings when building contrib on RHEL 9 beta"}],"string":"[\n {\n \"msg_contents\": \"Hi list,\\n\\njust compiled PostgreSQL head on RedHat 9 beta. Generally seems to be fine but when building contrib, a lot of these warnings pop up:\\n\\n/usr/lib64/perl5/CORE/inline.h:2635:5: warning: '}' and ')' tokens terminating statement expression appear in different macro expansion contexts [-Wcompound-token-split-by-macro]\\n ENV_UNLOCK;\\n ^~~~~~~~~~\\n/usr/lib64/perl5/CORE/perl.h:2925:31: note: expanded from macro 'ENV_UNLOCK'\\n# define ENV_UNLOCK MUTEX_UNLOCK(&PL_env_mutex)\\n ^~~~~~~~~~~~~~~~~~~~~~~~~~~\\n/usr/lib64/perl5/CORE/thread.h:227:5: note: expanded from macro 'MUTEX_UNLOCK'\\n } STMT_END\\n ^\\n/usr/lib64/perl5/CORE/inline.h:2635:5: note: ')' token is here\\n ENV_UNLOCK;\\n ^~~~~~~~~~\\n\\nFull log attached.\\n\\nNote: I had to install two different packages compared to RHEL 7 and 8:\\nperl-FindBin and perl-Opcode\\n\\nRegards\\nDaniel\",\n \"msg_date\": \"Wed, 3 Nov 2021 16:54:52 +0000\",\n \"msg_from\": \"\\\"Daniel Westermann (DWE)\\\" <daniel.westermann@dbi-services.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Perl warnings when building contrib on RHEL 9 beta\"\n },\n {\n \"msg_contents\": \"\\\"Daniel Westermann (DWE)\\\" <daniel.westermann@dbi-services.com> writes:\\n> just compiled PostgreSQL head on RedHat 9 beta. Generally seems to be fine but when building contrib, a lot of these warnings pop up:\\n\\n> /usr/lib64/perl5/CORE/inline.h:2635:5: warning: '}' and ')' tokens terminating statement expression appear in different macro expansion contexts [-Wcompound-token-split-by-macro]\\n\\nYeah, I've been seeing that on macOS as well, with the latest Xcode;\\nand several other buildfarm members are showing it too.\\n\\nI think this is a new clang warning that the Perl headers haven't\\ncaught up with yet. I'm not sure there is anything we can do about\\nit other than add -Wno-compound-token-split-by-macro.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 03 Nov 2021 13:27:42 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Perl warnings when building contrib on RHEL 9 beta\"\n }\n]","isConversational":false}}},{"rowIdx":53044,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"The code in gistvacuum.c is closely based on similar code in nbtree.c,\nexcept that it only acquires an exclusive lock -- not a\nsuper-exclusive lock. I suspect that that's because it seemed\nunnecessary; nbtree plain index scans have their own special reasons\nfor this, that don't apply to GiST. Namely: nbtree plain index scans\nthat don't use an MVCC snapshot clearly need some other interlock to\nprotect against concurrent recycling of pointed-to-by-leaf-page TIDs.\nAnd so as a general rule nbtree VACUUM needs a full\nsuper-exclusive/cleanup lock, just in case there is a plain index scan\nthat uses some other kind of snapshot (logical replication, say).\n\nTo say the same thing another way: nbtree follows \"the third rule\"\ndescribed by \"62.4. Index Locking Considerations\" in the docs [1], but\nGiST does not. The idea that GiST's behavior is okay here does seem\nconsistent with what the same docs go on to say about it: \"When using\nan MVCC-compliant snapshot, there is no problem because the new\noccupant of the slot is certain to be too new to pass the snapshot\ntest\".\n\nBut what about index-only scans, which GiST also supports? I think\nthat the rules are different there, even though index-only scans use\nan MVCC snapshot.\n\nThe (admittedly undocumented) reason why we can never drop the leaf\npage pin for an index-only scan in nbtree (but can do so for plain\nindex scans) also relates to heap interlocking -- but with a twist.\nHere's the twist: the second heap pass by VACUUM can set visibility\nmap bits independently of the first (once LP_DEAD items from the first\npass over the heap are set to LP_UNUSED, which renders the page\nall-visible) -- this all happens at the end of\nlazy_vacuum_heap_page(). That's why index-only scans can't just assume\nthat VACUUM won't have deleted the TID from the leaf page they're\nscanning immediately after they're done reading it. VACUUM could even\nmanage to set the visibility map bit for a relevant heap page inside\nlazy_vacuum_heap_page(), before the index-only scan can read the\nvisibility map. If that is allowed to happen, the index-only would\ngive wrong answers if one of the TID references held in local memory\nby the index-only scan happens to be marked LP_UNUSED inside\nlazy_vacuum_heap_page(). IOW, it looks like we run the risk of a\nconcurrently recycled dead-to-everybody TID becoming visible during\nGiST index-only scans, just because we have no interlock.\n\nIn summary:\n\nUUIC this is only safe for nbtree because 1.) It acquires a\nsuper-exclusive lock when vacuuming leaf pages, and 2.) Index-only\nscans never drop their pin on the leaf page when accessing the\nvisibility map \"in-sync\" with the scan (of course we hope not to\naccess the heap proper at all for index-only scans). These precautions\nare both necessary to make the race condition I describe impossible,\nbecause they ensure that VACUUM cannot reach lazy_vacuum_heap_page()\nuntil after our index-only scan reads the visibility map (and then has\nto read the heap for at least that one dead-to-all TID, discovering\nthat the TID is dead to its snapshot). Why wouldn't GiST need to take\nthe same precautions, though?\n\n[1] https://www.postgresql.org/docs/devel/index-locking.html\n--\nPeter Geoghegan\n\n\n","msg_date":"Wed, 3 Nov 2021 16:33:08 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"Why doesn't GiST VACUUM require a super-exclusive lock,\n like nbtree VACUUM?"},{"msg_contents":" 04.11.2021, 04:33, \"Peter Geoghegan\" <pg@bowt.ie>:But what about index-only scans, which GiST also supports? I thinkthat the rules are different there, even though index-only scans usean MVCC snapshot.Let's enumerate steps how things can go wrong.Backend1: Index-Only scan returns tid and xs_hitup with index_tuple1 on index_page1 pointing to heap_tuple1 on page1Backend2: Remove index_tuple1 and heap_tuple1Backend3: Mark page1 all-visibleBackend1: Thinks that page1 is all-visible and shows index_tuple1 as visible To avoid this Backend1 must hold pin on index_page1 until it's done with checking visibility, and Backend2 must do LockBufferForCleanup(index_page1). Do I get things right? Best regards, Andrey Borodin. \n","msg_date":"Thu, 04 Nov 2021 20:52:22 +0500","msg_from":"Andrey Borodin <x4mmm@yandex-team.ru>","msg_from_op":false,"msg_subject":"Re: Why doesn't GiST VACUUM require a super-exclusive lock,\n like nbtree VACUUM?"},{"msg_contents":"On Thu, Nov 4, 2021 at 8:52 AM Andrey Borodin <x4mmm@yandex-team.ru> wrote:\n> Let's enumerate steps how things can go wrong.\n>\n> Backend1: Index-Only scan returns tid and xs_hitup with index_tuple1 on index_page1 pointing to heap_tuple1 on page1\n>\n> Backend2: Remove index_tuple1 and heap_tuple1\n>\n> Backend3: Mark page1 all-visible\n> Backend1: Thinks that page1 is all-visible and shows index_tuple1 as visible\n>\n> To avoid this Backend1 must hold pin on index_page1 until it's done with checking visibility, and Backend2 must do LockBufferForCleanup(index_page1). Do I get things right?\n\nAlmost. Backend3 is actually Backend2 here (there is no 3) -- it runs\nVACUUM throughout.\n\nNote that it's not particularly likely that Backend2/VACUUM will \"win\"\nthis race, because it typically has to do much more work than\nBackend1. It has to actually remove the index tuples from the leaf\npage, then any other index work (for this and other indexes). Then it\nhas to arrive back in vacuumlazy.c to set the VM bit in\nlazy_vacuum_heap_page(). That's a pretty unlikely scenario. And even\nif it happened it would only happen once (until the next time we get\nunlucky). What are the chances of somebody noticing a more or less\nonce-off, slightly wrong answer?\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Thu, 4 Nov 2021 08:58:29 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"Re: Why doesn't GiST VACUUM require a super-exclusive lock, like\n nbtree VACUUM?"},{"msg_contents":"\n\n> 4 нояб. 2021 г., в 20:58, Peter Geoghegan <pg@bowt.ie> написал(а):\n> That's a pretty unlikely scenario. And even\n> if it happened it would only happen once (until the next time we get\n> unlucky). What are the chances of somebody noticing a more or less\n> once-off, slightly wrong answer?\n\nI'd say next to impossible, yet not impossible. Or, perhaps, I do not see protection from this.\n\nMoreover there's a \"microvacuum\". It kills tuples with BUFFER_LOCK_SHARE. AFAIU it should take cleanup lock on buffer too?\n\nBest regards, Andrey Borodin.\n\n","msg_date":"Fri, 5 Nov 2021 15:26:09 +0500","msg_from":"Andrey Borodin <x4mmm@yandex-team.ru>","msg_from_op":false,"msg_subject":"Re: Why doesn't GiST VACUUM require a super-exclusive lock, like\n nbtree VACUUM?"},{"msg_contents":"On Fri, Nov 5, 2021 at 3:26 AM Andrey Borodin <x4mmm@yandex-team.ru> wrote:\n> > 4 нояб. 2021 г., в 20:58, Peter Geoghegan <pg@bowt.ie> написал(а):\n> > That's a pretty unlikely scenario. And even\n> > if it happened it would only happen once (until the next time we get\n> > unlucky). What are the chances of somebody noticing a more or less\n> > once-off, slightly wrong answer?\n>\n> I'd say next to impossible, yet not impossible. Or, perhaps, I do not see protection from this.\n\nI think that that's probably all correct -- I would certainly make the\nsame guess. It's very unlikely to happen, and when it does happen it\nhappens only once.\n\n> Moreover there's a \"microvacuum\". It kills tuples with BUFFER_LOCK_SHARE. AFAIU it should take cleanup lock on buffer too?\n\nNo, because there is no heap vacuuming involved (because that doesn't\nhappen outside lazyvacuum.c). The work that VACUUM does inside\nlazy_vacuum_heap_rel() is part of the problem here -- we need an\ninterlock between that work, and index-only scans. Making LP_DEAD\nitems in heap pages LP_UNUSED is only ever possible during a VACUUM\noperation (I'm sure you know why). AFAICT there would be no bug at all\nwithout that detail.\n\nI believe that there have been several historic reasons why we need a\ncleanup lock during nbtree VACUUM, and that there is only one\nremaining reason for it today. So the history is unusually complicated. But\nAFAICT it's always some kind of \"interlock with heapam VACUUM\" issue,\nwith TID recycling, with no protection from our MVCC snapshot. I would\nsay that that's the \"real problem\" here, when I get to first principles.\n\nAttached draft patch attempts to explain things in this area within\nthe nbtree README. There is a much shorter comment about it within\nvacuumlazy.c. I am concerned about GiST index-only scans themselves,\nof course, but I discovered this issue when thinking carefully about\nthe concurrency rules for VACUUM -- I think it's valuable to formalize\nand justify the general rules that index access methods must follow.\n\nWe can talk about this some more in NYC. See you soon!\n--\nPeter Geoghegan","msg_date":"Tue, 30 Nov 2021 17:09:14 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"Re: Why doesn't GiST VACUUM require a super-exclusive lock, like\n nbtree VACUUM?"},{"msg_contents":"On Tue, Nov 30, 2021 at 5:09 PM Peter Geoghegan <pg@bowt.ie> wrote:\n> I believe that there have been several historic reasons why we need a\n> cleanup lock during nbtree VACUUM, and that there is only one\n> remaining reason for it today. So the history is unusually complicated.\n\nMinor correction: we actually also have to worry about plain index\nscans that don't use an MVCC snapshot, which is possible within\nnbtree. It's quite likely when using logical replication, actually.\nSee the patch for more.\n\nLike with the index-only scan case, a non-MVCC snapshot + plain nbtree\nindex scan cannot rely on heap access within the index scan node -- it\nwon't reliably notice that any newer heap tuples (that are really the\nresult of concurrent TID recycling) are not actually visible to its\nMVCC snapshot -- because there isn't an MVCC snapshot. The only\ndifference in the index-only scan scenario is that we use the\nvisibility map (not the heap) -- which is racey in a way that makes\nour MVCC snapshot (IOSs always have an MVCC snapshot) an ineffective\nprotection.\n\nIn summary, to be safe against confusion from concurrent TID recycling\nduring index/index-only scans, we can do either of the following\nthings:\n\n1. Hold a pin of our leaf page while accessing the heap -- that'll\ndefinitely conflict with the cleanup lock that nbtree VACUUM will\ninevitably try to acquire on our leaf page.\n\nOR:\n\n2. Hold an MVCC snapshot, AND do an actual heap page access during the\nplain index scan -- do both together.\n\nWith approach 2, our plain index scan must determine visibility using\nreal XIDs (against something like a dirty snapshot), rather than using\na visibility map bit. That is also necessary because the VM might\nbecome invalid or ambiguous, in a way that's clearly not possible when\nlooking at full heap tuple headers with XIDs -- concurrent recycling\nbecomes safe if we know that we'll reliably notice it and not give\nwrong answers.\n\nDoes that make sense?\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Tue, 30 Nov 2021 17:39:19 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"Re: Why doesn't GiST VACUUM require a super-exclusive lock, like\n nbtree VACUUM?"},{"msg_contents":"On Tue, Nov 30, 2021 at 5:09 PM Peter Geoghegan <pg@bowt.ie> wrote:\n> Attached draft patch attempts to explain things in this area within\n> the nbtree README. There is a much shorter comment about it within\n> vacuumlazy.c. I am concerned about GiST index-only scans themselves,\n> of course, but I discovered this issue when thinking carefully about\n> the concurrency rules for VACUUM -- I think it's valuable to formalize\n> and justify the general rules that index access methods must follow.\n\nI pushed a commit that described how this works for nbtree, in the README file.\n\nI think that there might be an even more subtle race condition in\nnbtree itself, though, during recovery. We no longer do a \"pin scan\"\nduring recovery these days (see commits 9f83468b, 3e4b7d87, and\n687f2cd7 for full information). I think that it might be necessary to\ndo that, just for the benefit of index-only scans -- if it's necessary\nduring original execution, then why not during recovery?\n\nThe work to remove \"pin scans\" was justified by pointing out that we\nno longer use various kinds of snapshots during recovery, but it said\nnothing about index-only scans, which need the TID recycling interlock\n(i.e. need to hold onto a leaf page while accessing the heap in sync)\neven with an MVCC snapshot. It's easy to imagine how it might have\nbeen missed: nobody ever documented the general issue with index-only\nscans, until now. Commit 2ed5b87f recognized they were unsafe for the\noptimization that it added (to avoid blocking VACUUM), but never\nexplained why they were unsafe.\n\nGoing back to doing pin scans during recovery seems deeply\nunappealing, especially to fix a totally narrow race condition.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Wed, 8 Dec 2021 17:40:37 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"Re: Why doesn't GiST VACUUM require a super-exclusive lock, like\n nbtree VACUUM?"}],"string":"[\n {\n \"msg_contents\": \"The code in gistvacuum.c is closely based on similar code in nbtree.c,\\nexcept that it only acquires an exclusive lock -- not a\\nsuper-exclusive lock. I suspect that that's because it seemed\\nunnecessary; nbtree plain index scans have their own special reasons\\nfor this, that don't apply to GiST. Namely: nbtree plain index scans\\nthat don't use an MVCC snapshot clearly need some other interlock to\\nprotect against concurrent recycling of pointed-to-by-leaf-page TIDs.\\nAnd so as a general rule nbtree VACUUM needs a full\\nsuper-exclusive/cleanup lock, just in case there is a plain index scan\\nthat uses some other kind of snapshot (logical replication, say).\\n\\nTo say the same thing another way: nbtree follows \\\"the third rule\\\"\\ndescribed by \\\"62.4. Index Locking Considerations\\\" in the docs [1], but\\nGiST does not. The idea that GiST's behavior is okay here does seem\\nconsistent with what the same docs go on to say about it: \\\"When using\\nan MVCC-compliant snapshot, there is no problem because the new\\noccupant of the slot is certain to be too new to pass the snapshot\\ntest\\\".\\n\\nBut what about index-only scans, which GiST also supports? I think\\nthat the rules are different there, even though index-only scans use\\nan MVCC snapshot.\\n\\nThe (admittedly undocumented) reason why we can never drop the leaf\\npage pin for an index-only scan in nbtree (but can do so for plain\\nindex scans) also relates to heap interlocking -- but with a twist.\\nHere's the twist: the second heap pass by VACUUM can set visibility\\nmap bits independently of the first (once LP_DEAD items from the first\\npass over the heap are set to LP_UNUSED, which renders the page\\nall-visible) -- this all happens at the end of\\nlazy_vacuum_heap_page(). That's why index-only scans can't just assume\\nthat VACUUM won't have deleted the TID from the leaf page they're\\nscanning immediately after they're done reading it. VACUUM could even\\nmanage to set the visibility map bit for a relevant heap page inside\\nlazy_vacuum_heap_page(), before the index-only scan can read the\\nvisibility map. If that is allowed to happen, the index-only would\\ngive wrong answers if one of the TID references held in local memory\\nby the index-only scan happens to be marked LP_UNUSED inside\\nlazy_vacuum_heap_page(). IOW, it looks like we run the risk of a\\nconcurrently recycled dead-to-everybody TID becoming visible during\\nGiST index-only scans, just because we have no interlock.\\n\\nIn summary:\\n\\nUUIC this is only safe for nbtree because 1.) It acquires a\\nsuper-exclusive lock when vacuuming leaf pages, and 2.) Index-only\\nscans never drop their pin on the leaf page when accessing the\\nvisibility map \\\"in-sync\\\" with the scan (of course we hope not to\\naccess the heap proper at all for index-only scans). These precautions\\nare both necessary to make the race condition I describe impossible,\\nbecause they ensure that VACUUM cannot reach lazy_vacuum_heap_page()\\nuntil after our index-only scan reads the visibility map (and then has\\nto read the heap for at least that one dead-to-all TID, discovering\\nthat the TID is dead to its snapshot). Why wouldn't GiST need to take\\nthe same precautions, though?\\n\\n[1] https://www.postgresql.org/docs/devel/index-locking.html\\n--\\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Wed, 3 Nov 2021 16:33:08 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Why doesn't GiST VACUUM require a super-exclusive lock,\\n like nbtree VACUUM?\"\n },\n {\n \"msg_contents\": \" 04.11.2021, 04:33, \\\"Peter Geoghegan\\\" <pg@bowt.ie>:But what about index-only scans, which GiST also supports? I thinkthat the rules are different there, even though index-only scans usean MVCC snapshot.Let's enumerate steps how things can go wrong.Backend1: Index-Only scan returns tid and xs_hitup with index_tuple1 on index_page1 pointing to heap_tuple1 on page1Backend2: Remove index_tuple1 and heap_tuple1Backend3: Mark page1 all-visibleBackend1: Thinks that page1 is all-visible and shows index_tuple1 as visible To avoid this Backend1 must hold pin on index_page1 until it's done with checking visibility, and Backend2 must do LockBufferForCleanup(index_page1). Do I get things right? Best regards, Andrey Borodin. \\n\",\n \"msg_date\": \"Thu, 04 Nov 2021 20:52:22 +0500\",\n \"msg_from\": \"Andrey Borodin <x4mmm@yandex-team.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Why doesn't GiST VACUUM require a super-exclusive lock,\\n like nbtree VACUUM?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 4, 2021 at 8:52 AM Andrey Borodin <x4mmm@yandex-team.ru> wrote:\\n> Let's enumerate steps how things can go wrong.\\n>\\n> Backend1: Index-Only scan returns tid and xs_hitup with index_tuple1 on index_page1 pointing to heap_tuple1 on page1\\n>\\n> Backend2: Remove index_tuple1 and heap_tuple1\\n>\\n> Backend3: Mark page1 all-visible\\n> Backend1: Thinks that page1 is all-visible and shows index_tuple1 as visible\\n>\\n> To avoid this Backend1 must hold pin on index_page1 until it's done with checking visibility, and Backend2 must do LockBufferForCleanup(index_page1). Do I get things right?\\n\\nAlmost. Backend3 is actually Backend2 here (there is no 3) -- it runs\\nVACUUM throughout.\\n\\nNote that it's not particularly likely that Backend2/VACUUM will \\\"win\\\"\\nthis race, because it typically has to do much more work than\\nBackend1. It has to actually remove the index tuples from the leaf\\npage, then any other index work (for this and other indexes). Then it\\nhas to arrive back in vacuumlazy.c to set the VM bit in\\nlazy_vacuum_heap_page(). That's a pretty unlikely scenario. And even\\nif it happened it would only happen once (until the next time we get\\nunlucky). What are the chances of somebody noticing a more or less\\nonce-off, slightly wrong answer?\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Thu, 4 Nov 2021 08:58:29 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Why doesn't GiST VACUUM require a super-exclusive lock, like\\n nbtree VACUUM?\"\n },\n {\n \"msg_contents\": \"\\n\\n> 4 нояб. 2021 г., в 20:58, Peter Geoghegan <pg@bowt.ie> написал(а):\\n> That's a pretty unlikely scenario. And even\\n> if it happened it would only happen once (until the next time we get\\n> unlucky). What are the chances of somebody noticing a more or less\\n> once-off, slightly wrong answer?\\n\\nI'd say next to impossible, yet not impossible. Or, perhaps, I do not see protection from this.\\n\\nMoreover there's a \\\"microvacuum\\\". It kills tuples with BUFFER_LOCK_SHARE. AFAIU it should take cleanup lock on buffer too?\\n\\nBest regards, Andrey Borodin.\\n\\n\",\n \"msg_date\": \"Fri, 5 Nov 2021 15:26:09 +0500\",\n \"msg_from\": \"Andrey Borodin <x4mmm@yandex-team.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Why doesn't GiST VACUUM require a super-exclusive lock, like\\n nbtree VACUUM?\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 5, 2021 at 3:26 AM Andrey Borodin <x4mmm@yandex-team.ru> wrote:\\n> > 4 нояб. 2021 г., в 20:58, Peter Geoghegan <pg@bowt.ie> написал(а):\\n> > That's a pretty unlikely scenario. And even\\n> > if it happened it would only happen once (until the next time we get\\n> > unlucky). What are the chances of somebody noticing a more or less\\n> > once-off, slightly wrong answer?\\n>\\n> I'd say next to impossible, yet not impossible. Or, perhaps, I do not see protection from this.\\n\\nI think that that's probably all correct -- I would certainly make the\\nsame guess. It's very unlikely to happen, and when it does happen it\\nhappens only once.\\n\\n> Moreover there's a \\\"microvacuum\\\". It kills tuples with BUFFER_LOCK_SHARE. AFAIU it should take cleanup lock on buffer too?\\n\\nNo, because there is no heap vacuuming involved (because that doesn't\\nhappen outside lazyvacuum.c). The work that VACUUM does inside\\nlazy_vacuum_heap_rel() is part of the problem here -- we need an\\ninterlock between that work, and index-only scans. Making LP_DEAD\\nitems in heap pages LP_UNUSED is only ever possible during a VACUUM\\noperation (I'm sure you know why). AFAICT there would be no bug at all\\nwithout that detail.\\n\\nI believe that there have been several historic reasons why we need a\\ncleanup lock during nbtree VACUUM, and that there is only one\\nremaining reason for it today. So the history is unusually complicated. But\\nAFAICT it's always some kind of \\\"interlock with heapam VACUUM\\\" issue,\\nwith TID recycling, with no protection from our MVCC snapshot. I would\\nsay that that's the \\\"real problem\\\" here, when I get to first principles.\\n\\nAttached draft patch attempts to explain things in this area within\\nthe nbtree README. There is a much shorter comment about it within\\nvacuumlazy.c. I am concerned about GiST index-only scans themselves,\\nof course, but I discovered this issue when thinking carefully about\\nthe concurrency rules for VACUUM -- I think it's valuable to formalize\\nand justify the general rules that index access methods must follow.\\n\\nWe can talk about this some more in NYC. See you soon!\\n--\\nPeter Geoghegan\",\n \"msg_date\": \"Tue, 30 Nov 2021 17:09:14 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Why doesn't GiST VACUUM require a super-exclusive lock, like\\n nbtree VACUUM?\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 30, 2021 at 5:09 PM Peter Geoghegan <pg@bowt.ie> wrote:\\n> I believe that there have been several historic reasons why we need a\\n> cleanup lock during nbtree VACUUM, and that there is only one\\n> remaining reason for it today. So the history is unusually complicated.\\n\\nMinor correction: we actually also have to worry about plain index\\nscans that don't use an MVCC snapshot, which is possible within\\nnbtree. It's quite likely when using logical replication, actually.\\nSee the patch for more.\\n\\nLike with the index-only scan case, a non-MVCC snapshot + plain nbtree\\nindex scan cannot rely on heap access within the index scan node -- it\\nwon't reliably notice that any newer heap tuples (that are really the\\nresult of concurrent TID recycling) are not actually visible to its\\nMVCC snapshot -- because there isn't an MVCC snapshot. The only\\ndifference in the index-only scan scenario is that we use the\\nvisibility map (not the heap) -- which is racey in a way that makes\\nour MVCC snapshot (IOSs always have an MVCC snapshot) an ineffective\\nprotection.\\n\\nIn summary, to be safe against confusion from concurrent TID recycling\\nduring index/index-only scans, we can do either of the following\\nthings:\\n\\n1. Hold a pin of our leaf page while accessing the heap -- that'll\\ndefinitely conflict with the cleanup lock that nbtree VACUUM will\\ninevitably try to acquire on our leaf page.\\n\\nOR:\\n\\n2. Hold an MVCC snapshot, AND do an actual heap page access during the\\nplain index scan -- do both together.\\n\\nWith approach 2, our plain index scan must determine visibility using\\nreal XIDs (against something like a dirty snapshot), rather than using\\na visibility map bit. That is also necessary because the VM might\\nbecome invalid or ambiguous, in a way that's clearly not possible when\\nlooking at full heap tuple headers with XIDs -- concurrent recycling\\nbecomes safe if we know that we'll reliably notice it and not give\\nwrong answers.\\n\\nDoes that make sense?\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Tue, 30 Nov 2021 17:39:19 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Why doesn't GiST VACUUM require a super-exclusive lock, like\\n nbtree VACUUM?\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 30, 2021 at 5:09 PM Peter Geoghegan <pg@bowt.ie> wrote:\\n> Attached draft patch attempts to explain things in this area within\\n> the nbtree README. There is a much shorter comment about it within\\n> vacuumlazy.c. I am concerned about GiST index-only scans themselves,\\n> of course, but I discovered this issue when thinking carefully about\\n> the concurrency rules for VACUUM -- I think it's valuable to formalize\\n> and justify the general rules that index access methods must follow.\\n\\nI pushed a commit that described how this works for nbtree, in the README file.\\n\\nI think that there might be an even more subtle race condition in\\nnbtree itself, though, during recovery. We no longer do a \\\"pin scan\\\"\\nduring recovery these days (see commits 9f83468b, 3e4b7d87, and\\n687f2cd7 for full information). I think that it might be necessary to\\ndo that, just for the benefit of index-only scans -- if it's necessary\\nduring original execution, then why not during recovery?\\n\\nThe work to remove \\\"pin scans\\\" was justified by pointing out that we\\nno longer use various kinds of snapshots during recovery, but it said\\nnothing about index-only scans, which need the TID recycling interlock\\n(i.e. need to hold onto a leaf page while accessing the heap in sync)\\neven with an MVCC snapshot. It's easy to imagine how it might have\\nbeen missed: nobody ever documented the general issue with index-only\\nscans, until now. Commit 2ed5b87f recognized they were unsafe for the\\noptimization that it added (to avoid blocking VACUUM), but never\\nexplained why they were unsafe.\\n\\nGoing back to doing pin scans during recovery seems deeply\\nunappealing, especially to fix a totally narrow race condition.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Wed, 8 Dec 2021 17:40:37 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Why doesn't GiST VACUUM require a super-exclusive lock, like\\n nbtree VACUUM?\"\n }\n]","isConversational":false}}},{"rowIdx":53045,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nsqlsmith triggers the following assertion when testing REL_14_STABLE:\n\n TRAP: FailedAssertion(\"a1 <= a2\", File: \"brin_minmax_multi.c\", Line: 1879, PID: 631814)\n\nI can reproduce it with the following query on a fresh regression\ndatabase:\n\n insert into public.brintest_multi (float4col) values (real 'nan');\n\nThe branch was at f6162c020c while testing, backtrace below.\n\nregards,\nAndreas\n\n(gdb) bt\n#0 0x00007f703cc46ce1 in raise () from /lib/x86_64-linux-gnu/libc.so.6\n#1 0x00007f703cc30537 in abort () from /lib/x86_64-linux-gnu/libc.so.6\n#2 0x000055b17517c521 in ExceptionalCondition at assert.c:69\n#3 0x000055b174d25871 in brin_minmax_multi_distance_float4 (fcinfo=<optimized out>) at brin_minmax_multi.c:1879\n#4 0x000055b1751853ba in FunctionCall2Coll (flinfo=flinfo@entry=0x55b176913d10, collation=collation@entry=0, arg1=<optimized out>, \n arg2=<optimized out>) at fmgr.c:1160\n#5 0x000055b174d23a41 in build_distances (distanceFn=distanceFn@entry=0x55b176913d10, colloid=0, \n eranges=eranges@entry=0x55b17693fad0, neranges=neranges@entry=5) at brin_minmax_multi.c:1352\n#6 0x000055b174d256be in compactify_ranges (max_values=32, ranges=0x55b176945708, bdesc=<optimized out>) at brin_minmax_multi.c:1822\n#7 brin_minmax_multi_serialize (bdesc=<optimized out>, src=94220687005448, dst=0x55b17693ae10) at brin_minmax_multi.c:2386\n#8 0x000055b174d2ae0d in brin_form_tuple (brdesc=brdesc@entry=0x55b176914988, blkno=blkno@entry=15, \n tuple=tuple@entry=0x55b17693aaa0, size=size@entry=0x7ffffcc6d4f8) at brin_tuple.c:165\n#9 0x000055b174d20a5f in brininsert (idxRel=0x7f703b2ae170, values=0x7ffffcc6d610, nulls=0x7ffffcc6d5f0, heaptid=0x55b17690da78, \n\n\n","msg_date":"Thu, 04 Nov 2021 09:46:49 +0100","msg_from":"Andreas Seltenreich <andreas.seltenreich@credativ.de>","msg_from_op":true,"msg_subject":"[sqlsmith] Failed assertion in brin_minmax_multi_distance_float4 on\n REL_14_STABLE"},{"msg_contents":"On Thu, Nov 04, 2021 at 09:46:49AM +0100, Andreas Seltenreich wrote:\n> sqlsmith triggers the following assertion when testing REL_14_STABLE:\n> \n> TRAP: FailedAssertion(\"a1 <= a2\", File: \"brin_minmax_multi.c\", Line: 1879, PID: 631814)\n> \n> I can reproduce it with the following query on a fresh regression\n> database:\n> \n> insert into public.brintest_multi (float4col) values (real 'nan');\n> \n> The branch was at f6162c020c while testing, backtrace below.\n\nI couldn't reproduce this, but it reminds me of this one, which we also had\ntrouble reproducing.\n\nhttps://www.postgresql.org/message-id/flat/20210913004447.GA17931%40ahch-to\n\nCould you send a \"bt full\" ?\n\n> (gdb) bt\n\n-- \nJustin\n\n\n","msg_date":"Thu, 4 Nov 2021 11:53:57 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: [sqlsmith] Failed assertion in brin_minmax_multi_distance_float4\n on REL_14_STABLE"},{"msg_contents":"Hi,\n\nOn 11/4/21 17:53, Justin Pryzby wrote:\n> On Thu, Nov 04, 2021 at 09:46:49AM +0100, Andreas Seltenreich wrote:\n>> sqlsmith triggers the following assertion when testing REL_14_STABLE:\n>>\n>> TRAP: FailedAssertion(\"a1 <= a2\", File: \"brin_minmax_multi.c\", Line: 1879, PID: 631814)\n>>\n>> I can reproduce it with the following query on a fresh regression\n>> database:\n>>\n>> insert into public.brintest_multi (float4col) values (real 'nan');\n>>\n>> The branch was at f6162c020c while testing, backtrace below.\n> \n> I couldn't reproduce this, but it reminds me of this one, which we also had\n> trouble reproducing.\n> \n\nI can reproduce that just fine - all I had to do was 'make installcheck' \nand then connect to the regression db and run the insert.\n\nIt seems to be a simple case of confusion in handling of NaN values. We \ndo sort them correctly (by calling float4_lt), but given two values\n\n arg1 = nan (0x400000)\n arg2 = 0.0909090936\n\nthen a simple comparison does not give the expected result\n\n (arg1 < arg2)\n (arg1 == arg2)\n (arg1 > arg2)\n\nall evaluate to false, which is why the assert fails. So I guess the \ndistance function for float4 (and probably float8 too) need a couple \nmore lines checking NaN.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 4 Nov 2021 23:56:07 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: [sqlsmith] Failed assertion in brin_minmax_multi_distance_float4\n on REL_14_STABLE"},{"msg_contents":"\n\nOn 11/4/21 23:56, Tomas Vondra wrote:\n> Hi,\n> \n> On 11/4/21 17:53, Justin Pryzby wrote:\n>> On Thu, Nov 04, 2021 at 09:46:49AM +0100, Andreas Seltenreich wrote:\n>>> sqlsmith triggers the following assertion when testing REL_14_STABLE:\n>>>\n>>> TRAP: FailedAssertion(\"a1 <= a2\", File: \"brin_minmax_multi.c\", \n>>> Line: 1879, PID: 631814)\n>>>\n>>> I can reproduce it with the following query on a fresh regression\n>>> database:\n>>>\n>>> insert into public.brintest_multi (float4col) values (real 'nan');\n>>>\n>>> The branch was at f6162c020c while testing, backtrace below.\n>>\n>> I couldn't reproduce this, but it reminds me of this one, which we \n>> also had\n>> trouble reproducing.\n>>\n> \n> I can reproduce that just fine - all I had to do was 'make installcheck' \n> and then connect to the regression db and run the insert.\n> \n> It seems to be a simple case of confusion in handling of NaN values. We \n> do sort them correctly (by calling float4_lt), but given two values\n> \n> arg1 = nan (0x400000)\n> arg2 = 0.0909090936\n> \n> then a simple comparison does not give the expected result\n> \n> (arg1 < arg2)\n> (arg1 == arg2)\n> (arg1 > arg2)\n> \n> all evaluate to false, which is why the assert fails. So I guess the \n> distance function for float4 (and probably float8 too) need a couple \n> more lines checking NaN.\n> \n\nHere's a patch that should fix this. It simply handled NaN and returns \ndistance 0.0 for two NaN values and Infinity for NaN vs. something else. \nI did check what happens with Infinity values, but I think those are \nfine - we can actually calculate distance with them, and the assert \nworks for them too.\n\nI'll improve the regression tests to include a couple NaN/Infinity cases \ntomorrow, and then I'll get this pushed.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 5 Nov 2021 02:09:59 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: [sqlsmith] Failed assertion in brin_minmax_multi_distance_float4\n on REL_14_STABLE"},{"msg_contents":"On 11/5/21 02:09, Tomas Vondra wrote:\n>\n> Here's a patch that should fix this. \n >\n\nMeh, forgot the attachment, ofc.\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Fri, 5 Nov 2021 02:10:55 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: [sqlsmith] Failed assertion in brin_minmax_multi_distance_float4\n on REL_14_STABLE"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nsqlsmith triggers the following assertion when testing REL_14_STABLE:\\n\\n TRAP: FailedAssertion(\\\"a1 <= a2\\\", File: \\\"brin_minmax_multi.c\\\", Line: 1879, PID: 631814)\\n\\nI can reproduce it with the following query on a fresh regression\\ndatabase:\\n\\n insert into public.brintest_multi (float4col) values (real 'nan');\\n\\nThe branch was at f6162c020c while testing, backtrace below.\\n\\nregards,\\nAndreas\\n\\n(gdb) bt\\n#0 0x00007f703cc46ce1 in raise () from /lib/x86_64-linux-gnu/libc.so.6\\n#1 0x00007f703cc30537 in abort () from /lib/x86_64-linux-gnu/libc.so.6\\n#2 0x000055b17517c521 in ExceptionalCondition at assert.c:69\\n#3 0x000055b174d25871 in brin_minmax_multi_distance_float4 (fcinfo=<optimized out>) at brin_minmax_multi.c:1879\\n#4 0x000055b1751853ba in FunctionCall2Coll (flinfo=flinfo@entry=0x55b176913d10, collation=collation@entry=0, arg1=<optimized out>, \\n arg2=<optimized out>) at fmgr.c:1160\\n#5 0x000055b174d23a41 in build_distances (distanceFn=distanceFn@entry=0x55b176913d10, colloid=0, \\n eranges=eranges@entry=0x55b17693fad0, neranges=neranges@entry=5) at brin_minmax_multi.c:1352\\n#6 0x000055b174d256be in compactify_ranges (max_values=32, ranges=0x55b176945708, bdesc=<optimized out>) at brin_minmax_multi.c:1822\\n#7 brin_minmax_multi_serialize (bdesc=<optimized out>, src=94220687005448, dst=0x55b17693ae10) at brin_minmax_multi.c:2386\\n#8 0x000055b174d2ae0d in brin_form_tuple (brdesc=brdesc@entry=0x55b176914988, blkno=blkno@entry=15, \\n tuple=tuple@entry=0x55b17693aaa0, size=size@entry=0x7ffffcc6d4f8) at brin_tuple.c:165\\n#9 0x000055b174d20a5f in brininsert (idxRel=0x7f703b2ae170, values=0x7ffffcc6d610, nulls=0x7ffffcc6d5f0, heaptid=0x55b17690da78, \\n\\n\\n\",\n \"msg_date\": \"Thu, 04 Nov 2021 09:46:49 +0100\",\n \"msg_from\": \"Andreas Seltenreich <andreas.seltenreich@credativ.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[sqlsmith] Failed assertion in brin_minmax_multi_distance_float4 on\\n REL_14_STABLE\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 04, 2021 at 09:46:49AM +0100, Andreas Seltenreich wrote:\\n> sqlsmith triggers the following assertion when testing REL_14_STABLE:\\n> \\n> TRAP: FailedAssertion(\\\"a1 <= a2\\\", File: \\\"brin_minmax_multi.c\\\", Line: 1879, PID: 631814)\\n> \\n> I can reproduce it with the following query on a fresh regression\\n> database:\\n> \\n> insert into public.brintest_multi (float4col) values (real 'nan');\\n> \\n> The branch was at f6162c020c while testing, backtrace below.\\n\\nI couldn't reproduce this, but it reminds me of this one, which we also had\\ntrouble reproducing.\\n\\nhttps://www.postgresql.org/message-id/flat/20210913004447.GA17931%40ahch-to\\n\\nCould you send a \\\"bt full\\\" ?\\n\\n> (gdb) bt\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Thu, 4 Nov 2021 11:53:57 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [sqlsmith] Failed assertion in brin_minmax_multi_distance_float4\\n on REL_14_STABLE\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 11/4/21 17:53, Justin Pryzby wrote:\\n> On Thu, Nov 04, 2021 at 09:46:49AM +0100, Andreas Seltenreich wrote:\\n>> sqlsmith triggers the following assertion when testing REL_14_STABLE:\\n>>\\n>> TRAP: FailedAssertion(\\\"a1 <= a2\\\", File: \\\"brin_minmax_multi.c\\\", Line: 1879, PID: 631814)\\n>>\\n>> I can reproduce it with the following query on a fresh regression\\n>> database:\\n>>\\n>> insert into public.brintest_multi (float4col) values (real 'nan');\\n>>\\n>> The branch was at f6162c020c while testing, backtrace below.\\n> \\n> I couldn't reproduce this, but it reminds me of this one, which we also had\\n> trouble reproducing.\\n> \\n\\nI can reproduce that just fine - all I had to do was 'make installcheck' \\nand then connect to the regression db and run the insert.\\n\\nIt seems to be a simple case of confusion in handling of NaN values. We \\ndo sort them correctly (by calling float4_lt), but given two values\\n\\n arg1 = nan (0x400000)\\n arg2 = 0.0909090936\\n\\nthen a simple comparison does not give the expected result\\n\\n (arg1 < arg2)\\n (arg1 == arg2)\\n (arg1 > arg2)\\n\\nall evaluate to false, which is why the assert fails. So I guess the \\ndistance function for float4 (and probably float8 too) need a couple \\nmore lines checking NaN.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 4 Nov 2021 23:56:07 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [sqlsmith] Failed assertion in brin_minmax_multi_distance_float4\\n on REL_14_STABLE\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 11/4/21 23:56, Tomas Vondra wrote:\\n> Hi,\\n> \\n> On 11/4/21 17:53, Justin Pryzby wrote:\\n>> On Thu, Nov 04, 2021 at 09:46:49AM +0100, Andreas Seltenreich wrote:\\n>>> sqlsmith triggers the following assertion when testing REL_14_STABLE:\\n>>>\\n>>> TRAP: FailedAssertion(\\\"a1 <= a2\\\", File: \\\"brin_minmax_multi.c\\\", \\n>>> Line: 1879, PID: 631814)\\n>>>\\n>>> I can reproduce it with the following query on a fresh regression\\n>>> database:\\n>>>\\n>>> insert into public.brintest_multi (float4col) values (real 'nan');\\n>>>\\n>>> The branch was at f6162c020c while testing, backtrace below.\\n>>\\n>> I couldn't reproduce this, but it reminds me of this one, which we \\n>> also had\\n>> trouble reproducing.\\n>>\\n> \\n> I can reproduce that just fine - all I had to do was 'make installcheck' \\n> and then connect to the regression db and run the insert.\\n> \\n> It seems to be a simple case of confusion in handling of NaN values. We \\n> do sort them correctly (by calling float4_lt), but given two values\\n> \\n> arg1 = nan (0x400000)\\n> arg2 = 0.0909090936\\n> \\n> then a simple comparison does not give the expected result\\n> \\n> (arg1 < arg2)\\n> (arg1 == arg2)\\n> (arg1 > arg2)\\n> \\n> all evaluate to false, which is why the assert fails. So I guess the \\n> distance function for float4 (and probably float8 too) need a couple \\n> more lines checking NaN.\\n> \\n\\nHere's a patch that should fix this. It simply handled NaN and returns \\ndistance 0.0 for two NaN values and Infinity for NaN vs. something else. \\nI did check what happens with Infinity values, but I think those are \\nfine - we can actually calculate distance with them, and the assert \\nworks for them too.\\n\\nI'll improve the regression tests to include a couple NaN/Infinity cases \\ntomorrow, and then I'll get this pushed.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Nov 2021 02:09:59 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [sqlsmith] Failed assertion in brin_minmax_multi_distance_float4\\n on REL_14_STABLE\"\n },\n {\n \"msg_contents\": \"On 11/5/21 02:09, Tomas Vondra wrote:\\n>\\n> Here's a patch that should fix this. \\n >\\n\\nMeh, forgot the attachment, ofc.\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Fri, 5 Nov 2021 02:10:55 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [sqlsmith] Failed assertion in brin_minmax_multi_distance_float4\\n on REL_14_STABLE\"\n }\n]","isConversational":false}}},{"rowIdx":53046,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hey,\n\nI've noticed that the current documentation doesn't mention IF EXISTS clause for ALTER STATISTICS in the synopsis section, where PG supports it.\nhttps://www.postgresql.org/docs/14/sql-alterstatistics.html\n(Only for the last item, that is ALTER STATISTICS .. SET STATISTICS; for the others, PG just throws a syntax error.)\n\nI'm from the Citus team and noticed this while bug fixing, and I wonder if it is intentional or not. If it's intentionally supported while the other ALTER STATISTICS statement types are not supported, it would be good to mention that in the documentation.\n\nBest,\nAhmet\n\n\n\n\n\n\n\n\nHey,\n\n\n\n\nI've noticed that the current documentation doesn't mention IF EXISTS clause for ALTER STATISTICS in the synopsis section, where PG supports it. \n\nhttps://www.postgresql.org/docs/14/sql-alterstatistics.html\n\n\n(Only for the last item, that is ALTER STATISTICS .. SET STATISTICS; for the others, PG just throws a syntax error.)\n\n\n\n\nI'm from the Citus team and noticed this while bug fixing, and I wonder if it is intentional or not. If it's intentionally supported while the other ALTER STATISTICS statement types are not supported, it would be good to mention that in the documentation.\n\n\n\n\nBest,\nAhmet","msg_date":"Thu, 4 Nov 2021 13:35:43 +0000","msg_from":"Ahmet Gedemenli <agedemenli@microsoft.com>","msg_from_op":true,"msg_subject":"Possible Documentation Update for ALTER STATISTICS"},{"msg_contents":"Hi Ahmet,\n\nOn 11/4/21 14:35, Ahmet Gedemenli wrote:\n> Hey,\n> \n> I've noticed that the current documentation doesn't mention IF EXISTS \n> clause for ALTER STATISTICS in the synopsis section, where PG supports it.\n> https://www.postgresql.org/docs/14/sql-alterstatistics.html \n> <https://www.postgresql.org/docs/14/sql-alterstatistics.html>\n> (Only for the last item, that is ALTER STATISTICS .. SET STATISTICS; for \n> the others, PG just throws a syntax error.)\n> \n> I'm from the Citus team and noticed this while bug fixing, and I wonder \n> if it is intentional or not. If it's intentionally supported while the \n> other ALTER STATISTICS statement types are not supported, it would be \n> good to mention that in the documentation.\n> \n\nWell, it's intentional in the sense that support for IF EXISTS in ALTER \ncommands is rather spotty. For OWNER TO it's not supported at all, and \nfor the other (RENAME & SET SCHEMA) it's supported only for some object \ntypes. So we added the minimum grammar and never got around to add it.\n\nSo you're right we should update the docs for the SET STATISTICS case to \nshow it's supported in 14. I'll do that shortly.\n\nFor 15+ we could improve this to allow IF EXISTS in the other cases. For \nRENAME and SET SCHEMA it's fairly easy (see attached fix), for OWNER TO \nit's going to be more work because the AlterOwnerStmt does not have the \nmissing_ok flag.\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Thu, 4 Nov 2021 23:07:47 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Possible Documentation Update for ALTER STATISTICS"}],"string":"[\n {\n \"msg_contents\": \"Hey,\\n\\nI've noticed that the current documentation doesn't mention IF EXISTS clause for ALTER STATISTICS in the synopsis section, where PG supports it.\\nhttps://www.postgresql.org/docs/14/sql-alterstatistics.html\\n(Only for the last item, that is ALTER STATISTICS .. SET STATISTICS; for the others, PG just throws a syntax error.)\\n\\nI'm from the Citus team and noticed this while bug fixing, and I wonder if it is intentional or not. If it's intentionally supported while the other ALTER STATISTICS statement types are not supported, it would be good to mention that in the documentation.\\n\\nBest,\\nAhmet\\n\\n\\n\\n\\n\\n\\n\\n\\nHey,\\n\\n\\n\\n\\nI've noticed that the current documentation doesn't mention IF EXISTS clause for ALTER STATISTICS in the synopsis section, where PG supports it. \\n\\nhttps://www.postgresql.org/docs/14/sql-alterstatistics.html\\n\\n\\n(Only for the last item, that is ALTER STATISTICS .. SET STATISTICS; for the others, PG just throws a syntax error.)\\n\\n\\n\\n\\nI'm from the Citus team and noticed this while bug fixing, and I wonder if it is intentional or not. If it's intentionally supported while the other ALTER STATISTICS statement types are not supported, it would be good to mention that in the documentation.\\n\\n\\n\\n\\nBest,\\nAhmet\",\n \"msg_date\": \"Thu, 4 Nov 2021 13:35:43 +0000\",\n \"msg_from\": \"Ahmet Gedemenli <agedemenli@microsoft.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Possible Documentation Update for ALTER STATISTICS\"\n },\n {\n \"msg_contents\": \"Hi Ahmet,\\n\\nOn 11/4/21 14:35, Ahmet Gedemenli wrote:\\n> Hey,\\n> \\n> I've noticed that the current documentation doesn't mention IF EXISTS \\n> clause for ALTER STATISTICS in the synopsis section, where PG supports it.\\n> https://www.postgresql.org/docs/14/sql-alterstatistics.html \\n> <https://www.postgresql.org/docs/14/sql-alterstatistics.html>\\n> (Only for the last item, that is ALTER STATISTICS .. SET STATISTICS; for \\n> the others, PG just throws a syntax error.)\\n> \\n> I'm from the Citus team and noticed this while bug fixing, and I wonder \\n> if it is intentional or not. If it's intentionally supported while the \\n> other ALTER STATISTICS statement types are not supported, it would be \\n> good to mention that in the documentation.\\n> \\n\\nWell, it's intentional in the sense that support for IF EXISTS in ALTER \\ncommands is rather spotty. For OWNER TO it's not supported at all, and \\nfor the other (RENAME & SET SCHEMA) it's supported only for some object \\ntypes. So we added the minimum grammar and never got around to add it.\\n\\nSo you're right we should update the docs for the SET STATISTICS case to \\nshow it's supported in 14. I'll do that shortly.\\n\\nFor 15+ we could improve this to allow IF EXISTS in the other cases. For \\nRENAME and SET SCHEMA it's fairly easy (see attached fix), for OWNER TO \\nit's going to be more work because the AlterOwnerStmt does not have the \\nmissing_ok flag.\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Thu, 4 Nov 2021 23:07:47 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Possible Documentation Update for ALTER STATISTICS\"\n }\n]","isConversational":false}}},{"rowIdx":53047,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI see that \"procsignal\" and \"ProcSignal\" are being used in the code\ncomments which look inconsistent. IMO, \"ProcSignal\" is the right word\nto use and let's be consistent across the code comments. Attaching a\ntiny patch for that.\n\nThoughts?\n\nRegards,\nBharath Rupireddy.","msg_date":"Fri, 5 Nov 2021 11:22:27 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"consistently use \"ProcSignal\" instead of \"procsignal\" in code\n comments"},{"msg_contents":"\n\nOn 2021/11/05 14:52, Bharath Rupireddy wrote:\n> Hi,\n> \n> I see that \"procsignal\" and \"ProcSignal\" are being used in the code\n> comments which look inconsistent. IMO, \"ProcSignal\" is the right word\n> to use and let's be consistent across the code comments. Attaching a\n> tiny patch for that.\n> \n> Thoughts?\n\nI'm fine with this. Barring any objection, I will commit the patch.\n\nRegards,\n\n-- \nFujii Masao\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Tue, 9 Nov 2021 12:51:39 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: consistently use \"ProcSignal\" instead of \"procsignal\" in code\n comments"},{"msg_contents":"On Tue, Nov 09, 2021 at 12:51:39PM +0900, Fujii Masao wrote:\n> I'm fine with this. Barring any objection, I will commit the patch.\n\nI have to admit that the timing is kind of odd, or strange, or both,\nbecause I was just going through my backlog from -hackers, and just\nbumped on this thread like 15 minutes ago.. And the patch was fine on\nconsistency ground, so I have just applied it now.\n--\nMichael","msg_date":"Tue, 9 Nov 2021 13:01:41 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: consistently use \"ProcSignal\" instead of \"procsignal\" in code\n comments"},{"msg_contents":"\n\nOn 2021/11/09 13:01, Michael Paquier wrote:\n> On Tue, Nov 09, 2021 at 12:51:39PM +0900, Fujii Masao wrote:\n>> I'm fine with this. Barring any objection, I will commit the patch.\n> \n> I have to admit that the timing is kind of odd, or strange, or both,\n> because I was just going through my backlog from -hackers, and just\n> bumped on this thread like 15 minutes ago..\n\n:)\n\n\n> And the patch was fine on\n> consistency ground, so I have just applied it now.\n\nThanks!\n\nRegards,\n\n-- \nFujii Masao\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Tue, 9 Nov 2021 13:03:58 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: consistently use \"ProcSignal\" instead of \"procsignal\" in code\n comments"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI see that \\\"procsignal\\\" and \\\"ProcSignal\\\" are being used in the code\\ncomments which look inconsistent. IMO, \\\"ProcSignal\\\" is the right word\\nto use and let's be consistent across the code comments. Attaching a\\ntiny patch for that.\\n\\nThoughts?\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Fri, 5 Nov 2021 11:22:27 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"consistently use \\\"ProcSignal\\\" instead of \\\"procsignal\\\" in code\\n comments\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2021/11/05 14:52, Bharath Rupireddy wrote:\\n> Hi,\\n> \\n> I see that \\\"procsignal\\\" and \\\"ProcSignal\\\" are being used in the code\\n> comments which look inconsistent. IMO, \\\"ProcSignal\\\" is the right word\\n> to use and let's be consistent across the code comments. Attaching a\\n> tiny patch for that.\\n> \\n> Thoughts?\\n\\nI'm fine with this. Barring any objection, I will commit the patch.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 12:51:39 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: consistently use \\\"ProcSignal\\\" instead of \\\"procsignal\\\" in code\\n comments\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 09, 2021 at 12:51:39PM +0900, Fujii Masao wrote:\\n> I'm fine with this. Barring any objection, I will commit the patch.\\n\\nI have to admit that the timing is kind of odd, or strange, or both,\\nbecause I was just going through my backlog from -hackers, and just\\nbumped on this thread like 15 minutes ago.. And the patch was fine on\\nconsistency ground, so I have just applied it now.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 9 Nov 2021 13:01:41 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: consistently use \\\"ProcSignal\\\" instead of \\\"procsignal\\\" in code\\n comments\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2021/11/09 13:01, Michael Paquier wrote:\\n> On Tue, Nov 09, 2021 at 12:51:39PM +0900, Fujii Masao wrote:\\n>> I'm fine with this. Barring any objection, I will commit the patch.\\n> \\n> I have to admit that the timing is kind of odd, or strange, or both,\\n> because I was just going through my backlog from -hackers, and just\\n> bumped on this thread like 15 minutes ago..\\n\\n:)\\n\\n\\n> And the patch was fine on\\n> consistency ground, so I have just applied it now.\\n\\nThanks!\\n\\nRegards,\\n\\n-- \\nFujii Masao\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 13:03:58 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: consistently use \\\"ProcSignal\\\" instead of \\\"procsignal\\\" in code\\n comments\"\n }\n]","isConversational":false}}},{"rowIdx":53048,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"We have amassed quite a lot of patches in the CF app, and while Jaime did a\nvery good job closing patches in the last CF there is still a lot to go\nthrough. I've attempted a brief per-patch triage here to see where we are with\nthese. Reading every email in the threads for these patches is a herculean\ntask, so I may very well have misunderstood things. If so I apologize and\nplease do jump and correct me. If not, then please also jump in and comment\nbelow (or on the relevant threads) to try and get closure.\n\nWhile I don't have the expertise to replicate Andres' famous patch triage [0]\nfrom a few years back, below is at least a try. These are the patches with 7\nor more commitfests under their belts (the record holder having 18), the next\ninstallment will be 6 CF's and down limited by when the Shiraz runs out.\n\n\n1608: schema variables, LET command\n===================================\nAfter 18 CF's and two very long threads it seems to be nearing completion\njudging by Tomas' review. There is an ask in that review for a second pass\nover the docs by a native speaker, any takers?\n\n\n1741: Index Skip Scan\n=====================\nAn often requested feature which has proven hard to reach consensus on an\nimplementation for. The thread(s) have stalled since May, is there any hope of\ntaking this further? Where do we go from here with this patch?\n\n\n1712: Remove self join on a unique column\n=========================================\nThis has moved from \"don't write bad queries\" to \"maybe we should do something\nabout this\". It seems like there is concensus that it can be worth paying the\nadded planner cost for this, especially if gated by a GUC to keep it out of\ninstallations where it doesn't apply. The regression on large join queries\nhasn't been benchmarked it seems (or I missed it), but the patch seems to have\nmatured and be quite ready. Any takers on getting it across the finish line?\n\n\n2161: standby recovery fails when re-replaying due to missing directory which\nwas removed in previous replay.\n=============================================================================\nTom and Robert seem to be in agreement that parts of this patchset are good,\nand that some parts are not. The thread has stalled and the patch no longer\napply, so unless someone feels like picking up the good pieces this seems like\na contender to close for now.\n\n\n2138: Incremental Materialized View Maintenance\n===============================================\nThere seems to be concensus on the thread that this is a feature that we want,\nand after initial design discussions there seems to be no disagreements with\nthe approach taken. The patch was marked ready for committer almost a year\nago, but have since been needs review (which seems correct). The size of the\npatchset and the length of the thread make it hard to gauge just far away it\nis, maybe the author or a review can summarize the current state and outline\nwhat is left for it to be committable.\n\n\n2215: pg_upgrade fails with non-standard ACL\n============================================\nThis thread has been stalled since March with open questions from Noah, but the\npatch seems to be in pretty good shape. Do you have any thoughts on the state\nof this patch Noah? Seems like a pretty good fix to get in.\n\n\n2218: Implement INSERT SET syntax\n=================================\nThe author has kept this patch updated, and has seemingly updated according to\nthe review comments. Tom: do you have any opinions on whether the updated\nversion addresses your concerns wrt the SELECT rewrite?\n\n\n2048: SQL:2011 application time\n===============================\nThe patch no longer applies, but there is a semi-active review discussion\nongoing. It doesn't seem like the patch is nearing completion, but the feature\nwould no doubt be welcome by many. Is this realistic for 15?\n\n\n2316: WITH SYSTEM VERSIONING Temporal Tables\n============================================\nThis patch no longer applies, and judging by the thread there is a pretty major\nre-design ongoing. I think we should close this entry and open a new one when\nthere is a new patch rather than push it along. What are your thoughts Simon?\n\n\n2096: psql - add SHOW_ALL_RESULTS option\n========================================\nPeter posted an updated version of Fabiens patch about a month ago (which at\nthis point no longer applies) fixing a few issues, but also point at old review\ncomments still unadressed. Since this was pushed, but had to be reverted, I\nassume there is a desire for the feature but it seems to need more work still.\n\n\n1651: GROUP BY optimization\n===========================\nThis is IMO a desired optimization, and after all the heavy lifting by Tomas\nthe patch seems to be in pretty good shape.\n\n\n2377: pg_ls_* functions for showing metadata and recurse (pg_ls_tmpdir to show\nshared filesets)\n==============================================================================\nThe question of what to do with lstat() on Windows is still left unanswered,\nbut the patchset has been split to up to be able to avoid it. Stephen and Tom,\nhaving done prior reviews do you have any thoughts on this?\n\n\n2349: global temporary table\n============================\nGTT has been up for discussion numerous times in tbe past, and I can't judge\nwhether this proposal has a better chance than previous ones. I do note the\npatch has a number of crashes reported lately, and no reviews from senior\ncontributors in a while, making it seem unlikely to be committed in this CF.\nSince the patch is very big, can it be teased apart and committed separately\nfor easier review?\n\n\n2433: Erase the distinctClause if the result is unique by definition\n====================================================================\n(parts of) The approach taken in this patch has been objected against in favor\nof work that Tom has proposed. Until that work materialize this patch is\nblocked, and thus I think we are better of closing it and re-opening it when it\ngets unstuck. Unless Tom has plans to hack on this shortly?\n\n\n2482: jsonpath syntax extensions\n================================\nThis thread has stalled, with no reviews so far a good year and half after\nbeing. Rather than continue to push this along, should we instead close it due\nto lack of interest? It's pretty big, can it be broken up into smaller\nfeatures for easier review?\n\n\n2490: Make message at end-of-recovery less scary\n================================================\nThis thread stalled, but has had recent interest. The patch no longer applies\nso while the patch has support it will be marked as returned with feedback\nduring this CF unless revived.\n\n\n2573: pg_dump - read data for some options from external file\n=============================================================\nThis patch has been heavily debated, but I think we've landed in a compromise\nthat has been agreed upon. The gist of the debate has been around the file\nformat and it's extensibility, the final version proposing an extensible yet\nsimple format which can be referenced from a larger configfile (rather than\nmaking the configfile now). Having done a lot of review myself I think the\npatch is ready with one exception: IMO the parsing code isn't maintainable\nenough to be committed (if another committer disagree with that, feel free to\njump in). I have on my TODO to implement it as a Bison parser to see what that\nwould look like.\n\n\n2553: should INSERT SELECT use a BulkInsertState?\n==================================================\nThe conclusion on the thread seems to be that there is little support for it\ndue to the risk of regressions, and that closing the patch might be the best\ncourse of action. If anyone is in favor of going ahead with it, now is the\ntime to speak up.\n\n\n2518: Corruption during WAL replay\n==================================\nBoth Robert and Tom have given positive remarks about this, and recent review\nconcerns raised turned out to have been handled. Robert, Tom: any thoughts on\nthe latest posted version?\n\n\n2601: Fast COPY FROM command for the foreign tables\n===================================================\nThis approach taken in this patch has stabilized and the benchmarks posted are\nvery promising. It seems pretty uncontroversial to add COPY to the FDW API SQL\nspec wise. Amit, Justin and Takayuki-san has done a lot of review and the\npatch is marked Ready for Committer. Any committer interested in taking this\non? Tomas?\n\n\n2602: ALTER SYSTEM READ { ONLY | WRITE }\n========================================\nAFAIK Robert is actively working on this and have committed parts of it\nalready, so there isn't much to add. I'm personally quite excited about this\nfeature so I'm looking forward to (hopefully) seeing it go in. Is the full\nfeature likely to make 15?\n\n\n2627: More scalable multixacts buffers and locking\n==================================================\nSkimming the thread makes it seem like this patch is still undergoing some\nfairly big changes. Thomas and Andrey, how far along is this from RFC?\n\n\n2716: fix spinlock contention in LogwrtResult\n=============================================\nThis addresses a bottleneck which definitely seems like one we want to fix, I\ndon't have a hard time imagining it impacting other production usecases then\nthe reported one. The goalposts moved early in the thread, and while there has\nbeen work on it the patch has stalled on Andres' review which raised some\nconcerns. Are you still pursuing this Alvaro?\n\n\n2684: enhancing plpgsql API for debugging and tracing\n======================================================\nThis hasn't seen much review, but what has been performed has been positive.\nIt does add some weight to plpgsql by passing the NS around, but maybe the\neffect is negligable. Given the size and complexity of the test module to use\nthis, it's perhaps not obvious to see who the target audience is. Maybe a\nbetter option would be to polish the test module and propose it for contrib/\nalong with the patch, and write the tests on using said module?\n\n\n2710: Fix behavior of geo_ops when NaN is involved\n==================================================\nTwo thirds of this bugfix patchset have been committed already, with the thread\nstalling on the remaining bit. Tom, Georgios: any thoughts on the last patch?\n\n\n2694: Automatic HASH and LIST partition creation\n================================================\nThe subject of automatic partition creation is pretty hot, so getting something\ndone in this area would be good. This proposal has gotten criticism over the\nsyntax chosen, and the thread has effectively stalled after Robert's review in\nJuly. It seems like this patch should be returned with feedback, and any new\nattempt at this should start by just discussing the syntax and only once agreed\nupon hack on the code.\n\n\nThanks for reading all this way, I hope it did something to help progress.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n[0] https://www.postgresql.org/message-id/flat/20180302075242.yfqkcgzbrmjysboa%40alap3.anarazel.de#b5015aab4ba3a95b67636faef74a1d51\n\n\n\n","msg_date":"Fri, 5 Nov 2021 22:15:49 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"Thanks for taking care of this.\n\n> 2433: Erase the distinctClause if the result is unique by definition\n> ====================================================================\n> (parts of) The approach taken in this patch has been objected against in favor\n> of work that Tom has proposed.\n\nActually no exactly. The 90% of the code is about UniqueKey and UniqueKey\nneeds a notnulls representation of Var during join, Tom is objecting\nthat part only.\n\n\n> Until that work materialize this patch is\n> blocked, and thus I think we are better of closing it and re-opening it when it\n> gets unstuck. Unless Tom has plans to hack on this shortly?\n\nBoth Tom and David have provided many insights on this topic. I'd like to hear\nthe decision from Tom and David.\n\n-- \nBest Regards\nAndy Fan (https://www.aliyun.com/)\n\n\n","msg_date":"Sat, 6 Nov 2021 09:00:33 +0800","msg_from":"Andy Fan <zhihui.fan1213@gmail.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"> 1741: Index Skip Scan\n> =====================\n> An often requested feature which has proven hard to reach consensus on an\n> implementation for. The thread(s) have stalled since May,\n\nThis statement is not accurate. Peter started a new thread in [1] for this\ntopic last month and then we had a discussion. Thanks for taking\ncare of this.\n\n[1] https://www.postgresql.org/message-id/flat/CAH2-WzmUscvoxVkokHxP%3DuPTDjSi0tJkFpUPD-CeA35dvn-CMw%40mail.gmail.com\n\nis there any hope of\n> taking this further? Where do we go from here with this patch?\n>\n\n-- \nBest Regards\nAndy Fan (https://www.aliyun.com/)\n\n\n","msg_date":"Sat, 6 Nov 2021 09:12:39 +0800","msg_from":"Andy Fan <zhihui.fan1213@gmail.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"On 11/5/21 22:15, Daniel Gustafsson wrote:\n> ...\n> 1651: GROUP BY optimization\n> ===========================\n> This is IMO a desired optimization, and after all the heavy lifting by Tomas\n> the patch seems to be in pretty good shape.\n> \n\nI agree it's desirable. To move the patch forward, I need some feedback \non the costing questions, mentioned in my last review, and the overall \napproach to planning (considering multiple group by variants).\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Sat, 6 Nov 2021 17:20:13 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"On Sat, Nov 6, 2021 at 6:16 AM Daniel Gustafsson <daniel@yesql.se> wrote:\n> 2601: Fast COPY FROM command for the foreign tables\n> ===================================================\n> This approach taken in this patch has stabilized and the benchmarks posted are\n> very promising. It seems pretty uncontroversial to add COPY to the FDW API SQL\n> spec wise. Amit, Justin and Takayuki-san has done a lot of review and the\n> patch is marked Ready for Committer. Any committer interested in taking this\n> on? Tomas?\n\nI’d like to work on this unless Tomas (or anyone else) want.\n\nBest regards,\nEtsuro Fujita\n\n\n","msg_date":"Sun, 7 Nov 2021 18:26:46 +0900","msg_from":"Etsuro Fujita <etsuro.fujita@gmail.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"> On 7 Nov 2021, at 10:26, Etsuro Fujita <etsuro.fujita@gmail.com> wrote:\n> \n> On Sat, Nov 6, 2021 at 6:16 AM Daniel Gustafsson <daniel@yesql.se> wrote:\n>> 2601: Fast COPY FROM command for the foreign tables\n>> ===================================================\n>> This approach taken in this patch has stabilized and the benchmarks posted are\n>> very promising. It seems pretty uncontroversial to add COPY to the FDW API SQL\n>> spec wise. Amit, Justin and Takayuki-san has done a lot of review and the\n>> patch is marked Ready for Committer. Any committer interested in taking this\n>> on? Tomas?\n> \n> I’d like to work on this unless Tomas (or anyone else) want.\n\nGreat, thanks for picking it up!\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Sun, 7 Nov 2021 17:05:17 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"> On 6 Nov 2021, at 17:20, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n> \n> On 11/5/21 22:15, Daniel Gustafsson wrote:\n>> ...\n>> 1651: GROUP BY optimization\n>> ===========================\n>> This is IMO a desired optimization, and after all the heavy lifting by Tomas\n>> the patch seems to be in pretty good shape.\n> \n> I agree it's desirable. To move the patch forward, I need some feedback on the\n> costing questions,\n\n\nTom or David perhaps have some thoughts on that?\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Sun, 7 Nov 2021 17:07:06 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"> On 6 Nov 2021, at 02:12, Andy Fan <zhihui.fan1213@gmail.com> wrote:\n> \n>> 1741: Index Skip Scan\n>> =====================\n>> An often requested feature which has proven hard to reach consensus on an\n>> implementation for. The thread(s) have stalled since May,\n> \n> This statement is not accurate. Peter started a new thread in [1] for this\n> topic last month and then we had a discussion. Thanks for taking\n> care of this.\n\nAha, I had missed that. Thanks to Dimitry for attaching this thread to the CF\nentry to make that clearer going forward.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Sun, 7 Nov 2021 17:11:39 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"On 2021-Nov-05, Daniel Gustafsson wrote:\n\n> 2716: fix spinlock contention in LogwrtResult\n> =============================================\n> This addresses a bottleneck which definitely seems like one we want to fix, I\n> don't have a hard time imagining it impacting other production usecases then\n> the reported one. The goalposts moved early in the thread, and while there has\n> been work on it the patch has stalled on Andres' review which raised some\n> concerns. Are you still pursuing this Alvaro?\n\nI am, but not in the immediate future. I am very absorbed by trying to\nmake MERGE work again for resubmission, so I haven't had time to invest\nin these other patches. I hope I'll be able to return with a new\nversion of this for the January commitfest.\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\"On the other flipper, one wrong move and we're Fatal Exceptions\"\n(T.U.X.: Term Unit X - http://www.thelinuxreview.com/TUX/)\n\n\n","msg_date":"Sun, 7 Nov 2021 18:41:26 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"Thanks for the summary.\n\nAt Fri, 5 Nov 2021 22:15:49 +0100, Daniel Gustafsson <daniel@yesql.se> wrote in \n> 2490: Make message at end-of-recovery less scary\n> ================================================\n> This thread stalled, but has had recent interest. The patch no longer applies\n> so while the patch has support it will be marked as returned with feedback\n> during this CF unless revived.\n\nI'll soon repost a rebased version.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Mon, 08 Nov 2021 11:35:12 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"On Fri, Nov 05, 2021 at 10:15:49PM +0100, Daniel Gustafsson wrote:\n> 2161: standby recovery fails when re-replaying due to missing directory which\n> was removed in previous replay.\n> =============================================================================\n> Tom and Robert seem to be in agreement that parts of this patchset are good,\n> and that some parts are not. The thread has stalled and the patch no longer\n> apply, so unless someone feels like picking up the good pieces this seems like\n> a contender to close for now.\n\nThat's in my area, so I have signed up as reviewer.\n\n> 2518: Corruption during WAL replay\n> ==================================\n> Both Robert and Tom have given positive remarks about this, and recent review\n> concerns raised turned out to have been handled. Robert, Tom: any thoughts on\n> the latest posted version?\n\nThe latest comments are from a couple of weeks ago. I'll take a look.\n\n> 2602: ALTER SYSTEM READ { ONLY | WRITE }\n> ========================================\n> AFAIK Robert is actively working on this and have committed parts of it\n> already, so there isn't much to add. I'm personally quite excited about this\n> feature so I'm looking forward to (hopefully) seeing it go in.\n\nSo am I.\n--\nMichael","msg_date":"Mon, 8 Nov 2021 14:43:43 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"At Mon, 8 Nov 2021 14:43:43 +0900, Michael Paquier <michael@paquier.xyz> wrote in \n> On Fri, Nov 05, 2021 at 10:15:49PM +0100, Daniel Gustafsson wrote:\n> > 2161: standby recovery fails when re-replaying due to missing directory which\n> > was removed in previous replay.\n> > =============================================================================\n> > Tom and Robert seem to be in agreement that parts of this patchset are good,\n> > and that some parts are not. The thread has stalled and the patch no longer\n> > apply, so unless someone feels like picking up the good pieces this seems like\n> > a contender to close for now.\n> \n> That's in my area, so I have signed up as reviewer.\n\nI noticed I'm one of the author^^; I'll also look into the comments\nand try to address them.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Mon, 08 Nov 2021 15:07:43 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"On Mon, Nov 08, 2021 at 03:07:43PM +0900, Kyotaro Horiguchi wrote:\n> I noticed I'm one of the author^^; I'll also look into the comments\n> and try to address them.\n\nCool, thanks. I'll come back to this thread from this point, then.\n--\nMichael","msg_date":"Tue, 9 Nov 2021 11:00:41 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"On Fri, 5 Nov 2021 22:15:49 +0100\nDaniel Gustafsson <daniel@yesql.se> wrote:\n\n> 2138: Incremental Materialized View Maintenance\n> ===============================================\n> There seems to be concensus on the thread that this is a feature that we want,\n> and after initial design discussions there seems to be no disagreements with\n> the approach taken. The patch was marked ready for committer almost a year\n> ago, but have since been needs review (which seems correct). The size of the\n> patchset and the length of the thread make it hard to gauge just far away it\n> is, maybe the author or a review can summarize the current state and outline\n> what is left for it to be committable.\n\nThanks for taking care of this.\n\nI've responded to it in the following thread, and described the recent discussions,\ncurrent status, summary of IVM feature and design, and past discussions.\n\nhttps://www.postgresql.org/message-id/20211129144826.c9d42c1f61495c6983d8b6b1%40sraoss.co.jp\n\nRegards,\nYugo Nagata\n\n-- \nYugo NAGATA <nagata@sraoss.co.jp>\n\n\n","msg_date":"Mon, 29 Nov 2021 14:58:31 +0900","msg_from":"Yugo NAGATA <nagata@sraoss.co.jp>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":">\n> > 2138: Incremental Materialized View Maintenance\n>\n> I've responded to it in the following thread, and described the recent\n> discussions,\n> current status, summary of IVM feature and design, and past discussions.\n>\n\nIVM is a wonderful feature, but some features were omitted because of\nits complexity, so maybe IVM will spend more time to completely solve what\nit wants to do.\nI did not understand, and didn´t find on docs, if a Materialized View is a\ntable, why I cannot change a specific record ?\nBecause if I have a way to create and refresh the entire view and update a\nsingle record it would give me all power of IVM is trying to.\n\nregards,\nMarcos\n\n> 2138: Incremental Materialized View Maintenance\nI've responded to it in the following thread, and described the recent discussions,\ncurrent status, summary of IVM feature and design, and past discussions. IVM is a wonderful feature, but some features were omitted because of its complexity, so maybe IVM will spend more time to completely solve what it wants to do.I did not understand, and didn´t find on docs, if a Materialized View is a table, why I cannot change a specific record ?Because if I have a way to create and refresh the entire view and update a single record it would give me all power of IVM is trying to. regards,Marcos","msg_date":"Mon, 29 Nov 2021 09:03:06 -0300","msg_from":"Marcos Pegoraro <marcos@f10.com.br>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"On Mon, 29 Nov 2021 09:03:06 -0300\nMarcos Pegoraro <marcos@f10.com.br> wrote:\n\n> >\n> > > 2138: Incremental Materialized View Maintenance\n> >\n> > I've responded to it in the following thread, and described the recent\n> > discussions,\n> > current status, summary of IVM feature and design, and past discussions.\n> >\n> \n> IVM is a wonderful feature, but some features were omitted because of\n> its complexity, so maybe IVM will spend more time to completely solve what\n> it wants to do.\n> I did not understand, and didn´t find on docs, if a Materialized View is a\n> table, why I cannot change a specific record ?\n> Because if I have a way to create and refresh the entire view and update a\n> single record it would give me all power of IVM is trying to.\n\nI think the reason why we can't update a materialized view directly is because\nit is basically a \"view\" and it should not contains any data irrelevant to its\ndefinition and underlying tables. If we would have a feature to update a\nmaterialized view direcly, maybe, it should behave as updatable-view as well\nas normal (virtual) views, although I am not sure....\n\n\n-- \nYugo NAGATA <nagata@sraoss.co.jp>\n\n\n","msg_date":"Wed, 1 Dec 2021 16:06:56 +0900","msg_from":"Yugo NAGATA <nagata@sraoss.co.jp>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":">\n> I think the reason why we can't update a materialized view directly is\n> because\n> it is basically a \"view\" and it should not contains any data irrelevant to\n> its\n> definition and underlying tables. If we would have a feature to update a\n> materialized view direcly, maybe, it should behave as updatable-view as\n> well\n> as normal (virtual) views, although I am not sure....\n>\n\nWell, I didn´t find any place where is detailed why those tables are not\nupdatable.\nAnd would be fine to be updated through triggers or cron jobs until IVM is\navailable.\nCheckValidRowMarkRel just gives an exception \"cannot lock rows in\nmaterialized view ...\", but why ?\nWhat are the differences between Materialized Views and tables ?\n\nI think the reason why we can't update a materialized view directly is because\nit is basically a \"view\" and it should not contains any data irrelevant to its\ndefinition and underlying tables. If we would have a feature to update a\nmaterialized view direcly, maybe, it should behave as updatable-view as well\nas normal (virtual) views, although I am not sure.... Well, I didn´t find any place where is detailed why those tables are not updatable.And would be fine to be updated through triggers or cron jobs until IVM is available.CheckValidRowMarkRel just gives an exception \"cannot lock rows in materialized view ...\", but why ?What are the differences between Materialized Views and tables ?","msg_date":"Wed, 1 Dec 2021 09:14:31 -0300","msg_from":"Marcos Pegoraro <marcos@f10.com.br>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"> On 6 Nov 2021, at 17:20, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n> \n> On 11/5/21 22:15, Daniel Gustafsson wrote:\n>> ...\n>> 1651: GROUP BY optimization\n>> ===========================\n>> This is IMO a desired optimization, and after all the heavy lifting by Tomas\n>> the patch seems to be in pretty good shape.\n> \n> I agree it's desirable. To move the patch forward, I need some feedback on the costing questions, mentioned in my last review, and the overall approach to planning (considering multiple group by variants).\n\nI'm moving this to the next CF with the hope that more feedback is received.\nIf you think \"Ready for Committer\" is a more suitable status, please do update\nit as I'm leaving it in Needs Review for now not having the full context.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Thu, 2 Dec 2021 12:06:48 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"\n\nOn 12/2/21 12:06, Daniel Gustafsson wrote:\n>> On 6 Nov 2021, at 17:20, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n>>\n>> On 11/5/21 22:15, Daniel Gustafsson wrote:\n>>> ...\n>>> 1651: GROUP BY optimization\n>>> ===========================\n>>> This is IMO a desired optimization, and after all the heavy lifting by Tomas\n>>> the patch seems to be in pretty good shape.\n>>\n>> I agree it's desirable. To move the patch forward, I need some feedback on the costing questions, mentioned in my last review, and the overall approach to planning (considering multiple group by variants).\n> \n> I'm moving this to the next CF with the hope that more feedback is received.\n> If you think \"Ready for Committer\" is a more suitable status, please do update\n> it as I'm leaving it in Needs Review for now not having the full context.\n\nI don't think anyone reviewed the costing, so I think \"needs review\" is \nappropriate.\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 2 Dec 2021 21:01:53 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"},{"msg_contents":"On Wed, 1 Dec 2021 09:14:31 -0300\nMarcos Pegoraro <marcos@f10.com.br> wrote:\n\n> >\n> > I think the reason why we can't update a materialized view directly is\n> > because\n> > it is basically a \"view\" and it should not contains any data irrelevant to\n> > its\n> > definition and underlying tables. If we would have a feature to update a\n> > materialized view direcly, maybe, it should behave as updatable-view as\n> > well\n> > as normal (virtual) views, although I am not sure....\n> >\n> \n> Well, I didn´t find any place where is detailed why those tables are not\n> updatable.\n> And would be fine to be updated through triggers or cron jobs until IVM is\n> available.\n> CheckValidRowMarkRel just gives an exception \"cannot lock rows in\n> materialized view ...\", but why ?\n> What are the differences between Materialized Views and tables ?\n\nIt would be that materialized views are related to its definition query\nand expected to have contents that is consistent with it, at least at some\npoint in time, I think.\n\nIn order to allow triggers to update materialized views, we'd need to\nmake OpenMatViewIncrementalMaintenance and CloseMatViewIncrementalMaintenance\npublic since there are static functions in matview.c. However, there is a\nconcern that it will make the contents of a materialized view completely\nunreliable [1]. Therefore, if we do it, we would need privilege management\nin some way.\n\n[1] https://www.postgresql.org/message-id/flat/CACjxUsP8J6bA4RKxbmwujTVMwMZrgR3AZ7yP5F2XkB-f9w7K7Q%40mail.gmail.com#efbee336d7651ce39bc5ff9574f92002\n\nRegards,\nYugo Nagata\n-- \nYugo NAGATA <nagata@sraoss.co.jp>\n\n\n","msg_date":"Fri, 3 Dec 2021 15:52:22 +0900","msg_from":"Yugo NAGATA <nagata@sraoss.co.jp>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 1"}],"string":"[\n {\n \"msg_contents\": \"We have amassed quite a lot of patches in the CF app, and while Jaime did a\\nvery good job closing patches in the last CF there is still a lot to go\\nthrough. I've attempted a brief per-patch triage here to see where we are with\\nthese. Reading every email in the threads for these patches is a herculean\\ntask, so I may very well have misunderstood things. If so I apologize and\\nplease do jump and correct me. If not, then please also jump in and comment\\nbelow (or on the relevant threads) to try and get closure.\\n\\nWhile I don't have the expertise to replicate Andres' famous patch triage [0]\\nfrom a few years back, below is at least a try. These are the patches with 7\\nor more commitfests under their belts (the record holder having 18), the next\\ninstallment will be 6 CF's and down limited by when the Shiraz runs out.\\n\\n\\n1608: schema variables, LET command\\n===================================\\nAfter 18 CF's and two very long threads it seems to be nearing completion\\njudging by Tomas' review. There is an ask in that review for a second pass\\nover the docs by a native speaker, any takers?\\n\\n\\n1741: Index Skip Scan\\n=====================\\nAn often requested feature which has proven hard to reach consensus on an\\nimplementation for. The thread(s) have stalled since May, is there any hope of\\ntaking this further? Where do we go from here with this patch?\\n\\n\\n1712: Remove self join on a unique column\\n=========================================\\nThis has moved from \\\"don't write bad queries\\\" to \\\"maybe we should do something\\nabout this\\\". It seems like there is concensus that it can be worth paying the\\nadded planner cost for this, especially if gated by a GUC to keep it out of\\ninstallations where it doesn't apply. The regression on large join queries\\nhasn't been benchmarked it seems (or I missed it), but the patch seems to have\\nmatured and be quite ready. Any takers on getting it across the finish line?\\n\\n\\n2161: standby recovery fails when re-replaying due to missing directory which\\nwas removed in previous replay.\\n=============================================================================\\nTom and Robert seem to be in agreement that parts of this patchset are good,\\nand that some parts are not. The thread has stalled and the patch no longer\\napply, so unless someone feels like picking up the good pieces this seems like\\na contender to close for now.\\n\\n\\n2138: Incremental Materialized View Maintenance\\n===============================================\\nThere seems to be concensus on the thread that this is a feature that we want,\\nand after initial design discussions there seems to be no disagreements with\\nthe approach taken. The patch was marked ready for committer almost a year\\nago, but have since been needs review (which seems correct). The size of the\\npatchset and the length of the thread make it hard to gauge just far away it\\nis, maybe the author or a review can summarize the current state and outline\\nwhat is left for it to be committable.\\n\\n\\n2215: pg_upgrade fails with non-standard ACL\\n============================================\\nThis thread has been stalled since March with open questions from Noah, but the\\npatch seems to be in pretty good shape. Do you have any thoughts on the state\\nof this patch Noah? Seems like a pretty good fix to get in.\\n\\n\\n2218: Implement INSERT SET syntax\\n=================================\\nThe author has kept this patch updated, and has seemingly updated according to\\nthe review comments. Tom: do you have any opinions on whether the updated\\nversion addresses your concerns wrt the SELECT rewrite?\\n\\n\\n2048: SQL:2011 application time\\n===============================\\nThe patch no longer applies, but there is a semi-active review discussion\\nongoing. It doesn't seem like the patch is nearing completion, but the feature\\nwould no doubt be welcome by many. Is this realistic for 15?\\n\\n\\n2316: WITH SYSTEM VERSIONING Temporal Tables\\n============================================\\nThis patch no longer applies, and judging by the thread there is a pretty major\\nre-design ongoing. I think we should close this entry and open a new one when\\nthere is a new patch rather than push it along. What are your thoughts Simon?\\n\\n\\n2096: psql - add SHOW_ALL_RESULTS option\\n========================================\\nPeter posted an updated version of Fabiens patch about a month ago (which at\\nthis point no longer applies) fixing a few issues, but also point at old review\\ncomments still unadressed. Since this was pushed, but had to be reverted, I\\nassume there is a desire for the feature but it seems to need more work still.\\n\\n\\n1651: GROUP BY optimization\\n===========================\\nThis is IMO a desired optimization, and after all the heavy lifting by Tomas\\nthe patch seems to be in pretty good shape.\\n\\n\\n2377: pg_ls_* functions for showing metadata and recurse (pg_ls_tmpdir to show\\nshared filesets)\\n==============================================================================\\nThe question of what to do with lstat() on Windows is still left unanswered,\\nbut the patchset has been split to up to be able to avoid it. Stephen and Tom,\\nhaving done prior reviews do you have any thoughts on this?\\n\\n\\n2349: global temporary table\\n============================\\nGTT has been up for discussion numerous times in tbe past, and I can't judge\\nwhether this proposal has a better chance than previous ones. I do note the\\npatch has a number of crashes reported lately, and no reviews from senior\\ncontributors in a while, making it seem unlikely to be committed in this CF.\\nSince the patch is very big, can it be teased apart and committed separately\\nfor easier review?\\n\\n\\n2433: Erase the distinctClause if the result is unique by definition\\n====================================================================\\n(parts of) The approach taken in this patch has been objected against in favor\\nof work that Tom has proposed. Until that work materialize this patch is\\nblocked, and thus I think we are better of closing it and re-opening it when it\\ngets unstuck. Unless Tom has plans to hack on this shortly?\\n\\n\\n2482: jsonpath syntax extensions\\n================================\\nThis thread has stalled, with no reviews so far a good year and half after\\nbeing. Rather than continue to push this along, should we instead close it due\\nto lack of interest? It's pretty big, can it be broken up into smaller\\nfeatures for easier review?\\n\\n\\n2490: Make message at end-of-recovery less scary\\n================================================\\nThis thread stalled, but has had recent interest. The patch no longer applies\\nso while the patch has support it will be marked as returned with feedback\\nduring this CF unless revived.\\n\\n\\n2573: pg_dump - read data for some options from external file\\n=============================================================\\nThis patch has been heavily debated, but I think we've landed in a compromise\\nthat has been agreed upon. The gist of the debate has been around the file\\nformat and it's extensibility, the final version proposing an extensible yet\\nsimple format which can be referenced from a larger configfile (rather than\\nmaking the configfile now). Having done a lot of review myself I think the\\npatch is ready with one exception: IMO the parsing code isn't maintainable\\nenough to be committed (if another committer disagree with that, feel free to\\njump in). I have on my TODO to implement it as a Bison parser to see what that\\nwould look like.\\n\\n\\n2553: should INSERT SELECT use a BulkInsertState?\\n==================================================\\nThe conclusion on the thread seems to be that there is little support for it\\ndue to the risk of regressions, and that closing the patch might be the best\\ncourse of action. If anyone is in favor of going ahead with it, now is the\\ntime to speak up.\\n\\n\\n2518: Corruption during WAL replay\\n==================================\\nBoth Robert and Tom have given positive remarks about this, and recent review\\nconcerns raised turned out to have been handled. Robert, Tom: any thoughts on\\nthe latest posted version?\\n\\n\\n2601: Fast COPY FROM command for the foreign tables\\n===================================================\\nThis approach taken in this patch has stabilized and the benchmarks posted are\\nvery promising. It seems pretty uncontroversial to add COPY to the FDW API SQL\\nspec wise. Amit, Justin and Takayuki-san has done a lot of review and the\\npatch is marked Ready for Committer. Any committer interested in taking this\\non? Tomas?\\n\\n\\n2602: ALTER SYSTEM READ { ONLY | WRITE }\\n========================================\\nAFAIK Robert is actively working on this and have committed parts of it\\nalready, so there isn't much to add. I'm personally quite excited about this\\nfeature so I'm looking forward to (hopefully) seeing it go in. Is the full\\nfeature likely to make 15?\\n\\n\\n2627: More scalable multixacts buffers and locking\\n==================================================\\nSkimming the thread makes it seem like this patch is still undergoing some\\nfairly big changes. Thomas and Andrey, how far along is this from RFC?\\n\\n\\n2716: fix spinlock contention in LogwrtResult\\n=============================================\\nThis addresses a bottleneck which definitely seems like one we want to fix, I\\ndon't have a hard time imagining it impacting other production usecases then\\nthe reported one. The goalposts moved early in the thread, and while there has\\nbeen work on it the patch has stalled on Andres' review which raised some\\nconcerns. Are you still pursuing this Alvaro?\\n\\n\\n2684: enhancing plpgsql API for debugging and tracing\\n======================================================\\nThis hasn't seen much review, but what has been performed has been positive.\\nIt does add some weight to plpgsql by passing the NS around, but maybe the\\neffect is negligable. Given the size and complexity of the test module to use\\nthis, it's perhaps not obvious to see who the target audience is. Maybe a\\nbetter option would be to polish the test module and propose it for contrib/\\nalong with the patch, and write the tests on using said module?\\n\\n\\n2710: Fix behavior of geo_ops when NaN is involved\\n==================================================\\nTwo thirds of this bugfix patchset have been committed already, with the thread\\nstalling on the remaining bit. Tom, Georgios: any thoughts on the last patch?\\n\\n\\n2694: Automatic HASH and LIST partition creation\\n================================================\\nThe subject of automatic partition creation is pretty hot, so getting something\\ndone in this area would be good. This proposal has gotten criticism over the\\nsyntax chosen, and the thread has effectively stalled after Robert's review in\\nJuly. It seems like this patch should be returned with feedback, and any new\\nattempt at this should start by just discussing the syntax and only once agreed\\nupon hack on the code.\\n\\n\\nThanks for reading all this way, I hope it did something to help progress.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n[0] https://www.postgresql.org/message-id/flat/20180302075242.yfqkcgzbrmjysboa%40alap3.anarazel.de#b5015aab4ba3a95b67636faef74a1d51\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Nov 2021 22:15:49 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"Thanks for taking care of this.\\n\\n> 2433: Erase the distinctClause if the result is unique by definition\\n> ====================================================================\\n> (parts of) The approach taken in this patch has been objected against in favor\\n> of work that Tom has proposed.\\n\\nActually no exactly. The 90% of the code is about UniqueKey and UniqueKey\\nneeds a notnulls representation of Var during join, Tom is objecting\\nthat part only.\\n\\n\\n> Until that work materialize this patch is\\n> blocked, and thus I think we are better of closing it and re-opening it when it\\n> gets unstuck. Unless Tom has plans to hack on this shortly?\\n\\nBoth Tom and David have provided many insights on this topic. I'd like to hear\\nthe decision from Tom and David.\\n\\n-- \\nBest Regards\\nAndy Fan (https://www.aliyun.com/)\\n\\n\\n\",\n \"msg_date\": \"Sat, 6 Nov 2021 09:00:33 +0800\",\n \"msg_from\": \"Andy Fan <zhihui.fan1213@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"> 1741: Index Skip Scan\\n> =====================\\n> An often requested feature which has proven hard to reach consensus on an\\n> implementation for. The thread(s) have stalled since May,\\n\\nThis statement is not accurate. Peter started a new thread in [1] for this\\ntopic last month and then we had a discussion. Thanks for taking\\ncare of this.\\n\\n[1] https://www.postgresql.org/message-id/flat/CAH2-WzmUscvoxVkokHxP%3DuPTDjSi0tJkFpUPD-CeA35dvn-CMw%40mail.gmail.com\\n\\nis there any hope of\\n> taking this further? Where do we go from here with this patch?\\n>\\n\\n-- \\nBest Regards\\nAndy Fan (https://www.aliyun.com/)\\n\\n\\n\",\n \"msg_date\": \"Sat, 6 Nov 2021 09:12:39 +0800\",\n \"msg_from\": \"Andy Fan <zhihui.fan1213@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"On 11/5/21 22:15, Daniel Gustafsson wrote:\\n> ...\\n> 1651: GROUP BY optimization\\n> ===========================\\n> This is IMO a desired optimization, and after all the heavy lifting by Tomas\\n> the patch seems to be in pretty good shape.\\n> \\n\\nI agree it's desirable. To move the patch forward, I need some feedback \\non the costing questions, mentioned in my last review, and the overall \\napproach to planning (considering multiple group by variants).\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sat, 6 Nov 2021 17:20:13 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 6, 2021 at 6:16 AM Daniel Gustafsson <daniel@yesql.se> wrote:\\n> 2601: Fast COPY FROM command for the foreign tables\\n> ===================================================\\n> This approach taken in this patch has stabilized and the benchmarks posted are\\n> very promising. It seems pretty uncontroversial to add COPY to the FDW API SQL\\n> spec wise. Amit, Justin and Takayuki-san has done a lot of review and the\\n> patch is marked Ready for Committer. Any committer interested in taking this\\n> on? Tomas?\\n\\nI’d like to work on this unless Tomas (or anyone else) want.\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n\",\n \"msg_date\": \"Sun, 7 Nov 2021 18:26:46 +0900\",\n \"msg_from\": \"Etsuro Fujita <etsuro.fujita@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"> On 7 Nov 2021, at 10:26, Etsuro Fujita <etsuro.fujita@gmail.com> wrote:\\n> \\n> On Sat, Nov 6, 2021 at 6:16 AM Daniel Gustafsson <daniel@yesql.se> wrote:\\n>> 2601: Fast COPY FROM command for the foreign tables\\n>> ===================================================\\n>> This approach taken in this patch has stabilized and the benchmarks posted are\\n>> very promising. It seems pretty uncontroversial to add COPY to the FDW API SQL\\n>> spec wise. Amit, Justin and Takayuki-san has done a lot of review and the\\n>> patch is marked Ready for Committer. Any committer interested in taking this\\n>> on? Tomas?\\n> \\n> I’d like to work on this unless Tomas (or anyone else) want.\\n\\nGreat, thanks for picking it up!\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Sun, 7 Nov 2021 17:05:17 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"> On 6 Nov 2021, at 17:20, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n> \\n> On 11/5/21 22:15, Daniel Gustafsson wrote:\\n>> ...\\n>> 1651: GROUP BY optimization\\n>> ===========================\\n>> This is IMO a desired optimization, and after all the heavy lifting by Tomas\\n>> the patch seems to be in pretty good shape.\\n> \\n> I agree it's desirable. To move the patch forward, I need some feedback on the\\n> costing questions,\\n\\n\\nTom or David perhaps have some thoughts on that?\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Sun, 7 Nov 2021 17:07:06 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"> On 6 Nov 2021, at 02:12, Andy Fan <zhihui.fan1213@gmail.com> wrote:\\n> \\n>> 1741: Index Skip Scan\\n>> =====================\\n>> An often requested feature which has proven hard to reach consensus on an\\n>> implementation for. The thread(s) have stalled since May,\\n> \\n> This statement is not accurate. Peter started a new thread in [1] for this\\n> topic last month and then we had a discussion. Thanks for taking\\n> care of this.\\n\\nAha, I had missed that. Thanks to Dimitry for attaching this thread to the CF\\nentry to make that clearer going forward.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Sun, 7 Nov 2021 17:11:39 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-05, Daniel Gustafsson wrote:\\n\\n> 2716: fix spinlock contention in LogwrtResult\\n> =============================================\\n> This addresses a bottleneck which definitely seems like one we want to fix, I\\n> don't have a hard time imagining it impacting other production usecases then\\n> the reported one. The goalposts moved early in the thread, and while there has\\n> been work on it the patch has stalled on Andres' review which raised some\\n> concerns. Are you still pursuing this Alvaro?\\n\\nI am, but not in the immediate future. I am very absorbed by trying to\\nmake MERGE work again for resubmission, so I haven't had time to invest\\nin these other patches. I hope I'll be able to return with a new\\nversion of this for the January commitfest.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\\"On the other flipper, one wrong move and we're Fatal Exceptions\\\"\\n(T.U.X.: Term Unit X - http://www.thelinuxreview.com/TUX/)\\n\\n\\n\",\n \"msg_date\": \"Sun, 7 Nov 2021 18:41:26 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"Thanks for the summary.\\n\\nAt Fri, 5 Nov 2021 22:15:49 +0100, Daniel Gustafsson <daniel@yesql.se> wrote in \\n> 2490: Make message at end-of-recovery less scary\\n> ================================================\\n> This thread stalled, but has had recent interest. The patch no longer applies\\n> so while the patch has support it will be marked as returned with feedback\\n> during this CF unless revived.\\n\\nI'll soon repost a rebased version.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Mon, 08 Nov 2021 11:35:12 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 05, 2021 at 10:15:49PM +0100, Daniel Gustafsson wrote:\\n> 2161: standby recovery fails when re-replaying due to missing directory which\\n> was removed in previous replay.\\n> =============================================================================\\n> Tom and Robert seem to be in agreement that parts of this patchset are good,\\n> and that some parts are not. The thread has stalled and the patch no longer\\n> apply, so unless someone feels like picking up the good pieces this seems like\\n> a contender to close for now.\\n\\nThat's in my area, so I have signed up as reviewer.\\n\\n> 2518: Corruption during WAL replay\\n> ==================================\\n> Both Robert and Tom have given positive remarks about this, and recent review\\n> concerns raised turned out to have been handled. Robert, Tom: any thoughts on\\n> the latest posted version?\\n\\nThe latest comments are from a couple of weeks ago. I'll take a look.\\n\\n> 2602: ALTER SYSTEM READ { ONLY | WRITE }\\n> ========================================\\n> AFAIK Robert is actively working on this and have committed parts of it\\n> already, so there isn't much to add. I'm personally quite excited about this\\n> feature so I'm looking forward to (hopefully) seeing it go in.\\n\\nSo am I.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 8 Nov 2021 14:43:43 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"At Mon, 8 Nov 2021 14:43:43 +0900, Michael Paquier <michael@paquier.xyz> wrote in \\n> On Fri, Nov 05, 2021 at 10:15:49PM +0100, Daniel Gustafsson wrote:\\n> > 2161: standby recovery fails when re-replaying due to missing directory which\\n> > was removed in previous replay.\\n> > =============================================================================\\n> > Tom and Robert seem to be in agreement that parts of this patchset are good,\\n> > and that some parts are not. The thread has stalled and the patch no longer\\n> > apply, so unless someone feels like picking up the good pieces this seems like\\n> > a contender to close for now.\\n> \\n> That's in my area, so I have signed up as reviewer.\\n\\nI noticed I'm one of the author^^; I'll also look into the comments\\nand try to address them.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Mon, 08 Nov 2021 15:07:43 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 08, 2021 at 03:07:43PM +0900, Kyotaro Horiguchi wrote:\\n> I noticed I'm one of the author^^; I'll also look into the comments\\n> and try to address them.\\n\\nCool, thanks. I'll come back to this thread from this point, then.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 9 Nov 2021 11:00:41 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"On Fri, 5 Nov 2021 22:15:49 +0100\\nDaniel Gustafsson <daniel@yesql.se> wrote:\\n\\n> 2138: Incremental Materialized View Maintenance\\n> ===============================================\\n> There seems to be concensus on the thread that this is a feature that we want,\\n> and after initial design discussions there seems to be no disagreements with\\n> the approach taken. The patch was marked ready for committer almost a year\\n> ago, but have since been needs review (which seems correct). The size of the\\n> patchset and the length of the thread make it hard to gauge just far away it\\n> is, maybe the author or a review can summarize the current state and outline\\n> what is left for it to be committable.\\n\\nThanks for taking care of this.\\n\\nI've responded to it in the following thread, and described the recent discussions,\\ncurrent status, summary of IVM feature and design, and past discussions.\\n\\nhttps://www.postgresql.org/message-id/20211129144826.c9d42c1f61495c6983d8b6b1%40sraoss.co.jp\\n\\nRegards,\\nYugo Nagata\\n\\n-- \\nYugo NAGATA <nagata@sraoss.co.jp>\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Nov 2021 14:58:31 +0900\",\n \"msg_from\": \"Yugo NAGATA <nagata@sraoss.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \">\\n> > 2138: Incremental Materialized View Maintenance\\n>\\n> I've responded to it in the following thread, and described the recent\\n> discussions,\\n> current status, summary of IVM feature and design, and past discussions.\\n>\\n\\nIVM is a wonderful feature, but some features were omitted because of\\nits complexity, so maybe IVM will spend more time to completely solve what\\nit wants to do.\\nI did not understand, and didn´t find on docs, if a Materialized View is a\\ntable, why I cannot change a specific record ?\\nBecause if I have a way to create and refresh the entire view and update a\\nsingle record it would give me all power of IVM is trying to.\\n\\nregards,\\nMarcos\\n\\n> 2138: Incremental Materialized View Maintenance\\nI've responded to it in the following thread, and described the recent discussions,\\ncurrent status, summary of IVM feature and design, and past discussions. IVM is a wonderful feature, but some features were omitted because of its complexity, so maybe IVM will spend more time to completely solve what it wants to do.I did not understand, and didn´t find on docs, if a Materialized View is a table, why I cannot change a specific record ?Because if I have a way to create and refresh the entire view and update a single record it would give me all power of IVM is trying to. regards,Marcos\",\n \"msg_date\": \"Mon, 29 Nov 2021 09:03:06 -0300\",\n \"msg_from\": \"Marcos Pegoraro <marcos@f10.com.br>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"On Mon, 29 Nov 2021 09:03:06 -0300\\nMarcos Pegoraro <marcos@f10.com.br> wrote:\\n\\n> >\\n> > > 2138: Incremental Materialized View Maintenance\\n> >\\n> > I've responded to it in the following thread, and described the recent\\n> > discussions,\\n> > current status, summary of IVM feature and design, and past discussions.\\n> >\\n> \\n> IVM is a wonderful feature, but some features were omitted because of\\n> its complexity, so maybe IVM will spend more time to completely solve what\\n> it wants to do.\\n> I did not understand, and didn´t find on docs, if a Materialized View is a\\n> table, why I cannot change a specific record ?\\n> Because if I have a way to create and refresh the entire view and update a\\n> single record it would give me all power of IVM is trying to.\\n\\nI think the reason why we can't update a materialized view directly is because\\nit is basically a \\\"view\\\" and it should not contains any data irrelevant to its\\ndefinition and underlying tables. If we would have a feature to update a\\nmaterialized view direcly, maybe, it should behave as updatable-view as well\\nas normal (virtual) views, although I am not sure....\\n\\n\\n-- \\nYugo NAGATA <nagata@sraoss.co.jp>\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Dec 2021 16:06:56 +0900\",\n \"msg_from\": \"Yugo NAGATA <nagata@sraoss.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \">\\n> I think the reason why we can't update a materialized view directly is\\n> because\\n> it is basically a \\\"view\\\" and it should not contains any data irrelevant to\\n> its\\n> definition and underlying tables. If we would have a feature to update a\\n> materialized view direcly, maybe, it should behave as updatable-view as\\n> well\\n> as normal (virtual) views, although I am not sure....\\n>\\n\\nWell, I didn´t find any place where is detailed why those tables are not\\nupdatable.\\nAnd would be fine to be updated through triggers or cron jobs until IVM is\\navailable.\\nCheckValidRowMarkRel just gives an exception \\\"cannot lock rows in\\nmaterialized view ...\\\", but why ?\\nWhat are the differences between Materialized Views and tables ?\\n\\nI think the reason why we can't update a materialized view directly is because\\nit is basically a \\\"view\\\" and it should not contains any data irrelevant to its\\ndefinition and underlying tables. If we would have a feature to update a\\nmaterialized view direcly, maybe, it should behave as updatable-view as well\\nas normal (virtual) views, although I am not sure.... Well, I didn´t find any place where is detailed why those tables are not updatable.And would be fine to be updated through triggers or cron jobs until IVM is available.CheckValidRowMarkRel just gives an exception \\\"cannot lock rows in materialized view ...\\\", but why ?What are the differences between Materialized Views and tables ?\",\n \"msg_date\": \"Wed, 1 Dec 2021 09:14:31 -0300\",\n \"msg_from\": \"Marcos Pegoraro <marcos@f10.com.br>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"> On 6 Nov 2021, at 17:20, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n> \\n> On 11/5/21 22:15, Daniel Gustafsson wrote:\\n>> ...\\n>> 1651: GROUP BY optimization\\n>> ===========================\\n>> This is IMO a desired optimization, and after all the heavy lifting by Tomas\\n>> the patch seems to be in pretty good shape.\\n> \\n> I agree it's desirable. To move the patch forward, I need some feedback on the costing questions, mentioned in my last review, and the overall approach to planning (considering multiple group by variants).\\n\\nI'm moving this to the next CF with the hope that more feedback is received.\\nIf you think \\\"Ready for Committer\\\" is a more suitable status, please do update\\nit as I'm leaving it in Needs Review for now not having the full context.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 2 Dec 2021 12:06:48 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 12/2/21 12:06, Daniel Gustafsson wrote:\\n>> On 6 Nov 2021, at 17:20, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n>>\\n>> On 11/5/21 22:15, Daniel Gustafsson wrote:\\n>>> ...\\n>>> 1651: GROUP BY optimization\\n>>> ===========================\\n>>> This is IMO a desired optimization, and after all the heavy lifting by Tomas\\n>>> the patch seems to be in pretty good shape.\\n>>\\n>> I agree it's desirable. To move the patch forward, I need some feedback on the costing questions, mentioned in my last review, and the overall approach to planning (considering multiple group by variants).\\n> \\n> I'm moving this to the next CF with the hope that more feedback is received.\\n> If you think \\\"Ready for Committer\\\" is a more suitable status, please do update\\n> it as I'm leaving it in Needs Review for now not having the full context.\\n\\nI don't think anyone reviewed the costing, so I think \\\"needs review\\\" is \\nappropriate.\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 2 Dec 2021 21:01:53 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n },\n {\n \"msg_contents\": \"On Wed, 1 Dec 2021 09:14:31 -0300\\nMarcos Pegoraro <marcos@f10.com.br> wrote:\\n\\n> >\\n> > I think the reason why we can't update a materialized view directly is\\n> > because\\n> > it is basically a \\\"view\\\" and it should not contains any data irrelevant to\\n> > its\\n> > definition and underlying tables. If we would have a feature to update a\\n> > materialized view direcly, maybe, it should behave as updatable-view as\\n> > well\\n> > as normal (virtual) views, although I am not sure....\\n> >\\n> \\n> Well, I didn´t find any place where is detailed why those tables are not\\n> updatable.\\n> And would be fine to be updated through triggers or cron jobs until IVM is\\n> available.\\n> CheckValidRowMarkRel just gives an exception \\\"cannot lock rows in\\n> materialized view ...\\\", but why ?\\n> What are the differences between Materialized Views and tables ?\\n\\nIt would be that materialized views are related to its definition query\\nand expected to have contents that is consistent with it, at least at some\\npoint in time, I think.\\n\\nIn order to allow triggers to update materialized views, we'd need to\\nmake OpenMatViewIncrementalMaintenance and CloseMatViewIncrementalMaintenance\\npublic since there are static functions in matview.c. However, there is a\\nconcern that it will make the contents of a materialized view completely\\nunreliable [1]. Therefore, if we do it, we would need privilege management\\nin some way.\\n\\n[1] https://www.postgresql.org/message-id/flat/CACjxUsP8J6bA4RKxbmwujTVMwMZrgR3AZ7yP5F2XkB-f9w7K7Q%40mail.gmail.com#efbee336d7651ce39bc5ff9574f92002\\n\\nRegards,\\nYugo Nagata\\n-- \\nYugo NAGATA <nagata@sraoss.co.jp>\\n\\n\\n\",\n \"msg_date\": \"Fri, 3 Dec 2021 15:52:22 +0900\",\n \"msg_from\": \"Yugo NAGATA <nagata@sraoss.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 1\"\n }\n]","isConversational":false}}},{"rowIdx":53049,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"First draft is up at\n\nhttps://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=39387228c8b7043d168bada0c64e9f59e83285f5\n\nAs usual, please send any comments/corrections by Sunday.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 05 Nov 2021 20:27:49 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Draft release notes for next week's releases"},{"msg_contents":"On Fri, Nov 05, 2021 at 08:27:49PM -0400, Tom Lane wrote:\n> First draft is up at\n> https://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=39387228c8b7043d168bada0c64e9f59e83285f5\n> As usual, please send any comments/corrections by Sunday.\n\n+ one more row than requested, so tha it can determine whether the\n\nA little typo: tha it\n\n+ Ensure prepared transactions are properly accounted for during\n+ promotion of a standby server (Michael Paquier, Andres Freund)\n\nEnsure *that*\n\n+ It seems unlikely that this bug has yet been hit in practice, as it\n\nRemove \"yet\" ?\n\n+ Previously this was allowed, but the collation effectively vanished\n+ into the ether because of the way collation lookup works: you could\n+ not use the collation, nor even drop it.\n\nIs \"vanished into the ether\" normal language style for the release notes ?\n\n+ <para>\n+ Unnecessarily stupid back-tracking logic could result in exponential\n+ time spent looking for a match. Fortunately the problem is masked\n+ in most cases by other optimizations; but it is possible to\n+ demonstrate it with fairly simple, if contrived, regexps.\n+ </para>\n+ </listitem>\n\nSame - it seems like this borrows too much from the commit message text (and\nstyle).\n\n-- \nJustin\n\n\n","msg_date":"Fri, 5 Nov 2021 21:48:48 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Draft release notes for next week's releases"},{"msg_contents":"Justin Pryzby <pryzby@telsasoft.com> writes:\n> On Fri, Nov 05, 2021 at 08:27:49PM -0400, Tom Lane wrote:\n>> + one more row than requested, so tha it can determine whether the\n\n> A little typo: tha it\n> [ etc ]\n\nUgh. These notes were prepared in a bit more haste than usual,\nand I guess it shows :-(. Will fix, thanks for proofreading!\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 06 Nov 2021 09:34:01 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Draft release notes for next week's releases"},{"msg_contents":"On Fri, Nov 5, 2021 at 5:27 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> As usual, please send any comments/corrections by Sunday.\n\nYou probably won't miss this, but just in case: Alexander's very\nrecent \"Reset lastOverflowedXid on standby when needed\" commit needs\nto be listed now.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Sat, 6 Nov 2021 11:43:40 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: Draft release notes for next week's releases"},{"msg_contents":"Peter Geoghegan <pg@bowt.ie> writes:\n> You probably won't miss this, but just in case: Alexander's very\n> recent \"Reset lastOverflowedXid on standby when needed\" commit needs\n> to be listed now.\n\nYup, I saw it.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 06 Nov 2021 15:03:47 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Draft release notes for next week's releases"}],"string":"[\n {\n \"msg_contents\": \"First draft is up at\\n\\nhttps://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=39387228c8b7043d168bada0c64e9f59e83285f5\\n\\nAs usual, please send any comments/corrections by Sunday.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 05 Nov 2021 20:27:49 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Draft release notes for next week's releases\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 05, 2021 at 08:27:49PM -0400, Tom Lane wrote:\\n> First draft is up at\\n> https://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=39387228c8b7043d168bada0c64e9f59e83285f5\\n> As usual, please send any comments/corrections by Sunday.\\n\\n+ one more row than requested, so tha it can determine whether the\\n\\nA little typo: tha it\\n\\n+ Ensure prepared transactions are properly accounted for during\\n+ promotion of a standby server (Michael Paquier, Andres Freund)\\n\\nEnsure *that*\\n\\n+ It seems unlikely that this bug has yet been hit in practice, as it\\n\\nRemove \\\"yet\\\" ?\\n\\n+ Previously this was allowed, but the collation effectively vanished\\n+ into the ether because of the way collation lookup works: you could\\n+ not use the collation, nor even drop it.\\n\\nIs \\\"vanished into the ether\\\" normal language style for the release notes ?\\n\\n+ <para>\\n+ Unnecessarily stupid back-tracking logic could result in exponential\\n+ time spent looking for a match. Fortunately the problem is masked\\n+ in most cases by other optimizations; but it is possible to\\n+ demonstrate it with fairly simple, if contrived, regexps.\\n+ </para>\\n+ </listitem>\\n\\nSame - it seems like this borrows too much from the commit message text (and\\nstyle).\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Nov 2021 21:48:48 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Draft release notes for next week's releases\"\n },\n {\n \"msg_contents\": \"Justin Pryzby <pryzby@telsasoft.com> writes:\\n> On Fri, Nov 05, 2021 at 08:27:49PM -0400, Tom Lane wrote:\\n>> + one more row than requested, so tha it can determine whether the\\n\\n> A little typo: tha it\\n> [ etc ]\\n\\nUgh. These notes were prepared in a bit more haste than usual,\\nand I guess it shows :-(. Will fix, thanks for proofreading!\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 06 Nov 2021 09:34:01 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Draft release notes for next week's releases\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 5, 2021 at 5:27 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> As usual, please send any comments/corrections by Sunday.\\n\\nYou probably won't miss this, but just in case: Alexander's very\\nrecent \\\"Reset lastOverflowedXid on standby when needed\\\" commit needs\\nto be listed now.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Sat, 6 Nov 2021 11:43:40 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Draft release notes for next week's releases\"\n },\n {\n \"msg_contents\": \"Peter Geoghegan <pg@bowt.ie> writes:\\n> You probably won't miss this, but just in case: Alexander's very\\n> recent \\\"Reset lastOverflowedXid on standby when needed\\\" commit needs\\n> to be listed now.\\n\\nYup, I saw it.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 06 Nov 2021 15:03:47 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Draft release notes for next week's releases\"\n }\n]","isConversational":false}}},{"rowIdx":53050,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"I recently pushed a commit which formalized what we expect of HOT\nchains during pruning -- see assertions and comments added by commit\n5cd7eb1f, \"Add various assertions to heap pruning code\". I notice that\nverify_heapam() doesn't really do anything with HOT chains, and I'd\nsay that that's a gap. Fortunately it should not be terribly difficult\nto fill in this gap.\n\nThe short version is this: I think that verify_heapam() should\nvalidate whole HOT chains, not just individual heap-only tuples. This\ncan be accomplished by taking a similar approach to the approach taken\nin heap_prune_chain() already, for pruning.\n\nFor a given heap page, pruning fills in its PruneState variable to\nbuild a coherent picture of the state of all HOT chains on the entire\npage, before modifying anything (barring hint bit setting). We only\nreach heap_page_prune_execute() when we have the whole picture about\nHOT chains. This high level approach is essential when pruning.\nConsider LP_REDIRECT items. We might need to set an existing\nLP_REDIRECT item to LP_DEAD, rendering an entire HOT chain dead.\nWhether or not that's the appropriate course of action for a given\nLP_REDIRECT item will depend on the state of the HOT chain as a whole.\n\nThe HOT chain could of course have several heap-only tuples, and only\none visible tuple means that the LP_REDIRECT root item must not be set\nLP_DEAD -- it should be made to link to the remaining live heap-only\ntuple instead (while the dead intermediate heap-only tuples become\nLP_UNUSED, never LP_DEAD). This is made even more tricky by the fact\nthat we have to consider HOT updaters with aborted transactions during\npruning.\n\nWe even have to consider the case where a HOT updater aborts, but\nanother HOT updater tries again and commits. Now we have a heap-only\ntuple that is \"disconnected from its original HOT chain\" (or will be\nuntil the next prune). Having a stray/disconnected heap-only tuple\nlike this ought to be fine -- the next pruning operation will get to\nit whenever (see the special handling for these aborted heap-only\ntuples at the top of heap_prune_chain()). But any disconnected\nheap-only tuple had better be from an aborted xact -- otherwise we\nhave corruption. This is corruption that looks like index corruption,\nbecause sequential scans still give correct answers (I've seen this\nseveral times before, ask me how). But it's not index corruption,\ncontrary to appearances -- it's heap corruption.\n\nHere are some specific checks I have in mind:\n\n* Right now, we throw an error when an LP_REDIRECT points to an\nLP_UNUSED item. But that existing test should go further, like the\nassertions I added recently. They should make sure that the\nLP_REDIRECT links directly to an LP_NORMAL item, which has tuple\nstorage. And, the stored tuple must be a heap-only tuple.\n\nNote, in particular, that it's not okay if the LP_REDIRECT points to\nan LP_DEAD item (which amcheck won't complain about today). That is at\nleast my position on it, and so far it looks like the right position\n(no assertion failures reported yet).\n\n* HOT chain validation: All visible/committed heap-only tuples should\nbe reachable through a \"recognizable root item for their HOT chain, or\nan earlier member of the HOT chain via its t_ctid link\". The root item\ncan either be an LP_REDIRECT, or a regular not-heap-only tuple -- but\nit can't be anything else.\n\n* All other heap-only tuples (all heap-only tuples that don't seem to\nbe part of a valid HOT chain) must be from aborted transactions -- as\nI said, this has to be corruption (heap corruption disguised as index\ncorruption, even).\n\nNote that we're very permissive about broken HOT chains -- in general\na heap-only tuple's t_ctid link can have pretty wild values, and\nthat's okay. For example, it can point past the current bounds of the\npage's line pointer array. Everything is still robust because of the\nway that we validate HOT chains during chain traversal -- we match\ntuple N's xmax to tuple N+1's xmin, all the time, inside code like\nheap_hot_search_buffer() that is routinely used during index scans.\nHowever, the root item of a HOT chain (usually an LP_REDIRECT)\nabsolutely must point to a valid heap-only tuple at all times --\nthat's very different indeed. That's the starting point for an index\nscan -- it is the TID that is stored in indexes that needs to be\nstable over time. And so pruning needs to leave behind coherent\nlooking whole entire HOT chains, just as we ought to verify that\npruning consistently gets it right.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Fri, 5 Nov 2021 19:51:12 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"amcheck's verify_heapam(), and HOT chain verification"},{"msg_contents":"\n\n> On Nov 5, 2021, at 7:51 PM, Peter Geoghegan <pg@bowt.ie> wrote:\n> \n> I recently pushed a commit which formalized what we expect of HOT\n> chains during pruning -- see assertions and comments added by commit\n> 5cd7eb1f, \"Add various assertions to heap pruning code\". I notice that\n> verify_heapam() doesn't really do anything with HOT chains, and I'd\n> say that that's a gap. Fortunately it should not be terribly difficult\n> to fill in this gap.\n> \n> The short version is this: I think that verify_heapam() should\n> validate whole HOT chains, not just individual heap-only tuples. This\n> can be accomplished by taking a similar approach to the approach taken\n> in heap_prune_chain() already, for pruning.\n> \n> For a given heap page, pruning fills in its PruneState variable to\n> build a coherent picture of the state of all HOT chains on the entire\n> page, before modifying anything (barring hint bit setting). We only\n> reach heap_page_prune_execute() when we have the whole picture about\n> HOT chains. This high level approach is essential when pruning.\n> Consider LP_REDIRECT items. We might need to set an existing\n> LP_REDIRECT item to LP_DEAD, rendering an entire HOT chain dead.\n> Whether or not that's the appropriate course of action for a given\n> LP_REDIRECT item will depend on the state of the HOT chain as a whole.\n> \n> The HOT chain could of course have several heap-only tuples, and only\n> one visible tuple means that the LP_REDIRECT root item must not be set\n> LP_DEAD -- it should be made to link to the remaining live heap-only\n> tuple instead (while the dead intermediate heap-only tuples become\n> LP_UNUSED, never LP_DEAD). This is made even more tricky by the fact\n> that we have to consider HOT updaters with aborted transactions during\n> pruning.\n> \n> We even have to consider the case where a HOT updater aborts, but\n> another HOT updater tries again and commits. Now we have a heap-only\n> tuple that is \"disconnected from its original HOT chain\" (or will be\n> until the next prune). Having a stray/disconnected heap-only tuple\n> like this ought to be fine -- the next pruning operation will get to\n> it whenever (see the special handling for these aborted heap-only\n> tuples at the top of heap_prune_chain()). But any disconnected\n> heap-only tuple had better be from an aborted xact -- otherwise we\n> have corruption. This is corruption that looks like index corruption,\n> because sequential scans still give correct answers (I've seen this\n> several times before, ask me how). But it's not index corruption,\n> contrary to appearances -- it's heap corruption.\n> \n> Here are some specific checks I have in mind:\n> \n> * Right now, we throw an error when an LP_REDIRECT points to an\n> LP_UNUSED item. But that existing test should go further, like the\n> assertions I added recently. They should make sure that the\n> LP_REDIRECT links directly to an LP_NORMAL item, which has tuple\n> storage. And, the stored tuple must be a heap-only tuple.\n> \n> Note, in particular, that it's not okay if the LP_REDIRECT points to\n> an LP_DEAD item (which amcheck won't complain about today). That is at\n> least my position on it, and so far it looks like the right position\n> (no assertion failures reported yet).\n> \n> * HOT chain validation: All visible/committed heap-only tuples should\n> be reachable through a \"recognizable root item for their HOT chain, or\n> an earlier member of the HOT chain via its t_ctid link\". The root item\n> can either be an LP_REDIRECT, or a regular not-heap-only tuple -- but\n> it can't be anything else.\n> \n> * All other heap-only tuples (all heap-only tuples that don't seem to\n> be part of a valid HOT chain) must be from aborted transactions -- as\n> I said, this has to be corruption (heap corruption disguised as index\n> corruption, even).\n> \n> Note that we're very permissive about broken HOT chains -- in general\n> a heap-only tuple's t_ctid link can have pretty wild values, and\n> that's okay. For example, it can point past the current bounds of the\n> page's line pointer array. Everything is still robust because of the\n> way that we validate HOT chains during chain traversal -- we match\n> tuple N's xmax to tuple N+1's xmin, all the time, inside code like\n> heap_hot_search_buffer() that is routinely used during index scans.\n> However, the root item of a HOT chain (usually an LP_REDIRECT)\n> absolutely must point to a valid heap-only tuple at all times --\n> that's very different indeed. That's the starting point for an index\n> scan -- it is the TID that is stored in indexes that needs to be\n> stable over time. And so pruning needs to leave behind coherent\n> looking whole entire HOT chains, just as we ought to verify that\n> pruning consistently gets it right.\n\nThanks, Peter, for this analysis. It's getting a bit late in the day, but I'll try this out tomorrow, I expect.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Fri, 5 Nov 2021 20:18:30 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: amcheck's verify_heapam(), and HOT chain verification"},{"msg_contents":"On Fri, Nov 5, 2021 at 8:18 PM Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> Thanks, Peter, for this analysis.\n\nIt's strategically important work. What you've done so far has every\nchance of catching corruption involving storage issues, which is\ngreat. But we ought to do more to catch certain known general patterns\nof corruption. The so-called \"freeze the dead\" bug (see commit\n9c2f0a6c for the final fix) is the single weirdest and scariest bug\nthat I can recall (there were a couple of incorrect bug fixes that had\nto be reverted), and it is very much the kind of thing that I'd like\nto see verify_heapam.c handle exhaustively. That would de-risk a lot\nof things. Note that the same variety of HOT chain corruption bugs\nbesides that one, so it's really a general class of corruption, not\none specific bug.\n\nThe heapallindexed verification option actually detected the \"freeze\nthe dead\" bug [1], although that was a total accident -- that checking\noption happens to use CREATE INDEX infrastructure, which happens to\nsanitize HOT chains by calling heap_get_root_tuples() and detecting\ncontradictions -- heap_get_root_tuples() is code that is right next to\nthe pruning code I mentioned. That level of detection is certainly\nbetter than nothing, but it's still not good enough -- we need this in\nverify_heapam, too.\n\nHere's why I think that we need very thorough HOT chain verification\nthat lives directly in verify_heapam:\n\nFirst of all, the heapallindexed option is rather expensive, whereas\nthe additional overhead for verify_heapam would be minimal -- you\nprobably wouldn't even need to make it optional. Second of all, why\nshould you need to use bt_index_check() to detect what is after all\nheap corruption? And third of all, the approach of detecting\ncorruption by outsourcing detection to heapam_index_build_range_scan()\n(which calls heap_get_root_tuples() itself) probably risks missing\ncorrupt HOT chains where the corruption doesn't look a certain way --\nit was not designed with amcheck in mind.\n\nheapam_index_build_range_scan() + heap_get_root_tuples() will notice a\nheap-only tuple without a parent, which is a good start. But what\nabout an LP_REDIRECT item whose lp_off (i.e. page offset number\nredirect) somehow points to any item on the page other than a valid\nheap-only tuple? Also doesn't catch contradictory commit states for\nheap-only tuples that form a hot chain through their t_ctid links. I'm\nsure that there are other things that I haven't thought of, too --\nthere is a lot going on here.\n\nThis HOT chain verification business is a little like the checking\nthat we do in verify_nbtree.c, actually. The individual tuples (in\nthis case heap tuples) may well not be corrupt (or demonstrably\ncorrupt) in isolation. But taken together, in a particular page offset\nnumber sequence order, they can nevertheless *imply* corruption --\nit's implied if the tuples tell a contradictory story about what's\ngoing on at the level of the whole page. A HOT chain LP_REDIRECT is\nafter all a little like a \"mini index\" stored on a heap page.\nSequential scans don't care about LP_REDIRECTs at all. But index scans\nexpect a great deal from LP_REDIRECTs.\n\n[1] https://www.postgresql.org/message-id/CAH2-Wznm4rCrhFAiwKPWTpEw2bXDtgROZK7jWWGucXeH3D1fmA@mail.gmail.com\n--\nPeter Geoghegan\n\n\n","msg_date":"Sat, 6 Nov 2021 13:03:17 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"Re: amcheck's verify_heapam(), and HOT chain verification"},{"msg_contents":"On Fri, Nov 5, 2021 at 7:51 PM Peter Geoghegan <pg@bowt.ie> wrote:\n> Here are some specific checks I have in mind:\n\nOne more for the list:\n\n* Validate PageIsAllVisible() for each page.\n\nIn other words, pg_visibility should be merged with verify_heapam.c\n(or at least pg_visibility 's pg_check_frozen() and pg_check_visible()\nfunctions should be moved, merged, or whatever). This would mean that\nverify_heapam() would directly check if the page-level PD_ALL_VISIBLE\nflag contradicts either the tuple headers of tuples with storage on\nthe page, the presence (or absence) of LP_DEAD stub line pointers on\nthe page, or the corresponding visibility map bit (e.g.,\nVISIBILITYMAP_ALL_VISIBLE) for the page.\n\nThere is value in teaching verify_heapam() about any possible problem,\nincluding with the visibility map, but it's certainly less valuable\nthan the HOT chain verification stuff -- and probably trickier to get\nright. I'm mentioning it now to be exhaustive, but it's less of a\npriority for me personally.\n\nI am quite willing to help out with all this, if you're interested.\n\nOne more thing about HOT chain validation:\n\nI can give you another example bug of the kind I'd expect\nverify_heapam() to catch only with full HOT chain validation. This one\nis a vintage MultiXact bug that has the same basic HOT chain\ncorruption, looks-like-index-corruption-but-isn't quality as the more\nmemorable freeze-the-dead bug (this one was fixed by commit 6bfa88ac):\n\nhttps://www.postgresql.org/message-id/CAM3SWZTMQiCi5PV5OWHb%2BbYkUcnCk%3DO67w0cSswPvV7XfUcU5g%40mail.gmail.com\n\nIn general I think that reviewing historic examples of pernicious\ncorruption bugs is a valuable exercise when designing tools like\namcheck. Maybe even revert the fix during testing, to be sure it would\nhave been caught had the final tool been available. History doesn't\nrepeat itself, but it does rhyme.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Sat, 6 Nov 2021 15:09:15 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"Re: amcheck's verify_heapam(), and HOT chain verification"},{"msg_contents":"\n\n> On Nov 6, 2021, at 3:09 PM, Peter Geoghegan <pg@bowt.ie> wrote:\n> \n> I am quite willing to help out with all this, if you're interested.\n\nYes, I am quite interested, though I will have to alternate between this work and the various patch sets that I've already submitted for this development cycle.\n\nI think we need a corruption generating framework that can be deployed on the buildfarm. What I have in mind is inspired by your comments about the \"freeze the dead\" bug. We need to be able to build versions of the database with known bugs enabled, both real-world bugs from the past and contrived bugs we write only for this purpose, so as to create non-trivial corruption on the build farm. I think it would be sufficient if such builds were run less frequently, perhaps triggered by commits to a corruption testing branch? We could keep the old bugs alive on that branch while periodically merging in everything else from HEAD? That seems a tad tiresome, but maybe it wouldn't be too bad if the intentional bugs were limited to just a few backend files, and as much as possible in code at the end of the file or in separate files to reduce merge conflicts?\n\nI'm cc'ing Andrew to get his thoughts about the buildfarm integration....\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Sun, 7 Nov 2021 09:30:23 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: amcheck's verify_heapam(), and HOT chain verification"},{"msg_contents":"On Sun, Nov 7, 2021 at 9:30 AM Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> Yes, I am quite interested, though I will have to alternate between this work and the various patch sets that I've already submitted for this development cycle.\n\nGreat! We still have a lot of work to do with HOT chain level\ninvariants. I have seen very clear evidence of that in the last 24\nhours:\n\nhttps://postgr.es/m/CAH2-Wzma=Y3O+LRx2Wj_HwGbbbeNwr6FoJzXni8hxOMw55pcZg@mail.gmail.com\n\n> I think we need a corruption generating framework that can be deployed on the buildfarm. What I have in mind is inspired by your comments about the \"freeze the dead\" bug.\n\nI'm not sure that that's truly necessary. IMV the important thing is\nthat we formalize the invariants, and have tooling that can test them.\nMaintaining tests that actually display specific broken behavior (as\nopposed to its absence) seems like it might be quite a burden.\n\nThere are many specific ways that these invariants might break, but\nthe specifics shouldn't matter -- the invariants should cut through\nthat (to the extent that that's possible). The \"freeze the dead\" bug\nis mostly useful as a way of framing the discussion (perhaps even in\ncode comments). I did this with a historic CREATE INDEX CONCURRENTLY\nbug, in code comments for heapallindexed verification. Verification\nusing heapallindexed actually detected two more CIC bugs years later\n-- not including the earlier fixes that didn't quite get everything\nright (thinking of the prepared xact CIC bugs found using amcheck).\nThe invariants that heapallindexed tests generalize to many different\nsituations, including situations that I couldn't have possibly\nanticipated with any kind of precision. Having the right high level\nidea is what really mattered.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Tue, 9 Nov 2021 10:12:01 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"Re: amcheck's verify_heapam(), and HOT chain verification"}],"string":"[\n {\n \"msg_contents\": \"I recently pushed a commit which formalized what we expect of HOT\\nchains during pruning -- see assertions and comments added by commit\\n5cd7eb1f, \\\"Add various assertions to heap pruning code\\\". I notice that\\nverify_heapam() doesn't really do anything with HOT chains, and I'd\\nsay that that's a gap. Fortunately it should not be terribly difficult\\nto fill in this gap.\\n\\nThe short version is this: I think that verify_heapam() should\\nvalidate whole HOT chains, not just individual heap-only tuples. This\\ncan be accomplished by taking a similar approach to the approach taken\\nin heap_prune_chain() already, for pruning.\\n\\nFor a given heap page, pruning fills in its PruneState variable to\\nbuild a coherent picture of the state of all HOT chains on the entire\\npage, before modifying anything (barring hint bit setting). We only\\nreach heap_page_prune_execute() when we have the whole picture about\\nHOT chains. This high level approach is essential when pruning.\\nConsider LP_REDIRECT items. We might need to set an existing\\nLP_REDIRECT item to LP_DEAD, rendering an entire HOT chain dead.\\nWhether or not that's the appropriate course of action for a given\\nLP_REDIRECT item will depend on the state of the HOT chain as a whole.\\n\\nThe HOT chain could of course have several heap-only tuples, and only\\none visible tuple means that the LP_REDIRECT root item must not be set\\nLP_DEAD -- it should be made to link to the remaining live heap-only\\ntuple instead (while the dead intermediate heap-only tuples become\\nLP_UNUSED, never LP_DEAD). This is made even more tricky by the fact\\nthat we have to consider HOT updaters with aborted transactions during\\npruning.\\n\\nWe even have to consider the case where a HOT updater aborts, but\\nanother HOT updater tries again and commits. Now we have a heap-only\\ntuple that is \\\"disconnected from its original HOT chain\\\" (or will be\\nuntil the next prune). Having a stray/disconnected heap-only tuple\\nlike this ought to be fine -- the next pruning operation will get to\\nit whenever (see the special handling for these aborted heap-only\\ntuples at the top of heap_prune_chain()). But any disconnected\\nheap-only tuple had better be from an aborted xact -- otherwise we\\nhave corruption. This is corruption that looks like index corruption,\\nbecause sequential scans still give correct answers (I've seen this\\nseveral times before, ask me how). But it's not index corruption,\\ncontrary to appearances -- it's heap corruption.\\n\\nHere are some specific checks I have in mind:\\n\\n* Right now, we throw an error when an LP_REDIRECT points to an\\nLP_UNUSED item. But that existing test should go further, like the\\nassertions I added recently. They should make sure that the\\nLP_REDIRECT links directly to an LP_NORMAL item, which has tuple\\nstorage. And, the stored tuple must be a heap-only tuple.\\n\\nNote, in particular, that it's not okay if the LP_REDIRECT points to\\nan LP_DEAD item (which amcheck won't complain about today). That is at\\nleast my position on it, and so far it looks like the right position\\n(no assertion failures reported yet).\\n\\n* HOT chain validation: All visible/committed heap-only tuples should\\nbe reachable through a \\\"recognizable root item for their HOT chain, or\\nan earlier member of the HOT chain via its t_ctid link\\\". The root item\\ncan either be an LP_REDIRECT, or a regular not-heap-only tuple -- but\\nit can't be anything else.\\n\\n* All other heap-only tuples (all heap-only tuples that don't seem to\\nbe part of a valid HOT chain) must be from aborted transactions -- as\\nI said, this has to be corruption (heap corruption disguised as index\\ncorruption, even).\\n\\nNote that we're very permissive about broken HOT chains -- in general\\na heap-only tuple's t_ctid link can have pretty wild values, and\\nthat's okay. For example, it can point past the current bounds of the\\npage's line pointer array. Everything is still robust because of the\\nway that we validate HOT chains during chain traversal -- we match\\ntuple N's xmax to tuple N+1's xmin, all the time, inside code like\\nheap_hot_search_buffer() that is routinely used during index scans.\\nHowever, the root item of a HOT chain (usually an LP_REDIRECT)\\nabsolutely must point to a valid heap-only tuple at all times --\\nthat's very different indeed. That's the starting point for an index\\nscan -- it is the TID that is stored in indexes that needs to be\\nstable over time. And so pruning needs to leave behind coherent\\nlooking whole entire HOT chains, just as we ought to verify that\\npruning consistently gets it right.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Nov 2021 19:51:12 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"amcheck's verify_heapam(), and HOT chain verification\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 5, 2021, at 7:51 PM, Peter Geoghegan <pg@bowt.ie> wrote:\\n> \\n> I recently pushed a commit which formalized what we expect of HOT\\n> chains during pruning -- see assertions and comments added by commit\\n> 5cd7eb1f, \\\"Add various assertions to heap pruning code\\\". I notice that\\n> verify_heapam() doesn't really do anything with HOT chains, and I'd\\n> say that that's a gap. Fortunately it should not be terribly difficult\\n> to fill in this gap.\\n> \\n> The short version is this: I think that verify_heapam() should\\n> validate whole HOT chains, not just individual heap-only tuples. This\\n> can be accomplished by taking a similar approach to the approach taken\\n> in heap_prune_chain() already, for pruning.\\n> \\n> For a given heap page, pruning fills in its PruneState variable to\\n> build a coherent picture of the state of all HOT chains on the entire\\n> page, before modifying anything (barring hint bit setting). We only\\n> reach heap_page_prune_execute() when we have the whole picture about\\n> HOT chains. This high level approach is essential when pruning.\\n> Consider LP_REDIRECT items. We might need to set an existing\\n> LP_REDIRECT item to LP_DEAD, rendering an entire HOT chain dead.\\n> Whether or not that's the appropriate course of action for a given\\n> LP_REDIRECT item will depend on the state of the HOT chain as a whole.\\n> \\n> The HOT chain could of course have several heap-only tuples, and only\\n> one visible tuple means that the LP_REDIRECT root item must not be set\\n> LP_DEAD -- it should be made to link to the remaining live heap-only\\n> tuple instead (while the dead intermediate heap-only tuples become\\n> LP_UNUSED, never LP_DEAD). This is made even more tricky by the fact\\n> that we have to consider HOT updaters with aborted transactions during\\n> pruning.\\n> \\n> We even have to consider the case where a HOT updater aborts, but\\n> another HOT updater tries again and commits. Now we have a heap-only\\n> tuple that is \\\"disconnected from its original HOT chain\\\" (or will be\\n> until the next prune). Having a stray/disconnected heap-only tuple\\n> like this ought to be fine -- the next pruning operation will get to\\n> it whenever (see the special handling for these aborted heap-only\\n> tuples at the top of heap_prune_chain()). But any disconnected\\n> heap-only tuple had better be from an aborted xact -- otherwise we\\n> have corruption. This is corruption that looks like index corruption,\\n> because sequential scans still give correct answers (I've seen this\\n> several times before, ask me how). But it's not index corruption,\\n> contrary to appearances -- it's heap corruption.\\n> \\n> Here are some specific checks I have in mind:\\n> \\n> * Right now, we throw an error when an LP_REDIRECT points to an\\n> LP_UNUSED item. But that existing test should go further, like the\\n> assertions I added recently. They should make sure that the\\n> LP_REDIRECT links directly to an LP_NORMAL item, which has tuple\\n> storage. And, the stored tuple must be a heap-only tuple.\\n> \\n> Note, in particular, that it's not okay if the LP_REDIRECT points to\\n> an LP_DEAD item (which amcheck won't complain about today). That is at\\n> least my position on it, and so far it looks like the right position\\n> (no assertion failures reported yet).\\n> \\n> * HOT chain validation: All visible/committed heap-only tuples should\\n> be reachable through a \\\"recognizable root item for their HOT chain, or\\n> an earlier member of the HOT chain via its t_ctid link\\\". The root item\\n> can either be an LP_REDIRECT, or a regular not-heap-only tuple -- but\\n> it can't be anything else.\\n> \\n> * All other heap-only tuples (all heap-only tuples that don't seem to\\n> be part of a valid HOT chain) must be from aborted transactions -- as\\n> I said, this has to be corruption (heap corruption disguised as index\\n> corruption, even).\\n> \\n> Note that we're very permissive about broken HOT chains -- in general\\n> a heap-only tuple's t_ctid link can have pretty wild values, and\\n> that's okay. For example, it can point past the current bounds of the\\n> page's line pointer array. Everything is still robust because of the\\n> way that we validate HOT chains during chain traversal -- we match\\n> tuple N's xmax to tuple N+1's xmin, all the time, inside code like\\n> heap_hot_search_buffer() that is routinely used during index scans.\\n> However, the root item of a HOT chain (usually an LP_REDIRECT)\\n> absolutely must point to a valid heap-only tuple at all times --\\n> that's very different indeed. That's the starting point for an index\\n> scan -- it is the TID that is stored in indexes that needs to be\\n> stable over time. And so pruning needs to leave behind coherent\\n> looking whole entire HOT chains, just as we ought to verify that\\n> pruning consistently gets it right.\\n\\nThanks, Peter, for this analysis. It's getting a bit late in the day, but I'll try this out tomorrow, I expect.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Nov 2021 20:18:30 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: amcheck's verify_heapam(), and HOT chain verification\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 5, 2021 at 8:18 PM Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> Thanks, Peter, for this analysis.\\n\\nIt's strategically important work. What you've done so far has every\\nchance of catching corruption involving storage issues, which is\\ngreat. But we ought to do more to catch certain known general patterns\\nof corruption. The so-called \\\"freeze the dead\\\" bug (see commit\\n9c2f0a6c for the final fix) is the single weirdest and scariest bug\\nthat I can recall (there were a couple of incorrect bug fixes that had\\nto be reverted), and it is very much the kind of thing that I'd like\\nto see verify_heapam.c handle exhaustively. That would de-risk a lot\\nof things. Note that the same variety of HOT chain corruption bugs\\nbesides that one, so it's really a general class of corruption, not\\none specific bug.\\n\\nThe heapallindexed verification option actually detected the \\\"freeze\\nthe dead\\\" bug [1], although that was a total accident -- that checking\\noption happens to use CREATE INDEX infrastructure, which happens to\\nsanitize HOT chains by calling heap_get_root_tuples() and detecting\\ncontradictions -- heap_get_root_tuples() is code that is right next to\\nthe pruning code I mentioned. That level of detection is certainly\\nbetter than nothing, but it's still not good enough -- we need this in\\nverify_heapam, too.\\n\\nHere's why I think that we need very thorough HOT chain verification\\nthat lives directly in verify_heapam:\\n\\nFirst of all, the heapallindexed option is rather expensive, whereas\\nthe additional overhead for verify_heapam would be minimal -- you\\nprobably wouldn't even need to make it optional. Second of all, why\\nshould you need to use bt_index_check() to detect what is after all\\nheap corruption? And third of all, the approach of detecting\\ncorruption by outsourcing detection to heapam_index_build_range_scan()\\n(which calls heap_get_root_tuples() itself) probably risks missing\\ncorrupt HOT chains where the corruption doesn't look a certain way --\\nit was not designed with amcheck in mind.\\n\\nheapam_index_build_range_scan() + heap_get_root_tuples() will notice a\\nheap-only tuple without a parent, which is a good start. But what\\nabout an LP_REDIRECT item whose lp_off (i.e. page offset number\\nredirect) somehow points to any item on the page other than a valid\\nheap-only tuple? Also doesn't catch contradictory commit states for\\nheap-only tuples that form a hot chain through their t_ctid links. I'm\\nsure that there are other things that I haven't thought of, too --\\nthere is a lot going on here.\\n\\nThis HOT chain verification business is a little like the checking\\nthat we do in verify_nbtree.c, actually. The individual tuples (in\\nthis case heap tuples) may well not be corrupt (or demonstrably\\ncorrupt) in isolation. But taken together, in a particular page offset\\nnumber sequence order, they can nevertheless *imply* corruption --\\nit's implied if the tuples tell a contradictory story about what's\\ngoing on at the level of the whole page. A HOT chain LP_REDIRECT is\\nafter all a little like a \\\"mini index\\\" stored on a heap page.\\nSequential scans don't care about LP_REDIRECTs at all. But index scans\\nexpect a great deal from LP_REDIRECTs.\\n\\n[1] https://www.postgresql.org/message-id/CAH2-Wznm4rCrhFAiwKPWTpEw2bXDtgROZK7jWWGucXeH3D1fmA@mail.gmail.com\\n--\\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Sat, 6 Nov 2021 13:03:17 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: amcheck's verify_heapam(), and HOT chain verification\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 5, 2021 at 7:51 PM Peter Geoghegan <pg@bowt.ie> wrote:\\n> Here are some specific checks I have in mind:\\n\\nOne more for the list:\\n\\n* Validate PageIsAllVisible() for each page.\\n\\nIn other words, pg_visibility should be merged with verify_heapam.c\\n(or at least pg_visibility 's pg_check_frozen() and pg_check_visible()\\nfunctions should be moved, merged, or whatever). This would mean that\\nverify_heapam() would directly check if the page-level PD_ALL_VISIBLE\\nflag contradicts either the tuple headers of tuples with storage on\\nthe page, the presence (or absence) of LP_DEAD stub line pointers on\\nthe page, or the corresponding visibility map bit (e.g.,\\nVISIBILITYMAP_ALL_VISIBLE) for the page.\\n\\nThere is value in teaching verify_heapam() about any possible problem,\\nincluding with the visibility map, but it's certainly less valuable\\nthan the HOT chain verification stuff -- and probably trickier to get\\nright. I'm mentioning it now to be exhaustive, but it's less of a\\npriority for me personally.\\n\\nI am quite willing to help out with all this, if you're interested.\\n\\nOne more thing about HOT chain validation:\\n\\nI can give you another example bug of the kind I'd expect\\nverify_heapam() to catch only with full HOT chain validation. This one\\nis a vintage MultiXact bug that has the same basic HOT chain\\ncorruption, looks-like-index-corruption-but-isn't quality as the more\\nmemorable freeze-the-dead bug (this one was fixed by commit 6bfa88ac):\\n\\nhttps://www.postgresql.org/message-id/CAM3SWZTMQiCi5PV5OWHb%2BbYkUcnCk%3DO67w0cSswPvV7XfUcU5g%40mail.gmail.com\\n\\nIn general I think that reviewing historic examples of pernicious\\ncorruption bugs is a valuable exercise when designing tools like\\namcheck. Maybe even revert the fix during testing, to be sure it would\\nhave been caught had the final tool been available. History doesn't\\nrepeat itself, but it does rhyme.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Sat, 6 Nov 2021 15:09:15 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: amcheck's verify_heapam(), and HOT chain verification\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 6, 2021, at 3:09 PM, Peter Geoghegan <pg@bowt.ie> wrote:\\n> \\n> I am quite willing to help out with all this, if you're interested.\\n\\nYes, I am quite interested, though I will have to alternate between this work and the various patch sets that I've already submitted for this development cycle.\\n\\nI think we need a corruption generating framework that can be deployed on the buildfarm. What I have in mind is inspired by your comments about the \\\"freeze the dead\\\" bug. We need to be able to build versions of the database with known bugs enabled, both real-world bugs from the past and contrived bugs we write only for this purpose, so as to create non-trivial corruption on the build farm. I think it would be sufficient if such builds were run less frequently, perhaps triggered by commits to a corruption testing branch? We could keep the old bugs alive on that branch while periodically merging in everything else from HEAD? That seems a tad tiresome, but maybe it wouldn't be too bad if the intentional bugs were limited to just a few backend files, and as much as possible in code at the end of the file or in separate files to reduce merge conflicts?\\n\\nI'm cc'ing Andrew to get his thoughts about the buildfarm integration....\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Sun, 7 Nov 2021 09:30:23 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: amcheck's verify_heapam(), and HOT chain verification\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 7, 2021 at 9:30 AM Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> Yes, I am quite interested, though I will have to alternate between this work and the various patch sets that I've already submitted for this development cycle.\\n\\nGreat! We still have a lot of work to do with HOT chain level\\ninvariants. I have seen very clear evidence of that in the last 24\\nhours:\\n\\nhttps://postgr.es/m/CAH2-Wzma=Y3O+LRx2Wj_HwGbbbeNwr6FoJzXni8hxOMw55pcZg@mail.gmail.com\\n\\n> I think we need a corruption generating framework that can be deployed on the buildfarm. What I have in mind is inspired by your comments about the \\\"freeze the dead\\\" bug.\\n\\nI'm not sure that that's truly necessary. IMV the important thing is\\nthat we formalize the invariants, and have tooling that can test them.\\nMaintaining tests that actually display specific broken behavior (as\\nopposed to its absence) seems like it might be quite a burden.\\n\\nThere are many specific ways that these invariants might break, but\\nthe specifics shouldn't matter -- the invariants should cut through\\nthat (to the extent that that's possible). The \\\"freeze the dead\\\" bug\\nis mostly useful as a way of framing the discussion (perhaps even in\\ncode comments). I did this with a historic CREATE INDEX CONCURRENTLY\\nbug, in code comments for heapallindexed verification. Verification\\nusing heapallindexed actually detected two more CIC bugs years later\\n-- not including the earlier fixes that didn't quite get everything\\nright (thinking of the prepared xact CIC bugs found using amcheck).\\nThe invariants that heapallindexed tests generalize to many different\\nsituations, including situations that I couldn't have possibly\\nanticipated with any kind of precision. Having the right high level\\nidea is what really mattered.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 10:12:01 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: amcheck's verify_heapam(), and HOT chain verification\"\n }\n]","isConversational":false}}},{"rowIdx":53051,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"On Sun, Oct 24, 2021 at 04:35:02PM -0700, Noah Misch wrote:\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kittiwake&dt=2021-10-24%2012%3A01%3A10\n> got an interesting v9.6 failure [...]:\n> \n> 2021-10-24 14:25:29.263 CEST [34569:175] pgbench ERROR: could not read two-phase state from xlog at 0/158F4E0\n> 2021-10-24 14:25:29.263 CEST [34569:176] pgbench STATEMENT: COMMIT PREPARED 'c1';\n\nAs a first step, let's report the actual XLogReadRecord() error message.\nAttached. All the other sites that expect no error already do this.","msg_date":"Sat, 6 Nov 2021 18:31:57 -0700","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Sat, Nov 06, 2021 at 06:31:57PM -0700, Noah Misch wrote:\n> As a first step, let's report the actual XLogReadRecord() error message.\n> Attached.\n\nGood catch! This looks good.\n\n> All the other sites that expect no error already do this.\n\nIndeed. Looking closer, I think that we'd better improve\nDecodingContextFindStartpoint(),\npg_logical_replication_slot_advance(), XLogSendLogical() as well as\npg_logical_slot_get_changes_guts() to follow a format closer to what\nyou have in your patch, with an error message that describes the\ncontext where the problem has been found, instead of just elog()'ing\nwhat XLogReadRecord() returns.\n--\nMichael","msg_date":"Mon, 8 Nov 2021 13:42:46 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"\n\n> 7 нояб. 2021 г., в 06:31, Noah Misch <noah@leadboat.com> написал(а):\n> \n> As a first step, let's report the actual XLogReadRecord() error message.\n> Attached. All the other sites that expect no error already do this.\n\nBTW some time ago I've spotted a good number of related unreported errors [0].\n\n\n[0] https://www.postgresql.org/message-id/A0A36AEE-3476-4326-B877-EE2B55BAEEED%40yandex-team.ru\n\n","msg_date":"Mon, 8 Nov 2021 16:38:47 +0500","msg_from":"Andrey Borodin <x4mmm@yandex-team.ru>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Mon, Nov 08, 2021 at 01:42:46PM +0900, Michael Paquier wrote:\n> Indeed. Looking closer, I think that we'd better improve\n> DecodingContextFindStartpoint(),\n> pg_logical_replication_slot_advance(), XLogSendLogical() as well as\n> pg_logical_slot_get_changes_guts() to follow a format closer to what\n> you have in your patch, with an error message that describes the\n> context where the problem has been found, instead of just elog()'ing\n> what XLogReadRecord() returns.\n\nFYI, I have just begun a new thread about those ones:\nhttps://www.postgresql.org/message-id/YYnTH6OyOwQcAdkw@paquier.xyz\n--\nMichael","msg_date":"Tue, 9 Nov 2021 10:51:40 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Mon, Nov 08, 2021 at 01:42:46PM +0900, Michael Paquier wrote:\n> On Sat, Nov 06, 2021 at 06:31:57PM -0700, Noah Misch wrote:\n> > On Sun, Oct 24, 2021 at 04:35:02PM -0700, Noah Misch wrote:\n> > > https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kittiwake&dt=2021-10-24%2012%3A01%3A10\n> > > got an interesting v9.6 failure [...]:\n> > > \n> > > 2021-10-24 14:25:29.263 CEST [34569:175] pgbench ERROR: could not read two-phase state from xlog at 0/158F4E0\n> > > 2021-10-24 14:25:29.263 CEST [34569:176] pgbench STATEMENT: COMMIT PREPARED 'c1';\n\nTom Lane reported another instance today:\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-11%2013%3A29%3A58\n\nEach of the three failures happened on a sparc64 Debian+gcc machine. I had\ntried ~8000 iterations on thorntail, another sparc64 Debian+gcc animal,\nwithout reproducing this.\n\n> > As a first step, let's report the actual XLogReadRecord() error message.\n> > Attached.\n> \n> Good catch! This looks good.\n\nPushed.\n\n\n","msg_date":"Thu, 11 Nov 2021 17:22:34 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"Noah Misch <noah@leadboat.com> writes:\n> Tom Lane reported another instance today:\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-11%2013%3A29%3A58\n\n> Each of the three failures happened on a sparc64 Debian+gcc machine. I had\n> tried ~8000 iterations on thorntail, another sparc64 Debian+gcc animal,\n> without reproducing this.\n\n>>> As a first step, let's report the actual XLogReadRecord() error message.\n>>> Attached.\n\n>> Good catch! This looks good.\n\n> Pushed.\n\nWell, we didn't have to wait too long [1]:\n\n# at t/003_cic_2pc.pl line 143.\n# 'pgbench: error: client 0 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/159EF88: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 5890048\n# pgbench: error: client 2 script 3 aborted in command 2 query 0: ERROR: canceling statement due to lock timeout\n# pgbench: fatal: Run was aborted; the above results are incomplete.\n\nI suppose \"unexpected pageaddr 0/0\" is most easily explained by supposing\nthat XlogReadTwoPhaseData tried to read a WAL page that hadn't been\nwritten out yet. Have we got any synchronization around that?\n\n\t\t\tregards, tom lane\n\n[1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-17%2013%3A01%3A24\n\n\n","msg_date":"Wed, 17 Nov 2021 17:47:10 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Wed, Nov 17, 2021 at 05:47:10PM -0500, Tom Lane wrote:\n> Noah Misch <noah@leadboat.com> writes:\n> > Each of the three failures happened on a sparc64 Debian+gcc machine. I had\n> > tried ~8000 iterations on thorntail, another sparc64 Debian+gcc animal,\n> > without reproducing this.\n\n> # 'pgbench: error: client 0 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/159EF88: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 5890048\n> [1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-17%2013%3A01%3A24\n\nTwo others:\nERROR: could not read two-phase state from WAL at 0/16F1850: invalid record length at 0/16F1850: wanted 24, got 0\n-- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-12%2013%3A01%3A15\nERROR: could not read two-phase state from WAL at 0/1668020: incorrect resource manager data checksum in record at 0/1668020\n-- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kittiwake&dt=2021-11-16%2015%3A00%3A52\n\n> I suppose \"unexpected pageaddr 0/0\" is most easily explained by supposing\n> that XlogReadTwoPhaseData tried to read a WAL page that hadn't been\n> written out yet. Have we got any synchronization around that?\n\nIf the WAL address isn't on disk, that error doesn't happen. Instead,\nread_local_xlog_page() blocks waiting for the WAL to become available. It's\nstill possible that we make the WAL region exist, but it somehow doesn't\ncontain the right data until shortly later. FinishPreparedTransaction() takes\nTwoPhaseStateLock and looks for an entry having gxact->valid. EndPrepare()\nfills gxact->prepare_end_lsn, then calls MarkAsPrepared() to set gxact->valid\nunder TwoPhaseStateLock. All freelist (freeGXacts) interaction holds\nTwoPhaseStateLock. I'm not seeing a gap in that synchronization.\n\nI don't have a great theory, but here are candidates to examine next:\n\n- Run with wal_debug=on to confirm logged write location matches read location.\n- Run \"PGDATA=contrib/amcheck/tmp_check/t_003_cic_2pc_CIC_2PC_test_data/pgdata\n pg_waldump -s 0/01000000\" at the end of the test.\n- Dump WAL page binary image at the point of failure.\n- Log which branches in XLogReadRecord() are taken.\n\nWhat else might help?\n\n\n","msg_date":"Wed, 17 Nov 2021 23:05:06 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"\n\n> 18 нояб. 2021 г., в 12:05, Noah Misch <noah@leadboat.com> написал(а):\n> \n> What else might help?\n\nLet's add more tests that check survival of 2PC through crash recovery? We do now only one restart. Maybe it worth to do 4 or 8?\n\nBest regards, Andrey Borodin.\n\n","msg_date":"Thu, 18 Nov 2021 17:48:02 +0500","msg_from":"Andrey Borodin <x4mmm@yandex-team.ru>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"Andrey Borodin <x4mmm@yandex-team.ru> writes:\n> Let's add more tests that check survival of 2PC through crash recovery? We do now only one restart. Maybe it worth to do 4 or 8?\n\nThat seems a little premature when we can't explain the failure\nwe have. Also, buildfarm cycles aren't free.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 18 Nov 2021 09:44:04 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Wed, Nov 17, 2021 at 11:05:06PM -0800, Noah Misch wrote:\n> On Wed, Nov 17, 2021 at 05:47:10PM -0500, Tom Lane wrote:\n> > Noah Misch <noah@leadboat.com> writes:\n> > > Each of the three failures happened on a sparc64 Debian+gcc machine. I had\n> > > tried ~8000 iterations on thorntail, another sparc64 Debian+gcc animal,\n> > > without reproducing this.\n> \n> > # 'pgbench: error: client 0 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/159EF88: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 5890048\n> > [1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-17%2013%3A01%3A24\n> \n> Two others:\n> ERROR: could not read two-phase state from WAL at 0/16F1850: invalid record length at 0/16F1850: wanted 24, got 0\n> -- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-12%2013%3A01%3A15\n> ERROR: could not read two-phase state from WAL at 0/1668020: incorrect resource manager data checksum in record at 0/1668020\n> -- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kittiwake&dt=2021-11-16%2015%3A00%3A52\n> \n> > I suppose \"unexpected pageaddr 0/0\" is most easily explained by supposing\n> > that XlogReadTwoPhaseData tried to read a WAL page that hadn't been\n> > written out yet. Have we got any synchronization around that?\n> \n> If the WAL address isn't on disk, that error doesn't happen. Instead,\n> read_local_xlog_page() blocks waiting for the WAL to become available. It's\n> still possible that we make the WAL region exist, but it somehow doesn't\n> contain the right data until shortly later. FinishPreparedTransaction() takes\n> TwoPhaseStateLock and looks for an entry having gxact->valid. EndPrepare()\n> fills gxact->prepare_end_lsn, then calls MarkAsPrepared() to set gxact->valid\n> under TwoPhaseStateLock. All freelist (freeGXacts) interaction holds\n> TwoPhaseStateLock. I'm not seeing a gap in that synchronization.\n> \n> I don't have a great theory, but here are candidates to examine next:\n> \n> - Run with wal_debug=on to confirm logged write location matches read location.\n> - Run \"PGDATA=contrib/amcheck/tmp_check/t_003_cic_2pc_CIC_2PC_test_data/pgdata\n> pg_waldump -s 0/01000000\" at the end of the test.\n> - Dump WAL page binary image at the point of failure.\n> - Log which branches in XLogReadRecord() are taken.\n\nTom Turelinckx, are you able to provide remote access to kittiwake or\ntadarida? I'd use it to attempt the above things. All else being equal,\nkittiwake is more relevant since it's still supported upstream.\n\nThe setup of your buildfarm animals provides a clue. I understand kittiwake\nis the same as ibisbill except for build options, and tadarida is the same as\nmussurana except for build options. ibisbill and mussurana haven't failed, so\none of these is likely needed to reproduce:\n\n absence of --enable-cassert\n CFLAGS='-g -O2 -fstack-protector -Wformat -Werror=format-security '\n CPPFLAGS='-Wdate-time -D_FORTIFY_SOURCE=2'\n LDFLAGS='-Wl,-z,relro -Wl,-z,now'\n\nHowever, I can't reproduce this on thorntail, even if I use tadarida's\n./configure options, CFLAGS, CPPFLAGS, and LDFLAGS.\n\n\n","msg_date":"Fri, 19 Nov 2021 21:18:23 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"Noah Misch <noah@leadboat.com> writes:\n> Tom Turelinckx, are you able to provide remote access to kittiwake or\n> tadarida? I'd use it to attempt the above things. All else being equal,\n> kittiwake is more relevant since it's still supported upstream.\n\nsnapper just exhibited the same failure, too:\n\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=snapper&dt=2021-11-18%2016%3A09%3A49\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 20 Nov 2021 00:22:28 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"I wrote:\n> snapper just exhibited the same failure, too:\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=snapper&dt=2021-11-18%2016%3A09%3A49\n\nI grepped the buildfarm logs for all recent (last 3 months) occurrences of\n'could not read two-phase state'. Here's the results:\n\n sysname | branch | snapshot | stage | l \n-----------+---------------+---------------------+-------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------\n kittiwake | REL9_6_STABLE | 2021-10-24 12:01:10 | pgbenchCheck | # 'client 1 aborted in state 3: ERROR: could not read two-phase state from xlog at 0/158F4E0\n kittiwake | REL_13_STABLE | 2021-10-26 12:51:11 | ContribCheck-en_US.utf8 | # 'pgbench: error: client 3 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/168C8D8\n kittiwake | REL_14_STABLE | 2021-11-08 15:42:35 | ContribCheck-en_US.utf8 | # 'pgbench: error: client 0 script 0 aborted in command 3 query 0: ERROR: could not read two-phase state from WAL at 0/17ABF48\n kittiwake | REL_13_STABLE | 2021-11-16 15:00:52 | ContribCheck-en_US.utf8 | # 'pgbench: error: client 3 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/1668020: incorrect resource manager data checksum in record at 0/1668020\n snapper | REL_14_STABLE | 2021-11-18 16:09:49 | contrib-amcheckCheck | # 'pgbench: error: client 3 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/1770328: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 7798784\n tadarida | REL_11_STABLE | 2021-11-11 13:29:58 | pgbenchCheck | # 'client 3 aborted in command 3 (SQL) of script 0; ERROR: could not read two-phase state from WAL at 0/1716C68\n tadarida | REL_10_STABLE | 2021-11-12 13:01:15 | pgbenchCheck | # 'client 4 aborted in command 3 of script 0; ERROR: could not read two-phase state from WAL at 0/16F1850: invalid record length at 0/16F1850: wanted 24, got 0\n tadarida | HEAD | 2021-11-17 13:01:24 | contrib-amcheckCheck | # 'pgbench: error: client 0 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/159EF88: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 5890048\n\nSo not all are exactly 'unexpected pageaddr 0/0', but they do all\nlook like we read garbage data.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 20 Nov 2021 00:55:16 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Fri, Nov 19, 2021 at 09:18:23PM -0800, Noah Misch wrote:\n> On Wed, Nov 17, 2021 at 11:05:06PM -0800, Noah Misch wrote:\n> > On Wed, Nov 17, 2021 at 05:47:10PM -0500, Tom Lane wrote:\n> > > Noah Misch <noah@leadboat.com> writes:\n> > > > Each of the three failures happened on a sparc64 Debian+gcc machine. I had\n> > > > tried ~8000 iterations on thorntail, another sparc64 Debian+gcc animal,\n> > > > without reproducing this.\n> > \n> > > # 'pgbench: error: client 0 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/159EF88: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 5890048\n> > > [1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-17%2013%3A01%3A24\n> > \n> > Two others:\n> > ERROR: could not read two-phase state from WAL at 0/16F1850: invalid record length at 0/16F1850: wanted 24, got 0\n> > -- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-12%2013%3A01%3A15\n> > ERROR: could not read two-phase state from WAL at 0/1668020: incorrect resource manager data checksum in record at 0/1668020\n> > -- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kittiwake&dt=2021-11-16%2015%3A00%3A52\n\n> > I don't have a great theory, but here are candidates to examine next:\n> > \n> > - Run with wal_debug=on to confirm logged write location matches read location.\n> > - Run \"PGDATA=contrib/amcheck/tmp_check/t_003_cic_2pc_CIC_2PC_test_data/pgdata\n> > pg_waldump -s 0/01000000\" at the end of the test.\n> > - Dump WAL page binary image at the point of failure.\n> > - Log which branches in XLogReadRecord() are taken.\n> \n> Tom Turelinckx, are you able to provide remote access to kittiwake or\n> tadarida? I'd use it to attempt the above things. All else being equal,\n> kittiwake is more relevant since it's still supported upstream.\n\nThanks for setting up access. Summary: this kernel has a bug in I/O syscalls.\nHow practical is it to update that kernel? (Userland differs across these\nanimals, but the kernel does not.) The kernel on buildfarm member thorntail\ndoesn't exhibit the bug.\n\nFor specifics of the kernel bug, see the attached test program. In brief, the\nbug arises if one process is write()ing or pwrite()ing a file at about the\nsame time that another process is read()ing or pread()ing the same. POSIX\nsays the reader should see the data as it existed before the write or the\nnewly-written data. On this kernel, the reader can see zeros instead. That\nleads to the $SUBJECT failure. PostgreSQL processes write out a given WAL\nblock 20-30 times in ~10ms, and COMMIT PREPARED reads that block. The writers\naren't changing the bytes of interest to COMMIT PREPARED, but the zeros from\nthe kernel bug yield the failure. We could opt to work around that by writing\nonly the not-already-written portion of a WAL block, but I doubt that's\nworthwhile unless it happens to be a performance win anyway.\n\nSeparately, while I don't know of relevance to PostgreSQL, I was interested to\nsee that CentOS 7 pwrite()/pread() fail to have the POSIX-required atomicity.\n\n> The setup of your buildfarm animals provides a clue. I understand kittiwake\n> is the same as ibisbill except for build options, and tadarida is the same as\n> mussurana except for build options. ibisbill and mussurana haven't failed, so\n> one of these is likely needed to reproduce:\n> \n> absence of --enable-cassert\n> CFLAGS='-g -O2 -fstack-protector -Wformat -Werror=format-security '\n> CPPFLAGS='-Wdate-time -D_FORTIFY_SOURCE=2'\n> LDFLAGS='-Wl,-z,relro -Wl,-z,now'\n\nThat was a red herring. ibisbill and mussurana don't use --with-tap-tests.\nAdding --with-tap-tests has been enough to make their configurations witness\nthe same kinds of failures.\n\nnm","msg_date":"Sat, 15 Jan 2022 23:12:10 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Sun, Jan 16, 2022 at 8:12 PM Noah Misch <noah@leadboat.com> wrote:\n> For specifics of the kernel bug, see the attached test program. In brief, the\n> bug arises if one process is write()ing or pwrite()ing a file at about the\n> same time that another process is read()ing or pread()ing the same. POSIX\n> says the reader should see the data as it existed before the write or the\n> newly-written data. On this kernel, the reader can see zeros instead. That\n> leads to the $SUBJECT failure. PostgreSQL processes write out a given WAL\n> block 20-30 times in ~10ms, and COMMIT PREPARED reads that block. The writers\n> aren't changing the bytes of interest to COMMIT PREPARED, but the zeros from\n> the kernel bug yield the failure. We could opt to work around that by writing\n> only the not-already-written portion of a WAL block, but I doubt that's\n> worthwhile unless it happens to be a performance win anyway.\n>\n> Separately, while I don't know of relevance to PostgreSQL, I was interested to\n> see that CentOS 7 pwrite()/pread() fail to have the POSIX-required atomicity.\n\nFWIW there was some related discussion over here:\n\nhttps://www.postgresql.org/message-id/flat/17064-bb0d7904ef72add3%40postgresql.org\n\n\n","msg_date":"Sun, 16 Jan 2022 22:19:30 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"Cancel that kernel upgrade idea. I no longer expect it to help...\n\nOn Sun, Jan 16, 2022 at 10:19:30PM +1300, Thomas Munro wrote:\n> On Sun, Jan 16, 2022 at 8:12 PM Noah Misch <noah@leadboat.com> wrote:\n> > For specifics of the kernel bug, see the attached test program. In brief, the\n> > bug arises if one process is write()ing or pwrite()ing a file at about the\n> > same time that another process is read()ing or pread()ing the same. POSIX\n> > says the reader should see the data as it existed before the write or the\n> > newly-written data. On this kernel, the reader can see zeros instead. That\n> > leads to the $SUBJECT failure. PostgreSQL processes write out a given WAL\n> > block 20-30 times in ~10ms, and COMMIT PREPARED reads that block. The writers\n> > aren't changing the bytes of interest to COMMIT PREPARED, but the zeros from\n> > the kernel bug yield the failure.\n\nThe difference between kittiwake and thorntail comes from thorntail using xfs\nand kittiwake using ext4. Running the io-rectitude.c tests on an ext4\npartition on thorntail, I see the zeros bug just like I do on kittiwake. I\ndon't see the zeros bug on ppc64 or x86_64, just sparc64 so far:\n\n * ext4, Linux 3.10.0-1160.49.1.el7.x86_64 (CentOS 7.9.2009):\n * pwrite/pread is non-atomic if count>16 (no -D switches)\n * write/read is atomic (-DUSE_SEEK -DXLOG_BLCKSZ=8192000)\n * pwrite/pread is free from zeros bug (-DCHANGE_CONTENT=0)\n * write/read is free from zeros bug (-DUSE_SEEK -DCHANGE_CONTENT=0)\n *\n * ext4, Linux 4.9.0-13-sparc64-smp (Debian):\n * pwrite/pread is non-atomic if count>4 (no -D switches)\n * write/read is non-atomic if count>4 (-DUSE_SEEK)\n * write/read IS atomic w/o REOPEN (-DUSE_SEEK -DREOPEN=0 -DXLOG_BLCKSZ=8192000)\n * pwrite/pread has zeros bug for count>127 (-DCHANGE_CONTENT=0)\n * pwrite/pread w/ O_SYNC has zeros bug (-DCHANGE_CONTENT=0 -DOPEN_FLAGS=O_SYNC)\n * far less frequent w/ O_SYNC, but it still happens\n * pwrite/pread w/o REOPEN also has zeros bug for count>127 (-DCHANGE_CONTENT=0 -DREOPEN=0)\n * write/read has zeros bug for count>127 (-DUSE_SEEK -DCHANGE_CONTENT=0)\n * write/read w/ O_SYNC has zeros bug (-DUSE_SEEK -DCHANGE_CONTENT=0 -DOPEN_FLAGS=O_SYNC)\n * write/read w/o REOPEN is free from zeros bug (-DUSE_SEEK -DCHANGE_CONTENT=0 -DREOPEN=0)\n *\n * ext4, Linux 5.15.0-2-sparc64-smp (Debian bookworm/sid):\n * [behaviors match the previous kernel exactly]\n *\n * ext4, Linux 5.15.0-2-powerpc64 (Debian bookworm/sid):\n * [atomicity matches previous kernel, but zeros bug does not]\n * pwrite/pread is non-atomic if count>4 (no -D switches)\n * write/read is non-atomic if count>4 (-DUSE_SEEK)\n * write/read IS atomic w/o REOPEN (-DUSE_SEEK -DREOPEN=0 -DXLOG_BLCKSZ=8192000)\n * pwrite/pread is free from zeros bug (-DCHANGE_CONTENT=0)\n * write/read is free from zeros bug (-DUSE_SEEK -DCHANGE_CONTENT=0)\n *\n * ext4, Linux 5.15.5-0-virt x86_64 (Alpine):\n * [behaviors match the previous kernel exactly]\n *\n * xfs, Linux 5.15.0-2-sparc64-smp (Debian bookworm/sid):\n * pwrite/pread is atomic (-DXLOG_BLCKSZ=8192000)\n * write/read is atomic (-DUSE_SEEK -DXLOG_BLCKSZ=8192000)\n * pwrite/pread is free from zeros bug (-DCHANGE_CONTENT=0)\n * write/read is free from zeros bug (-DUSE_SEEK -DCHANGE_CONTENT=0)\n\n> > We could opt to work around that by writing\n> > only the not-already-written portion of a WAL block, but I doubt that's\n> > worthwhile unless it happens to be a performance win anyway.\n\nMy next steps:\n\n- Report a Debian bug for the sparc64+ext4 zeros problem.\n\n- Try to falsify the idea that \"write only the not-already-written portion of\n a WAL block\" is an effective workaround. Specifically, modify the test\n program to have the writer process mutate offsets [N-k,N-1] and [N+1,N+k]\n while the reader process reads offset N. If the reader sees a zero, that\n workaround is ineffective.\n\n- Implement the workaround, if I didn't falsify its effectiveness. If it\n doesn't hurt performance on x86_64, we can use it unconditionally.\n Otherwise, limit its use to sparc64 Linux.\n\n> > Separately, while I don't know of relevance to PostgreSQL, I was interested to\n> > see that CentOS 7 pwrite()/pread() fail to have the POSIX-required atomicity.\n> \n> FWIW there was some related discussion over here:\n> \n> https://www.postgresql.org/message-id/flat/17064-bb0d7904ef72add3%40postgresql.org\n\nThat gave me the idea to test different filesystems. Thanks. Incidentally, I\nfind https://utcc.utoronto.ca/~cks/space/blog/unix/WriteNotVeryAtomic is\nmistaken about POSIX requirements. There's no precedent for POSIX writing\n\"two threads\" when it means \"two threads of the same process\". Moreover, the\npart about \"shall also apply whenever a file descriptor is successfully\nclosed, however caused (for example [...] process termination)\" would be\nsuperfluous in a requirement specific to threads of one process. Having said\nthat, if the most-prominent POSIX regular file implementation (ext4 on x86_64)\ndoesn't implement a POSIX requirement, that has the same practical\nconsequences for PostgreSQL as POSIX not requiring it.\n\nI now see newer Linux ext4 has drifted further away from POSIX atomicity,\ncompared to CentOS 7. In CentOS 7 ext4, plain write()/read() was still\natomic. By Linux 5.15.5, those abandoned atomicity.\n\n\n","msg_date":"Sun, 16 Jan 2022 13:02:41 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Mon, Jan 17, 2022 at 10:02 AM Noah Misch <noah@leadboat.com> wrote:\n> - Report a Debian bug for the sparc64+ext4 zeros problem.\n\nI suspect that 027_stream_regress.pl hits this kernel bug with high\nprobability[1]. I wonder if the owner of kittiwake and tadarida would\nconsider setting up an xfs file system? Or alternatively, since ext4\ndidn't support concurrent writes until recently, I wonder if there is\nan option somewhere to turn the new concurrency stuff off, or failing\nthat, if we could temporarily downgrade the kernel to an older version\nthat does inode-level read/write locking.\n\n[1] https://www.postgresql.org/message-id/CA%2BhUKG%2BeuZ%3Ddc27ZB%3Ds74x0q%3DzU%3D2%3Dvs8%2B6TkJoTUiCPUd2dQA%40mail.gmail.com\n\n\n","msg_date":"Fri, 21 Jan 2022 08:34:22 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Fri, Jan 21, 2022 at 08:34:22AM +1300, Thomas Munro wrote:\n> On Mon, Jan 17, 2022 at 10:02 AM Noah Misch <noah@leadboat.com> wrote:\n> > - Report a Debian bug for the sparc64+ext4 zeros problem.\n> \n> I suspect that 027_stream_regress.pl hits this kernel bug with high\n> probability[1]. I wonder if the owner of kittiwake and tadarida would\n> consider setting up an xfs file system? Or alternatively, since ext4\n> didn't support concurrent writes until recently, I wonder if there is\n> an option somewhere to turn the new concurrency stuff off, or failing\n> that, if we could temporarily downgrade the kernel to an older version\n> that does inode-level read/write locking.\n\nIf the write-only-new-bytes approach works, I think we'd want to revert those\nchanges. Perhaps a cheaper stopgap is to make the affected tests skip on\nsparc Linux. Is that worth doing? (Could even limit the skip to ext4,\ne.g. by testing \"df -x ext4 . >/dev/null\".)\n\n> [1] https://www.postgresql.org/message-id/CA%2BhUKG%2BeuZ%3Ddc27ZB%3Ds74x0q%3DzU%3D2%3Dvs8%2B6TkJoTUiCPUd2dQA%40mail.gmail.com\n\n\n","msg_date":"Thu, 20 Jan 2022 19:30:42 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Sun, Jan 16, 2022 at 01:02:41PM -0800, Noah Misch wrote:\n> My next steps:\n> \n> - Report a Debian bug for the sparc64+ext4 zeros problem.\n\n(Not done yet.)\n\n> - Try to falsify the idea that \"write only the not-already-written portion of\n> a WAL block\" is an effective workaround. Specifically, modify the test\n> program to have the writer process mutate offsets [N-k,N-1] and [N+1,N+k]\n> while the reader process reads offset N. If the reader sees a zero, that\n> workaround is ineffective.\n\nThe reader did not see a zero. In addition to bytes outside the write being\nimmune to the zeros bug, the first and last forty bytes of a write were immune\nto the zeros bug.\n\n> - Implement the workaround, if I didn't falsify its effectiveness. If it\n> doesn't hurt performance on x86_64, we can use it unconditionally.\n> Otherwise, limit its use to sparc64 Linux.\n\nAttached. With this, kittiwake has survived 8.5hr of 003_cic_2pc.pl. Without\nthe patch, it failed many times, always within 1.3hr. For easier review, this\npatch uses the new behavior on all platforms. Before commit and back-patch, I\nplan to limit use of the new behavior to sparc Linux. Future work can\nbenchmark the new behavior and, if it performs well, make it unconditional in\nv15+. I would expect performance to be unchanged or slightly better, because\nthe new behavior requests less futile work from the OS.","msg_date":"Sat, 22 Jan 2022 10:52:41 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Sun, Jan 23, 2022 at 7:52 AM Noah Misch <noah@leadboat.com> wrote:\n> Attached. With this, kittiwake has survived 8.5hr of 003_cic_2pc.pl. Without\n> the patch, it failed many times, always within 1.3hr. For easier review, this\n> patch uses the new behavior on all platforms. Before commit and back-patch, I\n> plan to limit use of the new behavior to sparc Linux. Future work can\n> benchmark the new behavior and, if it performs well, make it unconditional in\n> v15+. I would expect performance to be unchanged or slightly better, because\n> the new behavior requests less futile work from the OS.\n\nOne detail is that wal_level=open_datasync, wal_senders=0,\nwal_level=minimal will panic, because O_DIRECT requires fs\npage-aligned access (and fails in various other ways on other OSes, eg\nexpensive read-before-write every time). That's an ultra-niche\nconcern likely affecting nobody, especially when multiplied by the\nodds that anyone is using that stack at all (considering that\nstreaming rep has apparently been borked for years on linux/sparc/ext4\nand nobody told us).\n\nI was +1 for the control file locking change in that other thread (I\nview the atomicity stuff as Linux-realpolitik-vs-POSIX, not to mention\nthat we run Windows too, which requires separate analysis). I'm less\nsure it makes sense to do anything to support the presumed bogus\nzeroes bug for (probably) no real users, especially before we've even\nreported it and heard some analysis, for example acceptance that it's\nbroken and confirmation that this really is just a sparc problem.\n\n\n","msg_date":"Mon, 24 Jan 2022 09:42:13 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"Hi,\n\nOn 2022-01-24 09:42:13 +1300, Thomas Munro wrote:\n> On Sun, Jan 23, 2022 at 7:52 AM Noah Misch <noah@leadboat.com> wrote:\n> > Future work can benchmark the new behavior and, if it performs well, make\n> > it unconditional in v15+. I would expect performance to be unchanged or\n> > slightly better, because the new behavior requests less futile work from\n> > the OS.\n\nI doubt it'll be generally applicable. Turning a write operation into a\nread-write isn't free. Yes, often enough it's likely that the prior page is\nstill in cache, but I don't think we can rely on that in general.\n\nIt also just fundamentally locks us into never using O_DIRECT in anger. I\ndon't think that's a good direction.\n\n\n> One detail is that wal_level=open_datasync, wal_senders=0,\n> wal_level=minimal will panic, because O_DIRECT requires fs\n> page-aligned access (and fails in various other ways on other OSes, eg\n> expensive read-before-write every time). That's an ultra-niche\n> concern likely affecting nobody, especially when multiplied by the\n> odds that anyone is using that stack at all (considering that\n> streaming rep has apparently been borked for years on linux/sparc/ext4\n> and nobody told us).\n\nSeems like the patch should at least make this error out?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sun, 23 Jan 2022 14:26:53 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Mon, Jan 24, 2022 at 09:42:13AM +1300, Thomas Munro wrote:\n> I'm less\n> sure it makes sense to do anything to support the presumed bogus\n> zeroes bug for (probably) no real users, especially before we've even\n> reported it and heard some analysis, for example acceptance that it's\n> broken and confirmation that this really is just a sparc problem.\n\nGot it. I've already done a bad thing leaving the buildfarm broken for three\nmonths, so I don't want to let the buildfarm wait for a kernel fix. These are\nthe two main options I'm seeing now:\n\n(a) Modify the tests so the affected animals can skip affected tests by\nsetting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\n\n(b) Remove --enable-tap-tests from affected animals.\n\nDo you have a preference among those two or some other option that gets the\nbuildfarm green on a predictable schedule? I somewhat prefer (a), since\n--enable-tap-tests is where most of the interesting buildfarm reports happen\nthese days.\n\n\n","msg_date":"Sun, 23 Jan 2022 15:29:27 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Mon, Jan 24, 2022 at 12:29 PM Noah Misch <noah@leadboat.com> wrote:\n> On Mon, Jan 24, 2022 at 09:42:13AM +1300, Thomas Munro wrote:\n> > I'm less\n> > sure it makes sense to do anything to support the presumed bogus\n> > zeroes bug for (probably) no real users, especially before we've even\n> > reported it and heard some analysis, for example acceptance that it's\n> > broken and confirmation that this really is just a sparc problem.\n>\n> Got it. I've already done a bad thing leaving the buildfarm broken for three\n> months, so I don't want to let the buildfarm wait for a kernel fix. These are\n> the two main options I'm seeing now:\n>\n> (a) Modify the tests so the affected animals can skip affected tests by\n> setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\n>\n> (b) Remove --enable-tap-tests from affected animals.\n>\n> Do you have a preference among those two or some other option that gets the\n> buildfarm green on a predictable schedule? I somewhat prefer (a), since\n> --enable-tap-tests is where most of the interesting buildfarm reports happen\n> these days.\n\nTrying out a new idea: what if we could tell the buildfarm website\nthat a certain test is currently expected to fail for reasons we can't\nfix yet (configuration change needed but owner not responding, or\nbugfix from another project needed, etc)? That could cause it to be\ndisplayed in a different shade of green, or grey, or whatever? Other\nkinds of failures would still show as red. Perhaps this would be\nconfigured with a file in a git repo that any committer can push to.\n\n\n","msg_date":"Mon, 24 Jan 2022 12:49:16 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Mon, Jan 24, 2022 at 12:49:16PM +1300, Thomas Munro wrote:\n> On Mon, Jan 24, 2022 at 12:29 PM Noah Misch <noah@leadboat.com> wrote:\n> > On Mon, Jan 24, 2022 at 09:42:13AM +1300, Thomas Munro wrote:\n> > > I'm less\n> > > sure it makes sense to do anything to support the presumed bogus\n> > > zeroes bug for (probably) no real users, especially before we've even\n> > > reported it and heard some analysis, for example acceptance that it's\n> > > broken and confirmation that this really is just a sparc problem.\n> >\n> > Got it. I've already done a bad thing leaving the buildfarm broken for three\n> > months, so I don't want to let the buildfarm wait for a kernel fix. These are\n> > the two main options I'm seeing now:\n> >\n> > (a) Modify the tests so the affected animals can skip affected tests by\n> > setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\n> >\n> > (b) Remove --enable-tap-tests from affected animals.\n> >\n> > Do you have a preference among those two or some other option that gets the\n> > buildfarm green on a predictable schedule? I somewhat prefer (a), since\n> > --enable-tap-tests is where most of the interesting buildfarm reports happen\n> > these days.\n> \n> Trying out a new idea: what if we could tell the buildfarm website\n> that a certain test is currently expected to fail for reasons we can't\n> fix yet (configuration change needed but owner not responding, or\n> bugfix from another project needed, etc)? That could cause it to be\n> displayed in a different shade of green, or grey, or whatever? Other\n> kinds of failures would still show as red. Perhaps this would be\n> configured with a file in a git repo that any committer can push to.\n\nThat would be a better capability to use if we had it, agreed. Is it feasible\nto acquire that capability soon enough?\n\n\n","msg_date":"Sun, 23 Jan 2022 16:19:14 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"Noah Misch <noah@leadboat.com> writes:\n> On Mon, Jan 24, 2022 at 12:49:16PM +1300, Thomas Munro wrote:\n>> Trying out a new idea: what if we could tell the buildfarm website\n>> that a certain test is currently expected to fail for reasons we can't\n>> fix yet (configuration change needed but owner not responding, or\n>> bugfix from another project needed, etc)? That could cause it to be\n>> displayed in a different shade of green, or grey, or whatever? Other\n>> kinds of failures would still show as red. Perhaps this would be\n>> configured with a file in a git repo that any committer can push to.\n\n> That would be a better capability to use if we had it, agreed. Is it feasible\n> to acquire that capability soon enough?\n\nIt's not merely a website issue: you'd really rather that the\nbuildfarm animal runs the rest of the tests rather than going belly-up\nafter an expected failure. I think your suggestion about skipping\nproblematic tests based on an environment variable is more practical\nin the near term. We already have some cases like that, too, eg in\nsrc/bin/psql/t/010_tab_completion.pl.\n\nIn the long term I could get behind having some less ad-hoc way\nof skipping tests, but I don't think we can have that quickly.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 23 Jan 2022 19:28:47 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"\n\nOn January 23, 2022 3:29:27 PM PST\n>(a) Modify the tests so the affected animals can skip affected tests by\n>setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\n\nWhy not just detect the problem in the tap test and skip, rather than requiring multiple buildfarm configs to be changed as well as the test itself? \n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.\n\n\n","msg_date":"Sun, 23 Jan 2022 17:03:04 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Sun, Jan 23, 2022 at 05:03:04PM -0800, Andres Freund wrote:\n> On January 23, 2022 3:29:27 PM PST\n> >(a) Modify the tests so the affected animals can skip affected tests by\n> >setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\n> \n> Why not just detect the problem in the tap test and skip, rather than requiring multiple buildfarm configs to be changed as well as the test itself? \n\nEnd users running PostgreSQL test suites to acceptance-test their stack should\nconsider the affected stack unusable for PostgreSQL. Hence, I ruled out that\napproach, despite having implemented it at one point. Under some plausible\nset of goals, it is optimal.\n\n\n","msg_date":"Sun, 23 Jan 2022 17:17:59 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"Hi,\n\nOn 2022-01-23 17:17:59 -0800, Noah Misch wrote:\n> On Sun, Jan 23, 2022 at 05:03:04PM -0800, Andres Freund wrote:\n> > On January 23, 2022 3:29:27 PM PST\n> > >(a) Modify the tests so the affected animals can skip affected tests by\n> > >setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\n> > \n> > Why not just detect the problem in the tap test and skip, rather than requiring multiple buildfarm configs to be changed as well as the test itself? \n> \n> End users running PostgreSQL test suites to acceptance-test their stack should\n> consider the affected stack unusable for PostgreSQL.\n\nI'd bet that that's zero users ;)\n\n\n> Hence, I ruled out that\n> approach, despite having implemented it at one point. Under some plausible\n> set of goals, it is optimal.\n\nIt's not perfect due to the way we run our tests (seeing output is hard, it's\nnot aggregated), but marking the test as todo rather than SKIP seems like the\nmost appropriate test status. It's known to be a problem, we've not fixed it,\nbut we want to be able to run the tests.\n\nTest::more's description: \"If it's something the programmer hasn't done yet,\nuse TODO. This is for any code you haven't written yet, or bugs you have yet\nto fix, but want to put tests in your testing script (always a good idea).\"\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sun, 23 Jan 2022 17:40:54 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Sun, Jan 23, 2022 at 05:40:54PM -0800, Andres Freund wrote:\n> On 2022-01-23 17:17:59 -0800, Noah Misch wrote:\n> > On Sun, Jan 23, 2022 at 05:03:04PM -0800, Andres Freund wrote:\n> > > On January 23, 2022 3:29:27 PM PST\n> > > >(a) Modify the tests so the affected animals can skip affected tests by\n> > > >setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\n> > > \n> > > Why not just detect the problem in the tap test and skip, rather than requiring multiple buildfarm configs to be changed as well as the test itself? \n> > \n> > End users running PostgreSQL test suites to acceptance-test their stack should\n> > consider the affected stack unusable for PostgreSQL.\n> \n> I'd bet that that's zero users ;)\n\nWouldn't surprise me. I'm attaching what I had written and discarded. If\nnobody else hates it, I can live with it.\n\n> > Hence, I ruled out that\n> > approach, despite having implemented it at one point. Under some plausible\n> > set of goals, it is optimal.\n> \n> It's not perfect due to the way we run our tests (seeing output is hard, it's\n> not aggregated), but marking the test as todo rather than SKIP seems like the\n> most appropriate test status. It's known to be a problem, we've not fixed it,\n> but we want to be able to run the tests.\n> \n> Test::more's description: \"If it's something the programmer hasn't done yet,\n> use TODO. This is for any code you haven't written yet, or bugs you have yet\n> to fix, but want to put tests in your testing script (always a good idea).\"\n\nCould do that. Every run that doesn't get the flaky failure will print a\nmessage like \"TODO passed: 3-5\", though the test file could mitigate that by\ndeclaring the TODO only on configurations where we expect a failure. The\n027_stream_regress.pl trouble involves reaching a die(), not failing a test in\nthis sense, so that one would take more work.","msg_date":"Sun, 23 Jan 2022 18:10:07 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Sun, Jan 23, 2022 at 06:10:07PM -0800, Noah Misch wrote:\n> Could do that. Every run that doesn't get the flaky failure will print a\n> message like \"TODO passed: 3-5\", though the test file could mitigate that by\n> declaring the TODO only on configurations where we expect a failure. The\n> 027_stream_regress.pl trouble involves reaching a die(), not failing a test in\n> this sense, so that one would take more work.\n\nUsing a TODO has the advantage to allow the tests to run on a periodic\nbasis, even if they could fail in this unexpected way. We've had our\nload of issues in the past proper to specific architectures, so this\ncould help with the 2PC code paths and SPARC. Not to mention that we\ncould just let the code be as-is, and it would fix itself once the\nkernel is updated. So that sounds much better to me than a skip\nphase.\n--\nMichael","msg_date":"Mon, 24 Jan 2022 11:20:21 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Sun, Jan 23, 2022 at 06:10:07PM -0800, Noah Misch wrote:\n>> Could do that. Every run that doesn't get the flaky failure will print a\n>> message like \"TODO passed: 3-5\", though the test file could mitigate that by\n>> declaring the TODO only on configurations where we expect a failure. The\n>> 027_stream_regress.pl trouble involves reaching a die(), not failing a test in\n>> this sense, so that one would take more work.\n\n> Using a TODO has the advantage to allow the tests to run on a periodic\n> basis, even if they could fail in this unexpected way.\n\nI'm okay with this *if* the TODO marking can be constrained to platforms\nwhere we know there's a problem. Otherwise I'm afraid it will mask\nunexpected problems.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 23 Jan 2022 21:25:04 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"Hi,\n\nOn 2022-01-23 18:10:07 -0800, Noah Misch wrote:\n> On Sun, Jan 23, 2022 at 05:40:54PM -0800, Andres Freund wrote:\n> > Test::more's description: \"If it's something the programmer hasn't done yet,\n> > use TODO. This is for any code you haven't written yet, or bugs you have yet\n> > to fix, but want to put tests in your testing script (always a good idea).\"\n>\n> Could do that. Every run that doesn't get the flaky failure will print a\n> message like \"TODO passed: 3-5\", though the test file could mitigate that by\n> declaring the TODO only on configurations where we expect a failure.\n\nYea, that's what I was thinking we'd do.\n\n\n> The\n> 027_stream_regress.pl trouble involves reaching a die(), not failing a test in\n> this sense, so that one would take more work.\n\nI'm really not a perl person... But my understanding is that todo_skip() would\naddress this? I.e. something like\n\nTODO:\n{\n $todo_skip \"linux/sparc has unaddressed problems with partial page overwrites\"\n if ($^O eq 'linux' and $Config{archname'} ~= 'sparc');\n\n ok(whatever is broken);\n}\n\n(no idea if the above todo condition even approximates something working)\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sun, 23 Jan 2022 18:34:32 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On 2022-01-23 21:25:04 -0500, Tom Lane wrote:\n> Michael Paquier <michael@paquier.xyz> writes:\n> > On Sun, Jan 23, 2022 at 06:10:07PM -0800, Noah Misch wrote:\n> >> Could do that. Every run that doesn't get the flaky failure will print a\n> >> message like \"TODO passed: 3-5\", though the test file could mitigate that by\n> >> declaring the TODO only on configurations where we expect a failure. The\n> >> 027_stream_regress.pl trouble involves reaching a die(), not failing a test in\n> >> this sense, so that one would take more work.\n> \n> > Using a TODO has the advantage to allow the tests to run on a periodic\n> > basis, even if they could fail in this unexpected way.\n> \n> I'm okay with this *if* the TODO marking can be constrained to platforms\n> where we know there's a problem. Otherwise I'm afraid it will mask\n> unexpected problems.\n\nYep. Very weird syntax... TODO blocks only take effect if a local $TODO is set\n(to a text explaining the reason hopefully) or if there's a todo_skip. So\nsomething like\n\nTODO:\n{\n local $TODO = 'linux + sparc is borked' if $^O eq 'linux' and ...;\n\n tests...\n}\n\nwould run the tests and accept failures if the if matches, and otherwise not.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sun, 23 Jan 2022 18:38:34 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Sun, Jan 23, 2022 at 06:34:32PM -0800, Andres Freund wrote:\n> On 2022-01-23 18:10:07 -0800, Noah Misch wrote:\n> > On Sun, Jan 23, 2022 at 05:40:54PM -0800, Andres Freund wrote:\n> > > Test::more's description: \"If it's something the programmer hasn't done yet,\n> > > use TODO. This is for any code you haven't written yet, or bugs you have yet\n> > > to fix, but want to put tests in your testing script (always a good idea).\"\n> >\n> > Could do that. Every run that doesn't get the flaky failure will print a\n> > message like \"TODO passed: 3-5\", though the test file could mitigate that by\n> > declaring the TODO only on configurations where we expect a failure.\n> \n> Yea, that's what I was thinking we'd do.\n> \n> > The\n> > 027_stream_regress.pl trouble involves reaching a die(), not failing a test in\n> > this sense, so that one would take more work.\n> \n> I'm really not a perl person... But my understanding is that todo_skip() would\n> address this? I.e. something like\n> \n> TODO:\n> {\n> $todo_skip \"linux/sparc has unaddressed problems with partial page overwrites\"\n> if ($^O eq 'linux' and $Config{archname'} ~= 'sparc');\n> \n> ok(whatever is broken);\n> }\n\nYes. todo_skip() behaves much like regular skip(). The enclosed tests don't\nrun. Hence, it prevents die() and BAIL_OUT() failures. $TODO is a different\nbeast; tests still run, and it changes the reporting. For 003_cic_2pc.pl, I'm\nfine using $TODO so we continue to run all test commands and quietly log their\nresults. For 027_stream_regress.pl, which would need deep changes to use\n$TODO, it works to use any of todo_skip, skip, or skip_all. I prefer\nskip_all, because it prints the skip reason to gmake's stdout. (If the number\nof affected users is zero as theorized, the choice doesn't matter.) Any\nobjections? Here's the appearance of each strategy on gmake's stdout:\n\n=== todo_skip\n[23:56:45] t/003_cic_2pc.pl .. ok 60 ms ( 0.00 usr 0.00 sys + 0.05 cusr 0.00 csys = 0.05 CPU)\n[23:56:45]\nAll tests successful.\nFiles=1, Tests=5, 1 wallclock secs ( 0.01 usr 0.00 sys + 0.05 cusr 0.00 csys = 0.06 CPU)\nResult: PASS\n\n=== skip\n[23:55:47] t/003_cic_2pc.pl .. ok 59 ms ( 0.00 usr 0.00 sys + 0.05 cusr 0.00 csys = 0.05 CPU)\n[23:55:48]\nAll tests successful.\nFiles=1, Tests=5, 1 wallclock secs ( 0.01 usr 0.00 sys + 0.05 cusr 0.00 csys = 0.06 CPU)\nResult: PASS\n\n=== skip_all\n[23:31:04] t/003_cic_2pc.pl .. skipped: filesystem bug\n[23:31:04]\nFiles=1, Tests=0, 0 wallclock secs ( 0.02 usr 0.00 sys + 0.05 cusr 0.00 csys = 0.07 CPU)\nResult: NOTESTS\n\n=== $TODO, test 1 is expected fail, tests 2-5 are unexpected pass\n[23:32:32] t/003_cic_2pc.pl .. ok 1371 ms ( 0.01 usr 0.00 sys + 0.51 cusr 0.24 csys = 0.76 CPU)\n[23:32:33]\nAll tests successful.\n\nTest Summary Report\n-------------------\nt/003_cic_2pc.pl (Wstat: 0 Tests: 5 Failed: 0)\n TODO passed: 2-5\nFiles=1, Tests=5, 1 wallclock secs ( 0.02 usr 0.00 sys + 0.51 cusr 0.24 csys = 0.77 CPU)\nResult: PASS\n\n\n","msg_date":"Mon, 24 Jan 2022 00:02:43 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Mon, Jan 24, 2022 at 12:02:43AM -0800, Noah Misch wrote:\n> For 003_cic_2pc.pl, I'm\n> fine using $TODO so we continue to run all test commands and quietly log their\n> results. For 027_stream_regress.pl, which would need deep changes to use\n> $TODO, it works to use any of todo_skip, skip, or skip_all. I prefer\n> skip_all, because it prints the skip reason to gmake's stdout. (If the number\n> of affected users is zero as theorized, the choice doesn't matter.) Any\n> objections?\n\nI pushed that on 2022-01-26 as ce6d793, but it wasn't enough:\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=snapper&dt=2022-02-01%2013%3A29%3A58\n\nIn older Test::More versions, $TODO only works if tests know the right package\nin which to examine $that_pkg::TODO. In PostgreSQL v12+, all relevant\nfunctions set $Test::Builder::Level, so Test::More finds $main::TODO as\nintended. This would be one way to fix v11:\n\n--- a/src/test/perl/PostgresNode.pm\n+++ b/src/test/perl/PostgresNode.pm\n@@ -1869,4 +1869,6 @@ command_ok(...)\n sub command_checks_all\n {\n+\tlocal $Test::Builder::Level = $Test::Builder::Level + 1;\n+\n \tmy $self = shift;\n \ndiff --git a/src/test/perl/TestLib.pm b/src/test/perl/TestLib.pm\nindex 0c56156..2ab78b4 100644\n--- a/src/test/perl/TestLib.pm\n+++ b/src/test/perl/TestLib.pm\n@@ -598,4 +598,6 @@ sub command_fails_like\n sub command_checks_all\n {\n+\tlocal $Test::Builder::Level = $Test::Builder::Level + 1;\n+\n \tmy ($cmd, $expected_ret, $out, $err, $test_name) = @_;\n\nHowever, I'm inclined to use a less-fragile way, in all branches:\n\n--- a/src/bin/pgbench/t/023_cic_2pc.pl\n+++ b/src/bin/pgbench/t/023_cic_2pc.pl\n@@ -14 +14,2 @@ use Test::More tests => 6;\n-local $TODO = 'filesystem bug' if TestLib::has_wal_read_bug;\n+Test::More->builder->todo_start('filesystem bug')\n+ if TestLib::has_wal_read_bug;\n\nNone of this matters under newer Test::More. kittiwake's Test::More 1.302133\nis new enough not to care, but tadarida's Test::More 1.001014 is not. The\nTest::More->builder->todo_start() syntax works in 0.82, if not earlier.\n(PostgreSQL v10 requires Test::More 0.87.)\n\n\n","msg_date":"Tue, 1 Feb 2022 21:55:56 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"},{"msg_contents":"On Sun, Jan 16, 2022 at 01:02:41PM -0800, Noah Misch wrote:\n> My next steps:\n> \n> - Report a Debian bug for the sparc64+ext4 zeros problem.\n\nReported to Debian, then upstream:\nhttps://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1006157\nhttps://marc.info/?t=164539269900001\n\nLast week, someone confirmed the behavior on kernel 5.17.0-rc5. No other news.\n\n\n","msg_date":"Thu, 3 Mar 2022 17:41:39 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":true,"msg_subject":"Re: XLogReadRecord() error in XlogReadTwoPhaseData()"}],"string":"[\n {\n \"msg_contents\": \"On Sun, Oct 24, 2021 at 04:35:02PM -0700, Noah Misch wrote:\\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kittiwake&dt=2021-10-24%2012%3A01%3A10\\n> got an interesting v9.6 failure [...]:\\n> \\n> 2021-10-24 14:25:29.263 CEST [34569:175] pgbench ERROR: could not read two-phase state from xlog at 0/158F4E0\\n> 2021-10-24 14:25:29.263 CEST [34569:176] pgbench STATEMENT: COMMIT PREPARED 'c1';\\n\\nAs a first step, let's report the actual XLogReadRecord() error message.\\nAttached. All the other sites that expect no error already do this.\",\n \"msg_date\": \"Sat, 6 Nov 2021 18:31:57 -0700\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 06, 2021 at 06:31:57PM -0700, Noah Misch wrote:\\n> As a first step, let's report the actual XLogReadRecord() error message.\\n> Attached.\\n\\nGood catch! This looks good.\\n\\n> All the other sites that expect no error already do this.\\n\\nIndeed. Looking closer, I think that we'd better improve\\nDecodingContextFindStartpoint(),\\npg_logical_replication_slot_advance(), XLogSendLogical() as well as\\npg_logical_slot_get_changes_guts() to follow a format closer to what\\nyou have in your patch, with an error message that describes the\\ncontext where the problem has been found, instead of just elog()'ing\\nwhat XLogReadRecord() returns.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 8 Nov 2021 13:42:46 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"\\n\\n> 7 нояб. 2021 г., в 06:31, Noah Misch <noah@leadboat.com> написал(а):\\n> \\n> As a first step, let's report the actual XLogReadRecord() error message.\\n> Attached. All the other sites that expect no error already do this.\\n\\nBTW some time ago I've spotted a good number of related unreported errors [0].\\n\\n\\n[0] https://www.postgresql.org/message-id/A0A36AEE-3476-4326-B877-EE2B55BAEEED%40yandex-team.ru\\n\\n\",\n \"msg_date\": \"Mon, 8 Nov 2021 16:38:47 +0500\",\n \"msg_from\": \"Andrey Borodin <x4mmm@yandex-team.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 08, 2021 at 01:42:46PM +0900, Michael Paquier wrote:\\n> Indeed. Looking closer, I think that we'd better improve\\n> DecodingContextFindStartpoint(),\\n> pg_logical_replication_slot_advance(), XLogSendLogical() as well as\\n> pg_logical_slot_get_changes_guts() to follow a format closer to what\\n> you have in your patch, with an error message that describes the\\n> context where the problem has been found, instead of just elog()'ing\\n> what XLogReadRecord() returns.\\n\\nFYI, I have just begun a new thread about those ones:\\nhttps://www.postgresql.org/message-id/YYnTH6OyOwQcAdkw@paquier.xyz\\n--\\nMichael\",\n \"msg_date\": \"Tue, 9 Nov 2021 10:51:40 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 08, 2021 at 01:42:46PM +0900, Michael Paquier wrote:\\n> On Sat, Nov 06, 2021 at 06:31:57PM -0700, Noah Misch wrote:\\n> > On Sun, Oct 24, 2021 at 04:35:02PM -0700, Noah Misch wrote:\\n> > > https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kittiwake&dt=2021-10-24%2012%3A01%3A10\\n> > > got an interesting v9.6 failure [...]:\\n> > > \\n> > > 2021-10-24 14:25:29.263 CEST [34569:175] pgbench ERROR: could not read two-phase state from xlog at 0/158F4E0\\n> > > 2021-10-24 14:25:29.263 CEST [34569:176] pgbench STATEMENT: COMMIT PREPARED 'c1';\\n\\nTom Lane reported another instance today:\\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-11%2013%3A29%3A58\\n\\nEach of the three failures happened on a sparc64 Debian+gcc machine. I had\\ntried ~8000 iterations on thorntail, another sparc64 Debian+gcc animal,\\nwithout reproducing this.\\n\\n> > As a first step, let's report the actual XLogReadRecord() error message.\\n> > Attached.\\n> \\n> Good catch! This looks good.\\n\\nPushed.\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 17:22:34 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"Noah Misch <noah@leadboat.com> writes:\\n> Tom Lane reported another instance today:\\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-11%2013%3A29%3A58\\n\\n> Each of the three failures happened on a sparc64 Debian+gcc machine. I had\\n> tried ~8000 iterations on thorntail, another sparc64 Debian+gcc animal,\\n> without reproducing this.\\n\\n>>> As a first step, let's report the actual XLogReadRecord() error message.\\n>>> Attached.\\n\\n>> Good catch! This looks good.\\n\\n> Pushed.\\n\\nWell, we didn't have to wait too long [1]:\\n\\n# at t/003_cic_2pc.pl line 143.\\n# 'pgbench: error: client 0 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/159EF88: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 5890048\\n# pgbench: error: client 2 script 3 aborted in command 2 query 0: ERROR: canceling statement due to lock timeout\\n# pgbench: fatal: Run was aborted; the above results are incomplete.\\n\\nI suppose \\\"unexpected pageaddr 0/0\\\" is most easily explained by supposing\\nthat XlogReadTwoPhaseData tried to read a WAL page that hadn't been\\nwritten out yet. Have we got any synchronization around that?\\n\\n\\t\\t\\tregards, tom lane\\n\\n[1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-17%2013%3A01%3A24\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 17:47:10 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 05:47:10PM -0500, Tom Lane wrote:\\n> Noah Misch <noah@leadboat.com> writes:\\n> > Each of the three failures happened on a sparc64 Debian+gcc machine. I had\\n> > tried ~8000 iterations on thorntail, another sparc64 Debian+gcc animal,\\n> > without reproducing this.\\n\\n> # 'pgbench: error: client 0 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/159EF88: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 5890048\\n> [1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-17%2013%3A01%3A24\\n\\nTwo others:\\nERROR: could not read two-phase state from WAL at 0/16F1850: invalid record length at 0/16F1850: wanted 24, got 0\\n-- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-12%2013%3A01%3A15\\nERROR: could not read two-phase state from WAL at 0/1668020: incorrect resource manager data checksum in record at 0/1668020\\n-- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kittiwake&dt=2021-11-16%2015%3A00%3A52\\n\\n> I suppose \\\"unexpected pageaddr 0/0\\\" is most easily explained by supposing\\n> that XlogReadTwoPhaseData tried to read a WAL page that hadn't been\\n> written out yet. Have we got any synchronization around that?\\n\\nIf the WAL address isn't on disk, that error doesn't happen. Instead,\\nread_local_xlog_page() blocks waiting for the WAL to become available. It's\\nstill possible that we make the WAL region exist, but it somehow doesn't\\ncontain the right data until shortly later. FinishPreparedTransaction() takes\\nTwoPhaseStateLock and looks for an entry having gxact->valid. EndPrepare()\\nfills gxact->prepare_end_lsn, then calls MarkAsPrepared() to set gxact->valid\\nunder TwoPhaseStateLock. All freelist (freeGXacts) interaction holds\\nTwoPhaseStateLock. I'm not seeing a gap in that synchronization.\\n\\nI don't have a great theory, but here are candidates to examine next:\\n\\n- Run with wal_debug=on to confirm logged write location matches read location.\\n- Run \\\"PGDATA=contrib/amcheck/tmp_check/t_003_cic_2pc_CIC_2PC_test_data/pgdata\\n pg_waldump -s 0/01000000\\\" at the end of the test.\\n- Dump WAL page binary image at the point of failure.\\n- Log which branches in XLogReadRecord() are taken.\\n\\nWhat else might help?\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 23:05:06 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"\\n\\n> 18 нояб. 2021 г., в 12:05, Noah Misch <noah@leadboat.com> написал(а):\\n> \\n> What else might help?\\n\\nLet's add more tests that check survival of 2PC through crash recovery? We do now only one restart. Maybe it worth to do 4 or 8?\\n\\nBest regards, Andrey Borodin.\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 17:48:02 +0500\",\n \"msg_from\": \"Andrey Borodin <x4mmm@yandex-team.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"Andrey Borodin <x4mmm@yandex-team.ru> writes:\\n> Let's add more tests that check survival of 2PC through crash recovery? We do now only one restart. Maybe it worth to do 4 or 8?\\n\\nThat seems a little premature when we can't explain the failure\\nwe have. Also, buildfarm cycles aren't free.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 09:44:04 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 11:05:06PM -0800, Noah Misch wrote:\\n> On Wed, Nov 17, 2021 at 05:47:10PM -0500, Tom Lane wrote:\\n> > Noah Misch <noah@leadboat.com> writes:\\n> > > Each of the three failures happened on a sparc64 Debian+gcc machine. I had\\n> > > tried ~8000 iterations on thorntail, another sparc64 Debian+gcc animal,\\n> > > without reproducing this.\\n> \\n> > # 'pgbench: error: client 0 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/159EF88: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 5890048\\n> > [1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-17%2013%3A01%3A24\\n> \\n> Two others:\\n> ERROR: could not read two-phase state from WAL at 0/16F1850: invalid record length at 0/16F1850: wanted 24, got 0\\n> -- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-12%2013%3A01%3A15\\n> ERROR: could not read two-phase state from WAL at 0/1668020: incorrect resource manager data checksum in record at 0/1668020\\n> -- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kittiwake&dt=2021-11-16%2015%3A00%3A52\\n> \\n> > I suppose \\\"unexpected pageaddr 0/0\\\" is most easily explained by supposing\\n> > that XlogReadTwoPhaseData tried to read a WAL page that hadn't been\\n> > written out yet. Have we got any synchronization around that?\\n> \\n> If the WAL address isn't on disk, that error doesn't happen. Instead,\\n> read_local_xlog_page() blocks waiting for the WAL to become available. It's\\n> still possible that we make the WAL region exist, but it somehow doesn't\\n> contain the right data until shortly later. FinishPreparedTransaction() takes\\n> TwoPhaseStateLock and looks for an entry having gxact->valid. EndPrepare()\\n> fills gxact->prepare_end_lsn, then calls MarkAsPrepared() to set gxact->valid\\n> under TwoPhaseStateLock. All freelist (freeGXacts) interaction holds\\n> TwoPhaseStateLock. I'm not seeing a gap in that synchronization.\\n> \\n> I don't have a great theory, but here are candidates to examine next:\\n> \\n> - Run with wal_debug=on to confirm logged write location matches read location.\\n> - Run \\\"PGDATA=contrib/amcheck/tmp_check/t_003_cic_2pc_CIC_2PC_test_data/pgdata\\n> pg_waldump -s 0/01000000\\\" at the end of the test.\\n> - Dump WAL page binary image at the point of failure.\\n> - Log which branches in XLogReadRecord() are taken.\\n\\nTom Turelinckx, are you able to provide remote access to kittiwake or\\ntadarida? I'd use it to attempt the above things. All else being equal,\\nkittiwake is more relevant since it's still supported upstream.\\n\\nThe setup of your buildfarm animals provides a clue. I understand kittiwake\\nis the same as ibisbill except for build options, and tadarida is the same as\\nmussurana except for build options. ibisbill and mussurana haven't failed, so\\none of these is likely needed to reproduce:\\n\\n absence of --enable-cassert\\n CFLAGS='-g -O2 -fstack-protector -Wformat -Werror=format-security '\\n CPPFLAGS='-Wdate-time -D_FORTIFY_SOURCE=2'\\n LDFLAGS='-Wl,-z,relro -Wl,-z,now'\\n\\nHowever, I can't reproduce this on thorntail, even if I use tadarida's\\n./configure options, CFLAGS, CPPFLAGS, and LDFLAGS.\\n\\n\\n\",\n \"msg_date\": \"Fri, 19 Nov 2021 21:18:23 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"Noah Misch <noah@leadboat.com> writes:\\n> Tom Turelinckx, are you able to provide remote access to kittiwake or\\n> tadarida? I'd use it to attempt the above things. All else being equal,\\n> kittiwake is more relevant since it's still supported upstream.\\n\\nsnapper just exhibited the same failure, too:\\n\\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=snapper&dt=2021-11-18%2016%3A09%3A49\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 20 Nov 2021 00:22:28 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> snapper just exhibited the same failure, too:\\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=snapper&dt=2021-11-18%2016%3A09%3A49\\n\\nI grepped the buildfarm logs for all recent (last 3 months) occurrences of\\n'could not read two-phase state'. Here's the results:\\n\\n sysname | branch | snapshot | stage | l \\n-----------+---------------+---------------------+-------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------\\n kittiwake | REL9_6_STABLE | 2021-10-24 12:01:10 | pgbenchCheck | # 'client 1 aborted in state 3: ERROR: could not read two-phase state from xlog at 0/158F4E0\\n kittiwake | REL_13_STABLE | 2021-10-26 12:51:11 | ContribCheck-en_US.utf8 | # 'pgbench: error: client 3 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/168C8D8\\n kittiwake | REL_14_STABLE | 2021-11-08 15:42:35 | ContribCheck-en_US.utf8 | # 'pgbench: error: client 0 script 0 aborted in command 3 query 0: ERROR: could not read two-phase state from WAL at 0/17ABF48\\n kittiwake | REL_13_STABLE | 2021-11-16 15:00:52 | ContribCheck-en_US.utf8 | # 'pgbench: error: client 3 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/1668020: incorrect resource manager data checksum in record at 0/1668020\\n snapper | REL_14_STABLE | 2021-11-18 16:09:49 | contrib-amcheckCheck | # 'pgbench: error: client 3 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/1770328: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 7798784\\n tadarida | REL_11_STABLE | 2021-11-11 13:29:58 | pgbenchCheck | # 'client 3 aborted in command 3 (SQL) of script 0; ERROR: could not read two-phase state from WAL at 0/1716C68\\n tadarida | REL_10_STABLE | 2021-11-12 13:01:15 | pgbenchCheck | # 'client 4 aborted in command 3 of script 0; ERROR: could not read two-phase state from WAL at 0/16F1850: invalid record length at 0/16F1850: wanted 24, got 0\\n tadarida | HEAD | 2021-11-17 13:01:24 | contrib-amcheckCheck | # 'pgbench: error: client 0 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/159EF88: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 5890048\\n\\nSo not all are exactly 'unexpected pageaddr 0/0', but they do all\\nlook like we read garbage data.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 20 Nov 2021 00:55:16 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 19, 2021 at 09:18:23PM -0800, Noah Misch wrote:\\n> On Wed, Nov 17, 2021 at 11:05:06PM -0800, Noah Misch wrote:\\n> > On Wed, Nov 17, 2021 at 05:47:10PM -0500, Tom Lane wrote:\\n> > > Noah Misch <noah@leadboat.com> writes:\\n> > > > Each of the three failures happened on a sparc64 Debian+gcc machine. I had\\n> > > > tried ~8000 iterations on thorntail, another sparc64 Debian+gcc animal,\\n> > > > without reproducing this.\\n> > \\n> > > # 'pgbench: error: client 0 script 1 aborted in command 4 query 0: ERROR: could not read two-phase state from WAL at 0/159EF88: unexpected pageaddr 0/0 in log segment 000000010000000000000001, offset 5890048\\n> > > [1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-17%2013%3A01%3A24\\n> > \\n> > Two others:\\n> > ERROR: could not read two-phase state from WAL at 0/16F1850: invalid record length at 0/16F1850: wanted 24, got 0\\n> > -- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-12%2013%3A01%3A15\\n> > ERROR: could not read two-phase state from WAL at 0/1668020: incorrect resource manager data checksum in record at 0/1668020\\n> > -- https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kittiwake&dt=2021-11-16%2015%3A00%3A52\\n\\n> > I don't have a great theory, but here are candidates to examine next:\\n> > \\n> > - Run with wal_debug=on to confirm logged write location matches read location.\\n> > - Run \\\"PGDATA=contrib/amcheck/tmp_check/t_003_cic_2pc_CIC_2PC_test_data/pgdata\\n> > pg_waldump -s 0/01000000\\\" at the end of the test.\\n> > - Dump WAL page binary image at the point of failure.\\n> > - Log which branches in XLogReadRecord() are taken.\\n> \\n> Tom Turelinckx, are you able to provide remote access to kittiwake or\\n> tadarida? I'd use it to attempt the above things. All else being equal,\\n> kittiwake is more relevant since it's still supported upstream.\\n\\nThanks for setting up access. Summary: this kernel has a bug in I/O syscalls.\\nHow practical is it to update that kernel? (Userland differs across these\\nanimals, but the kernel does not.) The kernel on buildfarm member thorntail\\ndoesn't exhibit the bug.\\n\\nFor specifics of the kernel bug, see the attached test program. In brief, the\\nbug arises if one process is write()ing or pwrite()ing a file at about the\\nsame time that another process is read()ing or pread()ing the same. POSIX\\nsays the reader should see the data as it existed before the write or the\\nnewly-written data. On this kernel, the reader can see zeros instead. That\\nleads to the $SUBJECT failure. PostgreSQL processes write out a given WAL\\nblock 20-30 times in ~10ms, and COMMIT PREPARED reads that block. The writers\\naren't changing the bytes of interest to COMMIT PREPARED, but the zeros from\\nthe kernel bug yield the failure. We could opt to work around that by writing\\nonly the not-already-written portion of a WAL block, but I doubt that's\\nworthwhile unless it happens to be a performance win anyway.\\n\\nSeparately, while I don't know of relevance to PostgreSQL, I was interested to\\nsee that CentOS 7 pwrite()/pread() fail to have the POSIX-required atomicity.\\n\\n> The setup of your buildfarm animals provides a clue. I understand kittiwake\\n> is the same as ibisbill except for build options, and tadarida is the same as\\n> mussurana except for build options. ibisbill and mussurana haven't failed, so\\n> one of these is likely needed to reproduce:\\n> \\n> absence of --enable-cassert\\n> CFLAGS='-g -O2 -fstack-protector -Wformat -Werror=format-security '\\n> CPPFLAGS='-Wdate-time -D_FORTIFY_SOURCE=2'\\n> LDFLAGS='-Wl,-z,relro -Wl,-z,now'\\n\\nThat was a red herring. ibisbill and mussurana don't use --with-tap-tests.\\nAdding --with-tap-tests has been enough to make their configurations witness\\nthe same kinds of failures.\\n\\nnm\",\n \"msg_date\": \"Sat, 15 Jan 2022 23:12:10 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 16, 2022 at 8:12 PM Noah Misch <noah@leadboat.com> wrote:\\n> For specifics of the kernel bug, see the attached test program. In brief, the\\n> bug arises if one process is write()ing or pwrite()ing a file at about the\\n> same time that another process is read()ing or pread()ing the same. POSIX\\n> says the reader should see the data as it existed before the write or the\\n> newly-written data. On this kernel, the reader can see zeros instead. That\\n> leads to the $SUBJECT failure. PostgreSQL processes write out a given WAL\\n> block 20-30 times in ~10ms, and COMMIT PREPARED reads that block. The writers\\n> aren't changing the bytes of interest to COMMIT PREPARED, but the zeros from\\n> the kernel bug yield the failure. We could opt to work around that by writing\\n> only the not-already-written portion of a WAL block, but I doubt that's\\n> worthwhile unless it happens to be a performance win anyway.\\n>\\n> Separately, while I don't know of relevance to PostgreSQL, I was interested to\\n> see that CentOS 7 pwrite()/pread() fail to have the POSIX-required atomicity.\\n\\nFWIW there was some related discussion over here:\\n\\nhttps://www.postgresql.org/message-id/flat/17064-bb0d7904ef72add3%40postgresql.org\\n\\n\\n\",\n \"msg_date\": \"Sun, 16 Jan 2022 22:19:30 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"Cancel that kernel upgrade idea. I no longer expect it to help...\\n\\nOn Sun, Jan 16, 2022 at 10:19:30PM +1300, Thomas Munro wrote:\\n> On Sun, Jan 16, 2022 at 8:12 PM Noah Misch <noah@leadboat.com> wrote:\\n> > For specifics of the kernel bug, see the attached test program. In brief, the\\n> > bug arises if one process is write()ing or pwrite()ing a file at about the\\n> > same time that another process is read()ing or pread()ing the same. POSIX\\n> > says the reader should see the data as it existed before the write or the\\n> > newly-written data. On this kernel, the reader can see zeros instead. That\\n> > leads to the $SUBJECT failure. PostgreSQL processes write out a given WAL\\n> > block 20-30 times in ~10ms, and COMMIT PREPARED reads that block. The writers\\n> > aren't changing the bytes of interest to COMMIT PREPARED, but the zeros from\\n> > the kernel bug yield the failure.\\n\\nThe difference between kittiwake and thorntail comes from thorntail using xfs\\nand kittiwake using ext4. Running the io-rectitude.c tests on an ext4\\npartition on thorntail, I see the zeros bug just like I do on kittiwake. I\\ndon't see the zeros bug on ppc64 or x86_64, just sparc64 so far:\\n\\n * ext4, Linux 3.10.0-1160.49.1.el7.x86_64 (CentOS 7.9.2009):\\n * pwrite/pread is non-atomic if count>16 (no -D switches)\\n * write/read is atomic (-DUSE_SEEK -DXLOG_BLCKSZ=8192000)\\n * pwrite/pread is free from zeros bug (-DCHANGE_CONTENT=0)\\n * write/read is free from zeros bug (-DUSE_SEEK -DCHANGE_CONTENT=0)\\n *\\n * ext4, Linux 4.9.0-13-sparc64-smp (Debian):\\n * pwrite/pread is non-atomic if count>4 (no -D switches)\\n * write/read is non-atomic if count>4 (-DUSE_SEEK)\\n * write/read IS atomic w/o REOPEN (-DUSE_SEEK -DREOPEN=0 -DXLOG_BLCKSZ=8192000)\\n * pwrite/pread has zeros bug for count>127 (-DCHANGE_CONTENT=0)\\n * pwrite/pread w/ O_SYNC has zeros bug (-DCHANGE_CONTENT=0 -DOPEN_FLAGS=O_SYNC)\\n * far less frequent w/ O_SYNC, but it still happens\\n * pwrite/pread w/o REOPEN also has zeros bug for count>127 (-DCHANGE_CONTENT=0 -DREOPEN=0)\\n * write/read has zeros bug for count>127 (-DUSE_SEEK -DCHANGE_CONTENT=0)\\n * write/read w/ O_SYNC has zeros bug (-DUSE_SEEK -DCHANGE_CONTENT=0 -DOPEN_FLAGS=O_SYNC)\\n * write/read w/o REOPEN is free from zeros bug (-DUSE_SEEK -DCHANGE_CONTENT=0 -DREOPEN=0)\\n *\\n * ext4, Linux 5.15.0-2-sparc64-smp (Debian bookworm/sid):\\n * [behaviors match the previous kernel exactly]\\n *\\n * ext4, Linux 5.15.0-2-powerpc64 (Debian bookworm/sid):\\n * [atomicity matches previous kernel, but zeros bug does not]\\n * pwrite/pread is non-atomic if count>4 (no -D switches)\\n * write/read is non-atomic if count>4 (-DUSE_SEEK)\\n * write/read IS atomic w/o REOPEN (-DUSE_SEEK -DREOPEN=0 -DXLOG_BLCKSZ=8192000)\\n * pwrite/pread is free from zeros bug (-DCHANGE_CONTENT=0)\\n * write/read is free from zeros bug (-DUSE_SEEK -DCHANGE_CONTENT=0)\\n *\\n * ext4, Linux 5.15.5-0-virt x86_64 (Alpine):\\n * [behaviors match the previous kernel exactly]\\n *\\n * xfs, Linux 5.15.0-2-sparc64-smp (Debian bookworm/sid):\\n * pwrite/pread is atomic (-DXLOG_BLCKSZ=8192000)\\n * write/read is atomic (-DUSE_SEEK -DXLOG_BLCKSZ=8192000)\\n * pwrite/pread is free from zeros bug (-DCHANGE_CONTENT=0)\\n * write/read is free from zeros bug (-DUSE_SEEK -DCHANGE_CONTENT=0)\\n\\n> > We could opt to work around that by writing\\n> > only the not-already-written portion of a WAL block, but I doubt that's\\n> > worthwhile unless it happens to be a performance win anyway.\\n\\nMy next steps:\\n\\n- Report a Debian bug for the sparc64+ext4 zeros problem.\\n\\n- Try to falsify the idea that \\\"write only the not-already-written portion of\\n a WAL block\\\" is an effective workaround. Specifically, modify the test\\n program to have the writer process mutate offsets [N-k,N-1] and [N+1,N+k]\\n while the reader process reads offset N. If the reader sees a zero, that\\n workaround is ineffective.\\n\\n- Implement the workaround, if I didn't falsify its effectiveness. If it\\n doesn't hurt performance on x86_64, we can use it unconditionally.\\n Otherwise, limit its use to sparc64 Linux.\\n\\n> > Separately, while I don't know of relevance to PostgreSQL, I was interested to\\n> > see that CentOS 7 pwrite()/pread() fail to have the POSIX-required atomicity.\\n> \\n> FWIW there was some related discussion over here:\\n> \\n> https://www.postgresql.org/message-id/flat/17064-bb0d7904ef72add3%40postgresql.org\\n\\nThat gave me the idea to test different filesystems. Thanks. Incidentally, I\\nfind https://utcc.utoronto.ca/~cks/space/blog/unix/WriteNotVeryAtomic is\\nmistaken about POSIX requirements. There's no precedent for POSIX writing\\n\\\"two threads\\\" when it means \\\"two threads of the same process\\\". Moreover, the\\npart about \\\"shall also apply whenever a file descriptor is successfully\\nclosed, however caused (for example [...] process termination)\\\" would be\\nsuperfluous in a requirement specific to threads of one process. Having said\\nthat, if the most-prominent POSIX regular file implementation (ext4 on x86_64)\\ndoesn't implement a POSIX requirement, that has the same practical\\nconsequences for PostgreSQL as POSIX not requiring it.\\n\\nI now see newer Linux ext4 has drifted further away from POSIX atomicity,\\ncompared to CentOS 7. In CentOS 7 ext4, plain write()/read() was still\\natomic. By Linux 5.15.5, those abandoned atomicity.\\n\\n\\n\",\n \"msg_date\": \"Sun, 16 Jan 2022 13:02:41 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 17, 2022 at 10:02 AM Noah Misch <noah@leadboat.com> wrote:\\n> - Report a Debian bug for the sparc64+ext4 zeros problem.\\n\\nI suspect that 027_stream_regress.pl hits this kernel bug with high\\nprobability[1]. I wonder if the owner of kittiwake and tadarida would\\nconsider setting up an xfs file system? Or alternatively, since ext4\\ndidn't support concurrent writes until recently, I wonder if there is\\nan option somewhere to turn the new concurrency stuff off, or failing\\nthat, if we could temporarily downgrade the kernel to an older version\\nthat does inode-level read/write locking.\\n\\n[1] https://www.postgresql.org/message-id/CA%2BhUKG%2BeuZ%3Ddc27ZB%3Ds74x0q%3DzU%3D2%3Dvs8%2B6TkJoTUiCPUd2dQA%40mail.gmail.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 21 Jan 2022 08:34:22 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 21, 2022 at 08:34:22AM +1300, Thomas Munro wrote:\\n> On Mon, Jan 17, 2022 at 10:02 AM Noah Misch <noah@leadboat.com> wrote:\\n> > - Report a Debian bug for the sparc64+ext4 zeros problem.\\n> \\n> I suspect that 027_stream_regress.pl hits this kernel bug with high\\n> probability[1]. I wonder if the owner of kittiwake and tadarida would\\n> consider setting up an xfs file system? Or alternatively, since ext4\\n> didn't support concurrent writes until recently, I wonder if there is\\n> an option somewhere to turn the new concurrency stuff off, or failing\\n> that, if we could temporarily downgrade the kernel to an older version\\n> that does inode-level read/write locking.\\n\\nIf the write-only-new-bytes approach works, I think we'd want to revert those\\nchanges. Perhaps a cheaper stopgap is to make the affected tests skip on\\nsparc Linux. Is that worth doing? (Could even limit the skip to ext4,\\ne.g. by testing \\\"df -x ext4 . >/dev/null\\\".)\\n\\n> [1] https://www.postgresql.org/message-id/CA%2BhUKG%2BeuZ%3Ddc27ZB%3Ds74x0q%3DzU%3D2%3Dvs8%2B6TkJoTUiCPUd2dQA%40mail.gmail.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 20 Jan 2022 19:30:42 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 16, 2022 at 01:02:41PM -0800, Noah Misch wrote:\\n> My next steps:\\n> \\n> - Report a Debian bug for the sparc64+ext4 zeros problem.\\n\\n(Not done yet.)\\n\\n> - Try to falsify the idea that \\\"write only the not-already-written portion of\\n> a WAL block\\\" is an effective workaround. Specifically, modify the test\\n> program to have the writer process mutate offsets [N-k,N-1] and [N+1,N+k]\\n> while the reader process reads offset N. If the reader sees a zero, that\\n> workaround is ineffective.\\n\\nThe reader did not see a zero. In addition to bytes outside the write being\\nimmune to the zeros bug, the first and last forty bytes of a write were immune\\nto the zeros bug.\\n\\n> - Implement the workaround, if I didn't falsify its effectiveness. If it\\n> doesn't hurt performance on x86_64, we can use it unconditionally.\\n> Otherwise, limit its use to sparc64 Linux.\\n\\nAttached. With this, kittiwake has survived 8.5hr of 003_cic_2pc.pl. Without\\nthe patch, it failed many times, always within 1.3hr. For easier review, this\\npatch uses the new behavior on all platforms. Before commit and back-patch, I\\nplan to limit use of the new behavior to sparc Linux. Future work can\\nbenchmark the new behavior and, if it performs well, make it unconditional in\\nv15+. I would expect performance to be unchanged or slightly better, because\\nthe new behavior requests less futile work from the OS.\",\n \"msg_date\": \"Sat, 22 Jan 2022 10:52:41 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 23, 2022 at 7:52 AM Noah Misch <noah@leadboat.com> wrote:\\n> Attached. With this, kittiwake has survived 8.5hr of 003_cic_2pc.pl. Without\\n> the patch, it failed many times, always within 1.3hr. For easier review, this\\n> patch uses the new behavior on all platforms. Before commit and back-patch, I\\n> plan to limit use of the new behavior to sparc Linux. Future work can\\n> benchmark the new behavior and, if it performs well, make it unconditional in\\n> v15+. I would expect performance to be unchanged or slightly better, because\\n> the new behavior requests less futile work from the OS.\\n\\nOne detail is that wal_level=open_datasync, wal_senders=0,\\nwal_level=minimal will panic, because O_DIRECT requires fs\\npage-aligned access (and fails in various other ways on other OSes, eg\\nexpensive read-before-write every time). That's an ultra-niche\\nconcern likely affecting nobody, especially when multiplied by the\\nodds that anyone is using that stack at all (considering that\\nstreaming rep has apparently been borked for years on linux/sparc/ext4\\nand nobody told us).\\n\\nI was +1 for the control file locking change in that other thread (I\\nview the atomicity stuff as Linux-realpolitik-vs-POSIX, not to mention\\nthat we run Windows too, which requires separate analysis). I'm less\\nsure it makes sense to do anything to support the presumed bogus\\nzeroes bug for (probably) no real users, especially before we've even\\nreported it and heard some analysis, for example acceptance that it's\\nbroken and confirmation that this really is just a sparc problem.\\n\\n\\n\",\n \"msg_date\": \"Mon, 24 Jan 2022 09:42:13 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-01-24 09:42:13 +1300, Thomas Munro wrote:\\n> On Sun, Jan 23, 2022 at 7:52 AM Noah Misch <noah@leadboat.com> wrote:\\n> > Future work can benchmark the new behavior and, if it performs well, make\\n> > it unconditional in v15+. I would expect performance to be unchanged or\\n> > slightly better, because the new behavior requests less futile work from\\n> > the OS.\\n\\nI doubt it'll be generally applicable. Turning a write operation into a\\nread-write isn't free. Yes, often enough it's likely that the prior page is\\nstill in cache, but I don't think we can rely on that in general.\\n\\nIt also just fundamentally locks us into never using O_DIRECT in anger. I\\ndon't think that's a good direction.\\n\\n\\n> One detail is that wal_level=open_datasync, wal_senders=0,\\n> wal_level=minimal will panic, because O_DIRECT requires fs\\n> page-aligned access (and fails in various other ways on other OSes, eg\\n> expensive read-before-write every time). That's an ultra-niche\\n> concern likely affecting nobody, especially when multiplied by the\\n> odds that anyone is using that stack at all (considering that\\n> streaming rep has apparently been borked for years on linux/sparc/ext4\\n> and nobody told us).\\n\\nSeems like the patch should at least make this error out?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sun, 23 Jan 2022 14:26:53 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 24, 2022 at 09:42:13AM +1300, Thomas Munro wrote:\\n> I'm less\\n> sure it makes sense to do anything to support the presumed bogus\\n> zeroes bug for (probably) no real users, especially before we've even\\n> reported it and heard some analysis, for example acceptance that it's\\n> broken and confirmation that this really is just a sparc problem.\\n\\nGot it. I've already done a bad thing leaving the buildfarm broken for three\\nmonths, so I don't want to let the buildfarm wait for a kernel fix. These are\\nthe two main options I'm seeing now:\\n\\n(a) Modify the tests so the affected animals can skip affected tests by\\nsetting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\\n\\n(b) Remove --enable-tap-tests from affected animals.\\n\\nDo you have a preference among those two or some other option that gets the\\nbuildfarm green on a predictable schedule? I somewhat prefer (a), since\\n--enable-tap-tests is where most of the interesting buildfarm reports happen\\nthese days.\\n\\n\\n\",\n \"msg_date\": \"Sun, 23 Jan 2022 15:29:27 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 24, 2022 at 12:29 PM Noah Misch <noah@leadboat.com> wrote:\\n> On Mon, Jan 24, 2022 at 09:42:13AM +1300, Thomas Munro wrote:\\n> > I'm less\\n> > sure it makes sense to do anything to support the presumed bogus\\n> > zeroes bug for (probably) no real users, especially before we've even\\n> > reported it and heard some analysis, for example acceptance that it's\\n> > broken and confirmation that this really is just a sparc problem.\\n>\\n> Got it. I've already done a bad thing leaving the buildfarm broken for three\\n> months, so I don't want to let the buildfarm wait for a kernel fix. These are\\n> the two main options I'm seeing now:\\n>\\n> (a) Modify the tests so the affected animals can skip affected tests by\\n> setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\\n>\\n> (b) Remove --enable-tap-tests from affected animals.\\n>\\n> Do you have a preference among those two or some other option that gets the\\n> buildfarm green on a predictable schedule? I somewhat prefer (a), since\\n> --enable-tap-tests is where most of the interesting buildfarm reports happen\\n> these days.\\n\\nTrying out a new idea: what if we could tell the buildfarm website\\nthat a certain test is currently expected to fail for reasons we can't\\nfix yet (configuration change needed but owner not responding, or\\nbugfix from another project needed, etc)? That could cause it to be\\ndisplayed in a different shade of green, or grey, or whatever? Other\\nkinds of failures would still show as red. Perhaps this would be\\nconfigured with a file in a git repo that any committer can push to.\\n\\n\\n\",\n \"msg_date\": \"Mon, 24 Jan 2022 12:49:16 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 24, 2022 at 12:49:16PM +1300, Thomas Munro wrote:\\n> On Mon, Jan 24, 2022 at 12:29 PM Noah Misch <noah@leadboat.com> wrote:\\n> > On Mon, Jan 24, 2022 at 09:42:13AM +1300, Thomas Munro wrote:\\n> > > I'm less\\n> > > sure it makes sense to do anything to support the presumed bogus\\n> > > zeroes bug for (probably) no real users, especially before we've even\\n> > > reported it and heard some analysis, for example acceptance that it's\\n> > > broken and confirmation that this really is just a sparc problem.\\n> >\\n> > Got it. I've already done a bad thing leaving the buildfarm broken for three\\n> > months, so I don't want to let the buildfarm wait for a kernel fix. These are\\n> > the two main options I'm seeing now:\\n> >\\n> > (a) Modify the tests so the affected animals can skip affected tests by\\n> > setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\\n> >\\n> > (b) Remove --enable-tap-tests from affected animals.\\n> >\\n> > Do you have a preference among those two or some other option that gets the\\n> > buildfarm green on a predictable schedule? I somewhat prefer (a), since\\n> > --enable-tap-tests is where most of the interesting buildfarm reports happen\\n> > these days.\\n> \\n> Trying out a new idea: what if we could tell the buildfarm website\\n> that a certain test is currently expected to fail for reasons we can't\\n> fix yet (configuration change needed but owner not responding, or\\n> bugfix from another project needed, etc)? That could cause it to be\\n> displayed in a different shade of green, or grey, or whatever? Other\\n> kinds of failures would still show as red. Perhaps this would be\\n> configured with a file in a git repo that any committer can push to.\\n\\nThat would be a better capability to use if we had it, agreed. Is it feasible\\nto acquire that capability soon enough?\\n\\n\\n\",\n \"msg_date\": \"Sun, 23 Jan 2022 16:19:14 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"Noah Misch <noah@leadboat.com> writes:\\n> On Mon, Jan 24, 2022 at 12:49:16PM +1300, Thomas Munro wrote:\\n>> Trying out a new idea: what if we could tell the buildfarm website\\n>> that a certain test is currently expected to fail for reasons we can't\\n>> fix yet (configuration change needed but owner not responding, or\\n>> bugfix from another project needed, etc)? That could cause it to be\\n>> displayed in a different shade of green, or grey, or whatever? Other\\n>> kinds of failures would still show as red. Perhaps this would be\\n>> configured with a file in a git repo that any committer can push to.\\n\\n> That would be a better capability to use if we had it, agreed. Is it feasible\\n> to acquire that capability soon enough?\\n\\nIt's not merely a website issue: you'd really rather that the\\nbuildfarm animal runs the rest of the tests rather than going belly-up\\nafter an expected failure. I think your suggestion about skipping\\nproblematic tests based on an environment variable is more practical\\nin the near term. We already have some cases like that, too, eg in\\nsrc/bin/psql/t/010_tab_completion.pl.\\n\\nIn the long term I could get behind having some less ad-hoc way\\nof skipping tests, but I don't think we can have that quickly.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 23 Jan 2022 19:28:47 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"\\n\\nOn January 23, 2022 3:29:27 PM PST\\n>(a) Modify the tests so the affected animals can skip affected tests by\\n>setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\\n\\nWhy not just detect the problem in the tap test and skip, rather than requiring multiple buildfarm configs to be changed as well as the test itself? \\n-- \\nSent from my Android device with K-9 Mail. Please excuse my brevity.\\n\\n\\n\",\n \"msg_date\": \"Sun, 23 Jan 2022 17:03:04 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 23, 2022 at 05:03:04PM -0800, Andres Freund wrote:\\n> On January 23, 2022 3:29:27 PM PST\\n> >(a) Modify the tests so the affected animals can skip affected tests by\\n> >setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\\n> \\n> Why not just detect the problem in the tap test and skip, rather than requiring multiple buildfarm configs to be changed as well as the test itself? \\n\\nEnd users running PostgreSQL test suites to acceptance-test their stack should\\nconsider the affected stack unusable for PostgreSQL. Hence, I ruled out that\\napproach, despite having implemented it at one point. Under some plausible\\nset of goals, it is optimal.\\n\\n\\n\",\n \"msg_date\": \"Sun, 23 Jan 2022 17:17:59 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-01-23 17:17:59 -0800, Noah Misch wrote:\\n> On Sun, Jan 23, 2022 at 05:03:04PM -0800, Andres Freund wrote:\\n> > On January 23, 2022 3:29:27 PM PST\\n> > >(a) Modify the tests so the affected animals can skip affected tests by\\n> > >setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\\n> > \\n> > Why not just detect the problem in the tap test and skip, rather than requiring multiple buildfarm configs to be changed as well as the test itself? \\n> \\n> End users running PostgreSQL test suites to acceptance-test their stack should\\n> consider the affected stack unusable for PostgreSQL.\\n\\nI'd bet that that's zero users ;)\\n\\n\\n> Hence, I ruled out that\\n> approach, despite having implemented it at one point. Under some plausible\\n> set of goals, it is optimal.\\n\\nIt's not perfect due to the way we run our tests (seeing output is hard, it's\\nnot aggregated), but marking the test as todo rather than SKIP seems like the\\nmost appropriate test status. It's known to be a problem, we've not fixed it,\\nbut we want to be able to run the tests.\\n\\nTest::more's description: \\\"If it's something the programmer hasn't done yet,\\nuse TODO. This is for any code you haven't written yet, or bugs you have yet\\nto fix, but want to put tests in your testing script (always a good idea).\\\"\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sun, 23 Jan 2022 17:40:54 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 23, 2022 at 05:40:54PM -0800, Andres Freund wrote:\\n> On 2022-01-23 17:17:59 -0800, Noah Misch wrote:\\n> > On Sun, Jan 23, 2022 at 05:03:04PM -0800, Andres Freund wrote:\\n> > > On January 23, 2022 3:29:27 PM PST\\n> > > >(a) Modify the tests so the affected animals can skip affected tests by\\n> > > >setting an environment variable, named PG_TEST_HAS_WAL_READ_BUG or similar.\\n> > > \\n> > > Why not just detect the problem in the tap test and skip, rather than requiring multiple buildfarm configs to be changed as well as the test itself? \\n> > \\n> > End users running PostgreSQL test suites to acceptance-test their stack should\\n> > consider the affected stack unusable for PostgreSQL.\\n> \\n> I'd bet that that's zero users ;)\\n\\nWouldn't surprise me. I'm attaching what I had written and discarded. If\\nnobody else hates it, I can live with it.\\n\\n> > Hence, I ruled out that\\n> > approach, despite having implemented it at one point. Under some plausible\\n> > set of goals, it is optimal.\\n> \\n> It's not perfect due to the way we run our tests (seeing output is hard, it's\\n> not aggregated), but marking the test as todo rather than SKIP seems like the\\n> most appropriate test status. It's known to be a problem, we've not fixed it,\\n> but we want to be able to run the tests.\\n> \\n> Test::more's description: \\\"If it's something the programmer hasn't done yet,\\n> use TODO. This is for any code you haven't written yet, or bugs you have yet\\n> to fix, but want to put tests in your testing script (always a good idea).\\\"\\n\\nCould do that. Every run that doesn't get the flaky failure will print a\\nmessage like \\\"TODO passed: 3-5\\\", though the test file could mitigate that by\\ndeclaring the TODO only on configurations where we expect a failure. The\\n027_stream_regress.pl trouble involves reaching a die(), not failing a test in\\nthis sense, so that one would take more work.\",\n \"msg_date\": \"Sun, 23 Jan 2022 18:10:07 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 23, 2022 at 06:10:07PM -0800, Noah Misch wrote:\\n> Could do that. Every run that doesn't get the flaky failure will print a\\n> message like \\\"TODO passed: 3-5\\\", though the test file could mitigate that by\\n> declaring the TODO only on configurations where we expect a failure. The\\n> 027_stream_regress.pl trouble involves reaching a die(), not failing a test in\\n> this sense, so that one would take more work.\\n\\nUsing a TODO has the advantage to allow the tests to run on a periodic\\nbasis, even if they could fail in this unexpected way. We've had our\\nload of issues in the past proper to specific architectures, so this\\ncould help with the 2PC code paths and SPARC. Not to mention that we\\ncould just let the code be as-is, and it would fix itself once the\\nkernel is updated. So that sounds much better to me than a skip\\nphase.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 24 Jan 2022 11:20:21 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Sun, Jan 23, 2022 at 06:10:07PM -0800, Noah Misch wrote:\\n>> Could do that. Every run that doesn't get the flaky failure will print a\\n>> message like \\\"TODO passed: 3-5\\\", though the test file could mitigate that by\\n>> declaring the TODO only on configurations where we expect a failure. The\\n>> 027_stream_regress.pl trouble involves reaching a die(), not failing a test in\\n>> this sense, so that one would take more work.\\n\\n> Using a TODO has the advantage to allow the tests to run on a periodic\\n> basis, even if they could fail in this unexpected way.\\n\\nI'm okay with this *if* the TODO marking can be constrained to platforms\\nwhere we know there's a problem. Otherwise I'm afraid it will mask\\nunexpected problems.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 23 Jan 2022 21:25:04 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-01-23 18:10:07 -0800, Noah Misch wrote:\\n> On Sun, Jan 23, 2022 at 05:40:54PM -0800, Andres Freund wrote:\\n> > Test::more's description: \\\"If it's something the programmer hasn't done yet,\\n> > use TODO. This is for any code you haven't written yet, or bugs you have yet\\n> > to fix, but want to put tests in your testing script (always a good idea).\\\"\\n>\\n> Could do that. Every run that doesn't get the flaky failure will print a\\n> message like \\\"TODO passed: 3-5\\\", though the test file could mitigate that by\\n> declaring the TODO only on configurations where we expect a failure.\\n\\nYea, that's what I was thinking we'd do.\\n\\n\\n> The\\n> 027_stream_regress.pl trouble involves reaching a die(), not failing a test in\\n> this sense, so that one would take more work.\\n\\nI'm really not a perl person... But my understanding is that todo_skip() would\\naddress this? I.e. something like\\n\\nTODO:\\n{\\n $todo_skip \\\"linux/sparc has unaddressed problems with partial page overwrites\\\"\\n if ($^O eq 'linux' and $Config{archname'} ~= 'sparc');\\n\\n ok(whatever is broken);\\n}\\n\\n(no idea if the above todo condition even approximates something working)\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sun, 23 Jan 2022 18:34:32 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On 2022-01-23 21:25:04 -0500, Tom Lane wrote:\\n> Michael Paquier <michael@paquier.xyz> writes:\\n> > On Sun, Jan 23, 2022 at 06:10:07PM -0800, Noah Misch wrote:\\n> >> Could do that. Every run that doesn't get the flaky failure will print a\\n> >> message like \\\"TODO passed: 3-5\\\", though the test file could mitigate that by\\n> >> declaring the TODO only on configurations where we expect a failure. The\\n> >> 027_stream_regress.pl trouble involves reaching a die(), not failing a test in\\n> >> this sense, so that one would take more work.\\n> \\n> > Using a TODO has the advantage to allow the tests to run on a periodic\\n> > basis, even if they could fail in this unexpected way.\\n> \\n> I'm okay with this *if* the TODO marking can be constrained to platforms\\n> where we know there's a problem. Otherwise I'm afraid it will mask\\n> unexpected problems.\\n\\nYep. Very weird syntax... TODO blocks only take effect if a local $TODO is set\\n(to a text explaining the reason hopefully) or if there's a todo_skip. So\\nsomething like\\n\\nTODO:\\n{\\n local $TODO = 'linux + sparc is borked' if $^O eq 'linux' and ...;\\n\\n tests...\\n}\\n\\nwould run the tests and accept failures if the if matches, and otherwise not.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sun, 23 Jan 2022 18:38:34 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 23, 2022 at 06:34:32PM -0800, Andres Freund wrote:\\n> On 2022-01-23 18:10:07 -0800, Noah Misch wrote:\\n> > On Sun, Jan 23, 2022 at 05:40:54PM -0800, Andres Freund wrote:\\n> > > Test::more's description: \\\"If it's something the programmer hasn't done yet,\\n> > > use TODO. This is for any code you haven't written yet, or bugs you have yet\\n> > > to fix, but want to put tests in your testing script (always a good idea).\\\"\\n> >\\n> > Could do that. Every run that doesn't get the flaky failure will print a\\n> > message like \\\"TODO passed: 3-5\\\", though the test file could mitigate that by\\n> > declaring the TODO only on configurations where we expect a failure.\\n> \\n> Yea, that's what I was thinking we'd do.\\n> \\n> > The\\n> > 027_stream_regress.pl trouble involves reaching a die(), not failing a test in\\n> > this sense, so that one would take more work.\\n> \\n> I'm really not a perl person... But my understanding is that todo_skip() would\\n> address this? I.e. something like\\n> \\n> TODO:\\n> {\\n> $todo_skip \\\"linux/sparc has unaddressed problems with partial page overwrites\\\"\\n> if ($^O eq 'linux' and $Config{archname'} ~= 'sparc');\\n> \\n> ok(whatever is broken);\\n> }\\n\\nYes. todo_skip() behaves much like regular skip(). The enclosed tests don't\\nrun. Hence, it prevents die() and BAIL_OUT() failures. $TODO is a different\\nbeast; tests still run, and it changes the reporting. For 003_cic_2pc.pl, I'm\\nfine using $TODO so we continue to run all test commands and quietly log their\\nresults. For 027_stream_regress.pl, which would need deep changes to use\\n$TODO, it works to use any of todo_skip, skip, or skip_all. I prefer\\nskip_all, because it prints the skip reason to gmake's stdout. (If the number\\nof affected users is zero as theorized, the choice doesn't matter.) Any\\nobjections? Here's the appearance of each strategy on gmake's stdout:\\n\\n=== todo_skip\\n[23:56:45] t/003_cic_2pc.pl .. ok 60 ms ( 0.00 usr 0.00 sys + 0.05 cusr 0.00 csys = 0.05 CPU)\\n[23:56:45]\\nAll tests successful.\\nFiles=1, Tests=5, 1 wallclock secs ( 0.01 usr 0.00 sys + 0.05 cusr 0.00 csys = 0.06 CPU)\\nResult: PASS\\n\\n=== skip\\n[23:55:47] t/003_cic_2pc.pl .. ok 59 ms ( 0.00 usr 0.00 sys + 0.05 cusr 0.00 csys = 0.05 CPU)\\n[23:55:48]\\nAll tests successful.\\nFiles=1, Tests=5, 1 wallclock secs ( 0.01 usr 0.00 sys + 0.05 cusr 0.00 csys = 0.06 CPU)\\nResult: PASS\\n\\n=== skip_all\\n[23:31:04] t/003_cic_2pc.pl .. skipped: filesystem bug\\n[23:31:04]\\nFiles=1, Tests=0, 0 wallclock secs ( 0.02 usr 0.00 sys + 0.05 cusr 0.00 csys = 0.07 CPU)\\nResult: NOTESTS\\n\\n=== $TODO, test 1 is expected fail, tests 2-5 are unexpected pass\\n[23:32:32] t/003_cic_2pc.pl .. ok 1371 ms ( 0.01 usr 0.00 sys + 0.51 cusr 0.24 csys = 0.76 CPU)\\n[23:32:33]\\nAll tests successful.\\n\\nTest Summary Report\\n-------------------\\nt/003_cic_2pc.pl (Wstat: 0 Tests: 5 Failed: 0)\\n TODO passed: 2-5\\nFiles=1, Tests=5, 1 wallclock secs ( 0.02 usr 0.00 sys + 0.51 cusr 0.24 csys = 0.77 CPU)\\nResult: PASS\\n\\n\\n\",\n \"msg_date\": \"Mon, 24 Jan 2022 00:02:43 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 24, 2022 at 12:02:43AM -0800, Noah Misch wrote:\\n> For 003_cic_2pc.pl, I'm\\n> fine using $TODO so we continue to run all test commands and quietly log their\\n> results. For 027_stream_regress.pl, which would need deep changes to use\\n> $TODO, it works to use any of todo_skip, skip, or skip_all. I prefer\\n> skip_all, because it prints the skip reason to gmake's stdout. (If the number\\n> of affected users is zero as theorized, the choice doesn't matter.) Any\\n> objections?\\n\\nI pushed that on 2022-01-26 as ce6d793, but it wasn't enough:\\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=snapper&dt=2022-02-01%2013%3A29%3A58\\n\\nIn older Test::More versions, $TODO only works if tests know the right package\\nin which to examine $that_pkg::TODO. In PostgreSQL v12+, all relevant\\nfunctions set $Test::Builder::Level, so Test::More finds $main::TODO as\\nintended. This would be one way to fix v11:\\n\\n--- a/src/test/perl/PostgresNode.pm\\n+++ b/src/test/perl/PostgresNode.pm\\n@@ -1869,4 +1869,6 @@ command_ok(...)\\n sub command_checks_all\\n {\\n+\\tlocal $Test::Builder::Level = $Test::Builder::Level + 1;\\n+\\n \\tmy $self = shift;\\n \\ndiff --git a/src/test/perl/TestLib.pm b/src/test/perl/TestLib.pm\\nindex 0c56156..2ab78b4 100644\\n--- a/src/test/perl/TestLib.pm\\n+++ b/src/test/perl/TestLib.pm\\n@@ -598,4 +598,6 @@ sub command_fails_like\\n sub command_checks_all\\n {\\n+\\tlocal $Test::Builder::Level = $Test::Builder::Level + 1;\\n+\\n \\tmy ($cmd, $expected_ret, $out, $err, $test_name) = @_;\\n\\nHowever, I'm inclined to use a less-fragile way, in all branches:\\n\\n--- a/src/bin/pgbench/t/023_cic_2pc.pl\\n+++ b/src/bin/pgbench/t/023_cic_2pc.pl\\n@@ -14 +14,2 @@ use Test::More tests => 6;\\n-local $TODO = 'filesystem bug' if TestLib::has_wal_read_bug;\\n+Test::More->builder->todo_start('filesystem bug')\\n+ if TestLib::has_wal_read_bug;\\n\\nNone of this matters under newer Test::More. kittiwake's Test::More 1.302133\\nis new enough not to care, but tadarida's Test::More 1.001014 is not. The\\nTest::More->builder->todo_start() syntax works in 0.82, if not earlier.\\n(PostgreSQL v10 requires Test::More 0.87.)\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Feb 2022 21:55:56 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 16, 2022 at 01:02:41PM -0800, Noah Misch wrote:\\n> My next steps:\\n> \\n> - Report a Debian bug for the sparc64+ext4 zeros problem.\\n\\nReported to Debian, then upstream:\\nhttps://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1006157\\nhttps://marc.info/?t=164539269900001\\n\\nLast week, someone confirmed the behavior on kernel 5.17.0-rc5. No other news.\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Mar 2022 17:41:39 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: XLogReadRecord() error in XlogReadTwoPhaseData()\"\n }\n]","isConversational":false}}},{"rowIdx":53052,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Ok, subject was a bit philosophical but this message I just found is\nquite confusing.\n\n\"\"\"\nregression=# select cast(null as anyrange) &> cast(null as anymultirange);\nERROR: argument declared anymultirange is not a multirange type but type anymultirange\n\"\"\"\n\n-- \nJaime Casanova\nDirector de Servicios Profesionales\nSystemGuards - Consultores de PostgreSQL\n\n\n","msg_date":"Sun, 7 Nov 2021 01:17:57 -0500","msg_from":"Jaime Casanova <jcasanov@systemguards.com.ec>","msg_from_op":true,"msg_subject":"to be a multirange or not be, that's the question"},{"msg_contents":"On Saturday, November 6, 2021, Jaime Casanova <jcasanov@systemguards.com.ec>\nwrote:\n\n> Ok, subject was a bit philosophical but this message I just found is\n> quite confusing.\n>\n> \"\"\"\n> regression=# select cast(null as anyrange) &> cast(null as anymultirange);\n> ERROR: argument declared anymultirange is not a multirange type but type\n> anymultirange\n> \"\"\"\n>\n>\nI get that it is hard to parse but it is unambiguously correct. It does\nseem to presume that one understands how polymorphic pseudo-types work (the\nsupplied query was written without that knowledge applied). I can envision\nsome improvements here though it would depend a lot on the actual\nimplementation. I’m also curious as to why we get as far the operator\ninvocation when casting literals to polymorphic pseudo-types is supposedly\nan invalid thing to do.\n\nDavid J.\n\nOn Saturday, November 6, 2021, Jaime Casanova <jcasanov@systemguards.com.ec> wrote:Ok, subject was a bit philosophical but this message I just found is\nquite confusing.\n\n\"\"\"\nregression=# select cast(null as anyrange) &> cast(null as anymultirange);\nERROR: argument declared anymultirange is not a multirange type but type anymultirange\n\"\"\"\nI get that it is hard to parse but it is unambiguously correct. It does seem to presume that one understands how polymorphic pseudo-types work (the supplied query was written without that knowledge applied). I can envision some improvements here though it would depend a lot on the actual implementation. I’m also curious as to why we get as far the operator invocation when casting literals to polymorphic pseudo-types is supposedly an invalid thing to do.David J.","msg_date":"Sat, 6 Nov 2021 23:54:00 -0700","msg_from":"\"David G. Johnston\" <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: to be a multirange or not be, that's the question"}],"string":"[\n {\n \"msg_contents\": \"Ok, subject was a bit philosophical but this message I just found is\\nquite confusing.\\n\\n\\\"\\\"\\\"\\nregression=# select cast(null as anyrange) &> cast(null as anymultirange);\\nERROR: argument declared anymultirange is not a multirange type but type anymultirange\\n\\\"\\\"\\\"\\n\\n-- \\nJaime Casanova\\nDirector de Servicios Profesionales\\nSystemGuards - Consultores de PostgreSQL\\n\\n\\n\",\n \"msg_date\": \"Sun, 7 Nov 2021 01:17:57 -0500\",\n \"msg_from\": \"Jaime Casanova <jcasanov@systemguards.com.ec>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"to be a multirange or not be, that's the question\"\n },\n {\n \"msg_contents\": \"On Saturday, November 6, 2021, Jaime Casanova <jcasanov@systemguards.com.ec>\\nwrote:\\n\\n> Ok, subject was a bit philosophical but this message I just found is\\n> quite confusing.\\n>\\n> \\\"\\\"\\\"\\n> regression=# select cast(null as anyrange) &> cast(null as anymultirange);\\n> ERROR: argument declared anymultirange is not a multirange type but type\\n> anymultirange\\n> \\\"\\\"\\\"\\n>\\n>\\nI get that it is hard to parse but it is unambiguously correct. It does\\nseem to presume that one understands how polymorphic pseudo-types work (the\\nsupplied query was written without that knowledge applied). I can envision\\nsome improvements here though it would depend a lot on the actual\\nimplementation. I’m also curious as to why we get as far the operator\\ninvocation when casting literals to polymorphic pseudo-types is supposedly\\nan invalid thing to do.\\n\\nDavid J.\\n\\nOn Saturday, November 6, 2021, Jaime Casanova <jcasanov@systemguards.com.ec> wrote:Ok, subject was a bit philosophical but this message I just found is\\nquite confusing.\\n\\n\\\"\\\"\\\"\\nregression=# select cast(null as anyrange) &> cast(null as anymultirange);\\nERROR: argument declared anymultirange is not a multirange type but type anymultirange\\n\\\"\\\"\\\"\\nI get that it is hard to parse but it is unambiguously correct. It does seem to presume that one understands how polymorphic pseudo-types work (the supplied query was written without that knowledge applied). I can envision some improvements here though it would depend a lot on the actual implementation. I’m also curious as to why we get as far the operator invocation when casting literals to polymorphic pseudo-types is supposedly an invalid thing to do.David J.\",\n \"msg_date\": \"Sat, 6 Nov 2021 23:54:00 -0700\",\n \"msg_from\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: to be a multirange or not be, that's the question\"\n }\n]","isConversational":false}}},{"rowIdx":53053,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI have a question that is a specification of permission check\n(visibilityrule) for psql meta-command with schema option.\n\nAccording to the source code [1], there is no check if a schema\noption is added. As a result, a role that is not granted can see\nother roles' object names.\nWe might say it's okay because it's a name, not contents (data),\nbut It seems not preferable, I think.\n\nThe following is a reproducer using \\dX commands.\nNote: It is not only \\dX but also \\d because it uses the same\npermission check function (processSQLNamePattern).\n\nThe reproduction procedure (including some results):\n================================================\n-- Create role a, b as non-superuser\ncreate role a nosuperuser;\ncreate role b nosuperuser;\ngrant CREATE on database postgres to a;\n\n-- Create schema s_a, table hoge, and its extend stats by role a\nset role a;\ncreate schema s_a;\ncreate table s_a.hoge(a int, b int);\ncreate statistics s_a.hoge_ext on a,b from s_a.hoge;\nset search_path to public, s_a;\n\n-- Run \\dX and \\dX s_a.* by role a: OK (since schema s_a was created by role a)\n\\dX\n List of extended statistics\n Schema | Name | Definition | Ndistinct | Dependencies | MCV\n--------+----------+----------------+-----------+--------------+---------\n s_a | hoge_ext | a, b FROM hoge | defined | defined | defined\n(1 row)\n\n\\dX s_a.*\n List of extended statistics\n Schema | Name | Definition | Ndistinct | Dependencies | MCV\n--------+----------+----------------+-----------+--------------+---------\n s_a | hoge_ext | a, b FROM hoge | defined | defined | defined\n(1 row)\n\n-- Run \\dX by role b: OK\n-- (not displayed is fine since role b can't see info of role a)\nreset role;\nset role b;\n\\dX\n List of extended statistics\n Schema | Name | Definition | Ndistinct | Dependencies | MCV\n--------+------+------------+-----------+--------------+-----\n(0 rows)\n\n-- Run \\dX with schema by role b: OK?? (It should be NG?)\n-- this case is a point in my question\n\\dX s_a.*\n List of extended statistics\n Schema | Name | Definition | Ndistinct | Dependencies | MCV\n--------+----------+--------------------+-----------+--------------+---------\n s_a | hoge_ext | a, b FROM s_a.hoge | defined | defined | defined\n(1 row)\n\n-- clean-up\nreset role;\ndrop schema s_a cascade;\nrevoke CREATE on DATABASE postgres FROM a;\ndrop role a;\ndrop role b;\n================================================\n\n From the above results, I expected \"\\dX s_a.*\" doesn't show any info\nas same as \"\\dX\". but info is displayed. I'm wondering this behavior.\n\nI'm maybe missing something, but if this is a problem, I'll send a\npatch. Any comments are welcome!\n\n\n[1]: processSQLNamePattern in src/fe_utils/string_utils.c\n if (schemabuf.len > 2)\n {\n /* We have a schema pattern, so constrain the schemavar */\n\n /* Optimize away a \"*\" pattern */\n if (strcmp(schemabuf.data, \"^(.*)$\") != 0 && schemavar)\n {\n WHEREAND();\n appendPQExpBuffer(buf, \"%s OPERATOR(pg_catalog.~) \", schemavar);\n appendStringLiteralConn(buf, schemabuf.data, conn);\n if (PQserverVersion(conn) >= 120000)\n appendPQExpBufferStr(buf, \" COLLATE pg_catalog.default\");\n appendPQExpBufferChar(buf, '\\n');\n }\n }\n else\n {\n /* No schema pattern given, so select only visible objects */\n if (visibilityrule)\n {\n WHEREAND();\n appendPQExpBuffer(buf, \"%s\\n\", visibilityrule);\n }\n }\n\n\n\nThanks,\nTatsuro Yamada\n\n\n\n\n","msg_date":"Mon, 08 Nov 2021 11:43:34 +0900","msg_from":"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>","msg_from_op":true,"msg_subject":"Question about psql meta-command with schema option doesn't use\n visibilityrule"},{"msg_contents":"On Sunday, November 7, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\nwrote:\n\n>\n> According to the source code [1], there is no check if a schema\n> option is added. As a result, a role that is not granted can see\n> other roles' object names.\n> We might say it's okay because it's a name, not contents (data),\n> but It seems not preferable, I think.\n>\n\nNo, we are not interested in changing this long-standing documented\nbehavior. The contents of the catalogs are visible to all. So even if\nthis was something to consider, psql is not the correct scope.\n\nDavid J.\n\nOn Sunday, November 7, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp> wrote:\nAccording to the source code [1], there is no check if a schema\noption is added. As a result, a role that is not granted can see\nother roles' object names.\nWe might say it's okay because it's a name, not contents (data),\nbut It seems not preferable, I think.\nNo, we are not interested in changing this long-standing documented behavior. The contents of the catalogs are visible to all. So even if this was something to consider, psql is not the correct scope.David J.","msg_date":"Sun, 7 Nov 2021 20:09:44 -0700","msg_from":"\"David G. Johnston\" <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: Question about psql meta-command with schema option doesn't use\n visibilityrule"},{"msg_contents":"On Sunday, November 7, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\nwrote:\n\n>\n> I have a question that is a specification of permission check\n> (visibilityrule) for psql meta-command with schema option.\n>\n> From the above results, I expected \"\\dX s_a.*\" doesn't show any info\n> as same as \"\\dX\". but info is displayed. I'm wondering this behavior.\n>\n> I'm maybe missing something, but if this is a problem, I'll send a\n> patch. Any comments are welcome!\n>\n>\nVisibility means search_path, not permission. If s_a is not in the\nsearch_paths it objects are not visible unqualified but can be seen\n(catalog) when schema qualified.\n\nDavid J.\n\nOn Sunday, November 7, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp> wrote:\nI have a question that is a specification of permission check\n(visibilityrule) for psql meta-command with schema option.\nFrom the above results, I expected \"\\dX s_a.*\" doesn't show any info\nas same as \"\\dX\". but info is displayed. I'm wondering this behavior.\n\nI'm maybe missing something, but if this is a problem, I'll send a\npatch. Any comments are welcome!\nVisibility means search_path, not permission. If s_a is not in the search_paths it objects are not visible unqualified but can be seen (catalog) when schema qualified.David J.","msg_date":"Sun, 7 Nov 2021 20:40:21 -0700","msg_from":"\"David G. Johnston\" <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: Question about psql meta-command with schema option doesn't use\n visibilityrule"},{"msg_contents":"Hi David,\n\n> I have a question that is a specification of permission check\n> (visibilityrule) for psql meta-command with schema option.\n> \n> Visibility means search_path, not permission. If s_a is not in the search_paths it objects are not visible unqualified but can be seen (catalog) when schema qualified.\n\nThanks for your comments! I understood them:\n - all users can show System catalog (pg_catalog. *) is a\n specification, so it is not a bug\n - visibility and permission are not the same (I confused it before, oops)\n\nRegards,\nTatsuro Yamada\n\n\n\n","msg_date":"Mon, 08 Nov 2021 15:06:54 +0900","msg_from":"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>","msg_from_op":true,"msg_subject":"Re: Question about psql meta-command with schema option doesn't use\n visibilityrule"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI have a question that is a specification of permission check\\n(visibilityrule) for psql meta-command with schema option.\\n\\nAccording to the source code [1], there is no check if a schema\\noption is added. As a result, a role that is not granted can see\\nother roles' object names.\\nWe might say it's okay because it's a name, not contents (data),\\nbut It seems not preferable, I think.\\n\\nThe following is a reproducer using \\\\dX commands.\\nNote: It is not only \\\\dX but also \\\\d because it uses the same\\npermission check function (processSQLNamePattern).\\n\\nThe reproduction procedure (including some results):\\n================================================\\n-- Create role a, b as non-superuser\\ncreate role a nosuperuser;\\ncreate role b nosuperuser;\\ngrant CREATE on database postgres to a;\\n\\n-- Create schema s_a, table hoge, and its extend stats by role a\\nset role a;\\ncreate schema s_a;\\ncreate table s_a.hoge(a int, b int);\\ncreate statistics s_a.hoge_ext on a,b from s_a.hoge;\\nset search_path to public, s_a;\\n\\n-- Run \\\\dX and \\\\dX s_a.* by role a: OK (since schema s_a was created by role a)\\n\\\\dX\\n List of extended statistics\\n Schema | Name | Definition | Ndistinct | Dependencies | MCV\\n--------+----------+----------------+-----------+--------------+---------\\n s_a | hoge_ext | a, b FROM hoge | defined | defined | defined\\n(1 row)\\n\\n\\\\dX s_a.*\\n List of extended statistics\\n Schema | Name | Definition | Ndistinct | Dependencies | MCV\\n--------+----------+----------------+-----------+--------------+---------\\n s_a | hoge_ext | a, b FROM hoge | defined | defined | defined\\n(1 row)\\n\\n-- Run \\\\dX by role b: OK\\n-- (not displayed is fine since role b can't see info of role a)\\nreset role;\\nset role b;\\n\\\\dX\\n List of extended statistics\\n Schema | Name | Definition | Ndistinct | Dependencies | MCV\\n--------+------+------------+-----------+--------------+-----\\n(0 rows)\\n\\n-- Run \\\\dX with schema by role b: OK?? (It should be NG?)\\n-- this case is a point in my question\\n\\\\dX s_a.*\\n List of extended statistics\\n Schema | Name | Definition | Ndistinct | Dependencies | MCV\\n--------+----------+--------------------+-----------+--------------+---------\\n s_a | hoge_ext | a, b FROM s_a.hoge | defined | defined | defined\\n(1 row)\\n\\n-- clean-up\\nreset role;\\ndrop schema s_a cascade;\\nrevoke CREATE on DATABASE postgres FROM a;\\ndrop role a;\\ndrop role b;\\n================================================\\n\\n From the above results, I expected \\\"\\\\dX s_a.*\\\" doesn't show any info\\nas same as \\\"\\\\dX\\\". but info is displayed. I'm wondering this behavior.\\n\\nI'm maybe missing something, but if this is a problem, I'll send a\\npatch. Any comments are welcome!\\n\\n\\n[1]: processSQLNamePattern in src/fe_utils/string_utils.c\\n if (schemabuf.len > 2)\\n {\\n /* We have a schema pattern, so constrain the schemavar */\\n\\n /* Optimize away a \\\"*\\\" pattern */\\n if (strcmp(schemabuf.data, \\\"^(.*)$\\\") != 0 && schemavar)\\n {\\n WHEREAND();\\n appendPQExpBuffer(buf, \\\"%s OPERATOR(pg_catalog.~) \\\", schemavar);\\n appendStringLiteralConn(buf, schemabuf.data, conn);\\n if (PQserverVersion(conn) >= 120000)\\n appendPQExpBufferStr(buf, \\\" COLLATE pg_catalog.default\\\");\\n appendPQExpBufferChar(buf, '\\\\n');\\n }\\n }\\n else\\n {\\n /* No schema pattern given, so select only visible objects */\\n if (visibilityrule)\\n {\\n WHEREAND();\\n appendPQExpBuffer(buf, \\\"%s\\\\n\\\", visibilityrule);\\n }\\n }\\n\\n\\n\\nThanks,\\nTatsuro Yamada\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 08 Nov 2021 11:43:34 +0900\",\n \"msg_from\": \"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Question about psql meta-command with schema option doesn't use\\n visibilityrule\"\n },\n {\n \"msg_contents\": \"On Sunday, November 7, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\\nwrote:\\n\\n>\\n> According to the source code [1], there is no check if a schema\\n> option is added. As a result, a role that is not granted can see\\n> other roles' object names.\\n> We might say it's okay because it's a name, not contents (data),\\n> but It seems not preferable, I think.\\n>\\n\\nNo, we are not interested in changing this long-standing documented\\nbehavior. The contents of the catalogs are visible to all. So even if\\nthis was something to consider, psql is not the correct scope.\\n\\nDavid J.\\n\\nOn Sunday, November 7, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp> wrote:\\nAccording to the source code [1], there is no check if a schema\\noption is added. As a result, a role that is not granted can see\\nother roles' object names.\\nWe might say it's okay because it's a name, not contents (data),\\nbut It seems not preferable, I think.\\nNo, we are not interested in changing this long-standing documented behavior. The contents of the catalogs are visible to all. So even if this was something to consider, psql is not the correct scope.David J.\",\n \"msg_date\": \"Sun, 7 Nov 2021 20:09:44 -0700\",\n \"msg_from\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Question about psql meta-command with schema option doesn't use\\n visibilityrule\"\n },\n {\n \"msg_contents\": \"On Sunday, November 7, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\\nwrote:\\n\\n>\\n> I have a question that is a specification of permission check\\n> (visibilityrule) for psql meta-command with schema option.\\n>\\n> From the above results, I expected \\\"\\\\dX s_a.*\\\" doesn't show any info\\n> as same as \\\"\\\\dX\\\". but info is displayed. I'm wondering this behavior.\\n>\\n> I'm maybe missing something, but if this is a problem, I'll send a\\n> patch. Any comments are welcome!\\n>\\n>\\nVisibility means search_path, not permission. If s_a is not in the\\nsearch_paths it objects are not visible unqualified but can be seen\\n(catalog) when schema qualified.\\n\\nDavid J.\\n\\nOn Sunday, November 7, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp> wrote:\\nI have a question that is a specification of permission check\\n(visibilityrule) for psql meta-command with schema option.\\nFrom the above results, I expected \\\"\\\\dX s_a.*\\\" doesn't show any info\\nas same as \\\"\\\\dX\\\". but info is displayed. I'm wondering this behavior.\\n\\nI'm maybe missing something, but if this is a problem, I'll send a\\npatch. Any comments are welcome!\\nVisibility means search_path, not permission. If s_a is not in the search_paths it objects are not visible unqualified but can be seen (catalog) when schema qualified.David J.\",\n \"msg_date\": \"Sun, 7 Nov 2021 20:40:21 -0700\",\n \"msg_from\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Question about psql meta-command with schema option doesn't use\\n visibilityrule\"\n },\n {\n \"msg_contents\": \"Hi David,\\n\\n> I have a question that is a specification of permission check\\n> (visibilityrule) for psql meta-command with schema option.\\n> \\n> Visibility means search_path, not permission. If s_a is not in the search_paths it objects are not visible unqualified but can be seen (catalog) when schema qualified.\\n\\nThanks for your comments! I understood them:\\n - all users can show System catalog (pg_catalog. *) is a\\n specification, so it is not a bug\\n - visibility and permission are not the same (I confused it before, oops)\\n\\nRegards,\\nTatsuro Yamada\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 08 Nov 2021 15:06:54 +0900\",\n \"msg_from\": \"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Question about psql meta-command with schema option doesn't use\\n visibilityrule\"\n }\n]","isConversational":false}}},{"rowIdx":53054,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nConfiguration parameters for pg_stat_statements were not in the index,\nso I added them just like auto_explain configuration parameters.\n\nBest wishes,\n\n-- \nKen Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION","msg_date":"Mon, 08 Nov 2021 15:12:55 +0900","msg_from":"Ken Kato <katouknl@oss.nttdata.com>","msg_from_op":true,"msg_subject":"[PATCH] pg_stat_statements Configuration Parameters Documentation"},{"msg_contents":"\n\nOn 2021/11/08 15:12, Ken Kato wrote:\n> Hi,\n> \n> Configuration parameters for pg_stat_statements were not in the index,\n> so I added them just like auto_explain configuration parameters.\n\nThanks for the patch! LGTM.\n\nRegards,\n\n-- \nFujii Masao\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Mon, 8 Nov 2021 16:19:38 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] pg_stat_statements Configuration Parameters Documentation"},{"msg_contents":"On Mon, Nov 8, 2021 at 3:19 PM Fujii Masao <masao.fujii@oss.nttdata.com> wrote:\n>\n> On 2021/11/08 15:12, Ken Kato wrote:\n> > Hi,\n> >\n> > Configuration parameters for pg_stat_statements were not in the index,\n> > so I added them just like auto_explain configuration parameters.\n>\n> Thanks for the patch! LGTM.\n\n+1, patch looks good to me too.\n\n\n","msg_date":"Mon, 8 Nov 2021 15:23:05 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] pg_stat_statements Configuration Parameters Documentation"},{"msg_contents":"\n\nOn 2021/11/08 16:23, Julien Rouhaud wrote:\n> On Mon, Nov 8, 2021 at 3:19 PM Fujii Masao <masao.fujii@oss.nttdata.com> wrote:\n>>\n>> On 2021/11/08 15:12, Ken Kato wrote:\n>>> Hi,\n>>>\n>>> Configuration parameters for pg_stat_statements were not in the index,\n>>> so I added them just like auto_explain configuration parameters.\n>>\n>> Thanks for the patch! LGTM.\n> \n> +1, patch looks good to me too.\n\nSo I pushed the patch. Thanks!\n\nRegards,\n\n-- \nFujii Masao\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Tue, 9 Nov 2021 12:41:20 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] pg_stat_statements Configuration Parameters Documentation"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nConfiguration parameters for pg_stat_statements were not in the index,\\nso I added them just like auto_explain configuration parameters.\\n\\nBest wishes,\\n\\n-- \\nKen Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Mon, 08 Nov 2021 15:12:55 +0900\",\n \"msg_from\": \"Ken Kato <katouknl@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[PATCH] pg_stat_statements Configuration Parameters Documentation\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2021/11/08 15:12, Ken Kato wrote:\\n> Hi,\\n> \\n> Configuration parameters for pg_stat_statements were not in the index,\\n> so I added them just like auto_explain configuration parameters.\\n\\nThanks for the patch! LGTM.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 Nov 2021 16:19:38 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] pg_stat_statements Configuration Parameters Documentation\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 8, 2021 at 3:19 PM Fujii Masao <masao.fujii@oss.nttdata.com> wrote:\\n>\\n> On 2021/11/08 15:12, Ken Kato wrote:\\n> > Hi,\\n> >\\n> > Configuration parameters for pg_stat_statements were not in the index,\\n> > so I added them just like auto_explain configuration parameters.\\n>\\n> Thanks for the patch! LGTM.\\n\\n+1, patch looks good to me too.\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 Nov 2021 15:23:05 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] pg_stat_statements Configuration Parameters Documentation\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2021/11/08 16:23, Julien Rouhaud wrote:\\n> On Mon, Nov 8, 2021 at 3:19 PM Fujii Masao <masao.fujii@oss.nttdata.com> wrote:\\n>>\\n>> On 2021/11/08 15:12, Ken Kato wrote:\\n>>> Hi,\\n>>>\\n>>> Configuration parameters for pg_stat_statements were not in the index,\\n>>> so I added them just like auto_explain configuration parameters.\\n>>\\n>> Thanks for the patch! LGTM.\\n> \\n> +1, patch looks good to me too.\\n\\nSo I pushed the patch. Thanks!\\n\\nRegards,\\n\\n-- \\nFujii Masao\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 12:41:20 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] pg_stat_statements Configuration Parameters Documentation\"\n }\n]","isConversational":false}}},{"rowIdx":53055,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi All,\n\nWhile I was searching for methods to send the result of a query to the other side of the world, because it might be nice to go there once in a while, I noticed\nmy best option, SSL compression, has been disabled as of version 14. Could someone please explain to me, why compression is being considered unsafe / insecure?\n\nTransmissions to areas outside of Europe cost €5/mb at the moment and that makes SSL compression, or compression in general a vital component of d\nata transmissions.\n\nMight the underlying reason be, that certain people have shown interest in my libpq/PQblockwrite algorithms (\nhttps://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu) but felt turned down and are now persuading me to trade\nthe algorithms against SSL compression, than just say so please. I'll see what I can do.\n\nBest regards,\nMischa Baars.\n\n\n\n","msg_date":"Mon, 08 Nov 2021 08:41:42 +0100","msg_from":"\"Michael J. Baars\" <mjbaars1977.pgsql.hackers@gmail.com>","msg_from_op":true,"msg_subject":"SSL compression"},{"msg_contents":"At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\n>\n> Could someone please explain to me, why compression is being\n> considered unsafe / insecure?\n\nhttps://en.wikipedia.org/wiki/CRIME\n\n> Might the underlying reason be, that certain people have shown\n> interest in my libpq/PQblockwrite algorithms (\n> https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu)\n> but felt turned down and are now persuading me to trade the algorithms\n> against SSL compression, than just say so please. I'll see what I can\n> do.\n\nThe whole world is trying to move away from TLS compression (which has\nbeen removed from TLS 1.3). It has nothing to do with you.\n\n-- Abhijit\n\n\n","msg_date":"Mon, 8 Nov 2021 13:30:21 +0530","msg_from":"Abhijit Menon-Sen <ams@toroid.org>","msg_from_op":false,"msg_subject":"Re: SSL compression"},{"msg_contents":"On Mon, 2021-11-08 at 13:30 +0530, Abhijit Menon-Sen wrote:\n> At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\n> > Could someone please explain to me, why compression is being\n> > considered unsafe / insecure?\n> \n> https://en.wikipedia.org/wiki/CRIME\n> \n\nWell Abhijit, personally I don't see any connection between crime and compression. I do see however, that some people might feel safer communicating over an SSL\nENCRYPTED line doing their daily business, unjustified as that is, but they shouldn't be feeling safer communicating over a compressed line, that would be\nutterly stupid.\n\nThe sole purpose of compression is to reduce the size of a particular amount of data.\n\n> > Might the underlying reason be, that certain people have shown\n> > interest in my libpq/PQblockwrite algorithms (\n> > https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu)\n> > but felt turned down and are now persuading me to trade the algorithms\n> > against SSL compression, than just say so please. I'll see what I can\n> > do.\n> \n> The whole world is trying to move away from TLS compression (which has\n> been removed from TLS 1.3). It has nothing to do with you.\n\nAs I understand it, TLS is a predecessor of SSL. People are trying to move away from TLS, not from compression.\n\n> \n> -- Abhijit\n\n\n\n","msg_date":"Mon, 08 Nov 2021 10:10:55 +0100","msg_from":"\"Michael J. Baars\" <mjbaars1977.pgsql.hackers@gmail.com>","msg_from_op":true,"msg_subject":"Re: SSL compression"},{"msg_contents":"On Mon, Nov 8, 2021 at 10:11 AM Michael J. Baars <\nmjbaars1977.pgsql.hackers@gmail.com> wrote:\n\n> On Mon, 2021-11-08 at 13:30 +0530, Abhijit Menon-Sen wrote:\n> > At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\n> > > Could someone please explain to me, why compression is being\n> > > considered unsafe / insecure?\n> >\n> > https://en.wikipedia.org/wiki/CRIME\n> >\n>\n> Well Abhijit, personally I don't see any connection between crime and\n> compression. I do see however, that some people might feel safer\n> communicating over an SSL\n> ENCRYPTED line doing their daily business, unjustified as that is, but\n> they shouldn't be feeling safer communicating over a compressed line, that\n> would be\n> utterly stupid.\n>\n\nThis is talking about the CRIME attack which *explicitly* is about using\nthe fact that it's the *combination* of encryption and compression that\ncauses a problem. There are other similar attacks as well. If you read the\nlink posted, it will give you a pretty decent overview.\n\n\nThe sole purpose of compression is to reduce the size of a particular\n> amount of data.\n\n\nThis part is of course correct. The problem lies in the *interaction* of\ncompression and encryption. This is why doing compression as part of the\nencryption layer is fundamentally wrong.\n\nThe problem is that to make it secure, you have to encrypt first and then\ncompress. But encryption makes the compression a *lot* less efficient, so\nwhat most solutions did was compress first and then encrypt, which led to\nvulnerabilities.\n\nWould it be good to have the ability to do compression, independent of the\nencryption? Yes, definitely. But this is not a feature that PostgreSQL has,\nor ever had.\n\n\n> > Might the underlying reason be, that certain people have shown\n> > > interest in my libpq/PQblockwrite algorithms (\n> > >\n> https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu\n> )\n> > > but felt turned down and are now persuading me to trade the algorithms\n> > > against SSL compression, than just say so please. I'll see what I can\n> > > do.\n> >\n> > The whole world is trying to move away from TLS compression (which has\n> > been removed from TLS 1.3). It has nothing to do with you.\n>\n> As I understand it, TLS is a predecessor of SSL. People are trying to move\n> away from TLS, not from compression.\n>\n\nThen you don't understand it.\n\nSSL is the predecessor to TLS. TLS has more or less entirely replaced SSL\n-- only extremely outdated systems use SSL. TLSv1 followed after SSLv3.\n\nPeople are moving *to* TLS. And in particular, TLS v1.3 made significant\nand important changes. Compression was dropped from TLS v1.3 because it's\nconsidered obsolete.\n\nAgain, compression is a perfectly valid usecase. But it shouldn't be, and\nis no longer, a part of the encryption layer. Unfortunately, PostgreSQL\nused to rely on it being that, so once it was removed from there, there is\nno built-in support for compression.\n\nOne way to handle it, which is a bit ugly, is to tunnel the data through a\nVPN or something like that which can enable compression.\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/ <http://www.hagander.net/>\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\n\nOn Mon, Nov 8, 2021 at 10:11 AM Michael J. Baars <mjbaars1977.pgsql.hackers@gmail.com> wrote:On Mon, 2021-11-08 at 13:30 +0530, Abhijit Menon-Sen wrote:\n> At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\n> > Could someone please explain to me, why compression is being\n> > considered unsafe / insecure?\n> \n> https://en.wikipedia.org/wiki/CRIME\n> \n\nWell Abhijit, personally I don't see any connection between crime and compression. I do see however, that some people might feel safer communicating over an SSL\nENCRYPTED line doing their daily business, unjustified as that is, but they shouldn't be feeling safer communicating over a compressed line, that would be\nutterly stupid.This is talking about the CRIME attack which *explicitly* is about using the fact that it's the *combination* of encryption and compression that causes a problem. There are other similar attacks as well. If you read the link posted, it will give you a pretty decent overview.\nThe sole purpose of compression is to reduce the size of a particular amount of data. This part is of course correct. The problem lies in the *interaction* of compression and encryption. This is why doing compression as part of the encryption layer is fundamentally wrong.The problem is that to make it secure, you have to encrypt first and then compress. But encryption makes the compression a *lot* less efficient, so what most solutions did was compress first and then encrypt, which led to vulnerabilities.Would it be good to have the ability to do compression, independent of the encryption? Yes, definitely. But this is not a feature that PostgreSQL has, or ever had.\n> > Might the underlying reason be, that certain people have shown\n> > interest in my libpq/PQblockwrite algorithms (\n> > https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu)\n> > but felt turned down and are now persuading me to trade the algorithms\n> > against SSL compression, than just say so please. I'll see what I can\n> > do.\n> \n> The whole world is trying to move away from TLS compression (which has\n> been removed from TLS 1.3). It has nothing to do with you.\n\nAs I understand it, TLS is a predecessor of SSL. People are trying to move away from TLS, not from compression.Then you don't understand it.SSL is the predecessor to TLS. TLS has more or less entirely replaced SSL -- only extremely outdated systems use SSL. TLSv1 followed after SSLv3. People are moving *to* TLS. And in particular, TLS v1.3 made significant and important changes. Compression was dropped from TLS v1.3 because it's considered obsolete.Again, compression is a perfectly valid usecase. But it shouldn't be, and is no longer, a part of the encryption layer. Unfortunately, PostgreSQL used to rely on it being that, so once it was removed from there, there is no built-in support for compression.One way to handle it, which is a bit ugly, is to tunnel the data through a VPN or something like that which can enable compression.-- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/","msg_date":"Mon, 8 Nov 2021 10:20:33 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: SSL compression"},{"msg_contents":"At 2021-11-08 10:10:55 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\n>\n> > https://en.wikipedia.org/wiki/CRIME\n> \n> Well Abhijit, personally I don't see any connection between crime and\n> compression.\n\nReading the page I linked to may help you make the connection. If not,\ntry any of the numerous other pages that talk about vulnerabilities in\nTLS related to compression.\n\n> As I understand it, TLS is a predecessor of SSL.\n\nNo. SSL is the predecessor.\n\n> People are trying to move away from TLS, not from compression.\n\nNo.\n\n-- ams\n\n\n","msg_date":"Mon, 8 Nov 2021 15:00:48 +0530","msg_from":"Abhijit Menon-Sen <ams@toroid.org>","msg_from_op":false,"msg_subject":"Re: SSL compression"},{"msg_contents":"On Mon, 2021-11-08 at 10:20 +0100, Magnus Hagander wrote:\n> On Mon, Nov 8, 2021 at 10:11 AM Michael J. Baars <mjbaars1977.pgsql.hackers@gmail.com> wrote:\n> > On Mon, 2021-11-08 at 13:30 +0530, Abhijit Menon-Sen wrote:\n> > \n> > > At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\n> > \n> > > > Could someone please explain to me, why compression is being\n> > \n> > > > considered unsafe / insecure?\n> > \n> > > \n> > \n> > > https://en.wikipedia.org/wiki/CRIME\n> > \n> > > \n> > \n> > \n> > \n> > Well Abhijit, personally I don't see any connection between crime and compression. I do see however, that some people might feel safer communicating over an\n> > SSL\n> > \n> > ENCRYPTED line doing their daily business, unjustified as that is, but they shouldn't be feeling safer communicating over a compressed line, that would be\n> > \n> > utterly stupid.\n> \n> This is talking about the CRIME attack which *explicitly* is about using the fact that it's the *combination* of encryption and compression that causes a\n> problem. There are other similar attacks as well. If you read the link posted, it will give you a pretty decent overview.\n> \n> \n> > The sole purpose of compression is to reduce the size of a particular amount of data. \n> \n> This part is of course correct. The problem lies in the *interaction* of compression and encryption. This is why doing compression as part of the encryption\n> layer is fundamentally wrong.\n> \n> The problem is that to make it secure, you have to encrypt first and then compress. But encryption makes the compression a *lot* less efficient, so what most\n> solutions did was compress first and then encrypt, which led to vulnerabilities.\n\nYes, I am aware of the fact that it is pretty much useless to compress an already encrypted data source, the other way around indeed makes more sense.\nAfter reading the link, it looks to me that this attacker has for some reason, knowledge of the original unencrypted and uncompressed data or even uses its own\ndata (The attacker then observes the change in size of the compressed request payload, which contains both the secret cookie that is sent by the browser only\nto the target site, and variable content created by the attacker, as the variable content is altered), which is certainly never the case in a PostgreSQL data\ntransmission. Actually, I don't really see this happening. I've checked all the references, and in my opinion they're all pretty much non-believers. I have more\nfaith in huffman encoding and encryption than I have in these guys, that's for sure.\n> Would it be good to have the ability to do compression, independent of the encryption? Yes, definitely. But this is not a feature that PostgreSQL has, or ever\n> had.\n\nThat's too bad! Perhaps an idea for future development?\n> > > > Might the underlying reason be, that certain people have shown\n> > \n> > > > interest in my libpq/PQblockwrite algorithms (\n> > \n> > > > https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu)\n> > \n> > > > but felt turned down and are now persuading me to trade the algorithms\n> > \n> > > > against SSL compression, than just say so please. I'll see what I can\n> > \n> > > > do.\n> > \n> > > \n> > \n> > > The whole world is trying to move away from TLS compression (which has\n> > \n> > > been removed from TLS 1.3). It has nothing to do with you.\n> > \n> > \n> > \n> > As I understand it, TLS is a predecessor of SSL. People are trying to move away from TLS, not from compression.\n> \n> Then you don't understand it.\n> \n> SSL is the predecessor to TLS. TLS has more or less entirely replaced SSL -- only extremely outdated systems use SSL. TLSv1 followed after SSLv3. \n> \n> People are moving *to* TLS. And in particular, TLS v1.3 made significant and important changes. Compression was dropped from TLS v1.3 because it's considered\n> obsolete.\n> \n> Again, compression is a perfectly valid usecase. But it shouldn't be, and is no longer, a part of the encryption layer. Unfortunately, PostgreSQL used to rely\n> on it being that, so once it was removed from there, there is no built-in support for compression.\n> \n> One way to handle it, which is a bit ugly, is to tunnel the data through a VPN or something like that which can enable compression.\n> \n\nOn Mon, 2021-11-08 at 10:20 +0100, Magnus Hagander wrote:On Mon, Nov 8, 2021 at 10:11 AM Michael J. Baars <mjbaars1977.pgsql.hackers@gmail.com> wrote:On Mon, 2021-11-08 at 13:30 +0530, Abhijit Menon-Sen wrote:\n> At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\n> > Could someone please explain to me, why compression is being\n> > considered unsafe / insecure?\n> \n> https://en.wikipedia.org/wiki/CRIME\n> \n\nWell Abhijit, personally I don't see any connection between crime and compression. I do see however, that some people might feel safer communicating over an SSL\nENCRYPTED line doing their daily business, unjustified as that is, but they shouldn't be feeling safer communicating over a compressed line, that would be\nutterly stupid.This is talking about the CRIME attack which *explicitly* is about using the fact that it's the *combination* of encryption and compression that causes a problem. There are other similar attacks as well. If you read the link posted, it will give you a pretty decent overview.\nThe sole purpose of compression is to reduce the size of a particular amount of data. This part is of course correct. The problem lies in the *interaction* of compression and encryption. This is why doing compression as part of the encryption layer is fundamentally wrong.The problem is that to make it secure, you have to encrypt first and then compress. But encryption makes the compression a *lot* less efficient, so what most solutions did was compress first and then encrypt, which led to vulnerabilities.Yes, I am aware of the fact that it is pretty much useless to compress an already encrypted data source, the other way around indeed makes more sense.After reading the link, it looks to me that this attacker has for some reason, knowledge of the original unencrypted and uncompressed data or even uses its own data (The attacker then observes the change in size of the compressed request payload, which contains both the secret cookie that is sent by the browser only to the target site, and variable content created by the attacker, as the variable content is altered), which is certainly never the case in a PostgreSQL data transmission. Actually, I don't really see this happening. I've checked all the references, and in my opinion they're all pretty much non-believers. I have more faith in huffman encoding and encryption than I have in these guys, that's for sure.Would it be good to have the ability to do compression, independent of the encryption? Yes, definitely. But this is not a feature that PostgreSQL has, or ever had.That's too bad! Perhaps an idea for future development?\n> > Might the underlying reason be, that certain people have shown\n> > interest in my libpq/PQblockwrite algorithms (\n> > https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu)\n> > but felt turned down and are now persuading me to trade the algorithms\n> > against SSL compression, than just say so please. I'll see what I can\n> > do.\n> \n> The whole world is trying to move away from TLS compression (which has\n> been removed from TLS 1.3). It has nothing to do with you.\n\nAs I understand it, TLS is a predecessor of SSL. People are trying to move away from TLS, not from compression.Then you don't understand it.SSL is the predecessor to TLS. TLS has more or less entirely replaced SSL -- only extremely outdated systems use SSL. TLSv1 followed after SSLv3. People are moving *to* TLS. And in particular, TLS v1.3 made significant and important changes. Compression was dropped from TLS v1.3 because it's considered obsolete.Again, compression is a perfectly valid usecase. But it shouldn't be, and is no longer, a part of the encryption layer. Unfortunately, PostgreSQL used to rely on it being that, so once it was removed from there, there is no built-in support for compression.One way to handle it, which is a bit ugly, is to tunnel the data through a VPN or something like that which can enable compression.","msg_date":"Mon, 08 Nov 2021 12:19:31 +0100","msg_from":"\"Michael J. Baars\" <mjbaars1977.pgsql.hackers@gmail.com>","msg_from_op":true,"msg_subject":"Re: SSL compression"}],"string":"[\n {\n \"msg_contents\": \"Hi All,\\n\\nWhile I was searching for methods to send the result of a query to the other side of the world, because it might be nice to go there once in a while, I noticed\\nmy best option, SSL compression, has been disabled as of version 14. Could someone please explain to me, why compression is being considered unsafe / insecure?\\n\\nTransmissions to areas outside of Europe cost €5/mb at the moment and that makes SSL compression, or compression in general a vital component of d\\nata transmissions.\\n\\nMight the underlying reason be, that certain people have shown interest in my libpq/PQblockwrite algorithms (\\nhttps://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu) but felt turned down and are now persuading me to trade\\nthe algorithms against SSL compression, than just say so please. I'll see what I can do.\\n\\nBest regards,\\nMischa Baars.\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 08 Nov 2021 08:41:42 +0100\",\n \"msg_from\": \"\\\"Michael J. Baars\\\" <mjbaars1977.pgsql.hackers@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"SSL compression\"\n },\n {\n \"msg_contents\": \"At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\\n>\\n> Could someone please explain to me, why compression is being\\n> considered unsafe / insecure?\\n\\nhttps://en.wikipedia.org/wiki/CRIME\\n\\n> Might the underlying reason be, that certain people have shown\\n> interest in my libpq/PQblockwrite algorithms (\\n> https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu)\\n> but felt turned down and are now persuading me to trade the algorithms\\n> against SSL compression, than just say so please. I'll see what I can\\n> do.\\n\\nThe whole world is trying to move away from TLS compression (which has\\nbeen removed from TLS 1.3). It has nothing to do with you.\\n\\n-- Abhijit\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 Nov 2021 13:30:21 +0530\",\n \"msg_from\": \"Abhijit Menon-Sen <ams@toroid.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: SSL compression\"\n },\n {\n \"msg_contents\": \"On Mon, 2021-11-08 at 13:30 +0530, Abhijit Menon-Sen wrote:\\n> At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\\n> > Could someone please explain to me, why compression is being\\n> > considered unsafe / insecure?\\n> \\n> https://en.wikipedia.org/wiki/CRIME\\n> \\n\\nWell Abhijit, personally I don't see any connection between crime and compression. I do see however, that some people might feel safer communicating over an SSL\\nENCRYPTED line doing their daily business, unjustified as that is, but they shouldn't be feeling safer communicating over a compressed line, that would be\\nutterly stupid.\\n\\nThe sole purpose of compression is to reduce the size of a particular amount of data.\\n\\n> > Might the underlying reason be, that certain people have shown\\n> > interest in my libpq/PQblockwrite algorithms (\\n> > https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu)\\n> > but felt turned down and are now persuading me to trade the algorithms\\n> > against SSL compression, than just say so please. I'll see what I can\\n> > do.\\n> \\n> The whole world is trying to move away from TLS compression (which has\\n> been removed from TLS 1.3). It has nothing to do with you.\\n\\nAs I understand it, TLS is a predecessor of SSL. People are trying to move away from TLS, not from compression.\\n\\n> \\n> -- Abhijit\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 08 Nov 2021 10:10:55 +0100\",\n \"msg_from\": \"\\\"Michael J. Baars\\\" <mjbaars1977.pgsql.hackers@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: SSL compression\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 8, 2021 at 10:11 AM Michael J. Baars <\\nmjbaars1977.pgsql.hackers@gmail.com> wrote:\\n\\n> On Mon, 2021-11-08 at 13:30 +0530, Abhijit Menon-Sen wrote:\\n> > At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\\n> > > Could someone please explain to me, why compression is being\\n> > > considered unsafe / insecure?\\n> >\\n> > https://en.wikipedia.org/wiki/CRIME\\n> >\\n>\\n> Well Abhijit, personally I don't see any connection between crime and\\n> compression. I do see however, that some people might feel safer\\n> communicating over an SSL\\n> ENCRYPTED line doing their daily business, unjustified as that is, but\\n> they shouldn't be feeling safer communicating over a compressed line, that\\n> would be\\n> utterly stupid.\\n>\\n\\nThis is talking about the CRIME attack which *explicitly* is about using\\nthe fact that it's the *combination* of encryption and compression that\\ncauses a problem. There are other similar attacks as well. If you read the\\nlink posted, it will give you a pretty decent overview.\\n\\n\\nThe sole purpose of compression is to reduce the size of a particular\\n> amount of data.\\n\\n\\nThis part is of course correct. The problem lies in the *interaction* of\\ncompression and encryption. This is why doing compression as part of the\\nencryption layer is fundamentally wrong.\\n\\nThe problem is that to make it secure, you have to encrypt first and then\\ncompress. But encryption makes the compression a *lot* less efficient, so\\nwhat most solutions did was compress first and then encrypt, which led to\\nvulnerabilities.\\n\\nWould it be good to have the ability to do compression, independent of the\\nencryption? Yes, definitely. But this is not a feature that PostgreSQL has,\\nor ever had.\\n\\n\\n> > Might the underlying reason be, that certain people have shown\\n> > > interest in my libpq/PQblockwrite algorithms (\\n> > >\\n> https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu\\n> )\\n> > > but felt turned down and are now persuading me to trade the algorithms\\n> > > against SSL compression, than just say so please. I'll see what I can\\n> > > do.\\n> >\\n> > The whole world is trying to move away from TLS compression (which has\\n> > been removed from TLS 1.3). It has nothing to do with you.\\n>\\n> As I understand it, TLS is a predecessor of SSL. People are trying to move\\n> away from TLS, not from compression.\\n>\\n\\nThen you don't understand it.\\n\\nSSL is the predecessor to TLS. TLS has more or less entirely replaced SSL\\n-- only extremely outdated systems use SSL. TLSv1 followed after SSLv3.\\n\\nPeople are moving *to* TLS. And in particular, TLS v1.3 made significant\\nand important changes. Compression was dropped from TLS v1.3 because it's\\nconsidered obsolete.\\n\\nAgain, compression is a perfectly valid usecase. But it shouldn't be, and\\nis no longer, a part of the encryption layer. Unfortunately, PostgreSQL\\nused to rely on it being that, so once it was removed from there, there is\\nno built-in support for compression.\\n\\nOne way to handle it, which is a bit ugly, is to tunnel the data through a\\nVPN or something like that which can enable compression.\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/ <http://www.hagander.net/>\\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\\n\\nOn Mon, Nov 8, 2021 at 10:11 AM Michael J. Baars <mjbaars1977.pgsql.hackers@gmail.com> wrote:On Mon, 2021-11-08 at 13:30 +0530, Abhijit Menon-Sen wrote:\\n> At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\\n> > Could someone please explain to me, why compression is being\\n> > considered unsafe / insecure?\\n> \\n> https://en.wikipedia.org/wiki/CRIME\\n> \\n\\nWell Abhijit, personally I don't see any connection between crime and compression. I do see however, that some people might feel safer communicating over an SSL\\nENCRYPTED line doing their daily business, unjustified as that is, but they shouldn't be feeling safer communicating over a compressed line, that would be\\nutterly stupid.This is talking about the CRIME attack which *explicitly* is about using the fact that it's the *combination* of encryption and compression that causes a problem. There are other similar attacks as well. If you read the link posted, it will give you a pretty decent overview.\\nThe sole purpose of compression is to reduce the size of a particular amount of data. This part is of course correct. The problem lies in the *interaction* of compression and encryption. This is why doing compression as part of the encryption layer is fundamentally wrong.The problem is that to make it secure, you have to encrypt first and then compress. But encryption makes the compression a *lot* less efficient, so what most solutions did was compress first and then encrypt, which led to vulnerabilities.Would it be good to have the ability to do compression, independent of the encryption? Yes, definitely. But this is not a feature that PostgreSQL has, or ever had.\\n> > Might the underlying reason be, that certain people have shown\\n> > interest in my libpq/PQblockwrite algorithms (\\n> > https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu)\\n> > but felt turned down and are now persuading me to trade the algorithms\\n> > against SSL compression, than just say so please. I'll see what I can\\n> > do.\\n> \\n> The whole world is trying to move away from TLS compression (which has\\n> been removed from TLS 1.3). It has nothing to do with you.\\n\\nAs I understand it, TLS is a predecessor of SSL. People are trying to move away from TLS, not from compression.Then you don't understand it.SSL is the predecessor to TLS. TLS has more or less entirely replaced SSL -- only extremely outdated systems use SSL. TLSv1 followed after SSLv3. People are moving *to* TLS. And in particular, TLS v1.3 made significant and important changes. Compression was dropped from TLS v1.3 because it's considered obsolete.Again, compression is a perfectly valid usecase. But it shouldn't be, and is no longer, a part of the encryption layer. Unfortunately, PostgreSQL used to rely on it being that, so once it was removed from there, there is no built-in support for compression.One way to handle it, which is a bit ugly, is to tunnel the data through a VPN or something like that which can enable compression.-- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/\",\n \"msg_date\": \"Mon, 8 Nov 2021 10:20:33 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: SSL compression\"\n },\n {\n \"msg_contents\": \"At 2021-11-08 10:10:55 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\\n>\\n> > https://en.wikipedia.org/wiki/CRIME\\n> \\n> Well Abhijit, personally I don't see any connection between crime and\\n> compression.\\n\\nReading the page I linked to may help you make the connection. If not,\\ntry any of the numerous other pages that talk about vulnerabilities in\\nTLS related to compression.\\n\\n> As I understand it, TLS is a predecessor of SSL.\\n\\nNo. SSL is the predecessor.\\n\\n> People are trying to move away from TLS, not from compression.\\n\\nNo.\\n\\n-- ams\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 Nov 2021 15:00:48 +0530\",\n \"msg_from\": \"Abhijit Menon-Sen <ams@toroid.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: SSL compression\"\n },\n {\n \"msg_contents\": \"On Mon, 2021-11-08 at 10:20 +0100, Magnus Hagander wrote:\\n> On Mon, Nov 8, 2021 at 10:11 AM Michael J. Baars <mjbaars1977.pgsql.hackers@gmail.com> wrote:\\n> > On Mon, 2021-11-08 at 13:30 +0530, Abhijit Menon-Sen wrote:\\n> > \\n> > > At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\\n> > \\n> > > > Could someone please explain to me, why compression is being\\n> > \\n> > > > considered unsafe / insecure?\\n> > \\n> > > \\n> > \\n> > > https://en.wikipedia.org/wiki/CRIME\\n> > \\n> > > \\n> > \\n> > \\n> > \\n> > Well Abhijit, personally I don't see any connection between crime and compression. I do see however, that some people might feel safer communicating over an\\n> > SSL\\n> > \\n> > ENCRYPTED line doing their daily business, unjustified as that is, but they shouldn't be feeling safer communicating over a compressed line, that would be\\n> > \\n> > utterly stupid.\\n> \\n> This is talking about the CRIME attack which *explicitly* is about using the fact that it's the *combination* of encryption and compression that causes a\\n> problem. There are other similar attacks as well. If you read the link posted, it will give you a pretty decent overview.\\n> \\n> \\n> > The sole purpose of compression is to reduce the size of a particular amount of data. \\n> \\n> This part is of course correct. The problem lies in the *interaction* of compression and encryption. This is why doing compression as part of the encryption\\n> layer is fundamentally wrong.\\n> \\n> The problem is that to make it secure, you have to encrypt first and then compress. But encryption makes the compression a *lot* less efficient, so what most\\n> solutions did was compress first and then encrypt, which led to vulnerabilities.\\n\\nYes, I am aware of the fact that it is pretty much useless to compress an already encrypted data source, the other way around indeed makes more sense.\\nAfter reading the link, it looks to me that this attacker has for some reason, knowledge of the original unencrypted and uncompressed data or even uses its own\\ndata (The attacker then observes the change in size of the compressed request payload, which contains both the secret cookie that is sent by the browser only\\nto the target site, and variable content created by the attacker, as the variable content is altered), which is certainly never the case in a PostgreSQL data\\ntransmission. Actually, I don't really see this happening. I've checked all the references, and in my opinion they're all pretty much non-believers. I have more\\nfaith in huffman encoding and encryption than I have in these guys, that's for sure.\\n> Would it be good to have the ability to do compression, independent of the encryption? Yes, definitely. But this is not a feature that PostgreSQL has, or ever\\n> had.\\n\\nThat's too bad! Perhaps an idea for future development?\\n> > > > Might the underlying reason be, that certain people have shown\\n> > \\n> > > > interest in my libpq/PQblockwrite algorithms (\\n> > \\n> > > > https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu)\\n> > \\n> > > > but felt turned down and are now persuading me to trade the algorithms\\n> > \\n> > > > against SSL compression, than just say so please. I'll see what I can\\n> > \\n> > > > do.\\n> > \\n> > > \\n> > \\n> > > The whole world is trying to move away from TLS compression (which has\\n> > \\n> > > been removed from TLS 1.3). It has nothing to do with you.\\n> > \\n> > \\n> > \\n> > As I understand it, TLS is a predecessor of SSL. People are trying to move away from TLS, not from compression.\\n> \\n> Then you don't understand it.\\n> \\n> SSL is the predecessor to TLS. TLS has more or less entirely replaced SSL -- only extremely outdated systems use SSL. TLSv1 followed after SSLv3. \\n> \\n> People are moving *to* TLS. And in particular, TLS v1.3 made significant and important changes. Compression was dropped from TLS v1.3 because it's considered\\n> obsolete.\\n> \\n> Again, compression is a perfectly valid usecase. But it shouldn't be, and is no longer, a part of the encryption layer. Unfortunately, PostgreSQL used to rely\\n> on it being that, so once it was removed from there, there is no built-in support for compression.\\n> \\n> One way to handle it, which is a bit ugly, is to tunnel the data through a VPN or something like that which can enable compression.\\n> \\n\\nOn Mon, 2021-11-08 at 10:20 +0100, Magnus Hagander wrote:On Mon, Nov 8, 2021 at 10:11 AM Michael J. Baars <mjbaars1977.pgsql.hackers@gmail.com> wrote:On Mon, 2021-11-08 at 13:30 +0530, Abhijit Menon-Sen wrote:\\n> At 2021-11-08 08:41:42 +0100, mjbaars1977.pgsql.hackers@gmail.com wrote:\\n> > Could someone please explain to me, why compression is being\\n> > considered unsafe / insecure?\\n> \\n> https://en.wikipedia.org/wiki/CRIME\\n> \\n\\nWell Abhijit, personally I don't see any connection between crime and compression. I do see however, that some people might feel safer communicating over an SSL\\nENCRYPTED line doing their daily business, unjustified as that is, but they shouldn't be feeling safer communicating over a compressed line, that would be\\nutterly stupid.This is talking about the CRIME attack which *explicitly* is about using the fact that it's the *combination* of encryption and compression that causes a problem. There are other similar attacks as well. If you read the link posted, it will give you a pretty decent overview.\\nThe sole purpose of compression is to reduce the size of a particular amount of data. This part is of course correct. The problem lies in the *interaction* of compression and encryption. This is why doing compression as part of the encryption layer is fundamentally wrong.The problem is that to make it secure, you have to encrypt first and then compress. But encryption makes the compression a *lot* less efficient, so what most solutions did was compress first and then encrypt, which led to vulnerabilities.Yes, I am aware of the fact that it is pretty much useless to compress an already encrypted data source, the other way around indeed makes more sense.After reading the link, it looks to me that this attacker has for some reason, knowledge of the original unencrypted and uncompressed data or even uses its own data (The attacker then observes the change in size of the compressed request payload, which contains both the secret cookie that is sent by the browser only to the target site, and variable content created by the attacker, as the variable content is altered), which is certainly never the case in a PostgreSQL data transmission. Actually, I don't really see this happening. I've checked all the references, and in my opinion they're all pretty much non-believers. I have more faith in huffman encoding and encryption than I have in these guys, that's for sure.Would it be good to have the ability to do compression, independent of the encryption? Yes, definitely. But this is not a feature that PostgreSQL has, or ever had.That's too bad! Perhaps an idea for future development?\\n> > Might the underlying reason be, that certain people have shown\\n> > interest in my libpq/PQblockwrite algorithms (\\n> > https://www.postgresql.org/message-id/c7cccd0777f39c53b9514e3824badf276759fa87.camel%40cyberfiber.eu)\\n> > but felt turned down and are now persuading me to trade the algorithms\\n> > against SSL compression, than just say so please. I'll see what I can\\n> > do.\\n> \\n> The whole world is trying to move away from TLS compression (which has\\n> been removed from TLS 1.3). It has nothing to do with you.\\n\\nAs I understand it, TLS is a predecessor of SSL. People are trying to move away from TLS, not from compression.Then you don't understand it.SSL is the predecessor to TLS. TLS has more or less entirely replaced SSL -- only extremely outdated systems use SSL. TLSv1 followed after SSLv3. People are moving *to* TLS. And in particular, TLS v1.3 made significant and important changes. Compression was dropped from TLS v1.3 because it's considered obsolete.Again, compression is a perfectly valid usecase. But it shouldn't be, and is no longer, a part of the encryption layer. Unfortunately, PostgreSQL used to rely on it being that, so once it was removed from there, there is no built-in support for compression.One way to handle it, which is a bit ugly, is to tunnel the data through a VPN or something like that which can enable compression.\",\n \"msg_date\": \"Mon, 08 Nov 2021 12:19:31 +0100\",\n \"msg_from\": \"\\\"Michael J. Baars\\\" <mjbaars1977.pgsql.hackers@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: SSL compression\"\n }\n]","isConversational":false}}},{"rowIdx":53056,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello,\n\nExpanded data types (types that have a different in-memory working\nrepresentation than on disk) are pretty essential to a couple of projects\nthat I work on, and while the prose documentation is great at explaining\nthe rationale and technical overview, there isn't a complete,\nself-contained, simple example of an expanded data type. Readers of the\ndocumentation are encouraged to study the array type code.\n\nTo wrap my own head around it, and to have a sort of template to quickly\nspin up future expanded types, I wrote a simple extension that does nothing\nuseful other than show how to create and handle an expanded type.\n\nhttps://github.com/michelp/pgexpanded\n\nI posted this last week to psql-general, and Pavel Stehule suggested it\nwould be a good candidate for inclusion in the contrib collection and that\nI write this proposal to suggest the extension get into the next commit\nfest.\n\nI have not contributed to the core of postgres in the past so please excuse\nmy ignorance on the process. Does anyone have any comments or objections\nto me pushing this process further along?\n\n-Michel\n\nHello,Expanded data types (types that have a different in-memory working representation than on disk) are pretty essential to a couple of projects that I work on, and while the prose documentation is great at explaining the rationale and technical overview, there isn't a complete, self-contained, simple example of an expanded data type. Readers of the documentation are encouraged to study the array type code.To wrap my own head around it, and to have a sort of template to quickly spin up future expanded types, I wrote a simple extension that does nothing useful other than show how to create and handle an expanded type.https://github.com/michelp/pgexpandedI posted this last week to psql-general, and Pavel Stehule suggested it would be a good candidate for inclusion in the contrib collection and that I write this proposal to suggest the extension get into the next commit fest.I have not contributed to the core of postgres in the past so please excuse my ignorance on the process. Does anyone have any comments or objections to me pushing this process further along?-Michel","msg_date":"Mon, 8 Nov 2021 08:02:20 -0800","msg_from":"Michel Pelletier <pelletier.michel@gmail.com>","msg_from_op":true,"msg_subject":"Proposal for adding an example \"expanded\" data type to contrib"},{"msg_contents":"On 2021-Nov-08, Michel Pelletier wrote:\n\n> I posted this last week to psql-general, and Pavel Stehule suggested it\n> would be a good candidate for inclusion in the contrib collection and that\n> I write this proposal to suggest the extension get into the next commit\n> fest.\n\nHmm, I think src/test/modules/ would be a better place for it than\ncontrib/ -- alongside plsample, which has a similar intention.\n\n-- \nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Mon, 8 Nov 2021 13:07:08 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Proposal for adding an example \"expanded\" data type to contrib"},{"msg_contents":"On Mon, Nov 8, 2021 at 8:07 AM Alvaro Herrera <alvherre@alvh.no-ip.org>\nwrote:\n\n> On 2021-Nov-08, Michel Pelletier wrote:\n>\n> > I posted this last week to psql-general, and Pavel Stehule suggested it\n> > would be a good candidate for inclusion in the contrib collection and\n> that\n> > I write this proposal to suggest the extension get into the next commit\n> > fest.\n>\n> Hmm, I think src/test/modules/ would be a better place for it than\n> contrib/ -- alongside plsample, which has a similar intention.\n>\n\nAh excellent suggestion, that dir had slipped my mind. I see the readme\nsays:\n\n\"src/test/modules contains PostgreSQL extensions that are primarily or\nentirely intended for testing PostgreSQL and/or to serve as example code.\"\n\nso that seems quite appropriate, thanks!\n\n-Michel\n\n\n> --\n> Álvaro Herrera Valdivia, Chile —\n> https://www.EnterpriseDB.com/\n>\n\nOn Mon, Nov 8, 2021 at 8:07 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:On 2021-Nov-08, Michel Pelletier wrote:\n\n> I posted this last week to psql-general, and Pavel Stehule suggested it\n> would be a good candidate for inclusion in the contrib collection and that\n> I write this proposal to suggest the extension get into the next commit\n> fest.\n\nHmm, I think src/test/modules/ would be a better place for it than\ncontrib/ -- alongside plsample, which has a similar intention.Ah excellent suggestion, that dir had slipped my mind. I see the readme says: \"src/test/modules contains PostgreSQL extensions that are primarily or entirely intended for testing PostgreSQL and/or to serve as example code.\"so that seems quite appropriate, thanks!-Michel \n\n-- \nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/","msg_date":"Mon, 8 Nov 2021 08:12:52 -0800","msg_from":"Michel Pelletier <pelletier.michel@gmail.com>","msg_from_op":true,"msg_subject":"Re: Proposal for adding an example \"expanded\" data type to contrib"}],"string":"[\n {\n \"msg_contents\": \"Hello,\\n\\nExpanded data types (types that have a different in-memory working\\nrepresentation than on disk) are pretty essential to a couple of projects\\nthat I work on, and while the prose documentation is great at explaining\\nthe rationale and technical overview, there isn't a complete,\\nself-contained, simple example of an expanded data type. Readers of the\\ndocumentation are encouraged to study the array type code.\\n\\nTo wrap my own head around it, and to have a sort of template to quickly\\nspin up future expanded types, I wrote a simple extension that does nothing\\nuseful other than show how to create and handle an expanded type.\\n\\nhttps://github.com/michelp/pgexpanded\\n\\nI posted this last week to psql-general, and Pavel Stehule suggested it\\nwould be a good candidate for inclusion in the contrib collection and that\\nI write this proposal to suggest the extension get into the next commit\\nfest.\\n\\nI have not contributed to the core of postgres in the past so please excuse\\nmy ignorance on the process. Does anyone have any comments or objections\\nto me pushing this process further along?\\n\\n-Michel\\n\\nHello,Expanded data types (types that have a different in-memory working representation than on disk) are pretty essential to a couple of projects that I work on, and while the prose documentation is great at explaining the rationale and technical overview, there isn't a complete, self-contained, simple example of an expanded data type. Readers of the documentation are encouraged to study the array type code.To wrap my own head around it, and to have a sort of template to quickly spin up future expanded types, I wrote a simple extension that does nothing useful other than show how to create and handle an expanded type.https://github.com/michelp/pgexpandedI posted this last week to psql-general, and Pavel Stehule suggested it would be a good candidate for inclusion in the contrib collection and that I write this proposal to suggest the extension get into the next commit fest.I have not contributed to the core of postgres in the past so please excuse my ignorance on the process. Does anyone have any comments or objections to me pushing this process further along?-Michel\",\n \"msg_date\": \"Mon, 8 Nov 2021 08:02:20 -0800\",\n \"msg_from\": \"Michel Pelletier <pelletier.michel@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Proposal for adding an example \\\"expanded\\\" data type to contrib\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-08, Michel Pelletier wrote:\\n\\n> I posted this last week to psql-general, and Pavel Stehule suggested it\\n> would be a good candidate for inclusion in the contrib collection and that\\n> I write this proposal to suggest the extension get into the next commit\\n> fest.\\n\\nHmm, I think src/test/modules/ would be a better place for it than\\ncontrib/ -- alongside plsample, which has a similar intention.\\n\\n-- \\nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 Nov 2021 13:07:08 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal for adding an example \\\"expanded\\\" data type to contrib\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 8, 2021 at 8:07 AM Alvaro Herrera <alvherre@alvh.no-ip.org>\\nwrote:\\n\\n> On 2021-Nov-08, Michel Pelletier wrote:\\n>\\n> > I posted this last week to psql-general, and Pavel Stehule suggested it\\n> > would be a good candidate for inclusion in the contrib collection and\\n> that\\n> > I write this proposal to suggest the extension get into the next commit\\n> > fest.\\n>\\n> Hmm, I think src/test/modules/ would be a better place for it than\\n> contrib/ -- alongside plsample, which has a similar intention.\\n>\\n\\nAh excellent suggestion, that dir had slipped my mind. I see the readme\\nsays:\\n\\n\\\"src/test/modules contains PostgreSQL extensions that are primarily or\\nentirely intended for testing PostgreSQL and/or to serve as example code.\\\"\\n\\nso that seems quite appropriate, thanks!\\n\\n-Michel\\n\\n\\n> --\\n> Álvaro Herrera Valdivia, Chile —\\n> https://www.EnterpriseDB.com/\\n>\\n\\nOn Mon, Nov 8, 2021 at 8:07 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:On 2021-Nov-08, Michel Pelletier wrote:\\n\\n> I posted this last week to psql-general, and Pavel Stehule suggested it\\n> would be a good candidate for inclusion in the contrib collection and that\\n> I write this proposal to suggest the extension get into the next commit\\n> fest.\\n\\nHmm, I think src/test/modules/ would be a better place for it than\\ncontrib/ -- alongside plsample, which has a similar intention.Ah excellent suggestion, that dir had slipped my mind. I see the readme says: \\\"src/test/modules contains PostgreSQL extensions that are primarily or entirely intended for testing PostgreSQL and/or to serve as example code.\\\"so that seems quite appropriate, thanks!-Michel \\n\\n-- \\nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\",\n \"msg_date\": \"Mon, 8 Nov 2021 08:12:52 -0800\",\n \"msg_from\": \"Michel Pelletier <pelletier.michel@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal for adding an example \\\"expanded\\\" data type to contrib\"\n }\n]","isConversational":false}}},{"rowIdx":53057,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Motivation:\n\nI'm working on a columnar compression AM[0]. Currently, it uses generic\nxlog, which works for crash recovery and physical replication, but not\nlogical decoding/replication.\n\nExtensible rmgr would enable the table AM to support its own\nredo/decode hooks and WAL format, so that it could support crash\nrecovery, physical replication, and logical replication.\n\nBackground:\n\nI submitted another patch[0] to add new logical records, which could be\nused to support logical decoding directly, without the need for\nextensible rmgr and without any assumptions about the table AM. This\nwas designed to be easy to use, but inefficient. Amit raised\nconcerns[1] about whether it could meet the needs of zheap. Andres\nsuggested (off-list) that it would be better to just tackle the\nextensible rmgr problem.\n\nThe idea for extensible rmgr has been proposed before[3]. The biggest\nargument against it seemed to be that there was no complete use\ncase[4], so the worry was that something would be left out. Columnar is\ncomplete enough that I think it qualifies as a good use case.\n\nA subsequent proposal[5] was shot down because of a (potential?) need\nfor catalog access[6]. The attached patch does not use the catalog;\ninstead, it relies on table AM authors choosing IDs that don't conflict\nwith each other. This seems like a reasonable answer, considering that\nthere will likely be very few table AMs that go far enough to fully\nsupport WAL including decoding.\n\nAre there any other major arguments/objections that I missed?\n\nProposal:\n\nThe attached patch (against v14, so it's easier to test columnar) is\nsomewhat like a simplified version of [3] combined with refactoring to\nmake decoding a part of the rmgr.\n\n * adds a new RmgrData method rm_decode\n * refactors decode.c to use\nthe new method\n * add a layer of indirection GetRmgr to find an rmgr\n \n * fast path to find builtin rmgr in RmgrTable\n * to find a custom\nrmgr, traverses list of custom rmgrs that\n are currently loaded\n(unlikely to ever be more than a few)\n * rmgr IDs from 0-127 are\nreserved for builtin rmgrs\n * rmgr IDs from 128-255 are reserved for\ncustom rmgrs\n * table AM authors need to avoid collisions between\nrmgr IDs\n\nI have tested with columnar using a simple WAL format for logical\ndecoding only, and I'm still using generic xlog for recovery and\nphysical replication. I haven't tested the redo path, or how easy it\nmight be to do something like generic xlog.\n\nQuestions:\n\n0. Do we want to go this route, or something simpler like my other\nproposal, which introduces new logical record types[0]?\n\n1. I am allocating the custom rmgr list in TopMemoryContext, and it\nonly works when loading as a part of shared_preload_libraries. This\navoids the need for shared memory in Simon's patch[3]. Is that the\nright thing to do?\n\n2. If we go this route, what do we do with generic xlog? It seems like\na half feature, since it doesn't work with logical decoding.\n\n3. If the custom rmgr throws an error during redo, the server won't\nstart. Should we have a GUC to turn non-builtin redo into a no-op to\nreduce the impact of bugs in the implementation of a custom rmgr?\n\n4. Do we want to encourage index AMs to use this mechanism as well? I\ndidn't really look into how suitable it is, but at a high level it\nseems reasonable.\n\nRegards,\n\tJeff Davis\n\n[0] \nhttps://postgr.es/m/20ee0b0ae6958804a88fe9580157587720faf664.camel@j-davis.com\n[1] \nhttps://postgr.es/m/CAA4eK1JVDnbQ80ULdZuhzQkzr_yYhfON-tg%3Dd1U7aWjK_R1ixQ%40mail.gmail.com\n[2] https://github.com/citusdata/citus/tree/master/src/backend/columnar\n[3] https://postgr.es/m/1229541840.4793.79.camel%40ebony.2ndQuadrant\n[4] https://postgr.es/m/20992.1232667957%40sss.pgh.pa.us\n[5] https://postgr.es/m/1266774840.7341.29872.camel%40ebony\n[6] https://postgr.es/m/26134.1266776040%40sss.pgh.pa.us","msg_date":"Mon, 08 Nov 2021 15:36:21 -0800","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":true,"msg_subject":"Extensible Rmgr for Table AMs"},{"msg_contents":"On Mon, 2021-11-08 at 15:36 -0800, Jeff Davis wrote:\n> The attached patch (against v14, so it's easier to test columnar) is\n> somewhat like a simplified version of [3] combined with refactoring\n> to\n> make decoding a part of the rmgr.\n\nI created a wiki page here: \nhttps://wiki.postgresql.org/wiki/ExtensibleRmgr\n\nTo coordinate reservation of RmgrIds, to avoid conflicts. I don't\nexpect it to be a practical problem given how much work it takes to\ncreate a new table AM that needs full WAL support, but might as well\nhave some transparency on how to choose a new RmgrId.\n\nI also updated the patch to point to the wiki page in the comments, and\nadded in a new RM_EXPERIMENTAL_ID that can be used while an extension\nis still in development. Hopefully this will prevent people reserving\nlots of RmgrIds for extensions that never get released.\n\nRegards,\n\tJeff Davis","msg_date":"Tue, 09 Nov 2021 17:38:29 -0800","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":true,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Mon, Nov 8, 2021 at 6:36 PM Jeff Davis <pgsql@j-davis.com> wrote:\n> Extensible rmgr would enable the table AM to support its own\n> redo/decode hooks and WAL format, so that it could support crash\n> recovery, physical replication, and logical replication.\n\nWithout taking a position on your implementation, which I have not\nstudied, I like this idea in concept and I think it's an important\ngoal.\n\n> Are there any other major arguments/objections that I missed?\n\nISTR some discussion of the fact that uninstalling the extension that\nuses this facility, or failing to install it on your standby, will\nlead to an unusable database. Personally, I don't see that as a big\nproblem: we should just document that if you choose to use an\nextension like this, then (1) it needs to be installed on all standbys\nand (2) if you ever want to get rid of it, you need to stop using it,\ndrop all the objects created with it, and then wait until all the WAL\npreviously generated by that extension is gone not only from pg_wal\nbut from any archived WAL files or backups that you intend to use with\nthat cluster before you actually nuke it. Users who don't want to\nabide by those restrictions need not install such extensions. Users\nwho don't read the documentation might end up sad, but it doesn't seem\nparticularly likely, and it's hardly the only part of the\ndocumentation that users shouldn't ignore.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 10 Nov 2021 09:58:14 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Mon, 2021-11-08 at 15:36 -0800, Jeff Davis wrote:\n> The attached patch (against v14, so it's easier to test columnar) is\n> somewhat like a simplified version of [3] combined with refactoring\n> to\n> make decoding a part of the rmgr.\n\nNew patches attached (v3). Essentially the content as v2, but split\ninto 3 patches and rebased.\n\nReview on patch 0001 would be helpful. It makes decoding just another\nmethod of rmgr, which makes a lot of sense to me from a code\norganization standpoint regardless of the other patches. Is there any\nreason not to do that?\n\nThe other patches then make rmgr extensible, which in turn makes\ndecoding extensible and solves the logical replication problem for\ncustom table AMs. The most obvious way to make rmgr extensible would be\nto expand the rmgr table, but I was concerned about making that dynamic\n(right now the structure is entirely constant and perhaps that's\nimportant for some optimizations?). So, at the cost of complexity I\nmade a separate, dynamic rmgr table to hold the custom entries.\n\nThe custom rmgr API would probably be marked \"experimental\" for a\nwhile, and I don't expect lots of people to use it given the challenges\nof a production-quality table AM. But it's still important, because\nwithout it a table AM has no chance to participate in logical\nreplication.\n\nRegards,\n\tJeff Davis","msg_date":"Mon, 17 Jan 2022 00:42:06 -0800","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":true,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"Hi,\n\nOn Mon, Jan 17, 2022 at 12:42:06AM -0800, Jeff Davis wrote:\n> On Mon, 2021-11-08 at 15:36 -0800, Jeff Davis wrote:\n> > The attached patch (against v14, so it's easier to test columnar) is\n> > somewhat like a simplified version of [3] combined with refactoring\n> > to\n> > make decoding a part of the rmgr.\n> \n> New patches attached (v3). Essentially the content as v2, but split\n> into 3 patches and rebased.\n> \n> Review on patch 0001 would be helpful. It makes decoding just another\n> method of rmgr, which makes a lot of sense to me from a code\n> organization standpoint regardless of the other patches. Is there any\n> reason not to do that?\n\nI think that it's a clean and sensible approach, so +1 for me.\n\nThere's a bit of 0003 present in 002:\n\ndiff --git a/src/backend/access/transam/xlogreader.c b/src/backend/access/transam/xlogreader.c\nindex 35029cf97d..612b1b3723 100644\n--- a/src/backend/access/transam/xlogreader.c\n+++ b/src/backend/access/transam/xlogreader.c\n@@ -738,7 +738,7 @@ ValidXLogRecordHeader(XLogReaderState *state, XLogRecPtr RecPtr,\n (uint32) SizeOfXLogRecord, record->xl_tot_len);\n return false;\n }\n- if (record->xl_rmid > RM_MAX_ID)\n+ if (record->xl_rmid > RM_MAX_ID && record->xl_rmid < RM_CUSTOM_MIN_ID)\n {\n\n> \n> The other patches then make rmgr extensible, which in turn makes\n> decoding extensible and solves the logical replication problem for\n> custom table AMs. The most obvious way to make rmgr extensible would be\n> to expand the rmgr table, but I was concerned about making that dynamic\n> (right now the structure is entirely constant and perhaps that's\n> important for some optimizations?). So, at the cost of complexity I\n> made a separate, dynamic rmgr table to hold the custom entries.\n\nI'm not sure about performance overhead, but allowing extension to access the\nmain table directly seems a bit scary. We definitely accepted some overhead\nwith the various extensible support, and this new GetRmgr() doesn't look like\nit's going to cost a lot for the builtin case, especially compared to the cost\nof any of the code associated with the rmgr.\n\nAlso, to answer your initial email I think it's a better way to go compared to\nyour previous approach, given the limitations and performance of the generic\nxlog infrastructure, and hopefully index AMs should be able to rely on this\ntoo.\n\nA few random comments on 0003:\n\n #define RM_MAX_ID (RM_NEXT_ID - 1)\n+#define RM_CUSTOM_MIN_ID 128\n+#define RM_CUSTOM_MAX_ID UINT8_MAX\n\nIt would be a good idea to add a StaticAssertStmt here to make sure that\nthere's no overlap in the ranges.\n\n+\n+/*\n+ * RmgrId to use for extensions that require an RmgrId, but are still in\n+ * development and have not reserved their own unique RmgrId yet. See:\n+ * https://wiki.postgresql.org/wiki/ExtensibleRmgr\n+ */\n+#define RM_EXPERIMENTAL_ID 128\n\nI'm a bit dubious about the whole \"register your ID in the Wiki\" approach. It\nmight not be a problem since there probably won't be hundreds of users, and I\ndon't have any better suggestion since it has to be consistent across nodes.\n\n+ elog(LOG, \"registering customer rmgr \\\"%s\\\" with ID %d\",\n+ rmgr->rm_name, rmid);\n\nShould it be a DEBUG message instead? Also s/customer/custom/\n\n+RmgrData\n+GetCustomRmgr(RmgrId rmid)\n+{\n+ if (rmid < RM_CUSTOM_MIN_ID || rmid > RM_CUSTOM_MAX_ID)\n+ elog(PANIC, \"custom rmgr id %d out of range\", rmid);\n\nShould this be an assert?\n\n+#define GetBuiltinRmgr(rmid) RmgrTable[(rmid)]\n+#define GetRmgr(rmid) ((rmid < RM_CUSTOM_MIN_ID) ? \\\n+ GetBuiltinRmgr(rmid) : GetCustomRmgr(rmid))\n\nrmid should be protected in the macro.\n\n> The custom rmgr API would probably be marked \"experimental\" for a\n> while, and I don't expect lots of people to use it given the challenges\n> of a production-quality table AM. But it's still important, because\n> without it a table AM has no chance to participate in logical\n> replication.\n\nHow you plan to mark it experimental?\n\n\n","msg_date":"Tue, 18 Jan 2022 17:53:32 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Tue, 2022-01-18 at 17:53 +0800, Julien Rouhaud wrote:\n> > Review on patch 0001 would be helpful. It makes decoding just\n> > another\n> > method of rmgr, which makes a lot of sense to me from a code\n> > organization standpoint regardless of the other patches. Is there\n> > any\n> > reason not to do that?\n> \n> I think that it's a clean and sensible approach, so +1 for me.\n\nThank you, committed 0001. Other patches not committed yet.\n\n> There's a bit of 0003 present in 002:\n\nI just squashed 0002 and 0003 together. Not large enough to keep\nseparate.\n\n> A few random comments on 0003:\n> \n> #define RM_MAX_ID (RM_NEXT_ID - 1)\n> +#define RM_CUSTOM_MIN_ID 128\n> +#define RM_CUSTOM_MAX_ID UINT8_MAX\n> \n> It would be a good idea to add a StaticAssertStmt here to make sure\n> that\n> there's no overlap in the ranges.\n\nDone.\n\n> +\n> +/*\n> + * RmgrId to use for extensions that require an RmgrId, but are\n> still in\n> + * development and have not reserved their own unique RmgrId yet.\n> See:\n> + * https://wiki.postgresql.org/wiki/ExtensibleRmgr\n> + */\n> +#define RM_EXPERIMENTAL_ID 128\n> \n> I'm a bit dubious about the whole \"register your ID in the Wiki\"\n> approach. It\n> might not be a problem since there probably won't be hundreds of\n> users, and I\n> don't have any better suggestion since it has to be consistent across\n> nodes.\n\nAgree, but I don't see a better approach, either.\n\nI do some sanity checking, which should catch collisions when they\nhappen.\n\n> + elog(LOG, \"registering customer rmgr \\\"%s\\\" with ID %d\",\n> + rmgr->rm_name, rmid);\n> \n> Should it be a DEBUG message instead? Also s/customer/custom/\n\nIt seems like a fairly important thing to have in the log. Only a\ncouple extensions will ever encounter this message, and only at server\nstart.\n\nTypo fixed.\n\n> +RmgrData\n> +GetCustomRmgr(RmgrId rmid)\n> +{\n> + if (rmid < RM_CUSTOM_MIN_ID || rmid > RM_CUSTOM_MAX_ID)\n> + elog(PANIC, \"custom rmgr id %d out of range\", rmid);\n> \n> Should this be an assert?\n\nIf we make it an Assert, then it won't be caught in production builds.\n\n> +#define GetBuiltinRmgr(rmid) RmgrTable[(rmid)]\n> +#define GetRmgr(rmid) ((rmid < RM_CUSTOM_MIN_ID) ? \\\n> + GetBuiltinRmgr(rmid) : GetCustomRmgr(rmid))\n> \n> rmid should be protected in the macro.\n\nDone.\n\n> How you plan to mark it experimental?\n\nI suppose it doesn't need to be marked explicitly -- there are other\nAPIs that change. For instance, the ProcessUtility_hook changed, and\nthat's used much more widely.\n\nAs long as we generally agree that some kind of custom rmgrs are the\nway to go, if we change the API or implementation around from version\nto version I can easily update my table access method.\n\nRegards,\n\tJeff Davis","msg_date":"Mon, 31 Jan 2022 18:36:59 -0800","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":true,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"Hi,\n\nOn Mon, Jan 31, 2022 at 06:36:59PM -0800, Jeff Davis wrote:\n> On Tue, 2022-01-18 at 17:53 +0800, Julien Rouhaud wrote:\n> \n> > There's a bit of 0003 present in 002:\n> \n> I just squashed 0002 and 0003 together. Not large enough to keep\n> separate.\n\nAgreed.\n\n> > + elog(LOG, \"registering customer rmgr \\\"%s\\\" with ID %d\",\n> > + rmgr->rm_name, rmid);\n> > \n> > Should it be a DEBUG message instead? Also s/customer/custom/\n> \n> It seems like a fairly important thing to have in the log. Only a\n> couple extensions will ever encounter this message, and only at server\n> start.\n\nOk.\n\n> > +RmgrData\n> > +GetCustomRmgr(RmgrId rmid)\n> > +{\n> > + if (rmid < RM_CUSTOM_MIN_ID || rmid > RM_CUSTOM_MAX_ID)\n> > + elog(PANIC, \"custom rmgr id %d out of range\", rmid);\n> > \n> > Should this be an assert?\n> \n> If we make it an Assert, then it won't be caught in production builds.\n\nSure, but it seems something so fundamental that it should get done right\nduring the development phase rather than spending cycles in optimized builds to\ncheck for it.\nBut even if that happened it would get caught by the final \"not found\" PANIC\nanyway, and for end users I don't think that getting this error instead would\nmake much difference.\n\n> > How you plan to mark it experimental?\n> \n> I suppose it doesn't need to be marked explicitly -- there are other\n> APIs that change. For instance, the ProcessUtility_hook changed, and\n> that's used much more widely.\n> As long as we generally agree that some kind of custom rmgrs are the\n> way to go, if we change the API or implementation around from version\n> to version I can easily update my table access method.\n\nAgreed, API breaks happen often and that's not really a problem for third-party\nextensions, especially if that comes with hard compile failure.\n\nOther than that the patch looks good to me, as you said we just need a decision\non whether custom rmgrs are wanted or not.\n\n\n","msg_date":"Tue, 1 Feb 2022 12:39:38 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Tue, Feb 01, 2022 at 12:39:38PM +0800, Julien Rouhaud wrote:\n> \n> Other than that the patch looks good to me, as you said we just need a decision\n> on whether custom rmgrs are wanted or not.\n\nOne last thing, did you do some benchmark with a couple custom rmgr to see how\nmuch the O(n) access is showing up in profiles?\n\n\n","msg_date":"Tue, 1 Feb 2022 20:45:50 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Tue, 2022-02-01 at 20:45 +0800, Julien Rouhaud wrote:\n> On Tue, Feb 01, 2022 at 12:39:38PM +0800, Julien Rouhaud wrote:\n> > \n> > Other than that the patch looks good to me, as you said we just\n> > need a decision\n> > on whether custom rmgrs are wanted or not.\n> \n> One last thing, did you do some benchmark with a couple custom rmgr\n> to see how\n> much the O(n) access is showing up in profiles?\n\nWhat kind of a test case would be reasonable there? You mean having a\nlot of custom rmgrs?\n\nI was expecting that few people would have more than one custom rmgr\nloaded anyway, so a sparse array or hashtable seemed wasteful. If\ncustom rmgrs become popular we probably need to have a larger ID space\nanyway, but it seems like overengineering to do so now.\n\nRegards,\n\tJeff Davis\n\n\n\n\n","msg_date":"Tue, 01 Feb 2022 15:38:32 -0800","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":true,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"Hi,\n\nOn Tue, Feb 01, 2022 at 03:38:32PM -0800, Jeff Davis wrote:\n> On Tue, 2022-02-01 at 20:45 +0800, Julien Rouhaud wrote:\n> > \n> > One last thing, did you do some benchmark with a couple custom rmgr\n> > to see how\n> > much the O(n) access is showing up in profiles?\n> \n> What kind of a test case would be reasonable there? You mean having a\n> lot of custom rmgrs?\n> \n> I was expecting that few people would have more than one custom rmgr\n> loaded anyway, so a sparse array or hashtable seemed wasteful. If\n> custom rmgrs become popular we probably need to have a larger ID space\n> anyway, but it seems like overengineering to do so now.\n\nI agree that having dozen of custom rmgrs doesn't seem likely, but I also have\nno idea of how much overhead you get by not doing a direct array access. I\nthink it would be informative to benchmark something like simple OLTP write\nworkload on a fast storage (or a ramdisk, or with fsync off...), with the used\nrmgr being the 1st and the 2nd custom rmgr. Both scenario still seems\nplausible and shouldn't degenerate on good hardware.\n\n\n","msg_date":"Thu, 3 Feb 2022 13:34:18 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Thu, Feb 3, 2022 at 12:34 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> I agree that having dozen of custom rmgrs doesn't seem likely, but I also have\n> no idea of how much overhead you get by not doing a direct array access. I\n> think it would be informative to benchmark something like simple OLTP write\n> workload on a fast storage (or a ramdisk, or with fsync off...), with the used\n> rmgr being the 1st and the 2nd custom rmgr. Both scenario still seems\n> plausible and shouldn't degenerate on good hardware.\n\nI think it would be hard to measure the overhead of this approach on a\nmacrobenchmark. That having been said, I find this a surprising\nimplementation choice. I think that the approaches that are most worth\nconsidering are:\n\n(1) reallocate the array if needed so that we can continue to just do\nRmgrTable[rmid]\n(2) have one array for builtins and a second array for extensions and\ndo rmid < RM_CUSTOM_MIN_ID ? BuiltinRmgrTable[rmid] :\nExtensionRmgrTable[rmid]\n(3) change RmgrTable to be an array of pointers to structs rather than\nan an array of structs. then the structs don't move around and can be\nconst, but the pointers can be moved into a larger array if required\n\nI'm not really sure which is best. My intuition for what will be\ncheapest on modern hardware is pretty shaky. However, I can't see how\nit can be the thing the patch is doing now; a linear search seems like\nit has to be the slowest option.\n\n--\nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 4 Feb 2022 09:10:42 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"Hi,\n\nOn Fri, Feb 04, 2022 at 09:10:42AM -0500, Robert Haas wrote:\n> On Thu, Feb 3, 2022 at 12:34 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> > I agree that having dozen of custom rmgrs doesn't seem likely, but I also have\n> > no idea of how much overhead you get by not doing a direct array access. I\n> > think it would be informative to benchmark something like simple OLTP write\n> > workload on a fast storage (or a ramdisk, or with fsync off...), with the used\n> > rmgr being the 1st and the 2nd custom rmgr. Both scenario still seems\n> > plausible and shouldn't degenerate on good hardware.\n> \n> I think it would be hard to measure the overhead of this approach on a\n> macrobenchmark.\n\nYeah that's also my initial thought, but I wouldn't be terribly surprised to be\nwrong.\n\n> That having been said, I find this a surprising\n> implementation choice. I think that the approaches that are most worth\n> considering are:\n> \n> (1) reallocate the array if needed so that we can continue to just do\n> RmgrTable[rmid]\n> (2) have one array for builtins and a second array for extensions and\n> do rmid < RM_CUSTOM_MIN_ID ? BuiltinRmgrTable[rmid] :\n> ExtensionRmgrTable[rmid]\n> (3) change RmgrTable to be an array of pointers to structs rather than\n> an an array of structs. then the structs don't move around and can be\n> const, but the pointers can be moved into a larger array if required\n> \n> I'm not really sure which is best. My intuition for what will be\n> cheapest on modern hardware is pretty shaky. However, I can't see how\n> it can be the thing the patch is doing now; a linear search seems like\n> it has to be the slowest option.\n\nI guess the idea was to have a compromise between letting rmgr authors choose\narbitrary ids to avoid any conflicts, especially with private implementations,\nwithout wasting too much memory. But those approaches would be pretty much\nincompatible with the current definition:\n\n+#define RM_CUSTOM_MIN_ID 128\n+#define RM_CUSTOM_MAX_ID UINT8_MAX\n\neven if you only allocate up to the max id found, nothing guarantees that you\nwon't get a quite high id.\n\n\n","msg_date":"Fri, 4 Feb 2022 22:48:01 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Fri, Feb 4, 2022 at 9:48 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> I guess the idea was to have a compromise between letting rmgr authors choose\n> arbitrary ids to avoid any conflicts, especially with private implementations,\n> without wasting too much memory. But those approaches would be pretty much\n> incompatible with the current definition:\n>\n> +#define RM_CUSTOM_MIN_ID 128\n> +#define RM_CUSTOM_MAX_ID UINT8_MAX\n>\n> even if you only allocate up to the max id found, nothing guarantees that you\n> won't get a quite high id.\n\nRight, which I guess raises another question: if the maximum is\nUINT8_MAX, which BTW I find perfectly reasonable, why are we not just\ndefining this as an array of size 256? There's no point in adding code\ncomplexity to save a few kB of memory.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 4 Feb 2022 09:53:09 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Fri, Feb 04, 2022 at 09:53:09AM -0500, Robert Haas wrote:\n> On Fri, Feb 4, 2022 at 9:48 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> > I guess the idea was to have a compromise between letting rmgr authors choose\n> > arbitrary ids to avoid any conflicts, especially with private implementations,\n> > without wasting too much memory. But those approaches would be pretty much\n> > incompatible with the current definition:\n> >\n> > +#define RM_CUSTOM_MIN_ID 128\n> > +#define RM_CUSTOM_MAX_ID UINT8_MAX\n> >\n> > even if you only allocate up to the max id found, nothing guarantees that you\n> > won't get a quite high id.\n> \n> Right, which I guess raises another question: if the maximum is\n> UINT8_MAX, which BTW I find perfectly reasonable, why are we not just\n> defining this as an array of size 256? There's no point in adding code\n> complexity to save a few kB of memory.\n\nAgreed, especially if combined with your suggested approach 3 (array of\npointers).\n\n\n","msg_date":"Fri, 4 Feb 2022 22:56:58 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Fri, 2022-02-04 at 22:56 +0800, Julien Rouhaud wrote:\n> > Right, which I guess raises another question: if the maximum is\n> > UINT8_MAX, which BTW I find perfectly reasonable, why are we not\n> > just\n> > defining this as an array of size 256? There's no point in adding\n> > code\n> > complexity to save a few kB of memory.\n> \n> Agreed, especially if combined with your suggested approach 3 (array\n> of\n> pointers).\n\nImplemented and attached. I also updated pg_waldump and pg_rewind to do\nsomething reasonable.\n\nAdditionally, I now have a reasonably complete implementation of a\ncustom resource manager now:\n\nhttps://github.com/citusdata/citus/tree/custom-rmgr-15\n\n(Not committed or intended to actually be used right now -- just a\nPOC.)\n\nOffline, Andres mentioned that I should test recovery performance if we\ntake your approach, because making the RmgrTable non-const could impact\noptimizations. Not sure if that would be a practical concern compared\nto all the other work done at REDO time.\n\nRegards,\n\tJeff Davis","msg_date":"Wed, 23 Mar 2022 21:43:08 -0700","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":true,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"Hi,\n\nOn 2022-03-23 21:43:08 -0700, Jeff Davis wrote:\n> /* must be kept in sync with RmgrData definition in xlog_internal.h */\n> #define PG_RMGR(symname,name,redo,desc,identify,startup,cleanup,mask,decode) \\\n> -\t{ name, redo, desc, identify, startup, cleanup, mask, decode },\n> +\t&(struct RmgrData){ name, redo, desc, identify, startup, cleanup, mask, decode },\n> \n> -const RmgrData RmgrTable[RM_MAX_ID + 1] = {\n> +static RmgrData *RmgrTable[RM_MAX_ID + 1] = {\n> #include \"access/rmgrlist.h\"\n> };\n\nI think this has been discussed before, but to me it's not obvious that it's a\ngood idea to change RmgrTable from RmgrData to RmgrData *. That adds an\nindirection, without obvious benefit.\n\n\n> +\n> +/*\n> + * Start up all resource managers.\n> + */\n> +void\n> +StartupResourceManagers()\n\n(void)\n\n\n> +void\n> +RegisterCustomRmgr(RmgrId rmid, RmgrData *rmgr)\n> +{\n> +\tif (rmid < RM_MIN_CUSTOM_ID)\n> +\t\tereport(PANIC, errmsg(\"custom rmgr id %d is out of range\", rmid));\n> +\n> +\tif (!process_shared_preload_libraries_in_progress)\n> +\t\tereport(ERROR,\n> +\t\t\t\t(errmsg(\"custom rmgr must be registered while initializing modules in shared_preload_libraries\")));\n> +\n> +\tereport(LOG,\n> +\t\t\t(errmsg(\"registering custom rmgr \\\"%s\\\" with ID %d\",\n> +\t\t\t\t\trmgr->rm_name, rmid)));\n> +\n> +\tif (RmgrTable[rmid] != NULL)\n> +\t\tereport(PANIC,\n> +\t\t\t\t(errmsg(\"custom rmgr ID %d already registered with name \\\"%s\\\"\",\n> +\t\t\t\t\t\trmid, RmgrTable[rmid]->rm_name)));\n> +\n> +\t/* check for existing rmgr with the same name */\n> +\tfor (int i = 0; i <= RM_MAX_ID; i++)\n> +\t{\n> +\t\tconst RmgrData *existing_rmgr = RmgrTable[i];\n> +\n> +\t\tif (existing_rmgr == NULL)\n> +\t\t\tcontinue;\n> +\n> +\t\tif (!strcmp(existing_rmgr->rm_name, rmgr->rm_name))\n> +\t\t\tereport(PANIC,\n> +\t\t\t\t\t(errmsg(\"custom rmgr \\\"%s\\\" has the same name as builtin rmgr\",\n> +\t\t\t\t\t\t\texisting_rmgr->rm_name)));\n> +\t}\n> +\n> +\t/* register it */\n> +\tRmgrTable[rmid] = rmgr;\n> +}\n\nRandom idea: Might be worth emitting the id->name mapping just after a redo\nlocation is determined, to make it easier to debug things.\n\n\n> +RmgrData *\n> +GetRmgr(RmgrId rmid)\n> +{\n> +\treturn RmgrTable[rmid];\n> +}\n\nGiven this is so simple, why incur the cost of a function call? Rather than\ncontinuing to expose RmgrTable and move GetRmgr() into the header, as a static\ninline? As-is this also prevent the compiler from optimizing across repeated\nGetRmgr() calls (which often won't be possible anyway, but still)..\n\n\n> -\tif (record->xl_rmid > RM_MAX_ID)\n> +\tif (record->xl_rmid > RM_MAX_BUILTIN_ID && record->xl_rmid < RM_MIN_CUSTOM_ID)\n> \t{\n> \t\treport_invalid_record(state,\n> \t\t\t\t\t\t\t \"invalid resource manager ID %u at %X/%X\",\n\nShouldn't this continue to enforce RM_MAX_ID as well?\n\n\n> @@ -1604,12 +1603,7 @@ PerformWalRecovery(void)\n> \n> \t\tInRedo = true;\n> \n> -\t\t/* Initialize resource managers */\n> -\t\tfor (rmid = 0; rmid <= RM_MAX_ID; rmid++)\n> -\t\t{\n> -\t\t\tif (RmgrTable[rmid].rm_startup != NULL)\n> -\t\t\t\tRmgrTable[rmid].rm_startup();\n> -\t\t}\n> +\t\tStartupResourceManagers();\n\nPersonally I'd rather name it ResourceManagersStartup() or RmgrStartup().\n\n\n> @@ -1871,7 +1860,7 @@ ApplyWalRecord(XLogReaderState *xlogreader, XLogRecord *record, TimeLineID *repl\n> \t\txlogrecovery_redo(xlogreader, *replayTLI);\n> \n> \t/* Now apply the WAL record itself */\n> -\tRmgrTable[record->xl_rmid].rm_redo(xlogreader);\n> +\tGetRmgr(record->xl_rmid)->rm_redo(xlogreader);\n> \n> \t/*\n> \t * After redo, check whether the backup pages associated with the WAL\n\nSo we have just added one indirect call and one pointer indirection\n(previously RmgrTable could be resolved by the linker, now it needs to be\ndereferenced again), that's not too bad. I was afraid there'd be multiple\ncalls.\n\n\n> @@ -2101,16 +2090,16 @@ xlog_outdesc(StringInfo buf, XLogReaderState *record)\n> \tuint8\t\tinfo = XLogRecGetInfo(record);\n> \tconst char *id;\n> \n> -\tappendStringInfoString(buf, RmgrTable[rmid].rm_name);\n> +\tappendStringInfoString(buf, GetRmgr(rmid)->rm_name);\n> \tappendStringInfoChar(buf, '/');\n> \n> -\tid = RmgrTable[rmid].rm_identify(info);\n> +\tid = GetRmgr(rmid)->rm_identify(info);\n> \tif (id == NULL)\n> \t\tappendStringInfo(buf, \"UNKNOWN (%X): \", info & ~XLR_INFO_MASK);\n> \telse\n> \t\tappendStringInfo(buf, \"%s: \", id);\n> \n> -\tRmgrTable[rmid].rm_desc(buf, record);\n> +\tGetRmgr(rmid)->rm_desc(buf, record);\n> }\n\nLike here. It's obviously not as performance critical as replay. But it's\nstill a shame to add 3 calls to GetRmgr, that each then need to dereference\nRmgrTable. The compiler won't be able to optimize any of that away.\n\n\n> @@ -117,8 +117,8 @@ LogicalDecodingProcessRecord(LogicalDecodingContext *ctx, XLogReaderState *recor\n> \n> \trmid = XLogRecGetRmid(record);\n> \n> -\tif (RmgrTable[rmid].rm_decode != NULL)\n> -\t\tRmgrTable[rmid].rm_decode(ctx, &buf);\n> +\tif (GetRmgr(rmid)->rm_decode != NULL)\n> +\t\tGetRmgr(rmid)->rm_decode(ctx, &buf);\n> \telse\n> \t{\n> \t\t/* just deal with xid, and done */\n\nThis one might actually matter, overhead wise.\n\n\n> @@ -473,7 +473,7 @@ static void\n> XLogDumpDisplayRecord(XLogDumpConfig *config, XLogReaderState *record)\n> {\n> \tconst char *id;\n> -\tconst RmgrDescData *desc = &RmgrDescTable[XLogRecGetRmid(record)];\n> +\tconst RmgrDescData *desc = GetRmgrDesc(XLogRecGetRmid(record));\n> \tuint32\t\trec_len;\n> \tuint32\t\tfpi_len;\n> \tRelFileNode rnode;\n> @@ -658,7 +658,7 @@ XLogDumpDisplayStats(XLogDumpConfig *config, XLogDumpStats *stats)\n> \t * calculate column totals.\n> \t */\n> \n> -\tfor (ri = 0; ri < RM_NEXT_ID; ri++)\n> +\tfor (ri = 0; ri < RM_MAX_ID; ri++)\n> \t{\n> \t\ttotal_count += stats->rmgr_stats[ri].count;\n> \t\ttotal_rec_len += stats->rmgr_stats[ri].rec_len;\n\n> @@ -694,6 +694,9 @@ XLogDumpDisplayStats(XLogDumpConfig *config, XLogDumpStats *stats)\n> \t\t\tfpi_len = stats->rmgr_stats[ri].fpi_len;\n> \t\t\ttot_len = rec_len + fpi_len;\n> \n> +\t\t\tif (ri > RM_MAX_BUILTIN_ID && count == 0)\n> +\t\t\t\tcontinue;\n> +\n\nAh, I see. I had written a concerned comment about the previous hunk...\n\n\n> \t\t\tXLogDumpStatsRow(desc->rm_name,\n> \t\t\t\t\t\t\t count, total_count, rec_len, total_rec_len,\n> \t\t\t\t\t\t\t fpi_len, total_fpi_len, tot_len, total_len);\n> @@ -913,16 +916,16 @@ main(int argc, char **argv)\n> \t\t\t\t\t\texit(EXIT_SUCCESS);\n> \t\t\t\t\t}\n> \n> -\t\t\t\t\tfor (i = 0; i <= RM_MAX_ID; i++)\n> +\t\t\t\t\tfor (i = 0; i <= RM_MAX_BUILTIN_ID; i++)\n> \t\t\t\t\t{\n> -\t\t\t\t\t\tif (pg_strcasecmp(optarg, RmgrDescTable[i].rm_name) == 0)\n> +\t\t\t\t\t\tif (pg_strcasecmp(optarg, GetRmgrDesc(i)->rm_name) == 0)\n> \t\t\t\t\t\t{\n> \t\t\t\t\t\t\tconfig.filter_by_rmgr[i] = true;\n> \t\t\t\t\t\t\tconfig.filter_by_rmgr_enabled = true;\n> \t\t\t\t\t\t\tbreak;\n> \t\t\t\t\t\t}\n> \t\t\t\t\t}\n> -\t\t\t\t\tif (i > RM_MAX_ID)\n> +\t\t\t\t\tif (i > RM_MAX_BUILTIN_ID)\n> \t\t\t\t\t{\n> \t\t\t\t\t\tpg_log_error(\"resource manager \\\"%s\\\" does not exist\",\n> \t\t\t\t\t\t\t\t\t optarg);\n\nSo we can't filter by rmgr id for non-builtin rmgr's? That sucks. Maybe add a\ncustom:<i> syntax? Or generally allow numerical identifiers in addition to the\nnames?\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 28 Mar 2022 11:00:45 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Mon, 2022-03-28 at 11:00 -0700, Andres Freund wrote:\n> I think this has been discussed before, but to me it's not obvious\n> that it's a\n> good idea to change RmgrTable from RmgrData to RmgrData *. That adds\n> an\n> indirection, without obvious benefit.\n\nI did some performance tests. I created a narrow table, took a base\nbackup, loaded 100M records, finished the base backup. Then I recovered\nusing the different build combinations (patched/unpatched, clang/gcc).\n\n compiler run1 run2\n unpatched: gcc 102s 106s\n patched: gcc 107s 105s\n unpatched: clang 109s 110s\n patched: clang 109s 111s\n\nI don't see a clear signal that this patch worsens performance. The\n102s run was the very first run, so I suspect it was just due to the\nprocessor starting out cold. Let me know if you think the test is\ntesting the right paths.\n\nPerhaps I should address your other perf concerns around GetRmgr (make\nit static inline, reduce number of calls), and then leave the\nindirection for the sake of cleanliness?\n\nIf you are still concerned, I can switch back to separate tables to\neliminate the indirection for built-in rmgrs. Separate rmgr tables\nstill require a branch (to decide which table to access), but it should\nbe a highly predictable one.\n\nRegards,\n\tJeff Davis\n\n\n\n\n","msg_date":"Thu, 31 Mar 2022 14:20:51 -0700","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":true,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Mon, 2022-03-28 at 11:00 -0700, Andres Freund wrote:\n> > +/*\n> > + * Start up all resource managers.\n> > + */\n> > +void\n> > +StartupResourceManagers()\n> \n> (void)\n\nFixed.\n\n> Random idea: Might be worth emitting the id->name mapping just after\n> a redo\n> location is determined, to make it easier to debug things.\n\nNot quite sure I understood this idea, do you mean dump all rmgrs and\nthe IDs when performing recovery?\n\n> > +RmgrData *\n> > +GetRmgr(RmgrId rmid)\n> > +{\n> > +\treturn RmgrTable[rmid];\n> > +}\n> \n> Given this is so simple, why incur the cost of a function call?\n> Rather than\n> continuing to expose RmgrTable and move GetRmgr() into the header, as\n> a static\n> inline?\n\nMade it static inline.\n\n> Shouldn't this continue to enforce RM_MAX_ID as well?\n\nDone.\n\n> Personally I'd rather name it ResourceManagersStartup() or\n> RmgrStartup().\n\nDone.\n\n> Like here. It's obviously not as performance critical as replay. But\n> it's\n> still a shame to add 3 calls to GetRmgr, that each then need to\n> dereference\n> RmgrTable. The compiler won't be able to optimize any of that away.\n\nChanged to only call it once and save it in a variable for each call\nsite where it makes sense.\n\n> So we can't filter by rmgr id for non-builtin rmgr's? That sucks.\n> Maybe add a\n> custom:<i> syntax? Or generally allow numerical identifiers in\n> addition to the\n> names?\n\nGood idea. I changed it to allow \"custom###\" to mean the custom rmgr\nwith ID ### (3 digits).\n\nI still may change it to go back to two RmgrTables (one for builtin and\none for custom) to remove the lingering performance doubts. Other than\nthat, and some cleanup, this is pretty close to the version I intend to\ncommit.\n\nRegards,\n\tJeff Davis","msg_date":"Sun, 03 Apr 2022 20:03:08 -0700","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":true,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"Hi,\n\nOn 2022-03-31 14:20:51 -0700, Jeff Davis wrote:\n> If you are still concerned, I can switch back to separate tables to\n> eliminate the indirection for built-in rmgrs. Separate rmgr tables\n> still require a branch (to decide which table to access), but it should\n> be a highly predictable one.\n\nI still think the easiest and fastest would be to just make RmgrTable longer,\nand not const. When registering, copy the caller provided struct into the\nrespective RmgrData element. Yes, we'd waste a bit of space at the end of the\narray, but it's typically not going to be touched and thus not be backed by\n\"actual\" memory.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sun, 3 Apr 2022 21:33:37 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Mon, Apr 4, 2022 at 8:33 AM Jeff Davis <pgsql@j-davis.com> wrote:\n>\n> I still may change it to go back to two RmgrTables (one for builtin and\n> one for custom) to remove the lingering performance doubts. Other than\n> that, and some cleanup, this is pretty close to the version I intend to\n> commit.\n\nI just had a quick look over the v6 patch, few comments:\n\n1) Why can't rmid be chosen by the extensions in sequential order from\n(129 till 255), say, to start with a columnar extension can choose\n129, another extension can register 130 and so on right? This way, the\nRmgrStartup, RmgrCleanup, and XLogDumpDisplayRecord can avoid looping\nover all the 256 entries, with a global variable like\ncurrent_max_rmid(representing the current number of rmgers)? Is there\nany specific reason to allow them to choose rmgrid randomly between\n129 till 255? Am I missing some discussion here?\n2) RM_MAX_ID is now UINT8_MAX - do we need to make it configurable? or\ndo we think that only a few (127) custom rmgrs can exist?\n3) Do we need to talk about extensible rmgrs and\nhttps://wiki.postgresql.org/wiki/ExtensibleRmgr in documentation\nsomewhere? This will help extension developers to refer when needed.\n4) Do we need to change this description?\n <para>\n The default value of this setting is the empty string, which disables\n the feature. It can be set to <literal>all</literal> to check all\n records, or to a comma-separated list of resource managers to check\n only records originating from those resource managers. Currently,\n the supported resource managers are <literal>heap</literal>,\n <literal>heap2</literal>, <literal>btree</literal>,\n<literal>hash</literal>,\n <literal>gin</literal>, <literal>gist</literal>,\n<literal>sequence</literal>,\n <literal>spgist</literal>, <literal>brin</literal>, and\n<literal>generic</literal>. Onl\n superusers can change this setting.\n5) Do we need to add a PGDLLIMPORT qualifier to RegisterCustomRmgr()\n(possibly for RmgrStartup, RmgrTable, RmgrCleanup as well?) so that\nthe Windows versions of extensions can use this feature?\n6) Should the below strcmp pg_strcmp? The custom rmgrs can still use\nrm_name with different cases and below code fail to catch it?\n+ if (!strcmp(existing_rmgr->rm_name, rmgr->rm_name))\n+ ereport(PANIC,\n+ (errmsg(\"custom rmgr \\\"%s\\\" has the same name as builtin rmgr\",\n+ existing_rmgr->rm_name)));\n7) What's the intention of the below code? It seems like below it\nchecks if there's already a rmgr with the given name (RM_MAX_ID). Had\nit been RM_MAX_BUILTIN_ID instead of RM_MAX_ID, the error message\ndoes make sense. Is the intention here to not have duplicate rmgrs in\nthe entire RM_MAX_ID(both builtin and custom)?\n+ /* check for existing rmgr with the same name */\n+ for (int i = 0; i <= RM_MAX_ID; i++)\n+ {\n+ const RmgrData *existing_rmgr = RmgrTable[i];\n+\n+ if (existing_rmgr == NULL)\n+ continue;\n+\n+ if (!strcmp(existing_rmgr->rm_name, rmgr->rm_name))\n+ ereport(PANIC,\n+ (errmsg(\"custom rmgr \\\"%s\\\" has the same name as builtin rmgr\",\n+ existing_rmgr->rm_name)));\n8) Per https://wiki.postgresql.org/wiki/ExtensibleRmgr, 128 is\nreserved, can we have it as a macro? Or something like (RM_MAX_ID/2+1)\n+#define RM_MIN_CUSTOM_ID 128\n9) Thinking if there's a way to test the code, maybe exposing\nRegisterCustomRmgr as a function? I think we need to have a dummy\nextension that will be used for testing this kind of patches/features\nbut that's a separate discussion IMO.\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Mon, 4 Apr 2022 11:15:23 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Sun, 2022-04-03 at 21:33 -0700, Andres Freund wrote:\n> I still think the easiest and fastest would be to just make RmgrTable\n> longer,\n> and not const. When registering, copy the caller provided struct into\n> the\n> respective RmgrData element. Yes, we'd waste a bit of space at the\n> end of the\n> array, but it's typically not going to be touched and thus not be\n> backed by\n> \"actual\" memory.\n\nSounds good to me. I tried to break down the performance between these\napproaches and didn't get a clear signal, so I'll go with your\nintuition here.\n\nPosting new patch in response to Bharath's review, which will include\nthis change.\n\nNote that GetRmgr() also has an unlikely branch where it tests for the\nvalidity of the rmgr before using it, so that it can throw a nice error\nmessage if someone forgot to include the module in\nshared_preload_libraries. I expect this will be highly predictable and\ntherefore not a problem.\n\nRegards,\n\tJeff Davis\n\n\n\n\n","msg_date":"Mon, 04 Apr 2022 17:10:51 -0700","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":true,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Mon, 2022-04-04 at 11:15 +0530, Bharath Rupireddy wrote:\n> 1) Why can't rmid be chosen by the extensions in sequential order\n> from\n> (129 till 255), say, to start with a columnar extension can choose\n> 129, another extension can register 130 and so on right?\n\nI'm not sure what you mean by \"chosen by the extensions in sequential\norder\". If you mean:\n\n(a) that it would depend on the library load order, and that postgres\nwould assign the ID; then that won't work because the rmgr IDs need to\nbe stable across restarts and config changes\n\n(b) that there should be a convention where people reserve IDs on the\nwiki page in sequential order; then that's fine with me\n\n> 2) RM_MAX_ID is now UINT8_MAX - do we need to make it configurable?\n> or\n> do we think that only a few (127) custom rmgrs can exist?\n\n127 should be plenty for quite a long time. If we need more, we can\ncome up with a different solution (which is OK because the WAL format\nchanges in new major versions).\n\n> 3) Do we need to talk about extensible rmgrs and\n> https://wiki.postgresql.org/wiki/ExtensibleRmgr in documentation\n> somewhere? This will help extension developers to refer when needed.\n\nWe don't have user-facing documentation for every extension hook, and I\ndon't think it would be a good idea to document this one (at least in\nthe user-facing docs). Otherwise, we'd have to document the whole\nresource manager interface, and that seems like way too much of the\ninternals.\n\n> 4) Do we need to change this description?\n> <para>\n> The default value of this setting is the empty string, which\n> disables\n> the feature. It can be set to <literal>all</literal> to\n> check all\n> records, or to a comma-separated list of resource managers to\n> check\n> only records originating from those resource\n> managers. Currently,\n> the supported resource managers are <literal>heap</literal>,\n> <literal>heap2</literal>, <literal>btree</literal>,\n> <literal>hash</literal>,\n> <literal>gin</literal>, <literal>gist</literal>,\n> <literal>sequence</literal>,\n> <literal>spgist</literal>, <literal>brin</literal>, and\n> <literal>generic</literal>. Onl\n> superusers can change this setting.\n\nYes, you're right, I updated the docs for pg_waldump and the\nwal_consistency_checking GUC.\n\n> 5) Do we need to add a PGDLLIMPORT qualifier to RegisterCustomRmgr()\n> (possibly for RmgrStartup, RmgrTable, RmgrCleanup as well?) so that\n> the Windows versions of extensions can use this feature?\n\nThat seems to only be required for variables, not functions.\n\nI don't think we want to mark RmgrTable this way, because it's not\nintended to be accessed by extensions directly. It's accessed by\nGetRmgr(), which is 'static inline', but that's also not intended to be\nused by extensions.\n\n> 6) Should the below strcmp pg_strcmp? The custom rmgrs can still use\n> rm_name with different cases and below code fail to catch it?\n> + if (!strcmp(existing_rmgr->rm_name, rmgr->rm_name))\n> + ereport(PANIC,\n> + (errmsg(\"custom rmgr \\\"%s\\\" has the same name as builtin rmgr\",\n> + existing_rmgr->rm_name)));\n\nThere are already a lot of places where users can choose identifiers\nthat differ only in uppercase/lowercase. I don't see a reason to make\nan exception here.\n\n> 7) What's the intention of the below code? It seems like below it\n> checks if there's already a rmgr with the given name (RM_MAX_ID). Had\n> it been RM_MAX_BUILTIN_ID instead of RM_MAX_ID, the error message\n> does make sense. Is the intention here to not have duplicate rmgrs in\n> the entire RM_MAX_ID(both builtin and custom)?\n\nThank you. I redid a lot of the error messages and checked them out to\nmake sure they are helpful.\n\n> 8) Per https://wiki.postgresql.org/wiki/ExtensibleRmgr, 128 is\n> reserved, can we have it as a macro? Or something like\n> (RM_MAX_ID/2+1)\n\nIt's already defined as RM_EXPERIMENTAL_ID. Is that what you meant or\ndid I misunderstand?\n\nI don't see the point in defining it as RM_MAX_ID/2+1.\n\n> 9) Thinking if there's a way to test the code, maybe exposing\n> RegisterCustomRmgr as a function? I think we need to have a dummy\n> extension that will be used for testing this kind of patches/features\n> but that's a separate discussion IMO.\n\nIt's already exposed as a C function in xlog_internal.h.\n\nYou mean as a SQL function? I don't think that makes sense. It can only\nbe called while processing shared_preload_libraries (on server startup)\nanyway.\n\nOther changes this version:\n\n * implemented Andres's proposed performance change[1]\n * fix wal_consistency_checking\n * general cleanup\n\nRegards,\n\tJeff Davis\n\n[1] \nhttps://postgr.es/m/20220404043337.ocjnni7hknjsibhg@alap3.anarazel.de","msg_date":"Mon, 04 Apr 2022 17:25:55 -0700","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":true,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Tue, Apr 5, 2022 at 5:55 AM Jeff Davis <pgsql@j-davis.com> wrote:\n>\n> On Mon, 2022-04-04 at 11:15 +0530, Bharath Rupireddy wrote:\n> > 1) Why can't rmid be chosen by the extensions in sequential order\n> > from\n> > (129 till 255), say, to start with a columnar extension can choose\n> > 129, another extension can register 130 and so on right?\n>\n> I'm not sure what you mean by \"chosen by the extensions in sequential\n> order\". If you mean:\n>\n> (b) that there should be a convention where people reserve IDs on the\n> wiki page in sequential order; then that's fine with me\n\nYes, I meant this. If we do this, then a global variable\ncurrent_max_rmid can be maintained (by RegisterCustomRmgr) and the\nother functions RmgrStartup, RmgrCleanup and RegisterCustomRmgr can\navoid for-loops with full range RM_MAX_ID (for (int id = 0; id <\ncurrent_max_rmid; id++)).\n\n> > 3) Do we need to talk about extensible rmgrs and\n> > https://wiki.postgresql.org/wiki/ExtensibleRmgr in documentation\n> > somewhere? This will help extension developers to refer when needed.\n>\n> We don't have user-facing documentation for every extension hook, and I\n> don't think it would be a good idea to document this one (at least in\n> the user-facing docs). Otherwise, we'd have to document the whole\n> resource manager interface, and that seems like way too much of the\n> internals.\n\nWith the custom rmgrs, now it's time to have it in the documentation\n(probably not immediately, even after this patch gets in) describing\nwhat rmgrs are, what are built-in and custom rmgrs and how and when an\nextension needs to register, load and use rmgrs etc.\n\n> > 4) Do we need to change this description?\n> > <para>\n> > <literal>generic</literal>. Onl\n> > superusers can change this setting.\n>\n> Yes, you're right, I updated the docs for pg_waldump and the\n> wal_consistency_checking GUC.\n\n+ <para>\n+ Extensions may define additional resource managers with modules loaded\n+ in <xref linkend=\"guc-shared-preload-libraries\"/>. However, the names\n+ of the custom resource managers are not available at the time the\n\n+ errhint(\"Include the extension module that implements this resource\nmanager in shared_preload_libraries.\")));\n\nWith the above, do you mean to say that only the extensions that load\ntheir library via shared_preload_libraries are allowed to have custom\nrmgrs? How about extensions using {session, local}_preload_libraries,\nLOAD command? How about extensions that don't load the shared library\nvia {session, local, shared}_preload_libraries or LOAD command and let\nany of its functions load the shared library on first use?\n\n> > 5) Do we need to add a PGDLLIMPORT qualifier to RegisterCustomRmgr()\n> > (possibly for RmgrStartup, RmgrTable, RmgrCleanup as well?) so that\n> > the Windows versions of extensions can use this feature?\n>\n> That seems to only be required for variables, not functions.\n\nYou are right. For instance, the functions DefineCustomXXX are not\nusing PGDLLIMPORT qualifer, I believe they can be used by extensions\non WINDOWS.\n\n> > 6) Should the below strcmp pg_strcmp? The custom rmgrs can still use\n> > rm_name with different cases and below code fail to catch it?\n> > + if (!strcmp(existing_rmgr->rm_name, rmgr->rm_name))\n> > + ereport(PANIC,\n> > + (errmsg(\"custom rmgr \\\"%s\\\" has the same name as builtin rmgr\",\n> > + existing_rmgr->rm_name)));\n>\n> There are already a lot of places where users can choose identifiers\n> that differ only in uppercase/lowercase. I don't see a reason to make\n> an exception here.\n\nI think postgres parses the user-specified strings and converts them\nto lowercase unless they are double-quoted, that is why we see strcmp,\nnot pg_strcasecmp. But, the custom rmgr name isn't parsed by postgres\nright?\n\nFor instance - extension 1, with id 130 and name 'rmgr_foo' and\nextension 2 with id 131 and name 'Rmgr_foo'/'RMGR_FOO'...., then the\nfollowing code may not catch it right? Do you want rmgr names to be\ncase-sensitive/insensitive?\n\n+ if (!strcmp(RmgrTable[i].rm_name, rmgr->rm_name))\n+ ereport(PANIC,\n+ (errmsg(\"failed to register custom resource manager \\\"%s\\\" with ID\n%d\", rmgr->rm_name, rmid),\n+ errdetail(\"Existing resource manager with ID %d has the same name.\", i)));\n\n> > 8) Per https://wiki.postgresql.org/wiki/ExtensibleRmgr, 128 is\n> > reserved, can we have it as a macro? Or something like\n> > (RM_MAX_ID/2+1)\n>\n> It's already defined as RM_EXPERIMENTAL_ID. Is that what you meant or\n> did I misunderstand?\n\n+#define RM_MAX_ID UINT8_MAX\n+#define RM_MAX_BUILTIN_ID (RM_NEXT_ID - 1)\n+#define RM_MIN_CUSTOM_ID 128\n+#define RM_N_CUSTOM_IDS (RM_MAX_ID - RM_MIN_CUSTOM_ID + 1)\n+#define RMID_IS_BUILTIN(rmid) ((rmid) <= RM_MAX_BUILTIN_ID)\n+#define RMID_IS_CUSTOM(rmid) ((rmid) >= RM_MIN_CUSTOM_ID && (rmid) < RM_MAX_ID)\n+#define RMID_IS_VALID(rmid) (RMID_IS_BUILTIN((rmid)) || RMID_IS_CUSTOM((rmid)))\n\nTo me the above seems complex, why can't we just have the following,\nit's okay even if we don't use some of the macros in the *.c files?\ntotal rmgrs = 256\nbuilt in = 128 (index 0 - 127)\ncustom = 128 (index 128 - 255)\n\n#define RM_N_IDS (UINT8_MAX + 1)\n#define RM_N_BUILTIN_IDS (RM_N_IDS / 2)\n#define RM_N_CUSTOM_IDS (RM_N_IDS / 2)\n#define RM_MAX_ID (RM_N_IDS - 1)\n#define RM_MAX_BUILTIN_ID (RM_NEXT_ID - 1)\n#define RM_MIN_CUSTOM_ID (RM_N_IDS / 2)\n#define RM_MAX_CUSTOM_ID RM_MAX_ID\n#define RMID_IS_BUILTIN(rmid) ((rmid) <= RM_MAX_BUILTIN_ID)\n#define RMID_IS_CUSTOM(rmid) ((rmid) >= RM_MIN_CUSTOM_ID && (rmid) <=\nRM_MAX_CUSTOM_ID)\n#define RMID_IS_VALID(rmid) (RMID_IS_BUILTIN((rmid)) || RMID_IS_CUSTOM((rmid)))\n\nIf okay, you can specify in a comment that out of UINT8_MAX + 1 rmgrs,\nfirst half rmgr ids are supposed to be reserved for built in rmgrs and\nsecond half rmgr ids for extensions.\n\n> I don't see the point in defining it as RM_MAX_ID/2+1.\n>\n> > 9) Thinking if there's a way to test the code, maybe exposing\n> > RegisterCustomRmgr as a function? I think we need to have a dummy\n> > extension that will be used for testing this kind of patches/features\n> > but that's a separate discussion IMO.\n>\n> It's already exposed as a C function in xlog_internal.h.\n>\n> You mean as a SQL function? I don't think that makes sense. It can only\n> be called while processing shared_preload_libraries (on server startup)\n> anyway.\n\nYes, an SQL function showing all of the available rmgrs. With the\ncustom rmgrs, I think, an SQL function like pg_resource_managers()\nreturning a set of {name, id, (if possible - loaded_by_extension_name,\nrm_redo, rm_desc, rm_identify ... api names)}\n\nSome more comments:\n\n1) Is there any specific reason to have a function to emit a PANIC\nmessage? And it's being used only in one place in GetRmgr.\n+void\n+RmgrNotFound(RmgrId rmid)\n\n2) How about \"If rm_name is NULL, the corresponding RmgrTable[] entry\nis considered invalid.\"?\n+ * RmgrTable[] is indexed by RmgrId values (see rmgrlist.h). If rm_name is\n+ * NULL, the structure is considered invalid.\n\n3) How about adding errdetail something like \"Custom resource manager\nID must be between %d and %d (both inclusive)\", RM_MAX_BUILTIN_ID,\nRM_MAX_ID)?\n+ if (rmid < RM_MIN_CUSTOM_ID)\n+ ereport(PANIC, errmsg(\"custom resource manager ID %d is out of range\", rmid));\n\n4) How about adding errdetail something like \"Custom resource manager\nmust have a name\"\n+ if (rmgr->rm_name == NULL)\n+ ereport(PANIC, errmsg(\"custom resource manager is invalid\"));\n\n5) How about \"successfully registered custom resource manager\"?\n+ (errmsg(\"registered custom resource manager \\\"%s\\\" with ID %d\",\n\n6) Shouldn't this be ri <= RM_NEXT_ID?\n- for (ri = 0; ri < RM_NEXT_ID; ri++)\n+ for (ri = 0; ri < RM_MAX_ID; ri++)\n\n- for (ri = 0; ri < RM_NEXT_ID; ri++)\n+ for (ri = 0; ri <= RM_MAX_ID; ri++)\n\n7) Unrelated to this patch, why is there always an extra slot RM_MAX_ID + 1?\n\n-const RmgrDescData RmgrDescTable[RM_MAX_ID + 1] = {\n+static const RmgrDescData RmgrDescTable[RM_MAX_BUILTIN_ID + 1] = {\n #include \"access/rmgrlist.h\"\n\n-const RmgrData RmgrTable[RM_MAX_ID + 1] = {\n+RmgrData RmgrTable[RM_MAX_ID + 1] = {\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Tue, 5 Apr 2022 15:12:09 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Tue, 2022-04-05 at 15:12 +0530, Bharath Rupireddy wrote:\n> Yes, I meant this. If we do this, then a global variable\n> current_max_rmid can be maintained (by RegisterCustomRmgr) and the\n> other functions RmgrStartup, RmgrCleanup and RegisterCustomRmgr can\n> avoid for-loops with full range RM_MAX_ID (for (int id = 0; id <\n> current_max_rmid; id++)).\n\nThere are only 255 entries, so the loops in those functions aren't a\nsignificant overhead.\n\nIf we expand it, then we can come up a different solution. We can\nchange it easily enough later, because the WAL format changes between\nreleases and so do the APIs for any extension that might be using this.\n\n> With the custom rmgrs, now it's time to have it in the documentation\n> (probably not immediately, even after this patch gets in) describing\n> what rmgrs are, what are built-in and custom rmgrs and how and when\n> an\n> extension needs to register, load and use rmgrs etc.\n\nAdded a documentation section right after Generic WAL.\n\n> With the above, do you mean to say that only the extensions that load\n> their library via shared_preload_libraries are allowed to have custom\n> rmgrs? How about extensions using {session, local}_preload_libraries,\n> LOAD command? How about extensions that don't load the shared library\n> via {session, local, shared}_preload_libraries or LOAD command and\n> let\n> any of its functions load the shared library on first use?\n\nCorrect. The call to RegisterCustomRmgr *must* happen in the postmaster\nbefore any other processes (even the startup process) are launched.\nOtherwise recovery wouldn't work. The only place to make that happen is\nshared_preload_libraries.\n\n> For instance - extension 1, with id 130 and name 'rmgr_foo' and\n> extension 2 with id 131 and name 'Rmgr_foo'/'RMGR_FOO'...., then the\n> following code may not catch it right? Do you want rmgr names to be\n> case-sensitive/insensitive?\n\nYou convinced me. pg_waldump.c uses pg_strcasecmp(), so I'll use it as\nwell.\n\n> #define RM_N_IDS (UINT8_MAX + 1)\n> #define RM_N_BUILTIN_IDS (RM_N_IDS / 2)\n> #define RM_N_CUSTOM_IDS (RM_N_IDS / 2)\n> #define RM_MAX_ID (RM_N_IDS - 1)\n> #define RM_MAX_BUILTIN_ID (RM_NEXT_ID - 1)\n> #define RM_MIN_CUSTOM_ID (RM_N_IDS / 2)\n> #define RM_MAX_CUSTOM_ID RM_MAX_ID\n> #define RMID_IS_BUILTIN(rmid) ((rmid) <= RM_MAX_BUILTIN_ID)\n> #define RMID_IS_CUSTOM(rmid) ((rmid) >= RM_MIN_CUSTOM_ID && (rmid) <=\n> RM_MAX_CUSTOM_ID)\n> #define RMID_IS_VALID(rmid) (RMID_IS_BUILTIN((rmid)) ||\n> RMID_IS_CUSTOM((rmid)))\n\nI improved this a bit. But I didn't use your version that's based on\ndivision, that was more confusing to me.\n\n> Yes, an SQL function showing all of the available rmgrs. With the\n> custom rmgrs, I think, an SQL function like pg_resource_managers()\n> returning a set of {name, id, (if possible -\n> loaded_by_extension_name,\n> rm_redo, rm_desc, rm_identify ... api names)}\n\nAdded.\n\n> \n> Some more comments:\n> \n> 1) Is there any specific reason to have a function to emit a PANIC\n> message? And it's being used only in one place in GetRmgr.\n> +void\n> +RmgrNotFound(RmgrId rmid)\n\nGood call -- better to raise an error, and it can get escalated if it\nhappens in the wrong place. Done.\n\n> 2) How about \"If rm_name is NULL, the corresponding RmgrTable[] entry\n> is considered invalid.\"?\n> + * RmgrTable[] is indexed by RmgrId values (see rmgrlist.h). If\n> rm_name is\n> + * NULL, the structure is considered invalid.\n\nDone.\n\n> 3) How about adding errdetail something like \"Custom resource manager\n> ID must be between %d and %d (both inclusive)\", RM_MAX_BUILTIN_ID,\n> RM_MAX_ID)?\n> + if (rmid < RM_MIN_CUSTOM_ID)\n> + ereport(PANIC, errmsg(\"custom resource manager ID %d is out of\n> range\", rmid));\n\nDone.\n\n> 4) How about adding errdetail something like \"Custom resource manager\n> must have a name\"\n> + if (rmgr->rm_name == NULL)\n> + ereport(PANIC, errmsg(\"custom resource manager is invalid\"));\n\nImproved this message.\n\n> 5) How about \"successfully registered custom resource manager\"?\n> + (errmsg(\"registered custom resource manager \\\"%s\\\" with ID %d\",\n\nThat's a little much, we don't want to sound too excited that it worked\n;-)\n\n> 6) Shouldn't this be ri <= RM_NEXT_ID?\n> - for (ri = 0; ri < RM_NEXT_ID; ri++)\n> + for (ri = 0; ri < RM_MAX_ID; ri++)\n> \n> - for (ri = 0; ri < RM_NEXT_ID; ri++)\n> + for (ri = 0; ri <= RM_MAX_ID; ri++)\n\nThat would leave out the custom rmgrs.\n\n> 7) Unrelated to this patch, why is there always an extra slot\n> RM_MAX_ID + 1?\n\nRM_MAX_ID is 255, but we need 256 entries from 0-255 (inclusive).\n\nGood suggestions, thank you for the review!\n\nI'm happy with this patch and I committed it. Changes from last\nversion:\n\n 1. I came up with a better solution (based on a suggestion from\nAndres) to handle wal_consistency_checking properly. We can't\nunderstand the custom resource managers when the configuration file is\nfirst loaded, so if we enounter an unknown resource manager, I set a\nflag and reprocess it after the shared_preload_libraries are loaded.\n\n 2. Added pg_get_wal_resource_managers().\n\n 3. Documentation.\n\nRegards,\n\tJeff Davis\n\n\n\n\n","msg_date":"Wed, 06 Apr 2022 23:15:27 -0700","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":true,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"Hi,\n\nOn 2022-04-06 23:15:27 -0700, Jeff Davis wrote:\n> I'm happy with this patch and I committed it.\n\nThat caused breakage with WAL_DEBUG enabled. Fixed that. But it was\nhalf-broken before, at least since 70e81861fad, because 'rmid' didn't refer to\nthe current record's rmid anymore, but to rmid from \"Initialize resource\nmanagers\" - a constant.\n\n\nCauses plenty new warnings here:\n\nIn file included from /home/andres/src/postgresql/src/include/access/xlogrecord.h:14,\n from /home/andres/src/postgresql/src/include/access/xlogreader.h:41,\n from /home/andres/src/postgresql/src/include/access/brin_xlog.h:17,\n from /home/andres/src/postgresql/src/backend/access/transam/rmgr.c:10:\n/home/andres/src/postgresql/src/backend/access/transam/rmgr.c: In function ‘RegisterCustomRmgr’:\n/home/andres/src/postgresql/src/include/access/rmgr.h:42:66: warning: comparison is always true due to limited range of data type [-Wtype-limits]\n 42 | (rmid) <= RM_MAX_CUSTOM_ID)\n | ^~\n/home/andres/src/postgresql/src/backend/access/transam/rmgr.c:104:14: note: in expansion of macro ‘RMID_IS_CUSTOM’\n 104 | if (!RMID_IS_CUSTOM(rmid))\n | ^~~~~~~~~~~~~~\nIn file included from /home/andres/src/postgresql/src/include/c.h:814,\n from /home/andres/src/postgresql/src/include/postgres.h:46,\n from /home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c:9:\n/home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c: In function ‘GetRmgrDesc’:\n/home/andres/src/postgresql/src/include/access/rmgr.h:42:66: warning: comparison is always true due to limited range of data type [-Wtype-limits]\n 42 | (rmid) <= RM_MAX_CUSTOM_ID)\n | ^~\n/home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c:89:9: note: in expansion of macro ‘Assert’\n 89 | Assert(RMID_IS_VALID(rmid));\n | ^~~~~~\n/home/andres/src/postgresql/src/include/access/rmgr.h:43:57: note: in expansion of macro ‘RMID_IS_CUSTOM’\n 43 | #define RMID_IS_VALID(rmid) (RMID_IS_BUILTIN((rmid)) || RMID_IS_CUSTOM((rmid)))\n | ^~~~~~~~~~~~~~\n/home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c:89:16: note: in expansion of macro ‘RMID_IS_VALID’\n 89 | Assert(RMID_IS_VALID(rmid));\n | ^~~~~~~~~~~~~\n/home/andres/src/postgresql/src/include/access/rmgr.h:42:66: warning: comparison is always true due to limited range of data type [-Wtype-limits]\n 42 | (rmid) <= RM_MAX_CUSTOM_ID)\n | ^~\n/home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c:89:9: note: in expansion of macro ‘Assert’\n 89 | Assert(RMID_IS_VALID(rmid));\n | ^~~~~~\n/home/andres/src/postgresql/src/include/access/rmgr.h:43:57: note: in expansion of macro ‘RMID_IS_CUSTOM’\n 43 | #define RMID_IS_VALID(rmid) (RMID_IS_BUILTIN((rmid)) || RMID_IS_CUSTOM((rmid)))\n | ^~~~~~~~~~~~~~\n/home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c:89:16: note: in expansion of macro ‘RMID_IS_VALID’\n 89 | Assert(RMID_IS_VALID(rmid));\n | ^~~~~~~~~~~~~\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 6 Apr 2022 23:35:05 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"Hi,\n\nOn 2022-04-06 23:35:05 -0700, Andres Freund wrote:\n> Causes plenty new warnings here:\n\nAnd my machine isn't alone. There's also:\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=lapwing&dt=2022-04-07%2006%3A40%3A14\n\nrmgrdesc.c:44:1: error: missing braces around initializer [-Werror=missing-braces]\nrmgrdesc.c:44:1: error: (near initialization for 'CustomNumericNames[0]') [-Werror=missing-braces]\nrmgrdesc.c:45:1: error: missing braces around initializer [-Werror=missing-braces]\nrmgrdesc.c:45:1: error: (near initialization for 'CustomRmgrDesc[0]') [-Werror=missing-braces]\ncc1: all warnings being treated as errors\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 6 Apr 2022 23:56:40 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"Hi,\n\nOn 2022-04-06 23:56:40 -0700, Andres Freund wrote:\n> On 2022-04-06 23:35:05 -0700, Andres Freund wrote:\n> > Causes plenty new warnings here:\n> \n> And my machine isn't alone. There's also:\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=lapwing&dt=2022-04-07%2006%3A40%3A14\n> \n> rmgrdesc.c:44:1: error: missing braces around initializer [-Werror=missing-braces]\n> rmgrdesc.c:44:1: error: (near initialization for 'CustomNumericNames[0]') [-Werror=missing-braces]\n> rmgrdesc.c:45:1: error: missing braces around initializer [-Werror=missing-braces]\n> rmgrdesc.c:45:1: error: (near initialization for 'CustomRmgrDesc[0]') [-Werror=missing-braces]\n> cc1: all warnings being treated as errors\n\n(pushing a fix for this one in a few)\n\nThere's also:\n\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=wrasse&dt=2022-04-07%2006%3A49%3A14\n\nccache /opt/developerstudio12.6/bin/cc -m64 -Xa -v -O findtimezone.o initdb.o localtime.o -L../../../src/port -L../../../src/common -L../../../src/fe_utils -lpgfeutils -L../../../src/common -lpgcommon -L../../../src/port -lpgport -L../../../src/interfaces/libpq -lpq -L/home/nm/sw/nopath/uuid-64/lib -L/home/nm/sw/nopath/openldap-64/lib -Wl,--as-needed -Wl,-R'/home/nm/farm/studio64v12_6/HEAD/inst/lib' -lpgcommon -lpgport -lxslt -lxml2 -lpam -lssl -lcrypto -lgssapi_krb5 -lz -lreadline -ltermcap -lnsl -lsocket -lm -o initdb\nUndefined\t\t\tfirst referenced\n symbol \t\t\t in file\nRmgrTable initdb.o\n\nIIRC sun studio is annoying because it emits references to global variables\nwhen they're mentioned in a static inline, even if that inline is unused.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 7 Apr 2022 00:45:34 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"Hi,\n\nOn 2022-04-07 00:45:34 -0700, Andres Freund wrote:\n> On 2022-04-06 23:56:40 -0700, Andres Freund wrote:\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=wrasse&dt=2022-04-07%2006%3A49%3A14\n> \n> ccache /opt/developerstudio12.6/bin/cc -m64 -Xa -v -O findtimezone.o initdb.o localtime.o -L../../../src/port -L../../../src/common -L../../../src/fe_utils -lpgfeutils -L../../../src/common -lpgcommon -L../../../src/port -lpgport -L../../../src/interfaces/libpq -lpq -L/home/nm/sw/nopath/uuid-64/lib -L/home/nm/sw/nopath/openldap-64/lib -Wl,--as-needed -Wl,-R'/home/nm/farm/studio64v12_6/HEAD/inst/lib' -lpgcommon -lpgport -lxslt -lxml2 -lpam -lssl -lcrypto -lgssapi_krb5 -lz -lreadline -ltermcap -lnsl -lsocket -lm -o initdb\n> Undefined\t\t\tfirst referenced\n> symbol \t\t\t in file\n> RmgrTable initdb.o\n> \n> IIRC sun studio is annoying because it emits references to global variables\n> when they're mentioned in a static inline, even if that inline is unused.\n\nLooks like a #ifndef FRONTEND around the inlines should work? Or just go back\nto the macros, but without the RM_MAX_CUSTOM_ID check that's implied due to\nthe type width anyway.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 7 Apr 2022 01:16:26 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"On Thu, Apr 7, 2022 at 11:45 AM Jeff Davis <pgsql@j-davis.com> wrote:\n> I'm happy with this patch and I committed it. Changes from last\n> version:\n\nHi Jeff, I think there's a typo [1] in pg_waldump.c, we might miss to\ntake account of the last custom resource mgr stats. Please fix it.\n\ndiff --git a/src/bin/pg_waldump/pg_waldump.c b/src/bin/pg_waldump/pg_waldump.c\nindex 7d92dcaf87..30ca7684bd 100644\n--- a/src/bin/pg_waldump/pg_waldump.c\n+++ b/src/bin/pg_waldump/pg_waldump.c\n@@ -720,7 +720,7 @@ XLogDumpDisplayStats(XLogDumpConfig *config,\nXLogDumpStats *stats)\n * calculate column totals.\n */\n\n- for (ri = 0; ri < RM_MAX_ID; ri++)\n+ for (ri = 0; ri <= RM_MAX_ID; ri++)\n {\n total_count += stats->rmgr_stats[ri].count;\n total_rec_len += stats->rmgr_stats[ri].rec_len;\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Thu, 7 Apr 2022 14:00:08 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"},{"msg_contents":"Jeff Davis <pgsql@j-davis.com> writes:\n> I'm happy with this patch and I committed it.\n\nwrasse is less happy:\n\nccache /opt/developerstudio12.6/bin/cc -m64 -Xa -v -O findtimezone.o initdb.o localtime.o -L../../../src/port -L../../../src/common -L../../../src/fe_utils -lpgfeutils -L../../../src/common -lpgcommon -L../../../src/port -lpgport -L../../../src/interfaces/libpq -lpq -L/home/nm/sw/nopath/uuid-64/lib -L/home/nm/sw/nopath/openldap-64/lib -Wl,--as-needed -Wl,-R'/home/nm/farm/studio64v12_6/HEAD/inst/lib' -lpgcommon -lpgport -lxslt -lxml2 -lpam -lssl -lcrypto -lgssapi_krb5 -lz -lreadline -ltermcap -lnsl -lsocket -lm -o initdb\nUndefined\t\t\tfirst referenced\n symbol \t\t\t in file\nRmgrTable initdb.o\nld: fatal: symbol referencing errors\n\nI think this means the static inline functions you added in\nxlog_internal.h need to be wrapped in #ifndef FRONTEND.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 07 Apr 2022 10:05:15 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Extensible Rmgr for Table AMs"}],"string":"[\n {\n \"msg_contents\": \"Motivation:\\n\\nI'm working on a columnar compression AM[0]. Currently, it uses generic\\nxlog, which works for crash recovery and physical replication, but not\\nlogical decoding/replication.\\n\\nExtensible rmgr would enable the table AM to support its own\\nredo/decode hooks and WAL format, so that it could support crash\\nrecovery, physical replication, and logical replication.\\n\\nBackground:\\n\\nI submitted another patch[0] to add new logical records, which could be\\nused to support logical decoding directly, without the need for\\nextensible rmgr and without any assumptions about the table AM. This\\nwas designed to be easy to use, but inefficient. Amit raised\\nconcerns[1] about whether it could meet the needs of zheap. Andres\\nsuggested (off-list) that it would be better to just tackle the\\nextensible rmgr problem.\\n\\nThe idea for extensible rmgr has been proposed before[3]. The biggest\\nargument against it seemed to be that there was no complete use\\ncase[4], so the worry was that something would be left out. Columnar is\\ncomplete enough that I think it qualifies as a good use case.\\n\\nA subsequent proposal[5] was shot down because of a (potential?) need\\nfor catalog access[6]. The attached patch does not use the catalog;\\ninstead, it relies on table AM authors choosing IDs that don't conflict\\nwith each other. This seems like a reasonable answer, considering that\\nthere will likely be very few table AMs that go far enough to fully\\nsupport WAL including decoding.\\n\\nAre there any other major arguments/objections that I missed?\\n\\nProposal:\\n\\nThe attached patch (against v14, so it's easier to test columnar) is\\nsomewhat like a simplified version of [3] combined with refactoring to\\nmake decoding a part of the rmgr.\\n\\n * adds a new RmgrData method rm_decode\\n * refactors decode.c to use\\nthe new method\\n * add a layer of indirection GetRmgr to find an rmgr\\n \\n * fast path to find builtin rmgr in RmgrTable\\n * to find a custom\\nrmgr, traverses list of custom rmgrs that\\n are currently loaded\\n(unlikely to ever be more than a few)\\n * rmgr IDs from 0-127 are\\nreserved for builtin rmgrs\\n * rmgr IDs from 128-255 are reserved for\\ncustom rmgrs\\n * table AM authors need to avoid collisions between\\nrmgr IDs\\n\\nI have tested with columnar using a simple WAL format for logical\\ndecoding only, and I'm still using generic xlog for recovery and\\nphysical replication. I haven't tested the redo path, or how easy it\\nmight be to do something like generic xlog.\\n\\nQuestions:\\n\\n0. Do we want to go this route, or something simpler like my other\\nproposal, which introduces new logical record types[0]?\\n\\n1. I am allocating the custom rmgr list in TopMemoryContext, and it\\nonly works when loading as a part of shared_preload_libraries. This\\navoids the need for shared memory in Simon's patch[3]. Is that the\\nright thing to do?\\n\\n2. If we go this route, what do we do with generic xlog? It seems like\\na half feature, since it doesn't work with logical decoding.\\n\\n3. If the custom rmgr throws an error during redo, the server won't\\nstart. Should we have a GUC to turn non-builtin redo into a no-op to\\nreduce the impact of bugs in the implementation of a custom rmgr?\\n\\n4. Do we want to encourage index AMs to use this mechanism as well? I\\ndidn't really look into how suitable it is, but at a high level it\\nseems reasonable.\\n\\nRegards,\\n\\tJeff Davis\\n\\n[0] \\nhttps://postgr.es/m/20ee0b0ae6958804a88fe9580157587720faf664.camel@j-davis.com\\n[1] \\nhttps://postgr.es/m/CAA4eK1JVDnbQ80ULdZuhzQkzr_yYhfON-tg%3Dd1U7aWjK_R1ixQ%40mail.gmail.com\\n[2] https://github.com/citusdata/citus/tree/master/src/backend/columnar\\n[3] https://postgr.es/m/1229541840.4793.79.camel%40ebony.2ndQuadrant\\n[4] https://postgr.es/m/20992.1232667957%40sss.pgh.pa.us\\n[5] https://postgr.es/m/1266774840.7341.29872.camel%40ebony\\n[6] https://postgr.es/m/26134.1266776040%40sss.pgh.pa.us\",\n \"msg_date\": \"Mon, 08 Nov 2021 15:36:21 -0800\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Mon, 2021-11-08 at 15:36 -0800, Jeff Davis wrote:\\n> The attached patch (against v14, so it's easier to test columnar) is\\n> somewhat like a simplified version of [3] combined with refactoring\\n> to\\n> make decoding a part of the rmgr.\\n\\nI created a wiki page here: \\nhttps://wiki.postgresql.org/wiki/ExtensibleRmgr\\n\\nTo coordinate reservation of RmgrIds, to avoid conflicts. I don't\\nexpect it to be a practical problem given how much work it takes to\\ncreate a new table AM that needs full WAL support, but might as well\\nhave some transparency on how to choose a new RmgrId.\\n\\nI also updated the patch to point to the wiki page in the comments, and\\nadded in a new RM_EXPERIMENTAL_ID that can be used while an extension\\nis still in development. Hopefully this will prevent people reserving\\nlots of RmgrIds for extensions that never get released.\\n\\nRegards,\\n\\tJeff Davis\",\n \"msg_date\": \"Tue, 09 Nov 2021 17:38:29 -0800\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 8, 2021 at 6:36 PM Jeff Davis <pgsql@j-davis.com> wrote:\\n> Extensible rmgr would enable the table AM to support its own\\n> redo/decode hooks and WAL format, so that it could support crash\\n> recovery, physical replication, and logical replication.\\n\\nWithout taking a position on your implementation, which I have not\\nstudied, I like this idea in concept and I think it's an important\\ngoal.\\n\\n> Are there any other major arguments/objections that I missed?\\n\\nISTR some discussion of the fact that uninstalling the extension that\\nuses this facility, or failing to install it on your standby, will\\nlead to an unusable database. Personally, I don't see that as a big\\nproblem: we should just document that if you choose to use an\\nextension like this, then (1) it needs to be installed on all standbys\\nand (2) if you ever want to get rid of it, you need to stop using it,\\ndrop all the objects created with it, and then wait until all the WAL\\npreviously generated by that extension is gone not only from pg_wal\\nbut from any archived WAL files or backups that you intend to use with\\nthat cluster before you actually nuke it. Users who don't want to\\nabide by those restrictions need not install such extensions. Users\\nwho don't read the documentation might end up sad, but it doesn't seem\\nparticularly likely, and it's hardly the only part of the\\ndocumentation that users shouldn't ignore.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 09:58:14 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Mon, 2021-11-08 at 15:36 -0800, Jeff Davis wrote:\\n> The attached patch (against v14, so it's easier to test columnar) is\\n> somewhat like a simplified version of [3] combined with refactoring\\n> to\\n> make decoding a part of the rmgr.\\n\\nNew patches attached (v3). Essentially the content as v2, but split\\ninto 3 patches and rebased.\\n\\nReview on patch 0001 would be helpful. It makes decoding just another\\nmethod of rmgr, which makes a lot of sense to me from a code\\norganization standpoint regardless of the other patches. Is there any\\nreason not to do that?\\n\\nThe other patches then make rmgr extensible, which in turn makes\\ndecoding extensible and solves the logical replication problem for\\ncustom table AMs. The most obvious way to make rmgr extensible would be\\nto expand the rmgr table, but I was concerned about making that dynamic\\n(right now the structure is entirely constant and perhaps that's\\nimportant for some optimizations?). So, at the cost of complexity I\\nmade a separate, dynamic rmgr table to hold the custom entries.\\n\\nThe custom rmgr API would probably be marked \\\"experimental\\\" for a\\nwhile, and I don't expect lots of people to use it given the challenges\\nof a production-quality table AM. But it's still important, because\\nwithout it a table AM has no chance to participate in logical\\nreplication.\\n\\nRegards,\\n\\tJeff Davis\",\n \"msg_date\": \"Mon, 17 Jan 2022 00:42:06 -0800\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Mon, Jan 17, 2022 at 12:42:06AM -0800, Jeff Davis wrote:\\n> On Mon, 2021-11-08 at 15:36 -0800, Jeff Davis wrote:\\n> > The attached patch (against v14, so it's easier to test columnar) is\\n> > somewhat like a simplified version of [3] combined with refactoring\\n> > to\\n> > make decoding a part of the rmgr.\\n> \\n> New patches attached (v3). Essentially the content as v2, but split\\n> into 3 patches and rebased.\\n> \\n> Review on patch 0001 would be helpful. It makes decoding just another\\n> method of rmgr, which makes a lot of sense to me from a code\\n> organization standpoint regardless of the other patches. Is there any\\n> reason not to do that?\\n\\nI think that it's a clean and sensible approach, so +1 for me.\\n\\nThere's a bit of 0003 present in 002:\\n\\ndiff --git a/src/backend/access/transam/xlogreader.c b/src/backend/access/transam/xlogreader.c\\nindex 35029cf97d..612b1b3723 100644\\n--- a/src/backend/access/transam/xlogreader.c\\n+++ b/src/backend/access/transam/xlogreader.c\\n@@ -738,7 +738,7 @@ ValidXLogRecordHeader(XLogReaderState *state, XLogRecPtr RecPtr,\\n (uint32) SizeOfXLogRecord, record->xl_tot_len);\\n return false;\\n }\\n- if (record->xl_rmid > RM_MAX_ID)\\n+ if (record->xl_rmid > RM_MAX_ID && record->xl_rmid < RM_CUSTOM_MIN_ID)\\n {\\n\\n> \\n> The other patches then make rmgr extensible, which in turn makes\\n> decoding extensible and solves the logical replication problem for\\n> custom table AMs. The most obvious way to make rmgr extensible would be\\n> to expand the rmgr table, but I was concerned about making that dynamic\\n> (right now the structure is entirely constant and perhaps that's\\n> important for some optimizations?). So, at the cost of complexity I\\n> made a separate, dynamic rmgr table to hold the custom entries.\\n\\nI'm not sure about performance overhead, but allowing extension to access the\\nmain table directly seems a bit scary. We definitely accepted some overhead\\nwith the various extensible support, and this new GetRmgr() doesn't look like\\nit's going to cost a lot for the builtin case, especially compared to the cost\\nof any of the code associated with the rmgr.\\n\\nAlso, to answer your initial email I think it's a better way to go compared to\\nyour previous approach, given the limitations and performance of the generic\\nxlog infrastructure, and hopefully index AMs should be able to rely on this\\ntoo.\\n\\nA few random comments on 0003:\\n\\n #define RM_MAX_ID (RM_NEXT_ID - 1)\\n+#define RM_CUSTOM_MIN_ID 128\\n+#define RM_CUSTOM_MAX_ID UINT8_MAX\\n\\nIt would be a good idea to add a StaticAssertStmt here to make sure that\\nthere's no overlap in the ranges.\\n\\n+\\n+/*\\n+ * RmgrId to use for extensions that require an RmgrId, but are still in\\n+ * development and have not reserved their own unique RmgrId yet. See:\\n+ * https://wiki.postgresql.org/wiki/ExtensibleRmgr\\n+ */\\n+#define RM_EXPERIMENTAL_ID 128\\n\\nI'm a bit dubious about the whole \\\"register your ID in the Wiki\\\" approach. It\\nmight not be a problem since there probably won't be hundreds of users, and I\\ndon't have any better suggestion since it has to be consistent across nodes.\\n\\n+ elog(LOG, \\\"registering customer rmgr \\\\\\\"%s\\\\\\\" with ID %d\\\",\\n+ rmgr->rm_name, rmid);\\n\\nShould it be a DEBUG message instead? Also s/customer/custom/\\n\\n+RmgrData\\n+GetCustomRmgr(RmgrId rmid)\\n+{\\n+ if (rmid < RM_CUSTOM_MIN_ID || rmid > RM_CUSTOM_MAX_ID)\\n+ elog(PANIC, \\\"custom rmgr id %d out of range\\\", rmid);\\n\\nShould this be an assert?\\n\\n+#define GetBuiltinRmgr(rmid) RmgrTable[(rmid)]\\n+#define GetRmgr(rmid) ((rmid < RM_CUSTOM_MIN_ID) ? \\\\\\n+ GetBuiltinRmgr(rmid) : GetCustomRmgr(rmid))\\n\\nrmid should be protected in the macro.\\n\\n> The custom rmgr API would probably be marked \\\"experimental\\\" for a\\n> while, and I don't expect lots of people to use it given the challenges\\n> of a production-quality table AM. But it's still important, because\\n> without it a table AM has no chance to participate in logical\\n> replication.\\n\\nHow you plan to mark it experimental?\\n\\n\\n\",\n \"msg_date\": \"Tue, 18 Jan 2022 17:53:32 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Tue, 2022-01-18 at 17:53 +0800, Julien Rouhaud wrote:\\n> > Review on patch 0001 would be helpful. It makes decoding just\\n> > another\\n> > method of rmgr, which makes a lot of sense to me from a code\\n> > organization standpoint regardless of the other patches. Is there\\n> > any\\n> > reason not to do that?\\n> \\n> I think that it's a clean and sensible approach, so +1 for me.\\n\\nThank you, committed 0001. Other patches not committed yet.\\n\\n> There's a bit of 0003 present in 002:\\n\\nI just squashed 0002 and 0003 together. Not large enough to keep\\nseparate.\\n\\n> A few random comments on 0003:\\n> \\n> #define RM_MAX_ID (RM_NEXT_ID - 1)\\n> +#define RM_CUSTOM_MIN_ID 128\\n> +#define RM_CUSTOM_MAX_ID UINT8_MAX\\n> \\n> It would be a good idea to add a StaticAssertStmt here to make sure\\n> that\\n> there's no overlap in the ranges.\\n\\nDone.\\n\\n> +\\n> +/*\\n> + * RmgrId to use for extensions that require an RmgrId, but are\\n> still in\\n> + * development and have not reserved their own unique RmgrId yet.\\n> See:\\n> + * https://wiki.postgresql.org/wiki/ExtensibleRmgr\\n> + */\\n> +#define RM_EXPERIMENTAL_ID 128\\n> \\n> I'm a bit dubious about the whole \\\"register your ID in the Wiki\\\"\\n> approach. It\\n> might not be a problem since there probably won't be hundreds of\\n> users, and I\\n> don't have any better suggestion since it has to be consistent across\\n> nodes.\\n\\nAgree, but I don't see a better approach, either.\\n\\nI do some sanity checking, which should catch collisions when they\\nhappen.\\n\\n> + elog(LOG, \\\"registering customer rmgr \\\\\\\"%s\\\\\\\" with ID %d\\\",\\n> + rmgr->rm_name, rmid);\\n> \\n> Should it be a DEBUG message instead? Also s/customer/custom/\\n\\nIt seems like a fairly important thing to have in the log. Only a\\ncouple extensions will ever encounter this message, and only at server\\nstart.\\n\\nTypo fixed.\\n\\n> +RmgrData\\n> +GetCustomRmgr(RmgrId rmid)\\n> +{\\n> + if (rmid < RM_CUSTOM_MIN_ID || rmid > RM_CUSTOM_MAX_ID)\\n> + elog(PANIC, \\\"custom rmgr id %d out of range\\\", rmid);\\n> \\n> Should this be an assert?\\n\\nIf we make it an Assert, then it won't be caught in production builds.\\n\\n> +#define GetBuiltinRmgr(rmid) RmgrTable[(rmid)]\\n> +#define GetRmgr(rmid) ((rmid < RM_CUSTOM_MIN_ID) ? \\\\\\n> + GetBuiltinRmgr(rmid) : GetCustomRmgr(rmid))\\n> \\n> rmid should be protected in the macro.\\n\\nDone.\\n\\n> How you plan to mark it experimental?\\n\\nI suppose it doesn't need to be marked explicitly -- there are other\\nAPIs that change. For instance, the ProcessUtility_hook changed, and\\nthat's used much more widely.\\n\\nAs long as we generally agree that some kind of custom rmgrs are the\\nway to go, if we change the API or implementation around from version\\nto version I can easily update my table access method.\\n\\nRegards,\\n\\tJeff Davis\",\n \"msg_date\": \"Mon, 31 Jan 2022 18:36:59 -0800\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Mon, Jan 31, 2022 at 06:36:59PM -0800, Jeff Davis wrote:\\n> On Tue, 2022-01-18 at 17:53 +0800, Julien Rouhaud wrote:\\n> \\n> > There's a bit of 0003 present in 002:\\n> \\n> I just squashed 0002 and 0003 together. Not large enough to keep\\n> separate.\\n\\nAgreed.\\n\\n> > + elog(LOG, \\\"registering customer rmgr \\\\\\\"%s\\\\\\\" with ID %d\\\",\\n> > + rmgr->rm_name, rmid);\\n> > \\n> > Should it be a DEBUG message instead? Also s/customer/custom/\\n> \\n> It seems like a fairly important thing to have in the log. Only a\\n> couple extensions will ever encounter this message, and only at server\\n> start.\\n\\nOk.\\n\\n> > +RmgrData\\n> > +GetCustomRmgr(RmgrId rmid)\\n> > +{\\n> > + if (rmid < RM_CUSTOM_MIN_ID || rmid > RM_CUSTOM_MAX_ID)\\n> > + elog(PANIC, \\\"custom rmgr id %d out of range\\\", rmid);\\n> > \\n> > Should this be an assert?\\n> \\n> If we make it an Assert, then it won't be caught in production builds.\\n\\nSure, but it seems something so fundamental that it should get done right\\nduring the development phase rather than spending cycles in optimized builds to\\ncheck for it.\\nBut even if that happened it would get caught by the final \\\"not found\\\" PANIC\\nanyway, and for end users I don't think that getting this error instead would\\nmake much difference.\\n\\n> > How you plan to mark it experimental?\\n> \\n> I suppose it doesn't need to be marked explicitly -- there are other\\n> APIs that change. For instance, the ProcessUtility_hook changed, and\\n> that's used much more widely.\\n> As long as we generally agree that some kind of custom rmgrs are the\\n> way to go, if we change the API or implementation around from version\\n> to version I can easily update my table access method.\\n\\nAgreed, API breaks happen often and that's not really a problem for third-party\\nextensions, especially if that comes with hard compile failure.\\n\\nOther than that the patch looks good to me, as you said we just need a decision\\non whether custom rmgrs are wanted or not.\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Feb 2022 12:39:38 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Tue, Feb 01, 2022 at 12:39:38PM +0800, Julien Rouhaud wrote:\\n> \\n> Other than that the patch looks good to me, as you said we just need a decision\\n> on whether custom rmgrs are wanted or not.\\n\\nOne last thing, did you do some benchmark with a couple custom rmgr to see how\\nmuch the O(n) access is showing up in profiles?\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Feb 2022 20:45:50 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Tue, 2022-02-01 at 20:45 +0800, Julien Rouhaud wrote:\\n> On Tue, Feb 01, 2022 at 12:39:38PM +0800, Julien Rouhaud wrote:\\n> > \\n> > Other than that the patch looks good to me, as you said we just\\n> > need a decision\\n> > on whether custom rmgrs are wanted or not.\\n> \\n> One last thing, did you do some benchmark with a couple custom rmgr\\n> to see how\\n> much the O(n) access is showing up in profiles?\\n\\nWhat kind of a test case would be reasonable there? You mean having a\\nlot of custom rmgrs?\\n\\nI was expecting that few people would have more than one custom rmgr\\nloaded anyway, so a sparse array or hashtable seemed wasteful. If\\ncustom rmgrs become popular we probably need to have a larger ID space\\nanyway, but it seems like overengineering to do so now.\\n\\nRegards,\\n\\tJeff Davis\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 01 Feb 2022 15:38:32 -0800\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Tue, Feb 01, 2022 at 03:38:32PM -0800, Jeff Davis wrote:\\n> On Tue, 2022-02-01 at 20:45 +0800, Julien Rouhaud wrote:\\n> > \\n> > One last thing, did you do some benchmark with a couple custom rmgr\\n> > to see how\\n> > much the O(n) access is showing up in profiles?\\n> \\n> What kind of a test case would be reasonable there? You mean having a\\n> lot of custom rmgrs?\\n> \\n> I was expecting that few people would have more than one custom rmgr\\n> loaded anyway, so a sparse array or hashtable seemed wasteful. If\\n> custom rmgrs become popular we probably need to have a larger ID space\\n> anyway, but it seems like overengineering to do so now.\\n\\nI agree that having dozen of custom rmgrs doesn't seem likely, but I also have\\nno idea of how much overhead you get by not doing a direct array access. I\\nthink it would be informative to benchmark something like simple OLTP write\\nworkload on a fast storage (or a ramdisk, or with fsync off...), with the used\\nrmgr being the 1st and the 2nd custom rmgr. Both scenario still seems\\nplausible and shouldn't degenerate on good hardware.\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Feb 2022 13:34:18 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 3, 2022 at 12:34 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> I agree that having dozen of custom rmgrs doesn't seem likely, but I also have\\n> no idea of how much overhead you get by not doing a direct array access. I\\n> think it would be informative to benchmark something like simple OLTP write\\n> workload on a fast storage (or a ramdisk, or with fsync off...), with the used\\n> rmgr being the 1st and the 2nd custom rmgr. Both scenario still seems\\n> plausible and shouldn't degenerate on good hardware.\\n\\nI think it would be hard to measure the overhead of this approach on a\\nmacrobenchmark. That having been said, I find this a surprising\\nimplementation choice. I think that the approaches that are most worth\\nconsidering are:\\n\\n(1) reallocate the array if needed so that we can continue to just do\\nRmgrTable[rmid]\\n(2) have one array for builtins and a second array for extensions and\\ndo rmid < RM_CUSTOM_MIN_ID ? BuiltinRmgrTable[rmid] :\\nExtensionRmgrTable[rmid]\\n(3) change RmgrTable to be an array of pointers to structs rather than\\nan an array of structs. then the structs don't move around and can be\\nconst, but the pointers can be moved into a larger array if required\\n\\nI'm not really sure which is best. My intuition for what will be\\ncheapest on modern hardware is pretty shaky. However, I can't see how\\nit can be the thing the patch is doing now; a linear search seems like\\nit has to be the slowest option.\\n\\n--\\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 4 Feb 2022 09:10:42 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Fri, Feb 04, 2022 at 09:10:42AM -0500, Robert Haas wrote:\\n> On Thu, Feb 3, 2022 at 12:34 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> > I agree that having dozen of custom rmgrs doesn't seem likely, but I also have\\n> > no idea of how much overhead you get by not doing a direct array access. I\\n> > think it would be informative to benchmark something like simple OLTP write\\n> > workload on a fast storage (or a ramdisk, or with fsync off...), with the used\\n> > rmgr being the 1st and the 2nd custom rmgr. Both scenario still seems\\n> > plausible and shouldn't degenerate on good hardware.\\n> \\n> I think it would be hard to measure the overhead of this approach on a\\n> macrobenchmark.\\n\\nYeah that's also my initial thought, but I wouldn't be terribly surprised to be\\nwrong.\\n\\n> That having been said, I find this a surprising\\n> implementation choice. I think that the approaches that are most worth\\n> considering are:\\n> \\n> (1) reallocate the array if needed so that we can continue to just do\\n> RmgrTable[rmid]\\n> (2) have one array for builtins and a second array for extensions and\\n> do rmid < RM_CUSTOM_MIN_ID ? BuiltinRmgrTable[rmid] :\\n> ExtensionRmgrTable[rmid]\\n> (3) change RmgrTable to be an array of pointers to structs rather than\\n> an an array of structs. then the structs don't move around and can be\\n> const, but the pointers can be moved into a larger array if required\\n> \\n> I'm not really sure which is best. My intuition for what will be\\n> cheapest on modern hardware is pretty shaky. However, I can't see how\\n> it can be the thing the patch is doing now; a linear search seems like\\n> it has to be the slowest option.\\n\\nI guess the idea was to have a compromise between letting rmgr authors choose\\narbitrary ids to avoid any conflicts, especially with private implementations,\\nwithout wasting too much memory. But those approaches would be pretty much\\nincompatible with the current definition:\\n\\n+#define RM_CUSTOM_MIN_ID 128\\n+#define RM_CUSTOM_MAX_ID UINT8_MAX\\n\\neven if you only allocate up to the max id found, nothing guarantees that you\\nwon't get a quite high id.\\n\\n\\n\",\n \"msg_date\": \"Fri, 4 Feb 2022 22:48:01 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 4, 2022 at 9:48 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> I guess the idea was to have a compromise between letting rmgr authors choose\\n> arbitrary ids to avoid any conflicts, especially with private implementations,\\n> without wasting too much memory. But those approaches would be pretty much\\n> incompatible with the current definition:\\n>\\n> +#define RM_CUSTOM_MIN_ID 128\\n> +#define RM_CUSTOM_MAX_ID UINT8_MAX\\n>\\n> even if you only allocate up to the max id found, nothing guarantees that you\\n> won't get a quite high id.\\n\\nRight, which I guess raises another question: if the maximum is\\nUINT8_MAX, which BTW I find perfectly reasonable, why are we not just\\ndefining this as an array of size 256? There's no point in adding code\\ncomplexity to save a few kB of memory.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 4 Feb 2022 09:53:09 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 04, 2022 at 09:53:09AM -0500, Robert Haas wrote:\\n> On Fri, Feb 4, 2022 at 9:48 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> > I guess the idea was to have a compromise between letting rmgr authors choose\\n> > arbitrary ids to avoid any conflicts, especially with private implementations,\\n> > without wasting too much memory. But those approaches would be pretty much\\n> > incompatible with the current definition:\\n> >\\n> > +#define RM_CUSTOM_MIN_ID 128\\n> > +#define RM_CUSTOM_MAX_ID UINT8_MAX\\n> >\\n> > even if you only allocate up to the max id found, nothing guarantees that you\\n> > won't get a quite high id.\\n> \\n> Right, which I guess raises another question: if the maximum is\\n> UINT8_MAX, which BTW I find perfectly reasonable, why are we not just\\n> defining this as an array of size 256? There's no point in adding code\\n> complexity to save a few kB of memory.\\n\\nAgreed, especially if combined with your suggested approach 3 (array of\\npointers).\\n\\n\\n\",\n \"msg_date\": \"Fri, 4 Feb 2022 22:56:58 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Fri, 2022-02-04 at 22:56 +0800, Julien Rouhaud wrote:\\n> > Right, which I guess raises another question: if the maximum is\\n> > UINT8_MAX, which BTW I find perfectly reasonable, why are we not\\n> > just\\n> > defining this as an array of size 256? There's no point in adding\\n> > code\\n> > complexity to save a few kB of memory.\\n> \\n> Agreed, especially if combined with your suggested approach 3 (array\\n> of\\n> pointers).\\n\\nImplemented and attached. I also updated pg_waldump and pg_rewind to do\\nsomething reasonable.\\n\\nAdditionally, I now have a reasonably complete implementation of a\\ncustom resource manager now:\\n\\nhttps://github.com/citusdata/citus/tree/custom-rmgr-15\\n\\n(Not committed or intended to actually be used right now -- just a\\nPOC.)\\n\\nOffline, Andres mentioned that I should test recovery performance if we\\ntake your approach, because making the RmgrTable non-const could impact\\noptimizations. Not sure if that would be a practical concern compared\\nto all the other work done at REDO time.\\n\\nRegards,\\n\\tJeff Davis\",\n \"msg_date\": \"Wed, 23 Mar 2022 21:43:08 -0700\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-23 21:43:08 -0700, Jeff Davis wrote:\\n> /* must be kept in sync with RmgrData definition in xlog_internal.h */\\n> #define PG_RMGR(symname,name,redo,desc,identify,startup,cleanup,mask,decode) \\\\\\n> -\\t{ name, redo, desc, identify, startup, cleanup, mask, decode },\\n> +\\t&(struct RmgrData){ name, redo, desc, identify, startup, cleanup, mask, decode },\\n> \\n> -const RmgrData RmgrTable[RM_MAX_ID + 1] = {\\n> +static RmgrData *RmgrTable[RM_MAX_ID + 1] = {\\n> #include \\\"access/rmgrlist.h\\\"\\n> };\\n\\nI think this has been discussed before, but to me it's not obvious that it's a\\ngood idea to change RmgrTable from RmgrData to RmgrData *. That adds an\\nindirection, without obvious benefit.\\n\\n\\n> +\\n> +/*\\n> + * Start up all resource managers.\\n> + */\\n> +void\\n> +StartupResourceManagers()\\n\\n(void)\\n\\n\\n> +void\\n> +RegisterCustomRmgr(RmgrId rmid, RmgrData *rmgr)\\n> +{\\n> +\\tif (rmid < RM_MIN_CUSTOM_ID)\\n> +\\t\\tereport(PANIC, errmsg(\\\"custom rmgr id %d is out of range\\\", rmid));\\n> +\\n> +\\tif (!process_shared_preload_libraries_in_progress)\\n> +\\t\\tereport(ERROR,\\n> +\\t\\t\\t\\t(errmsg(\\\"custom rmgr must be registered while initializing modules in shared_preload_libraries\\\")));\\n> +\\n> +\\tereport(LOG,\\n> +\\t\\t\\t(errmsg(\\\"registering custom rmgr \\\\\\\"%s\\\\\\\" with ID %d\\\",\\n> +\\t\\t\\t\\t\\trmgr->rm_name, rmid)));\\n> +\\n> +\\tif (RmgrTable[rmid] != NULL)\\n> +\\t\\tereport(PANIC,\\n> +\\t\\t\\t\\t(errmsg(\\\"custom rmgr ID %d already registered with name \\\\\\\"%s\\\\\\\"\\\",\\n> +\\t\\t\\t\\t\\t\\trmid, RmgrTable[rmid]->rm_name)));\\n> +\\n> +\\t/* check for existing rmgr with the same name */\\n> +\\tfor (int i = 0; i <= RM_MAX_ID; i++)\\n> +\\t{\\n> +\\t\\tconst RmgrData *existing_rmgr = RmgrTable[i];\\n> +\\n> +\\t\\tif (existing_rmgr == NULL)\\n> +\\t\\t\\tcontinue;\\n> +\\n> +\\t\\tif (!strcmp(existing_rmgr->rm_name, rmgr->rm_name))\\n> +\\t\\t\\tereport(PANIC,\\n> +\\t\\t\\t\\t\\t(errmsg(\\\"custom rmgr \\\\\\\"%s\\\\\\\" has the same name as builtin rmgr\\\",\\n> +\\t\\t\\t\\t\\t\\t\\texisting_rmgr->rm_name)));\\n> +\\t}\\n> +\\n> +\\t/* register it */\\n> +\\tRmgrTable[rmid] = rmgr;\\n> +}\\n\\nRandom idea: Might be worth emitting the id->name mapping just after a redo\\nlocation is determined, to make it easier to debug things.\\n\\n\\n> +RmgrData *\\n> +GetRmgr(RmgrId rmid)\\n> +{\\n> +\\treturn RmgrTable[rmid];\\n> +}\\n\\nGiven this is so simple, why incur the cost of a function call? Rather than\\ncontinuing to expose RmgrTable and move GetRmgr() into the header, as a static\\ninline? As-is this also prevent the compiler from optimizing across repeated\\nGetRmgr() calls (which often won't be possible anyway, but still)..\\n\\n\\n> -\\tif (record->xl_rmid > RM_MAX_ID)\\n> +\\tif (record->xl_rmid > RM_MAX_BUILTIN_ID && record->xl_rmid < RM_MIN_CUSTOM_ID)\\n> \\t{\\n> \\t\\treport_invalid_record(state,\\n> \\t\\t\\t\\t\\t\\t\\t \\\"invalid resource manager ID %u at %X/%X\\\",\\n\\nShouldn't this continue to enforce RM_MAX_ID as well?\\n\\n\\n> @@ -1604,12 +1603,7 @@ PerformWalRecovery(void)\\n> \\n> \\t\\tInRedo = true;\\n> \\n> -\\t\\t/* Initialize resource managers */\\n> -\\t\\tfor (rmid = 0; rmid <= RM_MAX_ID; rmid++)\\n> -\\t\\t{\\n> -\\t\\t\\tif (RmgrTable[rmid].rm_startup != NULL)\\n> -\\t\\t\\t\\tRmgrTable[rmid].rm_startup();\\n> -\\t\\t}\\n> +\\t\\tStartupResourceManagers();\\n\\nPersonally I'd rather name it ResourceManagersStartup() or RmgrStartup().\\n\\n\\n> @@ -1871,7 +1860,7 @@ ApplyWalRecord(XLogReaderState *xlogreader, XLogRecord *record, TimeLineID *repl\\n> \\t\\txlogrecovery_redo(xlogreader, *replayTLI);\\n> \\n> \\t/* Now apply the WAL record itself */\\n> -\\tRmgrTable[record->xl_rmid].rm_redo(xlogreader);\\n> +\\tGetRmgr(record->xl_rmid)->rm_redo(xlogreader);\\n> \\n> \\t/*\\n> \\t * After redo, check whether the backup pages associated with the WAL\\n\\nSo we have just added one indirect call and one pointer indirection\\n(previously RmgrTable could be resolved by the linker, now it needs to be\\ndereferenced again), that's not too bad. I was afraid there'd be multiple\\ncalls.\\n\\n\\n> @@ -2101,16 +2090,16 @@ xlog_outdesc(StringInfo buf, XLogReaderState *record)\\n> \\tuint8\\t\\tinfo = XLogRecGetInfo(record);\\n> \\tconst char *id;\\n> \\n> -\\tappendStringInfoString(buf, RmgrTable[rmid].rm_name);\\n> +\\tappendStringInfoString(buf, GetRmgr(rmid)->rm_name);\\n> \\tappendStringInfoChar(buf, '/');\\n> \\n> -\\tid = RmgrTable[rmid].rm_identify(info);\\n> +\\tid = GetRmgr(rmid)->rm_identify(info);\\n> \\tif (id == NULL)\\n> \\t\\tappendStringInfo(buf, \\\"UNKNOWN (%X): \\\", info & ~XLR_INFO_MASK);\\n> \\telse\\n> \\t\\tappendStringInfo(buf, \\\"%s: \\\", id);\\n> \\n> -\\tRmgrTable[rmid].rm_desc(buf, record);\\n> +\\tGetRmgr(rmid)->rm_desc(buf, record);\\n> }\\n\\nLike here. It's obviously not as performance critical as replay. But it's\\nstill a shame to add 3 calls to GetRmgr, that each then need to dereference\\nRmgrTable. The compiler won't be able to optimize any of that away.\\n\\n\\n> @@ -117,8 +117,8 @@ LogicalDecodingProcessRecord(LogicalDecodingContext *ctx, XLogReaderState *recor\\n> \\n> \\trmid = XLogRecGetRmid(record);\\n> \\n> -\\tif (RmgrTable[rmid].rm_decode != NULL)\\n> -\\t\\tRmgrTable[rmid].rm_decode(ctx, &buf);\\n> +\\tif (GetRmgr(rmid)->rm_decode != NULL)\\n> +\\t\\tGetRmgr(rmid)->rm_decode(ctx, &buf);\\n> \\telse\\n> \\t{\\n> \\t\\t/* just deal with xid, and done */\\n\\nThis one might actually matter, overhead wise.\\n\\n\\n> @@ -473,7 +473,7 @@ static void\\n> XLogDumpDisplayRecord(XLogDumpConfig *config, XLogReaderState *record)\\n> {\\n> \\tconst char *id;\\n> -\\tconst RmgrDescData *desc = &RmgrDescTable[XLogRecGetRmid(record)];\\n> +\\tconst RmgrDescData *desc = GetRmgrDesc(XLogRecGetRmid(record));\\n> \\tuint32\\t\\trec_len;\\n> \\tuint32\\t\\tfpi_len;\\n> \\tRelFileNode rnode;\\n> @@ -658,7 +658,7 @@ XLogDumpDisplayStats(XLogDumpConfig *config, XLogDumpStats *stats)\\n> \\t * calculate column totals.\\n> \\t */\\n> \\n> -\\tfor (ri = 0; ri < RM_NEXT_ID; ri++)\\n> +\\tfor (ri = 0; ri < RM_MAX_ID; ri++)\\n> \\t{\\n> \\t\\ttotal_count += stats->rmgr_stats[ri].count;\\n> \\t\\ttotal_rec_len += stats->rmgr_stats[ri].rec_len;\\n\\n> @@ -694,6 +694,9 @@ XLogDumpDisplayStats(XLogDumpConfig *config, XLogDumpStats *stats)\\n> \\t\\t\\tfpi_len = stats->rmgr_stats[ri].fpi_len;\\n> \\t\\t\\ttot_len = rec_len + fpi_len;\\n> \\n> +\\t\\t\\tif (ri > RM_MAX_BUILTIN_ID && count == 0)\\n> +\\t\\t\\t\\tcontinue;\\n> +\\n\\nAh, I see. I had written a concerned comment about the previous hunk...\\n\\n\\n> \\t\\t\\tXLogDumpStatsRow(desc->rm_name,\\n> \\t\\t\\t\\t\\t\\t\\t count, total_count, rec_len, total_rec_len,\\n> \\t\\t\\t\\t\\t\\t\\t fpi_len, total_fpi_len, tot_len, total_len);\\n> @@ -913,16 +916,16 @@ main(int argc, char **argv)\\n> \\t\\t\\t\\t\\t\\texit(EXIT_SUCCESS);\\n> \\t\\t\\t\\t\\t}\\n> \\n> -\\t\\t\\t\\t\\tfor (i = 0; i <= RM_MAX_ID; i++)\\n> +\\t\\t\\t\\t\\tfor (i = 0; i <= RM_MAX_BUILTIN_ID; i++)\\n> \\t\\t\\t\\t\\t{\\n> -\\t\\t\\t\\t\\t\\tif (pg_strcasecmp(optarg, RmgrDescTable[i].rm_name) == 0)\\n> +\\t\\t\\t\\t\\t\\tif (pg_strcasecmp(optarg, GetRmgrDesc(i)->rm_name) == 0)\\n> \\t\\t\\t\\t\\t\\t{\\n> \\t\\t\\t\\t\\t\\t\\tconfig.filter_by_rmgr[i] = true;\\n> \\t\\t\\t\\t\\t\\t\\tconfig.filter_by_rmgr_enabled = true;\\n> \\t\\t\\t\\t\\t\\t\\tbreak;\\n> \\t\\t\\t\\t\\t\\t}\\n> \\t\\t\\t\\t\\t}\\n> -\\t\\t\\t\\t\\tif (i > RM_MAX_ID)\\n> +\\t\\t\\t\\t\\tif (i > RM_MAX_BUILTIN_ID)\\n> \\t\\t\\t\\t\\t{\\n> \\t\\t\\t\\t\\t\\tpg_log_error(\\\"resource manager \\\\\\\"%s\\\\\\\" does not exist\\\",\\n> \\t\\t\\t\\t\\t\\t\\t\\t\\t optarg);\\n\\nSo we can't filter by rmgr id for non-builtin rmgr's? That sucks. Maybe add a\\ncustom:<i> syntax? Or generally allow numerical identifiers in addition to the\\nnames?\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 11:00:45 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Mon, 2022-03-28 at 11:00 -0700, Andres Freund wrote:\\n> I think this has been discussed before, but to me it's not obvious\\n> that it's a\\n> good idea to change RmgrTable from RmgrData to RmgrData *. That adds\\n> an\\n> indirection, without obvious benefit.\\n\\nI did some performance tests. I created a narrow table, took a base\\nbackup, loaded 100M records, finished the base backup. Then I recovered\\nusing the different build combinations (patched/unpatched, clang/gcc).\\n\\n compiler run1 run2\\n unpatched: gcc 102s 106s\\n patched: gcc 107s 105s\\n unpatched: clang 109s 110s\\n patched: clang 109s 111s\\n\\nI don't see a clear signal that this patch worsens performance. The\\n102s run was the very first run, so I suspect it was just due to the\\nprocessor starting out cold. Let me know if you think the test is\\ntesting the right paths.\\n\\nPerhaps I should address your other perf concerns around GetRmgr (make\\nit static inline, reduce number of calls), and then leave the\\nindirection for the sake of cleanliness?\\n\\nIf you are still concerned, I can switch back to separate tables to\\neliminate the indirection for built-in rmgrs. Separate rmgr tables\\nstill require a branch (to decide which table to access), but it should\\nbe a highly predictable one.\\n\\nRegards,\\n\\tJeff Davis\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Mar 2022 14:20:51 -0700\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Mon, 2022-03-28 at 11:00 -0700, Andres Freund wrote:\\n> > +/*\\n> > + * Start up all resource managers.\\n> > + */\\n> > +void\\n> > +StartupResourceManagers()\\n> \\n> (void)\\n\\nFixed.\\n\\n> Random idea: Might be worth emitting the id->name mapping just after\\n> a redo\\n> location is determined, to make it easier to debug things.\\n\\nNot quite sure I understood this idea, do you mean dump all rmgrs and\\nthe IDs when performing recovery?\\n\\n> > +RmgrData *\\n> > +GetRmgr(RmgrId rmid)\\n> > +{\\n> > +\\treturn RmgrTable[rmid];\\n> > +}\\n> \\n> Given this is so simple, why incur the cost of a function call?\\n> Rather than\\n> continuing to expose RmgrTable and move GetRmgr() into the header, as\\n> a static\\n> inline?\\n\\nMade it static inline.\\n\\n> Shouldn't this continue to enforce RM_MAX_ID as well?\\n\\nDone.\\n\\n> Personally I'd rather name it ResourceManagersStartup() or\\n> RmgrStartup().\\n\\nDone.\\n\\n> Like here. It's obviously not as performance critical as replay. But\\n> it's\\n> still a shame to add 3 calls to GetRmgr, that each then need to\\n> dereference\\n> RmgrTable. The compiler won't be able to optimize any of that away.\\n\\nChanged to only call it once and save it in a variable for each call\\nsite where it makes sense.\\n\\n> So we can't filter by rmgr id for non-builtin rmgr's? That sucks.\\n> Maybe add a\\n> custom:<i> syntax? Or generally allow numerical identifiers in\\n> addition to the\\n> names?\\n\\nGood idea. I changed it to allow \\\"custom###\\\" to mean the custom rmgr\\nwith ID ### (3 digits).\\n\\nI still may change it to go back to two RmgrTables (one for builtin and\\none for custom) to remove the lingering performance doubts. Other than\\nthat, and some cleanup, this is pretty close to the version I intend to\\ncommit.\\n\\nRegards,\\n\\tJeff Davis\",\n \"msg_date\": \"Sun, 03 Apr 2022 20:03:08 -0700\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-31 14:20:51 -0700, Jeff Davis wrote:\\n> If you are still concerned, I can switch back to separate tables to\\n> eliminate the indirection for built-in rmgrs. Separate rmgr tables\\n> still require a branch (to decide which table to access), but it should\\n> be a highly predictable one.\\n\\nI still think the easiest and fastest would be to just make RmgrTable longer,\\nand not const. When registering, copy the caller provided struct into the\\nrespective RmgrData element. Yes, we'd waste a bit of space at the end of the\\narray, but it's typically not going to be touched and thus not be backed by\\n\\\"actual\\\" memory.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sun, 3 Apr 2022 21:33:37 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Mon, Apr 4, 2022 at 8:33 AM Jeff Davis <pgsql@j-davis.com> wrote:\\n>\\n> I still may change it to go back to two RmgrTables (one for builtin and\\n> one for custom) to remove the lingering performance doubts. Other than\\n> that, and some cleanup, this is pretty close to the version I intend to\\n> commit.\\n\\nI just had a quick look over the v6 patch, few comments:\\n\\n1) Why can't rmid be chosen by the extensions in sequential order from\\n(129 till 255), say, to start with a columnar extension can choose\\n129, another extension can register 130 and so on right? This way, the\\nRmgrStartup, RmgrCleanup, and XLogDumpDisplayRecord can avoid looping\\nover all the 256 entries, with a global variable like\\ncurrent_max_rmid(representing the current number of rmgers)? Is there\\nany specific reason to allow them to choose rmgrid randomly between\\n129 till 255? Am I missing some discussion here?\\n2) RM_MAX_ID is now UINT8_MAX - do we need to make it configurable? or\\ndo we think that only a few (127) custom rmgrs can exist?\\n3) Do we need to talk about extensible rmgrs and\\nhttps://wiki.postgresql.org/wiki/ExtensibleRmgr in documentation\\nsomewhere? This will help extension developers to refer when needed.\\n4) Do we need to change this description?\\n <para>\\n The default value of this setting is the empty string, which disables\\n the feature. It can be set to <literal>all</literal> to check all\\n records, or to a comma-separated list of resource managers to check\\n only records originating from those resource managers. Currently,\\n the supported resource managers are <literal>heap</literal>,\\n <literal>heap2</literal>, <literal>btree</literal>,\\n<literal>hash</literal>,\\n <literal>gin</literal>, <literal>gist</literal>,\\n<literal>sequence</literal>,\\n <literal>spgist</literal>, <literal>brin</literal>, and\\n<literal>generic</literal>. Onl\\n superusers can change this setting.\\n5) Do we need to add a PGDLLIMPORT qualifier to RegisterCustomRmgr()\\n(possibly for RmgrStartup, RmgrTable, RmgrCleanup as well?) so that\\nthe Windows versions of extensions can use this feature?\\n6) Should the below strcmp pg_strcmp? The custom rmgrs can still use\\nrm_name with different cases and below code fail to catch it?\\n+ if (!strcmp(existing_rmgr->rm_name, rmgr->rm_name))\\n+ ereport(PANIC,\\n+ (errmsg(\\\"custom rmgr \\\\\\\"%s\\\\\\\" has the same name as builtin rmgr\\\",\\n+ existing_rmgr->rm_name)));\\n7) What's the intention of the below code? It seems like below it\\nchecks if there's already a rmgr with the given name (RM_MAX_ID). Had\\nit been RM_MAX_BUILTIN_ID instead of RM_MAX_ID, the error message\\ndoes make sense. Is the intention here to not have duplicate rmgrs in\\nthe entire RM_MAX_ID(both builtin and custom)?\\n+ /* check for existing rmgr with the same name */\\n+ for (int i = 0; i <= RM_MAX_ID; i++)\\n+ {\\n+ const RmgrData *existing_rmgr = RmgrTable[i];\\n+\\n+ if (existing_rmgr == NULL)\\n+ continue;\\n+\\n+ if (!strcmp(existing_rmgr->rm_name, rmgr->rm_name))\\n+ ereport(PANIC,\\n+ (errmsg(\\\"custom rmgr \\\\\\\"%s\\\\\\\" has the same name as builtin rmgr\\\",\\n+ existing_rmgr->rm_name)));\\n8) Per https://wiki.postgresql.org/wiki/ExtensibleRmgr, 128 is\\nreserved, can we have it as a macro? Or something like (RM_MAX_ID/2+1)\\n+#define RM_MIN_CUSTOM_ID 128\\n9) Thinking if there's a way to test the code, maybe exposing\\nRegisterCustomRmgr as a function? I think we need to have a dummy\\nextension that will be used for testing this kind of patches/features\\nbut that's a separate discussion IMO.\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Apr 2022 11:15:23 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Sun, 2022-04-03 at 21:33 -0700, Andres Freund wrote:\\n> I still think the easiest and fastest would be to just make RmgrTable\\n> longer,\\n> and not const. When registering, copy the caller provided struct into\\n> the\\n> respective RmgrData element. Yes, we'd waste a bit of space at the\\n> end of the\\n> array, but it's typically not going to be touched and thus not be\\n> backed by\\n> \\\"actual\\\" memory.\\n\\nSounds good to me. I tried to break down the performance between these\\napproaches and didn't get a clear signal, so I'll go with your\\nintuition here.\\n\\nPosting new patch in response to Bharath's review, which will include\\nthis change.\\n\\nNote that GetRmgr() also has an unlikely branch where it tests for the\\nvalidity of the rmgr before using it, so that it can throw a nice error\\nmessage if someone forgot to include the module in\\nshared_preload_libraries. I expect this will be highly predictable and\\ntherefore not a problem.\\n\\nRegards,\\n\\tJeff Davis\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 04 Apr 2022 17:10:51 -0700\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Mon, 2022-04-04 at 11:15 +0530, Bharath Rupireddy wrote:\\n> 1) Why can't rmid be chosen by the extensions in sequential order\\n> from\\n> (129 till 255), say, to start with a columnar extension can choose\\n> 129, another extension can register 130 and so on right?\\n\\nI'm not sure what you mean by \\\"chosen by the extensions in sequential\\norder\\\". If you mean:\\n\\n(a) that it would depend on the library load order, and that postgres\\nwould assign the ID; then that won't work because the rmgr IDs need to\\nbe stable across restarts and config changes\\n\\n(b) that there should be a convention where people reserve IDs on the\\nwiki page in sequential order; then that's fine with me\\n\\n> 2) RM_MAX_ID is now UINT8_MAX - do we need to make it configurable?\\n> or\\n> do we think that only a few (127) custom rmgrs can exist?\\n\\n127 should be plenty for quite a long time. If we need more, we can\\ncome up with a different solution (which is OK because the WAL format\\nchanges in new major versions).\\n\\n> 3) Do we need to talk about extensible rmgrs and\\n> https://wiki.postgresql.org/wiki/ExtensibleRmgr in documentation\\n> somewhere? This will help extension developers to refer when needed.\\n\\nWe don't have user-facing documentation for every extension hook, and I\\ndon't think it would be a good idea to document this one (at least in\\nthe user-facing docs). Otherwise, we'd have to document the whole\\nresource manager interface, and that seems like way too much of the\\ninternals.\\n\\n> 4) Do we need to change this description?\\n> <para>\\n> The default value of this setting is the empty string, which\\n> disables\\n> the feature. It can be set to <literal>all</literal> to\\n> check all\\n> records, or to a comma-separated list of resource managers to\\n> check\\n> only records originating from those resource\\n> managers. Currently,\\n> the supported resource managers are <literal>heap</literal>,\\n> <literal>heap2</literal>, <literal>btree</literal>,\\n> <literal>hash</literal>,\\n> <literal>gin</literal>, <literal>gist</literal>,\\n> <literal>sequence</literal>,\\n> <literal>spgist</literal>, <literal>brin</literal>, and\\n> <literal>generic</literal>. Onl\\n> superusers can change this setting.\\n\\nYes, you're right, I updated the docs for pg_waldump and the\\nwal_consistency_checking GUC.\\n\\n> 5) Do we need to add a PGDLLIMPORT qualifier to RegisterCustomRmgr()\\n> (possibly for RmgrStartup, RmgrTable, RmgrCleanup as well?) so that\\n> the Windows versions of extensions can use this feature?\\n\\nThat seems to only be required for variables, not functions.\\n\\nI don't think we want to mark RmgrTable this way, because it's not\\nintended to be accessed by extensions directly. It's accessed by\\nGetRmgr(), which is 'static inline', but that's also not intended to be\\nused by extensions.\\n\\n> 6) Should the below strcmp pg_strcmp? The custom rmgrs can still use\\n> rm_name with different cases and below code fail to catch it?\\n> + if (!strcmp(existing_rmgr->rm_name, rmgr->rm_name))\\n> + ereport(PANIC,\\n> + (errmsg(\\\"custom rmgr \\\\\\\"%s\\\\\\\" has the same name as builtin rmgr\\\",\\n> + existing_rmgr->rm_name)));\\n\\nThere are already a lot of places where users can choose identifiers\\nthat differ only in uppercase/lowercase. I don't see a reason to make\\nan exception here.\\n\\n> 7) What's the intention of the below code? It seems like below it\\n> checks if there's already a rmgr with the given name (RM_MAX_ID). Had\\n> it been RM_MAX_BUILTIN_ID instead of RM_MAX_ID, the error message\\n> does make sense. Is the intention here to not have duplicate rmgrs in\\n> the entire RM_MAX_ID(both builtin and custom)?\\n\\nThank you. I redid a lot of the error messages and checked them out to\\nmake sure they are helpful.\\n\\n> 8) Per https://wiki.postgresql.org/wiki/ExtensibleRmgr, 128 is\\n> reserved, can we have it as a macro? Or something like\\n> (RM_MAX_ID/2+1)\\n\\nIt's already defined as RM_EXPERIMENTAL_ID. Is that what you meant or\\ndid I misunderstand?\\n\\nI don't see the point in defining it as RM_MAX_ID/2+1.\\n\\n> 9) Thinking if there's a way to test the code, maybe exposing\\n> RegisterCustomRmgr as a function? I think we need to have a dummy\\n> extension that will be used for testing this kind of patches/features\\n> but that's a separate discussion IMO.\\n\\nIt's already exposed as a C function in xlog_internal.h.\\n\\nYou mean as a SQL function? I don't think that makes sense. It can only\\nbe called while processing shared_preload_libraries (on server startup)\\nanyway.\\n\\nOther changes this version:\\n\\n * implemented Andres's proposed performance change[1]\\n * fix wal_consistency_checking\\n * general cleanup\\n\\nRegards,\\n\\tJeff Davis\\n\\n[1] \\nhttps://postgr.es/m/20220404043337.ocjnni7hknjsibhg@alap3.anarazel.de\",\n \"msg_date\": \"Mon, 04 Apr 2022 17:25:55 -0700\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2022 at 5:55 AM Jeff Davis <pgsql@j-davis.com> wrote:\\n>\\n> On Mon, 2022-04-04 at 11:15 +0530, Bharath Rupireddy wrote:\\n> > 1) Why can't rmid be chosen by the extensions in sequential order\\n> > from\\n> > (129 till 255), say, to start with a columnar extension can choose\\n> > 129, another extension can register 130 and so on right?\\n>\\n> I'm not sure what you mean by \\\"chosen by the extensions in sequential\\n> order\\\". If you mean:\\n>\\n> (b) that there should be a convention where people reserve IDs on the\\n> wiki page in sequential order; then that's fine with me\\n\\nYes, I meant this. If we do this, then a global variable\\ncurrent_max_rmid can be maintained (by RegisterCustomRmgr) and the\\nother functions RmgrStartup, RmgrCleanup and RegisterCustomRmgr can\\navoid for-loops with full range RM_MAX_ID (for (int id = 0; id <\\ncurrent_max_rmid; id++)).\\n\\n> > 3) Do we need to talk about extensible rmgrs and\\n> > https://wiki.postgresql.org/wiki/ExtensibleRmgr in documentation\\n> > somewhere? This will help extension developers to refer when needed.\\n>\\n> We don't have user-facing documentation for every extension hook, and I\\n> don't think it would be a good idea to document this one (at least in\\n> the user-facing docs). Otherwise, we'd have to document the whole\\n> resource manager interface, and that seems like way too much of the\\n> internals.\\n\\nWith the custom rmgrs, now it's time to have it in the documentation\\n(probably not immediately, even after this patch gets in) describing\\nwhat rmgrs are, what are built-in and custom rmgrs and how and when an\\nextension needs to register, load and use rmgrs etc.\\n\\n> > 4) Do we need to change this description?\\n> > <para>\\n> > <literal>generic</literal>. Onl\\n> > superusers can change this setting.\\n>\\n> Yes, you're right, I updated the docs for pg_waldump and the\\n> wal_consistency_checking GUC.\\n\\n+ <para>\\n+ Extensions may define additional resource managers with modules loaded\\n+ in <xref linkend=\\\"guc-shared-preload-libraries\\\"/>. However, the names\\n+ of the custom resource managers are not available at the time the\\n\\n+ errhint(\\\"Include the extension module that implements this resource\\nmanager in shared_preload_libraries.\\\")));\\n\\nWith the above, do you mean to say that only the extensions that load\\ntheir library via shared_preload_libraries are allowed to have custom\\nrmgrs? How about extensions using {session, local}_preload_libraries,\\nLOAD command? How about extensions that don't load the shared library\\nvia {session, local, shared}_preload_libraries or LOAD command and let\\nany of its functions load the shared library on first use?\\n\\n> > 5) Do we need to add a PGDLLIMPORT qualifier to RegisterCustomRmgr()\\n> > (possibly for RmgrStartup, RmgrTable, RmgrCleanup as well?) so that\\n> > the Windows versions of extensions can use this feature?\\n>\\n> That seems to only be required for variables, not functions.\\n\\nYou are right. For instance, the functions DefineCustomXXX are not\\nusing PGDLLIMPORT qualifer, I believe they can be used by extensions\\non WINDOWS.\\n\\n> > 6) Should the below strcmp pg_strcmp? The custom rmgrs can still use\\n> > rm_name with different cases and below code fail to catch it?\\n> > + if (!strcmp(existing_rmgr->rm_name, rmgr->rm_name))\\n> > + ereport(PANIC,\\n> > + (errmsg(\\\"custom rmgr \\\\\\\"%s\\\\\\\" has the same name as builtin rmgr\\\",\\n> > + existing_rmgr->rm_name)));\\n>\\n> There are already a lot of places where users can choose identifiers\\n> that differ only in uppercase/lowercase. I don't see a reason to make\\n> an exception here.\\n\\nI think postgres parses the user-specified strings and converts them\\nto lowercase unless they are double-quoted, that is why we see strcmp,\\nnot pg_strcasecmp. But, the custom rmgr name isn't parsed by postgres\\nright?\\n\\nFor instance - extension 1, with id 130 and name 'rmgr_foo' and\\nextension 2 with id 131 and name 'Rmgr_foo'/'RMGR_FOO'...., then the\\nfollowing code may not catch it right? Do you want rmgr names to be\\ncase-sensitive/insensitive?\\n\\n+ if (!strcmp(RmgrTable[i].rm_name, rmgr->rm_name))\\n+ ereport(PANIC,\\n+ (errmsg(\\\"failed to register custom resource manager \\\\\\\"%s\\\\\\\" with ID\\n%d\\\", rmgr->rm_name, rmid),\\n+ errdetail(\\\"Existing resource manager with ID %d has the same name.\\\", i)));\\n\\n> > 8) Per https://wiki.postgresql.org/wiki/ExtensibleRmgr, 128 is\\n> > reserved, can we have it as a macro? Or something like\\n> > (RM_MAX_ID/2+1)\\n>\\n> It's already defined as RM_EXPERIMENTAL_ID. Is that what you meant or\\n> did I misunderstand?\\n\\n+#define RM_MAX_ID UINT8_MAX\\n+#define RM_MAX_BUILTIN_ID (RM_NEXT_ID - 1)\\n+#define RM_MIN_CUSTOM_ID 128\\n+#define RM_N_CUSTOM_IDS (RM_MAX_ID - RM_MIN_CUSTOM_ID + 1)\\n+#define RMID_IS_BUILTIN(rmid) ((rmid) <= RM_MAX_BUILTIN_ID)\\n+#define RMID_IS_CUSTOM(rmid) ((rmid) >= RM_MIN_CUSTOM_ID && (rmid) < RM_MAX_ID)\\n+#define RMID_IS_VALID(rmid) (RMID_IS_BUILTIN((rmid)) || RMID_IS_CUSTOM((rmid)))\\n\\nTo me the above seems complex, why can't we just have the following,\\nit's okay even if we don't use some of the macros in the *.c files?\\ntotal rmgrs = 256\\nbuilt in = 128 (index 0 - 127)\\ncustom = 128 (index 128 - 255)\\n\\n#define RM_N_IDS (UINT8_MAX + 1)\\n#define RM_N_BUILTIN_IDS (RM_N_IDS / 2)\\n#define RM_N_CUSTOM_IDS (RM_N_IDS / 2)\\n#define RM_MAX_ID (RM_N_IDS - 1)\\n#define RM_MAX_BUILTIN_ID (RM_NEXT_ID - 1)\\n#define RM_MIN_CUSTOM_ID (RM_N_IDS / 2)\\n#define RM_MAX_CUSTOM_ID RM_MAX_ID\\n#define RMID_IS_BUILTIN(rmid) ((rmid) <= RM_MAX_BUILTIN_ID)\\n#define RMID_IS_CUSTOM(rmid) ((rmid) >= RM_MIN_CUSTOM_ID && (rmid) <=\\nRM_MAX_CUSTOM_ID)\\n#define RMID_IS_VALID(rmid) (RMID_IS_BUILTIN((rmid)) || RMID_IS_CUSTOM((rmid)))\\n\\nIf okay, you can specify in a comment that out of UINT8_MAX + 1 rmgrs,\\nfirst half rmgr ids are supposed to be reserved for built in rmgrs and\\nsecond half rmgr ids for extensions.\\n\\n> I don't see the point in defining it as RM_MAX_ID/2+1.\\n>\\n> > 9) Thinking if there's a way to test the code, maybe exposing\\n> > RegisterCustomRmgr as a function? I think we need to have a dummy\\n> > extension that will be used for testing this kind of patches/features\\n> > but that's a separate discussion IMO.\\n>\\n> It's already exposed as a C function in xlog_internal.h.\\n>\\n> You mean as a SQL function? I don't think that makes sense. It can only\\n> be called while processing shared_preload_libraries (on server startup)\\n> anyway.\\n\\nYes, an SQL function showing all of the available rmgrs. With the\\ncustom rmgrs, I think, an SQL function like pg_resource_managers()\\nreturning a set of {name, id, (if possible - loaded_by_extension_name,\\nrm_redo, rm_desc, rm_identify ... api names)}\\n\\nSome more comments:\\n\\n1) Is there any specific reason to have a function to emit a PANIC\\nmessage? And it's being used only in one place in GetRmgr.\\n+void\\n+RmgrNotFound(RmgrId rmid)\\n\\n2) How about \\\"If rm_name is NULL, the corresponding RmgrTable[] entry\\nis considered invalid.\\\"?\\n+ * RmgrTable[] is indexed by RmgrId values (see rmgrlist.h). If rm_name is\\n+ * NULL, the structure is considered invalid.\\n\\n3) How about adding errdetail something like \\\"Custom resource manager\\nID must be between %d and %d (both inclusive)\\\", RM_MAX_BUILTIN_ID,\\nRM_MAX_ID)?\\n+ if (rmid < RM_MIN_CUSTOM_ID)\\n+ ereport(PANIC, errmsg(\\\"custom resource manager ID %d is out of range\\\", rmid));\\n\\n4) How about adding errdetail something like \\\"Custom resource manager\\nmust have a name\\\"\\n+ if (rmgr->rm_name == NULL)\\n+ ereport(PANIC, errmsg(\\\"custom resource manager is invalid\\\"));\\n\\n5) How about \\\"successfully registered custom resource manager\\\"?\\n+ (errmsg(\\\"registered custom resource manager \\\\\\\"%s\\\\\\\" with ID %d\\\",\\n\\n6) Shouldn't this be ri <= RM_NEXT_ID?\\n- for (ri = 0; ri < RM_NEXT_ID; ri++)\\n+ for (ri = 0; ri < RM_MAX_ID; ri++)\\n\\n- for (ri = 0; ri < RM_NEXT_ID; ri++)\\n+ for (ri = 0; ri <= RM_MAX_ID; ri++)\\n\\n7) Unrelated to this patch, why is there always an extra slot RM_MAX_ID + 1?\\n\\n-const RmgrDescData RmgrDescTable[RM_MAX_ID + 1] = {\\n+static const RmgrDescData RmgrDescTable[RM_MAX_BUILTIN_ID + 1] = {\\n #include \\\"access/rmgrlist.h\\\"\\n\\n-const RmgrData RmgrTable[RM_MAX_ID + 1] = {\\n+RmgrData RmgrTable[RM_MAX_ID + 1] = {\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Apr 2022 15:12:09 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Tue, 2022-04-05 at 15:12 +0530, Bharath Rupireddy wrote:\\n> Yes, I meant this. If we do this, then a global variable\\n> current_max_rmid can be maintained (by RegisterCustomRmgr) and the\\n> other functions RmgrStartup, RmgrCleanup and RegisterCustomRmgr can\\n> avoid for-loops with full range RM_MAX_ID (for (int id = 0; id <\\n> current_max_rmid; id++)).\\n\\nThere are only 255 entries, so the loops in those functions aren't a\\nsignificant overhead.\\n\\nIf we expand it, then we can come up a different solution. We can\\nchange it easily enough later, because the WAL format changes between\\nreleases and so do the APIs for any extension that might be using this.\\n\\n> With the custom rmgrs, now it's time to have it in the documentation\\n> (probably not immediately, even after this patch gets in) describing\\n> what rmgrs are, what are built-in and custom rmgrs and how and when\\n> an\\n> extension needs to register, load and use rmgrs etc.\\n\\nAdded a documentation section right after Generic WAL.\\n\\n> With the above, do you mean to say that only the extensions that load\\n> their library via shared_preload_libraries are allowed to have custom\\n> rmgrs? How about extensions using {session, local}_preload_libraries,\\n> LOAD command? How about extensions that don't load the shared library\\n> via {session, local, shared}_preload_libraries or LOAD command and\\n> let\\n> any of its functions load the shared library on first use?\\n\\nCorrect. The call to RegisterCustomRmgr *must* happen in the postmaster\\nbefore any other processes (even the startup process) are launched.\\nOtherwise recovery wouldn't work. The only place to make that happen is\\nshared_preload_libraries.\\n\\n> For instance - extension 1, with id 130 and name 'rmgr_foo' and\\n> extension 2 with id 131 and name 'Rmgr_foo'/'RMGR_FOO'...., then the\\n> following code may not catch it right? Do you want rmgr names to be\\n> case-sensitive/insensitive?\\n\\nYou convinced me. pg_waldump.c uses pg_strcasecmp(), so I'll use it as\\nwell.\\n\\n> #define RM_N_IDS (UINT8_MAX + 1)\\n> #define RM_N_BUILTIN_IDS (RM_N_IDS / 2)\\n> #define RM_N_CUSTOM_IDS (RM_N_IDS / 2)\\n> #define RM_MAX_ID (RM_N_IDS - 1)\\n> #define RM_MAX_BUILTIN_ID (RM_NEXT_ID - 1)\\n> #define RM_MIN_CUSTOM_ID (RM_N_IDS / 2)\\n> #define RM_MAX_CUSTOM_ID RM_MAX_ID\\n> #define RMID_IS_BUILTIN(rmid) ((rmid) <= RM_MAX_BUILTIN_ID)\\n> #define RMID_IS_CUSTOM(rmid) ((rmid) >= RM_MIN_CUSTOM_ID && (rmid) <=\\n> RM_MAX_CUSTOM_ID)\\n> #define RMID_IS_VALID(rmid) (RMID_IS_BUILTIN((rmid)) ||\\n> RMID_IS_CUSTOM((rmid)))\\n\\nI improved this a bit. But I didn't use your version that's based on\\ndivision, that was more confusing to me.\\n\\n> Yes, an SQL function showing all of the available rmgrs. With the\\n> custom rmgrs, I think, an SQL function like pg_resource_managers()\\n> returning a set of {name, id, (if possible -\\n> loaded_by_extension_name,\\n> rm_redo, rm_desc, rm_identify ... api names)}\\n\\nAdded.\\n\\n> \\n> Some more comments:\\n> \\n> 1) Is there any specific reason to have a function to emit a PANIC\\n> message? And it's being used only in one place in GetRmgr.\\n> +void\\n> +RmgrNotFound(RmgrId rmid)\\n\\nGood call -- better to raise an error, and it can get escalated if it\\nhappens in the wrong place. Done.\\n\\n> 2) How about \\\"If rm_name is NULL, the corresponding RmgrTable[] entry\\n> is considered invalid.\\\"?\\n> + * RmgrTable[] is indexed by RmgrId values (see rmgrlist.h). If\\n> rm_name is\\n> + * NULL, the structure is considered invalid.\\n\\nDone.\\n\\n> 3) How about adding errdetail something like \\\"Custom resource manager\\n> ID must be between %d and %d (both inclusive)\\\", RM_MAX_BUILTIN_ID,\\n> RM_MAX_ID)?\\n> + if (rmid < RM_MIN_CUSTOM_ID)\\n> + ereport(PANIC, errmsg(\\\"custom resource manager ID %d is out of\\n> range\\\", rmid));\\n\\nDone.\\n\\n> 4) How about adding errdetail something like \\\"Custom resource manager\\n> must have a name\\\"\\n> + if (rmgr->rm_name == NULL)\\n> + ereport(PANIC, errmsg(\\\"custom resource manager is invalid\\\"));\\n\\nImproved this message.\\n\\n> 5) How about \\\"successfully registered custom resource manager\\\"?\\n> + (errmsg(\\\"registered custom resource manager \\\\\\\"%s\\\\\\\" with ID %d\\\",\\n\\nThat's a little much, we don't want to sound too excited that it worked\\n;-)\\n\\n> 6) Shouldn't this be ri <= RM_NEXT_ID?\\n> - for (ri = 0; ri < RM_NEXT_ID; ri++)\\n> + for (ri = 0; ri < RM_MAX_ID; ri++)\\n> \\n> - for (ri = 0; ri < RM_NEXT_ID; ri++)\\n> + for (ri = 0; ri <= RM_MAX_ID; ri++)\\n\\nThat would leave out the custom rmgrs.\\n\\n> 7) Unrelated to this patch, why is there always an extra slot\\n> RM_MAX_ID + 1?\\n\\nRM_MAX_ID is 255, but we need 256 entries from 0-255 (inclusive).\\n\\nGood suggestions, thank you for the review!\\n\\nI'm happy with this patch and I committed it. Changes from last\\nversion:\\n\\n 1. I came up with a better solution (based on a suggestion from\\nAndres) to handle wal_consistency_checking properly. We can't\\nunderstand the custom resource managers when the configuration file is\\nfirst loaded, so if we enounter an unknown resource manager, I set a\\nflag and reprocess it after the shared_preload_libraries are loaded.\\n\\n 2. Added pg_get_wal_resource_managers().\\n\\n 3. Documentation.\\n\\nRegards,\\n\\tJeff Davis\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 06 Apr 2022 23:15:27 -0700\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-04-06 23:15:27 -0700, Jeff Davis wrote:\\n> I'm happy with this patch and I committed it.\\n\\nThat caused breakage with WAL_DEBUG enabled. Fixed that. But it was\\nhalf-broken before, at least since 70e81861fad, because 'rmid' didn't refer to\\nthe current record's rmid anymore, but to rmid from \\\"Initialize resource\\nmanagers\\\" - a constant.\\n\\n\\nCauses plenty new warnings here:\\n\\nIn file included from /home/andres/src/postgresql/src/include/access/xlogrecord.h:14,\\n from /home/andres/src/postgresql/src/include/access/xlogreader.h:41,\\n from /home/andres/src/postgresql/src/include/access/brin_xlog.h:17,\\n from /home/andres/src/postgresql/src/backend/access/transam/rmgr.c:10:\\n/home/andres/src/postgresql/src/backend/access/transam/rmgr.c: In function ‘RegisterCustomRmgr’:\\n/home/andres/src/postgresql/src/include/access/rmgr.h:42:66: warning: comparison is always true due to limited range of data type [-Wtype-limits]\\n 42 | (rmid) <= RM_MAX_CUSTOM_ID)\\n | ^~\\n/home/andres/src/postgresql/src/backend/access/transam/rmgr.c:104:14: note: in expansion of macro ‘RMID_IS_CUSTOM’\\n 104 | if (!RMID_IS_CUSTOM(rmid))\\n | ^~~~~~~~~~~~~~\\nIn file included from /home/andres/src/postgresql/src/include/c.h:814,\\n from /home/andres/src/postgresql/src/include/postgres.h:46,\\n from /home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c:9:\\n/home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c: In function ‘GetRmgrDesc’:\\n/home/andres/src/postgresql/src/include/access/rmgr.h:42:66: warning: comparison is always true due to limited range of data type [-Wtype-limits]\\n 42 | (rmid) <= RM_MAX_CUSTOM_ID)\\n | ^~\\n/home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c:89:9: note: in expansion of macro ‘Assert’\\n 89 | Assert(RMID_IS_VALID(rmid));\\n | ^~~~~~\\n/home/andres/src/postgresql/src/include/access/rmgr.h:43:57: note: in expansion of macro ‘RMID_IS_CUSTOM’\\n 43 | #define RMID_IS_VALID(rmid) (RMID_IS_BUILTIN((rmid)) || RMID_IS_CUSTOM((rmid)))\\n | ^~~~~~~~~~~~~~\\n/home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c:89:16: note: in expansion of macro ‘RMID_IS_VALID’\\n 89 | Assert(RMID_IS_VALID(rmid));\\n | ^~~~~~~~~~~~~\\n/home/andres/src/postgresql/src/include/access/rmgr.h:42:66: warning: comparison is always true due to limited range of data type [-Wtype-limits]\\n 42 | (rmid) <= RM_MAX_CUSTOM_ID)\\n | ^~\\n/home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c:89:9: note: in expansion of macro ‘Assert’\\n 89 | Assert(RMID_IS_VALID(rmid));\\n | ^~~~~~\\n/home/andres/src/postgresql/src/include/access/rmgr.h:43:57: note: in expansion of macro ‘RMID_IS_CUSTOM’\\n 43 | #define RMID_IS_VALID(rmid) (RMID_IS_BUILTIN((rmid)) || RMID_IS_CUSTOM((rmid)))\\n | ^~~~~~~~~~~~~~\\n/home/andres/src/postgresql/src/bin/pg_waldump/rmgrdesc.c:89:16: note: in expansion of macro ‘RMID_IS_VALID’\\n 89 | Assert(RMID_IS_VALID(rmid));\\n | ^~~~~~~~~~~~~\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Apr 2022 23:35:05 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-04-06 23:35:05 -0700, Andres Freund wrote:\\n> Causes plenty new warnings here:\\n\\nAnd my machine isn't alone. There's also:\\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=lapwing&dt=2022-04-07%2006%3A40%3A14\\n\\nrmgrdesc.c:44:1: error: missing braces around initializer [-Werror=missing-braces]\\nrmgrdesc.c:44:1: error: (near initialization for 'CustomNumericNames[0]') [-Werror=missing-braces]\\nrmgrdesc.c:45:1: error: missing braces around initializer [-Werror=missing-braces]\\nrmgrdesc.c:45:1: error: (near initialization for 'CustomRmgrDesc[0]') [-Werror=missing-braces]\\ncc1: all warnings being treated as errors\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Apr 2022 23:56:40 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-04-06 23:56:40 -0700, Andres Freund wrote:\\n> On 2022-04-06 23:35:05 -0700, Andres Freund wrote:\\n> > Causes plenty new warnings here:\\n> \\n> And my machine isn't alone. There's also:\\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=lapwing&dt=2022-04-07%2006%3A40%3A14\\n> \\n> rmgrdesc.c:44:1: error: missing braces around initializer [-Werror=missing-braces]\\n> rmgrdesc.c:44:1: error: (near initialization for 'CustomNumericNames[0]') [-Werror=missing-braces]\\n> rmgrdesc.c:45:1: error: missing braces around initializer [-Werror=missing-braces]\\n> rmgrdesc.c:45:1: error: (near initialization for 'CustomRmgrDesc[0]') [-Werror=missing-braces]\\n> cc1: all warnings being treated as errors\\n\\n(pushing a fix for this one in a few)\\n\\nThere's also:\\n\\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=wrasse&dt=2022-04-07%2006%3A49%3A14\\n\\nccache /opt/developerstudio12.6/bin/cc -m64 -Xa -v -O findtimezone.o initdb.o localtime.o -L../../../src/port -L../../../src/common -L../../../src/fe_utils -lpgfeutils -L../../../src/common -lpgcommon -L../../../src/port -lpgport -L../../../src/interfaces/libpq -lpq -L/home/nm/sw/nopath/uuid-64/lib -L/home/nm/sw/nopath/openldap-64/lib -Wl,--as-needed -Wl,-R'/home/nm/farm/studio64v12_6/HEAD/inst/lib' -lpgcommon -lpgport -lxslt -lxml2 -lpam -lssl -lcrypto -lgssapi_krb5 -lz -lreadline -ltermcap -lnsl -lsocket -lm -o initdb\\nUndefined\\t\\t\\tfirst referenced\\n symbol \\t\\t\\t in file\\nRmgrTable initdb.o\\n\\nIIRC sun studio is annoying because it emits references to global variables\\nwhen they're mentioned in a static inline, even if that inline is unused.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Apr 2022 00:45:34 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-04-07 00:45:34 -0700, Andres Freund wrote:\\n> On 2022-04-06 23:56:40 -0700, Andres Freund wrote:\\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=wrasse&dt=2022-04-07%2006%3A49%3A14\\n> \\n> ccache /opt/developerstudio12.6/bin/cc -m64 -Xa -v -O findtimezone.o initdb.o localtime.o -L../../../src/port -L../../../src/common -L../../../src/fe_utils -lpgfeutils -L../../../src/common -lpgcommon -L../../../src/port -lpgport -L../../../src/interfaces/libpq -lpq -L/home/nm/sw/nopath/uuid-64/lib -L/home/nm/sw/nopath/openldap-64/lib -Wl,--as-needed -Wl,-R'/home/nm/farm/studio64v12_6/HEAD/inst/lib' -lpgcommon -lpgport -lxslt -lxml2 -lpam -lssl -lcrypto -lgssapi_krb5 -lz -lreadline -ltermcap -lnsl -lsocket -lm -o initdb\\n> Undefined\\t\\t\\tfirst referenced\\n> symbol \\t\\t\\t in file\\n> RmgrTable initdb.o\\n> \\n> IIRC sun studio is annoying because it emits references to global variables\\n> when they're mentioned in a static inline, even if that inline is unused.\\n\\nLooks like a #ifndef FRONTEND around the inlines should work? Or just go back\\nto the macros, but without the RM_MAX_CUSTOM_ID check that's implied due to\\nthe type width anyway.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Apr 2022 01:16:26 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"On Thu, Apr 7, 2022 at 11:45 AM Jeff Davis <pgsql@j-davis.com> wrote:\\n> I'm happy with this patch and I committed it. Changes from last\\n> version:\\n\\nHi Jeff, I think there's a typo [1] in pg_waldump.c, we might miss to\\ntake account of the last custom resource mgr stats. Please fix it.\\n\\ndiff --git a/src/bin/pg_waldump/pg_waldump.c b/src/bin/pg_waldump/pg_waldump.c\\nindex 7d92dcaf87..30ca7684bd 100644\\n--- a/src/bin/pg_waldump/pg_waldump.c\\n+++ b/src/bin/pg_waldump/pg_waldump.c\\n@@ -720,7 +720,7 @@ XLogDumpDisplayStats(XLogDumpConfig *config,\\nXLogDumpStats *stats)\\n * calculate column totals.\\n */\\n\\n- for (ri = 0; ri < RM_MAX_ID; ri++)\\n+ for (ri = 0; ri <= RM_MAX_ID; ri++)\\n {\\n total_count += stats->rmgr_stats[ri].count;\\n total_rec_len += stats->rmgr_stats[ri].rec_len;\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Apr 2022 14:00:08 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n },\n {\n \"msg_contents\": \"Jeff Davis <pgsql@j-davis.com> writes:\\n> I'm happy with this patch and I committed it.\\n\\nwrasse is less happy:\\n\\nccache /opt/developerstudio12.6/bin/cc -m64 -Xa -v -O findtimezone.o initdb.o localtime.o -L../../../src/port -L../../../src/common -L../../../src/fe_utils -lpgfeutils -L../../../src/common -lpgcommon -L../../../src/port -lpgport -L../../../src/interfaces/libpq -lpq -L/home/nm/sw/nopath/uuid-64/lib -L/home/nm/sw/nopath/openldap-64/lib -Wl,--as-needed -Wl,-R'/home/nm/farm/studio64v12_6/HEAD/inst/lib' -lpgcommon -lpgport -lxslt -lxml2 -lpam -lssl -lcrypto -lgssapi_krb5 -lz -lreadline -ltermcap -lnsl -lsocket -lm -o initdb\\nUndefined\\t\\t\\tfirst referenced\\n symbol \\t\\t\\t in file\\nRmgrTable initdb.o\\nld: fatal: symbol referencing errors\\n\\nI think this means the static inline functions you added in\\nxlog_internal.h need to be wrapped in #ifndef FRONTEND.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Apr 2022 10:05:15 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extensible Rmgr for Table AMs\"\n }\n]","isConversational":false}}},{"rowIdx":53058,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\n\nI have noticed $subject while looking at a patch in this area:\nhttps://www.postgresql.org/message-id/YYiqxtEqhgb5gIWL@paquier.xyz\n\nWhile we don't expect the four callers of XLogReadRecord() related to\nlogical decoding to fail in the code paths changed by the patch\nattached, I think that it would be good to provide more context to\nusers rather than just emitting an elog(ERROR) coming directly from\nthe WAL reader. This way, it becomes a bit easier to guess from where\nthe failure is coming.\n\nA patch is attached to improve all those elog()s. I have tried not to\nbe too fancy about those new error strings, but I am sure that these\ncould be tweaked.\n\nThoughts?\n--\nMichael","msg_date":"Tue, 9 Nov 2021 10:47:11 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":true,"msg_subject":"Improve error context after some failed XLogReadRecord()"},{"msg_contents":"At Tue, 9 Nov 2021 10:47:11 +0900, Michael Paquier <michael@paquier.xyz> wrote in \n> Hi all,\n> \n> I have noticed $subject while looking at a patch in this area:\n> https://www.postgresql.org/message-id/YYiqxtEqhgb5gIWL@paquier.xyz\n> \n> While we don't expect the four callers of XLogReadRecord() related to\n> logical decoding to fail in the code paths changed by the patch\n> attached, I think that it would be good to provide more context to\n> users rather than just emitting an elog(ERROR) coming directly from\n> the WAL reader. This way, it becomes a bit easier to guess from where\n> the failure is coming.\n> \n> A patch is attached to improve all those elog()s. I have tried not to\n> be too fancy about those new error strings, but I am sure that these\n> could be tweaked.\n> \n> Thoughts?\n\nAs the whole, I agree that any substantial message is far better than\n\"%s\" in regard to identifying the line that emitted a log line.\n\n+\t\t\t\telog(ERROR, \"could not find record to advance replication slot: %s\",\n\nI prefer \"could not read record while advancing replcation slot\",\nbecause it is not finding a record but it is reading successive\nrecords. However, I don't strongly object to the current wording.\n\n+\t\telog(ERROR, \"could not find record to send logically-decoded data: %s\", errm);\n\nI prefer something like the above for the same reason. How about\n\"could not read record while sending logically-decoded data:\". The\nsame with above, I'm not strongly opposed to the current wording.\n\nOtherwise it looks good to me.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Tue, 09 Nov 2021 16:58:11 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Improve error context after some failed XLogReadRecord()"},{"msg_contents":"On Tue, Nov 09, 2021 at 04:58:11PM +0900, Kyotaro Horiguchi wrote:\n> Otherwise it looks good to me.\n\nWhat you are proposing is better than what I had in those two\ncontexts, so I have just used what you have suggested for both of\nthem. Thanks!\n--\nMichael","msg_date":"Wed, 10 Nov 2021 15:59:16 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":true,"msg_subject":"Re: Improve error context after some failed XLogReadRecord()"}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\n\\nI have noticed $subject while looking at a patch in this area:\\nhttps://www.postgresql.org/message-id/YYiqxtEqhgb5gIWL@paquier.xyz\\n\\nWhile we don't expect the four callers of XLogReadRecord() related to\\nlogical decoding to fail in the code paths changed by the patch\\nattached, I think that it would be good to provide more context to\\nusers rather than just emitting an elog(ERROR) coming directly from\\nthe WAL reader. This way, it becomes a bit easier to guess from where\\nthe failure is coming.\\n\\nA patch is attached to improve all those elog()s. I have tried not to\\nbe too fancy about those new error strings, but I am sure that these\\ncould be tweaked.\\n\\nThoughts?\\n--\\nMichael\",\n \"msg_date\": \"Tue, 9 Nov 2021 10:47:11 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Improve error context after some failed XLogReadRecord()\"\n },\n {\n \"msg_contents\": \"At Tue, 9 Nov 2021 10:47:11 +0900, Michael Paquier <michael@paquier.xyz> wrote in \\n> Hi all,\\n> \\n> I have noticed $subject while looking at a patch in this area:\\n> https://www.postgresql.org/message-id/YYiqxtEqhgb5gIWL@paquier.xyz\\n> \\n> While we don't expect the four callers of XLogReadRecord() related to\\n> logical decoding to fail in the code paths changed by the patch\\n> attached, I think that it would be good to provide more context to\\n> users rather than just emitting an elog(ERROR) coming directly from\\n> the WAL reader. This way, it becomes a bit easier to guess from where\\n> the failure is coming.\\n> \\n> A patch is attached to improve all those elog()s. I have tried not to\\n> be too fancy about those new error strings, but I am sure that these\\n> could be tweaked.\\n> \\n> Thoughts?\\n\\nAs the whole, I agree that any substantial message is far better than\\n\\\"%s\\\" in regard to identifying the line that emitted a log line.\\n\\n+\\t\\t\\t\\telog(ERROR, \\\"could not find record to advance replication slot: %s\\\",\\n\\nI prefer \\\"could not read record while advancing replcation slot\\\",\\nbecause it is not finding a record but it is reading successive\\nrecords. However, I don't strongly object to the current wording.\\n\\n+\\t\\telog(ERROR, \\\"could not find record to send logically-decoded data: %s\\\", errm);\\n\\nI prefer something like the above for the same reason. How about\\n\\\"could not read record while sending logically-decoded data:\\\". The\\nsame with above, I'm not strongly opposed to the current wording.\\n\\nOtherwise it looks good to me.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Tue, 09 Nov 2021 16:58:11 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Improve error context after some failed XLogReadRecord()\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 09, 2021 at 04:58:11PM +0900, Kyotaro Horiguchi wrote:\\n> Otherwise it looks good to me.\\n\\nWhat you are proposing is better than what I had in those two\\ncontexts, so I have just used what you have suggested for both of\\nthem. Thanks!\\n--\\nMichael\",\n \"msg_date\": \"Wed, 10 Nov 2021 15:59:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Improve error context after some failed XLogReadRecord()\"\n }\n]","isConversational":false}}},{"rowIdx":53059,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Below is the second pass over the patches in the current commitfest, this time\nthe patches which have been around for 5 or 6 commitfests. The usual\ndisclaimers that I very well might have missed something still apply.\n\n\n2773: libpq compression\n=======================\nThis patch intended to provide libpq connection compression to \"replace SSL\ncompression\" which was doomed when the patch was written, and have since been\nremoved altogether. The initial approach didn't get much traction but there\nwas significant discussion and work, which has since fizzled out. The patch\nhas been updated but there hasn't been meaningful review the past months, the\nlast comments seem to imply there being a fair amount of questionmarks left in\nhere. Robert, having been very involved in this do you have any thoughts on\nwhere we are and where to go (if at all IYO)?\n\n\n2765: Add extra statistics to explain for Nested Loop\n=====================================================\nThere has been almost exclusively positive feedback on this idea, and the patch\nhas matured over several rounds of fixes. Any committer keen on picking this\nup and getting it across the finishline?\n\n\n2780: Allow COPY \"text\" to output a header and add header matching mode to COPY\nFROM\n===============================================================================\nThe original patch was rejected for lacking matching, but it has since been\naddressed in an updated patch which has seen a lot of review. Reading the\nthread it seems this should really be in Ready for Committer.\n\n\n2815: CREATE INDEX CONCURRENTLY on partitioned table\n====================================================\nThis spawned off from the work on REINDEX CONCURRENTLY but seems to have\nstalled and is awaiting more review.\n\n\n2800: CLUSTER on partitioned table\n==================================\nThe the patch above, it too spawned off from the REINDEX CONCURRENTLY thread\nbut seems to have stalled awaiting more review. There was scepticism raised\nover the usefulness vs code weight tradeoff, but the response to this haven't\nreally been followed up on. Alvaro and Michael, do you have any changed\nthoughts on Justins patch reorg or do you stil feel this isn't worth going\nahead with?\n\n\n2897: Faster pglz compression\n=============================\nTomas did benchmarking recently which shows promising speedups and is actively\nworking on getting it ready for commit.\n\n\n2906: Row filtering for logical replication\n===========================================\nThis patch is undergoing lots of active review.\n\n\n2831: Extended statistics / estimate Var op Var clauses\n=======================================================\nThe patch and approach has been approved by reviewers, it's down to the final\npush over the finishline from what I can gather. Seems to be awaiting\nbenchmarks by Tomas on the relative costs of the new operations in planning.\n\n\n2918: New default role allowing to change per-role/database settings\n====================================================================\nStephen, you had comments on the patch which have since been addressed. Any\nthoughts on the new revision?\n\n\n2911: Dynamic result sets from procedures\n=========================================\nThis patch is gated on the SHOW_ALL_RESULTS patch, and is thus not moving until\nthere is further progress on that one.\n\n\n2901: SQL/JSON: functions\n2901: SQL/JSON: JSON_TABLE\n==========================\nGrouping these two together as they are in very similar states. The patches\nseems to mostly move forward by rebases fixing smaller things. Andrew, do you\nhave a sense of how far these are from a committable state?\n\n\n2838: brin: avoid errors processing work item following concurrent reindex\n==========================================================================\nAlvaro proposed an alternative fix, and the thread has since stalled. Should\nthis fix be applied and the CF entry closed?\n\n\n2877: Fix firing of RI triggers during cross-partition updates of partitioned\ntables\n=============================================================================\nThis is fixing a live issue in the backbranches, and would be good to get\nclosure on. Amit, you mentioned a problem with the patch involving deferrable\nforeign constraints, is this all fixed in the latest posted revision? (sorry\nif that's spelled out and I missed it)\n\n\n2900: On client connection event trigger\n========================================\nThis implements a client login event trigger commonly used for auditing and\nconnection setup tasks. I've reviewed this a fair bit and I think it's pretty\nmuch committable, but there is an ongoing discussion on whether it needs an\nescape-hatch GUC to disable such triggers in case they brick the database. I\nam against such a GUC but there has been a lot of support voiced in the thread\nfor it. If we can put that to rest I think we can close this patch.\n\n\n2903: Parallel Hash Full Join\n=============================\nThis patch is blocked on a reproducible crash which is being worked on.\n\n\n2864: Fix pg_rewind race condition just after promotion\n2869: Mitigate pg_rewind race condition, if config file is enlarged\nconcurrently.\n===================================================================\nGrouping similar entries. Both fix live bugs (one reported in 2015), and both\npatches are reviewed and approved since months ago. Seems like we should be\nable to close these ones in the CF.\n\n\n2824: Functions for sorting GiST build of gist_btree types\n==========================================================\nThis seems to get close to be committable after bug investigation by Heikki.\n\n\n2837: pg_stat_statements and \"IN\" conditions\n============================================\nThis patch aims to consolidate (based on a configurable threshold) SELECT ..\nWHERE c IN (a, b, ...) queries to variations of the same query from cluttering\nup pg_stat_statements. Seems like a good idea, and the patch has been tested\nand reviewed to the point where it seems ready for committer to me.\n\n\n2839: possibility to rename root namespace in plpgsql\n=====================================================\nThis patch addresses a feature request for making plpgsql easier to work with,\nbut the approach has been questioned with the thread stalling since March in\none of these discussion. Is there interest in picking this up or should we\nclose it as Returned with Feedback?\n\n\n2863: Printing backtrace of postgres processes\n==============================================\nThis patch is being actively worked on and reviewed.\n\n\n2851: Consider parallel for LATERAL subqueries having LIMIT/OFFSET\n==================================================================\nA nice little fix for supporting parallel query execution for LATERAL ..\nLIMIT/OFFSET queries. Recently marked ready for committer, so let's try and\nget this in.\n\n\n2861: pg_upgrade test for binary compatibility of core data types\n=================================================================\nThis patch is also actively worked on and reviewed, and bits from this thread\nhave already landed. Do you think this can land in its entirety in this CF\nMichael?\n\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Tue, 9 Nov 2021 15:12:39 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"Daniel Gustafsson <daniel@yesql.se> writes:\n> 2773: libpq compression\n> =======================\n> This patch intended to provide libpq connection compression to \"replace SSL\n> compression\" which was doomed when the patch was written, and have since been\n> removed altogether. The initial approach didn't get much traction but there\n> was significant discussion and work, which has since fizzled out. The patch\n> has been updated but there hasn't been meaningful review the past months, the\n> last comments seem to imply there being a fair amount of questionmarks left in\n> here. Robert, having been very involved in this do you have any thoughts on\n> where we are and where to go (if at all IYO)?\n\nI'm not Robert, but I still have an opinion here, and that it's that this\nfeature would at best be an attractive nuisance. If you need compression\non a database session, it probably means that the connection is over the\nopen internet, which means that you need encryption even more. And we\nknow that compression and encryption do not play well together. The\nreason compression was taken out of the latest TLS standards is not that\nthey wouldn't have liked to have it, nor that applying compression in a\nseparate code layer would be any safer. I fear offering this would\nmerely lead people to build CVE-worthy setups.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 09 Nov 2021 10:02:30 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"Hi\n\n2839: possibility to rename root namespace in plpgsql\n> =====================================================\n> This patch addresses a feature request for making plpgsql easier to work\n> with,\n> but the approach has been questioned with the thread stalling since March\n> in\n> one of these discussion. Is there interest in picking this up or should we\n> close it as Returned with Feedback?\n>\n\nThis feature is from the category \"nice to have, but we can live without it\nwell\". There are few requests per decade about this feature. I think my\ndesign is good enough, but surely it is not the only one.\n\nWe cannot use inspiration from Oracle's PL/SQL or from Ada, because this\nrequires a feature that is not supported in Postgres (pragmas). Without\npragma is relatively hard to find good design. PL/pgSQL options are not\nnative elements of language (I think so it is the best option), and other\ndesigns are not fully correct (although they can look more natural). I\ndon't see any passion (negative or positive) about this feature, and I\nthink this feature can be returned without feedback. Maybe somebody will\ncome with other proposals, maybe we will have pragmas one day.\n\nMuch more important is extending the debug support of PL/pgSQL. This patch\ncan simplify implementations of PL/pgSQL debuggers, and this feature can be\na real benefit for PL/pgSQL developers.\n\nRegards\n\nPavel\n\nHi\n2839: possibility to rename root namespace in plpgsql\n=====================================================\nThis patch addresses a feature request for making plpgsql easier to work with,\nbut the approach has been questioned with the thread stalling since March in\none of these discussion. Is there interest in picking this up or should we\nclose it as Returned with Feedback?This feature is from the category \"nice to have, but we can live without it well\". There are few requests per decade about this feature. I think my design is good enough, but surely it is not the only one.We cannot use inspiration from Oracle's PL/SQL or from Ada, because this requires a feature that is not supported in Postgres (pragmas). Without pragma is relatively hard to find good design. PL/pgSQL options are not native elements of language (I think so it is the best option), and other designs are not fully correct (although they can look more natural). I don't see any passion (negative or positive) about this feature, and I think this feature can be returned without feedback. Maybe somebody will come with other proposals, maybe we will have pragmas one day.Much more important is extending the debug support of PL/pgSQL. This patch can simplify implementations of PL/pgSQL debuggers, and this feature can be a real benefit for PL/pgSQL developers.RegardsPavel","msg_date":"Tue, 9 Nov 2021 18:11:06 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"Greetings,\n\n* Tom Lane (tgl@sss.pgh.pa.us) wrote:\n> Daniel Gustafsson <daniel@yesql.se> writes:\n> > 2773: libpq compression\n> > =======================\n> > This patch intended to provide libpq connection compression to \"replace SSL\n> > compression\" which was doomed when the patch was written, and have since been\n> > removed altogether. The initial approach didn't get much traction but there\n> > was significant discussion and work, which has since fizzled out. The patch\n> > has been updated but there hasn't been meaningful review the past months, the\n> > last comments seem to imply there being a fair amount of questionmarks left in\n> > here. Robert, having been very involved in this do you have any thoughts on\n> > where we are and where to go (if at all IYO)?\n> \n> I'm not Robert, but I still have an opinion here, and that it's that this\n> feature would at best be an attractive nuisance. If you need compression\n> on a database session, it probably means that the connection is over the\n> open internet, which means that you need encryption even more. And we\n> know that compression and encryption do not play well together. The\n> reason compression was taken out of the latest TLS standards is not that\n> they wouldn't have liked to have it, nor that applying compression in a\n> separate code layer would be any safer. I fear offering this would\n> merely lead people to build CVE-worthy setups.\n\nI've got an opinion on this also and I don't agree that needing\ncompression means you're on the open internet or that we shouldn't allow\nusers the choice to decide if they want to use compression and\nencryption together or not. Yes, there's potential risks there, but I\ndon't think those risks would lead to CVEs against PG for supporting a\nmode where we allow compression and then also allow encryption- if\nthat's a problem then it's an issue for the encryption library/method\nbeing used and isn't the fault of us for allowing that combination.\n\nFurther, the general issue with encryption+compression attacks is when\nthe attacker is able to inject chosen plaintext content into what ends\nup being compressed and encrypted, and then is able to observe the\ncompressed+encrypted traffic, and that's not nearly as easy to do when\nyou're talking about a database connection vs. a case where an attacker\nis able to convince a victim to go to a malicious website and through\nthat is able to do cross-site requests to inject the known plaintext\ninto requests to a trusted webpage and see the results coming back.\n\nThat we don't have any support for such a thing in the PG protocol and\nthe general case that app server to DB traffic can certainly be a lot\nand mean that compression is worthwhile even on a private network, and\nthat users may wish to encrypt even on a private network, are all good\nreasons why we should support both compression and encryption and allow\nusers to choose what makes sense for them in their environment.\n\nThanks,\n\nStephen","msg_date":"Tue, 9 Nov 2021 12:43:20 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"> Daniel Gustafsson <daniel@yesql.se> writes:\n> > 2773: libpq compression\n> > =======================\n> > This patch intended to provide libpq connection compression to\n> > \"replace SSL compression\" which was doomed when the patch was\n> written,\n> > and have since been removed altogether. The initial approach didn't\n> > get much traction but there was significant discussion and work, which\n> > has since fizzled out. The patch has been updated but there hasn't\n> > been meaningful review the past months, the last comments seem to\n> > imply there being a fair amount of questionmarks left in here.\n> > Robert, having been very involved in this do you have any thoughts on\n> where we are and where to go (if at all IYO)?\n> \n> I'm not Robert, but I still have an opinion here, and that it's that this feature\n> would at best be an attractive nuisance. If you need compression on a\n> database session, it probably means that the connection is over the open\n> internet, which means that you need encryption even more. And we know\n> that compression and encryption do not play well together. The reason\n> compression was taken out of the latest TLS standards is not that they\n> wouldn't have liked to have it, nor that applying compression in a separate\n> code layer would be any safer. I fear offering this would merely lead people\n> to build CVE-worthy setups.\n> \n\nI don't have a strong opinion on the patch but I'd just like to mention that there are definitely use cases for using compression on private networks (with ssl off). We have cases where we have a PG extension that hooks into the COPY ... FROM command and accept lz4-compressed data for the COPY data message specifically. This has proven invaluable to us even on private networks, because it can give a good compression ratio for very little CPU decompression overhead.\n\n\n","msg_date":"Tue, 9 Nov 2021 19:28:32 +0000","msg_from":"Floris Van Nee <florisvannee@Optiver.com>","msg_from_op":false,"msg_subject":"RE: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"\nOn 11/9/21 09:12, Daniel Gustafsson wrote:\n>\n> 2901: SQL/JSON: functions\n> 2901: SQL/JSON: JSON_TABLE\n> ==========================\n> Grouping these two together as they are in very similar states. The patches\n> seems to mostly move forward by rebases fixing smaller things. Andrew, do you\n> have a sense of how far these are from a committable state?\n>\n>\n\nFor various reasons I'm going to be hors de combat for most of the rest\nof the year, back at the beginning of January, although I will be around\nif necessary, and probably tinkering with one or two things. I think at\nleast the first set is not in bad shape, but it would benefit from some\nfresh eyeballs. Tom Kincaid is trying to arrange some extra resources\nfor that, so we might well see some progress soon.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Tue, 9 Nov 2021 16:34:34 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"On Tue, Nov 9, 2021 at 12:43:20PM -0500, Stephen Frost wrote:\n> * Tom Lane (tgl@sss.pgh.pa.us) wrote:\n> > Daniel Gustafsson <daniel@yesql.se> writes:\n> > I'm not Robert, but I still have an opinion here, and that it's that this\n> > feature would at best be an attractive nuisance. If you need compression\n> > on a database session, it probably means that the connection is over the\n> > open internet, which means that you need encryption even more. And we\n> > know that compression and encryption do not play well together. The\n> > reason compression was taken out of the latest TLS standards is not that\n> > they wouldn't have liked to have it, nor that applying compression in a\n> > separate code layer would be any safer. I fear offering this would\n> > merely lead people to build CVE-worthy setups.\n> \n> I've got an opinion on this also and I don't agree that needing\n> compression means you're on the open internet or that we shouldn't allow\n> users the choice to decide if they want to use compression and\n> encryption together or not. Yes, there's potential risks there, but I\n> don't think those risks would lead to CVEs against PG for supporting a\n> mode where we allow compression and then also allow encryption- if\n> that's a problem then it's an issue for the encryption library/method\n> being used and isn't the fault of us for allowing that combination.\n\nYeah, I am leaning this diretion too. On the one hand, the number of\nways to exploit compression-then-encryption seem to continue to grow,\nbut it is the complex behavior of HTTPS and connecting to multiple sites\nwhich seems to always be required, and we don't do that with the\nPostgres client/server protocol.\n\n-- \n Bruce Momjian <bruce@momjian.us> https://momjian.us\n EDB https://enterprisedb.com\n\n If only the physical world exists, free will is an illusion.\n\n\n\n","msg_date":"Tue, 9 Nov 2021 18:21:43 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"On Tue, Nov 09, 2021 at 03:12:39PM +0100, Daniel Gustafsson wrote:\n> 2800: CLUSTER on partitioned table\n> ==================================\n> The the patch above, it too spawned off from the REINDEX CONCURRENTLY thread\n> but seems to have stalled awaiting more review. There was scepticism raised\n> over the usefulness vs code weight tradeoff, but the response to this haven't\n> really been followed up on. Alvaro and Michael, do you have any changed\n> thoughts on Justins patch reorg or do you stil feel this isn't worth going\n> ahead with?\n\nMy feeling on this one is that the addition was not worth the\ncomplexity, and I got the impression that the handling of\npg_index.indisclustered was rather blurry when it came to multiple\nlayers of partitions.\n\n> 2861: pg_upgrade test for binary compatibility of core data types\n> =================================================================\n> This patch is also actively worked on and reviewed, and bits from this thread\n> have already landed. Do you think this can land in its entirety in this CF\n> Michael?\n\nThe remaining parts are the addition of a SQL test, which requires\nmore comments and documentation as the thing is complex to follow, and\nthe switch of the upgrade SQL queries to a file. I need to do a pass\nover it, but that seems worth having on its own. For the second\npoint, I wanted to double-check first with Andrew about how much code\nwe could cut in the buildfarm client code.\n\nThanks for the summary!\n--\nMichael","msg_date":"Wed, 10 Nov 2021 21:42:06 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"\n\n> Daniel Gustafsson <daniel@yesql.se> writes:\n> \n>> 2773: libpq compression\n>> =======================\n>> This patch intended to provide libpq connection compression to \"replace SSL\n>> compression\" which was doomed when the patch was written, and have since been\n>> removed altogether. The initial approach didn't get much traction but there\n>> was significant discussion and work, which has since fizzled out. The patch\n>> has been updated but there hasn't been meaningful review the past months, the\n>> last comments seem to imply there being a fair amount of questionmarks left in\n>> here. Robert, having been very involved in this do you have any thoughts on\n>> where we are and where to go (if at all IYO)?\n> \n> I'm not Robert, but I still have an opinion here, and that it's that this\n> feature would at best be an attractive nuisance. If you need compression\n> on a database session, it probably means that the connection is over the\n> open internet, which means that you need encryption even more. And we\n> know that compression and encryption do not play well together. The\n> reason compression was taken out of the latest TLS standards is not that\n> they wouldn't have liked to have it, nor that applying compression in a\n> separate code layer would be any safer. I fear offering this would\n> merely lead people to build CVE-worthy setups.\n> \n\nCompression is crucial for highly available setups. Replication traffic is often billed. Or route has bandwidth limits.\nAn entropy added by WAL headers makes CRIME attack against replication encryption impractical.\n\nBest regards, Andrey Borodin.\n\n\n","msg_date":"Wed, 10 Nov 2021 20:54:21 +0500","msg_from":"Andrey Borodin <x4mmm@yandex-team.ru>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"> On 9 Nov 2021, at 18:43, Stephen Frost <sfrost@snowman.net> wrote:\n> * Tom Lane (tgl@sss.pgh.pa.us) wrote:\n>> Daniel Gustafsson <daniel@yesql.se> writes:\n>>> 2773: libpq compression\n>>> =======================\n>>> This patch intended to provide libpq connection compression to \"replace SSL\n>>> compression\" which was doomed when the patch was written, and have since been\n>>> removed altogether. The initial approach didn't get much traction but there\n>>> was significant discussion and work, which has since fizzled out. The patch\n>>> has been updated but there hasn't been meaningful review the past months, the\n>>> last comments seem to imply there being a fair amount of questionmarks left in\n>>> here. Robert, having been very involved in this do you have any thoughts on\n>>> where we are and where to go (if at all IYO)?\n>> \n>> I'm not Robert, but I still have an opinion here, and that it's that this\n>> feature would at best be an attractive nuisance. If you need compression\n>> on a database session, it probably means that the connection is over the\n>> open internet, which means that you need encryption even more. And we\n>> know that compression and encryption do not play well together. The\n>> reason compression was taken out of the latest TLS standards is not that\n>> they wouldn't have liked to have it, nor that applying compression in a\n>> separate code layer would be any safer. I fear offering this would\n>> merely lead people to build CVE-worthy setups.\n> \n> I've got an opinion on this also and I don't agree that needing\n> compression means you're on the open internet or that we shouldn't allow\n> users the choice to decide if they want to use compression and\n> encryption together or not. Yes, there's potential risks there, but I\n> don't think those risks would lead to CVEs against PG for supporting a\n> mode where we allow compression and then also allow encryption- if\n> that's a problem then it's an issue for the encryption library/method\n> being used and isn't the fault of us for allowing that combination.\n\nI'm not so sure, if the parts aren't vulnerable separately on their own but the\ncombination of them + libpq allows for weakening encryption in a TLS stream I'd\nsay the CVE is on us.\n\n> Further, the general issue with encryption+compression attacks is when\n> the attacker is able to inject chosen plaintext content into what ends\n> up being compressed and encrypted, and then is able to observe the\n> compressed+encrypted traffic, and that's not nearly as easy to do when\n> you're talking about a database connection vs. a case where an attacker\n> is able to convince a victim to go to a malicious website and through\n> that is able to do cross-site requests to inject the known plaintext\n> into requests to a trusted webpage and see the results coming back.\n\nI agree with this, the known weaknesses rely on mechanisms that aren't readily\napplicable to postgres connections. I don't remember if the patch already does\nthis, but if not we should ensure we only negotiate compression after\nauthentication like how OpenSSH does [0].\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n[0] https://www.openssh.com/txt/draft-miller-secsh-compression-delayed-00.txt\n\n","msg_date":"Fri, 12 Nov 2021 17:04:08 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"On Tue, Nov 9, 2021 at 7:02 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> Daniel Gustafsson <daniel@yesql.se> writes:\n> > 2773: libpq compression\n> > =======================\n> > This patch intended to provide libpq connection compression to \"replace\n> SSL\n> > compression\" which was doomed when the patch was written, and have since\n> been\n> > removed altogether. The initial approach didn't get much traction but\n> there\n> > was significant discussion and work, which has since fizzled out. The\n> patch\n> > has been updated but there hasn't been meaningful review the past\n> months, the\n> > last comments seem to imply there being a fair amount of questionmarks\n> left in\n> > here. Robert, having been very involved in this do you have any\n> thoughts on\n> > where we are and where to go (if at all IYO)?\n>\n> I'm not Robert, but I still have an opinion here, and that it's that this\n> feature would at best be an attractive nuisance. If you need compression\n> on a database session, it probably means that the connection is over the\n> open internet, which means that you need encryption even more.\n>\n\nThis assumption is very vague. I personally had multiple cases when network\nbandwidth for app <--> Postgres communication, that were fixed wither via\nupgrades (spending more money) or making application developers cache more\ndata (optimization), or, usually, both. Never public internet connections\nwere involved.\n\nActually, any growing startup experiences such issues sooner or later.\nNetwork compression would be a great tool for many setups, and some other\npopular database systems offer it for a long.\n\nOn Tue, Nov 9, 2021 at 7:02 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:Daniel Gustafsson <daniel@yesql.se> writes:\n> 2773: libpq compression\n> =======================\n> This patch intended to provide libpq connection compression to \"replace SSL\n> compression\" which was doomed when the patch was written, and have since been\n> removed altogether. The initial approach didn't get much traction but there\n> was significant discussion and work, which has since fizzled out. The patch\n> has been updated but there hasn't been meaningful review the past months, the\n> last comments seem to imply there being a fair amount of questionmarks left in\n> here. Robert, having been very involved in this do you have any thoughts on\n> where we are and where to go (if at all IYO)?\n\nI'm not Robert, but I still have an opinion here, and that it's that this\nfeature would at best be an attractive nuisance. If you need compression\non a database session, it probably means that the connection is over the\nopen internet, which means that you need encryption even more. This assumption is very vague. I personally had multiple cases when network bandwidth for app <--> Postgres communication, that were fixed wither via upgrades (spending more money) or making application developers cache more data (optimization), or, usually, both. Never public internet connections were involved.Actually, any growing startup experiences such issues sooner or later. Network compression would be a great tool for many setups, and some other popular database systems offer it for a long.","msg_date":"Fri, 12 Nov 2021 13:14:18 -0800","msg_from":"Nikolay Samokhvalov <samokhvalov@gmail.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"\n\nOn 11/10/21 00:21, Bruce Momjian wrote:\n> On Tue, Nov 9, 2021 at 12:43:20PM -0500, Stephen Frost wrote:\n>> * Tom Lane (tgl@sss.pgh.pa.us) wrote:\n>>> Daniel Gustafsson <daniel@yesql.se> writes:\n>>> I'm not Robert, but I still have an opinion here, and that it's that this\n>>> feature would at best be an attractive nuisance. If you need compression\n>>> on a database session, it probably means that the connection is over the\n>>> open internet, which means that you need encryption even more. And we\n>>> know that compression and encryption do not play well together. The\n>>> reason compression was taken out of the latest TLS standards is not that\n>>> they wouldn't have liked to have it, nor that applying compression in a\n>>> separate code layer would be any safer. I fear offering this would\n>>> merely lead people to build CVE-worthy setups.\n>>\n>> I've got an opinion on this also and I don't agree that needing\n>> compression means you're on the open internet or that we shouldn't allow\n>> users the choice to decide if they want to use compression and\n>> encryption together or not. Yes, there's potential risks there, but I\n>> don't think those risks would lead to CVEs against PG for supporting a\n>> mode where we allow compression and then also allow encryption- if\n>> that's a problem then it's an issue for the encryption library/method\n>> being used and isn't the fault of us for allowing that combination.\n> \n> Yeah, I am leaning this diretion too. On the one hand, the number of\n> ways to exploit compression-then-encryption seem to continue to grow,\n> but it is the complex behavior of HTTPS and connecting to multiple sites\n> which seems to always be required, and we don't do that with the\n> Postgres client/server protocol.\n> \n\n+0.5\n\nI don't think the number of attacks grows all that much, actually. Or \nmore precisely, the new vulnerabilities are just variations on the CRIME \npattern adopting it to different contexts.\n\nFor example Voracle [1] applies to openvpn connections, but in principle \nit's the same scenario - observing changes to the compressed size, after \ntweaking the requests.\n\n[1] https://community.openvpn.net/openvpn/wiki/VORACLE\n\nSo it's true compression is a side-channel, but these attacks require \nother ingredients too - ability to MITM the connection, and ability to \ninject custom data.\n\nDoing MITM is probably harder than when attacking HTTPS, because DB \nconnections tend to be between app server and db server - so setting up \na rogue AP does not really work. Sometimes people / DBAs may connect to \nthe DB directly, although in a way that makes it hard to do modify the \ndata (which is why psql/ssh are immune to this). But it's possible.\n\nBut how hard is injecting custom data in a way that'd make CRIME-like \nissue possible? Yes, you could hijack webapp similarly to CRIME, in a \nway that initiates database requests with custom data. But we don't \ngenerally send both the supplied and actual password to the DB to do the \ncheck there - we just send one of them, and check it against the value \nstored in the DB.\n\nThe one (somewhat) plausible scenario I can think of is an application doing\n\n UPDATE t SET field = user_data WHERE id = 1;\n SELECT field FROM t WHERE ...;\n\nin which case we might learn if some other row in the result contains \nthe same value. Not sure how practical this could be, though.\n\nAFAIK OpenVPN still supports tunnel compression, but it's disabled by \ndefault. That probably says something about what they think about the \nfeature, although that might be due to having compression and not \nwanting to just remove it.\n\nOne interesting idea OpenVPN allows is asymmetric compression, i.e. \nallowing compression in just one direction. For example it might be \npossible for the client to send compressed data, but the server would \nnever send compressed. Or vice versa. That might still help with e.g. \nlarge imports or exports, depending on the use case, while maybe \nmitigating some issues. But it's clearly not a solid defense.\n\n\nregards\n\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Sat, 13 Nov 2021 01:09:52 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"On 11/10/21 16:54, Andrey Borodin wrote:\n> \n> \n>> Daniel Gustafsson <daniel@yesql.se> writes:\n>>\n>>> 2773: libpq compression\n>>> =======================\n>>> This patch intended to provide libpq connection compression to \"replace SSL\n>>> compression\" which was doomed when the patch was written, and have since been\n>>> removed altogether. The initial approach didn't get much traction but there\n>>> was significant discussion and work, which has since fizzled out. The patch\n>>> has been updated but there hasn't been meaningful review the past months, the\n>>> last comments seem to imply there being a fair amount of questionmarks left in\n>>> here. Robert, having been very involved in this do you have any thoughts on\n>>> where we are and where to go (if at all IYO)?\n>>\n>> I'm not Robert, but I still have an opinion here, and that it's that this\n>> feature would at best be an attractive nuisance. If you need compression\n>> on a database session, it probably means that the connection is over the\n>> open internet, which means that you need encryption even more. And we\n>> know that compression and encryption do not play well together. The\n>> reason compression was taken out of the latest TLS standards is not that\n>> they wouldn't have liked to have it, nor that applying compression in a\n>> separate code layer would be any safer. I fear offering this would\n>> merely lead people to build CVE-worthy setups.\n>>\n> \n> Compression is crucial for highly available setups. Replication traffic is often billed. Or route has bandwidth limits.\n> An entropy added by WAL headers makes CRIME attack against replication encryption impractical.\n\nI very much doubt WAL headers are a reliable protection against CRIME, \nbecause the entropy of the headers is likely fairly constant. So if you \ncompress the WAL stream, the WAL headers may change but the compression \nratio should be pretty similar. At least that's my guess.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Sat, 13 Nov 2021 01:14:31 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"\n\n> On 11/10/21 16:54, Andrey Borodin wrote:\n> \n>> Compression is crucial for highly available setups. Replication traffic is often billed. Or route has bandwidth limits.\n>> An entropy added by WAL headers makes CRIME attack against replication encryption impractical.\n> \n> I very much doubt WAL headers are a reliable protection against CRIME,\n> because the entropy of the headers is likely fairly constant. So if you\n> compress the WAL stream, the WAL headers may change but the compression\n> ratio should be pretty similar. At least that's my guess.\n\nI've thought more about it and I agree.\nTo reliably protect against CRIME entropy of WAL headers must be comparable with the entropy of possibly injected data.\nIf this would stand, probably, our WAL would need a really serious rework.\n\nMaybe just refuse to enable compression on SSL connection? If someone really needs both - they will just patch a server on their own.\nOr make a GUC \"yes_i_kwow_what_crime_is_give_grant_read_on_my_data_to_spies\".\n\nBest regards, Andrey Borodin.\n\n\n","msg_date":"Sun, 14 Nov 2021 14:59:51 +0500","msg_from":"Andrey Borodin <x4mmm@yandex-team.ru>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"Greetings,\n\n* Andrey Borodin (x4mmm@yandex-team.ru) wrote:\n> > On 11/10/21 16:54, Andrey Borodin wrote:\n> >> Compression is crucial for highly available setups. Replication traffic is often billed. Or route has bandwidth limits.\n> >> An entropy added by WAL headers makes CRIME attack against replication encryption impractical.\n> > \n> > I very much doubt WAL headers are a reliable protection against CRIME,\n> > because the entropy of the headers is likely fairly constant. So if you\n> > compress the WAL stream, the WAL headers may change but the compression\n> > ratio should be pretty similar. At least that's my guess.\n> \n> I've thought more about it and I agree.\n> To reliably protect against CRIME entropy of WAL headers must be comparable with the entropy of possibly injected data.\n> If this would stand, probably, our WAL would need a really serious rework.\n> \n> Maybe just refuse to enable compression on SSL connection? If someone really needs both - they will just patch a server on their own.\n> Or make a GUC \"yes_i_kwow_what_crime_is_give_grant_read_on_my_data_to_spies\".\n\nAttackers aren't likely to have the kind of isolated control over the\ndata in the WAL stream (which is a combination of data from lots of\nongoing activity in the system and isn't likely to be exactly what the\nattacker supplied at some higher level anyway) and the ability to read\nand analyze the WAL stream from a primary to a replica to be able to\neffectively attack it.\n\nNone of this kind of babysitting makes any sense to me.\n\nI'm certainly fine with documenting that there are cases where\ncompression of data sent by an attacker and then sent over an encrypted\nchannel has been able to break the encryption and let admins decide if\nit's appropriate or not for them to use in their environment.\n\nThanks,\n\nStephen","msg_date":"Sun, 14 Nov 2021 14:13:12 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"Stephen Frost <sfrost@snowman.net> writes:\n> Attackers aren't likely to have the kind of isolated control over the\n> data in the WAL stream (which is a combination of data from lots of\n> ongoing activity in the system and isn't likely to be exactly what the\n> attacker supplied at some higher level anyway) and the ability to read\n> and analyze the WAL stream from a primary to a replica to be able to\n> effectively attack it.\n\nYeah, I concur with that so far as WAL data goes. A hypothetical attacker\nwill not have control over xact IDs, tuple TIDs, etc, which will add\nenough entropy to the stream that extracting data payloads seems pretty\ninfeasible.\n\nMy concern upthread was about client-session connections, where such\nmitigation doesn't apply. (I wonder a bit about logical-replication\nstreams, too.)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 14 Nov 2021 14:26:33 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"On Tue, Nov 9, 2021 at 9:12 AM Daniel Gustafsson <daniel@yesql.se> wrote:\n> 2773: libpq compression\n> =======================\n> This patch intended to provide libpq connection compression to \"replace SSL\n> compression\" which was doomed when the patch was written, and have since been\n> removed altogether. The initial approach didn't get much traction but there\n> was significant discussion and work, which has since fizzled out. The patch\n> has been updated but there hasn't been meaningful review the past months, the\n> last comments seem to imply there being a fair amount of questionmarks left in\n> here. Robert, having been very involved in this do you have any thoughts on\n> where we are and where to go (if at all IYO)?\n\nIt's been a while since I looked at that patch, and my recollections\nare somewhat fuzzy. However, I think that my main feeling was that the\nquality of the patch didn't seem to be that close to what I felt would\nbe required for a commit. I think at that point there were still\nsignificant questions about the details of the approach as well as\ncode-quality and performance concerns. It may have gotten better since\nlast I looked.\n\nI do agree with the comments from others that there is room to debate\n(1) the usefulness of this and (2) the degree to which it might leave\nus on the hook for security problems that wouldn't be our\nresponsibility if we left this up to OpenSSL (or maybe some other SSL\nlibrary, if we ever get pluggability there, which I hope we do). But\nmy concerns were more pragmatic than theoretical.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 15 Nov 2021 09:32:29 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"> On 15 Nov 2021, at 15:32, Robert Haas <robertmhaas@gmail.com> wrote:\n> \n> On Tue, Nov 9, 2021 at 9:12 AM Daniel Gustafsson <daniel@yesql.se> wrote:\n>> 2773: libpq compression\n>> =======================\n>> This patch intended to provide libpq connection compression to \"replace SSL\n>> compression\" which was doomed when the patch was written, and have since been\n>> removed altogether. The initial approach didn't get much traction but there\n>> was significant discussion and work, which has since fizzled out. The patch\n>> has been updated but there hasn't been meaningful review the past months, the\n>> last comments seem to imply there being a fair amount of questionmarks left in\n>> here. Robert, having been very involved in this do you have any thoughts on\n>> where we are and where to go (if at all IYO)?\n> \n> It's been a while since I looked at that patch, and my recollections\n> are somewhat fuzzy.\n\nNo worries, thanks for sharing!\n\n> However, I think that my main feeling was that the\n> quality of the patch didn't seem to be that close to what I felt would\n> be required for a commit. I think at that point there were still\n> significant questions about the details of the approach as well as\n> code-quality and performance concerns. It may have gotten better since\n> last I looked.\n\nContext switching into this thread is hard, so it's not clear how much the\npatch has moved since you reviewed, or how the current patchset relates to the\nearlier one as there seems to have been a shift at a recent(ish) point in time.\nThe last message (which is unreadable HTML soup in the archives) talks about\nimplementing lz4 with \"Looking forward to post an update soon.\"\n\nGiven the discussions had and the doubts over approach I wonder if the best\ncourse of action with the CF entry in question is to close it, hash out the\nwhat and how on -hackers and reopen a new entry when there is a patch which.\nWe are not doing ourselves, or the CF process, any favors by moving an entry\nalong which isn't all that close to committable.\n\nThoughts (addressing everyone involved/interested)?\n\n> I do agree with the comments from others that there is room to debate\n> (1) the usefulness of this and (2) the degree to which it might leave\n> us on the hook for security problems that wouldn't be our\n> responsibility if we left this up to OpenSSL (or maybe some other SSL\n> library, if we ever get pluggability there, which I hope we do). But\n> my concerns were more pragmatic than theoretical.\n\nI share you hope here =) However I don't think we'll get much help with\ncompression from the TLS libraries. OpenSSL has deprecated transport\ncompression, NSS never implemented it in the first place, and TLS 1.3 removes\nit altogether. That being said, this is mainly due to protocols where chosen\nplaintext attacks are practical/possible. For our case we, IMHO, should look\nmore at how OpenSSH and other protocols closer to ours has built compression in\na safe way, to see if we can achieve the level of transport safety we require.\n\nBut, there is I think broad concensus that we at least must fully understand\nthe tradeoffs should we allow compression.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Mon, 15 Nov 2021 20:05:05 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"Greetings,\n\n* Daniel Gustafsson (daniel@yesql.se) wrote:\n> > On 15 Nov 2021, at 15:32, Robert Haas <robertmhaas@gmail.com> wrote:\n> > On Tue, Nov 9, 2021 at 9:12 AM Daniel Gustafsson <daniel@yesql.se> wrote:\n> >> 2773: libpq compression\n> >> =======================\n> >> This patch intended to provide libpq connection compression to \"replace SSL\n> >> compression\" which was doomed when the patch was written, and have since been\n> >> removed altogether. The initial approach didn't get much traction but there\n> >> was significant discussion and work, which has since fizzled out. The patch\n> >> has been updated but there hasn't been meaningful review the past months, the\n> >> last comments seem to imply there being a fair amount of questionmarks left in\n> >> here. Robert, having been very involved in this do you have any thoughts on\n> >> where we are and where to go (if at all IYO)?\n> > \n> > It's been a while since I looked at that patch, and my recollections\n> > are somewhat fuzzy.\n> \n> No worries, thanks for sharing!\n> \n> > However, I think that my main feeling was that the\n> > quality of the patch didn't seem to be that close to what I felt would\n> > be required for a commit. I think at that point there were still\n> > significant questions about the details of the approach as well as\n> > code-quality and performance concerns. It may have gotten better since\n> > last I looked.\n> \n> Context switching into this thread is hard, so it's not clear how much the\n> patch has moved since you reviewed, or how the current patchset relates to the\n> earlier one as there seems to have been a shift at a recent(ish) point in time.\n> The last message (which is unreadable HTML soup in the archives) talks about\n> implementing lz4 with \"Looking forward to post an update soon.\"\n\nNot sure where the best place to drop this comment is, but the patch\nseems to be doing exactly what I was hoping it didn't- it's just taking\nthe ultimate results and compressing them and sending them on to the\nclient, at least that's what it looks like in a quick look.\n\nWhat about when that data has already been compressed and stored\ncompressed and not manipulated at all? The initial post in that thread\ntalks about returning data from 'large JSON columns'. With this\napproach, we would likely end up taking compressed json/jsonb data,\ndecompressing it, and then re-compressing it to send to the client.\n\nThat seems ... less than ideal.\n\nI get that just compressing the entire stream is simpler and easier and\nsuch, but it's surely cheaper and more efficient to not decompress and\nthen recompress data that's already compressed. Finding a way to pass\nthrough data that's already compressed when stored as-is while also\nsupporting compression of everything else (in a sensible way- wouldn't\nmake sense to just compress each attribute independently since a 4 byte\ninteger isn't going to get smaller with compression) definitely\ncomplicates the overall idea but perhaps would be possible to do.\n\n> Given the discussions had and the doubts over approach I wonder if the best\n> course of action with the CF entry in question is to close it, hash out the\n> what and how on -hackers and reopen a new entry when there is a patch which.\n> We are not doing ourselves, or the CF process, any favors by moving an entry\n> along which isn't all that close to committable.\n> \n> Thoughts (addressing everyone involved/interested)?\n\nYeah, agreed with this in general.\n\nThanks,\n\nStephen","msg_date":"Mon, 15 Nov 2021 14:51:46 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"On Mon, Nov 15, 2021 at 2:51 PM Stephen Frost <sfrost@snowman.net> wrote:\n> I get that just compressing the entire stream is simpler and easier and\n> such, but it's surely cheaper and more efficient to not decompress and\n> then recompress data that's already compressed. Finding a way to pass\n> through data that's already compressed when stored as-is while also\n> supporting compression of everything else (in a sensible way- wouldn't\n> make sense to just compress each attribute independently since a 4 byte\n> integer isn't going to get smaller with compression) definitely\n> complicates the overall idea but perhaps would be possible to do.\n\nTo me, this feels like an attempt to move the goalposts far enough to\nkill the project. Sure, in a perfect world, that would be nice. But,\nwe don't do it anywhere else. If you try to store a JPEG into a bytea\ncolumn, we'll try to compress it just like we would any other data,\nand it may not work out. If you then take a pg_basebackup of the\ndatabase using -Z, there's no attempt made to avoid the overhead of\nCPU overhead of compressing those TOAST table pages that contain\nalready-compressed data and not the others. And it's easy to\nunderstand why that's the case: when you insert data into the\ndatabase, there's no way for the database to magically know whether\nthat data has been previously compressed by some means, and if so, how\neffectively. And when you back up a database, the backup doesn't know\nwhich relfilenodes contain TOAST tables or which pages of those\nrelfilenodes contain that is already pre-compressed. In both cases,\nyour options are either (1) shut off compression yourself or (2) hope\nthat the compressor doesn't waste too much effort on it.\n\nI think the same approach ought to be completely acceptable here. I\ndon't even really understand how we could do anything else. printtup()\njust gets datums, and it has no idea whether or how they are toasted.\nIt calls the type output functions which don't know that data is being\nprepared for transmission to the client as opposed to some other\nhypothetical way you could call that function, nor do they know what\ncompression method the client wants. It does not seem at all\nstraightforward to teach them that ... and even if they did, what\nthen? It's not like every column value is sent as a separate packet;\nthe whole row is a single protocol message, and some columns may be\ncompressed and others uncompressed. Trying to guess what to do about\nthat seems to boil down to a sheer guess. Unless you try to compress\nthat mixture of compressed and uncompressed values - and it's\nmoderately uncommon for every column of a table to be even be\ntoastable - you aren't going to know how well it will compress. You\ncould easily waste more CPU cycles trying to guess than you would have\nspent just doing what the user asked for.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 15 Nov 2021 15:23:17 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"On Mon, Nov 15, 2021 at 03:23:17PM -0500, Robert Haas wrote:\n> On Mon, Nov 15, 2021 at 2:51 PM Stephen Frost <sfrost@snowman.net> wrote:\n> > I get that just compressing the entire stream is simpler and easier and\n> > such, but it's surely cheaper and more efficient to not decompress and\n> > then recompress data that's already compressed. Finding a way to pass\n> > through data that's already compressed when stored as-is while also\n> > supporting compression of everything else (in a sensible way- wouldn't\n> > make sense to just compress each attribute independently since a 4 byte\n> > integer isn't going to get smaller with compression) definitely\n> > complicates the overall idea but perhaps would be possible to do.\n> \n> To me, this feels like an attempt to move the goalposts far enough to\n> kill the project.\n\nI don't know what version of the patch Stephen looked at, but the patches I\nlooked at were only compressing things like CopyData, which seems like it would\nallow using compression when it'd be effective, without adding overhead to\ncases where it won't.\n\nIMO, the patch set should be presented differently, to use the existing zlib\nand lz4 in 0001, and then add zstd in 0002 or later.\n\n-- \nJustin\n\n\n","msg_date":"Mon, 15 Nov 2021 14:55:00 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"Greetings,\n\n* Robert Haas (robertmhaas@gmail.com) wrote:\n> On Mon, Nov 15, 2021 at 2:51 PM Stephen Frost <sfrost@snowman.net> wrote:\n> > I get that just compressing the entire stream is simpler and easier and\n> > such, but it's surely cheaper and more efficient to not decompress and\n> > then recompress data that's already compressed. Finding a way to pass\n> > through data that's already compressed when stored as-is while also\n> > supporting compression of everything else (in a sensible way- wouldn't\n> > make sense to just compress each attribute independently since a 4 byte\n> > integer isn't going to get smaller with compression) definitely\n> > complicates the overall idea but perhaps would be possible to do.\n> \n> To me, this feels like an attempt to move the goalposts far enough to\n> kill the project. Sure, in a perfect world, that would be nice. But,\n> we don't do it anywhere else. If you try to store a JPEG into a bytea\n> column, we'll try to compress it just like we would any other data,\n> and it may not work out. If you then take a pg_basebackup of the\n> database using -Z, there's no attempt made to avoid the overhead of\n> CPU overhead of compressing those TOAST table pages that contain\n> already-compressed data and not the others. And it's easy to\n> understand why that's the case: when you insert data into the\n> database, there's no way for the database to magically know whether\n> that data has been previously compressed by some means, and if so, how\n> effectively. And when you back up a database, the backup doesn't know\n> which relfilenodes contain TOAST tables or which pages of those\n> relfilenodes contain that is already pre-compressed. In both cases,\n> your options are either (1) shut off compression yourself or (2) hope\n> that the compressor doesn't waste too much effort on it.\n\nI'll grant that perhaps it's just a different project.\n\nWhile, sure, we will try to compress things we don't understand by\ndefault, the admin does have ways to tell us to not do that and so it\nisn't always just guess-work. As for other parts of the system not\nbeing smarter- they should be. I'd argue that up until quite recently\nit didn't make sense to teach something like pg_basebackup about TOAST\ntables because our compression was rather lacking and it was actually\nuseful to compress TOAST tables (perhaps still is with the compression\nmethods we have now, but we'll hopefully add more). We should probably\nlook at trying to provide a way for pg_basebackup and other backup tools\nto identify TOAST tables to avoid trying to compress them as it's just\nwasted effort (though in that case, it's at least less wasted than here-\nwe don't decompress that TOAST data and then re-compress it).\n\nUsing bytea for everything-and-the-kitchen-sink also just generally\nseems like it's throwing away more information than we really should (if\nwe had a jpeg data type, for example, we could just have that default to\nnot being compressed and not spend time trying to compress jpegs).\n\nI also just generally disagree that an inefficiency in one part of the\nsystem justifies having it somewhere else.\n\n> I think the same approach ought to be completely acceptable here. I\n> don't even really understand how we could do anything else. printtup()\n> just gets datums, and it has no idea whether or how they are toasted.\n> It calls the type output functions which don't know that data is being\n> prepared for transmission to the client as opposed to some other\n> hypothetical way you could call that function, nor do they know what\n> compression method the client wants. It does not seem at all\n> straightforward to teach them that ... and even if they did, what\n> then? It's not like every column value is sent as a separate packet;\n> the whole row is a single protocol message, and some columns may be\n> compressed and others uncompressed. Trying to guess what to do about\n> that seems to boil down to a sheer guess. Unless you try to compress\n> that mixture of compressed and uncompressed values - and it's\n> moderately uncommon for every column of a table to be even be\n> toastable - you aren't going to know how well it will compress. You\n> could easily waste more CPU cycles trying to guess than you would have\n> spent just doing what the user asked for.\n\nI agree that we wouldn't want to be doing this all based on guesswork\nsomehow. Having a way to identify what's compressed and what isn't\nwould be needed, as well as what method was used to compress. It's not\nlike we don't know that at the source, we'd just need a way to pass that\nback through- probably more-or-less exactly as how it's stored today.\n\nTackling the actual protocol side of it seems like it'd be the more\nsensible place to start working through a design and I don't think\nthat'd be trivial to do, so probably not something to try and hash out\nin this particular thread. I'm happy to have at least voiced that\nthought.\n\nThanks,\n\nStephen","msg_date":"Mon, 15 Nov 2021 15:58:57 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"Hi! It’s been a while since the original patch release. Let me provide a brief \noverview of the current patch status.\n\nThe initial approach was to use the streaming compression to compress all \noutgoing and decompress all incoming bytes. However, after the long discussion \nin the thread, the initial approach has been changed.\n\nThe current implementation allows compressing only specific message types, \nuse the different compression algorithms for different message types, configure \nthe allowed compression methods and levels both for server- and client- sides \nvia GUC setting / connection string respectively.\n\nAlso, current implementation combines (when possible) multiple protocol messages \ninto the single CompressedData message for a better compression ratio.\n\n> \n> On 16 Nov 2021, at 01:23, Robert Haas <robertmhaas@gmail.com> wrote:\n> \n> To me, this feels like an attempt to move the goalposts far enough to\n> kill the project. Sure, in a perfect world, that would be nice. But,\n> we don't do it anywhere else. If you try to store a JPEG into a bytea\n> column, we'll try to compress it just like we would any other data,\n> and it may not work out. If you then take a pg_basebackup of the\n> database using -Z, there's no attempt made to avoid the overhead of\n> CPU overhead of compressing those TOAST table pages that contain\n> already-compressed data and not the others. And it's easy to\n> understand why that's the case: when you insert data into the\n> database, there's no way for the database to magically know whether\n> that data has been previously compressed by some means, and if so, how\n> effectively. And when you back up a database, the backup doesn't know\n> which relfilenodes contain TOAST tables or which pages of those\n> relfilenodes contain that is already pre-compressed. In both cases,\n> your options are either (1) shut off compression yourself or (2) hope\n> that the compressor doesn't waste too much effort on it.\n> \n> I think the same approach ought to be completely acceptable here. I\n> don't even really understand how we could do anything else. printtup()\n> just gets datums, and it has no idea whether or how they are toasted.\n> It calls the type output functions which don't know that data is being\n> prepared for transmission to the client as opposed to some other\n> hypothetical way you could call that function, nor do they know what\n> compression method the client wants. It does not seem at all\n> straightforward to teach them that ... and even if they did, what\n> then? It's not like every column value is sent as a separate packet;\n> the whole row is a single protocol message, and some columns may be\n> compressed and others uncompressed. Trying to guess what to do about\n> that seems to boil down to a sheer guess. Unless you try to compress\n> that mixture of compressed and uncompressed values - and it's\n> moderately uncommon for every column of a table to be even be\n> toastable - you aren't going to know how well it will compress. You\n> could easily waste more CPU cycles trying to guess than you would have\n> spent just doing what the user asked for.\n> \n\nAgree. From my POV, it is OK to use the protocol message type and length to decide \nshould it be compressed or not. Also, this can be optimized later without the need to change \nthe protocol.\n\nRegarding the LZ4 support patch, it still has some minor polishing to do. Basically, it only adds the LZ4 \nalgorithm support and does not change anything fundamentally. So I would appreciate someone doing \na review of the current patch version.\n\nThe original thread is quite huge so I guess that it makes it hard to catch up with the current patch status.\nI can make a new one with a detailed summary if that would help.\n\nThanks,\n\nDaniil Zakhlystov\n\n\n\n","msg_date":"Tue, 16 Nov 2021 22:40:47 +0500","msg_from":"Daniil Zakhlystov <usernamedt@yandex-team.ru>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"},{"msg_contents":"> \n> 2780: Allow COPY \"text\" to output a header and add header matching mode to COPY\n> FROM\n> ===============================================================================\n> The original patch was rejected for lacking matching, but it has since been\n> addressed in an updated patch which has seen a lot of review. Reading the\n> thread it seems this should really be in Ready for Committer.\n\nI don’t think there is anything more that needs to be done for this patch. I have been checking that it apply cleanly and that the tests still pass but it has already multiple rounds of reviews and is probably ready.\n\nBest regards,\nRémi\n\n","msg_date":"Mon, 29 Nov 2021 21:26:48 +0100","msg_from":"=?utf-8?Q?R=C3=A9mi_Lapeyre?= <remi.lapeyre@lenstra.fr>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 2"}],"string":"[\n {\n \"msg_contents\": \"Below is the second pass over the patches in the current commitfest, this time\\nthe patches which have been around for 5 or 6 commitfests. The usual\\ndisclaimers that I very well might have missed something still apply.\\n\\n\\n2773: libpq compression\\n=======================\\nThis patch intended to provide libpq connection compression to \\\"replace SSL\\ncompression\\\" which was doomed when the patch was written, and have since been\\nremoved altogether. The initial approach didn't get much traction but there\\nwas significant discussion and work, which has since fizzled out. The patch\\nhas been updated but there hasn't been meaningful review the past months, the\\nlast comments seem to imply there being a fair amount of questionmarks left in\\nhere. Robert, having been very involved in this do you have any thoughts on\\nwhere we are and where to go (if at all IYO)?\\n\\n\\n2765: Add extra statistics to explain for Nested Loop\\n=====================================================\\nThere has been almost exclusively positive feedback on this idea, and the patch\\nhas matured over several rounds of fixes. Any committer keen on picking this\\nup and getting it across the finishline?\\n\\n\\n2780: Allow COPY \\\"text\\\" to output a header and add header matching mode to COPY\\nFROM\\n===============================================================================\\nThe original patch was rejected for lacking matching, but it has since been\\naddressed in an updated patch which has seen a lot of review. Reading the\\nthread it seems this should really be in Ready for Committer.\\n\\n\\n2815: CREATE INDEX CONCURRENTLY on partitioned table\\n====================================================\\nThis spawned off from the work on REINDEX CONCURRENTLY but seems to have\\nstalled and is awaiting more review.\\n\\n\\n2800: CLUSTER on partitioned table\\n==================================\\nThe the patch above, it too spawned off from the REINDEX CONCURRENTLY thread\\nbut seems to have stalled awaiting more review. There was scepticism raised\\nover the usefulness vs code weight tradeoff, but the response to this haven't\\nreally been followed up on. Alvaro and Michael, do you have any changed\\nthoughts on Justins patch reorg or do you stil feel this isn't worth going\\nahead with?\\n\\n\\n2897: Faster pglz compression\\n=============================\\nTomas did benchmarking recently which shows promising speedups and is actively\\nworking on getting it ready for commit.\\n\\n\\n2906: Row filtering for logical replication\\n===========================================\\nThis patch is undergoing lots of active review.\\n\\n\\n2831: Extended statistics / estimate Var op Var clauses\\n=======================================================\\nThe patch and approach has been approved by reviewers, it's down to the final\\npush over the finishline from what I can gather. Seems to be awaiting\\nbenchmarks by Tomas on the relative costs of the new operations in planning.\\n\\n\\n2918: New default role allowing to change per-role/database settings\\n====================================================================\\nStephen, you had comments on the patch which have since been addressed. Any\\nthoughts on the new revision?\\n\\n\\n2911: Dynamic result sets from procedures\\n=========================================\\nThis patch is gated on the SHOW_ALL_RESULTS patch, and is thus not moving until\\nthere is further progress on that one.\\n\\n\\n2901: SQL/JSON: functions\\n2901: SQL/JSON: JSON_TABLE\\n==========================\\nGrouping these two together as they are in very similar states. The patches\\nseems to mostly move forward by rebases fixing smaller things. Andrew, do you\\nhave a sense of how far these are from a committable state?\\n\\n\\n2838: brin: avoid errors processing work item following concurrent reindex\\n==========================================================================\\nAlvaro proposed an alternative fix, and the thread has since stalled. Should\\nthis fix be applied and the CF entry closed?\\n\\n\\n2877: Fix firing of RI triggers during cross-partition updates of partitioned\\ntables\\n=============================================================================\\nThis is fixing a live issue in the backbranches, and would be good to get\\nclosure on. Amit, you mentioned a problem with the patch involving deferrable\\nforeign constraints, is this all fixed in the latest posted revision? (sorry\\nif that's spelled out and I missed it)\\n\\n\\n2900: On client connection event trigger\\n========================================\\nThis implements a client login event trigger commonly used for auditing and\\nconnection setup tasks. I've reviewed this a fair bit and I think it's pretty\\nmuch committable, but there is an ongoing discussion on whether it needs an\\nescape-hatch GUC to disable such triggers in case they brick the database. I\\nam against such a GUC but there has been a lot of support voiced in the thread\\nfor it. If we can put that to rest I think we can close this patch.\\n\\n\\n2903: Parallel Hash Full Join\\n=============================\\nThis patch is blocked on a reproducible crash which is being worked on.\\n\\n\\n2864: Fix pg_rewind race condition just after promotion\\n2869: Mitigate pg_rewind race condition, if config file is enlarged\\nconcurrently.\\n===================================================================\\nGrouping similar entries. Both fix live bugs (one reported in 2015), and both\\npatches are reviewed and approved since months ago. Seems like we should be\\nable to close these ones in the CF.\\n\\n\\n2824: Functions for sorting GiST build of gist_btree types\\n==========================================================\\nThis seems to get close to be committable after bug investigation by Heikki.\\n\\n\\n2837: pg_stat_statements and \\\"IN\\\" conditions\\n============================================\\nThis patch aims to consolidate (based on a configurable threshold) SELECT ..\\nWHERE c IN (a, b, ...) queries to variations of the same query from cluttering\\nup pg_stat_statements. Seems like a good idea, and the patch has been tested\\nand reviewed to the point where it seems ready for committer to me.\\n\\n\\n2839: possibility to rename root namespace in plpgsql\\n=====================================================\\nThis patch addresses a feature request for making plpgsql easier to work with,\\nbut the approach has been questioned with the thread stalling since March in\\none of these discussion. Is there interest in picking this up or should we\\nclose it as Returned with Feedback?\\n\\n\\n2863: Printing backtrace of postgres processes\\n==============================================\\nThis patch is being actively worked on and reviewed.\\n\\n\\n2851: Consider parallel for LATERAL subqueries having LIMIT/OFFSET\\n==================================================================\\nA nice little fix for supporting parallel query execution for LATERAL ..\\nLIMIT/OFFSET queries. Recently marked ready for committer, so let's try and\\nget this in.\\n\\n\\n2861: pg_upgrade test for binary compatibility of core data types\\n=================================================================\\nThis patch is also actively worked on and reviewed, and bits from this thread\\nhave already landed. Do you think this can land in its entirety in this CF\\nMichael?\\n\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 15:12:39 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"Daniel Gustafsson <daniel@yesql.se> writes:\\n> 2773: libpq compression\\n> =======================\\n> This patch intended to provide libpq connection compression to \\\"replace SSL\\n> compression\\\" which was doomed when the patch was written, and have since been\\n> removed altogether. The initial approach didn't get much traction but there\\n> was significant discussion and work, which has since fizzled out. The patch\\n> has been updated but there hasn't been meaningful review the past months, the\\n> last comments seem to imply there being a fair amount of questionmarks left in\\n> here. Robert, having been very involved in this do you have any thoughts on\\n> where we are and where to go (if at all IYO)?\\n\\nI'm not Robert, but I still have an opinion here, and that it's that this\\nfeature would at best be an attractive nuisance. If you need compression\\non a database session, it probably means that the connection is over the\\nopen internet, which means that you need encryption even more. And we\\nknow that compression and encryption do not play well together. The\\nreason compression was taken out of the latest TLS standards is not that\\nthey wouldn't have liked to have it, nor that applying compression in a\\nseparate code layer would be any safer. I fear offering this would\\nmerely lead people to build CVE-worthy setups.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 09 Nov 2021 10:02:30 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"Hi\\n\\n2839: possibility to rename root namespace in plpgsql\\n> =====================================================\\n> This patch addresses a feature request for making plpgsql easier to work\\n> with,\\n> but the approach has been questioned with the thread stalling since March\\n> in\\n> one of these discussion. Is there interest in picking this up or should we\\n> close it as Returned with Feedback?\\n>\\n\\nThis feature is from the category \\\"nice to have, but we can live without it\\nwell\\\". There are few requests per decade about this feature. I think my\\ndesign is good enough, but surely it is not the only one.\\n\\nWe cannot use inspiration from Oracle's PL/SQL or from Ada, because this\\nrequires a feature that is not supported in Postgres (pragmas). Without\\npragma is relatively hard to find good design. PL/pgSQL options are not\\nnative elements of language (I think so it is the best option), and other\\ndesigns are not fully correct (although they can look more natural). I\\ndon't see any passion (negative or positive) about this feature, and I\\nthink this feature can be returned without feedback. Maybe somebody will\\ncome with other proposals, maybe we will have pragmas one day.\\n\\nMuch more important is extending the debug support of PL/pgSQL. This patch\\ncan simplify implementations of PL/pgSQL debuggers, and this feature can be\\na real benefit for PL/pgSQL developers.\\n\\nRegards\\n\\nPavel\\n\\nHi\\n2839: possibility to rename root namespace in plpgsql\\n=====================================================\\nThis patch addresses a feature request for making plpgsql easier to work with,\\nbut the approach has been questioned with the thread stalling since March in\\none of these discussion. Is there interest in picking this up or should we\\nclose it as Returned with Feedback?This feature is from the category \\\"nice to have, but we can live without it well\\\". There are few requests per decade about this feature. I think my design is good enough, but surely it is not the only one.We cannot use inspiration from Oracle's PL/SQL or from Ada, because this requires a feature that is not supported in Postgres (pragmas). Without pragma is relatively hard to find good design. PL/pgSQL options are not native elements of language (I think so it is the best option), and other designs are not fully correct (although they can look more natural). I don't see any passion (negative or positive) about this feature, and I think this feature can be returned without feedback. Maybe somebody will come with other proposals, maybe we will have pragmas one day.Much more important is extending the debug support of PL/pgSQL. This patch can simplify implementations of PL/pgSQL debuggers, and this feature can be a real benefit for PL/pgSQL developers.RegardsPavel\",\n \"msg_date\": \"Tue, 9 Nov 2021 18:11:06 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\n* Tom Lane (tgl@sss.pgh.pa.us) wrote:\\n> Daniel Gustafsson <daniel@yesql.se> writes:\\n> > 2773: libpq compression\\n> > =======================\\n> > This patch intended to provide libpq connection compression to \\\"replace SSL\\n> > compression\\\" which was doomed when the patch was written, and have since been\\n> > removed altogether. The initial approach didn't get much traction but there\\n> > was significant discussion and work, which has since fizzled out. The patch\\n> > has been updated but there hasn't been meaningful review the past months, the\\n> > last comments seem to imply there being a fair amount of questionmarks left in\\n> > here. Robert, having been very involved in this do you have any thoughts on\\n> > where we are and where to go (if at all IYO)?\\n> \\n> I'm not Robert, but I still have an opinion here, and that it's that this\\n> feature would at best be an attractive nuisance. If you need compression\\n> on a database session, it probably means that the connection is over the\\n> open internet, which means that you need encryption even more. And we\\n> know that compression and encryption do not play well together. The\\n> reason compression was taken out of the latest TLS standards is not that\\n> they wouldn't have liked to have it, nor that applying compression in a\\n> separate code layer would be any safer. I fear offering this would\\n> merely lead people to build CVE-worthy setups.\\n\\nI've got an opinion on this also and I don't agree that needing\\ncompression means you're on the open internet or that we shouldn't allow\\nusers the choice to decide if they want to use compression and\\nencryption together or not. Yes, there's potential risks there, but I\\ndon't think those risks would lead to CVEs against PG for supporting a\\nmode where we allow compression and then also allow encryption- if\\nthat's a problem then it's an issue for the encryption library/method\\nbeing used and isn't the fault of us for allowing that combination.\\n\\nFurther, the general issue with encryption+compression attacks is when\\nthe attacker is able to inject chosen plaintext content into what ends\\nup being compressed and encrypted, and then is able to observe the\\ncompressed+encrypted traffic, and that's not nearly as easy to do when\\nyou're talking about a database connection vs. a case where an attacker\\nis able to convince a victim to go to a malicious website and through\\nthat is able to do cross-site requests to inject the known plaintext\\ninto requests to a trusted webpage and see the results coming back.\\n\\nThat we don't have any support for such a thing in the PG protocol and\\nthe general case that app server to DB traffic can certainly be a lot\\nand mean that compression is worthwhile even on a private network, and\\nthat users may wish to encrypt even on a private network, are all good\\nreasons why we should support both compression and encryption and allow\\nusers to choose what makes sense for them in their environment.\\n\\nThanks,\\n\\nStephen\",\n \"msg_date\": \"Tue, 9 Nov 2021 12:43:20 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"> Daniel Gustafsson <daniel@yesql.se> writes:\\n> > 2773: libpq compression\\n> > =======================\\n> > This patch intended to provide libpq connection compression to\\n> > \\\"replace SSL compression\\\" which was doomed when the patch was\\n> written,\\n> > and have since been removed altogether. The initial approach didn't\\n> > get much traction but there was significant discussion and work, which\\n> > has since fizzled out. The patch has been updated but there hasn't\\n> > been meaningful review the past months, the last comments seem to\\n> > imply there being a fair amount of questionmarks left in here.\\n> > Robert, having been very involved in this do you have any thoughts on\\n> where we are and where to go (if at all IYO)?\\n> \\n> I'm not Robert, but I still have an opinion here, and that it's that this feature\\n> would at best be an attractive nuisance. If you need compression on a\\n> database session, it probably means that the connection is over the open\\n> internet, which means that you need encryption even more. And we know\\n> that compression and encryption do not play well together. The reason\\n> compression was taken out of the latest TLS standards is not that they\\n> wouldn't have liked to have it, nor that applying compression in a separate\\n> code layer would be any safer. I fear offering this would merely lead people\\n> to build CVE-worthy setups.\\n> \\n\\nI don't have a strong opinion on the patch but I'd just like to mention that there are definitely use cases for using compression on private networks (with ssl off). We have cases where we have a PG extension that hooks into the COPY ... FROM command and accept lz4-compressed data for the COPY data message specifically. This has proven invaluable to us even on private networks, because it can give a good compression ratio for very little CPU decompression overhead.\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 19:28:32 +0000\",\n \"msg_from\": \"Floris Van Nee <florisvannee@Optiver.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"\\nOn 11/9/21 09:12, Daniel Gustafsson wrote:\\n>\\n> 2901: SQL/JSON: functions\\n> 2901: SQL/JSON: JSON_TABLE\\n> ==========================\\n> Grouping these two together as they are in very similar states. The patches\\n> seems to mostly move forward by rebases fixing smaller things. Andrew, do you\\n> have a sense of how far these are from a committable state?\\n>\\n>\\n\\nFor various reasons I'm going to be hors de combat for most of the rest\\nof the year, back at the beginning of January, although I will be around\\nif necessary, and probably tinkering with one or two things. I think at\\nleast the first set is not in bad shape, but it would benefit from some\\nfresh eyeballs. Tom Kincaid is trying to arrange some extra resources\\nfor that, so we might well see some progress soon.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 16:34:34 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 9, 2021 at 12:43:20PM -0500, Stephen Frost wrote:\\n> * Tom Lane (tgl@sss.pgh.pa.us) wrote:\\n> > Daniel Gustafsson <daniel@yesql.se> writes:\\n> > I'm not Robert, but I still have an opinion here, and that it's that this\\n> > feature would at best be an attractive nuisance. If you need compression\\n> > on a database session, it probably means that the connection is over the\\n> > open internet, which means that you need encryption even more. And we\\n> > know that compression and encryption do not play well together. The\\n> > reason compression was taken out of the latest TLS standards is not that\\n> > they wouldn't have liked to have it, nor that applying compression in a\\n> > separate code layer would be any safer. I fear offering this would\\n> > merely lead people to build CVE-worthy setups.\\n> \\n> I've got an opinion on this also and I don't agree that needing\\n> compression means you're on the open internet or that we shouldn't allow\\n> users the choice to decide if they want to use compression and\\n> encryption together or not. Yes, there's potential risks there, but I\\n> don't think those risks would lead to CVEs against PG for supporting a\\n> mode where we allow compression and then also allow encryption- if\\n> that's a problem then it's an issue for the encryption library/method\\n> being used and isn't the fault of us for allowing that combination.\\n\\nYeah, I am leaning this diretion too. On the one hand, the number of\\nways to exploit compression-then-encryption seem to continue to grow,\\nbut it is the complex behavior of HTTPS and connecting to multiple sites\\nwhich seems to always be required, and we don't do that with the\\nPostgres client/server protocol.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> https://momjian.us\\n EDB https://enterprisedb.com\\n\\n If only the physical world exists, free will is an illusion.\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 18:21:43 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 09, 2021 at 03:12:39PM +0100, Daniel Gustafsson wrote:\\n> 2800: CLUSTER on partitioned table\\n> ==================================\\n> The the patch above, it too spawned off from the REINDEX CONCURRENTLY thread\\n> but seems to have stalled awaiting more review. There was scepticism raised\\n> over the usefulness vs code weight tradeoff, but the response to this haven't\\n> really been followed up on. Alvaro and Michael, do you have any changed\\n> thoughts on Justins patch reorg or do you stil feel this isn't worth going\\n> ahead with?\\n\\nMy feeling on this one is that the addition was not worth the\\ncomplexity, and I got the impression that the handling of\\npg_index.indisclustered was rather blurry when it came to multiple\\nlayers of partitions.\\n\\n> 2861: pg_upgrade test for binary compatibility of core data types\\n> =================================================================\\n> This patch is also actively worked on and reviewed, and bits from this thread\\n> have already landed. Do you think this can land in its entirety in this CF\\n> Michael?\\n\\nThe remaining parts are the addition of a SQL test, which requires\\nmore comments and documentation as the thing is complex to follow, and\\nthe switch of the upgrade SQL queries to a file. I need to do a pass\\nover it, but that seems worth having on its own. For the second\\npoint, I wanted to double-check first with Andrew about how much code\\nwe could cut in the buildfarm client code.\\n\\nThanks for the summary!\\n--\\nMichael\",\n \"msg_date\": \"Wed, 10 Nov 2021 21:42:06 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"\\n\\n> Daniel Gustafsson <daniel@yesql.se> writes:\\n> \\n>> 2773: libpq compression\\n>> =======================\\n>> This patch intended to provide libpq connection compression to \\\"replace SSL\\n>> compression\\\" which was doomed when the patch was written, and have since been\\n>> removed altogether. The initial approach didn't get much traction but there\\n>> was significant discussion and work, which has since fizzled out. The patch\\n>> has been updated but there hasn't been meaningful review the past months, the\\n>> last comments seem to imply there being a fair amount of questionmarks left in\\n>> here. Robert, having been very involved in this do you have any thoughts on\\n>> where we are and where to go (if at all IYO)?\\n> \\n> I'm not Robert, but I still have an opinion here, and that it's that this\\n> feature would at best be an attractive nuisance. If you need compression\\n> on a database session, it probably means that the connection is over the\\n> open internet, which means that you need encryption even more. And we\\n> know that compression and encryption do not play well together. The\\n> reason compression was taken out of the latest TLS standards is not that\\n> they wouldn't have liked to have it, nor that applying compression in a\\n> separate code layer would be any safer. I fear offering this would\\n> merely lead people to build CVE-worthy setups.\\n> \\n\\nCompression is crucial for highly available setups. Replication traffic is often billed. Or route has bandwidth limits.\\nAn entropy added by WAL headers makes CRIME attack against replication encryption impractical.\\n\\nBest regards, Andrey Borodin.\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 20:54:21 +0500\",\n \"msg_from\": \"Andrey Borodin <x4mmm@yandex-team.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"> On 9 Nov 2021, at 18:43, Stephen Frost <sfrost@snowman.net> wrote:\\n> * Tom Lane (tgl@sss.pgh.pa.us) wrote:\\n>> Daniel Gustafsson <daniel@yesql.se> writes:\\n>>> 2773: libpq compression\\n>>> =======================\\n>>> This patch intended to provide libpq connection compression to \\\"replace SSL\\n>>> compression\\\" which was doomed when the patch was written, and have since been\\n>>> removed altogether. The initial approach didn't get much traction but there\\n>>> was significant discussion and work, which has since fizzled out. The patch\\n>>> has been updated but there hasn't been meaningful review the past months, the\\n>>> last comments seem to imply there being a fair amount of questionmarks left in\\n>>> here. Robert, having been very involved in this do you have any thoughts on\\n>>> where we are and where to go (if at all IYO)?\\n>> \\n>> I'm not Robert, but I still have an opinion here, and that it's that this\\n>> feature would at best be an attractive nuisance. If you need compression\\n>> on a database session, it probably means that the connection is over the\\n>> open internet, which means that you need encryption even more. And we\\n>> know that compression and encryption do not play well together. The\\n>> reason compression was taken out of the latest TLS standards is not that\\n>> they wouldn't have liked to have it, nor that applying compression in a\\n>> separate code layer would be any safer. I fear offering this would\\n>> merely lead people to build CVE-worthy setups.\\n> \\n> I've got an opinion on this also and I don't agree that needing\\n> compression means you're on the open internet or that we shouldn't allow\\n> users the choice to decide if they want to use compression and\\n> encryption together or not. Yes, there's potential risks there, but I\\n> don't think those risks would lead to CVEs against PG for supporting a\\n> mode where we allow compression and then also allow encryption- if\\n> that's a problem then it's an issue for the encryption library/method\\n> being used and isn't the fault of us for allowing that combination.\\n\\nI'm not so sure, if the parts aren't vulnerable separately on their own but the\\ncombination of them + libpq allows for weakening encryption in a TLS stream I'd\\nsay the CVE is on us.\\n\\n> Further, the general issue with encryption+compression attacks is when\\n> the attacker is able to inject chosen plaintext content into what ends\\n> up being compressed and encrypted, and then is able to observe the\\n> compressed+encrypted traffic, and that's not nearly as easy to do when\\n> you're talking about a database connection vs. a case where an attacker\\n> is able to convince a victim to go to a malicious website and through\\n> that is able to do cross-site requests to inject the known plaintext\\n> into requests to a trusted webpage and see the results coming back.\\n\\nI agree with this, the known weaknesses rely on mechanisms that aren't readily\\napplicable to postgres connections. I don't remember if the patch already does\\nthis, but if not we should ensure we only negotiate compression after\\nauthentication like how OpenSSH does [0].\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n[0] https://www.openssh.com/txt/draft-miller-secsh-compression-delayed-00.txt\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 17:04:08 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 9, 2021 at 7:02 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> Daniel Gustafsson <daniel@yesql.se> writes:\\n> > 2773: libpq compression\\n> > =======================\\n> > This patch intended to provide libpq connection compression to \\\"replace\\n> SSL\\n> > compression\\\" which was doomed when the patch was written, and have since\\n> been\\n> > removed altogether. The initial approach didn't get much traction but\\n> there\\n> > was significant discussion and work, which has since fizzled out. The\\n> patch\\n> > has been updated but there hasn't been meaningful review the past\\n> months, the\\n> > last comments seem to imply there being a fair amount of questionmarks\\n> left in\\n> > here. Robert, having been very involved in this do you have any\\n> thoughts on\\n> > where we are and where to go (if at all IYO)?\\n>\\n> I'm not Robert, but I still have an opinion here, and that it's that this\\n> feature would at best be an attractive nuisance. If you need compression\\n> on a database session, it probably means that the connection is over the\\n> open internet, which means that you need encryption even more.\\n>\\n\\nThis assumption is very vague. I personally had multiple cases when network\\nbandwidth for app <--> Postgres communication, that were fixed wither via\\nupgrades (spending more money) or making application developers cache more\\ndata (optimization), or, usually, both. Never public internet connections\\nwere involved.\\n\\nActually, any growing startup experiences such issues sooner or later.\\nNetwork compression would be a great tool for many setups, and some other\\npopular database systems offer it for a long.\\n\\nOn Tue, Nov 9, 2021 at 7:02 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:Daniel Gustafsson <daniel@yesql.se> writes:\\n> 2773: libpq compression\\n> =======================\\n> This patch intended to provide libpq connection compression to \\\"replace SSL\\n> compression\\\" which was doomed when the patch was written, and have since been\\n> removed altogether. The initial approach didn't get much traction but there\\n> was significant discussion and work, which has since fizzled out. The patch\\n> has been updated but there hasn't been meaningful review the past months, the\\n> last comments seem to imply there being a fair amount of questionmarks left in\\n> here. Robert, having been very involved in this do you have any thoughts on\\n> where we are and where to go (if at all IYO)?\\n\\nI'm not Robert, but I still have an opinion here, and that it's that this\\nfeature would at best be an attractive nuisance. If you need compression\\non a database session, it probably means that the connection is over the\\nopen internet, which means that you need encryption even more. This assumption is very vague. I personally had multiple cases when network bandwidth for app <--> Postgres communication, that were fixed wither via upgrades (spending more money) or making application developers cache more data (optimization), or, usually, both. Never public internet connections were involved.Actually, any growing startup experiences such issues sooner or later. Network compression would be a great tool for many setups, and some other popular database systems offer it for a long.\",\n \"msg_date\": \"Fri, 12 Nov 2021 13:14:18 -0800\",\n \"msg_from\": \"Nikolay Samokhvalov <samokhvalov@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 11/10/21 00:21, Bruce Momjian wrote:\\n> On Tue, Nov 9, 2021 at 12:43:20PM -0500, Stephen Frost wrote:\\n>> * Tom Lane (tgl@sss.pgh.pa.us) wrote:\\n>>> Daniel Gustafsson <daniel@yesql.se> writes:\\n>>> I'm not Robert, but I still have an opinion here, and that it's that this\\n>>> feature would at best be an attractive nuisance. If you need compression\\n>>> on a database session, it probably means that the connection is over the\\n>>> open internet, which means that you need encryption even more. And we\\n>>> know that compression and encryption do not play well together. The\\n>>> reason compression was taken out of the latest TLS standards is not that\\n>>> they wouldn't have liked to have it, nor that applying compression in a\\n>>> separate code layer would be any safer. I fear offering this would\\n>>> merely lead people to build CVE-worthy setups.\\n>>\\n>> I've got an opinion on this also and I don't agree that needing\\n>> compression means you're on the open internet or that we shouldn't allow\\n>> users the choice to decide if they want to use compression and\\n>> encryption together or not. Yes, there's potential risks there, but I\\n>> don't think those risks would lead to CVEs against PG for supporting a\\n>> mode where we allow compression and then also allow encryption- if\\n>> that's a problem then it's an issue for the encryption library/method\\n>> being used and isn't the fault of us for allowing that combination.\\n> \\n> Yeah, I am leaning this diretion too. On the one hand, the number of\\n> ways to exploit compression-then-encryption seem to continue to grow,\\n> but it is the complex behavior of HTTPS and connecting to multiple sites\\n> which seems to always be required, and we don't do that with the\\n> Postgres client/server protocol.\\n> \\n\\n+0.5\\n\\nI don't think the number of attacks grows all that much, actually. Or \\nmore precisely, the new vulnerabilities are just variations on the CRIME \\npattern adopting it to different contexts.\\n\\nFor example Voracle [1] applies to openvpn connections, but in principle \\nit's the same scenario - observing changes to the compressed size, after \\ntweaking the requests.\\n\\n[1] https://community.openvpn.net/openvpn/wiki/VORACLE\\n\\nSo it's true compression is a side-channel, but these attacks require \\nother ingredients too - ability to MITM the connection, and ability to \\ninject custom data.\\n\\nDoing MITM is probably harder than when attacking HTTPS, because DB \\nconnections tend to be between app server and db server - so setting up \\na rogue AP does not really work. Sometimes people / DBAs may connect to \\nthe DB directly, although in a way that makes it hard to do modify the \\ndata (which is why psql/ssh are immune to this). But it's possible.\\n\\nBut how hard is injecting custom data in a way that'd make CRIME-like \\nissue possible? Yes, you could hijack webapp similarly to CRIME, in a \\nway that initiates database requests with custom data. But we don't \\ngenerally send both the supplied and actual password to the DB to do the \\ncheck there - we just send one of them, and check it against the value \\nstored in the DB.\\n\\nThe one (somewhat) plausible scenario I can think of is an application doing\\n\\n UPDATE t SET field = user_data WHERE id = 1;\\n SELECT field FROM t WHERE ...;\\n\\nin which case we might learn if some other row in the result contains \\nthe same value. Not sure how practical this could be, though.\\n\\nAFAIK OpenVPN still supports tunnel compression, but it's disabled by \\ndefault. That probably says something about what they think about the \\nfeature, although that might be due to having compression and not \\nwanting to just remove it.\\n\\nOne interesting idea OpenVPN allows is asymmetric compression, i.e. \\nallowing compression in just one direction. For example it might be \\npossible for the client to send compressed data, but the server would \\nnever send compressed. Or vice versa. That might still help with e.g. \\nlarge imports or exports, depending on the use case, while maybe \\nmitigating some issues. But it's clearly not a solid defense.\\n\\n\\nregards\\n\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 01:09:52 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"On 11/10/21 16:54, Andrey Borodin wrote:\\n> \\n> \\n>> Daniel Gustafsson <daniel@yesql.se> writes:\\n>>\\n>>> 2773: libpq compression\\n>>> =======================\\n>>> This patch intended to provide libpq connection compression to \\\"replace SSL\\n>>> compression\\\" which was doomed when the patch was written, and have since been\\n>>> removed altogether. The initial approach didn't get much traction but there\\n>>> was significant discussion and work, which has since fizzled out. The patch\\n>>> has been updated but there hasn't been meaningful review the past months, the\\n>>> last comments seem to imply there being a fair amount of questionmarks left in\\n>>> here. Robert, having been very involved in this do you have any thoughts on\\n>>> where we are and where to go (if at all IYO)?\\n>>\\n>> I'm not Robert, but I still have an opinion here, and that it's that this\\n>> feature would at best be an attractive nuisance. If you need compression\\n>> on a database session, it probably means that the connection is over the\\n>> open internet, which means that you need encryption even more. And we\\n>> know that compression and encryption do not play well together. The\\n>> reason compression was taken out of the latest TLS standards is not that\\n>> they wouldn't have liked to have it, nor that applying compression in a\\n>> separate code layer would be any safer. I fear offering this would\\n>> merely lead people to build CVE-worthy setups.\\n>>\\n> \\n> Compression is crucial for highly available setups. Replication traffic is often billed. Or route has bandwidth limits.\\n> An entropy added by WAL headers makes CRIME attack against replication encryption impractical.\\n\\nI very much doubt WAL headers are a reliable protection against CRIME, \\nbecause the entropy of the headers is likely fairly constant. So if you \\ncompress the WAL stream, the WAL headers may change but the compression \\nratio should be pretty similar. At least that's my guess.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 01:14:31 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"\\n\\n> On 11/10/21 16:54, Andrey Borodin wrote:\\n> \\n>> Compression is crucial for highly available setups. Replication traffic is often billed. Or route has bandwidth limits.\\n>> An entropy added by WAL headers makes CRIME attack against replication encryption impractical.\\n> \\n> I very much doubt WAL headers are a reliable protection against CRIME,\\n> because the entropy of the headers is likely fairly constant. So if you\\n> compress the WAL stream, the WAL headers may change but the compression\\n> ratio should be pretty similar. At least that's my guess.\\n\\nI've thought more about it and I agree.\\nTo reliably protect against CRIME entropy of WAL headers must be comparable with the entropy of possibly injected data.\\nIf this would stand, probably, our WAL would need a really serious rework.\\n\\nMaybe just refuse to enable compression on SSL connection? If someone really needs both - they will just patch a server on their own.\\nOr make a GUC \\\"yes_i_kwow_what_crime_is_give_grant_read_on_my_data_to_spies\\\".\\n\\nBest regards, Andrey Borodin.\\n\\n\\n\",\n \"msg_date\": \"Sun, 14 Nov 2021 14:59:51 +0500\",\n \"msg_from\": \"Andrey Borodin <x4mmm@yandex-team.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\n* Andrey Borodin (x4mmm@yandex-team.ru) wrote:\\n> > On 11/10/21 16:54, Andrey Borodin wrote:\\n> >> Compression is crucial for highly available setups. Replication traffic is often billed. Or route has bandwidth limits.\\n> >> An entropy added by WAL headers makes CRIME attack against replication encryption impractical.\\n> > \\n> > I very much doubt WAL headers are a reliable protection against CRIME,\\n> > because the entropy of the headers is likely fairly constant. So if you\\n> > compress the WAL stream, the WAL headers may change but the compression\\n> > ratio should be pretty similar. At least that's my guess.\\n> \\n> I've thought more about it and I agree.\\n> To reliably protect against CRIME entropy of WAL headers must be comparable with the entropy of possibly injected data.\\n> If this would stand, probably, our WAL would need a really serious rework.\\n> \\n> Maybe just refuse to enable compression on SSL connection? If someone really needs both - they will just patch a server on their own.\\n> Or make a GUC \\\"yes_i_kwow_what_crime_is_give_grant_read_on_my_data_to_spies\\\".\\n\\nAttackers aren't likely to have the kind of isolated control over the\\ndata in the WAL stream (which is a combination of data from lots of\\nongoing activity in the system and isn't likely to be exactly what the\\nattacker supplied at some higher level anyway) and the ability to read\\nand analyze the WAL stream from a primary to a replica to be able to\\neffectively attack it.\\n\\nNone of this kind of babysitting makes any sense to me.\\n\\nI'm certainly fine with documenting that there are cases where\\ncompression of data sent by an attacker and then sent over an encrypted\\nchannel has been able to break the encryption and let admins decide if\\nit's appropriate or not for them to use in their environment.\\n\\nThanks,\\n\\nStephen\",\n \"msg_date\": \"Sun, 14 Nov 2021 14:13:12 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"Stephen Frost <sfrost@snowman.net> writes:\\n> Attackers aren't likely to have the kind of isolated control over the\\n> data in the WAL stream (which is a combination of data from lots of\\n> ongoing activity in the system and isn't likely to be exactly what the\\n> attacker supplied at some higher level anyway) and the ability to read\\n> and analyze the WAL stream from a primary to a replica to be able to\\n> effectively attack it.\\n\\nYeah, I concur with that so far as WAL data goes. A hypothetical attacker\\nwill not have control over xact IDs, tuple TIDs, etc, which will add\\nenough entropy to the stream that extracting data payloads seems pretty\\ninfeasible.\\n\\nMy concern upthread was about client-session connections, where such\\nmitigation doesn't apply. (I wonder a bit about logical-replication\\nstreams, too.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 14 Nov 2021 14:26:33 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 9, 2021 at 9:12 AM Daniel Gustafsson <daniel@yesql.se> wrote:\\n> 2773: libpq compression\\n> =======================\\n> This patch intended to provide libpq connection compression to \\\"replace SSL\\n> compression\\\" which was doomed when the patch was written, and have since been\\n> removed altogether. The initial approach didn't get much traction but there\\n> was significant discussion and work, which has since fizzled out. The patch\\n> has been updated but there hasn't been meaningful review the past months, the\\n> last comments seem to imply there being a fair amount of questionmarks left in\\n> here. Robert, having been very involved in this do you have any thoughts on\\n> where we are and where to go (if at all IYO)?\\n\\nIt's been a while since I looked at that patch, and my recollections\\nare somewhat fuzzy. However, I think that my main feeling was that the\\nquality of the patch didn't seem to be that close to what I felt would\\nbe required for a commit. I think at that point there were still\\nsignificant questions about the details of the approach as well as\\ncode-quality and performance concerns. It may have gotten better since\\nlast I looked.\\n\\nI do agree with the comments from others that there is room to debate\\n(1) the usefulness of this and (2) the degree to which it might leave\\nus on the hook for security problems that wouldn't be our\\nresponsibility if we left this up to OpenSSL (or maybe some other SSL\\nlibrary, if we ever get pluggability there, which I hope we do). But\\nmy concerns were more pragmatic than theoretical.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 09:32:29 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"> On 15 Nov 2021, at 15:32, Robert Haas <robertmhaas@gmail.com> wrote:\\n> \\n> On Tue, Nov 9, 2021 at 9:12 AM Daniel Gustafsson <daniel@yesql.se> wrote:\\n>> 2773: libpq compression\\n>> =======================\\n>> This patch intended to provide libpq connection compression to \\\"replace SSL\\n>> compression\\\" which was doomed when the patch was written, and have since been\\n>> removed altogether. The initial approach didn't get much traction but there\\n>> was significant discussion and work, which has since fizzled out. The patch\\n>> has been updated but there hasn't been meaningful review the past months, the\\n>> last comments seem to imply there being a fair amount of questionmarks left in\\n>> here. Robert, having been very involved in this do you have any thoughts on\\n>> where we are and where to go (if at all IYO)?\\n> \\n> It's been a while since I looked at that patch, and my recollections\\n> are somewhat fuzzy.\\n\\nNo worries, thanks for sharing!\\n\\n> However, I think that my main feeling was that the\\n> quality of the patch didn't seem to be that close to what I felt would\\n> be required for a commit. I think at that point there were still\\n> significant questions about the details of the approach as well as\\n> code-quality and performance concerns. It may have gotten better since\\n> last I looked.\\n\\nContext switching into this thread is hard, so it's not clear how much the\\npatch has moved since you reviewed, or how the current patchset relates to the\\nearlier one as there seems to have been a shift at a recent(ish) point in time.\\nThe last message (which is unreadable HTML soup in the archives) talks about\\nimplementing lz4 with \\\"Looking forward to post an update soon.\\\"\\n\\nGiven the discussions had and the doubts over approach I wonder if the best\\ncourse of action with the CF entry in question is to close it, hash out the\\nwhat and how on -hackers and reopen a new entry when there is a patch which.\\nWe are not doing ourselves, or the CF process, any favors by moving an entry\\nalong which isn't all that close to committable.\\n\\nThoughts (addressing everyone involved/interested)?\\n\\n> I do agree with the comments from others that there is room to debate\\n> (1) the usefulness of this and (2) the degree to which it might leave\\n> us on the hook for security problems that wouldn't be our\\n> responsibility if we left this up to OpenSSL (or maybe some other SSL\\n> library, if we ever get pluggability there, which I hope we do). But\\n> my concerns were more pragmatic than theoretical.\\n\\nI share you hope here =) However I don't think we'll get much help with\\ncompression from the TLS libraries. OpenSSL has deprecated transport\\ncompression, NSS never implemented it in the first place, and TLS 1.3 removes\\nit altogether. That being said, this is mainly due to protocols where chosen\\nplaintext attacks are practical/possible. For our case we, IMHO, should look\\nmore at how OpenSSH and other protocols closer to ours has built compression in\\na safe way, to see if we can achieve the level of transport safety we require.\\n\\nBut, there is I think broad concensus that we at least must fully understand\\nthe tradeoffs should we allow compression.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 20:05:05 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\n* Daniel Gustafsson (daniel@yesql.se) wrote:\\n> > On 15 Nov 2021, at 15:32, Robert Haas <robertmhaas@gmail.com> wrote:\\n> > On Tue, Nov 9, 2021 at 9:12 AM Daniel Gustafsson <daniel@yesql.se> wrote:\\n> >> 2773: libpq compression\\n> >> =======================\\n> >> This patch intended to provide libpq connection compression to \\\"replace SSL\\n> >> compression\\\" which was doomed when the patch was written, and have since been\\n> >> removed altogether. The initial approach didn't get much traction but there\\n> >> was significant discussion and work, which has since fizzled out. The patch\\n> >> has been updated but there hasn't been meaningful review the past months, the\\n> >> last comments seem to imply there being a fair amount of questionmarks left in\\n> >> here. Robert, having been very involved in this do you have any thoughts on\\n> >> where we are and where to go (if at all IYO)?\\n> > \\n> > It's been a while since I looked at that patch, and my recollections\\n> > are somewhat fuzzy.\\n> \\n> No worries, thanks for sharing!\\n> \\n> > However, I think that my main feeling was that the\\n> > quality of the patch didn't seem to be that close to what I felt would\\n> > be required for a commit. I think at that point there were still\\n> > significant questions about the details of the approach as well as\\n> > code-quality and performance concerns. It may have gotten better since\\n> > last I looked.\\n> \\n> Context switching into this thread is hard, so it's not clear how much the\\n> patch has moved since you reviewed, or how the current patchset relates to the\\n> earlier one as there seems to have been a shift at a recent(ish) point in time.\\n> The last message (which is unreadable HTML soup in the archives) talks about\\n> implementing lz4 with \\\"Looking forward to post an update soon.\\\"\\n\\nNot sure where the best place to drop this comment is, but the patch\\nseems to be doing exactly what I was hoping it didn't- it's just taking\\nthe ultimate results and compressing them and sending them on to the\\nclient, at least that's what it looks like in a quick look.\\n\\nWhat about when that data has already been compressed and stored\\ncompressed and not manipulated at all? The initial post in that thread\\ntalks about returning data from 'large JSON columns'. With this\\napproach, we would likely end up taking compressed json/jsonb data,\\ndecompressing it, and then re-compressing it to send to the client.\\n\\nThat seems ... less than ideal.\\n\\nI get that just compressing the entire stream is simpler and easier and\\nsuch, but it's surely cheaper and more efficient to not decompress and\\nthen recompress data that's already compressed. Finding a way to pass\\nthrough data that's already compressed when stored as-is while also\\nsupporting compression of everything else (in a sensible way- wouldn't\\nmake sense to just compress each attribute independently since a 4 byte\\ninteger isn't going to get smaller with compression) definitely\\ncomplicates the overall idea but perhaps would be possible to do.\\n\\n> Given the discussions had and the doubts over approach I wonder if the best\\n> course of action with the CF entry in question is to close it, hash out the\\n> what and how on -hackers and reopen a new entry when there is a patch which.\\n> We are not doing ourselves, or the CF process, any favors by moving an entry\\n> along which isn't all that close to committable.\\n> \\n> Thoughts (addressing everyone involved/interested)?\\n\\nYeah, agreed with this in general.\\n\\nThanks,\\n\\nStephen\",\n \"msg_date\": \"Mon, 15 Nov 2021 14:51:46 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 2:51 PM Stephen Frost <sfrost@snowman.net> wrote:\\n> I get that just compressing the entire stream is simpler and easier and\\n> such, but it's surely cheaper and more efficient to not decompress and\\n> then recompress data that's already compressed. Finding a way to pass\\n> through data that's already compressed when stored as-is while also\\n> supporting compression of everything else (in a sensible way- wouldn't\\n> make sense to just compress each attribute independently since a 4 byte\\n> integer isn't going to get smaller with compression) definitely\\n> complicates the overall idea but perhaps would be possible to do.\\n\\nTo me, this feels like an attempt to move the goalposts far enough to\\nkill the project. Sure, in a perfect world, that would be nice. But,\\nwe don't do it anywhere else. If you try to store a JPEG into a bytea\\ncolumn, we'll try to compress it just like we would any other data,\\nand it may not work out. If you then take a pg_basebackup of the\\ndatabase using -Z, there's no attempt made to avoid the overhead of\\nCPU overhead of compressing those TOAST table pages that contain\\nalready-compressed data and not the others. And it's easy to\\nunderstand why that's the case: when you insert data into the\\ndatabase, there's no way for the database to magically know whether\\nthat data has been previously compressed by some means, and if so, how\\neffectively. And when you back up a database, the backup doesn't know\\nwhich relfilenodes contain TOAST tables or which pages of those\\nrelfilenodes contain that is already pre-compressed. In both cases,\\nyour options are either (1) shut off compression yourself or (2) hope\\nthat the compressor doesn't waste too much effort on it.\\n\\nI think the same approach ought to be completely acceptable here. I\\ndon't even really understand how we could do anything else. printtup()\\njust gets datums, and it has no idea whether or how they are toasted.\\nIt calls the type output functions which don't know that data is being\\nprepared for transmission to the client as opposed to some other\\nhypothetical way you could call that function, nor do they know what\\ncompression method the client wants. It does not seem at all\\nstraightforward to teach them that ... and even if they did, what\\nthen? It's not like every column value is sent as a separate packet;\\nthe whole row is a single protocol message, and some columns may be\\ncompressed and others uncompressed. Trying to guess what to do about\\nthat seems to boil down to a sheer guess. Unless you try to compress\\nthat mixture of compressed and uncompressed values - and it's\\nmoderately uncommon for every column of a table to be even be\\ntoastable - you aren't going to know how well it will compress. You\\ncould easily waste more CPU cycles trying to guess than you would have\\nspent just doing what the user asked for.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 15:23:17 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 03:23:17PM -0500, Robert Haas wrote:\\n> On Mon, Nov 15, 2021 at 2:51 PM Stephen Frost <sfrost@snowman.net> wrote:\\n> > I get that just compressing the entire stream is simpler and easier and\\n> > such, but it's surely cheaper and more efficient to not decompress and\\n> > then recompress data that's already compressed. Finding a way to pass\\n> > through data that's already compressed when stored as-is while also\\n> > supporting compression of everything else (in a sensible way- wouldn't\\n> > make sense to just compress each attribute independently since a 4 byte\\n> > integer isn't going to get smaller with compression) definitely\\n> > complicates the overall idea but perhaps would be possible to do.\\n> \\n> To me, this feels like an attempt to move the goalposts far enough to\\n> kill the project.\\n\\nI don't know what version of the patch Stephen looked at, but the patches I\\nlooked at were only compressing things like CopyData, which seems like it would\\nallow using compression when it'd be effective, without adding overhead to\\ncases where it won't.\\n\\nIMO, the patch set should be presented differently, to use the existing zlib\\nand lz4 in 0001, and then add zstd in 0002 or later.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 14:55:00 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\n* Robert Haas (robertmhaas@gmail.com) wrote:\\n> On Mon, Nov 15, 2021 at 2:51 PM Stephen Frost <sfrost@snowman.net> wrote:\\n> > I get that just compressing the entire stream is simpler and easier and\\n> > such, but it's surely cheaper and more efficient to not decompress and\\n> > then recompress data that's already compressed. Finding a way to pass\\n> > through data that's already compressed when stored as-is while also\\n> > supporting compression of everything else (in a sensible way- wouldn't\\n> > make sense to just compress each attribute independently since a 4 byte\\n> > integer isn't going to get smaller with compression) definitely\\n> > complicates the overall idea but perhaps would be possible to do.\\n> \\n> To me, this feels like an attempt to move the goalposts far enough to\\n> kill the project. Sure, in a perfect world, that would be nice. But,\\n> we don't do it anywhere else. If you try to store a JPEG into a bytea\\n> column, we'll try to compress it just like we would any other data,\\n> and it may not work out. If you then take a pg_basebackup of the\\n> database using -Z, there's no attempt made to avoid the overhead of\\n> CPU overhead of compressing those TOAST table pages that contain\\n> already-compressed data and not the others. And it's easy to\\n> understand why that's the case: when you insert data into the\\n> database, there's no way for the database to magically know whether\\n> that data has been previously compressed by some means, and if so, how\\n> effectively. And when you back up a database, the backup doesn't know\\n> which relfilenodes contain TOAST tables or which pages of those\\n> relfilenodes contain that is already pre-compressed. In both cases,\\n> your options are either (1) shut off compression yourself or (2) hope\\n> that the compressor doesn't waste too much effort on it.\\n\\nI'll grant that perhaps it's just a different project.\\n\\nWhile, sure, we will try to compress things we don't understand by\\ndefault, the admin does have ways to tell us to not do that and so it\\nisn't always just guess-work. As for other parts of the system not\\nbeing smarter- they should be. I'd argue that up until quite recently\\nit didn't make sense to teach something like pg_basebackup about TOAST\\ntables because our compression was rather lacking and it was actually\\nuseful to compress TOAST tables (perhaps still is with the compression\\nmethods we have now, but we'll hopefully add more). We should probably\\nlook at trying to provide a way for pg_basebackup and other backup tools\\nto identify TOAST tables to avoid trying to compress them as it's just\\nwasted effort (though in that case, it's at least less wasted than here-\\nwe don't decompress that TOAST data and then re-compress it).\\n\\nUsing bytea for everything-and-the-kitchen-sink also just generally\\nseems like it's throwing away more information than we really should (if\\nwe had a jpeg data type, for example, we could just have that default to\\nnot being compressed and not spend time trying to compress jpegs).\\n\\nI also just generally disagree that an inefficiency in one part of the\\nsystem justifies having it somewhere else.\\n\\n> I think the same approach ought to be completely acceptable here. I\\n> don't even really understand how we could do anything else. printtup()\\n> just gets datums, and it has no idea whether or how they are toasted.\\n> It calls the type output functions which don't know that data is being\\n> prepared for transmission to the client as opposed to some other\\n> hypothetical way you could call that function, nor do they know what\\n> compression method the client wants. It does not seem at all\\n> straightforward to teach them that ... and even if they did, what\\n> then? It's not like every column value is sent as a separate packet;\\n> the whole row is a single protocol message, and some columns may be\\n> compressed and others uncompressed. Trying to guess what to do about\\n> that seems to boil down to a sheer guess. Unless you try to compress\\n> that mixture of compressed and uncompressed values - and it's\\n> moderately uncommon for every column of a table to be even be\\n> toastable - you aren't going to know how well it will compress. You\\n> could easily waste more CPU cycles trying to guess than you would have\\n> spent just doing what the user asked for.\\n\\nI agree that we wouldn't want to be doing this all based on guesswork\\nsomehow. Having a way to identify what's compressed and what isn't\\nwould be needed, as well as what method was used to compress. It's not\\nlike we don't know that at the source, we'd just need a way to pass that\\nback through- probably more-or-less exactly as how it's stored today.\\n\\nTackling the actual protocol side of it seems like it'd be the more\\nsensible place to start working through a design and I don't think\\nthat'd be trivial to do, so probably not something to try and hash out\\nin this particular thread. I'm happy to have at least voiced that\\nthought.\\n\\nThanks,\\n\\nStephen\",\n \"msg_date\": \"Mon, 15 Nov 2021 15:58:57 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"Hi! It’s been a while since the original patch release. Let me provide a brief \\noverview of the current patch status.\\n\\nThe initial approach was to use the streaming compression to compress all \\noutgoing and decompress all incoming bytes. However, after the long discussion \\nin the thread, the initial approach has been changed.\\n\\nThe current implementation allows compressing only specific message types, \\nuse the different compression algorithms for different message types, configure \\nthe allowed compression methods and levels both for server- and client- sides \\nvia GUC setting / connection string respectively.\\n\\nAlso, current implementation combines (when possible) multiple protocol messages \\ninto the single CompressedData message for a better compression ratio.\\n\\n> \\n> On 16 Nov 2021, at 01:23, Robert Haas <robertmhaas@gmail.com> wrote:\\n> \\n> To me, this feels like an attempt to move the goalposts far enough to\\n> kill the project. Sure, in a perfect world, that would be nice. But,\\n> we don't do it anywhere else. If you try to store a JPEG into a bytea\\n> column, we'll try to compress it just like we would any other data,\\n> and it may not work out. If you then take a pg_basebackup of the\\n> database using -Z, there's no attempt made to avoid the overhead of\\n> CPU overhead of compressing those TOAST table pages that contain\\n> already-compressed data and not the others. And it's easy to\\n> understand why that's the case: when you insert data into the\\n> database, there's no way for the database to magically know whether\\n> that data has been previously compressed by some means, and if so, how\\n> effectively. And when you back up a database, the backup doesn't know\\n> which relfilenodes contain TOAST tables or which pages of those\\n> relfilenodes contain that is already pre-compressed. In both cases,\\n> your options are either (1) shut off compression yourself or (2) hope\\n> that the compressor doesn't waste too much effort on it.\\n> \\n> I think the same approach ought to be completely acceptable here. I\\n> don't even really understand how we could do anything else. printtup()\\n> just gets datums, and it has no idea whether or how they are toasted.\\n> It calls the type output functions which don't know that data is being\\n> prepared for transmission to the client as opposed to some other\\n> hypothetical way you could call that function, nor do they know what\\n> compression method the client wants. It does not seem at all\\n> straightforward to teach them that ... and even if they did, what\\n> then? It's not like every column value is sent as a separate packet;\\n> the whole row is a single protocol message, and some columns may be\\n> compressed and others uncompressed. Trying to guess what to do about\\n> that seems to boil down to a sheer guess. Unless you try to compress\\n> that mixture of compressed and uncompressed values - and it's\\n> moderately uncommon for every column of a table to be even be\\n> toastable - you aren't going to know how well it will compress. You\\n> could easily waste more CPU cycles trying to guess than you would have\\n> spent just doing what the user asked for.\\n> \\n\\nAgree. From my POV, it is OK to use the protocol message type and length to decide \\nshould it be compressed or not. Also, this can be optimized later without the need to change \\nthe protocol.\\n\\nRegarding the LZ4 support patch, it still has some minor polishing to do. Basically, it only adds the LZ4 \\nalgorithm support and does not change anything fundamentally. So I would appreciate someone doing \\na review of the current patch version.\\n\\nThe original thread is quite huge so I guess that it makes it hard to catch up with the current patch status.\\nI can make a new one with a detailed summary if that would help.\\n\\nThanks,\\n\\nDaniil Zakhlystov\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 22:40:47 +0500\",\n \"msg_from\": \"Daniil Zakhlystov <usernamedt@yandex-team.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n },\n {\n \"msg_contents\": \"> \\n> 2780: Allow COPY \\\"text\\\" to output a header and add header matching mode to COPY\\n> FROM\\n> ===============================================================================\\n> The original patch was rejected for lacking matching, but it has since been\\n> addressed in an updated patch which has seen a lot of review. Reading the\\n> thread it seems this should really be in Ready for Committer.\\n\\nI don’t think there is anything more that needs to be done for this patch. I have been checking that it apply cleanly and that the tests still pass but it has already multiple rounds of reviews and is probably ready.\\n\\nBest regards,\\nRémi\\n\\n\",\n \"msg_date\": \"Mon, 29 Nov 2021 21:26:48 +0100\",\n \"msg_from\": \"=?utf-8?Q?R=C3=A9mi_Lapeyre?= <remi.lapeyre@lenstra.fr>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 2\"\n }\n]","isConversational":false}}},{"rowIdx":53060,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nWhile trying to add some new tests, I found that\nPostgreSQL::Test::Utils is not required. I felt\nPostgreSQL::Test::Utils can be removed from a lot of tap tests which\ndo not require it. I removed it, ran the tests and found the tests to\nbe executing fine. I have made a patch including the changes for it.\nIf this import can be removed, kindly accept the attached patch for\nthe same.\n\nRegards,\nVignesh","msg_date":"Tue, 9 Nov 2021 21:48:30 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Removed unused import modules from tap tests"},{"msg_contents":"On Tue, Nov 09, 2021 at 09:48:30PM +0530, vignesh C wrote:\n> While trying to add some new tests, I found that\n> PostgreSQL::Test::Utils is not required. I felt\n> PostgreSQL::Test::Utils can be removed from a lot of tap tests which\n> do not require it. I removed it, ran the tests and found the tests to\n> be executing fine. I have made a patch including the changes for it.\n> If this import can be removed, kindly accept the attached patch for\n> the same.\n\nI would not have bothered changing things if the names of the modules\nwere the same across stable branches to minimize merge conflicts.\n\nHowever, everything has changed on HEAD, so there is a good argument\nfor simplifying the tests as you are proposing here. Any thoughts\nfrom others?\n--\nMichael","msg_date":"Wed, 10 Nov 2021 12:34:53 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Removed unused import modules from tap tests"},{"msg_contents":"On 2021-Nov-10, Michael Paquier wrote:\n\n> I would not have bothered changing things if the names of the modules\n> were the same across stable branches to minimize merge conflicts.\n> \n> However, everything has changed on HEAD, so there is a good argument\n> for simplifying the tests as you are proposing here. Any thoughts\n> from others?\n\nI agree with your reasoning, but I wonder what's the *benefit* of\nremoving those includes. IOW, what's the reason not to simply drop the\npatch?\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Wed, 10 Nov 2021 09:37:23 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Removed unused import modules from tap tests"},{"msg_contents":"> On 10 Nov 2021, at 13:37, Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> \n> On 2021-Nov-10, Michael Paquier wrote:\n> \n>> I would not have bothered changing things if the names of the modules\n>> were the same across stable branches to minimize merge conflicts.\n>> \n>> However, everything has changed on HEAD, so there is a good argument\n>> for simplifying the tests as you are proposing here. Any thoughts\n>> from others?\n> \n> I agree with your reasoning..\n\nAgreed, if we are ever to do it now would be the time.\n\n> ..but I wonder what's the *benefit* of removing those includes. IOW, what's\n> the reason not to simply drop the patch?\n\n\nI think the value is mostly neatnikism, the actual effect on runtime is\nunlikely to be measureable. I won't argue against doing it, but I suspect\nwe'll just slowly add a lot of these back as tests evolve making excercise\nless useful.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Wed, 10 Nov 2021 13:59:31 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Removed unused import modules from tap tests"},{"msg_contents":"On Wed, Nov 10, 2021 at 6:07 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>\n> On 2021-Nov-10, Michael Paquier wrote:\n>\n> > I would not have bothered changing things if the names of the modules\n> > were the same across stable branches to minimize merge conflicts.\n> >\n> > However, everything has changed on HEAD, so there is a good argument\n> > for simplifying the tests as you are proposing here. Any thoughts\n> > from others?\n>\n> I agree with your reasoning, but I wonder what's the *benefit* of\n> removing those includes. IOW, what's the reason not to simply drop the\n> patch?\n\nThe idea was to clean up the unused import. I noticed that generally\nwhile adding new test files we copy from existing files, this results\nin existing unused imports also being added when new tests are added.\nIt is just a cleanup activity to remove it from the existing code and\nprobably can be taken care during the review so that it does not get\nadded in the new tests.\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 10 Nov 2021 20:01:40 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Removed unused import modules from tap tests"},{"msg_contents":"Daniel Gustafsson <daniel@yesql.se> writes:\n>> On 10 Nov 2021, at 13:37, Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>> ..but I wonder what's the *benefit* of removing those includes. IOW, what's\n>> the reason not to simply drop the patch?\n\n> I think the value is mostly neatnikism, the actual effect on runtime is\n> unlikely to be measureable. I won't argue against doing it, but I suspect\n> we'll just slowly add a lot of these back as tests evolve making excercise\n> less useful.\n\nYeah, that last was pretty much my reaction. I don't know enough about\nPerl to be sure how much an unused import costs, but I suspect you're\nright that it won't be measurable in context, considering that most of\nthese test scripts run at least one initdb.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 10 Nov 2021 09:53:27 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Removed unused import modules from tap tests"},{"msg_contents":"Tom Lane <tgl@sss.pgh.pa.us> writes:\n\n> Daniel Gustafsson <daniel@yesql.se> writes:\n>>> On 10 Nov 2021, at 13:37, Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>>> ..but I wonder what's the *benefit* of removing those includes. IOW, what's\n>>> the reason not to simply drop the patch?\n>\n>> I think the value is mostly neatnikism, the actual effect on runtime is\n>> unlikely to be measureable. I won't argue against doing it, but I suspect\n>> we'll just slowly add a lot of these back as tests evolve making excercise\n>> less useful.\n>\n> Yeah, that last was pretty much my reaction. I don't know enough about\n> Perl to be sure how much an unused import costs, but I suspect you're\n> right that it won't be measurable in context, considering that most of\n> these test scripts run at least one initdb.\n\nOn my laptop (using the \"performance\" CPU frequency governor to avoid\nscaling noise) loading PostgreSQL::Test::Utils takes about 80ms:\n\n$ time perl -I./src/test/perl -e 'use PostgreSQL::Test::Utils;'\n\nreal 0m0.079s\nuser 0m0.078s\nsys 0m0.001s\n\nHowever, if we're also loading PostgreSQL::Test::Cluster, the difference\nis negligible, because most of the time is spent loading modules used by\nboth:\n\n$ time perl -I./src/test/perl -e 'use PostgreSQL::Test::Cluster; use PostgreSQL::Test::Utils;'\n\nreal 0m0.092s\nuser 0m0.088s\nsys 0m0.004s\n\n$ time perl -I./src/test/perl -e 'use PostgreSQL::Test::Cluster;'\n\nreal 0m0.090s\nuser 0m0.081s\nsys 0m0.008s\n\n- ilmari\n\n\n","msg_date":"Wed, 10 Nov 2021 15:42:00 +0000","msg_from":"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org>","msg_from_op":false,"msg_subject":"Re: Removed unused import modules from tap tests"},{"msg_contents":"On Wed, Nov 10, 2021 at 9:53 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Yeah, that last was pretty much my reaction. I don't know enough about\n> Perl to be sure how much an unused import costs, but I suspect you're\n> right that it won't be measurable in context, considering that most of\n> these test scripts run at least one initdb.\n\nI think focusing on the runtime might be missing the point to some\nextent. Removing used #include directives from C source files is\ngenerally good practice just for cleanliness, and this is the same\nkind of thing. That said, the argument that future changes are likely\nto just put the same 'use' directive back seems to me to have some\nmerit. But we've probably spent more energy debating this than the\ntopic deserves. It wouldn't hurt to remove these, and it also won't\nhurt if we don't.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 10 Nov 2021 11:07:09 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Removed unused import modules from tap tests"},{"msg_contents":"\nOn 11/10/21 09:53, Tom Lane wrote:\n> Daniel Gustafsson <daniel@yesql.se> writes:\n>>> On 10 Nov 2021, at 13:37, Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>>> ..but I wonder what's the *benefit* of removing those includes. IOW, what's\n>>> the reason not to simply drop the patch?\n>> I think the value is mostly neatnikism, the actual effect on runtime is\n>> unlikely to be measureable. I won't argue against doing it, but I suspect\n>> we'll just slowly add a lot of these back as tests evolve making excercise\n>> less useful.\n> Yeah, that last was pretty much my reaction. I don't know enough about\n> Perl to be sure how much an unused import costs, but I suspect you're\n> right that it won't be measurable in context, considering that most of\n> these test scripts run at least one initdb.\n>\n> \t\t\t\n\n\n:Cluster uses :Utils, and perl is smart enough not to try to reprocess\nthe module. Thus the extra cost here is almost certainly very close to zero.\n\n\nThis is a perfectly reasonable piece of boilerplate to use at the top of\na TAP test:\n\n\n use strict;\n\n use warnings;\n\n use PostgreSQL:Test::Cluster;\n\n use PostgreSQL::Test::Utils;\n\n use Test::More;\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 10 Nov 2021 13:17:46 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Removed unused import modules from tap tests"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nWhile trying to add some new tests, I found that\\nPostgreSQL::Test::Utils is not required. I felt\\nPostgreSQL::Test::Utils can be removed from a lot of tap tests which\\ndo not require it. I removed it, ran the tests and found the tests to\\nbe executing fine. I have made a patch including the changes for it.\\nIf this import can be removed, kindly accept the attached patch for\\nthe same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Tue, 9 Nov 2021 21:48:30 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Removed unused import modules from tap tests\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 09, 2021 at 09:48:30PM +0530, vignesh C wrote:\\n> While trying to add some new tests, I found that\\n> PostgreSQL::Test::Utils is not required. I felt\\n> PostgreSQL::Test::Utils can be removed from a lot of tap tests which\\n> do not require it. I removed it, ran the tests and found the tests to\\n> be executing fine. I have made a patch including the changes for it.\\n> If this import can be removed, kindly accept the attached patch for\\n> the same.\\n\\nI would not have bothered changing things if the names of the modules\\nwere the same across stable branches to minimize merge conflicts.\\n\\nHowever, everything has changed on HEAD, so there is a good argument\\nfor simplifying the tests as you are proposing here. Any thoughts\\nfrom others?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 10 Nov 2021 12:34:53 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Removed unused import modules from tap tests\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-10, Michael Paquier wrote:\\n\\n> I would not have bothered changing things if the names of the modules\\n> were the same across stable branches to minimize merge conflicts.\\n> \\n> However, everything has changed on HEAD, so there is a good argument\\n> for simplifying the tests as you are proposing here. Any thoughts\\n> from others?\\n\\nI agree with your reasoning, but I wonder what's the *benefit* of\\nremoving those includes. IOW, what's the reason not to simply drop the\\npatch?\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 09:37:23 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Removed unused import modules from tap tests\"\n },\n {\n \"msg_contents\": \"> On 10 Nov 2021, at 13:37, Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> \\n> On 2021-Nov-10, Michael Paquier wrote:\\n> \\n>> I would not have bothered changing things if the names of the modules\\n>> were the same across stable branches to minimize merge conflicts.\\n>> \\n>> However, everything has changed on HEAD, so there is a good argument\\n>> for simplifying the tests as you are proposing here. Any thoughts\\n>> from others?\\n> \\n> I agree with your reasoning..\\n\\nAgreed, if we are ever to do it now would be the time.\\n\\n> ..but I wonder what's the *benefit* of removing those includes. IOW, what's\\n> the reason not to simply drop the patch?\\n\\n\\nI think the value is mostly neatnikism, the actual effect on runtime is\\nunlikely to be measureable. I won't argue against doing it, but I suspect\\nwe'll just slowly add a lot of these back as tests evolve making excercise\\nless useful.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 13:59:31 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Removed unused import modules from tap tests\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 10, 2021 at 6:07 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>\\n> On 2021-Nov-10, Michael Paquier wrote:\\n>\\n> > I would not have bothered changing things if the names of the modules\\n> > were the same across stable branches to minimize merge conflicts.\\n> >\\n> > However, everything has changed on HEAD, so there is a good argument\\n> > for simplifying the tests as you are proposing here. Any thoughts\\n> > from others?\\n>\\n> I agree with your reasoning, but I wonder what's the *benefit* of\\n> removing those includes. IOW, what's the reason not to simply drop the\\n> patch?\\n\\nThe idea was to clean up the unused import. I noticed that generally\\nwhile adding new test files we copy from existing files, this results\\nin existing unused imports also being added when new tests are added.\\nIt is just a cleanup activity to remove it from the existing code and\\nprobably can be taken care during the review so that it does not get\\nadded in the new tests.\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 20:01:40 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Removed unused import modules from tap tests\"\n },\n {\n \"msg_contents\": \"Daniel Gustafsson <daniel@yesql.se> writes:\\n>> On 10 Nov 2021, at 13:37, Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>> ..but I wonder what's the *benefit* of removing those includes. IOW, what's\\n>> the reason not to simply drop the patch?\\n\\n> I think the value is mostly neatnikism, the actual effect on runtime is\\n> unlikely to be measureable. I won't argue against doing it, but I suspect\\n> we'll just slowly add a lot of these back as tests evolve making excercise\\n> less useful.\\n\\nYeah, that last was pretty much my reaction. I don't know enough about\\nPerl to be sure how much an unused import costs, but I suspect you're\\nright that it won't be measurable in context, considering that most of\\nthese test scripts run at least one initdb.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 09:53:27 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Removed unused import modules from tap tests\"\n },\n {\n \"msg_contents\": \"Tom Lane <tgl@sss.pgh.pa.us> writes:\\n\\n> Daniel Gustafsson <daniel@yesql.se> writes:\\n>>> On 10 Nov 2021, at 13:37, Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>>> ..but I wonder what's the *benefit* of removing those includes. IOW, what's\\n>>> the reason not to simply drop the patch?\\n>\\n>> I think the value is mostly neatnikism, the actual effect on runtime is\\n>> unlikely to be measureable. I won't argue against doing it, but I suspect\\n>> we'll just slowly add a lot of these back as tests evolve making excercise\\n>> less useful.\\n>\\n> Yeah, that last was pretty much my reaction. I don't know enough about\\n> Perl to be sure how much an unused import costs, but I suspect you're\\n> right that it won't be measurable in context, considering that most of\\n> these test scripts run at least one initdb.\\n\\nOn my laptop (using the \\\"performance\\\" CPU frequency governor to avoid\\nscaling noise) loading PostgreSQL::Test::Utils takes about 80ms:\\n\\n$ time perl -I./src/test/perl -e 'use PostgreSQL::Test::Utils;'\\n\\nreal 0m0.079s\\nuser 0m0.078s\\nsys 0m0.001s\\n\\nHowever, if we're also loading PostgreSQL::Test::Cluster, the difference\\nis negligible, because most of the time is spent loading modules used by\\nboth:\\n\\n$ time perl -I./src/test/perl -e 'use PostgreSQL::Test::Cluster; use PostgreSQL::Test::Utils;'\\n\\nreal 0m0.092s\\nuser 0m0.088s\\nsys 0m0.004s\\n\\n$ time perl -I./src/test/perl -e 'use PostgreSQL::Test::Cluster;'\\n\\nreal 0m0.090s\\nuser 0m0.081s\\nsys 0m0.008s\\n\\n- ilmari\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 15:42:00 +0000\",\n \"msg_from\": \"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Removed unused import modules from tap tests\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 10, 2021 at 9:53 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Yeah, that last was pretty much my reaction. I don't know enough about\\n> Perl to be sure how much an unused import costs, but I suspect you're\\n> right that it won't be measurable in context, considering that most of\\n> these test scripts run at least one initdb.\\n\\nI think focusing on the runtime might be missing the point to some\\nextent. Removing used #include directives from C source files is\\ngenerally good practice just for cleanliness, and this is the same\\nkind of thing. That said, the argument that future changes are likely\\nto just put the same 'use' directive back seems to me to have some\\nmerit. But we've probably spent more energy debating this than the\\ntopic deserves. It wouldn't hurt to remove these, and it also won't\\nhurt if we don't.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 11:07:09 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Removed unused import modules from tap tests\"\n },\n {\n \"msg_contents\": \"\\nOn 11/10/21 09:53, Tom Lane wrote:\\n> Daniel Gustafsson <daniel@yesql.se> writes:\\n>>> On 10 Nov 2021, at 13:37, Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>>> ..but I wonder what's the *benefit* of removing those includes. IOW, what's\\n>>> the reason not to simply drop the patch?\\n>> I think the value is mostly neatnikism, the actual effect on runtime is\\n>> unlikely to be measureable. I won't argue against doing it, but I suspect\\n>> we'll just slowly add a lot of these back as tests evolve making excercise\\n>> less useful.\\n> Yeah, that last was pretty much my reaction. I don't know enough about\\n> Perl to be sure how much an unused import costs, but I suspect you're\\n> right that it won't be measurable in context, considering that most of\\n> these test scripts run at least one initdb.\\n>\\n> \\t\\t\\t\\n\\n\\n:Cluster uses :Utils, and perl is smart enough not to try to reprocess\\nthe module. Thus the extra cost here is almost certainly very close to zero.\\n\\n\\nThis is a perfectly reasonable piece of boilerplate to use at the top of\\na TAP test:\\n\\n\\n use strict;\\n\\n use warnings;\\n\\n use PostgreSQL:Test::Cluster;\\n\\n use PostgreSQL::Test::Utils;\\n\\n use Test::More;\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 13:17:46 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Removed unused import modules from tap tests\"\n }\n]","isConversational":false}}},{"rowIdx":53061,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\r\n\r\nAttached please find a draft of the release announcement for 2021-11-11. \r\nPlease provide any feedback no later than Thu, Nov 11, 2021 0:00 AoE[1].\r\n\r\nThanks!\r\n\r\nJonathan\r\n\r\n[1] https://en.wikipedia.org/wiki/Anywhere_on_Earth","msg_date":"Tue, 9 Nov 2021 11:25:37 -0500","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":true,"msg_subject":"2021-11-11 release announcement draft"},{"msg_contents":"Op 09-11-2021 om 17:25 schreef Jonathan S. Katz:\n> \n> Attached please find a draft of the release announcement for 2021-11-11. \n\n\"overflowed-subsraction\" should probably be\n\"overflowed-subtraction\"\n\n\nErik\n\n\n\n","msg_date":"Tue, 9 Nov 2021 18:10:11 +0100","msg_from":"Erik Rijkers <er@xs4all.nl>","msg_from_op":false,"msg_subject":"Re: 2021-11-11 release announcement draft"},{"msg_contents":"On 11/9/21 12:10 PM, Erik Rijkers wrote:\r\n> Op 09-11-2021 om 17:25 schreef Jonathan S. Katz:\r\n>>\r\n>> Attached please find a draft of the release announcement for 2021-11-11. \r\n> \r\n> \"overflowed-subsraction\" should probably be\r\n> \"overflowed-subtraction\"\r\n\r\nActually it should be \"overflowed-subtransaction\" but regardless, good \r\ncatch.\r\n\r\nFixed in the canonical copy. Thanks!\r\n\r\nJonathan","msg_date":"Tue, 9 Nov 2021 12:16:05 -0500","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":true,"msg_subject":"Re: 2021-11-11 release announcement draft"},{"msg_contents":"On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\n> * Fix is when creating a new range type with `CREATE TYPE` that could cause\n> problems for later event triggers or subsequent executions of the `CREATE TYPE`\n> command.\n\nI don't know what \"is when\" means, but it seems wrong.\n\n> * Fix restoration of a Portal's snapshot inside a subtransaction, which could\n> lead to a crash. For example, this coul doccur in PL/pgSQL when a `COMMIT` is\n\ncould occur\n\n-- \nJustin\n\n\n","msg_date":"Tue, 9 Nov 2021 14:19:37 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: 2021-11-11 release announcement draft"},{"msg_contents":"On 11/9/21 3:19 PM, Justin Pryzby wrote:\r\n> On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\r\n>> * Fix is when creating a new range type with `CREATE TYPE` that could cause\r\n>> problems for later event triggers or subsequent executions of the `CREATE TYPE`\r\n>> command.\r\n> \r\n> I don't know what \"is when\" means, but it seems wrong.\r\n> \r\n>> * Fix restoration of a Portal's snapshot inside a subtransaction, which could\r\n>> lead to a crash. For example, this coul doccur in PL/pgSQL when a `COMMIT` is\r\n> \r\n> could occur\r\n\r\nBoth fixed. Thanks!\r\n\r\nJonathan","msg_date":"Tue, 9 Nov 2021 15:45:34 -0500","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":true,"msg_subject":"Re: 2021-11-11 release announcement draft"},{"msg_contents":"On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\n> Attached please find a draft of the release announcement for 2021-11-11.\n> Please provide any feedback no later than Thu, Nov 11, 2021 0:00 AoE[1].\n\n> Bug Fixes and Improvements\n> --------------------------\n> \n> This update fixes over 40 bugs that were reported in the last several months.\n> Some of these issues only affect version 14, but could also apply to other\n> supported versions.\n\nThe part after \"but\" reads oddly to me. I would just delete the sentence.\n\n> * Fix `CREATE INDEX CONCURRENTLY` to wait for the latest prepared transactions.\n> If you have enabled prepared transactions (`max_prepared_transactions` > 0), you\n> should reindex any concurrently built indexes.\n\nThe other CONCURRENTLY fix (\"Avoid race condition\" in the release notes) is\nbroader, so please replace the above entry with something like this, not even\nmentioning prepared transactions:\n\n* Fix causes of `CREATE INDEX CONCURRENTLY` and `REINDEX CONCURRENTLY` writing\ncorrupt indexes. You should reindex any concurrently-built indexes.\n\n> * Disallow `ALTER INDEX index ALTER COLUMN col SET (options)`.\n\nI'd drop this since the feature was unusable before.\n\nPlease add something like:\n\n* Fix REINDEX CONCURRENTLY to preserve operator class parameters that were\nattached to the target index\n\n\n","msg_date":"Tue, 9 Nov 2021 19:39:11 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":false,"msg_subject":"Re: 2021-11-11 release announcement draft"},{"msg_contents":"On 11/9/21 10:39 PM, Noah Misch wrote:\r\n> On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\r\n>> Attached please find a draft of the release announcement for 2021-11-11.\r\n>> Please provide any feedback no later than Thu, Nov 11, 2021 0:00 AoE[1].\r\n> \r\n>> Bug Fixes and Improvements\r\n>> --------------------------\r\n>>\r\n>> This update fixes over 40 bugs that were reported in the last several months.\r\n>> Some of these issues only affect version 14, but could also apply to other\r\n>> supported versions.\r\n> \r\n> The part after \"but\" reads oddly to me. I would just delete the sentence.\r\n\r\nI'll change to:\r\n\r\n\"Some of these issues only affect PostgreSQL 14, but may also apply to \r\nother supported versions of PostgreSQL.\"\r\n\r\nI'm a strong -1 on dropping the sentence altogether given the release \r\nannouncement only highlights fixes known to be in $LATEST, but they may \r\nalso be fixed in releases < $LATEST.\r\n\r\n> * Fix causes of `CREATE INDEX CONCURRENTLY` and `REINDEX CONCURRENTLY` writing\r\n> corrupt indexes. You should reindex any concurrently-built indexes.\r\n\r\nDone.\r\n\r\n>> * Disallow `ALTER INDEX index ALTER COLUMN col SET (options)`.\r\n> \r\n> I'd drop this since the feature was unusable before.\r\n\r\nDone.\r\n\r\n> Please add something like:\r\n> \r\n> * Fix REINDEX CONCURRENTLY to preserve operator class parameters that were\r\n> attached to the target index\r\n\r\nI wondered how widely this affected users, which is why I had initially \r\nomitted it. I looked at the test cases provided in f255de9a45 and for \r\nthat case, the only reference I could find to the opsclass parameter was \r\non[1], and that was after a nontrivial bit of searching.\r\n\r\nI know operator classes are more frequently attached to indexes, but I \r\nhaven't seen parameters very often.\r\n\r\nHappy to be convinced otherwise (maybe this is used more than I \r\nrealize?), but I'm not sure if we need that bullet point in the release \r\nannouncement.\r\n\r\nThanks,\r\n\r\nJonathan\r\n\r\n[1] https://www.postgresql.org/docs/current/textsearch-indexes.html","msg_date":"Wed, 10 Nov 2021 14:07:43 -0500","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":true,"msg_subject":"Re: 2021-11-11 release announcement draft"},{"msg_contents":"\"Jonathan S. Katz\" <jkatz@postgresql.org> writes:\n> On 11/9/21 10:39 PM, Noah Misch wrote:\n>> Please add something like:\n>> * Fix REINDEX CONCURRENTLY to preserve operator class parameters that were\n>> attached to the target index\n\n> Happy to be convinced otherwise (maybe this is used more than I \n> realize?), but I'm not sure if we need that bullet point in the release \n> announcement.\n\nFWIW, I agree with Jonathan's choice here. Opclass parameters are a\nvery tiny minority use-case as of today. (Were they not, we'd probably\nhave found this bug sooner.)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 10 Nov 2021 14:37:03 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: 2021-11-11 release announcement draft"},{"msg_contents":"On Wed, Nov 10, 2021 at 11:07 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n> > * Fix causes of `CREATE INDEX CONCURRENTLY` and `REINDEX CONCURRENTLY` writing\n> > corrupt indexes. You should reindex any concurrently-built indexes.\n>\n> Done.\n\nAs far as I know (correct me if I'm mistaken), all of the CIC/RC bugs\nare detectable using amcheck's heapallindexed verification feature.\nWhy not say as much? Lots of people create lots of indexes with CIC,\nand so giving more actionable advice seems worthwhile.\n\n(Unfortunately the same cannot really be said for the 14-only bug in\nparallel VACUUM. Seeing what amcheck (or pg_amcheck) have to say is\ncertainly better than assuming all indexes are corrupt (when they\nmight have been manually VACUUMed), but it's not the same situation. I\ndon't think that it's all that unlikely that the B-Tree verification\nstuff would miss real problems, since even heapallindexed only notices\nthe absence of needed index tuples. Whereas the VACUUM bug is mostly\n(but not entirely) characterized by the *presence* of \"extra\", corrupt\nindex tuples -- tuples that VACUUM should have already removed, but\ndidn't.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Wed, 10 Nov 2021 15:33:08 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: 2021-11-11 release announcement draft"},{"msg_contents":"On Wed, Nov 10, 2021 at 02:07:43PM -0500, Jonathan S. Katz wrote:\n> On 11/9/21 10:39 PM, Noah Misch wrote:\n> >On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\n> >>Attached please find a draft of the release announcement for 2021-11-11.\n> >>Please provide any feedback no later than Thu, Nov 11, 2021 0:00 AoE[1].\n> >\n> >>Bug Fixes and Improvements\n> >>--------------------------\n> >>\n> >>This update fixes over 40 bugs that were reported in the last several months.\n> >>Some of these issues only affect version 14, but could also apply to other\n> >>supported versions.\n> >\n> >The part after \"but\" reads oddly to me. I would just delete the sentence.\n> \n> I'll change to:\n> \n> \"Some of these issues only affect PostgreSQL 14, but may also apply to other\n> supported versions of PostgreSQL.\"\n\nI think that's equivalent to, \"Some of these issues only affect PostgreSQL 14.\nThe issues that only affect PostgreSQL 14 may also apply to other supported\nversions of PostgreSQL.\" That contradiction is the problem. This spot calls\nfor \"Not all issues affect all supported versions of PostgreSQL.\" or\nequivalent.\n\n> I'm a strong -1 on dropping the sentence altogether given the release\n> announcement only highlights fixes known to be in $LATEST, but they may also\n> be fixed in releases < $LATEST.\n\nOkay.\n\n\n","msg_date":"Wed, 10 Nov 2021 19:47:16 -0800","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":false,"msg_subject":"Re: 2021-11-11 release announcement draft"},{"msg_contents":"\n\n> 11 нояб. 2021 г., в 04:33, Peter Geoghegan <pg@bowt.ie> написал(а):\n> \n> On Wed, Nov 10, 2021 at 11:07 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>>> * Fix causes of `CREATE INDEX CONCURRENTLY` and `REINDEX CONCURRENTLY` writing\n>>> corrupt indexes. You should reindex any concurrently-built indexes.\n>> \n>> Done.\n> \n> As far as I know (correct me if I'm mistaken), all of the CIC/RC bugs\n> are detectable using amcheck's heapallindexed verification feature.\nYes, but only for B-trees.\n\nBest regards, Andrey Borodin.\n\n\n\n","msg_date":"Thu, 11 Nov 2021 17:17:23 +0500","msg_from":"Andrey Borodin <x4mmm@yandex-team.ru>","msg_from_op":false,"msg_subject":"Re: 2021-11-11 release announcement draft"},{"msg_contents":"On 11/10/21 10:47 PM, Noah Misch wrote:\r\n> On Wed, Nov 10, 2021 at 02:07:43PM -0500, Jonathan S. Katz wrote:\r\n>> On 11/9/21 10:39 PM, Noah Misch wrote:\r\n>>> On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\r\n>>>> Attached please find a draft of the release announcement for 2021-11-11.\r\n>>>> Please provide any feedback no later than Thu, Nov 11, 2021 0:00 AoE[1].\r\n>>>\r\n>>>> Bug Fixes and Improvements\r\n>>>> --------------------------\r\n>>>>\r\n>>>> This update fixes over 40 bugs that were reported in the last several months.\r\n>>>> Some of these issues only affect version 14, but could also apply to other\r\n>>>> supported versions.\r\n>>>\r\n>>> The part after \"but\" reads oddly to me. I would just delete the sentence.\r\n>>\r\n>> I'll change to:\r\n>>\r\n>> \"Some of these issues only affect PostgreSQL 14, but may also apply to other\r\n>> supported versions of PostgreSQL.\"\r\n> \r\n> I think that's equivalent to, \"Some of these issues only affect PostgreSQL 14.\r\n> The issues that only affect PostgreSQL 14 may also apply to other supported\r\n> versions of PostgreSQL.\" That contradiction is the problem. This spot calls\r\n> for \"Not all issues affect all supported versions of PostgreSQL.\" or\r\n> equivalent.\r\n\r\nBased on that, I swapped it to:\r\n\r\n\"The issues listed below affect PostgreSQL 14. Some of these issues may \r\nalso affect other supported versions of PostgreSQL.\"\r\n\r\nwhich I believe encompasses all of your points.\r\n\r\nJonathan","msg_date":"Thu, 11 Nov 2021 08:54:41 -0500","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":true,"msg_subject":"Re: 2021-11-11 release announcement draft"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\r\\n\\r\\nAttached please find a draft of the release announcement for 2021-11-11. \\r\\nPlease provide any feedback no later than Thu, Nov 11, 2021 0:00 AoE[1].\\r\\n\\r\\nThanks!\\r\\n\\r\\nJonathan\\r\\n\\r\\n[1] https://en.wikipedia.org/wiki/Anywhere_on_Earth\",\n \"msg_date\": \"Tue, 9 Nov 2021 11:25:37 -0500\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"2021-11-11 release announcement draft\"\n },\n {\n \"msg_contents\": \"Op 09-11-2021 om 17:25 schreef Jonathan S. Katz:\\n> \\n> Attached please find a draft of the release announcement for 2021-11-11. \\n\\n\\\"overflowed-subsraction\\\" should probably be\\n\\\"overflowed-subtraction\\\"\\n\\n\\nErik\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 18:10:11 +0100\",\n \"msg_from\": \"Erik Rijkers <er@xs4all.nl>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 2021-11-11 release announcement draft\"\n },\n {\n \"msg_contents\": \"On 11/9/21 12:10 PM, Erik Rijkers wrote:\\r\\n> Op 09-11-2021 om 17:25 schreef Jonathan S. Katz:\\r\\n>>\\r\\n>> Attached please find a draft of the release announcement for 2021-11-11. \\r\\n> \\r\\n> \\\"overflowed-subsraction\\\" should probably be\\r\\n> \\\"overflowed-subtraction\\\"\\r\\n\\r\\nActually it should be \\\"overflowed-subtransaction\\\" but regardless, good \\r\\ncatch.\\r\\n\\r\\nFixed in the canonical copy. Thanks!\\r\\n\\r\\nJonathan\",\n \"msg_date\": \"Tue, 9 Nov 2021 12:16:05 -0500\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: 2021-11-11 release announcement draft\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\\n> * Fix is when creating a new range type with `CREATE TYPE` that could cause\\n> problems for later event triggers or subsequent executions of the `CREATE TYPE`\\n> command.\\n\\nI don't know what \\\"is when\\\" means, but it seems wrong.\\n\\n> * Fix restoration of a Portal's snapshot inside a subtransaction, which could\\n> lead to a crash. For example, this coul doccur in PL/pgSQL when a `COMMIT` is\\n\\ncould occur\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 14:19:37 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 2021-11-11 release announcement draft\"\n },\n {\n \"msg_contents\": \"On 11/9/21 3:19 PM, Justin Pryzby wrote:\\r\\n> On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\\r\\n>> * Fix is when creating a new range type with `CREATE TYPE` that could cause\\r\\n>> problems for later event triggers or subsequent executions of the `CREATE TYPE`\\r\\n>> command.\\r\\n> \\r\\n> I don't know what \\\"is when\\\" means, but it seems wrong.\\r\\n> \\r\\n>> * Fix restoration of a Portal's snapshot inside a subtransaction, which could\\r\\n>> lead to a crash. For example, this coul doccur in PL/pgSQL when a `COMMIT` is\\r\\n> \\r\\n> could occur\\r\\n\\r\\nBoth fixed. Thanks!\\r\\n\\r\\nJonathan\",\n \"msg_date\": \"Tue, 9 Nov 2021 15:45:34 -0500\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: 2021-11-11 release announcement draft\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\\n> Attached please find a draft of the release announcement for 2021-11-11.\\n> Please provide any feedback no later than Thu, Nov 11, 2021 0:00 AoE[1].\\n\\n> Bug Fixes and Improvements\\n> --------------------------\\n> \\n> This update fixes over 40 bugs that were reported in the last several months.\\n> Some of these issues only affect version 14, but could also apply to other\\n> supported versions.\\n\\nThe part after \\\"but\\\" reads oddly to me. I would just delete the sentence.\\n\\n> * Fix `CREATE INDEX CONCURRENTLY` to wait for the latest prepared transactions.\\n> If you have enabled prepared transactions (`max_prepared_transactions` > 0), you\\n> should reindex any concurrently built indexes.\\n\\nThe other CONCURRENTLY fix (\\\"Avoid race condition\\\" in the release notes) is\\nbroader, so please replace the above entry with something like this, not even\\nmentioning prepared transactions:\\n\\n* Fix causes of `CREATE INDEX CONCURRENTLY` and `REINDEX CONCURRENTLY` writing\\ncorrupt indexes. You should reindex any concurrently-built indexes.\\n\\n> * Disallow `ALTER INDEX index ALTER COLUMN col SET (options)`.\\n\\nI'd drop this since the feature was unusable before.\\n\\nPlease add something like:\\n\\n* Fix REINDEX CONCURRENTLY to preserve operator class parameters that were\\nattached to the target index\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 19:39:11 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 2021-11-11 release announcement draft\"\n },\n {\n \"msg_contents\": \"On 11/9/21 10:39 PM, Noah Misch wrote:\\r\\n> On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\\r\\n>> Attached please find a draft of the release announcement for 2021-11-11.\\r\\n>> Please provide any feedback no later than Thu, Nov 11, 2021 0:00 AoE[1].\\r\\n> \\r\\n>> Bug Fixes and Improvements\\r\\n>> --------------------------\\r\\n>>\\r\\n>> This update fixes over 40 bugs that were reported in the last several months.\\r\\n>> Some of these issues only affect version 14, but could also apply to other\\r\\n>> supported versions.\\r\\n> \\r\\n> The part after \\\"but\\\" reads oddly to me. I would just delete the sentence.\\r\\n\\r\\nI'll change to:\\r\\n\\r\\n\\\"Some of these issues only affect PostgreSQL 14, but may also apply to \\r\\nother supported versions of PostgreSQL.\\\"\\r\\n\\r\\nI'm a strong -1 on dropping the sentence altogether given the release \\r\\nannouncement only highlights fixes known to be in $LATEST, but they may \\r\\nalso be fixed in releases < $LATEST.\\r\\n\\r\\n> * Fix causes of `CREATE INDEX CONCURRENTLY` and `REINDEX CONCURRENTLY` writing\\r\\n> corrupt indexes. You should reindex any concurrently-built indexes.\\r\\n\\r\\nDone.\\r\\n\\r\\n>> * Disallow `ALTER INDEX index ALTER COLUMN col SET (options)`.\\r\\n> \\r\\n> I'd drop this since the feature was unusable before.\\r\\n\\r\\nDone.\\r\\n\\r\\n> Please add something like:\\r\\n> \\r\\n> * Fix REINDEX CONCURRENTLY to preserve operator class parameters that were\\r\\n> attached to the target index\\r\\n\\r\\nI wondered how widely this affected users, which is why I had initially \\r\\nomitted it. I looked at the test cases provided in f255de9a45 and for \\r\\nthat case, the only reference I could find to the opsclass parameter was \\r\\non[1], and that was after a nontrivial bit of searching.\\r\\n\\r\\nI know operator classes are more frequently attached to indexes, but I \\r\\nhaven't seen parameters very often.\\r\\n\\r\\nHappy to be convinced otherwise (maybe this is used more than I \\r\\nrealize?), but I'm not sure if we need that bullet point in the release \\r\\nannouncement.\\r\\n\\r\\nThanks,\\r\\n\\r\\nJonathan\\r\\n\\r\\n[1] https://www.postgresql.org/docs/current/textsearch-indexes.html\",\n \"msg_date\": \"Wed, 10 Nov 2021 14:07:43 -0500\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: 2021-11-11 release announcement draft\"\n },\n {\n \"msg_contents\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org> writes:\\n> On 11/9/21 10:39 PM, Noah Misch wrote:\\n>> Please add something like:\\n>> * Fix REINDEX CONCURRENTLY to preserve operator class parameters that were\\n>> attached to the target index\\n\\n> Happy to be convinced otherwise (maybe this is used more than I \\n> realize?), but I'm not sure if we need that bullet point in the release \\n> announcement.\\n\\nFWIW, I agree with Jonathan's choice here. Opclass parameters are a\\nvery tiny minority use-case as of today. (Were they not, we'd probably\\nhave found this bug sooner.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 14:37:03 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 2021-11-11 release announcement draft\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 10, 2021 at 11:07 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n> > * Fix causes of `CREATE INDEX CONCURRENTLY` and `REINDEX CONCURRENTLY` writing\\n> > corrupt indexes. You should reindex any concurrently-built indexes.\\n>\\n> Done.\\n\\nAs far as I know (correct me if I'm mistaken), all of the CIC/RC bugs\\nare detectable using amcheck's heapallindexed verification feature.\\nWhy not say as much? Lots of people create lots of indexes with CIC,\\nand so giving more actionable advice seems worthwhile.\\n\\n(Unfortunately the same cannot really be said for the 14-only bug in\\nparallel VACUUM. Seeing what amcheck (or pg_amcheck) have to say is\\ncertainly better than assuming all indexes are corrupt (when they\\nmight have been manually VACUUMed), but it's not the same situation. I\\ndon't think that it's all that unlikely that the B-Tree verification\\nstuff would miss real problems, since even heapallindexed only notices\\nthe absence of needed index tuples. Whereas the VACUUM bug is mostly\\n(but not entirely) characterized by the *presence* of \\\"extra\\\", corrupt\\nindex tuples -- tuples that VACUUM should have already removed, but\\ndidn't.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 15:33:08 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 2021-11-11 release announcement draft\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 10, 2021 at 02:07:43PM -0500, Jonathan S. Katz wrote:\\n> On 11/9/21 10:39 PM, Noah Misch wrote:\\n> >On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\\n> >>Attached please find a draft of the release announcement for 2021-11-11.\\n> >>Please provide any feedback no later than Thu, Nov 11, 2021 0:00 AoE[1].\\n> >\\n> >>Bug Fixes and Improvements\\n> >>--------------------------\\n> >>\\n> >>This update fixes over 40 bugs that were reported in the last several months.\\n> >>Some of these issues only affect version 14, but could also apply to other\\n> >>supported versions.\\n> >\\n> >The part after \\\"but\\\" reads oddly to me. I would just delete the sentence.\\n> \\n> I'll change to:\\n> \\n> \\\"Some of these issues only affect PostgreSQL 14, but may also apply to other\\n> supported versions of PostgreSQL.\\\"\\n\\nI think that's equivalent to, \\\"Some of these issues only affect PostgreSQL 14.\\nThe issues that only affect PostgreSQL 14 may also apply to other supported\\nversions of PostgreSQL.\\\" That contradiction is the problem. This spot calls\\nfor \\\"Not all issues affect all supported versions of PostgreSQL.\\\" or\\nequivalent.\\n\\n> I'm a strong -1 on dropping the sentence altogether given the release\\n> announcement only highlights fixes known to be in $LATEST, but they may also\\n> be fixed in releases < $LATEST.\\n\\nOkay.\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 19:47:16 -0800\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 2021-11-11 release announcement draft\"\n },\n {\n \"msg_contents\": \"\\n\\n> 11 нояб. 2021 г., в 04:33, Peter Geoghegan <pg@bowt.ie> написал(а):\\n> \\n> On Wed, Nov 10, 2021 at 11:07 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>>> * Fix causes of `CREATE INDEX CONCURRENTLY` and `REINDEX CONCURRENTLY` writing\\n>>> corrupt indexes. You should reindex any concurrently-built indexes.\\n>> \\n>> Done.\\n> \\n> As far as I know (correct me if I'm mistaken), all of the CIC/RC bugs\\n> are detectable using amcheck's heapallindexed verification feature.\\nYes, but only for B-trees.\\n\\nBest regards, Andrey Borodin.\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 17:17:23 +0500\",\n \"msg_from\": \"Andrey Borodin <x4mmm@yandex-team.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 2021-11-11 release announcement draft\"\n },\n {\n \"msg_contents\": \"On 11/10/21 10:47 PM, Noah Misch wrote:\\r\\n> On Wed, Nov 10, 2021 at 02:07:43PM -0500, Jonathan S. Katz wrote:\\r\\n>> On 11/9/21 10:39 PM, Noah Misch wrote:\\r\\n>>> On Tue, Nov 09, 2021 at 11:25:37AM -0500, Jonathan S. Katz wrote:\\r\\n>>>> Attached please find a draft of the release announcement for 2021-11-11.\\r\\n>>>> Please provide any feedback no later than Thu, Nov 11, 2021 0:00 AoE[1].\\r\\n>>>\\r\\n>>>> Bug Fixes and Improvements\\r\\n>>>> --------------------------\\r\\n>>>>\\r\\n>>>> This update fixes over 40 bugs that were reported in the last several months.\\r\\n>>>> Some of these issues only affect version 14, but could also apply to other\\r\\n>>>> supported versions.\\r\\n>>>\\r\\n>>> The part after \\\"but\\\" reads oddly to me. I would just delete the sentence.\\r\\n>>\\r\\n>> I'll change to:\\r\\n>>\\r\\n>> \\\"Some of these issues only affect PostgreSQL 14, but may also apply to other\\r\\n>> supported versions of PostgreSQL.\\\"\\r\\n> \\r\\n> I think that's equivalent to, \\\"Some of these issues only affect PostgreSQL 14.\\r\\n> The issues that only affect PostgreSQL 14 may also apply to other supported\\r\\n> versions of PostgreSQL.\\\" That contradiction is the problem. This spot calls\\r\\n> for \\\"Not all issues affect all supported versions of PostgreSQL.\\\" or\\r\\n> equivalent.\\r\\n\\r\\nBased on that, I swapped it to:\\r\\n\\r\\n\\\"The issues listed below affect PostgreSQL 14. Some of these issues may \\r\\nalso affect other supported versions of PostgreSQL.\\\"\\r\\n\\r\\nwhich I believe encompasses all of your points.\\r\\n\\r\\nJonathan\",\n \"msg_date\": \"Thu, 11 Nov 2021 08:54:41 -0500\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: 2021-11-11 release announcement draft\"\n }\n]","isConversational":false}}},{"rowIdx":53062,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Coverity complained about this code in bbstreamer_file.c,\nsaying that the get_gz_error() call could be accessing freed\nmemory:\n\n if (gzclose(mystreamer->gzfile) != 0)\n {\n pg_log_error(\"could not close compressed file \\\"%s\\\": %s\",\n mystreamer->pathname,\n get_gz_error(mystreamer->gzfile));\n exit(1);\n }\n\nI think it's right. This certainly isn't sanctioned by the zlib\nspecification [1], nor does it look like our other gzclose() calls,\nwhich simply consult errno. (Note that this coding is not at all new,\nbut Coverity thinks it is because 23a1c6578 moved it to a new file.)\n\nI set out to fix that, intending only to replace the use of\nget_gz_error() with %m, but the patch soon metastasized to the\npoint where I think it needs review :-(.\n\n0001 below addresses places that either didn't check the result of\ngzclose() at all, or neglected to print a useful error message.\nThe zlib spec doesn't quite promise that a failed gzclose() will\nset errno, but the cases where it might not seem like can't-happen\ncases that we needn't spend much effort on. So what I've done here\nis just to initialize errno to zero before gzclose(). If we ever\nget a report of \"could not close file: Success\" in one of these\nplaces, we'll know to look more closely.\n\n0002 below addresses a rather messier problem, which is that the\nerror handling in src/bin/pg_basebackup/walmethods.c is just\nappallingly bad. The design is awful, because it relies on errno\nto hold still for much longer than we can sanely expect --- notably,\nwalmethods.c really shouldn't be assuming much about what the callers\nmight do between calling an error-generating method and calling\ngetlasterror(). Worse, there are a lot of places that naively expect\nerrno to hold still across free() and similar operations. We know\nthat not to be the case on (at least) macOS. On top of that,\ncommit babbbb595 (LZ4 support) broke the design completely, because\nliblz4 doesn't use errno to report errors. Separately from the\nquestion of how to report errors, there were a number of places\nthat failed to detect errors in the first place, and more that were\nsloppy about setting the correct error code for a short write.\n\nWhat I chose to do to fix the design problem is to use a modified\nversion of the idea that existed in the tar-methods part of the file,\nwhich is to have both a string field and an errno field in the\nDirectoryMethodData and TarMethodData objects. If the string isn't\nNULL then it overrides the errno field, allowing us to report\nnon-errno problems. Then, a bunch of places have to remember to copy\nerrno into the lasterrno field, which is a bit annoying. On the other\nhand, we can get rid of logic that tries to save and restore errno\nacross functions that might change it, so this does buy back some\ncomplexity too.\n\nThere's at least one loose end here, which is that there's one\nplace in tar_close() that does a summary exit(1) on fsync failure,\nwithout even bothering with an error message. I added the\nmissing error report:\n \n /* Always fsync on close, so the padding gets fsynced */\n if (tar_sync(f) < 0)\n+ {\n+ /* XXX this seems pretty bogus; why is only this case fatal? */\n+ pg_log_fatal(\"could not fsync file \\\"%s\\\": %s\",\n+ tf->pathname, tar_getlasterror());\n exit(1);\n+ }\n\nbut this code seems about as fishy and ill-thought-out as anything\nelse I've touched here. Why is this different from the half-dozen\nother fsync-error cases in the same file? Why, if fsync failure\nhere is so catastrophic, is it okay to just return a normal failure\ncode when tar_close doesn't even get to this point because of some\nearlier error? At the very least it seems like it'd be preferable\nto do the exit(1) at the caller level.\n\nThe addition of the exit(1) seems to have been intentional in\n1e2fddfa3, so cc'ing Peter for comment.\n\n\t\t\tregards, tom lane\n\n[1] https://refspecs.linuxbase.org/LSB_3.0.0/LSB-Core-generic/LSB-Core-generic/zlib-gzclose-1.html","msg_date":"Tue, 09 Nov 2021 15:20:31 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Deficient error handling in pg_dump and pg_basebackup"},{"msg_contents":"On Tue, Nov 09, 2021 at 03:20:31PM -0500, Tom Lane wrote:\n> I think it's right. This certainly isn't sanctioned by the zlib\n> specification [1], nor does it look like our other gzclose() calls,\n> which simply consult errno. (Note that this coding is not at all new,\n> but Coverity thinks it is because 23a1c6578 moved it to a new file.)\n> \n> I set out to fix that, intending only to replace the use of\n> get_gz_error() with %m, but the patch soon metastasized to the\n> point where I think it needs review :-(.\n\nThis points gzclose_w() to gzwrite.c. Looking at its end, it could\nreturn Z_ERRNO after freeing everything when failing to close the\nstate's fd. So I agree that coverity is right about this\npossibility. That's unlikely going to happen, though.\n\n> 0001 below addresses places that either didn't check the result of\n> gzclose() at all, or neglected to print a useful error message.\n> The zlib spec doesn't quite promise that a failed gzclose() will\n> set errno, but the cases where it might not seem like can't-happen\n> cases that we needn't spend much effort on.\n\nThe zlib code does not do much regarding that, but system calls\nwould. Still that joins with the point you are doing below where\nfree() may not hold errno, and note that the last thing gzclose() does\nis a free(). :p\n\n> So what I've done here\n> is just to initialize errno to zero before gzclose(). If we ever\n> get a report of \"could not close file: Success\" in one of these\n> places, we'll know to look more closely.\n\nOkay.\n\n> On top of that,\n> commit babbbb595 (LZ4 support) broke the design completely, because\n> liblz4 doesn't use errno to report errors. Separately from the\n> question of how to report errors, there were a number of places\n> that failed to detect errors in the first place, and more that were\n> sloppy about setting the correct error code for a short write.\n\nThis one's my fault, sorry about that :/\n\nWhat you are doing here looks fine for the LZ4F_isError() paths.\n\n> What I chose to do to fix the design problem is to use a modified\n> version of the idea that existed in the tar-methods part of the file,\n> which is to have both a string field and an errno field in the\n> DirectoryMethodData and TarMethodData objects. If the string isn't\n> NULL then it overrides the errno field, allowing us to report\n> non-errno problems. Then, a bunch of places have to remember to copy\n> errno into the lasterrno field, which is a bit annoying. On the other\n> hand, we can get rid of logic that tries to save and restore errno\n> across functions that might change it, so this does buy back some\n> complexity too.\n\nThis makes the code a bit harder to follow when it comes to cascading\ncalls with tar_write_*() or tar_close(), but your approach looks\nsimple enough for this code. I don't think you have missed a spot\nafter scanning the whole area.\n\n> There's at least one loose end here, which is that there's one\n> place in tar_close() that does a summary exit(1) on fsync failure,\n> without even bothering with an error message. I added the\n> missing error report:\n> \n> /* Always fsync on close, so the padding gets fsynced */\n> if (tar_sync(f) < 0)\n> + {\n> + /* XXX this seems pretty bogus; why is only this case fatal? */\n> + pg_log_fatal(\"could not fsync file \\\"%s\\\": %s\",\n> + tf->pathname, tar_getlasterror());\n> exit(1);\n> + }\n\n> but this code seems about as fishy and ill-thought-out as anything\n> else I've touched here. Why is this different from the half-dozen\n> other fsync-error cases in the same file? Why, if fsync failure\n> here is so catastrophic, is it okay to just return a normal failure\n> code when tar_close doesn't even get to this point because of some\n> earlier error?\n\nHmm, I don't think that's fine. There is an extra one in tar_finish()\nthat would report a failure, but not exit() immediately. All the\ncallers of the sync callback in receivelog.c exit() on sight, but I am\nwondering if it would not be saner to just exit() from walmethods.c\neach time we see a failure with a fsync().\n\n> At the very least it seems like it'd be preferable\n> to do the exit(1) at the caller level.\n\nYou mean walmethods.c here, right?\n--\nMichael","msg_date":"Wed, 10 Nov 2021 14:03:11 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Deficient error handling in pg_dump and pg_basebackup"},{"msg_contents":"On Wed, Nov 10, 2021 at 02:03:11PM +0900, Michael Paquier wrote:\n>> but this code seems about as fishy and ill-thought-out as anything\n>> else I've touched here. Why is this different from the half-dozen\n>> other fsync-error cases in the same file? Why, if fsync failure\n>> here is so catastrophic, is it okay to just return a normal failure\n>> code when tar_close doesn't even get to this point because of some\n>> earlier error?\n> \n> Hmm, I don't think that's fine. There is an extra one in tar_finish()\n> that would report a failure, but not exit() immediately. All the\n> callers of the sync callback in receivelog.c exit() on sight, but I am\n> wondering if it would not be saner to just exit() from walmethods.c\n> each time we see a failure with a fsync().\n\nTaking the issues with fsync() aside, which could be improved as a\nseparate patch, is there anything I can do for this thread? The error\nhandling problems in walmethods.c introduced by the integration of LZ4\nare on me, so I'd like to fix it. 0002 depends on some parts of 0001,\nas well for walmethods.c and the new error code paths. And I have\nbeen through this code quite a lot recently.\n--\nMichael","msg_date":"Wed, 17 Nov 2021 11:44:19 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Deficient error handling in pg_dump and pg_basebackup"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> Taking the issues with fsync() aside, which could be improved as a\n> separate patch, is there anything I can do for this thread? The error\n> handling problems in walmethods.c introduced by the integration of LZ4\n> are on me, so I'd like to fix it. 0002 depends on some parts of 0001,\n> as well for walmethods.c and the new error code paths. And I have\n> been through this code quite a lot recently.\n\nI feel like doing an immediate exit() for these errors and not other\nones is a pretty terrible idea, mainly because it doesn't account for\nthe question of what to do with a failure that prevents us from getting\nto the fsync() call in the first place. So I'd like to see a better\ndesign rather than more quick hacking. I confess I don't have a clear\nidea of what \"a better design\" would look like.\n\nHowever, that's largely orthogonal to any of the things my proposed\npatches are trying to fix. If you want to review the patches without\nconsidering the fsync-error-handling problem, that'd be great.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 16 Nov 2021 22:26:11 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Deficient error handling in pg_dump and pg_basebackup"},{"msg_contents":"On Tue, Nov 16, 2021 at 10:26:11PM -0500, Tom Lane wrote:\n> I feel like doing an immediate exit() for these errors and not other\n> ones is a pretty terrible idea, mainly because it doesn't account for\n> the question of what to do with a failure that prevents us from getting\n> to the fsync() call in the first place. So I'd like to see a better\n> design rather than more quick hacking. I confess I don't have a\n> clear idea of what \"a better design\" would look like.\n\n[ .. thinks .. ]\n\nWe cannot really have something equivalent to data_sync_retry in the\nfrontends. But I'd like to think that it would be fine for\npg_basebackup to just exit() on this failure so as callers would be\nable to retry a base backup. pg_receivewal is more tricky though. An\nexit() would allow for flush retries of a previous WAL segment where\nthings failed, but that stands when using --no-loop (still the code\npath triggered by this option would not be used). When not using\n--no-loop, it may be better to actually just retry streaming from the\nprevious point so as the error should be reported from walmethods.c to\nthe upper stack anyway.\n\n> However, that's largely orthogonal to any of the things my proposed\n> patches are trying to fix. If you want to review the patches without\n> considering the fsync-error-handling problem, that'd be great.\n\nI have looked at them upthread, FWIW:\nhttps://www.postgresql.org/message-id/YYtSj5vlWp5faVXz@paquier.xyz\nYour proposals still look rather sane to me, after a second look.\n--\nMichael","msg_date":"Wed, 17 Nov 2021 14:24:59 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Deficient error handling in pg_dump and pg_basebackup"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Tue, Nov 16, 2021 at 10:26:11PM -0500, Tom Lane wrote:\n>> However, that's largely orthogonal to any of the things my proposed\n>> patches are trying to fix. If you want to review the patches without\n>> considering the fsync-error-handling problem, that'd be great.\n\n> I have looked at them upthread, FWIW:\n> https://www.postgresql.org/message-id/YYtSj5vlWp5faVXz@paquier.xyz\n> Your proposals still look rather sane to me, after a second look.\n\nPushed then; thanks for reviewing that. We can consider the fsync\nerror question at leisure.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 17 Nov 2021 14:19:20 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Deficient error handling in pg_dump and pg_basebackup"},{"msg_contents":"On Wed, Nov 17, 2021 at 02:19:20PM -0500, Tom Lane wrote:\n> Pushed then; thanks for reviewing that. We can consider the fsync\n> error question at leisure.\n\nFine by me. Thanks for the commit.\n--\nMichael","msg_date":"Thu, 18 Nov 2021 09:29:32 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Deficient error handling in pg_dump and pg_basebackup"},{"msg_contents":"On 09.11.21 21:20, Tom Lane wrote:\n> Why is this different from the half-dozen\n> other fsync-error cases in the same file? Why, if fsync failure\n> here is so catastrophic, is it okay to just return a normal failure\n> code when tar_close doesn't even get to this point because of some\n> earlier error? At the very least it seems like it'd be preferable\n> to do the exit(1) at the caller level.\n> \n> The addition of the exit(1) seems to have been intentional in\n> 1e2fddfa3, so cc'ing Peter for comment.\n\nThat commit addressed the behavior of fsync() failure in pg_receivewal \nand pg_recvlogical, which are long-running daemon processes, so this \nchange was analogous to the server-side changes at the time. I don't \nknow what the behavior of fsync() failure in pg_basebackup should be, so \ncalls that are only reachable from pg_basebackup are currently being \nhandled differently.\n\n\n","msg_date":"Fri, 19 Nov 2021 11:10:46 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Deficient error handling in pg_dump and pg_basebackup"}],"string":"[\n {\n \"msg_contents\": \"Coverity complained about this code in bbstreamer_file.c,\\nsaying that the get_gz_error() call could be accessing freed\\nmemory:\\n\\n if (gzclose(mystreamer->gzfile) != 0)\\n {\\n pg_log_error(\\\"could not close compressed file \\\\\\\"%s\\\\\\\": %s\\\",\\n mystreamer->pathname,\\n get_gz_error(mystreamer->gzfile));\\n exit(1);\\n }\\n\\nI think it's right. This certainly isn't sanctioned by the zlib\\nspecification [1], nor does it look like our other gzclose() calls,\\nwhich simply consult errno. (Note that this coding is not at all new,\\nbut Coverity thinks it is because 23a1c6578 moved it to a new file.)\\n\\nI set out to fix that, intending only to replace the use of\\nget_gz_error() with %m, but the patch soon metastasized to the\\npoint where I think it needs review :-(.\\n\\n0001 below addresses places that either didn't check the result of\\ngzclose() at all, or neglected to print a useful error message.\\nThe zlib spec doesn't quite promise that a failed gzclose() will\\nset errno, but the cases where it might not seem like can't-happen\\ncases that we needn't spend much effort on. So what I've done here\\nis just to initialize errno to zero before gzclose(). If we ever\\nget a report of \\\"could not close file: Success\\\" in one of these\\nplaces, we'll know to look more closely.\\n\\n0002 below addresses a rather messier problem, which is that the\\nerror handling in src/bin/pg_basebackup/walmethods.c is just\\nappallingly bad. The design is awful, because it relies on errno\\nto hold still for much longer than we can sanely expect --- notably,\\nwalmethods.c really shouldn't be assuming much about what the callers\\nmight do between calling an error-generating method and calling\\ngetlasterror(). Worse, there are a lot of places that naively expect\\nerrno to hold still across free() and similar operations. We know\\nthat not to be the case on (at least) macOS. On top of that,\\ncommit babbbb595 (LZ4 support) broke the design completely, because\\nliblz4 doesn't use errno to report errors. Separately from the\\nquestion of how to report errors, there were a number of places\\nthat failed to detect errors in the first place, and more that were\\nsloppy about setting the correct error code for a short write.\\n\\nWhat I chose to do to fix the design problem is to use a modified\\nversion of the idea that existed in the tar-methods part of the file,\\nwhich is to have both a string field and an errno field in the\\nDirectoryMethodData and TarMethodData objects. If the string isn't\\nNULL then it overrides the errno field, allowing us to report\\nnon-errno problems. Then, a bunch of places have to remember to copy\\nerrno into the lasterrno field, which is a bit annoying. On the other\\nhand, we can get rid of logic that tries to save and restore errno\\nacross functions that might change it, so this does buy back some\\ncomplexity too.\\n\\nThere's at least one loose end here, which is that there's one\\nplace in tar_close() that does a summary exit(1) on fsync failure,\\nwithout even bothering with an error message. I added the\\nmissing error report:\\n \\n /* Always fsync on close, so the padding gets fsynced */\\n if (tar_sync(f) < 0)\\n+ {\\n+ /* XXX this seems pretty bogus; why is only this case fatal? */\\n+ pg_log_fatal(\\\"could not fsync file \\\\\\\"%s\\\\\\\": %s\\\",\\n+ tf->pathname, tar_getlasterror());\\n exit(1);\\n+ }\\n\\nbut this code seems about as fishy and ill-thought-out as anything\\nelse I've touched here. Why is this different from the half-dozen\\nother fsync-error cases in the same file? Why, if fsync failure\\nhere is so catastrophic, is it okay to just return a normal failure\\ncode when tar_close doesn't even get to this point because of some\\nearlier error? At the very least it seems like it'd be preferable\\nto do the exit(1) at the caller level.\\n\\nThe addition of the exit(1) seems to have been intentional in\\n1e2fddfa3, so cc'ing Peter for comment.\\n\\n\\t\\t\\tregards, tom lane\\n\\n[1] https://refspecs.linuxbase.org/LSB_3.0.0/LSB-Core-generic/LSB-Core-generic/zlib-gzclose-1.html\",\n \"msg_date\": \"Tue, 09 Nov 2021 15:20:31 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Deficient error handling in pg_dump and pg_basebackup\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 09, 2021 at 03:20:31PM -0500, Tom Lane wrote:\\n> I think it's right. This certainly isn't sanctioned by the zlib\\n> specification [1], nor does it look like our other gzclose() calls,\\n> which simply consult errno. (Note that this coding is not at all new,\\n> but Coverity thinks it is because 23a1c6578 moved it to a new file.)\\n> \\n> I set out to fix that, intending only to replace the use of\\n> get_gz_error() with %m, but the patch soon metastasized to the\\n> point where I think it needs review :-(.\\n\\nThis points gzclose_w() to gzwrite.c. Looking at its end, it could\\nreturn Z_ERRNO after freeing everything when failing to close the\\nstate's fd. So I agree that coverity is right about this\\npossibility. That's unlikely going to happen, though.\\n\\n> 0001 below addresses places that either didn't check the result of\\n> gzclose() at all, or neglected to print a useful error message.\\n> The zlib spec doesn't quite promise that a failed gzclose() will\\n> set errno, but the cases where it might not seem like can't-happen\\n> cases that we needn't spend much effort on.\\n\\nThe zlib code does not do much regarding that, but system calls\\nwould. Still that joins with the point you are doing below where\\nfree() may not hold errno, and note that the last thing gzclose() does\\nis a free(). :p\\n\\n> So what I've done here\\n> is just to initialize errno to zero before gzclose(). If we ever\\n> get a report of \\\"could not close file: Success\\\" in one of these\\n> places, we'll know to look more closely.\\n\\nOkay.\\n\\n> On top of that,\\n> commit babbbb595 (LZ4 support) broke the design completely, because\\n> liblz4 doesn't use errno to report errors. Separately from the\\n> question of how to report errors, there were a number of places\\n> that failed to detect errors in the first place, and more that were\\n> sloppy about setting the correct error code for a short write.\\n\\nThis one's my fault, sorry about that :/\\n\\nWhat you are doing here looks fine for the LZ4F_isError() paths.\\n\\n> What I chose to do to fix the design problem is to use a modified\\n> version of the idea that existed in the tar-methods part of the file,\\n> which is to have both a string field and an errno field in the\\n> DirectoryMethodData and TarMethodData objects. If the string isn't\\n> NULL then it overrides the errno field, allowing us to report\\n> non-errno problems. Then, a bunch of places have to remember to copy\\n> errno into the lasterrno field, which is a bit annoying. On the other\\n> hand, we can get rid of logic that tries to save and restore errno\\n> across functions that might change it, so this does buy back some\\n> complexity too.\\n\\nThis makes the code a bit harder to follow when it comes to cascading\\ncalls with tar_write_*() or tar_close(), but your approach looks\\nsimple enough for this code. I don't think you have missed a spot\\nafter scanning the whole area.\\n\\n> There's at least one loose end here, which is that there's one\\n> place in tar_close() that does a summary exit(1) on fsync failure,\\n> without even bothering with an error message. I added the\\n> missing error report:\\n> \\n> /* Always fsync on close, so the padding gets fsynced */\\n> if (tar_sync(f) < 0)\\n> + {\\n> + /* XXX this seems pretty bogus; why is only this case fatal? */\\n> + pg_log_fatal(\\\"could not fsync file \\\\\\\"%s\\\\\\\": %s\\\",\\n> + tf->pathname, tar_getlasterror());\\n> exit(1);\\n> + }\\n\\n> but this code seems about as fishy and ill-thought-out as anything\\n> else I've touched here. Why is this different from the half-dozen\\n> other fsync-error cases in the same file? Why, if fsync failure\\n> here is so catastrophic, is it okay to just return a normal failure\\n> code when tar_close doesn't even get to this point because of some\\n> earlier error?\\n\\nHmm, I don't think that's fine. There is an extra one in tar_finish()\\nthat would report a failure, but not exit() immediately. All the\\ncallers of the sync callback in receivelog.c exit() on sight, but I am\\nwondering if it would not be saner to just exit() from walmethods.c\\neach time we see a failure with a fsync().\\n\\n> At the very least it seems like it'd be preferable\\n> to do the exit(1) at the caller level.\\n\\nYou mean walmethods.c here, right?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 10 Nov 2021 14:03:11 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Deficient error handling in pg_dump and pg_basebackup\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 10, 2021 at 02:03:11PM +0900, Michael Paquier wrote:\\n>> but this code seems about as fishy and ill-thought-out as anything\\n>> else I've touched here. Why is this different from the half-dozen\\n>> other fsync-error cases in the same file? Why, if fsync failure\\n>> here is so catastrophic, is it okay to just return a normal failure\\n>> code when tar_close doesn't even get to this point because of some\\n>> earlier error?\\n> \\n> Hmm, I don't think that's fine. There is an extra one in tar_finish()\\n> that would report a failure, but not exit() immediately. All the\\n> callers of the sync callback in receivelog.c exit() on sight, but I am\\n> wondering if it would not be saner to just exit() from walmethods.c\\n> each time we see a failure with a fsync().\\n\\nTaking the issues with fsync() aside, which could be improved as a\\nseparate patch, is there anything I can do for this thread? The error\\nhandling problems in walmethods.c introduced by the integration of LZ4\\nare on me, so I'd like to fix it. 0002 depends on some parts of 0001,\\nas well for walmethods.c and the new error code paths. And I have\\nbeen through this code quite a lot recently.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 17 Nov 2021 11:44:19 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Deficient error handling in pg_dump and pg_basebackup\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> Taking the issues with fsync() aside, which could be improved as a\\n> separate patch, is there anything I can do for this thread? The error\\n> handling problems in walmethods.c introduced by the integration of LZ4\\n> are on me, so I'd like to fix it. 0002 depends on some parts of 0001,\\n> as well for walmethods.c and the new error code paths. And I have\\n> been through this code quite a lot recently.\\n\\nI feel like doing an immediate exit() for these errors and not other\\nones is a pretty terrible idea, mainly because it doesn't account for\\nthe question of what to do with a failure that prevents us from getting\\nto the fsync() call in the first place. So I'd like to see a better\\ndesign rather than more quick hacking. I confess I don't have a clear\\nidea of what \\\"a better design\\\" would look like.\\n\\nHowever, that's largely orthogonal to any of the things my proposed\\npatches are trying to fix. If you want to review the patches without\\nconsidering the fsync-error-handling problem, that'd be great.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 22:26:11 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Deficient error handling in pg_dump and pg_basebackup\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 10:26:11PM -0500, Tom Lane wrote:\\n> I feel like doing an immediate exit() for these errors and not other\\n> ones is a pretty terrible idea, mainly because it doesn't account for\\n> the question of what to do with a failure that prevents us from getting\\n> to the fsync() call in the first place. So I'd like to see a better\\n> design rather than more quick hacking. I confess I don't have a\\n> clear idea of what \\\"a better design\\\" would look like.\\n\\n[ .. thinks .. ]\\n\\nWe cannot really have something equivalent to data_sync_retry in the\\nfrontends. But I'd like to think that it would be fine for\\npg_basebackup to just exit() on this failure so as callers would be\\nable to retry a base backup. pg_receivewal is more tricky though. An\\nexit() would allow for flush retries of a previous WAL segment where\\nthings failed, but that stands when using --no-loop (still the code\\npath triggered by this option would not be used). When not using\\n--no-loop, it may be better to actually just retry streaming from the\\nprevious point so as the error should be reported from walmethods.c to\\nthe upper stack anyway.\\n\\n> However, that's largely orthogonal to any of the things my proposed\\n> patches are trying to fix. If you want to review the patches without\\n> considering the fsync-error-handling problem, that'd be great.\\n\\nI have looked at them upthread, FWIW:\\nhttps://www.postgresql.org/message-id/YYtSj5vlWp5faVXz@paquier.xyz\\nYour proposals still look rather sane to me, after a second look.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 17 Nov 2021 14:24:59 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Deficient error handling in pg_dump and pg_basebackup\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Tue, Nov 16, 2021 at 10:26:11PM -0500, Tom Lane wrote:\\n>> However, that's largely orthogonal to any of the things my proposed\\n>> patches are trying to fix. If you want to review the patches without\\n>> considering the fsync-error-handling problem, that'd be great.\\n\\n> I have looked at them upthread, FWIW:\\n> https://www.postgresql.org/message-id/YYtSj5vlWp5faVXz@paquier.xyz\\n> Your proposals still look rather sane to me, after a second look.\\n\\nPushed then; thanks for reviewing that. We can consider the fsync\\nerror question at leisure.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 14:19:20 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Deficient error handling in pg_dump and pg_basebackup\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 02:19:20PM -0500, Tom Lane wrote:\\n> Pushed then; thanks for reviewing that. We can consider the fsync\\n> error question at leisure.\\n\\nFine by me. Thanks for the commit.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 18 Nov 2021 09:29:32 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Deficient error handling in pg_dump and pg_basebackup\"\n },\n {\n \"msg_contents\": \"On 09.11.21 21:20, Tom Lane wrote:\\n> Why is this different from the half-dozen\\n> other fsync-error cases in the same file? Why, if fsync failure\\n> here is so catastrophic, is it okay to just return a normal failure\\n> code when tar_close doesn't even get to this point because of some\\n> earlier error? At the very least it seems like it'd be preferable\\n> to do the exit(1) at the caller level.\\n> \\n> The addition of the exit(1) seems to have been intentional in\\n> 1e2fddfa3, so cc'ing Peter for comment.\\n\\nThat commit addressed the behavior of fsync() failure in pg_receivewal \\nand pg_recvlogical, which are long-running daemon processes, so this \\nchange was analogous to the server-side changes at the time. I don't \\nknow what the behavior of fsync() failure in pg_basebackup should be, so \\ncalls that are only reachable from pg_basebackup are currently being \\nhandled differently.\\n\\n\\n\",\n \"msg_date\": \"Fri, 19 Nov 2021 11:10:46 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Deficient error handling in pg_dump and pg_basebackup\"\n }\n]","isConversational":false}}},{"rowIdx":53063,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI notice, belatedly, that 141fd1b tweaked the order of some members of\nstruct catctup, moving c_list and my_cache from earlier locations to the\nend of the struct. [0]\n\nThat resulted in this sequence at the end of the struct:\n\n...\n HeapTupleData tuple; /* tuple management header */\n struct catclist *c_list; /* containing CatCList, or NULL if none */\n CatCache *my_cache; /* link to owning catcache */\n /* properly aligned tuple data follows, unless a negative entry */\n}\n\n\nIt works, of course, with tuple->t_data pointed appropriately to the aligned\nspot after the struct. But does this add a new variant to the possible\nsituations described in the htup.h comments?\n\n * * Part of a palloc'd tuple: the HeapTupleData itself and the tuple\n * form a single palloc'd chunk. t_data points to the memory location\n * immediately following the HeapTupleData struct (at offset\nHEAPTUPLESIZE). ...\n *\n * * Separately allocated tuple: t_data points to a palloc'd chunk that\n * is not adjacent to the HeapTupleData. (This case is deprecated since\n * it's difficult to tell apart from case #1. ...\n\n\nAm I right in thinking that created a new case where the HeapTupleData\nand the tuple are not adjacent, but also not separately allocated?\nWould that perhaps warrant a mention in those comments in htup.h?\n\nRegards,\n-Chap\n\n\n[0]\nhttps://github.com/postgres/postgres/commit/141fd1b#diff-0f0061d745a9e6bcdf8b59ba304d4bae77c59ce400fb5ffd062cc36be077e8e7L116\n\n\n","msg_date":"Tue, 9 Nov 2021 16:49:00 -0500","msg_from":"Chapman Flack <chap@anastigmatix.net>","msg_from_op":true,"msg_subject":"Did this add to the \"ways in which [HeapTupleData] is used\"?"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI notice, belatedly, that 141fd1b tweaked the order of some members of\\nstruct catctup, moving c_list and my_cache from earlier locations to the\\nend of the struct. [0]\\n\\nThat resulted in this sequence at the end of the struct:\\n\\n...\\n HeapTupleData tuple; /* tuple management header */\\n struct catclist *c_list; /* containing CatCList, or NULL if none */\\n CatCache *my_cache; /* link to owning catcache */\\n /* properly aligned tuple data follows, unless a negative entry */\\n}\\n\\n\\nIt works, of course, with tuple->t_data pointed appropriately to the aligned\\nspot after the struct. But does this add a new variant to the possible\\nsituations described in the htup.h comments?\\n\\n * * Part of a palloc'd tuple: the HeapTupleData itself and the tuple\\n * form a single palloc'd chunk. t_data points to the memory location\\n * immediately following the HeapTupleData struct (at offset\\nHEAPTUPLESIZE). ...\\n *\\n * * Separately allocated tuple: t_data points to a palloc'd chunk that\\n * is not adjacent to the HeapTupleData. (This case is deprecated since\\n * it's difficult to tell apart from case #1. ...\\n\\n\\nAm I right in thinking that created a new case where the HeapTupleData\\nand the tuple are not adjacent, but also not separately allocated?\\nWould that perhaps warrant a mention in those comments in htup.h?\\n\\nRegards,\\n-Chap\\n\\n\\n[0]\\nhttps://github.com/postgres/postgres/commit/141fd1b#diff-0f0061d745a9e6bcdf8b59ba304d4bae77c59ce400fb5ffd062cc36be077e8e7L116\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Nov 2021 16:49:00 -0500\",\n \"msg_from\": \"Chapman Flack <chap@anastigmatix.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Did this add to the \\\"ways in which [HeapTupleData] is used\\\"?\"\n }\n]","isConversational":false}}},{"rowIdx":53064,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"ISTM that the recently-introduced new frontend logging support\n(common/logging.h et al) could use a second pass now that we have\nsome experience with it. There are two points that are bugging me:\n\n1. The distinction between \"error\" and \"fatal\" levels seems squishy\nto the point of uselessness. I think we should either get rid of it\nentirely, or make an effort to use \"fatal\" exactly for the cases that\nare going to give up and exit right away. Of the approximately 830\npg_log_error calls in HEAD, I count at least 450 that are immediately\nfollowed by exit(1), and so should be pg_log_fatal if this distinction\nmeans anything at all. OTOH, if we decide it doesn't mean anything,\nthere are only about 90 pg_log_fatal calls to convert. I lean\nslightly to the \"get rid of the distinction\" option, not only because\nit'd be a much smaller patch but also because I don't care to expend\nbrain cells on the which-to-use question while reviewing future\npatches.\n\n2. What is the preferred style for adding extra lines to log messages?\nI see a lot of direct prints to stderr:\n\n\t\tpg_log_error(\"missing required argument: database name\");\n\t\tfprintf(stderr, _(\"Try \\\"%s --help\\\" for more information.\\n\"), progname);\n\t\texit(1);\n\nbut a number of places have chosen to do this:\n\n\t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n\t\tpg_log_error(\"query was: %s\", todo);\n\nand some places got creative and did this:\n\n\t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n\t\tpg_log_info(\"query was: %s\", sql.data);\n\nI think this ought to be cleaned up so we have a more-or-less uniform\napproach. Aside from the randomly different source code, each of\nthese choices has different implications for whether the extra line\ngets printed or not depending on verbosity level, and that seems bad.\n\nOne plausible choice is to drop the first style (which surely has\nlittle to recommend it) and use either the second or third style\ndepending on whether you think the addendum ought to appear at the\nsame or higher verbosity level as the main message. But we'd\nstill be at hazard of people making randomly different choices\nabout that in identical cases, as indeed the second and third\nexamples show.\n\nAnother idea is to reduce such cases to one call:\n\n\t\tpg_log_error(\"query failed: %s\\nquery was: %s\",\n\t\t PQerrorMessage(conn), sql.data);\n\nbut I don't much like that: it knows more than it should about\nthe presentation format, and it can't support hiding the detail\nportion at lower verbosity levels.\n\nSo I'm not totally satisfied with these ideas, but I don't\nimmediately have a better one.\n\nThoughts?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 09 Nov 2021 17:20:41 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Frontend error logging style"},{"msg_contents":"At Tue, 09 Nov 2021 17:20:41 -0500, Tom Lane <tgl@sss.pgh.pa.us> wrote in \n> ISTM that the recently-introduced new frontend logging support\n> (common/logging.h et al) could use a second pass now that we have\n> some experience with it. There are two points that are bugging me:\n> \n> 1. The distinction between \"error\" and \"fatal\" levels seems squishy\n> to the point of uselessness. I think we should either get rid of it\n> entirely, or make an effort to use \"fatal\" exactly for the cases that\n> are going to give up and exit right away. Of the approximately 830\n> pg_log_error calls in HEAD, I count at least 450 that are immediately\n> followed by exit(1), and so should be pg_log_fatal if this distinction\n> means anything at all. OTOH, if we decide it doesn't mean anything,\n> there are only about 90 pg_log_fatal calls to convert. I lean\n> slightly to the \"get rid of the distinction\" option, not only because\n> it'd be a much smaller patch but also because I don't care to expend\n> brain cells on the which-to-use question while reviewing future\n> patches.\n\nAgreed. On backends, FATAL is bound to a behavior defferent from\nERROR. I doubt of the necessity of the difference between the two on\na console application (or frontend in PG-term). I would prefer to get\nrid of the FATAL level.\n\nI once thought ERROR should cause outoright end of an application, but\nthings don't seem so simple..\n\n> 2. What is the preferred style for adding extra lines to log messages?\n> I see a lot of direct prints to stderr:\n> \n> \t\tpg_log_error(\"missing required argument: database name\");\n> \t\tfprintf(stderr, _(\"Try \\\"%s --help\\\" for more information.\\n\"), progname);\n> \t\texit(1);\n> \n> but a number of places have chosen to do this:\n> \n> \t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n> \t\tpg_log_error(\"query was: %s\", todo);\n> \n> and some places got creative and did this:\n> \n> \t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n> \t\tpg_log_info(\"query was: %s\", sql.data);\n\nIn the first place the pg_log_info doesn't work at all since pgbench\ndoesn't make the log-level go below PG_LOG_INFO. (--quiet works the\ndifferent way..) In that case the pg_log_info can even be said to be\nused just to avoid the message from being prepended by a severity tag.\nI think we should fix that.\n\n> I think this ought to be cleaned up so we have a more-or-less uniform\n> approach. Aside from the randomly different source code, each of\n> these choices has different implications for whether the extra line\n> gets printed or not depending on verbosity level, and that seems bad.\n> \n> One plausible choice is to drop the first style (which surely has\n> little to recommend it) and use either the second or third style\n> depending on whether you think the addendum ought to appear at the\n> same or higher verbosity level as the main message. But we'd\n> still be at hazard of people making randomly different choices\n> about that in identical cases, as indeed the second and third\n> examples show.\n> \n> Another idea is to reduce such cases to one call:\n> \n> \t\tpg_log_error(\"query failed: %s\\nquery was: %s\",\n> \t\t PQerrorMessage(conn), sql.data);\n> \n> but I don't much like that: it knows more than it should about\n> the presentation format, and it can't support hiding the detail\n> portion at lower verbosity levels.\n> \n> So I'm not totally satisfied with these ideas, but I don't\n> immediately have a better one.\n> \n> Thoughts?\n\nHonestly I don't like that:p\n\nThe cause of the confusion is we are using verbosity to hide *a part\nof* a grouped messages. We make distinction beween the two concepts\nin the backend code but not in frontend. At least in pgbench, it\nseems to me as if the severity tag prepended by the pg_log_hoge\nfunctions is mere a label just to be shown on console.\n\nThat being said it is a kind of impssible to expect users to specify\nthe two independent levels on command line.\n\nCouldn't we have a function, say, named as \"pg_log_error_detail\",\nwhich prints the message when __pg_log_level is *above* ERROR and\nprepended by \"detail:\" tag? (or not showing the tag, keeping the\ncurrent behavior on surface.) We would have the same for\npg_log_warning. (pg_log_error_detail behaves a bit differetnly from\npg_log_info in this definition.)\n\n> \t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n> \t\tpg_log_error_detail(\"query was: %s\", sql.data);\n\nSo pg_log_generic(_v) have an additional parameter, say, detailed bool.\n\npg_log_generic(enum pg_log_level level, bool detailed, const char *pg_restrict fmt,...)\n\nConsidering that the level is used only to identify severity tag\nstring, it is somewhat strange that the detailed flag conflicts with\nit. But I'm not sure it is sane to add instead an extra-and-hidden log\nlevel to pg_log_level..\n\n\nOr as a simpler way, we could just have aliases for pg_log_info().\n\n#define pg_log_error_detail(...) pg_log_info(__VA_ARGS__)\n#define pg_log_warning_detail(...) pg_log_info(__VA_ARGS__)\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Wed, 10 Nov 2021 12:04:12 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On Tue, Nov 9, 2021 at 5:20 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> 1. The distinction between \"error\" and \"fatal\" levels seems squishy\n> to the point of uselessness.\n>\n> 2. What is the preferred style for adding extra lines to log messages?\n\nI agree with this list of problems. I think that the end game here is\ngetting to be able to use ereport() and friends in the frontend, which\nwould require confronting both of these issues at a deep level. We\ndon't necessarily have to do that now, though, but I think it's an\nargument against just nuking \"fatal\" from orbit. What I think we ought\nto be driving towards is having pg_log_fatal() forcibly exit, and\npg_log_error() do the same unless the error is somehow caught. That\nmight require more changes than you or whoever wants to do right now,\nso perhaps what we ought to do is just enforce the policy you\nsuggested before: if we're going to exit immediately afterward, it's\nfatal; if not, it's an error.\n\nI have been wondering for some time about trying to make the frontend\nand backend facilities symmetric and using case to distinguish. That\nis, if you see:\n\nERROR: this stinks\nDETAIL: It smells very bad.\nCONTEXT: garbage dump\n\n...well then that's a backend message. And if you see:\n\nerror: this stinks\ndetail: It smells very bad.\ncontext: garbage dump\n\n...well then that's a frontend message. I don't completely love that\nway of making a distinction, but I think we need something, and that's\npretty nearly the present practice at least for the primary message.\nWe don't really have a solid convention for detail/context/hint on the\nFE side; this is one idea.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 10 Nov 2021 10:28:35 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> I agree with this list of problems. I think that the end game here is\n> getting to be able to use ereport() and friends in the frontend, which\n> would require confronting both of these issues at a deep level. We\n> don't necessarily have to do that now, though, but I think it's an\n> argument against just nuking \"fatal\" from orbit. What I think we ought\n> to be driving towards is having pg_log_fatal() forcibly exit, and\n> pg_log_error() do the same unless the error is somehow caught.\n\nPerhaps. The usage that I'm concerned about is exemplified by this\ncommon pattern:\n\n\t\tpg_log_error(\"%s\", PQerrorMessage(conn));\n\t\tPQfinish(conn);\n\t\texit(1);\n\nie there's some cleanup you want to do between emitting the error and\nactually exiting. (I grant that maybe someday the PQfinish could be\ndone in an atexit handler or the like, but that's way more redesign\nthan I want to do right now.) In the case where you *don't* have\nany cleanup to do, though, it'd be nice to just write one line not\ntwo.\n\nThat leads me to think that we should redefine pg_log_fatal as\n\n#define pg_log_fatal(...) (pg_log_error(__VA_ARGS__), exit(1))\n\nWe still want to get rid of the distinct PG_LOG_FATAL log level,\nbecause whether you are using this form of pg_log_fatal or writing\nexit() separately is a coding detail that should not manifest\nas visibly distinct user output.\n\nThis proposal doesn't move us very far towards your endgame,\nbut I think it's a reasonable incremental step, and it makes\nthe difference between pg_log_error() and pg_log_fatal()\nclear and useful.\n\n> I have been wondering for some time about trying to make the frontend\n> and backend facilities symmetric and using case to distinguish. That\n> is, if you see:\n\n> ERROR: this stinks\n> DETAIL: It smells very bad.\n> CONTEXT: garbage dump\n\n> ...well then that's a backend message. And if you see:\n\n> error: this stinks\n> detail: It smells very bad.\n> context: garbage dump\n\n> ...well then that's a frontend message. I don't completely love that\n> way of making a distinction, but I think we need something, and that's\n> pretty nearly the present practice at least for the primary message.\n> We don't really have a solid convention for detail/context/hint on the\n> FE side; this is one idea.\n\nHmm, interesting. Taking up my point #2, I'd been thinking about\nproposing that we convert\n\n\t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n\t\tpg_log_error(\"query was: %s\", todo);\n\nto\n\n\t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n\t\tpg_log_error_detail(\"Query was: %s\", todo);\n\nand similarly add a pg_log_info_detail() companion for pg_log_info(), etc.\nWith your point above, the prefix these'd print would be \"detail:\" not\n\"DETAIL:\", but otherwise it seems to match up pretty well.\n\nAlso, if we are modeling this on backend conventions, we should say that\nstyle rules for pg_log_error_detail match those for errdetail(), with\ncomplete sentences and so forth. I wonder whether \"detail:\" followed\nby a capitalized sentence would look weird ... but you did it above,\nso maybe that's what people would expect.\n\nAgain, there's more that could be done later, but this seems to be\nenough to address the cases that people have been inventing ad-hoc\nsolutions for.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 10 Nov 2021 11:15:25 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"I wrote:\n> Hmm, interesting. Taking up my point #2, I'd been thinking about\n> proposing that we convert\n> \t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n> \t\tpg_log_error(\"query was: %s\", todo);\n> to\n> \t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n> \t\tpg_log_error_detail(\"Query was: %s\", todo);\n\nAfter looking around a bit, I see that a lot of these add-on messages\nare more nearly hints than details, so we'd probably better support\nboth those cases right off the bat.\n\nTo move things along a bit, here's a draft patch to logging.h/.c only\nto implement what I'm envisioning. I don't think there's much point\nin doing the per-call-site gruntwork until we have agreement on what\nthe API is, so this seems like enough for discussion.\n\n(As a fervent hater of colorization, I don't have an opinion about\nwhether or how to colorize the \"detail:\" and \"hint:\" fragments.\nBut I'll happily take somebody else's adjustment to add that.)\n\n\t\t\tregards, tom lane","msg_date":"Wed, 10 Nov 2021 14:25:08 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"At Wed, 10 Nov 2021 14:25:08 -0500, Tom Lane <tgl@sss.pgh.pa.us> wrote in \n> I wrote:\n> > Hmm, interesting. Taking up my point #2, I'd been thinking about\n> > proposing that we convert\n> > \t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n> > \t\tpg_log_error(\"query was: %s\", todo);\n> > to\n> > \t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n> > \t\tpg_log_error_detail(\"Query was: %s\", todo);\n> \n> After looking around a bit, I see that a lot of these add-on messages\n> are more nearly hints than details, so we'd probably better support\n> both those cases right off the bat.\n\nSounds reasonable.\n\n> To move things along a bit, here's a draft patch to logging.h/.c only\n> to implement what I'm envisioning. I don't think there's much point\n> in doing the per-call-site gruntwork until we have agreement on what\n> the API is, so this seems like enough for discussion.\n> \n> (As a fervent hater of colorization, I don't have an opinion about\n> whether or how to colorize the \"detail:\" and \"hint:\" fragments.\n> But I'll happily take somebody else's adjustment to add that.)\n\n(:) I don't hate colorization so much, but I'm frequently annoyed by\nneeding to turn off colorization and I disgust when there's no easy\nway to disable that in a simple steps..)\n\nAren't DETAIL and HINT expected to be hidden at the targetted cutoff\nlevel? In other words, I suspect that people want to hide non-primary\nmessages for a lower verbosity level. On the other hand I'm not sure\nit is a proper behavior that log_level = WARNING causes ERROR messages\nare accompanied by DETAIL/HINT submessages...\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Thu, 11 Nov 2021 11:07:55 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Kyotaro Horiguchi <horikyota.ntt@gmail.com> writes:\n> Aren't DETAIL and HINT expected to be hidden at the targetted cutoff\n> level? In other words, I suspect that people want to hide non-primary\n> messages for a lower verbosity level. On the other hand I'm not sure\n> it is a proper behavior that log_level = WARNING causes ERROR messages\n> are accompanied by DETAIL/HINT submessages...\n\nI abandoned that idea in the draft patch. We could maybe do something\nabout it further down the line, but I'm not sure there's really any\ndemand.\n\nAs the patch is set up, you could theoretically do something like\n\n\tpg_log_error(\"blah blah\");\n\tpg_log_info_detail(\"Very boring detail goes here.\");\n\n(note the intentionally different log priorities). But that feels wrong\nto me --- it doesn't seem like individual call sites should be setting\nsuch policy. If we do arrange for a way to hide the optional message\nparts, I'd rather that the control were centralized in logging.c.\n\nIt certainly wouldn't be hard for logging.c to make different decisions\nabout what to print; the thing that's not clear to me is what the\nuser-level knob for it should look like. We already used up the\noption of more or fewer -v switches.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 10 Nov 2021 21:20:28 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"\nOn 09.11.21 23:20, Tom Lane wrote:\n> 1. The distinction between \"error\" and \"fatal\" levels seems squishy\n> to the point of uselessness. I think we should either get rid of it\n> entirely, or make an effort to use \"fatal\" exactly for the cases that\n> are going to give up and exit right away. Of the approximately 830\n> pg_log_error calls in HEAD, I count at least 450 that are immediately\n> followed by exit(1), and so should be pg_log_fatal if this distinction\n> means anything at all. OTOH, if we decide it doesn't mean anything,\n> there are only about 90 pg_log_fatal calls to convert. I lean\n> slightly to the \"get rid of the distinction\" option, not only because\n> it'd be a much smaller patch but also because I don't care to expend\n> brain cells on the which-to-use question while reviewing future\n> patches.\n\nThis logging system has been designed more generally, drawing some \ninspiration from Python and Java libraries, for example. It's up to the \nprogram using this to make sensible use of it. I think there are \nprograms such as pg_receivewal where there is a meaningful distinction \nbetween errors flowing by and a fatal exit. But with something like \npg_basebackup, there is no real difference, so that code sort of uses \npg_log_error by default, since any error is implicitly fatal. I see the \napparent inconsistency, but I don't think it's a real problem. Each \nprogram by itself has arguably sensible behavior.\n\n\n","msg_date":"Mon, 15 Nov 2021 20:05:15 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On 09.11.21 23:20, Tom Lane wrote:\n> 2. What is the preferred style for adding extra lines to log messages?\n> I see a lot of direct prints to stderr:\n> \n> \t\tpg_log_error(\"missing required argument: database name\");\n> \t\tfprintf(stderr, _(\"Try \\\"%s --help\\\" for more information.\\n\"), progname);\n> \t\texit(1);\n\nThis is mainly used for those few messages that are sort of a \"global \nstandard\".\n\n> but a number of places have chosen to do this:\n> \n> \t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n> \t\tpg_log_error(\"query was: %s\", todo);\n> \n> and some places got creative and did this:\n> \n> \t\tpg_log_error(\"query failed: %s\", PQerrorMessage(conn));\n> \t\tpg_log_info(\"query was: %s\", sql.data);\n\nI can't decide between those two. It depends on how each program \nhandles the log level internally.\n\nSome kind of \"detail\" system could be useful. But it would need to be \ntied into the log level, layout, etc. I have not seen much inspiration \nfor this kind of thing in other logging libraries, so I didn't do \nanything about it yet.\n\n\n","msg_date":"Mon, 15 Nov 2021 20:08:19 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On 10.11.21 16:28, Robert Haas wrote:\n> What I think we ought\n> to be driving towards is having pg_log_fatal() forcibly exit, and\n> pg_log_error() do the same unless the error is somehow caught.\n\nThis is specifically designed not to do any flow control. In the \nbackend, we have many instances, where log messages are issued with the \nwrong log level because the stronger log level would have flow control \nimpact that is not appropriate at the call site. I don't think we want \nmore of that, especially since the flow control requirements in the \nvaried suite of frontend programs is quite diverse. Moreover, we also \nrequire control over the exit codes in some cases, which this kind of \nAPI wouldn't allow.\n\nSeveral programs wrap, say, pg_log_fatal() into a pg_fatal(), that does \nlogging, cleanup, and exit, as the case may be. I think that's a good \nsolution. If someone wanted to write a more widely reusable pg_fatal(), \nwhy not, but in my previous attempts, this was quite complicated and \ndidn't turn out to be useful.\n\n\n\n","msg_date":"Mon, 15 Nov 2021 20:15:04 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On Mon, Nov 15, 2021 at 2:15 PM Peter Eisentraut\n<peter.eisentraut@enterprisedb.com> wrote:\n> This is specifically designed not to do any flow control. In the\n> backend, we have many instances, where log messages are issued with the\n> wrong log level because the stronger log level would have flow control\n> impact that is not appropriate at the call site. I don't think we want\n> more of that, especially since the flow control requirements in the\n> varied suite of frontend programs is quite diverse. Moreover, we also\n> require control over the exit codes in some cases, which this kind of\n> API wouldn't allow.\n\nI agree that the system we use in the backend isn't problem free, but\nmaking the frontend do something randomly different isn't an\nimprovement. I think that most people who are writing PostgreSQL\nfrontend code have also written a lot of backend code, and they are\nused to the way things work in the backend. More importantly, there's\nan increasing amount of code that wants to run in either environment.\nAnd there have been suggestions that we want to make more things, like\nmemory contexts, work that way. The design decision that you've made\nhere makes that harder, and results in stuff like this:\n\n[rhaas pgsql]$ git grep pg_log_fatal.*VA_ARGS\nsrc/bin/pg_rewind/pg_rewind.h:#define pg_fatal(...) do {\npg_log_fatal(__VA_ARGS__); exit(1); } while(0)\nsrc/bin/pg_waldump/pg_waldump.c:#define fatal_error(...) do {\npg_log_fatal(__VA_ARGS__); exit(EXIT_FAILURE); } while(0)\nsrc/include/lib/simplehash.h: do { pg_log_fatal(__VA_ARGS__);\nexit(1); } while(0)\n\nHaving different frontend utilities each invent their own\nslightly-different way of doing this makes it hard to reuse code, and\nhard to understand code. We need to find ways to make it more uniform,\nnot just observe that it isn't uniform today and give up.\n\nIOW, I think the fact that it's not designed to do any flow control is\na bad thing.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 15 Nov 2021 14:40:10 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> Several programs wrap, say, pg_log_fatal() into a pg_fatal(), that does \n> logging, cleanup, and exit, as the case may be. I think that's a good \n> solution.\n\nI agree, and my draft patch formalized that by turning pg_log_fatal into\nexactly that.\n\nThe question that I think is relevant here is what is the point of\nlabeling errors as \"error:\" or \"fatal:\" if we're not going to make any\nserious attempt to make that distinction meaningful. I'm not really\nbuying your argument that it's fine as-is. Anybody who thinks that\nthere's a difference is going to be very confused by the behavior they\nobserve. But, if we all know there's no difference, why have the\ndifference?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 15 Nov 2021 14:45:11 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On Mon, Nov 15, 2021 at 02:40:10PM -0500, Robert Haas wrote:\n> Having different frontend utilities each invent their own\n> slightly-different way of doing this makes it hard to reuse code, and\n> hard to understand code. We need to find ways to make it more uniform,\n> not just observe that it isn't uniform today and give up.\n\nI agree with this sentiment, but this is a bit more complex than just\ncalling exit() with pg_log_fatal(), no? pg_dump likes playing a lot\nwith its exit_nicely(), meaning that we may want to allow frontends to\nplug in callbacks.\n--\nMichael","msg_date":"Tue, 16 Nov 2021 12:02:08 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On Mon, Nov 15, 2021 at 10:02 PM Michael Paquier <michael@paquier.xyz> wrote:\n> On Mon, Nov 15, 2021 at 02:40:10PM -0500, Robert Haas wrote:\n> > Having different frontend utilities each invent their own\n> > slightly-different way of doing this makes it hard to reuse code, and\n> > hard to understand code. We need to find ways to make it more uniform,\n> > not just observe that it isn't uniform today and give up.\n>\n> I agree with this sentiment, but this is a bit more complex than just\n> calling exit() with pg_log_fatal(), no? pg_dump likes playing a lot\n> with its exit_nicely(), meaning that we may want to allow frontends to\n> plug in callbacks.\n\nYep.\n\nI think we need frontend facilities that look like the backend\nfacilities, so try/catch blocks, on-exit callbacks, and whatever else\nthere is. Otherwise code reuse is going to continue to be annoying.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 16 Nov 2021 10:18:38 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On 16.11.21 16:18, Robert Haas wrote:\n> I think we need frontend facilities that look like the backend\n> facilities, so try/catch blocks, on-exit callbacks, and whatever else\n> there is. Otherwise code reuse is going to continue to be annoying.\n\nIf people want to do that kind of thing (I'm undecided whether the \ncomplexity is worth it), then make it a different API. The pg_log_* \ncalls are for writing formatted output. They normalized existing \nhand-coded patterns at the time. We can wrap another API on top of them \nthat does flow control and output. The pg_log_* stuff is more on the \nlevel of syslog(), which also just outputs stuff. Nobody is suggesting \nthat syslog(LOG_EMERG) should exit the program automatically. But you \ncan wrap higher-level APIs such as ereport() on top of that that might \ndo that.\n\n\n","msg_date":"Fri, 19 Nov 2021 11:17:50 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On Fri, Nov 19, 2021 at 5:17 AM Peter Eisentraut\n<peter.eisentraut@enterprisedb.com> wrote:\n> If people want to do that kind of thing (I'm undecided whether the\n> complexity is worth it), then make it a different API. The pg_log_*\n> calls are for writing formatted output. They normalized existing\n> hand-coded patterns at the time. We can wrap another API on top of them\n> that does flow control and output. The pg_log_* stuff is more on the\n> level of syslog(), which also just outputs stuff. Nobody is suggesting\n> that syslog(LOG_EMERG) should exit the program automatically. But you\n> can wrap higher-level APIs such as ereport() on top of that that might\n> do that.\n\nYeah, that might be a way forward.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 19 Nov 2021 10:18:55 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Fri, Nov 19, 2021 at 5:17 AM Peter Eisentraut\n> <peter.eisentraut@enterprisedb.com> wrote:\n>> If people want to do that kind of thing (I'm undecided whether the\n>> complexity is worth it), then make it a different API. The pg_log_*\n>> calls are for writing formatted output. They normalized existing\n>> hand-coded patterns at the time. We can wrap another API on top of them\n>> that does flow control and output. The pg_log_* stuff is more on the\n>> level of syslog(), which also just outputs stuff. Nobody is suggesting\n>> that syslog(LOG_EMERG) should exit the program automatically. But you\n>> can wrap higher-level APIs such as ereport() on top of that that might\n>> do that.\n\n> Yeah, that might be a way forward.\n\nThis conversation seems to have tailed off without full resolution,\nbut I observe that pretty much everyone except Peter is on board\nwith defining pg_log_fatal as pg_log_error + exit(1). So I think\nwe should just do that, unless Peter wants to produce a finished\nalternative proposal.\n\nI've now gone through and fleshed out the patch I sketched upthread.\nThis exercise convinced me that we absolutely should do something\nvery like this, because\n\n(1) As it stands, this patch removes a net of just about 1000 lines\nof code. So we clearly need *something*.\n\n(2) The savings would be even more, except that people have invented\nmacros for pg_log_error + exit(1) in at least four places already.\n(None of them quite the same of course.) Here I've just fixed those\nmacro definitions to use pg_log_fatal. In the name of consistency, we\nshould probably get rid of those macros in favor of using pg_log_fatal\ndirectly; but I've not done so here, since this patch is eyewateringly\nlong and boring already.\n\n(3) The amount of inconsistency in how we add on details/hints right\nnow is even worse than I thought. For example, we've got places\nburying the lede like this:\n\n- pg_log_error(\"server version: %s; %s version: %s\",\n- remoteversion_str, progname, PG_VERSION);\n- fatal(\"aborting because of server version mismatch\");\n+ pg_log_error(\"aborting because of server version mismatch\");\n+ pg_log_error_detail(\"server version: %s; %s version: %s\",\n+ remoteversion_str, progname, PG_VERSION);\n+ exit(1);\n\nor misidentifying the primary message altogether, like this:\n\n pg_log_error(\"query failed: %s\",\n PQerrorMessage(AH->connection));\n- fatal(\"query was: %s\", query);\n+ pg_log_error_detail(\"Query was: %s\", query);\n+ exit(1);\n\nThoughts?\n\n\t\t\tregards, tom lane","msg_date":"Tue, 22 Feb 2022 18:23:19 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Hi,\n\nOn 2022-02-22 18:23:19 -0500, Tom Lane wrote:\n> Robert Haas <robertmhaas@gmail.com> writes:\n> > On Fri, Nov 19, 2021 at 5:17 AM Peter Eisentraut\n> > <peter.eisentraut@enterprisedb.com> wrote:\n> >> If people want to do that kind of thing (I'm undecided whether the\n> >> complexity is worth it), then make it a different API. The pg_log_*\n> >> calls are for writing formatted output. They normalized existing\n> >> hand-coded patterns at the time. We can wrap another API on top of them\n> >> that does flow control and output. The pg_log_* stuff is more on the\n> >> level of syslog(), which also just outputs stuff. Nobody is suggesting\n> >> that syslog(LOG_EMERG) should exit the program automatically. But you\n> >> can wrap higher-level APIs such as ereport() on top of that that might\n> >> do that.\n\nThat'd maybe be a convincing argument in a project that didn't have\nelog(FATAL).\n\n\n> > Yeah, that might be a way forward.\n> \n> This conversation seems to have tailed off without full resolution,\n> but I observe that pretty much everyone except Peter is on board\n> with defining pg_log_fatal as pg_log_error + exit(1). So I think\n> we should just do that, unless Peter wants to produce a finished\n> alternative proposal.\n\nWhat about adding a pg_fatal() that's pg_log_fatal() + exit()? That keeps\npg_log_* stuff \"log only\", but adds something adjacent enough to hopefully\nreduce future misunderstandings?\n\nTo further decrease the chance of such mistakes, we could rename\npg_log_fatal() to something else. Or add a a pg_log_fatal() function that's\nmarked pg_nodiscard(), so that each callsite explicitly has to be (void)\npg_log_fatal().\n\n\n> I've now gone through and fleshed out the patch I sketched upthread.\n> This exercise convinced me that we absolutely should do something\n> very like this, because\n> \n> (1) As it stands, this patch removes a net of just about 1000 lines\n> of code. So we clearly need *something*.\n\n> (2) The savings would be even more, except that people have invented\n> macros for pg_log_error + exit(1) in at least four places already.\n> (None of them quite the same of course.) Here I've just fixed those\n> macro definitions to use pg_log_fatal. In the name of consistency, we\n> should probably get rid of those macros in favor of using pg_log_fatal\n> directly; but I've not done so here, since this patch is eyewateringly\n> long and boring already.\n\nAgreed.\n\n\nI don't care that much about the naming here.\n\n\nFor a moment I was wondering whether there'd be a benefit for having the \"pure\nlogging\" stuff be in a different static library than the erroring variant. So\nthat e.g. libpq's check for exit() doesn't run into trouble. But then I\nremembered that libpq doesn't link against common...\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 22 Feb 2022 19:11:37 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> What about adding a pg_fatal() that's pg_log_fatal() + exit()? That keeps\n> pg_log_* stuff \"log only\", but adds something adjacent enough to hopefully\n> reduce future misunderstandings?\n\nI'd be okay with that, except that pg_upgrade already has a pg_fatal\n(because it has its *own* logging system, just in case you thought\nthis wasn't enough of a mess yet). I'm in favor of aligning\npg_upgrade's logging with the rest, but I'd hoped to leave that for\nlater. Making the names collide would be bad even as a short-term\nthing, I fear.\n\nLooks like libpq_pipeline.c has its own pg_fatal, too.\n\nI'm not against choosing some name other than pg_log_fatal, but that\nparticular suggestion has got conflicts. Got any other ideas?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 22 Feb 2022 22:44:25 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On 2022-02-22 22:44:25 -0500, Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n> > What about adding a pg_fatal() that's pg_log_fatal() + exit()? That keeps\n> > pg_log_* stuff \"log only\", but adds something adjacent enough to hopefully\n> > reduce future misunderstandings?\n> \n> I'd be okay with that, except that pg_upgrade already has a pg_fatal\n> (because it has its *own* logging system, just in case you thought\n> this wasn't enough of a mess yet). I'm in favor of aligning\n> pg_upgrade's logging with the rest, but I'd hoped to leave that for\n> later. Making the names collide would be bad even as a short-term\n> thing, I fear.\n\nI guess we could name pg_upgrade's out of the way...\n\n\n> I'm not against choosing some name other than pg_log_fatal, but that\n> particular suggestion has got conflicts. Got any other ideas?\n\nMaybe pg_fatal_exit(), pg_exit_fatal() or pg_fatal_error()?\n\n\n","msg_date":"Tue, 22 Feb 2022 20:06:21 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"I wrote:\n> Andres Freund <andres@anarazel.de> writes:\n>> What about adding a pg_fatal() that's pg_log_fatal() + exit()? That keeps\n>> pg_log_* stuff \"log only\", but adds something adjacent enough to hopefully\n>> reduce future misunderstandings?\n\n> I'd be okay with that, except that pg_upgrade already has a pg_fatal\n> (because it has its *own* logging system, just in case you thought\n> this wasn't enough of a mess yet).\n\nWait a moment. After looking closer, I realize that pg_upgrade's\npg_fatal could trivially be turned into a macro; and the other two\nexisting definitions already are macros. That would remove the risk\nof link-time symbol collisions that I was worried about. As a bonus,\nit'd substantially reduce the number of changes needed to make\npg_upgrade use logging.c, whenever somebody wants to make that happen.\n\nSo I now propose modifying yesterday's patch thus:\n\n* Reinstantiate the PG_LOG_FATAL enum value, add support macros\npg_log_fatal, pg_log_fatal_hint, pg_log_fatal_detail.\n\n* Define pg_fatal as pg_log_fatal + exit(1). (This would essentially\nmove pg_rewind's definition into logging.h. pg_upgrade will\ndefine it slightly differently, but the semantics end up the same.)\n\n* Adjust call sites to match.\n\nI do like this idea because it would not break any existing code\nthat expects pg_log_fatal to return. There is likely to be some\nof that in outstanding patches, and this approach would merely\nrender it less-than-idiomatic rather than outright broken.\n\nUpdating the patch is going to be a bit tedious, so I'm not\ngoing to do it without buy-in that this solution would be\nokay to commit. Any objections?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 23 Feb 2022 10:30:02 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Hi,\n\nOn 2022-02-23 10:30:02 -0500, Tom Lane wrote:\n> So I now propose modifying yesterday's patch thus:\n> \n> * Reinstantiate the PG_LOG_FATAL enum value, add support macros\n> pg_log_fatal, pg_log_fatal_hint, pg_log_fatal_detail.\n> \n> * Define pg_fatal as pg_log_fatal + exit(1). (This would essentially\n> move pg_rewind's definition into logging.h. pg_upgrade will\n> define it slightly differently, but the semantics end up the same.)\n> \n> * Adjust call sites to match.\n\n+1\n\n\n> I do like this idea because it would not break any existing code\n> that expects pg_log_fatal to return. There is likely to be some\n> of that in outstanding patches, and this approach would merely\n> render it less-than-idiomatic rather than outright broken.\n> \n> Updating the patch is going to be a bit tedious, so I'm not\n> going to do it without buy-in that this solution would be\n> okay to commit. Any objections?\n\n+1\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 23 Feb 2022 08:47:20 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On Wed, Feb 23, 2022, at 12:30 PM, Tom Lane wrote:\n> Updating the patch is going to be a bit tedious, so I'm not\n> going to do it without buy-in that this solution would be\n> okay to commit. Any objections?\nNo. I was confused by the multiple log functions spread in the client programs\nwhile I was working on pg_subscriber [1]. Your proposal looks good to me.\n\n[1] https://postgr.es/m/5ac50071-f2ed-4ace-a8fd-b892cffd33eb%40www.fastmail.com\n\n\n--\nEuler Taveira\nEDB https://www.enterprisedb.com/\n\nOn Wed, Feb 23, 2022, at 12:30 PM, Tom Lane wrote:Updating the patch is going to be a bit tedious, so I'm notgoing to do it without buy-in that this solution would beokay to commit. Any objections?No. I was confused by the multiple log functions spread in the client programswhile I was working on pg_subscriber [1]. Your proposal looks good to me.[1] https://postgr.es/m/5ac50071-f2ed-4ace-a8fd-b892cffd33eb%40www.fastmail.com--Euler TaveiraEDB https://www.enterprisedb.com/","msg_date":"Wed, 23 Feb 2022 14:21:09 -0300","msg_from":"\"Euler Taveira\" <euler@eulerto.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On 23.02.22 00:23, Tom Lane wrote:\n> This conversation seems to have tailed off without full resolution,\n> but I observe that pretty much everyone except Peter is on board\n> with defining pg_log_fatal as pg_log_error + exit(1). So I think\n> we should just do that, unless Peter wants to produce a finished\n> alternative proposal.\n> \n> I've now gone through and fleshed out the patch I sketched upthread.\n\nThis patch already illustrates a couple of things that are wrong with \nthis approach:\n\n- It doesn't allow any other way of exiting. For example, in pg_dump, \nyou have removed a few exit_nicely() calls. It's not clear why that is \nvalid or whether it would always be valid for all call sites.\n\n- It doesn't allow other exit codes. For example, in psql, you have \nchanged a pg_log_fatal() call to pg_log_error() because it needed \nanother exit code. This slides us right back into that annoying \nsituation where in the backend we have to log error messages using \nelog(LOG) because the flow control is tangled up with the log level.\n\nMy suggestion is to just get rid of pg_log_fatal() and replace them all \nwith pg_log_error().\n\n\n\n","msg_date":"Thu, 24 Feb 2022 14:06:18 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Hi,\n\nOn 2022-02-24 14:06:18 +0100, Peter Eisentraut wrote:\n> My suggestion is to just get rid of pg_log_fatal() and replace them all with\n> pg_log_error().\n\n-1. This ignores that already several places came up with their slightly\ndifferent versions of fatal exit handlers. We don't gain anything by not\nstandardizing on one notion of a fatal error wrapper.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 24 Feb 2022 07:03:11 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> This patch already illustrates a couple of things that are wrong with \n> this approach:\n\nAll of these objections are a bit moot with my followup proposal, no?\n\n> - It doesn't allow any other way of exiting. For example, in pg_dump, \n> you have removed a few exit_nicely() calls. It's not clear why that is \n> valid or whether it would always be valid for all call sites.\n\nAs the patch stood, I'd hacked it so that pg_log_fatal called\nexit_nicely() not exit() within pg_dump. I agree that's not a great\nsolution, but I think that the correct fix is to get rid of exit_nicely\nin favor of doing the requisite cleanup in an atexit hook. pg_dump\nis already at great hazard of somebody calling exit() not exit_nicely(),\neither through failure to pay attention to that undocumented convention,\nor because some code it imports from someplace like src/common calls\nexit() directly. You need not look any further than pg_malloc()\nto see that there are live problems there.\n\nI figured this could be addressed in a separate patch, though,\nbecause to the extent that missing the exit-nicely callbacks is\na real bug, it's already a bug.\n\n> My suggestion is to just get rid of pg_log_fatal() and replace them all \n> with pg_log_error().\n\nI'm on board with dropping the separate FATAL log level if there's\nconsensus to do so; I think it adds more confusion than anything else.\nI still want to have something like pg_fatal(), though, because it's\nimpossible to deny the usefulness of such an abbreviation. It's already\nbeen reinvented three or four times independently.\n\nIndependently of either of those points, I still want to make the changes\nI proposed w.r.t. making explicit concepts of \"detail\" and \"hint\"\naddendums. What we have right now in the frontend code is an impossible\nmishmash of three or four ways of filling that lack, all inconsistent.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 24 Feb 2022 10:10:41 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"I wrote:\n> I'm on board with dropping the separate FATAL log level if there's\n> consensus to do so; I think it adds more confusion than anything else.\n\nI feel that the reasonable alternatives are either to drop the FATAL\nlog level, or try to make it actually mean something by consistently\nusing it for errors that are indeed fatal. I had a go at doing the\nlatter, but eventually concluded that that way madness lies. It's\nnot too hard to use \"pg_log_fatal\" when there's an exit(1) right\nafter it, but there are quite a lot of cases where a subroutine\nreports an error and returns a failure code to its caller, whereupon\nthe caller exits. Either the subroutine has to make an unwarranted\nassumption about what its callers will do, or we need to make an API\nchange to allow the subroutine itself to exit(), or we are going to\npresent a user experience that is inconsistently different depending\non internal implementation details.\n\nJust to add insult to injury, once I'd gotten done with the easy\npart (use pg_log_fatal where there's an adjacent exit()), a whole\nlot of TAP tests fell over, because they were expecting \"error:\"\nnot \"fatal:\". The reverse direction of s/fatal/error/g only breaks\none TAP test, which says a lot about how many places are troubling\nto make this distinction now.\n\nI conclude that we ought to drop the separate FATAL level and just\nuse ERROR instead.\n\nThe attached revision does that, standardizes on pg_fatal() as the\nabbreviation for pg_log_error() + exit(1), and invents detail/hint\nfeatures as per previous discussion.\n\n\t\t\tregards, tom lane","msg_date":"Fri, 25 Feb 2022 12:15:25 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On Fri, Feb 25, 2022 at 12:15:25PM -0500, Tom Lane wrote:\n> I feel that the reasonable alternatives are either to drop the FATAL\n> log level, or try to make it actually mean something by consistently\n> using it for errors that are indeed fatal. I had a go at doing the\n> latter, but eventually concluded that that way madness lies. It's\n> not too hard to use \"pg_log_fatal\" when there's an exit(1) right\n> after it, but there are quite a lot of cases where a subroutine\n> reports an error and returns a failure code to its caller, whereupon\n> the caller exits. Either the subroutine has to make an unwarranted\n> assumption about what its callers will do, or we need to make an API\n> change to allow the subroutine itself to exit(), or we are going to\n> present a user experience that is inconsistently different depending\n> on internal implementation details.\n\nNice code cut.\n\n> I conclude that we ought to drop the separate FATAL level and just\n> use ERROR instead.\n\nFWIW, I have found the uses of pg_log_error() and pg_log_fatal()\nrather confusing in some of the src/bin/ tools, so I am in favor of\nremoving completely one or the other, and getting rid of FATAL is the\nbest choice.\n\n> The attached revision does that, standardizes on pg_fatal() as the\n> abbreviation for pg_log_error() + exit(1), and invents detail/hint\n> features as per previous discussion.\n\n+#define pg_log_warning_hint(...) do { \\\n+ if (likely(__pg_log_level <= PG_LOG_WARNING)) \\\n+ pg_log_generic(PG_LOG_WARNING, PG_LOG_HINT, __VA_ARGS__);\n\\\n } while(0)\n\nHm. I am not sure to like much this abstraction by having so many\nmacros that understand the log level through their name. Wouldn't it\nbe cleaner to have only one pg_log_hint() and one pg_log_detail() with\nthe log level passed as argument of the macro?\n--\nMichael","msg_date":"Sat, 26 Feb 2022 14:21:01 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> Hm. I am not sure to like much this abstraction by having so many\n> macros that understand the log level through their name. Wouldn't it\n> be cleaner to have only one pg_log_hint() and one pg_log_detail() with\n> the log level passed as argument of the macro?\n\nMmm ... that doesn't sound better to me. I think it wouldn't be\nobvious that pg_log_warning and pg_log_hint are fundamentally\ndifferent sorts of things: in the first, \"warning\" refers to\nan error severity level, while in the second, \"hint\" refers to\na message component. I'm not wedded to the way I did it in this\npatch, but I think we ought to maintain a notational distinction\nbetween those two sorts of concepts.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 26 Feb 2022 00:55:34 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Hi,\n\nOn 2022-02-25 12:15:25 -0500, Tom Lane wrote:\n> The attached revision does that, standardizes on pg_fatal() as the\n> abbreviation for pg_log_error() + exit(1), and invents detail/hint\n> features as per previous discussion.\n\nThis unfortunately has had some bitrot: http://cfbot.cputube.org/patch_37_3574.log\n\nMarked as waiting-on-author.\n\nGreetings,\n\nAndres\n\n\n","msg_date":"Mon, 21 Mar 2022 18:51:35 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2022-02-25 12:15:25 -0500, Tom Lane wrote:\n>> The attached revision does that, standardizes on pg_fatal() as the\n>> abbreviation for pg_log_error() + exit(1), and invents detail/hint\n>> features as per previous discussion.\n\n> This unfortunately has had some bitrot: http://cfbot.cputube.org/patch_37_3574.log\n\nYeah, I know. That patch touches enough places that it's sure to\nbreak every few days during a commitfest. I'm not real excited about\nmaintaining it reactively. The flip side of the coin is that pushing\nit will doubtless break a lot of other uncommitted patches. So I'm\nnot sure how to proceed.\n\nMaybe the plan could be to push it at the end of the commitfest? But\nwhat I'd really like at the moment is buy-in or rejection of the whole\nconcept, so I know whether it's worth spending more time on.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 21 Mar 2022 22:00:37 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Hi,\n\nOn 2022-03-21 22:00:37 -0400, Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n> > This unfortunately has had some bitrot: http://cfbot.cputube.org/patch_37_3574.log\n> \n> Yeah, I know. That patch touches enough places that it's sure to\n> break every few days during a commitfest. I'm not real excited about\n> maintaining it reactively.\n\nMakes sense. I'd leave it on waiting-on-author then, so that reviewers looking\nthrough the CF don't need to look at it?\n\n\n> Maybe the plan could be to push it at the end of the commitfest?\n\nWould make sense to me.\n\nArguably parts of it could be committed sooner than that, e.g. exit(-1). But\nthat'd not make it meaningfully easier to maintain, so ...\n\n\n> But what I'd really like at the moment is buy-in or rejection of the whole\n> concept, so I know whether it's worth spending more time on.\n\nI'm +1. I've not carefully looked through every single changed callsite, but\nIMO it looks like a clear improvement.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 21 Mar 2022 19:13:48 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Here's a rebase up to today's HEAD. I've fixed the merge problems,\nbut there may be some stray new error calls that could be converted\nto use pg_fatal() and haven't been. I don't want to do a full\nfresh scan of the code until we're about ready to commit this.\n\n\t\t\tregards, tom lane","msg_date":"Sun, 27 Mar 2022 16:19:53 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"> On 27 Mar 2022, at 22:19, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Here's a rebase up to today's HEAD. I've fixed the merge problems,\n> but there may be some stray new error calls that could be converted\n> to use pg_fatal() and haven't been. I don't want to do a full\n> fresh scan of the code until we're about ready to commit this.\n\nI focused on finding any changes in logic (exiting where it previously kept\ngoing after error, or vice versa) and was unable to find any which inspires\nconfidence. This was a large patch to read through, so I'm fairly sure I've\nmissed something.\n\nBelow are comments (nitpicks to a large extent) which either relates to this\npatch or to the messages in general. For the latter, feel free to deem them\nout of scope and we can take those comments separately from this work on\n-hackers.\n\nAs a side-note realization from reading this, there is a tremendous amount of\ncode paths in pg_dump which leads to a fatal error.. thats not necessarily a\nbad thing, it's just striking when reading them all at once.\n\n> @@ -508,24 +502,15 @@ writefile(char *path, char **lines)\n\n> \tif (fclose(out_file))\n> -\t{\n> -\t\tpg_log_error(\"could not write file \\\"%s\\\": %m\", path);\n> -\t\texit(1);\n> -\t}\n> +\t\tpg_fatal(\"could not write file \\\"%s\\\": %m\", path);\n> }\n\nShould we update this message to differentiate it from the fwrite error case?\nSomething like \"an error occurred during writing..\"\n\n> @@ -2057,10 +2004,7 @@ check_locale_name(int category, const char *locale, char **canonname)\n> \n> \tsave = setlocale(category, NULL);\n> \tif (!save)\n> -\t{\n> -\t\tpg_log_error(\"setlocale() failed\");\n> -\t\texit(1);\n> -\t}\n> +\t\tpg_fatal(\"setlocale() failed\");\n\nShould this gain a hint message for those users who have no idea what this\nreally means?\n\n> @@ -2127,15 +2067,14 @@ check_locale_encoding(const char *locale, int user_enc)\n> \t\t user_enc == PG_SQL_ASCII))\n> \t{\n> \t\tpg_log_error(\"encoding mismatch\");\n> -\t\tfprintf(stderr,\n> -\t\t\t\t_(\"The encoding you selected (%s) and the encoding that the\\n\"\n> -\t\t\t\t \"selected locale uses (%s) do not match. This would lead to\\n\"\n> -\t\t\t\t \"misbehavior in various character string processing functions.\\n\"\n> -\t\t\t\t \"Rerun %s and either do not specify an encoding explicitly,\\n\"\n> -\t\t\t\t \"or choose a matching combination.\\n\"),\n> -\t\t\t\tpg_encoding_to_char(user_enc),\n> -\t\t\t\tpg_encoding_to_char(locale_enc),\n> -\t\t\t\tprogname);\n> +\t\tpg_log_error_detail(\"The encoding you selected (%s) and the encoding that the \"\n> +\t\t\t\t\t\t\t\"selected locale uses (%s) do not match. This would lead to \"\n\nI will be sad to see double space after period go, but I think I'm rare in\npreferring that.\n\n> @@ -2344,10 +2276,7 @@ setup_pgdata(void)\n> \t * have embedded spaces.\n> \t */\n> \tif (setenv(\"PGDATA\", pg_data, 1) != 0)\n> -\t{\n> -\t\tpg_log_error(\"could not set environment\");\n> -\t\texit(1);\n> -\t}\n> +\t\tpg_fatal(\"could not set environment\");\n\nShould we be explicit about which env var?\n\n> @@ -3089,18 +2979,14 @@ main(int argc, char *argv[])\n> \t\t\t\telse if (strcmp(optarg, \"libc\") == 0)\n> \t\t\t\t\tlocale_provider = COLLPROVIDER_LIBC;\n> \t\t\t\telse\n> -\t\t\t\t{\n> -\t\t\t\t\tpg_log_error(\"unrecognized locale provider: %s\", optarg);\n> -\t\t\t\t\texit(1);\n> -\t\t\t\t}\n> +\t\t\t\t\tpg_fatal(\"unrecognized locale provider: %s\", optarg);\n\nShould this %s be within quotes to match how we usually emit user-input?\n\n> @@ -1123,9 +1097,9 @@ verify_btree_slot_handler(PGresult *res, PGconn *conn, void *context)\n> \t\t\tpg_log_warning(\"btree index \\\"%s.%s.%s\\\": btree checking function returned unexpected number of rows: %d\",\n> \t\t\t\t\t\t rel->datinfo->datname, rel->nspname, rel->relname, ntups);\n> \t\t\tif (opts.verbose)\n> -\t\t\t\tpg_log_info(\"query was: %s\", rel->sql);\n> -\t\t\tpg_log_warning(\"Are %s's and amcheck's versions compatible?\",\n> -\t\t\t\t\t\t progname);\n> +\t\t\t\tpg_log_warning_detail(\"Query was: %s\", rel->sql);\n> +\t\t\tpg_log_warning_hint(\"Are %s's and amcheck's versions compatible?\",\n> +\t\t\t\t\t\t\t\tprogname);\n\nShould \"amcheck's\" be a %s parameter to make translation reusable (which it\nmiht never be) and possibly avoid translation mistake?\n\n> --- a/src/bin/pg_basebackup/receivelog.c\n> +++ b/src/bin/pg_basebackup/receivelog.c\n> @@ -140,7 +140,7 @@ open_walfile(StreamCtl *stream, XLogRecPtr startpoint)\n> \t\t\t/* fsync file in case of a previous crash */\n> \t\t\tif (stream->walmethod->sync(f) != 0)\n> \t\t\t{\n> -\t\t\t\tpg_log_fatal(\"could not fsync existing write-ahead log file \\\"%s\\\": %s\",\n> +\t\t\t\tpg_log_error(\"could not fsync existing write-ahead log file \\\"%s\\\": %s\",\n> \t\t\t\t\t\t\t fn, stream->walmethod->getlasterror());\n> \t\t\t\tstream->walmethod->close(f, CLOSE_UNLINK);\n> \t\t\t\texit(1);\n\nIn the case where we already have IO related errors, couldn't the close() call\ncause an additional error message which potentially could be helpful for\ndebugging?\n\n> @@ -597,31 +570,19 @@ main(int argc, char *argv[])\n\n> \tif (ControlFile->data_checksum_version == 0 &&\n> \t\tmode == PG_MODE_CHECK)\n> -\t{\n> -\t\tpg_log_error(\"data checksums are not enabled in cluster\");\n> -\t\texit(1);\n> -\t}\n> +\t\tpg_fatal(\"data checksums are not enabled in cluster\");\n> \n> \tif (ControlFile->data_checksum_version == 0 &&\n> \t\tmode == PG_MODE_DISABLE)\n> -\t{\n> -\t\tpg_log_error(\"data checksums are already disabled in cluster\");\n> -\t\texit(1);\n> -\t}\n> +\t\tpg_fatal(\"data checksums are already disabled in cluster\");\n> \n> \tif (ControlFile->data_checksum_version > 0 &&\n> \t\tmode == PG_MODE_ENABLE)\n> -\t{\n> -\t\tpg_log_error(\"data checksums are already enabled in cluster\");\n> -\t\texit(1);\n> -\t}\n> +\t\tpg_fatal(\"data checksums are already enabled in cluster\");\n\nFatal seems sort of out place here, it's not really a case of \"something\nterrible happened\" but rather \"the preconditions weren't met\". Couldn't these\nbe a single pg_log_error erroring out with the reason in a pg_log_detail?\n\n> @@ -1323,10 +1320,10 @@ lockTableForWorker(ArchiveHandle *AH, TocEntry *te)\n> \tres = PQexec(AH->connection, query->data);\n> \n> \tif (!res || PQresultStatus(res) != PGRES_COMMAND_OK)\n> -\t\tfatal(\"could not obtain lock on relation \\\"%s\\\"\\n\"\n> -\t\t\t \"This usually means that someone requested an ACCESS EXCLUSIVE lock \"\n> -\t\t\t \"on the table after the pg_dump parent process had gotten the \"\n> -\t\t\t \"initial ACCESS SHARE lock on the table.\", qualId);\n> +\t\tpg_fatal(\"could not obtain lock on relation \\\"%s\\\"\\n\"\n> +\t\t\t\t \"This usually means that someone requested an ACCESS EXCLUSIVE lock \"\n> +\t\t\t\t \"on the table after the pg_dump parent process had gotten the \"\n> +\t\t\t\t \"initial ACCESS SHARE lock on the table.\", qualId);\n\nPunctuation and formatting in a pg_fatal feels a bit wrong, and smells of need\nfor _detail/_hint. I guess there isn't a way to do that here though, but still\na shame.\n\n> @@ -50,9 +50,10 @@ _check_database_version(ArchiveHandle *AH)\n> \t\t&& (remoteversion < AH->public.minRemoteVersion ||\n> \t\t\tremoteversion > AH->public.maxRemoteVersion))\n> \t{\n> -\t\tpg_log_error(\"server version: %s; %s version: %s\",\n> -\t\t\t\t\t remoteversion_str, progname, PG_VERSION);\n> -\t\tfatal(\"aborting because of server version mismatch\");\n> +\t\tpg_log_error(\"aborting because of server version mismatch\");\n> +\t\tpg_log_error_detail(\"server version: %s; %s version: %s\",\n> +\t\t\t\t\t\t\tremoteversion_str, progname, PG_VERSION);\n> +\t\texit(1);\n\nCapitalization of the _detail message?\n\n> @@ -166,7 +167,7 @@ ConnectDatabase(Archive *AHX,\n> \t\tAH->connection = PQconnectdbParams(keywords, values, true);\n> \n> \t\tif (!AH->connection)\n> -\t\t\tfatal(\"could not connect to database\");\n> +\t\t\tpg_fatal(\"could not connect to database\");\n\nShouldn't this print which database it failed to connect to?\n\n\n> @@ -721,7 +721,7 @@ setFilePath(ArchiveHandle *AH, char *buf, const char *relativeFilename)\n> \tdname = ctx->directory;\n> \n> \tif (strlen(dname) + 1 + strlen(relativeFilename) + 1 > MAXPGPATH)\n\nUnrelated, but shouldn't that be >= MAXPGPATH?\n\n> @@ -2033,8 +2027,8 @@ dumpTableData_copy(Archive *fout, const void *dcontext)\n> \t{\n> \t\t/* copy data transfer failed */\n> \t\tpg_log_error(\"Dumping the contents of table \\\"%s\\\" failed: PQgetCopyData() failed.\", classname);\n> -\t\tpg_log_error(\"Error message from server: %s\", PQerrorMessage(conn));\n> -\t\tpg_log_error(\"The command was: %s\", q->data);\n> +\t\tpg_log_error_detail(\"Error message from server: %s\", PQerrorMessage(conn));\n> +\t\tpg_log_error_detail(\"Command was: %s\", q->data);\n\nShouldn't the pg_log_error() call be with leading lowercase and no punctuation?\n(same in the following hunk)\n\n\n> @@ -11756,8 +11747,8 @@ dumpFunc(Archive *fout, const FuncInfo *finfo)\n> \t\telse if (provolatile[0] == PROVOLATILE_STABLE)\n> \t\t\tappendPQExpBufferStr(q, \" STABLE\");\n> \t\telse if (provolatile[0] != PROVOLATILE_VOLATILE)\n> -\t\t\tfatal(\"unrecognized provolatile value for function \\\"%s\\\"\",\n> -\t\t\t\t finfo->dobj.name);\n> +\t\t\tpg_fatal(\"unrecognized provolatile value for function \\\"%s\\\"\",\n> +\t\t\t\t\t finfo->dobj.name);\n> \t}\n> \n> \tif (proisstrict[0] == 't')\n> @@ -11806,8 +11797,8 @@ dumpFunc(Archive *fout, const FuncInfo *finfo)\n> \t\telse if (proparallel[0] == PROPARALLEL_RESTRICTED)\n> \t\t\tappendPQExpBufferStr(q, \" PARALLEL RESTRICTED\");\n> \t\telse if (proparallel[0] != PROPARALLEL_UNSAFE)\n> -\t\t\tfatal(\"unrecognized proparallel value for function \\\"%s\\\"\",\n> -\t\t\t\t finfo->dobj.name);\n> +\t\t\tpg_fatal(\"unrecognized proparallel value for function \\\"%s\\\"\",\n> +\t\t\t\t\t finfo->dobj.name);\n\nWe should probably move the provolatile and propallel keywords to a %s param so\nwe can reuse the translation.\n\n> @@ -14951,18 +14942,18 @@ createViewAsClause(Archive *fout, const TableInfo *tbinfo)\n\n> -\t\tfatal(\"definition of view \\\"%s\\\" appears to be empty (length zero)\",\n> -\t\t\t tbinfo->dobj.name);\n> +\t\tpg_fatal(\"definition of view \\\"%s\\\" appears to be empty (length zero)\",\n> +\t\t\t\t tbinfo->dobj.name);\n\nI'm not sure we need to provide a definition of empty here, most readers will\nprobably understand that already =)\n\n> @@ -16602,13 +16593,10 @@ dumpSequence(Archive *fout, const TableInfo *tbinfo)\n> \tres = ExecuteSqlQuery(fout, query->data, PGRES_TUPLES_OK);\n> \n> \tif (PQntuples(res) != 1)\n> -\t{\n> -\t\tpg_log_error(ngettext(\"query to get data of sequence \\\"%s\\\" returned %d row (expected 1)\",\n> -\t\t\t\t\t\t\t \"query to get data of sequence \\\"%s\\\" returned %d rows (expected 1)\",\n> -\t\t\t\t\t\t\t PQntuples(res)),\n> -\t\t\t\t\t tbinfo->dobj.name, PQntuples(res));\n> -\t\texit_nicely(1);\n> -\t}\n> +\t\tpg_fatal(ngettext(\"query to get data of sequence \\\"%s\\\" returned %d row (expected 1)\",\n> +\t\t\t\t\t\t \"query to get data of sequence \\\"%s\\\" returned %d rows (expected 1)\",\n> +\t\t\t\t\t\t PQntuples(res)),\n> +\t\t\t\t tbinfo->dobj.name, PQntuples(res));\n\nThe ngettext() call seems a bit out of place here since we already know that\nsecond form will be taken given the check on PQntuples(res).\n\n> @@ -16824,13 +16812,10 @@ dumpSequenceData(Archive *fout, const TableDataInfo *tdinfo)\n> \tres = ExecuteSqlQuery(fout, query->data, PGRES_TUPLES_OK);\n> \n> \tif (PQntuples(res) != 1)\n> -\t{\n> -\t\tpg_log_error(ngettext(\"query to get data of sequence \\\"%s\\\" returned %d row (expected 1)\",\n> -\t\t\t\t\t\t\t \"query to get data of sequence \\\"%s\\\" returned %d rows (expected 1)\",\n> -\t\t\t\t\t\t\t PQntuples(res)),\n> -\t\t\t\t\t tbinfo->dobj.name, PQntuples(res));\n> -\t\texit_nicely(1);\n> -\t}\n> +\t\tpg_fatal(ngettext(\"query to get data of sequence \\\"%s\\\" returned %d row (expected 1)\",\n> +\t\t\t\t\t\t \"query to get data of sequence \\\"%s\\\" returned %d rows (expected 1)\",\n> +\t\t\t\t\t\t PQntuples(res)),\n> +\t\t\t\t tbinfo->dobj.name, PQntuples(res));\n\nSame as above, PQntuples(res) cannot be 1 so there os just one form used no?\n\n> --- a/src/bin/pg_dump/pg_dumpall.c\n> +++ b/src/bin/pg_dump/pg_dumpall.c\n> @@ -200,16 +200,15 @@ main(int argc, char *argv[])\n> \t\t\tstrlcpy(full_path, progname, sizeof(full_path));\n> \n> \t\tif (ret == -1)\n> -\t\t\tpg_log_error(\"The program \\\"%s\\\" is needed by %s but was not found in the\\n\"\n> -\t\t\t\t\t\t \"same directory as \\\"%s\\\".\\n\"\n> -\t\t\t\t\t\t \"Check your installation.\",\n> -\t\t\t\t\t\t \"pg_dump\", progname, full_path);\n> +\t\t\tpg_fatal(\"The program \\\"%s\\\" is needed by %s but was not found in the\\n\"\n> +\t\t\t\t\t \"same directory as \\\"%s\\\".\\n\"\n> +\t\t\t\t\t \"Check your installation.\",\n> +\t\t\t\t\t \"pg_dump\", progname, full_path);\n> \t\telse\n> -\t\t\tpg_log_error(\"The program \\\"%s\\\" was found by \\\"%s\\\"\\n\"\n> -\t\t\t\t\t\t \"but was not the same version as %s.\\n\"\n> -\t\t\t\t\t\t \"Check your installation.\",\n> -\t\t\t\t\t\t \"pg_dump\", full_path, progname);\n> -\t\texit_nicely(1);\n> +\t\t\tpg_fatal(\"The program \\\"%s\\\" was found by \\\"%s\\\"\\n\"\n> +\t\t\t\t\t \"but was not the same version as %s.\\n\"\n> +\t\t\t\t\t \"Check your installation.\",\n\n\"Check your installation\" seems superfluous here given the error. (we do this\nin a number of places)\n\n\n> @@ -1611,9 +1572,9 @@ connectDatabase(const char *dbname, const char *connection_string,\n> \t\t&& (server_version < 90200 ||\n> \t\t\t(server_version / 100) > (my_version / 100)))\n> \t{\n> -\t\tpg_log_error(\"server version: %s; %s version: %s\",\n> -\t\t\t\t\t remoteversion_str, progname, PG_VERSION);\n> \t\tpg_log_error(\"aborting because of server version mismatch\");\n> +\t\tpg_log_error_detail(\"server version: %s; %s version: %s\",\n> +\t\t\t\t\t\t\tremoteversion_str, progname, PG_VERSION);\n\nCapitalization?\n\n> @@ -388,13 +387,13 @@ WALDumpReadPage(XLogReaderState *state, XLogRecPtr targetPagePtr, int reqLen,\n> \t\tif (errinfo.wre_errno != 0)\n> \t\t{\n> \t\t\terrno = errinfo.wre_errno;\n> -\t\t\tfatal_error(\"could not read from file %s, offset %d: %m\",\n> -\t\t\t\t\t\tfname, errinfo.wre_off);\n> +\t\t\tpg_fatal(\"could not read from file %s, offset %d: %m\",\n> +\t\t\t\t\t fname, errinfo.wre_off);\n> \t\t}\n> \t\telse\n> -\t\t\tfatal_error(\"could not read from file %s, offset %d: read %d of %d\",\n> -\t\t\t\t\t\tfname, errinfo.wre_off, errinfo.wre_read,\n> -\t\t\t\t\t\terrinfo.wre_req);\n> +\t\t\tpg_fatal(\"could not read from file %s, offset %d: read %d of %d\",\n> +\t\t\t\t\t fname, errinfo.wre_off, errinfo.wre_read,\n> +\t\t\t\t\t errinfo.wre_req);\n\nFilename %s should be quoted like \\\"%s\\\".\n\n> @@ -7379,7 +7254,7 @@ main(int argc, char **argv)\n> \tTHREAD_BARRIER_DESTROY(&barrier);\n> \n> \tif (exit_code != 0)\n> -\t\tpg_log_fatal(\"Run was aborted; the above results are incomplete.\");\n> +\t\tpg_log_error(\"Run was aborted; the above results are incomplete.\");\n\nCapitalization and punctuation?\n\n> @@ -781,10 +772,7 @@ process_psqlrc(char *argv0)\n> \tchar\t *envrc = getenv(\"PSQLRC\");\n> \n> \tif (find_my_exec(argv0, my_exec_path) < 0)\n> -\t{\n> -\t\tpg_log_fatal(\"could not find own program executable\");\n> -\t\texit(EXIT_FAILURE);\n> -\t}\n> +\t\tpg_fatal(\"could not find own program executable\");\n\nI have a feeling this will be confusing to a fair few users. Not sure how to\nimprove it, but there seems to be some room for it.\n\n> @@ -144,16 +145,10 @@ main(int argc, char *argv[])\n> \tif (alldb)\n> \t{\n> \t\tif (dbname)\n> -\t\t{\n> -\t\t\tpg_log_error(\"cannot cluster all databases and a specific one at the same time\");\n> -\t\t\texit(1);\n> -\t\t}\n> +\t\t\tpg_fatal(\"cannot cluster all databases and a specific one at the same time\");\n> \n> \t\tif (tables.head != NULL)\n> -\t\t{\n> -\t\t\tpg_log_error(\"cannot cluster specific table(s) in all databases\");\n> -\t\t\texit(1);\n> -\t\t}\n> +\t\t\tpg_fatal(\"cannot cluster specific table(s) in all databases\");\n\nAn ngettext() candidate perhaps? There are more like this in main() hunks further down omitted for brevity here.\n\n> diff --git a/src/fe_utils/connect_utils.c b/src/fe_utils/connect_utils.c\n> index a30c66f13a..f2e583f9fa 100644\n> --- a/src/fe_utils/connect_utils.c\n> +++ b/src/fe_utils/connect_utils.c\n> @@ -88,11 +88,8 @@ connectDatabase(const ConnParams *cparams, const char *progname,\n> \t\tconn = PQconnectdbParams(keywords, values, true);\n> \n> \t\tif (!conn)\n> -\t\t{\n> -\t\t\tpg_log_error(\"could not connect to database %s: out of memory\",\n> -\t\t\t\t\t\t cparams->dbname);\n> -\t\t\texit(1);\n> -\t\t}\n> +\t\t\tpg_fatal(\"could not connect to database %s: out of memory\",\n> +\t\t\t\t\t cparams->dbname);\n\nDatabase name quoted like \\\"%s\\\" to match other messages?\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Tue, 29 Mar 2022 12:13:38 +0200","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On 27.03.22 22:19, Tom Lane wrote:\n> Here's a rebase up to today's HEAD. I've fixed the merge problems,\n> but there may be some stray new error calls that could be converted\n> to use pg_fatal() and haven't been. I don't want to do a full\n> fresh scan of the code until we're about ready to commit this.\n\nThis looks like a good improvement to me.\n\nI think I would want the program name/location also in front of the \ndetail and hint lines. I need to think about this a bit more. This \nshouldn't hold up this patch; it would be a quick localized change. \n(I'm also thinking about providing a separate color code for the \nsecondary messages. Again, this could be a quick follow-up patch.)\n\nThe one change I didn't like was\n\n- pg_log_error(\"The program \\\"%s\\\" is needed by %s \nbut was not found in the\\n\"\n- \"same directory as \n\\\"%s\\\".\\n\"\n- \"Check your installation.\",\n+ pg_log_error(\"the program \\\"%s\\\" is needed by %s \nbut was not found in the same directory as \\\"%s\\\"\",\n \"postgres\", progname, \nfull_path);\n\nThis appears to prioritize the guideline \"don't punctuate error message \nas full sentence\" over what should be the actual guideline \"don't make \nthe error message a full sentence\".\n\nThere are other occurrences of a similar message that were not changed \nin the same way by the patch. Maybe we should leave this one alone in \nthis patch and consider rewriting the message instead.\n\n\n","msg_date":"Tue, 29 Mar 2022 16:24:18 +0200","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On 29.03.22 12:13, Daniel Gustafsson wrote:\n\nMost of these should probably be addressed separately from Tom's patch.\n\n>> @@ -508,24 +502,15 @@ writefile(char *path, char **lines)\n> \n>> \tif (fclose(out_file))\n>> -\t{\n>> -\t\tpg_log_error(\"could not write file \\\"%s\\\": %m\", path);\n>> -\t\texit(1);\n>> -\t}\n>> +\t\tpg_fatal(\"could not write file \\\"%s\\\": %m\", path);\n>> }\n> \n> Should we update this message to differentiate it from the fwrite error case?\n> Something like \"an error occurred during writing..\"\n\nShould be \"could not close ...\", no?\n\n>> @@ -2057,10 +2004,7 @@ check_locale_name(int category, const char *locale, char **canonname)\n>>\n>> \tsave = setlocale(category, NULL);\n>> \tif (!save)\n>> -\t{\n>> -\t\tpg_log_error(\"setlocale() failed\");\n>> -\t\texit(1);\n>> -\t}\n>> +\t\tpg_fatal(\"setlocale() failed\");\n> \n> Should this gain a hint message for those users who have no idea what this\n> really means?\n\nMy setlocale() man page says:\n\nERRORS\n No errors are defined.\n\nSo uh ... ;-)\n\n>> @@ -3089,18 +2979,14 @@ main(int argc, char *argv[])\n>> \t\t\t\telse if (strcmp(optarg, \"libc\") == 0)\n>> \t\t\t\t\tlocale_provider = COLLPROVIDER_LIBC;\n>> \t\t\t\telse\n>> -\t\t\t\t{\n>> -\t\t\t\t\tpg_log_error(\"unrecognized locale provider: %s\", optarg);\n>> -\t\t\t\t\texit(1);\n>> -\t\t\t\t}\n>> +\t\t\t\t\tpg_fatal(\"unrecognized locale provider: %s\", optarg);\n> \n> Should this %s be within quotes to match how we usually emit user-input?\n\nUsually not done after colon, but could be.\n\n>> @@ -1123,9 +1097,9 @@ verify_btree_slot_handler(PGresult *res, PGconn *conn, void *context)\n>> \t\t\tpg_log_warning(\"btree index \\\"%s.%s.%s\\\": btree checking function returned unexpected number of rows: %d\",\n>> \t\t\t\t\t\t rel->datinfo->datname, rel->nspname, rel->relname, ntups);\n>> \t\t\tif (opts.verbose)\n>> -\t\t\t\tpg_log_info(\"query was: %s\", rel->sql);\n>> -\t\t\tpg_log_warning(\"Are %s's and amcheck's versions compatible?\",\n>> -\t\t\t\t\t\t progname);\n>> +\t\t\t\tpg_log_warning_detail(\"Query was: %s\", rel->sql);\n>> +\t\t\tpg_log_warning_hint(\"Are %s's and amcheck's versions compatible?\",\n>> +\t\t\t\t\t\t\t\tprogname);\n> \n> Should \"amcheck's\" be a %s parameter to make translation reusable (which it\n> miht never be) and possibly avoid translation mistake?\n\nWe don't have translations set up for contrib modules. Otherwise, this \nkind of thing would probably be something to look into.\n\n>> --- a/src/bin/pg_basebackup/receivelog.c\n>> +++ b/src/bin/pg_basebackup/receivelog.c\n>> @@ -140,7 +140,7 @@ open_walfile(StreamCtl *stream, XLogRecPtr startpoint)\n>> \t\t\t/* fsync file in case of a previous crash */\n>> \t\t\tif (stream->walmethod->sync(f) != 0)\n>> \t\t\t{\n>> -\t\t\t\tpg_log_fatal(\"could not fsync existing write-ahead log file \\\"%s\\\": %s\",\n>> +\t\t\t\tpg_log_error(\"could not fsync existing write-ahead log file \\\"%s\\\": %s\",\n>> \t\t\t\t\t\t\t fn, stream->walmethod->getlasterror());\n>> \t\t\t\tstream->walmethod->close(f, CLOSE_UNLINK);\n>> \t\t\t\texit(1);\n> \n> In the case where we already have IO related errors, couldn't the close() call\n> cause an additional error message which potentially could be helpful for\n> debugging?\n\nYeah, I think in general we have been sloppy with reporting file-closing \nerrors properly. Those presumably happen very rarely, but when they do, \nthings are probably very bad.\n\n>> @@ -597,31 +570,19 @@ main(int argc, char *argv[])\n> \n>> \tif (ControlFile->data_checksum_version == 0 &&\n>> \t\tmode == PG_MODE_CHECK)\n>> -\t{\n>> -\t\tpg_log_error(\"data checksums are not enabled in cluster\");\n>> -\t\texit(1);\n>> -\t}\n>> +\t\tpg_fatal(\"data checksums are not enabled in cluster\");\n\n> Fatal seems sort of out place here, it's not really a case of \"something\n> terrible happened\" but rather \"the preconditions weren't met\". Couldn't these\n> be a single pg_log_error erroring out with the reason in a pg_log_detail?\n\n\"fatal\" means error plus exit, so this seems ok. There is no separate \nlog level for really-bad-error.\n\n>> @@ -721,7 +721,7 @@ setFilePath(ArchiveHandle *AH, char *buf, const char *relativeFilename)\n>> \tdname = ctx->directory;\n>>\n>> \tif (strlen(dname) + 1 + strlen(relativeFilename) + 1 > MAXPGPATH)\n> \n> Unrelated, but shouldn't that be >= MAXPGPATH?\n\nSeems correct to me as is.\n\n>> @@ -14951,18 +14942,18 @@ createViewAsClause(Archive *fout, const TableInfo *tbinfo)\n> \n>> -\t\tfatal(\"definition of view \\\"%s\\\" appears to be empty (length zero)\",\n>> -\t\t\t tbinfo->dobj.name);\n>> +\t\tpg_fatal(\"definition of view \\\"%s\\\" appears to be empty (length zero)\",\n>> +\t\t\t\t tbinfo->dobj.name);\n> \n> I'm not sure we need to provide a definition of empty here, most readers will\n> probably understand that already =)\n\nIt could mean, contains no columns, or something similar. If I had to \nchange it, I would remove the \"empty\" and keep the \"length zero\".\n\n>> @@ -16602,13 +16593,10 @@ dumpSequence(Archive *fout, const TableInfo *tbinfo)\n>> \tres = ExecuteSqlQuery(fout, query->data, PGRES_TUPLES_OK);\n>>\n>> \tif (PQntuples(res) != 1)\n>> -\t{\n>> -\t\tpg_log_error(ngettext(\"query to get data of sequence \\\"%s\\\" returned %d row (expected 1)\",\n>> -\t\t\t\t\t\t\t \"query to get data of sequence \\\"%s\\\" returned %d rows (expected 1)\",\n>> -\t\t\t\t\t\t\t PQntuples(res)),\n>> -\t\t\t\t\t tbinfo->dobj.name, PQntuples(res));\n>> -\t\texit_nicely(1);\n>> -\t}\n>> +\t\tpg_fatal(ngettext(\"query to get data of sequence \\\"%s\\\" returned %d row (expected 1)\",\n>> +\t\t\t\t\t\t \"query to get data of sequence \\\"%s\\\" returned %d rows (expected 1)\",\n>> +\t\t\t\t\t\t PQntuples(res)),\n>> +\t\t\t\t tbinfo->dobj.name, PQntuples(res));\n> \n> The ngettext() call seems a bit out of place here since we already know that\n> second form will be taken given the check on PQntuples(res).\n\nSee \n<https://www.postgresql.org/message-id/flat/CALj2ACUfJKTmK5v%3DvF%2BH2iLkqM9Yvjsp6iXaCqAks6gDpzZh6g%40mail.gmail.com> \nfor a similar case that explains why this is still correct and necessary.\n\n>> @@ -144,16 +145,10 @@ main(int argc, char *argv[])\n>> \tif (alldb)\n>> \t{\n>> \t\tif (dbname)\n>> -\t\t{\n>> -\t\t\tpg_log_error(\"cannot cluster all databases and a specific one at the same time\");\n>> -\t\t\texit(1);\n>> -\t\t}\n>> +\t\t\tpg_fatal(\"cannot cluster all databases and a specific one at the same time\");\n>>\n>> \t\tif (tables.head != NULL)\n>> -\t\t{\n>> -\t\t\tpg_log_error(\"cannot cluster specific table(s) in all databases\");\n>> -\t\t\texit(1);\n>> -\t\t}\n>> +\t\t\tpg_fatal(\"cannot cluster specific table(s) in all databases\");\n> \n> An ngettext() candidate perhaps? There are more like this in main() hunks further down omitted for brevity here.\n\nI would just rephrase this as\n\n \"cannot cluster specific tables in all databases\"\n\nThis is still correct and sensible if the user specified just one table.\n\n\n","msg_date":"Tue, 29 Mar 2022 16:38:35 +0200","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"> On 29 Mar 2022, at 16:38, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\n> \n> On 29.03.22 12:13, Daniel Gustafsson wrote:\n> \n> Most of these should probably be addressed separately from Tom's patch.\n\nYeah, I think so too. I'll await input from Tom on how he wants to do, but I'm\nhappy to take these to a new thread. The main point of the review was to find\nlogic errors in the logging changes, these came as a bonus from reading them\nall in one place.\n\n>>> @@ -508,24 +502,15 @@ writefile(char *path, char **lines)\n>>> \tif (fclose(out_file))\n>>> -\t{\n>>> -\t\tpg_log_error(\"could not write file \\\"%s\\\": %m\", path);\n>>> -\t\texit(1);\n>>> -\t}\n>>> +\t\tpg_fatal(\"could not write file \\\"%s\\\": %m\", path);\n>>> }\n>> Should we update this message to differentiate it from the fwrite error case?\n>> Something like \"an error occurred during writing..\"\n> \n> Should be \"could not close ...\", no?\n\nYes, it should, not sure what I was thinking.\n\n>>> @@ -2057,10 +2004,7 @@ check_locale_name(int category, const char *locale, char **canonname)\n>>> \n>>> \tsave = setlocale(category, NULL);\n>>> \tif (!save)\n>>> -\t{\n>>> -\t\tpg_log_error(\"setlocale() failed\");\n>>> -\t\texit(1);\n>>> -\t}\n>>> +\t\tpg_fatal(\"setlocale() failed\");\n>> Should this gain a hint message for those users who have no idea what this\n>> really means?\n> \n> My setlocale() man page says:\n> \n> ERRORS\n> No errors are defined.\n> \n> So uh ... ;-)\n\nEven more reason to be confused then =) We have a mixed bag in the tree on how\nwe handle errors from setlocale, in this case we could reword it to say\nsomething like \"failed to retrieve locale for %s, category\" which IMO would be\nslightly more informative. Might be academic though since I guess it rarely, if\never, happens.\n\n>>> --- a/src/bin/pg_basebackup/receivelog.c\n>>> +++ b/src/bin/pg_basebackup/receivelog.c\n>>> @@ -140,7 +140,7 @@ open_walfile(StreamCtl *stream, XLogRecPtr startpoint)\n>>> \t\t\t/* fsync file in case of a previous crash */\n>>> \t\t\tif (stream->walmethod->sync(f) != 0)\n>>> \t\t\t{\n>>> -\t\t\t\tpg_log_fatal(\"could not fsync existing write-ahead log file \\\"%s\\\": %s\",\n>>> +\t\t\t\tpg_log_error(\"could not fsync existing write-ahead log file \\\"%s\\\": %s\",\n>>> \t\t\t\t\t\t\t fn, stream->walmethod->getlasterror());\n>>> \t\t\t\tstream->walmethod->close(f, CLOSE_UNLINK);\n>>> \t\t\t\texit(1);\n>> In the case where we already have IO related errors, couldn't the close() call\n>> cause an additional error message which potentially could be helpful for\n>> debugging?\n> \n> Yeah, I think in general we have been sloppy with reporting file-closing errors properly. Those presumably happen very rarely, but when they do, things are probably very bad.\n\nAgreed. I'm not sure if Tom wants to add net new loggings in this patchset,\nelse we could do this separately.\n\n>> The ngettext() call seems a bit out of place here since we already know that\n>> second form will be taken given the check on PQntuples(res).\n> \n> See <https://www.postgresql.org/message-id/flat/CALj2ACUfJKTmK5v%3DvF%2BH2iLkqM9Yvjsp6iXaCqAks6gDpzZh6g%40mail.gmail.com> for a similar case that explains why this is still correct and necessary.\n\nDoh, I knew that, sorry.\n\n>>> \t}\n>>> +\t\t\tpg_fatal(\"cannot cluster specific table(s) in all databases\");\n>> An ngettext() candidate perhaps? There are more like this in main() hunks further down omitted for brevity here.\n> \n> I would just rephrase this as\n> \n> \"cannot cluster specific tables in all databases\"\n> \n> This is still correct and sensible if the user specified just one table.\n\nThat's one way yes. Mostly I'm just a bit allergic to the \"foo(s)\" which my\nunscientifically based gut feeling am concerned about not necessarily always\nworking well for translations.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Wed, 30 Mar 2022 00:19:05 +0200","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Daniel Gustafsson <daniel@yesql.se> writes:\n> On 29 Mar 2022, at 16:38, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\n>> Most of these should probably be addressed separately from Tom's patch.\n\n> Yeah, I think so too.\n\nAgreed. I tried to confine my patch to mechanical changes, except\nfor changing places where the detail/hint features could be used.\n(I don't say that I yielded to temptation nowhere, because I don't\nrecall that for certain; but editing the message texts was not the\npoint of my patch.)\n\nFeel free to work on a followup editing patch though.\n\nAnother thing that occurred to me as I looked at your list is that\nI think a lot of these are nearly-can't-happen cases that we didn't\nput a lot of effort into devising great error messages for. Maybe\nwe should continue that approach, and in particular not put effort\ninto translating such messages. That would suggest inventing\n\"pg_fatal_internal\" and so forth, with the same functionality\nexcept for not being gettext triggers, comparable to\nerrmsg_internal. However, then somebody would have to make\njudgment calls about which messages not to bother translating.\nDo we want to go down that path?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 29 Mar 2022 18:38:30 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"FYI this patch no longer applies.\n\nGiven it's a Tom patch and due to its nature it'll bitrot rapidly\nanyways I'm don't see a point in updating the status though.\n\n\n","msg_date":"Wed, 30 Mar 2022 11:51:53 -0400","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Greg Stark <stark@mit.edu> writes:\n> FYI this patch no longer applies.\n\nNo surprise :-(\n\n> Given it's a Tom patch and due to its nature it'll bitrot rapidly\n> anyways I'm don't see a point in updating the status though.\n\nWe now seem to have buy-in on the concept, so my plan is to let\nthis sit till the end of the commitfest, then rebase and push.\nThat should avoid unnecessary churn for other pending patches.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 30 Mar 2022 12:18:39 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"> On 30 Mar 2022, at 00:38, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Daniel Gustafsson <daniel@yesql.se> writes:\n>> On 29 Mar 2022, at 16:38, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\n>>> Most of these should probably be addressed separately from Tom's patch.\n> \n>> Yeah, I think so too.\n> \n> Agreed. I tried to confine my patch to mechanical changes, except\n> for changing places where the detail/hint features could be used.\n> (I don't say that I yielded to temptation nowhere, because I don't\n> recall that for certain; but editing the message texts was not the\n> point of my patch.)\n> \n> Feel free to work on a followup editing patch though.\n\nThats my plan, once this lands I'll rebase the comments on top of your work and\nwe can have a separate discussion around them then.\n\n> Another thing that occurred to me as I looked at your list is that\n> I think a lot of these are nearly-can't-happen cases that we didn't\n> put a lot of effort into devising great error messages for. Maybe\n> we should continue that approach, and in particular not put effort\n> into translating such messages. That would suggest inventing\n> \"pg_fatal_internal\" and so forth, with the same functionality\n> except for not being gettext triggers, comparable to\n> errmsg_internal. However, then somebody would have to make\n> judgment calls about which messages not to bother translating.\n> Do we want to go down that path?\n\nI'm not sure. If, against the odds, these cases do happen then a user of a\nlocalized build probably really want to see the error in language they can\neasily understand and take in.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Tue, 5 Apr 2022 11:49:10 +0200","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Daniel Gustafsson <daniel@yesql.se> writes:\n> On 30 Mar 2022, at 00:38, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> Feel free to work on a followup editing patch though.\n\n> Thats my plan, once this lands I'll rebase the comments on top of your work and\n> we can have a separate discussion around them then.\n\nThe main patch is pushed now. I addressed the complaint Peter had\nabout the messages with \"Check your installation\" pseudo-hints\nby getting rid of them; I concur with your observation that those\nhints were basically useless. I also fixed the one place where the\nmessage should clearly be \"could not close\" not \"could not write\".\nMostly didn't yield to temptation anywhere else.\n\nOne other loose end is bothering me: I stuck with logging.h's\noriginal choice to put \"if (likely())\" or \"if (unlikely())\"\nconditionals into the macros, but I rather suspect that that's\njust a waste. I think we should put a centralized level check\ninto logging.c, and get rid of at least the \"if (likely())\"\nchecks, because those are going to succeed approximately 100.0%\nof the time. Maybe there's an argument for keeping the unlikely()\nones.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 08 Apr 2022 15:05:25 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"I wrote:\n> One other loose end is bothering me: I stuck with logging.h's\n> original choice to put \"if (likely())\" or \"if (unlikely())\"\n> conditionals into the macros, but I rather suspect that that's\n> just a waste. I think we should put a centralized level check\n> into logging.c, and get rid of at least the \"if (likely())\"\n> checks, because those are going to succeed approximately 100.0%\n> of the time. Maybe there's an argument for keeping the unlikely()\n> ones.\n\nConcretely, something like the attached. As a simple check,\nI looked at the compiled size of pg_dump. It went from\n\n text data bss dec hex filename\n 380298 4008 1384 385690 5e29a /home/postgres/testversion/bin/pg_dump\n\nto\n\n text data bss dec hex filename\n 374954 4008 1384 380346 5cdba src/bin/pg_dump/pg_dump\n\nfor a savings of about 5K or 1.5%. Not a huge amount, but\nnot nothing either, especially considering that the existing\ncoding isn't buying us anything.\n\n\t\t\tregards, tom lane","msg_date":"Fri, 08 Apr 2022 16:26:49 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On 08.04.22 22:26, Tom Lane wrote:\n> I wrote:\n>> One other loose end is bothering me: I stuck with logging.h's\n>> original choice to put \"if (likely())\" or \"if (unlikely())\"\n>> conditionals into the macros, but I rather suspect that that's\n>> just a waste. I think we should put a centralized level check\n>> into logging.c, and get rid of at least the \"if (likely())\"\n>> checks, because those are going to succeed approximately 100.0%\n>> of the time. Maybe there's an argument for keeping the unlikely()\n>> ones.\n> \n> Concretely, something like the attached. As a simple check,\n> I looked at the compiled size of pg_dump. It went from\n> \n> text data bss dec hex filename\n> 380298 4008 1384 385690 5e29a /home/postgres/testversion/bin/pg_dump\n> \n> to\n> \n> text data bss dec hex filename\n> 374954 4008 1384 380346 5cdba src/bin/pg_dump/pg_dump\n> \n> for a savings of about 5K or 1.5%. Not a huge amount, but\n> not nothing either, especially considering that the existing\n> coding isn't buying us anything.\n\nYeah, that seems ok to change. The previous coding style is more useful \nif you have a lot of debug messages in a hot code path, but that usually \ndoesn't apply to where this is used.\n\n\n","msg_date":"Mon, 11 Apr 2022 17:04:30 +0200","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> On 08.04.22 22:26, Tom Lane wrote:\n>>> I think we should put a centralized level check\n>>> into logging.c, and get rid of at least the \"if (likely())\"\n>>> checks, because those are going to succeed approximately 100.0%\n>>> of the time. Maybe there's an argument for keeping the unlikely()\n>>> ones.\n\n> Yeah, that seems ok to change. The previous coding style is more useful \n> if you have a lot of debug messages in a hot code path, but that usually \n> doesn't apply to where this is used.\n\nThe patch I presented keeps the unlikely() checks in the debug-level\nmacros. Do you think we should drop those too? I figured that avoiding\nevaluating the arguments would be worth something.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 11 Apr 2022 11:22:40 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"\nOn 11.04.22 17:22, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>> On 08.04.22 22:26, Tom Lane wrote:\n>>>> I think we should put a centralized level check\n>>>> into logging.c, and get rid of at least the \"if (likely())\"\n>>>> checks, because those are going to succeed approximately 100.0%\n>>>> of the time. Maybe there's an argument for keeping the unlikely()\n>>>> ones.\n> \n>> Yeah, that seems ok to change. The previous coding style is more useful\n>> if you have a lot of debug messages in a hot code path, but that usually\n>> doesn't apply to where this is used.\n> \n> The patch I presented keeps the unlikely() checks in the debug-level\n> macros. Do you think we should drop those too? I figured that avoiding\n> evaluating the arguments would be worth something.\n\nOh, that's right, the whole thing is to not evaluate the arguments if \nthe log level isn't adequate. We should probably keep that.\n\nIs the code size a big problem? ereport() has a bunch of extra code \naround each call as well. Does it have similar problems?\n\n\n","msg_date":"Tue, 12 Apr 2022 20:29:34 +0200","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> On 11.04.22 17:22, Tom Lane wrote:\n>> The patch I presented keeps the unlikely() checks in the debug-level\n>> macros. Do you think we should drop those too? I figured that avoiding\n>> evaluating the arguments would be worth something.\n\n> Oh, that's right, the whole thing is to not evaluate the arguments if \n> the log level isn't adequate. We should probably keep that.\n\nI don't think that's nearly as interesting in the frontend as in\nthe backend. Error messages in the backend frequently include\ncatalog lookups and the like in the arguments, but I think nearly\nall frontend messages are writing nothing more complicated than\nvariables and maybe some indirections or array fetches. So I'm\nnot all that worried about runtime, and would rather save some\ncode space.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 12 Apr 2022 14:56:13 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On 29.03.22 16:24, Peter Eisentraut wrote:\n> I think I would want the program name/location also in front of the \n> detail and hint lines. I need to think about this a bit more. This \n> shouldn't hold up this patch; it would be a quick localized change.\n\nAfter experiencing this for a bit now, I propose to make this change.\nIt lines up better and allows easier filtering of messages by program.\n\nExample:\n\nBefore:\n\ninitdb: error: locale \"zh_HK.Big5HKSCS\" requires unsupported encoding \"BIG5\"\ndetail: Encoding \"BIG5\" is not allowed as a server-side encoding.\nhint: Rerun initdb with a different locale selection.\n\nAfter:\n\ninitdb: error: locale \"zh_HK.Big5HKSCS\" requires unsupported encoding \"BIG5\"\ninitdb: detail: Encoding \"BIG5\" is not allowed as a server-side encoding.\ninitdb: hint: Rerun initdb with a different locale selection.","msg_date":"Mon, 23 May 2022 11:53:04 +0200","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"On 23.05.22 11:53, Peter Eisentraut wrote:\n> On 29.03.22 16:24, Peter Eisentraut wrote:\n>> I think I would want the program name/location also in front of the \n>> detail and hint lines. I need to think about this a bit more. This \n>> shouldn't hold up this patch; it would be a quick localized change.\n> \n> After experiencing this for a bit now, I propose to make this change.\n> It lines up better and allows easier filtering of messages by program.\n\nThis has been committed.\n\n\n","msg_date":"Mon, 30 May 2022 07:30:44 +0200","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Hi!\n\nCommit 9a374b77fb53 (Improve frontend error logging style.) replaced a \nline of file src/include/common/logging.h\n```\nextern PGDLLIMPORT enum pg_log_level __pg_log_level;\n```\nto\n```\nextern enum pg_log_level __pg_log_level;\n```\ni.e. it is rollback of commit 8ec569479fc28 (Apply PGDLLIMPORT markings \nbroadly.)\n\nIs it correct?\n\n-- \nWith best regards,\nDmitry Koval\n\nPostgres Professional: http://postgrespro.com\n\n\n","msg_date":"Tue, 11 Oct 2022 17:57:03 +0300","msg_from":"Dmitry Koval <d.koval@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: Frontend error logging style"},{"msg_contents":"Dmitry Koval <d.koval@postgrespro.ru> writes:\n> Hi!\n> Commit 9a374b77fb53 (Improve frontend error logging style.) replaced a \n> line of file src/include/common/logging.h\n> ```\n> extern PGDLLIMPORT enum pg_log_level __pg_log_level;\n> ```\n> to\n> ```\n> extern enum pg_log_level __pg_log_level;\n> ```\n> i.e. it is rollback of commit 8ec569479fc28 (Apply PGDLLIMPORT markings \n> broadly.)\n\n> Is it correct?\n\nYes, since that file is frontend-only.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 11 Oct 2022 11:08:47 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Frontend error logging style"}],"string":"[\n {\n \"msg_contents\": \"ISTM that the recently-introduced new frontend logging support\\n(common/logging.h et al) could use a second pass now that we have\\nsome experience with it. There are two points that are bugging me:\\n\\n1. The distinction between \\\"error\\\" and \\\"fatal\\\" levels seems squishy\\nto the point of uselessness. I think we should either get rid of it\\nentirely, or make an effort to use \\\"fatal\\\" exactly for the cases that\\nare going to give up and exit right away. Of the approximately 830\\npg_log_error calls in HEAD, I count at least 450 that are immediately\\nfollowed by exit(1), and so should be pg_log_fatal if this distinction\\nmeans anything at all. OTOH, if we decide it doesn't mean anything,\\nthere are only about 90 pg_log_fatal calls to convert. I lean\\nslightly to the \\\"get rid of the distinction\\\" option, not only because\\nit'd be a much smaller patch but also because I don't care to expend\\nbrain cells on the which-to-use question while reviewing future\\npatches.\\n\\n2. What is the preferred style for adding extra lines to log messages?\\nI see a lot of direct prints to stderr:\\n\\n\\t\\tpg_log_error(\\\"missing required argument: database name\\\");\\n\\t\\tfprintf(stderr, _(\\\"Try \\\\\\\"%s --help\\\\\\\" for more information.\\\\n\\\"), progname);\\n\\t\\texit(1);\\n\\nbut a number of places have chosen to do this:\\n\\n\\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n\\t\\tpg_log_error(\\\"query was: %s\\\", todo);\\n\\nand some places got creative and did this:\\n\\n\\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n\\t\\tpg_log_info(\\\"query was: %s\\\", sql.data);\\n\\nI think this ought to be cleaned up so we have a more-or-less uniform\\napproach. Aside from the randomly different source code, each of\\nthese choices has different implications for whether the extra line\\ngets printed or not depending on verbosity level, and that seems bad.\\n\\nOne plausible choice is to drop the first style (which surely has\\nlittle to recommend it) and use either the second or third style\\ndepending on whether you think the addendum ought to appear at the\\nsame or higher verbosity level as the main message. But we'd\\nstill be at hazard of people making randomly different choices\\nabout that in identical cases, as indeed the second and third\\nexamples show.\\n\\nAnother idea is to reduce such cases to one call:\\n\\n\\t\\tpg_log_error(\\\"query failed: %s\\\\nquery was: %s\\\",\\n\\t\\t PQerrorMessage(conn), sql.data);\\n\\nbut I don't much like that: it knows more than it should about\\nthe presentation format, and it can't support hiding the detail\\nportion at lower verbosity levels.\\n\\nSo I'm not totally satisfied with these ideas, but I don't\\nimmediately have a better one.\\n\\nThoughts?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 09 Nov 2021 17:20:41 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Frontend error logging style\"\n },\n {\n \"msg_contents\": \"At Tue, 09 Nov 2021 17:20:41 -0500, Tom Lane <tgl@sss.pgh.pa.us> wrote in \\n> ISTM that the recently-introduced new frontend logging support\\n> (common/logging.h et al) could use a second pass now that we have\\n> some experience with it. There are two points that are bugging me:\\n> \\n> 1. The distinction between \\\"error\\\" and \\\"fatal\\\" levels seems squishy\\n> to the point of uselessness. I think we should either get rid of it\\n> entirely, or make an effort to use \\\"fatal\\\" exactly for the cases that\\n> are going to give up and exit right away. Of the approximately 830\\n> pg_log_error calls in HEAD, I count at least 450 that are immediately\\n> followed by exit(1), and so should be pg_log_fatal if this distinction\\n> means anything at all. OTOH, if we decide it doesn't mean anything,\\n> there are only about 90 pg_log_fatal calls to convert. I lean\\n> slightly to the \\\"get rid of the distinction\\\" option, not only because\\n> it'd be a much smaller patch but also because I don't care to expend\\n> brain cells on the which-to-use question while reviewing future\\n> patches.\\n\\nAgreed. On backends, FATAL is bound to a behavior defferent from\\nERROR. I doubt of the necessity of the difference between the two on\\na console application (or frontend in PG-term). I would prefer to get\\nrid of the FATAL level.\\n\\nI once thought ERROR should cause outoright end of an application, but\\nthings don't seem so simple..\\n\\n> 2. What is the preferred style for adding extra lines to log messages?\\n> I see a lot of direct prints to stderr:\\n> \\n> \\t\\tpg_log_error(\\\"missing required argument: database name\\\");\\n> \\t\\tfprintf(stderr, _(\\\"Try \\\\\\\"%s --help\\\\\\\" for more information.\\\\n\\\"), progname);\\n> \\t\\texit(1);\\n> \\n> but a number of places have chosen to do this:\\n> \\n> \\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n> \\t\\tpg_log_error(\\\"query was: %s\\\", todo);\\n> \\n> and some places got creative and did this:\\n> \\n> \\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n> \\t\\tpg_log_info(\\\"query was: %s\\\", sql.data);\\n\\nIn the first place the pg_log_info doesn't work at all since pgbench\\ndoesn't make the log-level go below PG_LOG_INFO. (--quiet works the\\ndifferent way..) In that case the pg_log_info can even be said to be\\nused just to avoid the message from being prepended by a severity tag.\\nI think we should fix that.\\n\\n> I think this ought to be cleaned up so we have a more-or-less uniform\\n> approach. Aside from the randomly different source code, each of\\n> these choices has different implications for whether the extra line\\n> gets printed or not depending on verbosity level, and that seems bad.\\n> \\n> One plausible choice is to drop the first style (which surely has\\n> little to recommend it) and use either the second or third style\\n> depending on whether you think the addendum ought to appear at the\\n> same or higher verbosity level as the main message. But we'd\\n> still be at hazard of people making randomly different choices\\n> about that in identical cases, as indeed the second and third\\n> examples show.\\n> \\n> Another idea is to reduce such cases to one call:\\n> \\n> \\t\\tpg_log_error(\\\"query failed: %s\\\\nquery was: %s\\\",\\n> \\t\\t PQerrorMessage(conn), sql.data);\\n> \\n> but I don't much like that: it knows more than it should about\\n> the presentation format, and it can't support hiding the detail\\n> portion at lower verbosity levels.\\n> \\n> So I'm not totally satisfied with these ideas, but I don't\\n> immediately have a better one.\\n> \\n> Thoughts?\\n\\nHonestly I don't like that:p\\n\\nThe cause of the confusion is we are using verbosity to hide *a part\\nof* a grouped messages. We make distinction beween the two concepts\\nin the backend code but not in frontend. At least in pgbench, it\\nseems to me as if the severity tag prepended by the pg_log_hoge\\nfunctions is mere a label just to be shown on console.\\n\\nThat being said it is a kind of impssible to expect users to specify\\nthe two independent levels on command line.\\n\\nCouldn't we have a function, say, named as \\\"pg_log_error_detail\\\",\\nwhich prints the message when __pg_log_level is *above* ERROR and\\nprepended by \\\"detail:\\\" tag? (or not showing the tag, keeping the\\ncurrent behavior on surface.) We would have the same for\\npg_log_warning. (pg_log_error_detail behaves a bit differetnly from\\npg_log_info in this definition.)\\n\\n> \\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n> \\t\\tpg_log_error_detail(\\\"query was: %s\\\", sql.data);\\n\\nSo pg_log_generic(_v) have an additional parameter, say, detailed bool.\\n\\npg_log_generic(enum pg_log_level level, bool detailed, const char *pg_restrict fmt,...)\\n\\nConsidering that the level is used only to identify severity tag\\nstring, it is somewhat strange that the detailed flag conflicts with\\nit. But I'm not sure it is sane to add instead an extra-and-hidden log\\nlevel to pg_log_level..\\n\\n\\nOr as a simpler way, we could just have aliases for pg_log_info().\\n\\n#define pg_log_error_detail(...) pg_log_info(__VA_ARGS__)\\n#define pg_log_warning_detail(...) pg_log_info(__VA_ARGS__)\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 12:04:12 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 9, 2021 at 5:20 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> 1. The distinction between \\\"error\\\" and \\\"fatal\\\" levels seems squishy\\n> to the point of uselessness.\\n>\\n> 2. What is the preferred style for adding extra lines to log messages?\\n\\nI agree with this list of problems. I think that the end game here is\\ngetting to be able to use ereport() and friends in the frontend, which\\nwould require confronting both of these issues at a deep level. We\\ndon't necessarily have to do that now, though, but I think it's an\\nargument against just nuking \\\"fatal\\\" from orbit. What I think we ought\\nto be driving towards is having pg_log_fatal() forcibly exit, and\\npg_log_error() do the same unless the error is somehow caught. That\\nmight require more changes than you or whoever wants to do right now,\\nso perhaps what we ought to do is just enforce the policy you\\nsuggested before: if we're going to exit immediately afterward, it's\\nfatal; if not, it's an error.\\n\\nI have been wondering for some time about trying to make the frontend\\nand backend facilities symmetric and using case to distinguish. That\\nis, if you see:\\n\\nERROR: this stinks\\nDETAIL: It smells very bad.\\nCONTEXT: garbage dump\\n\\n...well then that's a backend message. And if you see:\\n\\nerror: this stinks\\ndetail: It smells very bad.\\ncontext: garbage dump\\n\\n...well then that's a frontend message. I don't completely love that\\nway of making a distinction, but I think we need something, and that's\\npretty nearly the present practice at least for the primary message.\\nWe don't really have a solid convention for detail/context/hint on the\\nFE side; this is one idea.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 10:28:35 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> I agree with this list of problems. I think that the end game here is\\n> getting to be able to use ereport() and friends in the frontend, which\\n> would require confronting both of these issues at a deep level. We\\n> don't necessarily have to do that now, though, but I think it's an\\n> argument against just nuking \\\"fatal\\\" from orbit. What I think we ought\\n> to be driving towards is having pg_log_fatal() forcibly exit, and\\n> pg_log_error() do the same unless the error is somehow caught.\\n\\nPerhaps. The usage that I'm concerned about is exemplified by this\\ncommon pattern:\\n\\n\\t\\tpg_log_error(\\\"%s\\\", PQerrorMessage(conn));\\n\\t\\tPQfinish(conn);\\n\\t\\texit(1);\\n\\nie there's some cleanup you want to do between emitting the error and\\nactually exiting. (I grant that maybe someday the PQfinish could be\\ndone in an atexit handler or the like, but that's way more redesign\\nthan I want to do right now.) In the case where you *don't* have\\nany cleanup to do, though, it'd be nice to just write one line not\\ntwo.\\n\\nThat leads me to think that we should redefine pg_log_fatal as\\n\\n#define pg_log_fatal(...) (pg_log_error(__VA_ARGS__), exit(1))\\n\\nWe still want to get rid of the distinct PG_LOG_FATAL log level,\\nbecause whether you are using this form of pg_log_fatal or writing\\nexit() separately is a coding detail that should not manifest\\nas visibly distinct user output.\\n\\nThis proposal doesn't move us very far towards your endgame,\\nbut I think it's a reasonable incremental step, and it makes\\nthe difference between pg_log_error() and pg_log_fatal()\\nclear and useful.\\n\\n> I have been wondering for some time about trying to make the frontend\\n> and backend facilities symmetric and using case to distinguish. That\\n> is, if you see:\\n\\n> ERROR: this stinks\\n> DETAIL: It smells very bad.\\n> CONTEXT: garbage dump\\n\\n> ...well then that's a backend message. And if you see:\\n\\n> error: this stinks\\n> detail: It smells very bad.\\n> context: garbage dump\\n\\n> ...well then that's a frontend message. I don't completely love that\\n> way of making a distinction, but I think we need something, and that's\\n> pretty nearly the present practice at least for the primary message.\\n> We don't really have a solid convention for detail/context/hint on the\\n> FE side; this is one idea.\\n\\nHmm, interesting. Taking up my point #2, I'd been thinking about\\nproposing that we convert\\n\\n\\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n\\t\\tpg_log_error(\\\"query was: %s\\\", todo);\\n\\nto\\n\\n\\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n\\t\\tpg_log_error_detail(\\\"Query was: %s\\\", todo);\\n\\nand similarly add a pg_log_info_detail() companion for pg_log_info(), etc.\\nWith your point above, the prefix these'd print would be \\\"detail:\\\" not\\n\\\"DETAIL:\\\", but otherwise it seems to match up pretty well.\\n\\nAlso, if we are modeling this on backend conventions, we should say that\\nstyle rules for pg_log_error_detail match those for errdetail(), with\\ncomplete sentences and so forth. I wonder whether \\\"detail:\\\" followed\\nby a capitalized sentence would look weird ... but you did it above,\\nso maybe that's what people would expect.\\n\\nAgain, there's more that could be done later, but this seems to be\\nenough to address the cases that people have been inventing ad-hoc\\nsolutions for.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 11:15:25 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Hmm, interesting. Taking up my point #2, I'd been thinking about\\n> proposing that we convert\\n> \\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n> \\t\\tpg_log_error(\\\"query was: %s\\\", todo);\\n> to\\n> \\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n> \\t\\tpg_log_error_detail(\\\"Query was: %s\\\", todo);\\n\\nAfter looking around a bit, I see that a lot of these add-on messages\\nare more nearly hints than details, so we'd probably better support\\nboth those cases right off the bat.\\n\\nTo move things along a bit, here's a draft patch to logging.h/.c only\\nto implement what I'm envisioning. I don't think there's much point\\nin doing the per-call-site gruntwork until we have agreement on what\\nthe API is, so this seems like enough for discussion.\\n\\n(As a fervent hater of colorization, I don't have an opinion about\\nwhether or how to colorize the \\\"detail:\\\" and \\\"hint:\\\" fragments.\\nBut I'll happily take somebody else's adjustment to add that.)\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Wed, 10 Nov 2021 14:25:08 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"At Wed, 10 Nov 2021 14:25:08 -0500, Tom Lane <tgl@sss.pgh.pa.us> wrote in \\n> I wrote:\\n> > Hmm, interesting. Taking up my point #2, I'd been thinking about\\n> > proposing that we convert\\n> > \\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n> > \\t\\tpg_log_error(\\\"query was: %s\\\", todo);\\n> > to\\n> > \\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n> > \\t\\tpg_log_error_detail(\\\"Query was: %s\\\", todo);\\n> \\n> After looking around a bit, I see that a lot of these add-on messages\\n> are more nearly hints than details, so we'd probably better support\\n> both those cases right off the bat.\\n\\nSounds reasonable.\\n\\n> To move things along a bit, here's a draft patch to logging.h/.c only\\n> to implement what I'm envisioning. I don't think there's much point\\n> in doing the per-call-site gruntwork until we have agreement on what\\n> the API is, so this seems like enough for discussion.\\n> \\n> (As a fervent hater of colorization, I don't have an opinion about\\n> whether or how to colorize the \\\"detail:\\\" and \\\"hint:\\\" fragments.\\n> But I'll happily take somebody else's adjustment to add that.)\\n\\n(:) I don't hate colorization so much, but I'm frequently annoyed by\\nneeding to turn off colorization and I disgust when there's no easy\\nway to disable that in a simple steps..)\\n\\nAren't DETAIL and HINT expected to be hidden at the targetted cutoff\\nlevel? In other words, I suspect that people want to hide non-primary\\nmessages for a lower verbosity level. On the other hand I'm not sure\\nit is a proper behavior that log_level = WARNING causes ERROR messages\\nare accompanied by DETAIL/HINT submessages...\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 11:07:55 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com> writes:\\n> Aren't DETAIL and HINT expected to be hidden at the targetted cutoff\\n> level? In other words, I suspect that people want to hide non-primary\\n> messages for a lower verbosity level. On the other hand I'm not sure\\n> it is a proper behavior that log_level = WARNING causes ERROR messages\\n> are accompanied by DETAIL/HINT submessages...\\n\\nI abandoned that idea in the draft patch. We could maybe do something\\nabout it further down the line, but I'm not sure there's really any\\ndemand.\\n\\nAs the patch is set up, you could theoretically do something like\\n\\n\\tpg_log_error(\\\"blah blah\\\");\\n\\tpg_log_info_detail(\\\"Very boring detail goes here.\\\");\\n\\n(note the intentionally different log priorities). But that feels wrong\\nto me --- it doesn't seem like individual call sites should be setting\\nsuch policy. If we do arrange for a way to hide the optional message\\nparts, I'd rather that the control were centralized in logging.c.\\n\\nIt certainly wouldn't be hard for logging.c to make different decisions\\nabout what to print; the thing that's not clear to me is what the\\nuser-level knob for it should look like. We already used up the\\noption of more or fewer -v switches.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 21:20:28 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"\\nOn 09.11.21 23:20, Tom Lane wrote:\\n> 1. The distinction between \\\"error\\\" and \\\"fatal\\\" levels seems squishy\\n> to the point of uselessness. I think we should either get rid of it\\n> entirely, or make an effort to use \\\"fatal\\\" exactly for the cases that\\n> are going to give up and exit right away. Of the approximately 830\\n> pg_log_error calls in HEAD, I count at least 450 that are immediately\\n> followed by exit(1), and so should be pg_log_fatal if this distinction\\n> means anything at all. OTOH, if we decide it doesn't mean anything,\\n> there are only about 90 pg_log_fatal calls to convert. I lean\\n> slightly to the \\\"get rid of the distinction\\\" option, not only because\\n> it'd be a much smaller patch but also because I don't care to expend\\n> brain cells on the which-to-use question while reviewing future\\n> patches.\\n\\nThis logging system has been designed more generally, drawing some \\ninspiration from Python and Java libraries, for example. It's up to the \\nprogram using this to make sensible use of it. I think there are \\nprograms such as pg_receivewal where there is a meaningful distinction \\nbetween errors flowing by and a fatal exit. But with something like \\npg_basebackup, there is no real difference, so that code sort of uses \\npg_log_error by default, since any error is implicitly fatal. I see the \\napparent inconsistency, but I don't think it's a real problem. Each \\nprogram by itself has arguably sensible behavior.\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 20:05:15 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On 09.11.21 23:20, Tom Lane wrote:\\n> 2. What is the preferred style for adding extra lines to log messages?\\n> I see a lot of direct prints to stderr:\\n> \\n> \\t\\tpg_log_error(\\\"missing required argument: database name\\\");\\n> \\t\\tfprintf(stderr, _(\\\"Try \\\\\\\"%s --help\\\\\\\" for more information.\\\\n\\\"), progname);\\n> \\t\\texit(1);\\n\\nThis is mainly used for those few messages that are sort of a \\\"global \\nstandard\\\".\\n\\n> but a number of places have chosen to do this:\\n> \\n> \\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n> \\t\\tpg_log_error(\\\"query was: %s\\\", todo);\\n> \\n> and some places got creative and did this:\\n> \\n> \\t\\tpg_log_error(\\\"query failed: %s\\\", PQerrorMessage(conn));\\n> \\t\\tpg_log_info(\\\"query was: %s\\\", sql.data);\\n\\nI can't decide between those two. It depends on how each program \\nhandles the log level internally.\\n\\nSome kind of \\\"detail\\\" system could be useful. But it would need to be \\ntied into the log level, layout, etc. I have not seen much inspiration \\nfor this kind of thing in other logging libraries, so I didn't do \\nanything about it yet.\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 20:08:19 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On 10.11.21 16:28, Robert Haas wrote:\\n> What I think we ought\\n> to be driving towards is having pg_log_fatal() forcibly exit, and\\n> pg_log_error() do the same unless the error is somehow caught.\\n\\nThis is specifically designed not to do any flow control. In the \\nbackend, we have many instances, where log messages are issued with the \\nwrong log level because the stronger log level would have flow control \\nimpact that is not appropriate at the call site. I don't think we want \\nmore of that, especially since the flow control requirements in the \\nvaried suite of frontend programs is quite diverse. Moreover, we also \\nrequire control over the exit codes in some cases, which this kind of \\nAPI wouldn't allow.\\n\\nSeveral programs wrap, say, pg_log_fatal() into a pg_fatal(), that does \\nlogging, cleanup, and exit, as the case may be. I think that's a good \\nsolution. If someone wanted to write a more widely reusable pg_fatal(), \\nwhy not, but in my previous attempts, this was quite complicated and \\ndidn't turn out to be useful.\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 20:15:04 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 2:15 PM Peter Eisentraut\\n<peter.eisentraut@enterprisedb.com> wrote:\\n> This is specifically designed not to do any flow control. In the\\n> backend, we have many instances, where log messages are issued with the\\n> wrong log level because the stronger log level would have flow control\\n> impact that is not appropriate at the call site. I don't think we want\\n> more of that, especially since the flow control requirements in the\\n> varied suite of frontend programs is quite diverse. Moreover, we also\\n> require control over the exit codes in some cases, which this kind of\\n> API wouldn't allow.\\n\\nI agree that the system we use in the backend isn't problem free, but\\nmaking the frontend do something randomly different isn't an\\nimprovement. I think that most people who are writing PostgreSQL\\nfrontend code have also written a lot of backend code, and they are\\nused to the way things work in the backend. More importantly, there's\\nan increasing amount of code that wants to run in either environment.\\nAnd there have been suggestions that we want to make more things, like\\nmemory contexts, work that way. The design decision that you've made\\nhere makes that harder, and results in stuff like this:\\n\\n[rhaas pgsql]$ git grep pg_log_fatal.*VA_ARGS\\nsrc/bin/pg_rewind/pg_rewind.h:#define pg_fatal(...) do {\\npg_log_fatal(__VA_ARGS__); exit(1); } while(0)\\nsrc/bin/pg_waldump/pg_waldump.c:#define fatal_error(...) do {\\npg_log_fatal(__VA_ARGS__); exit(EXIT_FAILURE); } while(0)\\nsrc/include/lib/simplehash.h: do { pg_log_fatal(__VA_ARGS__);\\nexit(1); } while(0)\\n\\nHaving different frontend utilities each invent their own\\nslightly-different way of doing this makes it hard to reuse code, and\\nhard to understand code. We need to find ways to make it more uniform,\\nnot just observe that it isn't uniform today and give up.\\n\\nIOW, I think the fact that it's not designed to do any flow control is\\na bad thing.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 14:40:10 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> Several programs wrap, say, pg_log_fatal() into a pg_fatal(), that does \\n> logging, cleanup, and exit, as the case may be. I think that's a good \\n> solution.\\n\\nI agree, and my draft patch formalized that by turning pg_log_fatal into\\nexactly that.\\n\\nThe question that I think is relevant here is what is the point of\\nlabeling errors as \\\"error:\\\" or \\\"fatal:\\\" if we're not going to make any\\nserious attempt to make that distinction meaningful. I'm not really\\nbuying your argument that it's fine as-is. Anybody who thinks that\\nthere's a difference is going to be very confused by the behavior they\\nobserve. But, if we all know there's no difference, why have the\\ndifference?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 14:45:11 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 02:40:10PM -0500, Robert Haas wrote:\\n> Having different frontend utilities each invent their own\\n> slightly-different way of doing this makes it hard to reuse code, and\\n> hard to understand code. We need to find ways to make it more uniform,\\n> not just observe that it isn't uniform today and give up.\\n\\nI agree with this sentiment, but this is a bit more complex than just\\ncalling exit() with pg_log_fatal(), no? pg_dump likes playing a lot\\nwith its exit_nicely(), meaning that we may want to allow frontends to\\nplug in callbacks.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 16 Nov 2021 12:02:08 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 10:02 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> On Mon, Nov 15, 2021 at 02:40:10PM -0500, Robert Haas wrote:\\n> > Having different frontend utilities each invent their own\\n> > slightly-different way of doing this makes it hard to reuse code, and\\n> > hard to understand code. We need to find ways to make it more uniform,\\n> > not just observe that it isn't uniform today and give up.\\n>\\n> I agree with this sentiment, but this is a bit more complex than just\\n> calling exit() with pg_log_fatal(), no? pg_dump likes playing a lot\\n> with its exit_nicely(), meaning that we may want to allow frontends to\\n> plug in callbacks.\\n\\nYep.\\n\\nI think we need frontend facilities that look like the backend\\nfacilities, so try/catch blocks, on-exit callbacks, and whatever else\\nthere is. Otherwise code reuse is going to continue to be annoying.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 10:18:38 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On 16.11.21 16:18, Robert Haas wrote:\\n> I think we need frontend facilities that look like the backend\\n> facilities, so try/catch blocks, on-exit callbacks, and whatever else\\n> there is. Otherwise code reuse is going to continue to be annoying.\\n\\nIf people want to do that kind of thing (I'm undecided whether the \\ncomplexity is worth it), then make it a different API. The pg_log_* \\ncalls are for writing formatted output. They normalized existing \\nhand-coded patterns at the time. We can wrap another API on top of them \\nthat does flow control and output. The pg_log_* stuff is more on the \\nlevel of syslog(), which also just outputs stuff. Nobody is suggesting \\nthat syslog(LOG_EMERG) should exit the program automatically. But you \\ncan wrap higher-level APIs such as ereport() on top of that that might \\ndo that.\\n\\n\\n\",\n \"msg_date\": \"Fri, 19 Nov 2021 11:17:50 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 19, 2021 at 5:17 AM Peter Eisentraut\\n<peter.eisentraut@enterprisedb.com> wrote:\\n> If people want to do that kind of thing (I'm undecided whether the\\n> complexity is worth it), then make it a different API. The pg_log_*\\n> calls are for writing formatted output. They normalized existing\\n> hand-coded patterns at the time. We can wrap another API on top of them\\n> that does flow control and output. The pg_log_* stuff is more on the\\n> level of syslog(), which also just outputs stuff. Nobody is suggesting\\n> that syslog(LOG_EMERG) should exit the program automatically. But you\\n> can wrap higher-level APIs such as ereport() on top of that that might\\n> do that.\\n\\nYeah, that might be a way forward.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 19 Nov 2021 10:18:55 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Fri, Nov 19, 2021 at 5:17 AM Peter Eisentraut\\n> <peter.eisentraut@enterprisedb.com> wrote:\\n>> If people want to do that kind of thing (I'm undecided whether the\\n>> complexity is worth it), then make it a different API. The pg_log_*\\n>> calls are for writing formatted output. They normalized existing\\n>> hand-coded patterns at the time. We can wrap another API on top of them\\n>> that does flow control and output. The pg_log_* stuff is more on the\\n>> level of syslog(), which also just outputs stuff. Nobody is suggesting\\n>> that syslog(LOG_EMERG) should exit the program automatically. But you\\n>> can wrap higher-level APIs such as ereport() on top of that that might\\n>> do that.\\n\\n> Yeah, that might be a way forward.\\n\\nThis conversation seems to have tailed off without full resolution,\\nbut I observe that pretty much everyone except Peter is on board\\nwith defining pg_log_fatal as pg_log_error + exit(1). So I think\\nwe should just do that, unless Peter wants to produce a finished\\nalternative proposal.\\n\\nI've now gone through and fleshed out the patch I sketched upthread.\\nThis exercise convinced me that we absolutely should do something\\nvery like this, because\\n\\n(1) As it stands, this patch removes a net of just about 1000 lines\\nof code. So we clearly need *something*.\\n\\n(2) The savings would be even more, except that people have invented\\nmacros for pg_log_error + exit(1) in at least four places already.\\n(None of them quite the same of course.) Here I've just fixed those\\nmacro definitions to use pg_log_fatal. In the name of consistency, we\\nshould probably get rid of those macros in favor of using pg_log_fatal\\ndirectly; but I've not done so here, since this patch is eyewateringly\\nlong and boring already.\\n\\n(3) The amount of inconsistency in how we add on details/hints right\\nnow is even worse than I thought. For example, we've got places\\nburying the lede like this:\\n\\n- pg_log_error(\\\"server version: %s; %s version: %s\\\",\\n- remoteversion_str, progname, PG_VERSION);\\n- fatal(\\\"aborting because of server version mismatch\\\");\\n+ pg_log_error(\\\"aborting because of server version mismatch\\\");\\n+ pg_log_error_detail(\\\"server version: %s; %s version: %s\\\",\\n+ remoteversion_str, progname, PG_VERSION);\\n+ exit(1);\\n\\nor misidentifying the primary message altogether, like this:\\n\\n pg_log_error(\\\"query failed: %s\\\",\\n PQerrorMessage(AH->connection));\\n- fatal(\\\"query was: %s\\\", query);\\n+ pg_log_error_detail(\\\"Query was: %s\\\", query);\\n+ exit(1);\\n\\nThoughts?\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Tue, 22 Feb 2022 18:23:19 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-22 18:23:19 -0500, Tom Lane wrote:\\n> Robert Haas <robertmhaas@gmail.com> writes:\\n> > On Fri, Nov 19, 2021 at 5:17 AM Peter Eisentraut\\n> > <peter.eisentraut@enterprisedb.com> wrote:\\n> >> If people want to do that kind of thing (I'm undecided whether the\\n> >> complexity is worth it), then make it a different API. The pg_log_*\\n> >> calls are for writing formatted output. They normalized existing\\n> >> hand-coded patterns at the time. We can wrap another API on top of them\\n> >> that does flow control and output. The pg_log_* stuff is more on the\\n> >> level of syslog(), which also just outputs stuff. Nobody is suggesting\\n> >> that syslog(LOG_EMERG) should exit the program automatically. But you\\n> >> can wrap higher-level APIs such as ereport() on top of that that might\\n> >> do that.\\n\\nThat'd maybe be a convincing argument in a project that didn't have\\nelog(FATAL).\\n\\n\\n> > Yeah, that might be a way forward.\\n> \\n> This conversation seems to have tailed off without full resolution,\\n> but I observe that pretty much everyone except Peter is on board\\n> with defining pg_log_fatal as pg_log_error + exit(1). So I think\\n> we should just do that, unless Peter wants to produce a finished\\n> alternative proposal.\\n\\nWhat about adding a pg_fatal() that's pg_log_fatal() + exit()? That keeps\\npg_log_* stuff \\\"log only\\\", but adds something adjacent enough to hopefully\\nreduce future misunderstandings?\\n\\nTo further decrease the chance of such mistakes, we could rename\\npg_log_fatal() to something else. Or add a a pg_log_fatal() function that's\\nmarked pg_nodiscard(), so that each callsite explicitly has to be (void)\\npg_log_fatal().\\n\\n\\n> I've now gone through and fleshed out the patch I sketched upthread.\\n> This exercise convinced me that we absolutely should do something\\n> very like this, because\\n> \\n> (1) As it stands, this patch removes a net of just about 1000 lines\\n> of code. So we clearly need *something*.\\n\\n> (2) The savings would be even more, except that people have invented\\n> macros for pg_log_error + exit(1) in at least four places already.\\n> (None of them quite the same of course.) Here I've just fixed those\\n> macro definitions to use pg_log_fatal. In the name of consistency, we\\n> should probably get rid of those macros in favor of using pg_log_fatal\\n> directly; but I've not done so here, since this patch is eyewateringly\\n> long and boring already.\\n\\nAgreed.\\n\\n\\nI don't care that much about the naming here.\\n\\n\\nFor a moment I was wondering whether there'd be a benefit for having the \\\"pure\\nlogging\\\" stuff be in a different static library than the erroring variant. So\\nthat e.g. libpq's check for exit() doesn't run into trouble. But then I\\nremembered that libpq doesn't link against common...\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 22 Feb 2022 19:11:37 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> What about adding a pg_fatal() that's pg_log_fatal() + exit()? That keeps\\n> pg_log_* stuff \\\"log only\\\", but adds something adjacent enough to hopefully\\n> reduce future misunderstandings?\\n\\nI'd be okay with that, except that pg_upgrade already has a pg_fatal\\n(because it has its *own* logging system, just in case you thought\\nthis wasn't enough of a mess yet). I'm in favor of aligning\\npg_upgrade's logging with the rest, but I'd hoped to leave that for\\nlater. Making the names collide would be bad even as a short-term\\nthing, I fear.\\n\\nLooks like libpq_pipeline.c has its own pg_fatal, too.\\n\\nI'm not against choosing some name other than pg_log_fatal, but that\\nparticular suggestion has got conflicts. Got any other ideas?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 22 Feb 2022 22:44:25 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On 2022-02-22 22:44:25 -0500, Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n> > What about adding a pg_fatal() that's pg_log_fatal() + exit()? That keeps\\n> > pg_log_* stuff \\\"log only\\\", but adds something adjacent enough to hopefully\\n> > reduce future misunderstandings?\\n> \\n> I'd be okay with that, except that pg_upgrade already has a pg_fatal\\n> (because it has its *own* logging system, just in case you thought\\n> this wasn't enough of a mess yet). I'm in favor of aligning\\n> pg_upgrade's logging with the rest, but I'd hoped to leave that for\\n> later. Making the names collide would be bad even as a short-term\\n> thing, I fear.\\n\\nI guess we could name pg_upgrade's out of the way...\\n\\n\\n> I'm not against choosing some name other than pg_log_fatal, but that\\n> particular suggestion has got conflicts. Got any other ideas?\\n\\nMaybe pg_fatal_exit(), pg_exit_fatal() or pg_fatal_error()?\\n\\n\\n\",\n \"msg_date\": \"Tue, 22 Feb 2022 20:06:21 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n>> What about adding a pg_fatal() that's pg_log_fatal() + exit()? That keeps\\n>> pg_log_* stuff \\\"log only\\\", but adds something adjacent enough to hopefully\\n>> reduce future misunderstandings?\\n\\n> I'd be okay with that, except that pg_upgrade already has a pg_fatal\\n> (because it has its *own* logging system, just in case you thought\\n> this wasn't enough of a mess yet).\\n\\nWait a moment. After looking closer, I realize that pg_upgrade's\\npg_fatal could trivially be turned into a macro; and the other two\\nexisting definitions already are macros. That would remove the risk\\nof link-time symbol collisions that I was worried about. As a bonus,\\nit'd substantially reduce the number of changes needed to make\\npg_upgrade use logging.c, whenever somebody wants to make that happen.\\n\\nSo I now propose modifying yesterday's patch thus:\\n\\n* Reinstantiate the PG_LOG_FATAL enum value, add support macros\\npg_log_fatal, pg_log_fatal_hint, pg_log_fatal_detail.\\n\\n* Define pg_fatal as pg_log_fatal + exit(1). (This would essentially\\nmove pg_rewind's definition into logging.h. pg_upgrade will\\ndefine it slightly differently, but the semantics end up the same.)\\n\\n* Adjust call sites to match.\\n\\nI do like this idea because it would not break any existing code\\nthat expects pg_log_fatal to return. There is likely to be some\\nof that in outstanding patches, and this approach would merely\\nrender it less-than-idiomatic rather than outright broken.\\n\\nUpdating the patch is going to be a bit tedious, so I'm not\\ngoing to do it without buy-in that this solution would be\\nokay to commit. Any objections?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 10:30:02 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-23 10:30:02 -0500, Tom Lane wrote:\\n> So I now propose modifying yesterday's patch thus:\\n> \\n> * Reinstantiate the PG_LOG_FATAL enum value, add support macros\\n> pg_log_fatal, pg_log_fatal_hint, pg_log_fatal_detail.\\n> \\n> * Define pg_fatal as pg_log_fatal + exit(1). (This would essentially\\n> move pg_rewind's definition into logging.h. pg_upgrade will\\n> define it slightly differently, but the semantics end up the same.)\\n> \\n> * Adjust call sites to match.\\n\\n+1\\n\\n\\n> I do like this idea because it would not break any existing code\\n> that expects pg_log_fatal to return. There is likely to be some\\n> of that in outstanding patches, and this approach would merely\\n> render it less-than-idiomatic rather than outright broken.\\n> \\n> Updating the patch is going to be a bit tedious, so I'm not\\n> going to do it without buy-in that this solution would be\\n> okay to commit. Any objections?\\n\\n+1\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 08:47:20 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 23, 2022, at 12:30 PM, Tom Lane wrote:\\n> Updating the patch is going to be a bit tedious, so I'm not\\n> going to do it without buy-in that this solution would be\\n> okay to commit. Any objections?\\nNo. I was confused by the multiple log functions spread in the client programs\\nwhile I was working on pg_subscriber [1]. Your proposal looks good to me.\\n\\n[1] https://postgr.es/m/5ac50071-f2ed-4ace-a8fd-b892cffd33eb%40www.fastmail.com\\n\\n\\n--\\nEuler Taveira\\nEDB https://www.enterprisedb.com/\\n\\nOn Wed, Feb 23, 2022, at 12:30 PM, Tom Lane wrote:Updating the patch is going to be a bit tedious, so I'm notgoing to do it without buy-in that this solution would beokay to commit. Any objections?No. I was confused by the multiple log functions spread in the client programswhile I was working on pg_subscriber [1]. Your proposal looks good to me.[1] https://postgr.es/m/5ac50071-f2ed-4ace-a8fd-b892cffd33eb%40www.fastmail.com--Euler TaveiraEDB https://www.enterprisedb.com/\",\n \"msg_date\": \"Wed, 23 Feb 2022 14:21:09 -0300\",\n \"msg_from\": \"\\\"Euler Taveira\\\" <euler@eulerto.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On 23.02.22 00:23, Tom Lane wrote:\\n> This conversation seems to have tailed off without full resolution,\\n> but I observe that pretty much everyone except Peter is on board\\n> with defining pg_log_fatal as pg_log_error + exit(1). So I think\\n> we should just do that, unless Peter wants to produce a finished\\n> alternative proposal.\\n> \\n> I've now gone through and fleshed out the patch I sketched upthread.\\n\\nThis patch already illustrates a couple of things that are wrong with \\nthis approach:\\n\\n- It doesn't allow any other way of exiting. For example, in pg_dump, \\nyou have removed a few exit_nicely() calls. It's not clear why that is \\nvalid or whether it would always be valid for all call sites.\\n\\n- It doesn't allow other exit codes. For example, in psql, you have \\nchanged a pg_log_fatal() call to pg_log_error() because it needed \\nanother exit code. This slides us right back into that annoying \\nsituation where in the backend we have to log error messages using \\nelog(LOG) because the flow control is tangled up with the log level.\\n\\nMy suggestion is to just get rid of pg_log_fatal() and replace them all \\nwith pg_log_error().\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 14:06:18 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-24 14:06:18 +0100, Peter Eisentraut wrote:\\n> My suggestion is to just get rid of pg_log_fatal() and replace them all with\\n> pg_log_error().\\n\\n-1. This ignores that already several places came up with their slightly\\ndifferent versions of fatal exit handlers. We don't gain anything by not\\nstandardizing on one notion of a fatal error wrapper.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 07:03:11 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> This patch already illustrates a couple of things that are wrong with \\n> this approach:\\n\\nAll of these objections are a bit moot with my followup proposal, no?\\n\\n> - It doesn't allow any other way of exiting. For example, in pg_dump, \\n> you have removed a few exit_nicely() calls. It's not clear why that is \\n> valid or whether it would always be valid for all call sites.\\n\\nAs the patch stood, I'd hacked it so that pg_log_fatal called\\nexit_nicely() not exit() within pg_dump. I agree that's not a great\\nsolution, but I think that the correct fix is to get rid of exit_nicely\\nin favor of doing the requisite cleanup in an atexit hook. pg_dump\\nis already at great hazard of somebody calling exit() not exit_nicely(),\\neither through failure to pay attention to that undocumented convention,\\nor because some code it imports from someplace like src/common calls\\nexit() directly. You need not look any further than pg_malloc()\\nto see that there are live problems there.\\n\\nI figured this could be addressed in a separate patch, though,\\nbecause to the extent that missing the exit-nicely callbacks is\\na real bug, it's already a bug.\\n\\n> My suggestion is to just get rid of pg_log_fatal() and replace them all \\n> with pg_log_error().\\n\\nI'm on board with dropping the separate FATAL log level if there's\\nconsensus to do so; I think it adds more confusion than anything else.\\nI still want to have something like pg_fatal(), though, because it's\\nimpossible to deny the usefulness of such an abbreviation. It's already\\nbeen reinvented three or four times independently.\\n\\nIndependently of either of those points, I still want to make the changes\\nI proposed w.r.t. making explicit concepts of \\\"detail\\\" and \\\"hint\\\"\\naddendums. What we have right now in the frontend code is an impossible\\nmishmash of three or four ways of filling that lack, all inconsistent.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 10:10:41 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> I'm on board with dropping the separate FATAL log level if there's\\n> consensus to do so; I think it adds more confusion than anything else.\\n\\nI feel that the reasonable alternatives are either to drop the FATAL\\nlog level, or try to make it actually mean something by consistently\\nusing it for errors that are indeed fatal. I had a go at doing the\\nlatter, but eventually concluded that that way madness lies. It's\\nnot too hard to use \\\"pg_log_fatal\\\" when there's an exit(1) right\\nafter it, but there are quite a lot of cases where a subroutine\\nreports an error and returns a failure code to its caller, whereupon\\nthe caller exits. Either the subroutine has to make an unwarranted\\nassumption about what its callers will do, or we need to make an API\\nchange to allow the subroutine itself to exit(), or we are going to\\npresent a user experience that is inconsistently different depending\\non internal implementation details.\\n\\nJust to add insult to injury, once I'd gotten done with the easy\\npart (use pg_log_fatal where there's an adjacent exit()), a whole\\nlot of TAP tests fell over, because they were expecting \\\"error:\\\"\\nnot \\\"fatal:\\\". The reverse direction of s/fatal/error/g only breaks\\none TAP test, which says a lot about how many places are troubling\\nto make this distinction now.\\n\\nI conclude that we ought to drop the separate FATAL level and just\\nuse ERROR instead.\\n\\nThe attached revision does that, standardizes on pg_fatal() as the\\nabbreviation for pg_log_error() + exit(1), and invents detail/hint\\nfeatures as per previous discussion.\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Fri, 25 Feb 2022 12:15:25 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 12:15:25PM -0500, Tom Lane wrote:\\n> I feel that the reasonable alternatives are either to drop the FATAL\\n> log level, or try to make it actually mean something by consistently\\n> using it for errors that are indeed fatal. I had a go at doing the\\n> latter, but eventually concluded that that way madness lies. It's\\n> not too hard to use \\\"pg_log_fatal\\\" when there's an exit(1) right\\n> after it, but there are quite a lot of cases where a subroutine\\n> reports an error and returns a failure code to its caller, whereupon\\n> the caller exits. Either the subroutine has to make an unwarranted\\n> assumption about what its callers will do, or we need to make an API\\n> change to allow the subroutine itself to exit(), or we are going to\\n> present a user experience that is inconsistently different depending\\n> on internal implementation details.\\n\\nNice code cut.\\n\\n> I conclude that we ought to drop the separate FATAL level and just\\n> use ERROR instead.\\n\\nFWIW, I have found the uses of pg_log_error() and pg_log_fatal()\\nrather confusing in some of the src/bin/ tools, so I am in favor of\\nremoving completely one or the other, and getting rid of FATAL is the\\nbest choice.\\n\\n> The attached revision does that, standardizes on pg_fatal() as the\\n> abbreviation for pg_log_error() + exit(1), and invents detail/hint\\n> features as per previous discussion.\\n\\n+#define pg_log_warning_hint(...) do { \\\\\\n+ if (likely(__pg_log_level <= PG_LOG_WARNING)) \\\\\\n+ pg_log_generic(PG_LOG_WARNING, PG_LOG_HINT, __VA_ARGS__);\\n\\\\\\n } while(0)\\n\\nHm. I am not sure to like much this abstraction by having so many\\nmacros that understand the log level through their name. Wouldn't it\\nbe cleaner to have only one pg_log_hint() and one pg_log_detail() with\\nthe log level passed as argument of the macro?\\n--\\nMichael\",\n \"msg_date\": \"Sat, 26 Feb 2022 14:21:01 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> Hm. I am not sure to like much this abstraction by having so many\\n> macros that understand the log level through their name. Wouldn't it\\n> be cleaner to have only one pg_log_hint() and one pg_log_detail() with\\n> the log level passed as argument of the macro?\\n\\nMmm ... that doesn't sound better to me. I think it wouldn't be\\nobvious that pg_log_warning and pg_log_hint are fundamentally\\ndifferent sorts of things: in the first, \\\"warning\\\" refers to\\nan error severity level, while in the second, \\\"hint\\\" refers to\\na message component. I'm not wedded to the way I did it in this\\npatch, but I think we ought to maintain a notational distinction\\nbetween those two sorts of concepts.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 00:55:34 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-25 12:15:25 -0500, Tom Lane wrote:\\n> The attached revision does that, standardizes on pg_fatal() as the\\n> abbreviation for pg_log_error() + exit(1), and invents detail/hint\\n> features as per previous discussion.\\n\\nThis unfortunately has had some bitrot: http://cfbot.cputube.org/patch_37_3574.log\\n\\nMarked as waiting-on-author.\\n\\nGreetings,\\n\\nAndres\\n\\n\\n\",\n \"msg_date\": \"Mon, 21 Mar 2022 18:51:35 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2022-02-25 12:15:25 -0500, Tom Lane wrote:\\n>> The attached revision does that, standardizes on pg_fatal() as the\\n>> abbreviation for pg_log_error() + exit(1), and invents detail/hint\\n>> features as per previous discussion.\\n\\n> This unfortunately has had some bitrot: http://cfbot.cputube.org/patch_37_3574.log\\n\\nYeah, I know. That patch touches enough places that it's sure to\\nbreak every few days during a commitfest. I'm not real excited about\\nmaintaining it reactively. The flip side of the coin is that pushing\\nit will doubtless break a lot of other uncommitted patches. So I'm\\nnot sure how to proceed.\\n\\nMaybe the plan could be to push it at the end of the commitfest? But\\nwhat I'd really like at the moment is buy-in or rejection of the whole\\nconcept, so I know whether it's worth spending more time on.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 21 Mar 2022 22:00:37 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-21 22:00:37 -0400, Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n> > This unfortunately has had some bitrot: http://cfbot.cputube.org/patch_37_3574.log\\n> \\n> Yeah, I know. That patch touches enough places that it's sure to\\n> break every few days during a commitfest. I'm not real excited about\\n> maintaining it reactively.\\n\\nMakes sense. I'd leave it on waiting-on-author then, so that reviewers looking\\nthrough the CF don't need to look at it?\\n\\n\\n> Maybe the plan could be to push it at the end of the commitfest?\\n\\nWould make sense to me.\\n\\nArguably parts of it could be committed sooner than that, e.g. exit(-1). But\\nthat'd not make it meaningfully easier to maintain, so ...\\n\\n\\n> But what I'd really like at the moment is buy-in or rejection of the whole\\n> concept, so I know whether it's worth spending more time on.\\n\\nI'm +1. I've not carefully looked through every single changed callsite, but\\nIMO it looks like a clear improvement.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 21 Mar 2022 19:13:48 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Here's a rebase up to today's HEAD. I've fixed the merge problems,\\nbut there may be some stray new error calls that could be converted\\nto use pg_fatal() and haven't been. I don't want to do a full\\nfresh scan of the code until we're about ready to commit this.\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Sun, 27 Mar 2022 16:19:53 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"> On 27 Mar 2022, at 22:19, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Here's a rebase up to today's HEAD. I've fixed the merge problems,\\n> but there may be some stray new error calls that could be converted\\n> to use pg_fatal() and haven't been. I don't want to do a full\\n> fresh scan of the code until we're about ready to commit this.\\n\\nI focused on finding any changes in logic (exiting where it previously kept\\ngoing after error, or vice versa) and was unable to find any which inspires\\nconfidence. This was a large patch to read through, so I'm fairly sure I've\\nmissed something.\\n\\nBelow are comments (nitpicks to a large extent) which either relates to this\\npatch or to the messages in general. For the latter, feel free to deem them\\nout of scope and we can take those comments separately from this work on\\n-hackers.\\n\\nAs a side-note realization from reading this, there is a tremendous amount of\\ncode paths in pg_dump which leads to a fatal error.. thats not necessarily a\\nbad thing, it's just striking when reading them all at once.\\n\\n> @@ -508,24 +502,15 @@ writefile(char *path, char **lines)\\n\\n> \\tif (fclose(out_file))\\n> -\\t{\\n> -\\t\\tpg_log_error(\\\"could not write file \\\\\\\"%s\\\\\\\": %m\\\", path);\\n> -\\t\\texit(1);\\n> -\\t}\\n> +\\t\\tpg_fatal(\\\"could not write file \\\\\\\"%s\\\\\\\": %m\\\", path);\\n> }\\n\\nShould we update this message to differentiate it from the fwrite error case?\\nSomething like \\\"an error occurred during writing..\\\"\\n\\n> @@ -2057,10 +2004,7 @@ check_locale_name(int category, const char *locale, char **canonname)\\n> \\n> \\tsave = setlocale(category, NULL);\\n> \\tif (!save)\\n> -\\t{\\n> -\\t\\tpg_log_error(\\\"setlocale() failed\\\");\\n> -\\t\\texit(1);\\n> -\\t}\\n> +\\t\\tpg_fatal(\\\"setlocale() failed\\\");\\n\\nShould this gain a hint message for those users who have no idea what this\\nreally means?\\n\\n> @@ -2127,15 +2067,14 @@ check_locale_encoding(const char *locale, int user_enc)\\n> \\t\\t user_enc == PG_SQL_ASCII))\\n> \\t{\\n> \\t\\tpg_log_error(\\\"encoding mismatch\\\");\\n> -\\t\\tfprintf(stderr,\\n> -\\t\\t\\t\\t_(\\\"The encoding you selected (%s) and the encoding that the\\\\n\\\"\\n> -\\t\\t\\t\\t \\\"selected locale uses (%s) do not match. This would lead to\\\\n\\\"\\n> -\\t\\t\\t\\t \\\"misbehavior in various character string processing functions.\\\\n\\\"\\n> -\\t\\t\\t\\t \\\"Rerun %s and either do not specify an encoding explicitly,\\\\n\\\"\\n> -\\t\\t\\t\\t \\\"or choose a matching combination.\\\\n\\\"),\\n> -\\t\\t\\t\\tpg_encoding_to_char(user_enc),\\n> -\\t\\t\\t\\tpg_encoding_to_char(locale_enc),\\n> -\\t\\t\\t\\tprogname);\\n> +\\t\\tpg_log_error_detail(\\\"The encoding you selected (%s) and the encoding that the \\\"\\n> +\\t\\t\\t\\t\\t\\t\\t\\\"selected locale uses (%s) do not match. This would lead to \\\"\\n\\nI will be sad to see double space after period go, but I think I'm rare in\\npreferring that.\\n\\n> @@ -2344,10 +2276,7 @@ setup_pgdata(void)\\n> \\t * have embedded spaces.\\n> \\t */\\n> \\tif (setenv(\\\"PGDATA\\\", pg_data, 1) != 0)\\n> -\\t{\\n> -\\t\\tpg_log_error(\\\"could not set environment\\\");\\n> -\\t\\texit(1);\\n> -\\t}\\n> +\\t\\tpg_fatal(\\\"could not set environment\\\");\\n\\nShould we be explicit about which env var?\\n\\n> @@ -3089,18 +2979,14 @@ main(int argc, char *argv[])\\n> \\t\\t\\t\\telse if (strcmp(optarg, \\\"libc\\\") == 0)\\n> \\t\\t\\t\\t\\tlocale_provider = COLLPROVIDER_LIBC;\\n> \\t\\t\\t\\telse\\n> -\\t\\t\\t\\t{\\n> -\\t\\t\\t\\t\\tpg_log_error(\\\"unrecognized locale provider: %s\\\", optarg);\\n> -\\t\\t\\t\\t\\texit(1);\\n> -\\t\\t\\t\\t}\\n> +\\t\\t\\t\\t\\tpg_fatal(\\\"unrecognized locale provider: %s\\\", optarg);\\n\\nShould this %s be within quotes to match how we usually emit user-input?\\n\\n> @@ -1123,9 +1097,9 @@ verify_btree_slot_handler(PGresult *res, PGconn *conn, void *context)\\n> \\t\\t\\tpg_log_warning(\\\"btree index \\\\\\\"%s.%s.%s\\\\\\\": btree checking function returned unexpected number of rows: %d\\\",\\n> \\t\\t\\t\\t\\t\\t rel->datinfo->datname, rel->nspname, rel->relname, ntups);\\n> \\t\\t\\tif (opts.verbose)\\n> -\\t\\t\\t\\tpg_log_info(\\\"query was: %s\\\", rel->sql);\\n> -\\t\\t\\tpg_log_warning(\\\"Are %s's and amcheck's versions compatible?\\\",\\n> -\\t\\t\\t\\t\\t\\t progname);\\n> +\\t\\t\\t\\tpg_log_warning_detail(\\\"Query was: %s\\\", rel->sql);\\n> +\\t\\t\\tpg_log_warning_hint(\\\"Are %s's and amcheck's versions compatible?\\\",\\n> +\\t\\t\\t\\t\\t\\t\\t\\tprogname);\\n\\nShould \\\"amcheck's\\\" be a %s parameter to make translation reusable (which it\\nmiht never be) and possibly avoid translation mistake?\\n\\n> --- a/src/bin/pg_basebackup/receivelog.c\\n> +++ b/src/bin/pg_basebackup/receivelog.c\\n> @@ -140,7 +140,7 @@ open_walfile(StreamCtl *stream, XLogRecPtr startpoint)\\n> \\t\\t\\t/* fsync file in case of a previous crash */\\n> \\t\\t\\tif (stream->walmethod->sync(f) != 0)\\n> \\t\\t\\t{\\n> -\\t\\t\\t\\tpg_log_fatal(\\\"could not fsync existing write-ahead log file \\\\\\\"%s\\\\\\\": %s\\\",\\n> +\\t\\t\\t\\tpg_log_error(\\\"could not fsync existing write-ahead log file \\\\\\\"%s\\\\\\\": %s\\\",\\n> \\t\\t\\t\\t\\t\\t\\t fn, stream->walmethod->getlasterror());\\n> \\t\\t\\t\\tstream->walmethod->close(f, CLOSE_UNLINK);\\n> \\t\\t\\t\\texit(1);\\n\\nIn the case where we already have IO related errors, couldn't the close() call\\ncause an additional error message which potentially could be helpful for\\ndebugging?\\n\\n> @@ -597,31 +570,19 @@ main(int argc, char *argv[])\\n\\n> \\tif (ControlFile->data_checksum_version == 0 &&\\n> \\t\\tmode == PG_MODE_CHECK)\\n> -\\t{\\n> -\\t\\tpg_log_error(\\\"data checksums are not enabled in cluster\\\");\\n> -\\t\\texit(1);\\n> -\\t}\\n> +\\t\\tpg_fatal(\\\"data checksums are not enabled in cluster\\\");\\n> \\n> \\tif (ControlFile->data_checksum_version == 0 &&\\n> \\t\\tmode == PG_MODE_DISABLE)\\n> -\\t{\\n> -\\t\\tpg_log_error(\\\"data checksums are already disabled in cluster\\\");\\n> -\\t\\texit(1);\\n> -\\t}\\n> +\\t\\tpg_fatal(\\\"data checksums are already disabled in cluster\\\");\\n> \\n> \\tif (ControlFile->data_checksum_version > 0 &&\\n> \\t\\tmode == PG_MODE_ENABLE)\\n> -\\t{\\n> -\\t\\tpg_log_error(\\\"data checksums are already enabled in cluster\\\");\\n> -\\t\\texit(1);\\n> -\\t}\\n> +\\t\\tpg_fatal(\\\"data checksums are already enabled in cluster\\\");\\n\\nFatal seems sort of out place here, it's not really a case of \\\"something\\nterrible happened\\\" but rather \\\"the preconditions weren't met\\\". Couldn't these\\nbe a single pg_log_error erroring out with the reason in a pg_log_detail?\\n\\n> @@ -1323,10 +1320,10 @@ lockTableForWorker(ArchiveHandle *AH, TocEntry *te)\\n> \\tres = PQexec(AH->connection, query->data);\\n> \\n> \\tif (!res || PQresultStatus(res) != PGRES_COMMAND_OK)\\n> -\\t\\tfatal(\\\"could not obtain lock on relation \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n> -\\t\\t\\t \\\"This usually means that someone requested an ACCESS EXCLUSIVE lock \\\"\\n> -\\t\\t\\t \\\"on the table after the pg_dump parent process had gotten the \\\"\\n> -\\t\\t\\t \\\"initial ACCESS SHARE lock on the table.\\\", qualId);\\n> +\\t\\tpg_fatal(\\\"could not obtain lock on relation \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n> +\\t\\t\\t\\t \\\"This usually means that someone requested an ACCESS EXCLUSIVE lock \\\"\\n> +\\t\\t\\t\\t \\\"on the table after the pg_dump parent process had gotten the \\\"\\n> +\\t\\t\\t\\t \\\"initial ACCESS SHARE lock on the table.\\\", qualId);\\n\\nPunctuation and formatting in a pg_fatal feels a bit wrong, and smells of need\\nfor _detail/_hint. I guess there isn't a way to do that here though, but still\\na shame.\\n\\n> @@ -50,9 +50,10 @@ _check_database_version(ArchiveHandle *AH)\\n> \\t\\t&& (remoteversion < AH->public.minRemoteVersion ||\\n> \\t\\t\\tremoteversion > AH->public.maxRemoteVersion))\\n> \\t{\\n> -\\t\\tpg_log_error(\\\"server version: %s; %s version: %s\\\",\\n> -\\t\\t\\t\\t\\t remoteversion_str, progname, PG_VERSION);\\n> -\\t\\tfatal(\\\"aborting because of server version mismatch\\\");\\n> +\\t\\tpg_log_error(\\\"aborting because of server version mismatch\\\");\\n> +\\t\\tpg_log_error_detail(\\\"server version: %s; %s version: %s\\\",\\n> +\\t\\t\\t\\t\\t\\t\\tremoteversion_str, progname, PG_VERSION);\\n> +\\t\\texit(1);\\n\\nCapitalization of the _detail message?\\n\\n> @@ -166,7 +167,7 @@ ConnectDatabase(Archive *AHX,\\n> \\t\\tAH->connection = PQconnectdbParams(keywords, values, true);\\n> \\n> \\t\\tif (!AH->connection)\\n> -\\t\\t\\tfatal(\\\"could not connect to database\\\");\\n> +\\t\\t\\tpg_fatal(\\\"could not connect to database\\\");\\n\\nShouldn't this print which database it failed to connect to?\\n\\n\\n> @@ -721,7 +721,7 @@ setFilePath(ArchiveHandle *AH, char *buf, const char *relativeFilename)\\n> \\tdname = ctx->directory;\\n> \\n> \\tif (strlen(dname) + 1 + strlen(relativeFilename) + 1 > MAXPGPATH)\\n\\nUnrelated, but shouldn't that be >= MAXPGPATH?\\n\\n> @@ -2033,8 +2027,8 @@ dumpTableData_copy(Archive *fout, const void *dcontext)\\n> \\t{\\n> \\t\\t/* copy data transfer failed */\\n> \\t\\tpg_log_error(\\\"Dumping the contents of table \\\\\\\"%s\\\\\\\" failed: PQgetCopyData() failed.\\\", classname);\\n> -\\t\\tpg_log_error(\\\"Error message from server: %s\\\", PQerrorMessage(conn));\\n> -\\t\\tpg_log_error(\\\"The command was: %s\\\", q->data);\\n> +\\t\\tpg_log_error_detail(\\\"Error message from server: %s\\\", PQerrorMessage(conn));\\n> +\\t\\tpg_log_error_detail(\\\"Command was: %s\\\", q->data);\\n\\nShouldn't the pg_log_error() call be with leading lowercase and no punctuation?\\n(same in the following hunk)\\n\\n\\n> @@ -11756,8 +11747,8 @@ dumpFunc(Archive *fout, const FuncInfo *finfo)\\n> \\t\\telse if (provolatile[0] == PROVOLATILE_STABLE)\\n> \\t\\t\\tappendPQExpBufferStr(q, \\\" STABLE\\\");\\n> \\t\\telse if (provolatile[0] != PROVOLATILE_VOLATILE)\\n> -\\t\\t\\tfatal(\\\"unrecognized provolatile value for function \\\\\\\"%s\\\\\\\"\\\",\\n> -\\t\\t\\t\\t finfo->dobj.name);\\n> +\\t\\t\\tpg_fatal(\\\"unrecognized provolatile value for function \\\\\\\"%s\\\\\\\"\\\",\\n> +\\t\\t\\t\\t\\t finfo->dobj.name);\\n> \\t}\\n> \\n> \\tif (proisstrict[0] == 't')\\n> @@ -11806,8 +11797,8 @@ dumpFunc(Archive *fout, const FuncInfo *finfo)\\n> \\t\\telse if (proparallel[0] == PROPARALLEL_RESTRICTED)\\n> \\t\\t\\tappendPQExpBufferStr(q, \\\" PARALLEL RESTRICTED\\\");\\n> \\t\\telse if (proparallel[0] != PROPARALLEL_UNSAFE)\\n> -\\t\\t\\tfatal(\\\"unrecognized proparallel value for function \\\\\\\"%s\\\\\\\"\\\",\\n> -\\t\\t\\t\\t finfo->dobj.name);\\n> +\\t\\t\\tpg_fatal(\\\"unrecognized proparallel value for function \\\\\\\"%s\\\\\\\"\\\",\\n> +\\t\\t\\t\\t\\t finfo->dobj.name);\\n\\nWe should probably move the provolatile and propallel keywords to a %s param so\\nwe can reuse the translation.\\n\\n> @@ -14951,18 +14942,18 @@ createViewAsClause(Archive *fout, const TableInfo *tbinfo)\\n\\n> -\\t\\tfatal(\\\"definition of view \\\\\\\"%s\\\\\\\" appears to be empty (length zero)\\\",\\n> -\\t\\t\\t tbinfo->dobj.name);\\n> +\\t\\tpg_fatal(\\\"definition of view \\\\\\\"%s\\\\\\\" appears to be empty (length zero)\\\",\\n> +\\t\\t\\t\\t tbinfo->dobj.name);\\n\\nI'm not sure we need to provide a definition of empty here, most readers will\\nprobably understand that already =)\\n\\n> @@ -16602,13 +16593,10 @@ dumpSequence(Archive *fout, const TableInfo *tbinfo)\\n> \\tres = ExecuteSqlQuery(fout, query->data, PGRES_TUPLES_OK);\\n> \\n> \\tif (PQntuples(res) != 1)\\n> -\\t{\\n> -\\t\\tpg_log_error(ngettext(\\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d row (expected 1)\\\",\\n> -\\t\\t\\t\\t\\t\\t\\t \\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d rows (expected 1)\\\",\\n> -\\t\\t\\t\\t\\t\\t\\t PQntuples(res)),\\n> -\\t\\t\\t\\t\\t tbinfo->dobj.name, PQntuples(res));\\n> -\\t\\texit_nicely(1);\\n> -\\t}\\n> +\\t\\tpg_fatal(ngettext(\\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d row (expected 1)\\\",\\n> +\\t\\t\\t\\t\\t\\t \\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d rows (expected 1)\\\",\\n> +\\t\\t\\t\\t\\t\\t PQntuples(res)),\\n> +\\t\\t\\t\\t tbinfo->dobj.name, PQntuples(res));\\n\\nThe ngettext() call seems a bit out of place here since we already know that\\nsecond form will be taken given the check on PQntuples(res).\\n\\n> @@ -16824,13 +16812,10 @@ dumpSequenceData(Archive *fout, const TableDataInfo *tdinfo)\\n> \\tres = ExecuteSqlQuery(fout, query->data, PGRES_TUPLES_OK);\\n> \\n> \\tif (PQntuples(res) != 1)\\n> -\\t{\\n> -\\t\\tpg_log_error(ngettext(\\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d row (expected 1)\\\",\\n> -\\t\\t\\t\\t\\t\\t\\t \\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d rows (expected 1)\\\",\\n> -\\t\\t\\t\\t\\t\\t\\t PQntuples(res)),\\n> -\\t\\t\\t\\t\\t tbinfo->dobj.name, PQntuples(res));\\n> -\\t\\texit_nicely(1);\\n> -\\t}\\n> +\\t\\tpg_fatal(ngettext(\\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d row (expected 1)\\\",\\n> +\\t\\t\\t\\t\\t\\t \\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d rows (expected 1)\\\",\\n> +\\t\\t\\t\\t\\t\\t PQntuples(res)),\\n> +\\t\\t\\t\\t tbinfo->dobj.name, PQntuples(res));\\n\\nSame as above, PQntuples(res) cannot be 1 so there os just one form used no?\\n\\n> --- a/src/bin/pg_dump/pg_dumpall.c\\n> +++ b/src/bin/pg_dump/pg_dumpall.c\\n> @@ -200,16 +200,15 @@ main(int argc, char *argv[])\\n> \\t\\t\\tstrlcpy(full_path, progname, sizeof(full_path));\\n> \\n> \\t\\tif (ret == -1)\\n> -\\t\\t\\tpg_log_error(\\\"The program \\\\\\\"%s\\\\\\\" is needed by %s but was not found in the\\\\n\\\"\\n> -\\t\\t\\t\\t\\t\\t \\\"same directory as \\\\\\\"%s\\\\\\\".\\\\n\\\"\\n> -\\t\\t\\t\\t\\t\\t \\\"Check your installation.\\\",\\n> -\\t\\t\\t\\t\\t\\t \\\"pg_dump\\\", progname, full_path);\\n> +\\t\\t\\tpg_fatal(\\\"The program \\\\\\\"%s\\\\\\\" is needed by %s but was not found in the\\\\n\\\"\\n> +\\t\\t\\t\\t\\t \\\"same directory as \\\\\\\"%s\\\\\\\".\\\\n\\\"\\n> +\\t\\t\\t\\t\\t \\\"Check your installation.\\\",\\n> +\\t\\t\\t\\t\\t \\\"pg_dump\\\", progname, full_path);\\n> \\t\\telse\\n> -\\t\\t\\tpg_log_error(\\\"The program \\\\\\\"%s\\\\\\\" was found by \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n> -\\t\\t\\t\\t\\t\\t \\\"but was not the same version as %s.\\\\n\\\"\\n> -\\t\\t\\t\\t\\t\\t \\\"Check your installation.\\\",\\n> -\\t\\t\\t\\t\\t\\t \\\"pg_dump\\\", full_path, progname);\\n> -\\t\\texit_nicely(1);\\n> +\\t\\t\\tpg_fatal(\\\"The program \\\\\\\"%s\\\\\\\" was found by \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n> +\\t\\t\\t\\t\\t \\\"but was not the same version as %s.\\\\n\\\"\\n> +\\t\\t\\t\\t\\t \\\"Check your installation.\\\",\\n\\n\\\"Check your installation\\\" seems superfluous here given the error. (we do this\\nin a number of places)\\n\\n\\n> @@ -1611,9 +1572,9 @@ connectDatabase(const char *dbname, const char *connection_string,\\n> \\t\\t&& (server_version < 90200 ||\\n> \\t\\t\\t(server_version / 100) > (my_version / 100)))\\n> \\t{\\n> -\\t\\tpg_log_error(\\\"server version: %s; %s version: %s\\\",\\n> -\\t\\t\\t\\t\\t remoteversion_str, progname, PG_VERSION);\\n> \\t\\tpg_log_error(\\\"aborting because of server version mismatch\\\");\\n> +\\t\\tpg_log_error_detail(\\\"server version: %s; %s version: %s\\\",\\n> +\\t\\t\\t\\t\\t\\t\\tremoteversion_str, progname, PG_VERSION);\\n\\nCapitalization?\\n\\n> @@ -388,13 +387,13 @@ WALDumpReadPage(XLogReaderState *state, XLogRecPtr targetPagePtr, int reqLen,\\n> \\t\\tif (errinfo.wre_errno != 0)\\n> \\t\\t{\\n> \\t\\t\\terrno = errinfo.wre_errno;\\n> -\\t\\t\\tfatal_error(\\\"could not read from file %s, offset %d: %m\\\",\\n> -\\t\\t\\t\\t\\t\\tfname, errinfo.wre_off);\\n> +\\t\\t\\tpg_fatal(\\\"could not read from file %s, offset %d: %m\\\",\\n> +\\t\\t\\t\\t\\t fname, errinfo.wre_off);\\n> \\t\\t}\\n> \\t\\telse\\n> -\\t\\t\\tfatal_error(\\\"could not read from file %s, offset %d: read %d of %d\\\",\\n> -\\t\\t\\t\\t\\t\\tfname, errinfo.wre_off, errinfo.wre_read,\\n> -\\t\\t\\t\\t\\t\\terrinfo.wre_req);\\n> +\\t\\t\\tpg_fatal(\\\"could not read from file %s, offset %d: read %d of %d\\\",\\n> +\\t\\t\\t\\t\\t fname, errinfo.wre_off, errinfo.wre_read,\\n> +\\t\\t\\t\\t\\t errinfo.wre_req);\\n\\nFilename %s should be quoted like \\\\\\\"%s\\\\\\\".\\n\\n> @@ -7379,7 +7254,7 @@ main(int argc, char **argv)\\n> \\tTHREAD_BARRIER_DESTROY(&barrier);\\n> \\n> \\tif (exit_code != 0)\\n> -\\t\\tpg_log_fatal(\\\"Run was aborted; the above results are incomplete.\\\");\\n> +\\t\\tpg_log_error(\\\"Run was aborted; the above results are incomplete.\\\");\\n\\nCapitalization and punctuation?\\n\\n> @@ -781,10 +772,7 @@ process_psqlrc(char *argv0)\\n> \\tchar\\t *envrc = getenv(\\\"PSQLRC\\\");\\n> \\n> \\tif (find_my_exec(argv0, my_exec_path) < 0)\\n> -\\t{\\n> -\\t\\tpg_log_fatal(\\\"could not find own program executable\\\");\\n> -\\t\\texit(EXIT_FAILURE);\\n> -\\t}\\n> +\\t\\tpg_fatal(\\\"could not find own program executable\\\");\\n\\nI have a feeling this will be confusing to a fair few users. Not sure how to\\nimprove it, but there seems to be some room for it.\\n\\n> @@ -144,16 +145,10 @@ main(int argc, char *argv[])\\n> \\tif (alldb)\\n> \\t{\\n> \\t\\tif (dbname)\\n> -\\t\\t{\\n> -\\t\\t\\tpg_log_error(\\\"cannot cluster all databases and a specific one at the same time\\\");\\n> -\\t\\t\\texit(1);\\n> -\\t\\t}\\n> +\\t\\t\\tpg_fatal(\\\"cannot cluster all databases and a specific one at the same time\\\");\\n> \\n> \\t\\tif (tables.head != NULL)\\n> -\\t\\t{\\n> -\\t\\t\\tpg_log_error(\\\"cannot cluster specific table(s) in all databases\\\");\\n> -\\t\\t\\texit(1);\\n> -\\t\\t}\\n> +\\t\\t\\tpg_fatal(\\\"cannot cluster specific table(s) in all databases\\\");\\n\\nAn ngettext() candidate perhaps? There are more like this in main() hunks further down omitted for brevity here.\\n\\n> diff --git a/src/fe_utils/connect_utils.c b/src/fe_utils/connect_utils.c\\n> index a30c66f13a..f2e583f9fa 100644\\n> --- a/src/fe_utils/connect_utils.c\\n> +++ b/src/fe_utils/connect_utils.c\\n> @@ -88,11 +88,8 @@ connectDatabase(const ConnParams *cparams, const char *progname,\\n> \\t\\tconn = PQconnectdbParams(keywords, values, true);\\n> \\n> \\t\\tif (!conn)\\n> -\\t\\t{\\n> -\\t\\t\\tpg_log_error(\\\"could not connect to database %s: out of memory\\\",\\n> -\\t\\t\\t\\t\\t\\t cparams->dbname);\\n> -\\t\\t\\texit(1);\\n> -\\t\\t}\\n> +\\t\\t\\tpg_fatal(\\\"could not connect to database %s: out of memory\\\",\\n> +\\t\\t\\t\\t\\t cparams->dbname);\\n\\nDatabase name quoted like \\\\\\\"%s\\\\\\\" to match other messages?\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 12:13:38 +0200\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On 27.03.22 22:19, Tom Lane wrote:\\n> Here's a rebase up to today's HEAD. I've fixed the merge problems,\\n> but there may be some stray new error calls that could be converted\\n> to use pg_fatal() and haven't been. I don't want to do a full\\n> fresh scan of the code until we're about ready to commit this.\\n\\nThis looks like a good improvement to me.\\n\\nI think I would want the program name/location also in front of the \\ndetail and hint lines. I need to think about this a bit more. This \\nshouldn't hold up this patch; it would be a quick localized change. \\n(I'm also thinking about providing a separate color code for the \\nsecondary messages. Again, this could be a quick follow-up patch.)\\n\\nThe one change I didn't like was\\n\\n- pg_log_error(\\\"The program \\\\\\\"%s\\\\\\\" is needed by %s \\nbut was not found in the\\\\n\\\"\\n- \\\"same directory as \\n\\\\\\\"%s\\\\\\\".\\\\n\\\"\\n- \\\"Check your installation.\\\",\\n+ pg_log_error(\\\"the program \\\\\\\"%s\\\\\\\" is needed by %s \\nbut was not found in the same directory as \\\\\\\"%s\\\\\\\"\\\",\\n \\\"postgres\\\", progname, \\nfull_path);\\n\\nThis appears to prioritize the guideline \\\"don't punctuate error message \\nas full sentence\\\" over what should be the actual guideline \\\"don't make \\nthe error message a full sentence\\\".\\n\\nThere are other occurrences of a similar message that were not changed \\nin the same way by the patch. Maybe we should leave this one alone in \\nthis patch and consider rewriting the message instead.\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 16:24:18 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On 29.03.22 12:13, Daniel Gustafsson wrote:\\n\\nMost of these should probably be addressed separately from Tom's patch.\\n\\n>> @@ -508,24 +502,15 @@ writefile(char *path, char **lines)\\n> \\n>> \\tif (fclose(out_file))\\n>> -\\t{\\n>> -\\t\\tpg_log_error(\\\"could not write file \\\\\\\"%s\\\\\\\": %m\\\", path);\\n>> -\\t\\texit(1);\\n>> -\\t}\\n>> +\\t\\tpg_fatal(\\\"could not write file \\\\\\\"%s\\\\\\\": %m\\\", path);\\n>> }\\n> \\n> Should we update this message to differentiate it from the fwrite error case?\\n> Something like \\\"an error occurred during writing..\\\"\\n\\nShould be \\\"could not close ...\\\", no?\\n\\n>> @@ -2057,10 +2004,7 @@ check_locale_name(int category, const char *locale, char **canonname)\\n>>\\n>> \\tsave = setlocale(category, NULL);\\n>> \\tif (!save)\\n>> -\\t{\\n>> -\\t\\tpg_log_error(\\\"setlocale() failed\\\");\\n>> -\\t\\texit(1);\\n>> -\\t}\\n>> +\\t\\tpg_fatal(\\\"setlocale() failed\\\");\\n> \\n> Should this gain a hint message for those users who have no idea what this\\n> really means?\\n\\nMy setlocale() man page says:\\n\\nERRORS\\n No errors are defined.\\n\\nSo uh ... ;-)\\n\\n>> @@ -3089,18 +2979,14 @@ main(int argc, char *argv[])\\n>> \\t\\t\\t\\telse if (strcmp(optarg, \\\"libc\\\") == 0)\\n>> \\t\\t\\t\\t\\tlocale_provider = COLLPROVIDER_LIBC;\\n>> \\t\\t\\t\\telse\\n>> -\\t\\t\\t\\t{\\n>> -\\t\\t\\t\\t\\tpg_log_error(\\\"unrecognized locale provider: %s\\\", optarg);\\n>> -\\t\\t\\t\\t\\texit(1);\\n>> -\\t\\t\\t\\t}\\n>> +\\t\\t\\t\\t\\tpg_fatal(\\\"unrecognized locale provider: %s\\\", optarg);\\n> \\n> Should this %s be within quotes to match how we usually emit user-input?\\n\\nUsually not done after colon, but could be.\\n\\n>> @@ -1123,9 +1097,9 @@ verify_btree_slot_handler(PGresult *res, PGconn *conn, void *context)\\n>> \\t\\t\\tpg_log_warning(\\\"btree index \\\\\\\"%s.%s.%s\\\\\\\": btree checking function returned unexpected number of rows: %d\\\",\\n>> \\t\\t\\t\\t\\t\\t rel->datinfo->datname, rel->nspname, rel->relname, ntups);\\n>> \\t\\t\\tif (opts.verbose)\\n>> -\\t\\t\\t\\tpg_log_info(\\\"query was: %s\\\", rel->sql);\\n>> -\\t\\t\\tpg_log_warning(\\\"Are %s's and amcheck's versions compatible?\\\",\\n>> -\\t\\t\\t\\t\\t\\t progname);\\n>> +\\t\\t\\t\\tpg_log_warning_detail(\\\"Query was: %s\\\", rel->sql);\\n>> +\\t\\t\\tpg_log_warning_hint(\\\"Are %s's and amcheck's versions compatible?\\\",\\n>> +\\t\\t\\t\\t\\t\\t\\t\\tprogname);\\n> \\n> Should \\\"amcheck's\\\" be a %s parameter to make translation reusable (which it\\n> miht never be) and possibly avoid translation mistake?\\n\\nWe don't have translations set up for contrib modules. Otherwise, this \\nkind of thing would probably be something to look into.\\n\\n>> --- a/src/bin/pg_basebackup/receivelog.c\\n>> +++ b/src/bin/pg_basebackup/receivelog.c\\n>> @@ -140,7 +140,7 @@ open_walfile(StreamCtl *stream, XLogRecPtr startpoint)\\n>> \\t\\t\\t/* fsync file in case of a previous crash */\\n>> \\t\\t\\tif (stream->walmethod->sync(f) != 0)\\n>> \\t\\t\\t{\\n>> -\\t\\t\\t\\tpg_log_fatal(\\\"could not fsync existing write-ahead log file \\\\\\\"%s\\\\\\\": %s\\\",\\n>> +\\t\\t\\t\\tpg_log_error(\\\"could not fsync existing write-ahead log file \\\\\\\"%s\\\\\\\": %s\\\",\\n>> \\t\\t\\t\\t\\t\\t\\t fn, stream->walmethod->getlasterror());\\n>> \\t\\t\\t\\tstream->walmethod->close(f, CLOSE_UNLINK);\\n>> \\t\\t\\t\\texit(1);\\n> \\n> In the case where we already have IO related errors, couldn't the close() call\\n> cause an additional error message which potentially could be helpful for\\n> debugging?\\n\\nYeah, I think in general we have been sloppy with reporting file-closing \\nerrors properly. Those presumably happen very rarely, but when they do, \\nthings are probably very bad.\\n\\n>> @@ -597,31 +570,19 @@ main(int argc, char *argv[])\\n> \\n>> \\tif (ControlFile->data_checksum_version == 0 &&\\n>> \\t\\tmode == PG_MODE_CHECK)\\n>> -\\t{\\n>> -\\t\\tpg_log_error(\\\"data checksums are not enabled in cluster\\\");\\n>> -\\t\\texit(1);\\n>> -\\t}\\n>> +\\t\\tpg_fatal(\\\"data checksums are not enabled in cluster\\\");\\n\\n> Fatal seems sort of out place here, it's not really a case of \\\"something\\n> terrible happened\\\" but rather \\\"the preconditions weren't met\\\". Couldn't these\\n> be a single pg_log_error erroring out with the reason in a pg_log_detail?\\n\\n\\\"fatal\\\" means error plus exit, so this seems ok. There is no separate \\nlog level for really-bad-error.\\n\\n>> @@ -721,7 +721,7 @@ setFilePath(ArchiveHandle *AH, char *buf, const char *relativeFilename)\\n>> \\tdname = ctx->directory;\\n>>\\n>> \\tif (strlen(dname) + 1 + strlen(relativeFilename) + 1 > MAXPGPATH)\\n> \\n> Unrelated, but shouldn't that be >= MAXPGPATH?\\n\\nSeems correct to me as is.\\n\\n>> @@ -14951,18 +14942,18 @@ createViewAsClause(Archive *fout, const TableInfo *tbinfo)\\n> \\n>> -\\t\\tfatal(\\\"definition of view \\\\\\\"%s\\\\\\\" appears to be empty (length zero)\\\",\\n>> -\\t\\t\\t tbinfo->dobj.name);\\n>> +\\t\\tpg_fatal(\\\"definition of view \\\\\\\"%s\\\\\\\" appears to be empty (length zero)\\\",\\n>> +\\t\\t\\t\\t tbinfo->dobj.name);\\n> \\n> I'm not sure we need to provide a definition of empty here, most readers will\\n> probably understand that already =)\\n\\nIt could mean, contains no columns, or something similar. If I had to \\nchange it, I would remove the \\\"empty\\\" and keep the \\\"length zero\\\".\\n\\n>> @@ -16602,13 +16593,10 @@ dumpSequence(Archive *fout, const TableInfo *tbinfo)\\n>> \\tres = ExecuteSqlQuery(fout, query->data, PGRES_TUPLES_OK);\\n>>\\n>> \\tif (PQntuples(res) != 1)\\n>> -\\t{\\n>> -\\t\\tpg_log_error(ngettext(\\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d row (expected 1)\\\",\\n>> -\\t\\t\\t\\t\\t\\t\\t \\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d rows (expected 1)\\\",\\n>> -\\t\\t\\t\\t\\t\\t\\t PQntuples(res)),\\n>> -\\t\\t\\t\\t\\t tbinfo->dobj.name, PQntuples(res));\\n>> -\\t\\texit_nicely(1);\\n>> -\\t}\\n>> +\\t\\tpg_fatal(ngettext(\\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d row (expected 1)\\\",\\n>> +\\t\\t\\t\\t\\t\\t \\\"query to get data of sequence \\\\\\\"%s\\\\\\\" returned %d rows (expected 1)\\\",\\n>> +\\t\\t\\t\\t\\t\\t PQntuples(res)),\\n>> +\\t\\t\\t\\t tbinfo->dobj.name, PQntuples(res));\\n> \\n> The ngettext() call seems a bit out of place here since we already know that\\n> second form will be taken given the check on PQntuples(res).\\n\\nSee \\n<https://www.postgresql.org/message-id/flat/CALj2ACUfJKTmK5v%3DvF%2BH2iLkqM9Yvjsp6iXaCqAks6gDpzZh6g%40mail.gmail.com> \\nfor a similar case that explains why this is still correct and necessary.\\n\\n>> @@ -144,16 +145,10 @@ main(int argc, char *argv[])\\n>> \\tif (alldb)\\n>> \\t{\\n>> \\t\\tif (dbname)\\n>> -\\t\\t{\\n>> -\\t\\t\\tpg_log_error(\\\"cannot cluster all databases and a specific one at the same time\\\");\\n>> -\\t\\t\\texit(1);\\n>> -\\t\\t}\\n>> +\\t\\t\\tpg_fatal(\\\"cannot cluster all databases and a specific one at the same time\\\");\\n>>\\n>> \\t\\tif (tables.head != NULL)\\n>> -\\t\\t{\\n>> -\\t\\t\\tpg_log_error(\\\"cannot cluster specific table(s) in all databases\\\");\\n>> -\\t\\t\\texit(1);\\n>> -\\t\\t}\\n>> +\\t\\t\\tpg_fatal(\\\"cannot cluster specific table(s) in all databases\\\");\\n> \\n> An ngettext() candidate perhaps? There are more like this in main() hunks further down omitted for brevity here.\\n\\nI would just rephrase this as\\n\\n \\\"cannot cluster specific tables in all databases\\\"\\n\\nThis is still correct and sensible if the user specified just one table.\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 16:38:35 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"> On 29 Mar 2022, at 16:38, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\\n> \\n> On 29.03.22 12:13, Daniel Gustafsson wrote:\\n> \\n> Most of these should probably be addressed separately from Tom's patch.\\n\\nYeah, I think so too. I'll await input from Tom on how he wants to do, but I'm\\nhappy to take these to a new thread. The main point of the review was to find\\nlogic errors in the logging changes, these came as a bonus from reading them\\nall in one place.\\n\\n>>> @@ -508,24 +502,15 @@ writefile(char *path, char **lines)\\n>>> \\tif (fclose(out_file))\\n>>> -\\t{\\n>>> -\\t\\tpg_log_error(\\\"could not write file \\\\\\\"%s\\\\\\\": %m\\\", path);\\n>>> -\\t\\texit(1);\\n>>> -\\t}\\n>>> +\\t\\tpg_fatal(\\\"could not write file \\\\\\\"%s\\\\\\\": %m\\\", path);\\n>>> }\\n>> Should we update this message to differentiate it from the fwrite error case?\\n>> Something like \\\"an error occurred during writing..\\\"\\n> \\n> Should be \\\"could not close ...\\\", no?\\n\\nYes, it should, not sure what I was thinking.\\n\\n>>> @@ -2057,10 +2004,7 @@ check_locale_name(int category, const char *locale, char **canonname)\\n>>> \\n>>> \\tsave = setlocale(category, NULL);\\n>>> \\tif (!save)\\n>>> -\\t{\\n>>> -\\t\\tpg_log_error(\\\"setlocale() failed\\\");\\n>>> -\\t\\texit(1);\\n>>> -\\t}\\n>>> +\\t\\tpg_fatal(\\\"setlocale() failed\\\");\\n>> Should this gain a hint message for those users who have no idea what this\\n>> really means?\\n> \\n> My setlocale() man page says:\\n> \\n> ERRORS\\n> No errors are defined.\\n> \\n> So uh ... ;-)\\n\\nEven more reason to be confused then =) We have a mixed bag in the tree on how\\nwe handle errors from setlocale, in this case we could reword it to say\\nsomething like \\\"failed to retrieve locale for %s, category\\\" which IMO would be\\nslightly more informative. Might be academic though since I guess it rarely, if\\never, happens.\\n\\n>>> --- a/src/bin/pg_basebackup/receivelog.c\\n>>> +++ b/src/bin/pg_basebackup/receivelog.c\\n>>> @@ -140,7 +140,7 @@ open_walfile(StreamCtl *stream, XLogRecPtr startpoint)\\n>>> \\t\\t\\t/* fsync file in case of a previous crash */\\n>>> \\t\\t\\tif (stream->walmethod->sync(f) != 0)\\n>>> \\t\\t\\t{\\n>>> -\\t\\t\\t\\tpg_log_fatal(\\\"could not fsync existing write-ahead log file \\\\\\\"%s\\\\\\\": %s\\\",\\n>>> +\\t\\t\\t\\tpg_log_error(\\\"could not fsync existing write-ahead log file \\\\\\\"%s\\\\\\\": %s\\\",\\n>>> \\t\\t\\t\\t\\t\\t\\t fn, stream->walmethod->getlasterror());\\n>>> \\t\\t\\t\\tstream->walmethod->close(f, CLOSE_UNLINK);\\n>>> \\t\\t\\t\\texit(1);\\n>> In the case where we already have IO related errors, couldn't the close() call\\n>> cause an additional error message which potentially could be helpful for\\n>> debugging?\\n> \\n> Yeah, I think in general we have been sloppy with reporting file-closing errors properly. Those presumably happen very rarely, but when they do, things are probably very bad.\\n\\nAgreed. I'm not sure if Tom wants to add net new loggings in this patchset,\\nelse we could do this separately.\\n\\n>> The ngettext() call seems a bit out of place here since we already know that\\n>> second form will be taken given the check on PQntuples(res).\\n> \\n> See <https://www.postgresql.org/message-id/flat/CALj2ACUfJKTmK5v%3DvF%2BH2iLkqM9Yvjsp6iXaCqAks6gDpzZh6g%40mail.gmail.com> for a similar case that explains why this is still correct and necessary.\\n\\nDoh, I knew that, sorry.\\n\\n>>> \\t}\\n>>> +\\t\\t\\tpg_fatal(\\\"cannot cluster specific table(s) in all databases\\\");\\n>> An ngettext() candidate perhaps? There are more like this in main() hunks further down omitted for brevity here.\\n> \\n> I would just rephrase this as\\n> \\n> \\\"cannot cluster specific tables in all databases\\\"\\n> \\n> This is still correct and sensible if the user specified just one table.\\n\\nThat's one way yes. Mostly I'm just a bit allergic to the \\\"foo(s)\\\" which my\\nunscientifically based gut feeling am concerned about not necessarily always\\nworking well for translations.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 00:19:05 +0200\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Daniel Gustafsson <daniel@yesql.se> writes:\\n> On 29 Mar 2022, at 16:38, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\\n>> Most of these should probably be addressed separately from Tom's patch.\\n\\n> Yeah, I think so too.\\n\\nAgreed. I tried to confine my patch to mechanical changes, except\\nfor changing places where the detail/hint features could be used.\\n(I don't say that I yielded to temptation nowhere, because I don't\\nrecall that for certain; but editing the message texts was not the\\npoint of my patch.)\\n\\nFeel free to work on a followup editing patch though.\\n\\nAnother thing that occurred to me as I looked at your list is that\\nI think a lot of these are nearly-can't-happen cases that we didn't\\nput a lot of effort into devising great error messages for. Maybe\\nwe should continue that approach, and in particular not put effort\\ninto translating such messages. That would suggest inventing\\n\\\"pg_fatal_internal\\\" and so forth, with the same functionality\\nexcept for not being gettext triggers, comparable to\\nerrmsg_internal. However, then somebody would have to make\\njudgment calls about which messages not to bother translating.\\nDo we want to go down that path?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 18:38:30 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"FYI this patch no longer applies.\\n\\nGiven it's a Tom patch and due to its nature it'll bitrot rapidly\\nanyways I'm don't see a point in updating the status though.\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 11:51:53 -0400\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Greg Stark <stark@mit.edu> writes:\\n> FYI this patch no longer applies.\\n\\nNo surprise :-(\\n\\n> Given it's a Tom patch and due to its nature it'll bitrot rapidly\\n> anyways I'm don't see a point in updating the status though.\\n\\nWe now seem to have buy-in on the concept, so my plan is to let\\nthis sit till the end of the commitfest, then rebase and push.\\nThat should avoid unnecessary churn for other pending patches.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 12:18:39 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"> On 30 Mar 2022, at 00:38, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Daniel Gustafsson <daniel@yesql.se> writes:\\n>> On 29 Mar 2022, at 16:38, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\\n>>> Most of these should probably be addressed separately from Tom's patch.\\n> \\n>> Yeah, I think so too.\\n> \\n> Agreed. I tried to confine my patch to mechanical changes, except\\n> for changing places where the detail/hint features could be used.\\n> (I don't say that I yielded to temptation nowhere, because I don't\\n> recall that for certain; but editing the message texts was not the\\n> point of my patch.)\\n> \\n> Feel free to work on a followup editing patch though.\\n\\nThats my plan, once this lands I'll rebase the comments on top of your work and\\nwe can have a separate discussion around them then.\\n\\n> Another thing that occurred to me as I looked at your list is that\\n> I think a lot of these are nearly-can't-happen cases that we didn't\\n> put a lot of effort into devising great error messages for. Maybe\\n> we should continue that approach, and in particular not put effort\\n> into translating such messages. That would suggest inventing\\n> \\\"pg_fatal_internal\\\" and so forth, with the same functionality\\n> except for not being gettext triggers, comparable to\\n> errmsg_internal. However, then somebody would have to make\\n> judgment calls about which messages not to bother translating.\\n> Do we want to go down that path?\\n\\nI'm not sure. If, against the odds, these cases do happen then a user of a\\nlocalized build probably really want to see the error in language they can\\neasily understand and take in.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Apr 2022 11:49:10 +0200\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Daniel Gustafsson <daniel@yesql.se> writes:\\n> On 30 Mar 2022, at 00:38, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Feel free to work on a followup editing patch though.\\n\\n> Thats my plan, once this lands I'll rebase the comments on top of your work and\\n> we can have a separate discussion around them then.\\n\\nThe main patch is pushed now. I addressed the complaint Peter had\\nabout the messages with \\\"Check your installation\\\" pseudo-hints\\nby getting rid of them; I concur with your observation that those\\nhints were basically useless. I also fixed the one place where the\\nmessage should clearly be \\\"could not close\\\" not \\\"could not write\\\".\\nMostly didn't yield to temptation anywhere else.\\n\\nOne other loose end is bothering me: I stuck with logging.h's\\noriginal choice to put \\\"if (likely())\\\" or \\\"if (unlikely())\\\"\\nconditionals into the macros, but I rather suspect that that's\\njust a waste. I think we should put a centralized level check\\ninto logging.c, and get rid of at least the \\\"if (likely())\\\"\\nchecks, because those are going to succeed approximately 100.0%\\nof the time. Maybe there's an argument for keeping the unlikely()\\nones.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 08 Apr 2022 15:05:25 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> One other loose end is bothering me: I stuck with logging.h's\\n> original choice to put \\\"if (likely())\\\" or \\\"if (unlikely())\\\"\\n> conditionals into the macros, but I rather suspect that that's\\n> just a waste. I think we should put a centralized level check\\n> into logging.c, and get rid of at least the \\\"if (likely())\\\"\\n> checks, because those are going to succeed approximately 100.0%\\n> of the time. Maybe there's an argument for keeping the unlikely()\\n> ones.\\n\\nConcretely, something like the attached. As a simple check,\\nI looked at the compiled size of pg_dump. It went from\\n\\n text data bss dec hex filename\\n 380298 4008 1384 385690 5e29a /home/postgres/testversion/bin/pg_dump\\n\\nto\\n\\n text data bss dec hex filename\\n 374954 4008 1384 380346 5cdba src/bin/pg_dump/pg_dump\\n\\nfor a savings of about 5K or 1.5%. Not a huge amount, but\\nnot nothing either, especially considering that the existing\\ncoding isn't buying us anything.\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Fri, 08 Apr 2022 16:26:49 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On 08.04.22 22:26, Tom Lane wrote:\\n> I wrote:\\n>> One other loose end is bothering me: I stuck with logging.h's\\n>> original choice to put \\\"if (likely())\\\" or \\\"if (unlikely())\\\"\\n>> conditionals into the macros, but I rather suspect that that's\\n>> just a waste. I think we should put a centralized level check\\n>> into logging.c, and get rid of at least the \\\"if (likely())\\\"\\n>> checks, because those are going to succeed approximately 100.0%\\n>> of the time. Maybe there's an argument for keeping the unlikely()\\n>> ones.\\n> \\n> Concretely, something like the attached. As a simple check,\\n> I looked at the compiled size of pg_dump. It went from\\n> \\n> text data bss dec hex filename\\n> 380298 4008 1384 385690 5e29a /home/postgres/testversion/bin/pg_dump\\n> \\n> to\\n> \\n> text data bss dec hex filename\\n> 374954 4008 1384 380346 5cdba src/bin/pg_dump/pg_dump\\n> \\n> for a savings of about 5K or 1.5%. Not a huge amount, but\\n> not nothing either, especially considering that the existing\\n> coding isn't buying us anything.\\n\\nYeah, that seems ok to change. The previous coding style is more useful \\nif you have a lot of debug messages in a hot code path, but that usually \\ndoesn't apply to where this is used.\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Apr 2022 17:04:30 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> On 08.04.22 22:26, Tom Lane wrote:\\n>>> I think we should put a centralized level check\\n>>> into logging.c, and get rid of at least the \\\"if (likely())\\\"\\n>>> checks, because those are going to succeed approximately 100.0%\\n>>> of the time. Maybe there's an argument for keeping the unlikely()\\n>>> ones.\\n\\n> Yeah, that seems ok to change. The previous coding style is more useful \\n> if you have a lot of debug messages in a hot code path, but that usually \\n> doesn't apply to where this is used.\\n\\nThe patch I presented keeps the unlikely() checks in the debug-level\\nmacros. Do you think we should drop those too? I figured that avoiding\\nevaluating the arguments would be worth something.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Apr 2022 11:22:40 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"\\nOn 11.04.22 17:22, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>> On 08.04.22 22:26, Tom Lane wrote:\\n>>>> I think we should put a centralized level check\\n>>>> into logging.c, and get rid of at least the \\\"if (likely())\\\"\\n>>>> checks, because those are going to succeed approximately 100.0%\\n>>>> of the time. Maybe there's an argument for keeping the unlikely()\\n>>>> ones.\\n> \\n>> Yeah, that seems ok to change. The previous coding style is more useful\\n>> if you have a lot of debug messages in a hot code path, but that usually\\n>> doesn't apply to where this is used.\\n> \\n> The patch I presented keeps the unlikely() checks in the debug-level\\n> macros. Do you think we should drop those too? I figured that avoiding\\n> evaluating the arguments would be worth something.\\n\\nOh, that's right, the whole thing is to not evaluate the arguments if \\nthe log level isn't adequate. We should probably keep that.\\n\\nIs the code size a big problem? ereport() has a bunch of extra code \\naround each call as well. Does it have similar problems?\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Apr 2022 20:29:34 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> On 11.04.22 17:22, Tom Lane wrote:\\n>> The patch I presented keeps the unlikely() checks in the debug-level\\n>> macros. Do you think we should drop those too? I figured that avoiding\\n>> evaluating the arguments would be worth something.\\n\\n> Oh, that's right, the whole thing is to not evaluate the arguments if \\n> the log level isn't adequate. We should probably keep that.\\n\\nI don't think that's nearly as interesting in the frontend as in\\nthe backend. Error messages in the backend frequently include\\ncatalog lookups and the like in the arguments, but I think nearly\\nall frontend messages are writing nothing more complicated than\\nvariables and maybe some indirections or array fetches. So I'm\\nnot all that worried about runtime, and would rather save some\\ncode space.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Apr 2022 14:56:13 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On 29.03.22 16:24, Peter Eisentraut wrote:\\n> I think I would want the program name/location also in front of the \\n> detail and hint lines. I need to think about this a bit more. This \\n> shouldn't hold up this patch; it would be a quick localized change.\\n\\nAfter experiencing this for a bit now, I propose to make this change.\\nIt lines up better and allows easier filtering of messages by program.\\n\\nExample:\\n\\nBefore:\\n\\ninitdb: error: locale \\\"zh_HK.Big5HKSCS\\\" requires unsupported encoding \\\"BIG5\\\"\\ndetail: Encoding \\\"BIG5\\\" is not allowed as a server-side encoding.\\nhint: Rerun initdb with a different locale selection.\\n\\nAfter:\\n\\ninitdb: error: locale \\\"zh_HK.Big5HKSCS\\\" requires unsupported encoding \\\"BIG5\\\"\\ninitdb: detail: Encoding \\\"BIG5\\\" is not allowed as a server-side encoding.\\ninitdb: hint: Rerun initdb with a different locale selection.\",\n \"msg_date\": \"Mon, 23 May 2022 11:53:04 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"On 23.05.22 11:53, Peter Eisentraut wrote:\\n> On 29.03.22 16:24, Peter Eisentraut wrote:\\n>> I think I would want the program name/location also in front of the \\n>> detail and hint lines. I need to think about this a bit more. This \\n>> shouldn't hold up this patch; it would be a quick localized change.\\n> \\n> After experiencing this for a bit now, I propose to make this change.\\n> It lines up better and allows easier filtering of messages by program.\\n\\nThis has been committed.\\n\\n\\n\",\n \"msg_date\": \"Mon, 30 May 2022 07:30:44 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Hi!\\n\\nCommit 9a374b77fb53 (Improve frontend error logging style.) replaced a \\nline of file src/include/common/logging.h\\n```\\nextern PGDLLIMPORT enum pg_log_level __pg_log_level;\\n```\\nto\\n```\\nextern enum pg_log_level __pg_log_level;\\n```\\ni.e. it is rollback of commit 8ec569479fc28 (Apply PGDLLIMPORT markings \\nbroadly.)\\n\\nIs it correct?\\n\\n-- \\nWith best regards,\\nDmitry Koval\\n\\nPostgres Professional: http://postgrespro.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 11 Oct 2022 17:57:03 +0300\",\n \"msg_from\": \"Dmitry Koval <d.koval@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Frontend error logging style\"\n },\n {\n \"msg_contents\": \"Dmitry Koval <d.koval@postgrespro.ru> writes:\\n> Hi!\\n> Commit 9a374b77fb53 (Improve frontend error logging style.) replaced a \\n> line of file src/include/common/logging.h\\n> ```\\n> extern PGDLLIMPORT enum pg_log_level __pg_log_level;\\n> ```\\n> to\\n> ```\\n> extern enum pg_log_level __pg_log_level;\\n> ```\\n> i.e. it is rollback of commit 8ec569479fc28 (Apply PGDLLIMPORT markings \\n> broadly.)\\n\\n> Is it correct?\\n\\nYes, since that file is frontend-only.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 11 Oct 2022 11:08:47 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Frontend error logging style\"\n }\n]","isConversational":false}}},{"rowIdx":53065,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\n\nThis is a follow-up of the discussion that happened here:\nhttps://www.postgresql.org/message-id/YYoJYO475dsqYhta@paquier.xyz\n\nRecently, we are getting warnings in the build of plperl when using at\nleast clang-12 because of their new flag -Wcompound-token-split-by-macro,\nall the warnings coming from Perl itself.\n\nUpstream has posted a fix about that but it will take time before this\npropagates across the buildfarm (five~ animals are now impacted?):\nhttps://www.nntp.perl.org/group/perl.perl5.changes/2021/07/msg57758.html\n\ndangomushi is impacted by that, and I'd like to stick an -Werror in it\nas it is proving to be good in detecting incorrect placeholders.\n\nTom has suggested that we could add -Wno-compound-token-split-by-macro\nto take care of the issue on our side, and attached is a patch to do\nso.\n\nAny objections? I'd like to get this back-patched.\n\nThanks,\n--\nMichael","msg_date":"Wed, 10 Nov 2021 07:35:21 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":true,"msg_subject":"Clean up build warnings of plperl with clang-12+"},{"msg_contents":"On Wed, Nov 10, 2021 at 07:35:21AM +0900, Michael Paquier wrote:\n> Tom has suggested that we could add -Wno-compound-token-split-by-macro\n> to take care of the issue on our side, and attached is a patch to do\n> so.\n> \n> Any objections? I'd like to get this back-patched.\n\nBackpatched this one as of 9ff47ea. That should allow the addition of\n-Werror on dangomushi.\n--\nMichael","msg_date":"Thu, 11 Nov 2021 11:43:17 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":true,"msg_subject":"Re: Clean up build warnings of plperl with clang-12+"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> Backpatched this one as of 9ff47ea. That should allow the addition of\n> -Werror on dangomushi.\n\nCool. I have also enabled -Werror on florican.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 10 Nov 2021 22:14:20 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Clean up build warnings of plperl with clang-12+"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> Backpatched this one as of 9ff47ea. That should allow the addition of\n> -Werror on dangomushi.\n\nI see you tried to do that by adding -Werror to CPPFLAGS. That doesn't\nwork because it breaks configure's tests. (I wonder if meson is any\nsmarter than autoconf about that.) The way I do it on my own\nanimals is like\n\nif ($branch eq 'HEAD' or $branch ge 'REL9_3')\n{\n # Add -Werror so we get errors for warning conditions.\n # Pre-9.3 PG doesn't compile cleanly with Sierra's cc.\n # Can't put this in CFLAGS because it breaks configure,\n # so use COPT instead.\n $conf{build_env}{COPT} = \"-Werror\";\n}\n\nGuess I could drop the branch check now.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 11 Nov 2021 13:18:32 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Clean up build warnings of plperl with clang-12+"},{"msg_contents":"\nOn 11/11/21 13:18, Tom Lane wrote:\n> Michael Paquier <michael@paquier.xyz> writes:\n>> Backpatched this one as of 9ff47ea. That should allow the addition of\n>> -Werror on dangomushi.\n> I see you tried to do that by adding -Werror to CPPFLAGS. That doesn't\n> work because it breaks configure's tests. (I wonder if meson is any\n> smarter than autoconf about that.) The way I do it on my own\n> animals is like\n>\n> if ($branch eq 'HEAD' or $branch ge 'REL9_3')\n> {\n> # Add -Werror so we get errors for warning conditions.\n> # Pre-9.3 PG doesn't compile cleanly with Sierra's cc.\n> # Can't put this in CFLAGS because it breaks configure,\n> # so use COPT instead.\n> $conf{build_env}{COPT} = \"-Werror\";\n> }\n>\n> Guess I could drop the branch check now.\n>\n> \t\n\n\n\nWouldn't it be better in any case just to add the clang fix for building\nplperl rather than globally?\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Thu, 11 Nov 2021 15:07:00 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Clean up build warnings of plperl with clang-12+"},{"msg_contents":"Andrew Dunstan <andrew@dunslane.net> writes:\n> Wouldn't it be better in any case just to add the clang fix for building\n> plperl rather than globally?\n\nThat'd be considerably messier wouldn't it? (And I wonder if the\nmeson system will be able to do it at all.) If you feel motivated\nto make that happen, go for it, but I don't.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 11 Nov 2021 15:51:35 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Clean up build warnings of plperl with clang-12+"},{"msg_contents":"On Thu, Nov 11, 2021 at 01:18:32PM -0500, Tom Lane wrote:\n> I see you tried to do that by adding -Werror to CPPFLAGS. That doesn't\n> work because it breaks configure's tests. (I wonder if meson is any\n> smarter than autoconf about that.) The way I do it on my own\n> animals is like\n\nNya. (Japanized Meh)\n\n> if ($branch eq 'HEAD' or $branch ge 'REL9_3')\n> {\n> # Add -Werror so we get errors for warning conditions.\n> # Pre-9.3 PG doesn't compile cleanly with Sierra's cc.\n> # Can't put this in CFLAGS because it breaks configure,\n> # so use COPT instead.\n> $conf{build_env}{COPT} = \"-Werror\";\n> }\n> \n> Guess I could drop the branch check now.\n\nThanks. I'll switch to something like that.\n--\nMichael","msg_date":"Fri, 12 Nov 2021 07:22:44 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":true,"msg_subject":"Re: Clean up build warnings of plperl with clang-12+"},{"msg_contents":"On Thu, Nov 11, 2021 at 03:51:35PM -0500, Tom Lane wrote:\n> That'd be considerably messier wouldn't it?\n\nYes, that would be a bit messier. For example, we could do that with\nsomething saved in Makefile.global.in by ./configure that plperl feeds\non to add this extra CFLAGS in its own local Makefile.\n\n> (And I wonder if the\n> meson system will be able to do it at all.) If you feel motivated\n> to make that happen, go for it, but I don't.\n\nNeither am I. 9ff47ea is the least invasive solution, at least.\n--\nMichael","msg_date":"Mon, 15 Nov 2021 12:39:39 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":true,"msg_subject":"Re: Clean up build warnings of plperl with clang-12+"}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\n\\nThis is a follow-up of the discussion that happened here:\\nhttps://www.postgresql.org/message-id/YYoJYO475dsqYhta@paquier.xyz\\n\\nRecently, we are getting warnings in the build of plperl when using at\\nleast clang-12 because of their new flag -Wcompound-token-split-by-macro,\\nall the warnings coming from Perl itself.\\n\\nUpstream has posted a fix about that but it will take time before this\\npropagates across the buildfarm (five~ animals are now impacted?):\\nhttps://www.nntp.perl.org/group/perl.perl5.changes/2021/07/msg57758.html\\n\\ndangomushi is impacted by that, and I'd like to stick an -Werror in it\\nas it is proving to be good in detecting incorrect placeholders.\\n\\nTom has suggested that we could add -Wno-compound-token-split-by-macro\\nto take care of the issue on our side, and attached is a patch to do\\nso.\\n\\nAny objections? I'd like to get this back-patched.\\n\\nThanks,\\n--\\nMichael\",\n \"msg_date\": \"Wed, 10 Nov 2021 07:35:21 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Clean up build warnings of plperl with clang-12+\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 10, 2021 at 07:35:21AM +0900, Michael Paquier wrote:\\n> Tom has suggested that we could add -Wno-compound-token-split-by-macro\\n> to take care of the issue on our side, and attached is a patch to do\\n> so.\\n> \\n> Any objections? I'd like to get this back-patched.\\n\\nBackpatched this one as of 9ff47ea. That should allow the addition of\\n-Werror on dangomushi.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 11 Nov 2021 11:43:17 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Clean up build warnings of plperl with clang-12+\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> Backpatched this one as of 9ff47ea. That should allow the addition of\\n> -Werror on dangomushi.\\n\\nCool. I have also enabled -Werror on florican.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 22:14:20 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Clean up build warnings of plperl with clang-12+\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> Backpatched this one as of 9ff47ea. That should allow the addition of\\n> -Werror on dangomushi.\\n\\nI see you tried to do that by adding -Werror to CPPFLAGS. That doesn't\\nwork because it breaks configure's tests. (I wonder if meson is any\\nsmarter than autoconf about that.) The way I do it on my own\\nanimals is like\\n\\nif ($branch eq 'HEAD' or $branch ge 'REL9_3')\\n{\\n # Add -Werror so we get errors for warning conditions.\\n # Pre-9.3 PG doesn't compile cleanly with Sierra's cc.\\n # Can't put this in CFLAGS because it breaks configure,\\n # so use COPT instead.\\n $conf{build_env}{COPT} = \\\"-Werror\\\";\\n}\\n\\nGuess I could drop the branch check now.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 13:18:32 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Clean up build warnings of plperl with clang-12+\"\n },\n {\n \"msg_contents\": \"\\nOn 11/11/21 13:18, Tom Lane wrote:\\n> Michael Paquier <michael@paquier.xyz> writes:\\n>> Backpatched this one as of 9ff47ea. That should allow the addition of\\n>> -Werror on dangomushi.\\n> I see you tried to do that by adding -Werror to CPPFLAGS. That doesn't\\n> work because it breaks configure's tests. (I wonder if meson is any\\n> smarter than autoconf about that.) The way I do it on my own\\n> animals is like\\n>\\n> if ($branch eq 'HEAD' or $branch ge 'REL9_3')\\n> {\\n> # Add -Werror so we get errors for warning conditions.\\n> # Pre-9.3 PG doesn't compile cleanly with Sierra's cc.\\n> # Can't put this in CFLAGS because it breaks configure,\\n> # so use COPT instead.\\n> $conf{build_env}{COPT} = \\\"-Werror\\\";\\n> }\\n>\\n> Guess I could drop the branch check now.\\n>\\n> \\t\\n\\n\\n\\nWouldn't it be better in any case just to add the clang fix for building\\nplperl rather than globally?\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 15:07:00 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Clean up build warnings of plperl with clang-12+\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew@dunslane.net> writes:\\n> Wouldn't it be better in any case just to add the clang fix for building\\n> plperl rather than globally?\\n\\nThat'd be considerably messier wouldn't it? (And I wonder if the\\nmeson system will be able to do it at all.) If you feel motivated\\nto make that happen, go for it, but I don't.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 15:51:35 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Clean up build warnings of plperl with clang-12+\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 01:18:32PM -0500, Tom Lane wrote:\\n> I see you tried to do that by adding -Werror to CPPFLAGS. That doesn't\\n> work because it breaks configure's tests. (I wonder if meson is any\\n> smarter than autoconf about that.) The way I do it on my own\\n> animals is like\\n\\nNya. (Japanized Meh)\\n\\n> if ($branch eq 'HEAD' or $branch ge 'REL9_3')\\n> {\\n> # Add -Werror so we get errors for warning conditions.\\n> # Pre-9.3 PG doesn't compile cleanly with Sierra's cc.\\n> # Can't put this in CFLAGS because it breaks configure,\\n> # so use COPT instead.\\n> $conf{build_env}{COPT} = \\\"-Werror\\\";\\n> }\\n> \\n> Guess I could drop the branch check now.\\n\\nThanks. I'll switch to something like that.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 12 Nov 2021 07:22:44 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Clean up build warnings of plperl with clang-12+\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 03:51:35PM -0500, Tom Lane wrote:\\n> That'd be considerably messier wouldn't it?\\n\\nYes, that would be a bit messier. For example, we could do that with\\nsomething saved in Makefile.global.in by ./configure that plperl feeds\\non to add this extra CFLAGS in its own local Makefile.\\n\\n> (And I wonder if the\\n> meson system will be able to do it at all.) If you feel motivated\\n> to make that happen, go for it, but I don't.\\n\\nNeither am I. 9ff47ea is the least invasive solution, at least.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 15 Nov 2021 12:39:39 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Clean up build warnings of plperl with clang-12+\"\n }\n]","isConversational":false}}},{"rowIdx":53066,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"I'm currently working on a GiST extension for PostgreSQL and I ran into strange ORDER BY results during my queries.\n\nBecause I can't find the problem in my source code, I want to investigate the issue by looking at the PostgreSQL source,\nmaybe inserting extra log messages.\nWhile trying to do this, I came to the conclusion that I need developer help, because I can't find the part of the source code\nwhere the ORDER BY logic is implemented.\n\nCould you please help, by telling me which files should I take a look at?\n\nThanks in advance!\n\n\n\n\n\n\n\n\nI'm currently working on a GiST extension for PostgreSQL and I ran into strange ORDER BY results during my queries.\n\n\n\n\nBecause I can't find the problem in my source code, I want to investigate the issue by looking at the PostgreSQL source,\n\nmaybe inserting extra log messages.\n\nWhile trying to do this, I came to the conclusion that I need developer help, because I can't find the part of the source code\n\nwhere the ORDER BY logic is implemented.\n\n\n\n\n\nCould you please help, by telling me which files should I take a look at?\n\n\n\n\nThanks in advance!","msg_date":"Wed, 10 Nov 2021 06:43:43 +0000","msg_from":"=?iso-8859-2?Q?Sajti_Zsolt_Zolt=E1n?= <qnwbq9@inf.elte.hu>","msg_from_op":true,"msg_subject":"ORDER BY logic in PostgreSQL source code"},{"msg_contents":"On Wed, 10 Nov 2021 at 11:57, Sajti Zsolt Zoltán <qnwbq9@inf.elte.hu> wrote:\n>\n> I'm currently working on a GiST extension for PostgreSQL and I ran into strange ORDER BY results during my queries.\n>\n> Because I can't find the problem in my source code, I want to investigate the issue by looking at the PostgreSQL source,\n> maybe inserting extra log messages.\n> While trying to do this, I came to the conclusion that I need developer help, because I can't find the part of the source code\n> where the ORDER BY logic is implemented.\n>\n> Could you please help, by telling me which files should I take a look at?\n>\n> Thanks in advance!\n\nI think you can start by looking into the function grouping_planner\nand explore further by going through the route of it.\n\n-- \nRegards,\nRafia Sabih\n\n\n","msg_date":"Wed, 10 Nov 2021 12:47:46 +0100","msg_from":"Rafia Sabih <rafia.pghackers@gmail.com>","msg_from_op":false,"msg_subject":"Re: ORDER BY logic in PostgreSQL source code"}],"string":"[\n {\n \"msg_contents\": \"I'm currently working on a GiST extension for PostgreSQL and I ran into strange ORDER BY results during my queries.\\n\\nBecause I can't find the problem in my source code, I want to investigate the issue by looking at the PostgreSQL source,\\nmaybe inserting extra log messages.\\nWhile trying to do this, I came to the conclusion that I need developer help, because I can't find the part of the source code\\nwhere the ORDER BY logic is implemented.\\n\\nCould you please help, by telling me which files should I take a look at?\\n\\nThanks in advance!\\n\\n\\n\\n\\n\\n\\n\\n\\nI'm currently working on a GiST extension for PostgreSQL and I ran into strange ORDER BY results during my queries.\\n\\n\\n\\n\\nBecause I can't find the problem in my source code, I want to investigate the issue by looking at the PostgreSQL source,\\n\\nmaybe inserting extra log messages.\\n\\nWhile trying to do this, I came to the conclusion that I need developer help, because I can't find the part of the source code\\n\\nwhere the ORDER BY logic is implemented.\\n\\n\\n\\n\\n\\nCould you please help, by telling me which files should I take a look at?\\n\\n\\n\\n\\nThanks in advance!\",\n \"msg_date\": \"Wed, 10 Nov 2021 06:43:43 +0000\",\n \"msg_from\": \"=?iso-8859-2?Q?Sajti_Zsolt_Zolt=E1n?= <qnwbq9@inf.elte.hu>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"ORDER BY logic in PostgreSQL source code\"\n },\n {\n \"msg_contents\": \"On Wed, 10 Nov 2021 at 11:57, Sajti Zsolt Zoltán <qnwbq9@inf.elte.hu> wrote:\\n>\\n> I'm currently working on a GiST extension for PostgreSQL and I ran into strange ORDER BY results during my queries.\\n>\\n> Because I can't find the problem in my source code, I want to investigate the issue by looking at the PostgreSQL source,\\n> maybe inserting extra log messages.\\n> While trying to do this, I came to the conclusion that I need developer help, because I can't find the part of the source code\\n> where the ORDER BY logic is implemented.\\n>\\n> Could you please help, by telling me which files should I take a look at?\\n>\\n> Thanks in advance!\\n\\nI think you can start by looking into the function grouping_planner\\nand explore further by going through the route of it.\\n\\n-- \\nRegards,\\nRafia Sabih\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 12:47:46 +0100\",\n \"msg_from\": \"Rafia Sabih <rafia.pghackers@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ORDER BY logic in PostgreSQL source code\"\n }\n]","isConversational":false}}},{"rowIdx":53067,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Do Aggregate functions/queries reuse the state value they generate for\nsubsequent queries? If not then an explanation would be greatly\nappreciated.\n\nDo Aggregate functions/queries reuse the state value they generate for subsequent queries? If not then an explanation would be greatly appreciated.","msg_date":"Wed, 10 Nov 2021 14:44:20 +0800","msg_from":"gg pw <lucenaalmirjames123@gmail.com>","msg_from_op":true,"msg_subject":"Reuse of State value in Aggregates"},{"msg_contents":"gg pw <lucenaalmirjames123@gmail.com> wrote:\n\n> Do Aggregate functions/queries reuse the state value they generate for subsequent queries? If not then an explanation would be greatly appreciated. \n\nThe subsequent queries are not supposed to feed the same input into the\naggregates, so their aggregate states will most likely be different. Maybe you\nneed to post an example where you think such reusing would be useful.\n\n-- \nAntonin Houska\nWeb: https://www.cybertec-postgresql.com\n\n\n","msg_date":"Wed, 10 Nov 2021 08:40:12 +0100","msg_from":"Antonin Houska <ah@cybertec.at>","msg_from_op":false,"msg_subject":"Re: Reuse of State value in Aggregates"},{"msg_contents":"That's sufficient information for me thanks! I don't really have a concrete\nexample at the moment. The idea just popped up in my head when I read that\nCTEs are reused inside parent queries.\n\nThat's sufficient information for me thanks! I don't really have a concrete example at the moment. The idea just popped up in my head when I read that CTEs are reused inside parent queries.","msg_date":"Mon, 15 Nov 2021 13:50:02 +0800","msg_from":"gg pw <lucenaalmirjames123@gmail.com>","msg_from_op":true,"msg_subject":"Re: Reuse of State value in Aggregates"}],"string":"[\n {\n \"msg_contents\": \"Do Aggregate functions/queries reuse the state value they generate for\\nsubsequent queries? If not then an explanation would be greatly\\nappreciated.\\n\\nDo Aggregate functions/queries reuse the state value they generate for subsequent queries? If not then an explanation would be greatly appreciated.\",\n \"msg_date\": \"Wed, 10 Nov 2021 14:44:20 +0800\",\n \"msg_from\": \"gg pw <lucenaalmirjames123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Reuse of State value in Aggregates\"\n },\n {\n \"msg_contents\": \"gg pw <lucenaalmirjames123@gmail.com> wrote:\\n\\n> Do Aggregate functions/queries reuse the state value they generate for subsequent queries? If not then an explanation would be greatly appreciated. \\n\\nThe subsequent queries are not supposed to feed the same input into the\\naggregates, so their aggregate states will most likely be different. Maybe you\\nneed to post an example where you think such reusing would be useful.\\n\\n-- \\nAntonin Houska\\nWeb: https://www.cybertec-postgresql.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 08:40:12 +0100\",\n \"msg_from\": \"Antonin Houska <ah@cybertec.at>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Reuse of State value in Aggregates\"\n },\n {\n \"msg_contents\": \"That's sufficient information for me thanks! I don't really have a concrete\\nexample at the moment. The idea just popped up in my head when I read that\\nCTEs are reused inside parent queries.\\n\\nThat's sufficient information for me thanks! I don't really have a concrete example at the moment. The idea just popped up in my head when I read that CTEs are reused inside parent queries.\",\n \"msg_date\": \"Mon, 15 Nov 2021 13:50:02 +0800\",\n \"msg_from\": \"gg pw <lucenaalmirjames123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Reuse of State value in Aggregates\"\n }\n]","isConversational":false}}},{"rowIdx":53068,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi\n\nI think I found a bug related to logical replication(REPLICA IDENTITY in specific).\nIf I change REPLICA IDENTITY after creating publication, the DELETE/UPDATE operations won't be replicated as expected. \n\nFor example:\n-- publisher\nCREATE TABLE tbl(a int, b int);\nALTER TABLE tbl ALTER COLUMN a SET NOT NULL;\nCREATE UNIQUE INDEX idx_a ON tbl(a);\nALTER TABLE tbl ALTER COLUMN b SET NOT NULL;\nCREATE UNIQUE INDEX idx_b ON tbl(b);\nALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_a;\nCREATE PUBLICATION pub FOR TABLE tbl;\nALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_b;\nINSERT INTO tbl VALUES (1,1);\n\n-- subscriber\nCREATE TABLE tbl(a int, b int);\nALTER TABLE tbl ALTER COLUMN b SET NOT NULL;\nCREATE UNIQUE INDEX idx_b ON tbl(b);\nALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_b;\nCREATE SUBSCRIPTION sub CONNECTION 'dbname=postgres port=5432' PUBLICATION pub;\nSELECT * FROM tbl;\n\n-- publisher\npostgres=# UPDATE tbl SET a=-a;\nUPDATE 1\npostgres=# SELECT * FROM tbl;\na | b\n----+---\n -1 | 1 \n(1 row)\n\n-- subscriber\npostgres=# SELECT * FROM tbl;\n a | b\n---+---\n 1 | 1\n(1 row)\n\n(a in subscriber should be -1)\n\nBut if I restart a psql client before executing UPDATE operation in publication, it works well. \nSo I think the problem is: when using \"ALTER TABLE ... REPLICA IDENTITY USING INDEX ...\", relcahe was not invalidated.\n\nAttach a patch to fix it, I also added a test case for it.(Hou helped me to write/review this patch, which is very kind of him)\n\nFYI: I also tested that same problem could be reproduced on PG10 ~ PG14. (Logical replication is introduced in PG10.)\nSo I think backpatch is needed here.\n\nRegards\nTang","msg_date":"Wed, 10 Nov 2021 07:11:57 +0000","msg_from":"\"tanghy.fnst@fujitsu.com\" <tanghy.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"[BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Wed, Nov 10, 2021 at 12:42 PM tanghy.fnst@fujitsu.com\n<tanghy.fnst@fujitsu.com> wrote:\n>\n> Hi\n>\n> I think I found a bug related to logical replication(REPLICA IDENTITY in specific).\n> If I change REPLICA IDENTITY after creating publication, the DELETE/UPDATE operations won't be replicated as expected.\n>\n> For example:\n> -- publisher\n> CREATE TABLE tbl(a int, b int);\n> ALTER TABLE tbl ALTER COLUMN a SET NOT NULL;\n> CREATE UNIQUE INDEX idx_a ON tbl(a);\n> ALTER TABLE tbl ALTER COLUMN b SET NOT NULL;\n> CREATE UNIQUE INDEX idx_b ON tbl(b);\n> ALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_a;\n> CREATE PUBLICATION pub FOR TABLE tbl;\n> ALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_b;\n> INSERT INTO tbl VALUES (1,1);\n>\n> -- subscriber\n> CREATE TABLE tbl(a int, b int);\n> ALTER TABLE tbl ALTER COLUMN b SET NOT NULL;\n> CREATE UNIQUE INDEX idx_b ON tbl(b);\n> ALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_b;\n> CREATE SUBSCRIPTION sub CONNECTION 'dbname=postgres port=5432' PUBLICATION pub;\n> SELECT * FROM tbl;\n>\n> -- publisher\n> postgres=# UPDATE tbl SET a=-a;\n> UPDATE 1\n> postgres=# SELECT * FROM tbl;\n> a | b\n> ----+---\n> -1 | 1\n> (1 row)\n>\n> -- subscriber\n> postgres=# SELECT * FROM tbl;\n> a | b\n> ---+---\n> 1 | 1\n> (1 row)\n>\n> (a in subscriber should be -1)\n>\n\nI don't understand the purpose of idx_b in the above test case, why is\nit required to reproduce the problem?\n\n@@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation rel,\nchar ri_type, Oid indexOid,\n CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self, pg_index_tuple);\n InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\n InvalidOid, is_internal);\n+ CacheInvalidateRelcache(rel);\n\nCatalogTupleUpdate internally calls heap_update which calls\nCacheInvalidateHeapTuple(), why is that not sufficient for\ninvalidation?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 10 Nov 2021 16:58:59 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Wed, Nov 10, 2021 7:29 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> On Wed, Nov 10, 2021 at 12:42 PM tanghy.fnst@fujitsu.com\r\n> <tanghy.fnst@fujitsu.com> wrote:\r\n> >\r\n> > Hi\r\n> >\r\n> > I think I found a bug related to logical replication(REPLICA IDENTITY in\r\n> specific).\r\n> > If I change REPLICA IDENTITY after creating publication, the\r\n> DELETE/UPDATE operations won't be replicated as expected.\r\n> >\r\n> > For example:\r\n> > -- publisher\r\n> > CREATE TABLE tbl(a int, b int);\r\n> > ALTER TABLE tbl ALTER COLUMN a SET NOT NULL; CREATE UNIQUE INDEX\r\n> idx_a\r\n> > ON tbl(a); ALTER TABLE tbl ALTER COLUMN b SET NOT NULL; CREATE UNIQUE\r\n> > INDEX idx_b ON tbl(b); ALTER TABLE tbl REPLICA IDENTITY USING INDEX\r\n> > idx_a; CREATE PUBLICATION pub FOR TABLE tbl; ALTER TABLE tbl REPLICA\r\n> > IDENTITY USING INDEX idx_b; INSERT INTO tbl VALUES (1,1);\r\n> >\r\n> > -- subscriber\r\n> > CREATE TABLE tbl(a int, b int);\r\n> > ALTER TABLE tbl ALTER COLUMN b SET NOT NULL; CREATE UNIQUE INDEX\r\n> idx_b\r\n> > ON tbl(b); ALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_b; CREATE\r\n> > SUBSCRIPTION sub CONNECTION 'dbname=postgres port=5432'\r\n> PUBLICATION\r\n> > pub; SELECT * FROM tbl;\r\n> >\r\n> > -- publisher\r\n> > postgres=# UPDATE tbl SET a=-a;\r\n> > UPDATE 1\r\n> > postgres=# SELECT * FROM tbl;\r\n> > a | b\r\n> > ----+---\r\n> > -1 | 1\r\n> > (1 row)\r\n> >\r\n> > -- subscriber\r\n> > postgres=# SELECT * FROM tbl;\r\n> > a | b\r\n> > ---+---\r\n> > 1 | 1\r\n> > (1 row)\r\n> >\r\n> > (a in subscriber should be -1)\r\n> >\r\n> \r\n> I don't understand the purpose of idx_b in the above test case, why is it\r\n> required to reproduce the problem?\r\n> @@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation rel, char\r\n> ri_type, Oid indexOid,\r\n> CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self, pg_index_tuple);\r\n> InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\r\n> InvalidOid, is_internal);\r\n> + CacheInvalidateRelcache(rel);\r\n> \r\n> CatalogTupleUpdate internally calls heap_update which calls\r\n> CacheInvalidateHeapTuple(), why is that not sufficient for invalidation?\r\n\r\nI think it's because the bug happens only when changing REPLICA IDENTITY index\r\nfrom one(idx_a) to another one(idx_b).\r\n\r\nWhen first time specify the REPLICA IDENTITY index, the code works well because\r\nit will invoke 'CatalogTupleUpdate(pg_class,...' Which will invalidate the\r\ntarget table's relcache.\r\n\r\n\tif (pg_class_form->relreplident != ri_type)\r\n\t{\r\n\t\tpg_class_form->relreplident = ri_type;\r\n\t\tCatalogTupleUpdate(pg_class, &pg_class_tuple->t_self, pg_class_tuple);\r\n\t}\r\n\r\nBut when changing REPLICA IDENTITY index, the code only invoke\r\n'CatalogTupleUpdate(pg_index,' which seems only invalidate the index's cache not the\r\ntarget table.\r\n\r\n\t\tif (dirty)\r\n\t\t{\r\n\t\t\tCatalogTupleUpdate(pg_index, &pg_index_tuple->t_self, pg_index_tuple);\r\n\t\t\tInvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\r\n\t\t\t\t\t\t\t\t\t\t InvalidOid, is_internal);\r\n\t\t}\r\n\r\nBest regards,\r\nHou zj\r\n","msg_date":"Thu, 11 Nov 2021 01:37:13 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Thu, Nov 11, 2021 at 7:07 AM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n>\n> On Wed, Nov 10, 2021 7:29 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > I don't understand the purpose of idx_b in the above test case, why is it\n> > required to reproduce the problem?\n> > @@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation rel, char\n> > ri_type, Oid indexOid,\n> > CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self, pg_index_tuple);\n> > InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\n> > InvalidOid, is_internal);\n> > + CacheInvalidateRelcache(rel);\n> >\n> > CatalogTupleUpdate internally calls heap_update which calls\n> > CacheInvalidateHeapTuple(), why is that not sufficient for invalidation?\n>\n> I think it's because the bug happens only when changing REPLICA IDENTITY index\n> from one(idx_a) to another one(idx_b).\n>\n\nOkay, but do we need to invalidate the rel cache each time the dirty\nflag is set? Can't we do it once outside the foreach index loop? I\nthink we can record whether to do relcache invalidation in a new\nboolean variable need_rel_inval or something like that. Also, it is\nbetter if you can add a comment on the lines of: \"Invalidate the\nrelcache for the table so that after we commit all sessions will\nrefresh the table's replica identity index before attempting any\nupdate on the table.\".\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 11 Nov 2021 09:37:32 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Thu, Nov 11, 2021 at 9:37 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Thu, Nov 11, 2021 at 7:07 AM houzj.fnst@fujitsu.com\n> <houzj.fnst@fujitsu.com> wrote:\n> >\n> > On Wed, Nov 10, 2021 7:29 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > >\n> > > I don't understand the purpose of idx_b in the above test case, why is it\n> > > required to reproduce the problem?\n> > > @@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation rel, char\n> > > ri_type, Oid indexOid,\n> > > CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self, pg_index_tuple);\n> > > InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\n> > > InvalidOid, is_internal);\n> > > + CacheInvalidateRelcache(rel);\n> > >\n> > > CatalogTupleUpdate internally calls heap_update which calls\n> > > CacheInvalidateHeapTuple(), why is that not sufficient for invalidation?\n> >\n> > I think it's because the bug happens only when changing REPLICA IDENTITY index\n> > from one(idx_a) to another one(idx_b).\n> >\n>\n> Okay, but do we need to invalidate the rel cache each time the dirty\n> flag is set? Can't we do it once outside the foreach index loop? I\n> think we can record whether to do relcache invalidation in a new\n> boolean variable need_rel_inval or something like that. Also, it is\n> better if you can add a comment on the lines of: \"Invalidate the\n> relcache for the table so that after we commit all sessions will\n> refresh the table's replica identity index before attempting any\n> update on the table.\".\n>\n\nFew comments on testcase added by the patch:\n==================================\n1.\n+$node_subscriber->safe_psql('postgres',\n+);\n\nIt is not clear what is the purpose served by this statement.\n\n2.\n+# replication of the table without replica identity index but not primary key\n+$node_subscriber->safe_psql('postgres',\n+ \"CREATE TABLE tab_replidentity_index(a int not null, b int not null)\");\n\nI think the part of the comment \"but not primary key\" is not required.\n\n3. Can we move this test to subscription/t/100_bugs?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 11 Nov 2021 09:59:15 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Thur, Nov 11, 2021 12:08 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> On Thu, Nov 11, 2021 at 7:07 AM houzj.fnst@fujitsu.com <houzj.fnst@fujitsu.com> wrote:\r\n> >\r\n> > On Wed, Nov 10, 2021 7:29 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> > >\r\n> > > I don't understand the purpose of idx_b in the above test case, why is it\r\n> > > required to reproduce the problem?\r\n> > > @@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation rel,\r\n> char\r\n> > > ri_type, Oid indexOid,\r\n> > > CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self,\r\n> pg_index_tuple);\r\n> > > InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\r\n> > > InvalidOid, is_internal);\r\n> > > + CacheInvalidateRelcache(rel);\r\n> > >\r\n> > > CatalogTupleUpdate internally calls heap_update which calls\r\n> > > CacheInvalidateHeapTuple(), why is that not sufficient for invalidation?\r\n> >\r\n> > I think it's because the bug happens only when changing REPLICA IDENTITY\r\n> index\r\n> > from one(idx_a) to another one(idx_b).\r\n> >\r\n> \r\n> Okay, but do we need to invalidate the rel cache each time the dirty\r\n> flag is set? Can't we do it once outside the foreach index loop? I\r\n> think we can record whether to do relcache invalidation in a new\r\n> boolean variable need_rel_inval or something like that. Also, it is\r\n> better if you can add a comment on the lines of: \"Invalidate the\r\n> relcache for the table so that after we commit all sessions will\r\n> refresh the table's replica identity index before attempting any\r\n> update on the table.\".\r\n\r\nI agree.\r\n\r\nAttach the new version fix patch which move the invalidation out of the loop\r\nand addressed the comments[1] for the testcases.\r\n\r\n[1] https://www.postgresql.org/message-id/CAA4eK1Ki-kUx52uRER3bZSC6bLc%3Diix%2BQnBxs1pZSHHYZOEF1A%40mail.gmail.com\r\n\r\nBest regards,\r\nHou zj","msg_date":"Thu, 11 Nov 2021 09:35:32 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Thursday, November 11, 2021 5:36 PM houzj.fnst@fujitsu.com wrote:\r\n> On Thur, Nov 11, 2021 12:08 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> > On Thu, Nov 11, 2021 at 7:07 AM houzj.fnst@fujitsu.com\r\n> <houzj.fnst@fujitsu.com> wrote:\r\n> > >\r\n> > > On Wed, Nov 10, 2021 7:29 PM Amit Kapila <amit.kapila16@gmail.com>\r\n> wrote:\r\n> > > >\r\n> > > > I don't understand the purpose of idx_b in the above test case,\r\n> > > > why is it required to reproduce the problem?\r\n> > > > @@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation\r\n> > > > rel,\r\n> > char\r\n> > > > ri_type, Oid indexOid,\r\n> > > > CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self,\r\n> > pg_index_tuple);\r\n> > > > InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\r\n> > > > InvalidOid, is_internal);\r\n> > > > + CacheInvalidateRelcache(rel);\r\n> > > >\r\n> > > > CatalogTupleUpdate internally calls heap_update which calls\r\n> > > > CacheInvalidateHeapTuple(), why is that not sufficient for invalidation?\r\n> > >\r\n> > > I think it's because the bug happens only when changing REPLICA\r\n> > > IDENTITY\r\n> > index\r\n> > > from one(idx_a) to another one(idx_b).\r\n> > >\r\n> >\r\n> > Okay, but do we need to invalidate the rel cache each time the dirty\r\n> > flag is set? Can't we do it once outside the foreach index loop? I\r\n> > think we can record whether to do relcache invalidation in a new\r\n> > boolean variable need_rel_inval or something like that. Also, it is\r\n> > better if you can add a comment on the lines of: \"Invalidate the\r\n> > relcache for the table so that after we commit all sessions will\r\n> > refresh the table's replica identity index before attempting any\r\n> > update on the table.\".\r\n> \r\n> I agree.\r\n> \r\n> Attach the new version fix patch which move the invalidation out of the loop and\r\n> addressed the comments[1] for the testcases.\r\n\r\nAlso attach the patches for back branch and remove some unnecessary\r\nchanges from pgindent.\r\n\r\nBest regards,\r\nHou zj","msg_date":"Thu, 11 Nov 2021 12:01:33 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Thu, Nov 11, 2021, at 9:01 AM, houzj.fnst@fujitsu.com wrote:\n> Also attach the patches for back branch and remove some unnecessary\n> changes from pgindent.\nI reviewed your patch and I think the fix could be simplified by\n\nif (OidIsValid(indexOid))\nCacheInvalidateRelcache(rel);\n\nIf indexOid is valid it is a REPLICA IDENTITY INDEX. A few lines above there is\na check for a valid indexOid that makes sure the index is already marked as a\nreplica identity; if so, it bail out. If it is not, the relation should be\ninvalidated. Am I missing something?\n\nI also modified your test case to include a DELETE command, wait the initial\ntable sync to avoid failing a subsequent test and improve some comments.\n\n\n--\nEuler Taveira\nEDB https://www.enterprisedb.com/","msg_date":"Thu, 11 Nov 2021 21:14:45 -0300","msg_from":"\"Euler Taveira\" <euler@eulerto.com>","msg_from_op":false,"msg_subject":"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Friday, November 12, 2021 8:15 AM Euler Taveira <euler@eulerto.com> wrote:\n> I reviewed your patch and I think the fix could be simplified by\n> \n> if (OidIsValid(indexOid))\n> CacheInvalidateRelcache(rel);\n> \n> If indexOid is valid it is a REPLICA IDENTITY INDEX. A few lines above there is\n> a check for a valid indexOid that makes sure the index is already marked as a\n> replica identity; if so, it bail out. If it is not, the relation should be\n> invalidated. Am I missing something?\n\nThanks for reviewing !\nBut I am not sure it's better to simplify the code like\n\"if (OidIsValid(indexOid)) CacheInvalidate\". After this change, it will\nalso invalidate the relcache when changing REPLICA IDENTITY from\nDEFAULT/FULL/NOTHING to USING INDEX which I think is unnecessary, because the\ninvalidate for these cases was already handled by\n\"CatalogTupleUpdate(pg_class\".\n\nSo, I still think the flag need_rel_inval is needed.\n\n> I also modified your test case to include a DELETE command, wait the initial\n> table sync to avoid failing a subsequent test and improve some comments.\nThanks!\n\nAttach the patch which changed the code use need_rel_inval flag.\nAlso ran pgperltidy for the testcases. I will post patches for backbranch\nif there are no more comments.\n\nBest regards,\nHou zj","msg_date":"Fri, 12 Nov 2021 02:45:57 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Friday, November 12, 2021 10:46 AM I wrote:\n> On Friday, November 12, 2021 8:15 AM Euler Taveira <euler@eulerto.com>\n> wrote:\n> > I reviewed your patch and I think the fix could be simplified by\n> >\n> > if (OidIsValid(indexOid))\n> > CacheInvalidateRelcache(rel);\n> >\n> > If indexOid is valid it is a REPLICA IDENTITY INDEX. A few lines above\n> > there is a check for a valid indexOid that makes sure the index is\n> > already marked as a replica identity; if so, it bail out. If it is\n> > not, the relation should be invalidated. Am I missing something?\n> \n> Thanks for reviewing !\n> But I am not sure it's better to simplify the code like \"if (OidIsValid(indexOid))\n> CacheInvalidate\". \n\nOh, I got confused with the logic in relation_mark_replica_identity, sorry for that.\nI now realize that you are right, we can just check \"if (OidIsValid(indexOid))\" here\nto simplify the code.\n\nAttach the new version patch which simplify the check as pointed out by Euler.\n\nBest regards,\nHou zj","msg_date":"Fri, 12 Nov 2021 04:57:33 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Fri, Nov 12, 2021 at 10:27 AM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n>\n> On Friday, November 12, 2021 10:46 AM I wrote:\n> > On Friday, November 12, 2021 8:15 AM Euler Taveira <euler@eulerto.com>\n> > wrote:\n> > > I reviewed your patch and I think the fix could be simplified by\n> > >\n> > > if (OidIsValid(indexOid))\n> > > CacheInvalidateRelcache(rel);\n> > >\n> > > If indexOid is valid it is a REPLICA IDENTITY INDEX. A few lines above\n> > > there is a check for a valid indexOid that makes sure the index is\n> > > already marked as a replica identity; if so, it bail out. If it is\n> > > not, the relation should be invalidated. Am I missing something?\n> >\n> > Thanks for reviewing !\n> > But I am not sure it's better to simplify the code like \"if (OidIsValid(indexOid))\n> > CacheInvalidate\".\n>\n> Oh, I got confused with the logic in relation_mark_replica_identity, sorry for that.\n> I now realize that you are right, we can just check \"if (OidIsValid(indexOid))\" here\n> to simplify the code.\n>\n\nBut won't that generate invalidation for the rel twice in the case\n(change Replica Identity from Nothing to some index) you mentioned in\nthe previous email?\n\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 12 Nov 2021 10:50:14 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Fri, Nov 12, 2021 at 10:50 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Fri, Nov 12, 2021 at 10:27 AM houzj.fnst@fujitsu.com\n> <houzj.fnst@fujitsu.com> wrote:\n> >\n> > On Friday, November 12, 2021 10:46 AM I wrote:\n> > > On Friday, November 12, 2021 8:15 AM Euler Taveira <euler@eulerto.com>\n> > > wrote:\n> > > > I reviewed your patch and I think the fix could be simplified by\n> > > >\n> > > > if (OidIsValid(indexOid))\n> > > > CacheInvalidateRelcache(rel);\n> > > >\n> > > > If indexOid is valid it is a REPLICA IDENTITY INDEX. A few lines above\n> > > > there is a check for a valid indexOid that makes sure the index is\n> > > > already marked as a replica identity; if so, it bail out. If it is\n> > > > not, the relation should be invalidated. Am I missing something?\n> > >\n> > > Thanks for reviewing !\n> > > But I am not sure it's better to simplify the code like \"if (OidIsValid(indexOid))\n> > > CacheInvalidate\".\n> >\n> > Oh, I got confused with the logic in relation_mark_replica_identity, sorry for that.\n> > I now realize that you are right, we can just check \"if (OidIsValid(indexOid))\" here\n> > to simplify the code.\n> >\n>\n> But won't that generate invalidation for the rel twice in the case\n> (change Replica Identity from Nothing to some index) you mentioned in\n> the previous email?\n>\n\nOh, I see the point. I think this is okay because\nAddRelcacheInvalidationMessage doesn't allow to add duplicate rel\ninvalidation. If that is the case I wonder why not simply register\ninvalidation without any check in the for loop as was the case with\nTang's original patch?\n\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 12 Nov 2021 11:02:55 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Fri, Nov 12, 2021 1:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> On Fri, Nov 12, 2021 at 10:50 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> > But won't that generate invalidation for the rel twice in the case\r\n> > (change Replica Identity from Nothing to some index) you mentioned in\r\n> > the previous email?\r\n> >\r\n> \r\n> Oh, I see the point. I think this is okay because\r\n> AddRelcacheInvalidationMessage doesn't allow to add duplicate rel\r\n> invalidation. If that is the case I wonder why not simply register\r\n> invalidation without any check in the for loop as was the case with\r\n> Tang's original patch?\r\n\r\nOK, I also think the code in Tang's original patch is fine.\r\nAttach the patch which register invalidation without any check in the for loop.\r\n\r\nBest regards,\r\nHou zj","msg_date":"Fri, 12 Nov 2021 06:10:07 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Fri, Nov 12, 2021, at 3:10 AM, houzj.fnst@fujitsu.com wrote:\n> On Fri, Nov 12, 2021 1:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Fri, Nov 12, 2021 at 10:50 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > But won't that generate invalidation for the rel twice in the case\n> > > (change Replica Identity from Nothing to some index) you mentioned in\n> > > the previous email?\n> > >\n> > \n> > Oh, I see the point. I think this is okay because\n> > AddRelcacheInvalidationMessage doesn't allow to add duplicate rel\n> > invalidation. If that is the case I wonder why not simply register\n> > invalidation without any check in the for loop as was the case with\n> > Tang's original patch?\n> \n> OK, I also think the code in Tang's original patch is fine.\n> Attach the patch which register invalidation without any check in the for loop.\nWFM.\n\n\n--\nEuler Taveira\nEDB https://www.enterprisedb.com/\n\nOn Fri, Nov 12, 2021, at 3:10 AM, houzj.fnst@fujitsu.com wrote:On Fri, Nov 12, 2021 1:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:> On Fri, Nov 12, 2021 at 10:50 AM Amit Kapila <amit.kapila16@gmail.com> wrote:> > But won't that generate invalidation for the rel twice in the case> > (change Replica Identity from Nothing to some index) you mentioned in> > the previous email?> >> > Oh, I see the point. I think this is okay because> AddRelcacheInvalidationMessage doesn't allow to add duplicate rel> invalidation. If that is the case I wonder why not simply register> invalidation without any check in the for loop as was the case with> Tang's original patch?OK, I also think the code in Tang's original patch is fine.Attach the patch which register invalidation without any check in the for loop.WFM.--Euler TaveiraEDB https://www.enterprisedb.com/","msg_date":"Fri, 12 Nov 2021 16:17:16 -0300","msg_from":"\"Euler Taveira\" <euler@eulerto.com>","msg_from_op":false,"msg_subject":"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Sat, Nov 13, 2021 at 12:47 AM Euler Taveira <euler@eulerto.com> wrote:\n>\n> On Fri, Nov 12, 2021, at 3:10 AM, houzj.fnst@fujitsu.com wrote:\n>\n> On Fri, Nov 12, 2021 1:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Fri, Nov 12, 2021 at 10:50 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > But won't that generate invalidation for the rel twice in the case\n> > > (change Replica Identity from Nothing to some index) you mentioned in\n> > > the previous email?\n> > >\n> >\n> > Oh, I see the point. I think this is okay because\n> > AddRelcacheInvalidationMessage doesn't allow to add duplicate rel\n> > invalidation. If that is the case I wonder why not simply register\n> > invalidation without any check in the for loop as was the case with\n> > Tang's original patch?\n>\n> OK, I also think the code in Tang's original patch is fine.\n> Attach the patch which register invalidation without any check in the for loop.\n>\n> WFM.\n>\n\nThe patch looks mostly good to me. I have slightly tweaked the\ncomments in the code (as per my previous suggestion) and test. Also, I\nhave slightly modified the commit message. If the attached looks good\nto you then kindly prepare patches for back-branches.\n\n-- \nWith Regards,\nAmit Kapila.","msg_date":"Sat, 13 Nov 2021 16:19:41 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Sat, Nov 13, 2021 6:50 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> \r\n> The patch looks mostly good to me. I have slightly tweaked the comments in\r\n> the code (as per my previous suggestion) and test. Also, I have slightly\r\n> modified the commit message. If the attached looks good to you then kindly\r\n> prepare patches for back-branches.\r\n\r\nThanks for reviewing, the latest patch looks good to me.\r\nAttach the patches for back branch (HEAD ~ v10).\r\n\r\nBest regards,\r\nHou zj","msg_date":"Mon, 15 Nov 2021 01:50:11 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY"},{"msg_contents":"On Mon, Nov 15, 2021 at 7:20 AM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n>\n> On Sat, Nov 13, 2021 6:50 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > The patch looks mostly good to me. I have slightly tweaked the comments in\n> > the code (as per my previous suggestion) and test. Also, I have slightly\n> > modified the commit message. If the attached looks good to you then kindly\n> > prepare patches for back-branches.\n>\n> Thanks for reviewing, the latest patch looks good to me.\n> Attach the patches for back branch (HEAD ~ v10).\n>\n\nPushed.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 16 Nov 2021 17:27:46 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY"}],"string":"[\n {\n \"msg_contents\": \"Hi\\n\\nI think I found a bug related to logical replication(REPLICA IDENTITY in specific).\\nIf I change REPLICA IDENTITY after creating publication, the DELETE/UPDATE operations won't be replicated as expected. \\n\\nFor example:\\n-- publisher\\nCREATE TABLE tbl(a int, b int);\\nALTER TABLE tbl ALTER COLUMN a SET NOT NULL;\\nCREATE UNIQUE INDEX idx_a ON tbl(a);\\nALTER TABLE tbl ALTER COLUMN b SET NOT NULL;\\nCREATE UNIQUE INDEX idx_b ON tbl(b);\\nALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_a;\\nCREATE PUBLICATION pub FOR TABLE tbl;\\nALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_b;\\nINSERT INTO tbl VALUES (1,1);\\n\\n-- subscriber\\nCREATE TABLE tbl(a int, b int);\\nALTER TABLE tbl ALTER COLUMN b SET NOT NULL;\\nCREATE UNIQUE INDEX idx_b ON tbl(b);\\nALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_b;\\nCREATE SUBSCRIPTION sub CONNECTION 'dbname=postgres port=5432' PUBLICATION pub;\\nSELECT * FROM tbl;\\n\\n-- publisher\\npostgres=# UPDATE tbl SET a=-a;\\nUPDATE 1\\npostgres=# SELECT * FROM tbl;\\na | b\\n----+---\\n -1 | 1 \\n(1 row)\\n\\n-- subscriber\\npostgres=# SELECT * FROM tbl;\\n a | b\\n---+---\\n 1 | 1\\n(1 row)\\n\\n(a in subscriber should be -1)\\n\\nBut if I restart a psql client before executing UPDATE operation in publication, it works well. \\nSo I think the problem is: when using \\\"ALTER TABLE ... REPLICA IDENTITY USING INDEX ...\\\", relcahe was not invalidated.\\n\\nAttach a patch to fix it, I also added a test case for it.(Hou helped me to write/review this patch, which is very kind of him)\\n\\nFYI: I also tested that same problem could be reproduced on PG10 ~ PG14. (Logical replication is introduced in PG10.)\\nSo I think backpatch is needed here.\\n\\nRegards\\nTang\",\n \"msg_date\": \"Wed, 10 Nov 2021 07:11:57 +0000\",\n \"msg_from\": \"\\\"tanghy.fnst@fujitsu.com\\\" <tanghy.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 10, 2021 at 12:42 PM tanghy.fnst@fujitsu.com\\n<tanghy.fnst@fujitsu.com> wrote:\\n>\\n> Hi\\n>\\n> I think I found a bug related to logical replication(REPLICA IDENTITY in specific).\\n> If I change REPLICA IDENTITY after creating publication, the DELETE/UPDATE operations won't be replicated as expected.\\n>\\n> For example:\\n> -- publisher\\n> CREATE TABLE tbl(a int, b int);\\n> ALTER TABLE tbl ALTER COLUMN a SET NOT NULL;\\n> CREATE UNIQUE INDEX idx_a ON tbl(a);\\n> ALTER TABLE tbl ALTER COLUMN b SET NOT NULL;\\n> CREATE UNIQUE INDEX idx_b ON tbl(b);\\n> ALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_a;\\n> CREATE PUBLICATION pub FOR TABLE tbl;\\n> ALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_b;\\n> INSERT INTO tbl VALUES (1,1);\\n>\\n> -- subscriber\\n> CREATE TABLE tbl(a int, b int);\\n> ALTER TABLE tbl ALTER COLUMN b SET NOT NULL;\\n> CREATE UNIQUE INDEX idx_b ON tbl(b);\\n> ALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_b;\\n> CREATE SUBSCRIPTION sub CONNECTION 'dbname=postgres port=5432' PUBLICATION pub;\\n> SELECT * FROM tbl;\\n>\\n> -- publisher\\n> postgres=# UPDATE tbl SET a=-a;\\n> UPDATE 1\\n> postgres=# SELECT * FROM tbl;\\n> a | b\\n> ----+---\\n> -1 | 1\\n> (1 row)\\n>\\n> -- subscriber\\n> postgres=# SELECT * FROM tbl;\\n> a | b\\n> ---+---\\n> 1 | 1\\n> (1 row)\\n>\\n> (a in subscriber should be -1)\\n>\\n\\nI don't understand the purpose of idx_b in the above test case, why is\\nit required to reproduce the problem?\\n\\n@@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation rel,\\nchar ri_type, Oid indexOid,\\n CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self, pg_index_tuple);\\n InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\\n InvalidOid, is_internal);\\n+ CacheInvalidateRelcache(rel);\\n\\nCatalogTupleUpdate internally calls heap_update which calls\\nCacheInvalidateHeapTuple(), why is that not sufficient for\\ninvalidation?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 16:58:59 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 10, 2021 7:29 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> On Wed, Nov 10, 2021 at 12:42 PM tanghy.fnst@fujitsu.com\\r\\n> <tanghy.fnst@fujitsu.com> wrote:\\r\\n> >\\r\\n> > Hi\\r\\n> >\\r\\n> > I think I found a bug related to logical replication(REPLICA IDENTITY in\\r\\n> specific).\\r\\n> > If I change REPLICA IDENTITY after creating publication, the\\r\\n> DELETE/UPDATE operations won't be replicated as expected.\\r\\n> >\\r\\n> > For example:\\r\\n> > -- publisher\\r\\n> > CREATE TABLE tbl(a int, b int);\\r\\n> > ALTER TABLE tbl ALTER COLUMN a SET NOT NULL; CREATE UNIQUE INDEX\\r\\n> idx_a\\r\\n> > ON tbl(a); ALTER TABLE tbl ALTER COLUMN b SET NOT NULL; CREATE UNIQUE\\r\\n> > INDEX idx_b ON tbl(b); ALTER TABLE tbl REPLICA IDENTITY USING INDEX\\r\\n> > idx_a; CREATE PUBLICATION pub FOR TABLE tbl; ALTER TABLE tbl REPLICA\\r\\n> > IDENTITY USING INDEX idx_b; INSERT INTO tbl VALUES (1,1);\\r\\n> >\\r\\n> > -- subscriber\\r\\n> > CREATE TABLE tbl(a int, b int);\\r\\n> > ALTER TABLE tbl ALTER COLUMN b SET NOT NULL; CREATE UNIQUE INDEX\\r\\n> idx_b\\r\\n> > ON tbl(b); ALTER TABLE tbl REPLICA IDENTITY USING INDEX idx_b; CREATE\\r\\n> > SUBSCRIPTION sub CONNECTION 'dbname=postgres port=5432'\\r\\n> PUBLICATION\\r\\n> > pub; SELECT * FROM tbl;\\r\\n> >\\r\\n> > -- publisher\\r\\n> > postgres=# UPDATE tbl SET a=-a;\\r\\n> > UPDATE 1\\r\\n> > postgres=# SELECT * FROM tbl;\\r\\n> > a | b\\r\\n> > ----+---\\r\\n> > -1 | 1\\r\\n> > (1 row)\\r\\n> >\\r\\n> > -- subscriber\\r\\n> > postgres=# SELECT * FROM tbl;\\r\\n> > a | b\\r\\n> > ---+---\\r\\n> > 1 | 1\\r\\n> > (1 row)\\r\\n> >\\r\\n> > (a in subscriber should be -1)\\r\\n> >\\r\\n> \\r\\n> I don't understand the purpose of idx_b in the above test case, why is it\\r\\n> required to reproduce the problem?\\r\\n> @@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation rel, char\\r\\n> ri_type, Oid indexOid,\\r\\n> CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self, pg_index_tuple);\\r\\n> InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\\r\\n> InvalidOid, is_internal);\\r\\n> + CacheInvalidateRelcache(rel);\\r\\n> \\r\\n> CatalogTupleUpdate internally calls heap_update which calls\\r\\n> CacheInvalidateHeapTuple(), why is that not sufficient for invalidation?\\r\\n\\r\\nI think it's because the bug happens only when changing REPLICA IDENTITY index\\r\\nfrom one(idx_a) to another one(idx_b).\\r\\n\\r\\nWhen first time specify the REPLICA IDENTITY index, the code works well because\\r\\nit will invoke 'CatalogTupleUpdate(pg_class,...' Which will invalidate the\\r\\ntarget table's relcache.\\r\\n\\r\\n\\tif (pg_class_form->relreplident != ri_type)\\r\\n\\t{\\r\\n\\t\\tpg_class_form->relreplident = ri_type;\\r\\n\\t\\tCatalogTupleUpdate(pg_class, &pg_class_tuple->t_self, pg_class_tuple);\\r\\n\\t}\\r\\n\\r\\nBut when changing REPLICA IDENTITY index, the code only invoke\\r\\n'CatalogTupleUpdate(pg_index,' which seems only invalidate the index's cache not the\\r\\ntarget table.\\r\\n\\r\\n\\t\\tif (dirty)\\r\\n\\t\\t{\\r\\n\\t\\t\\tCatalogTupleUpdate(pg_index, &pg_index_tuple->t_self, pg_index_tuple);\\r\\n\\t\\t\\tInvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\\r\\n\\t\\t\\t\\t\\t\\t\\t\\t\\t\\t InvalidOid, is_internal);\\r\\n\\t\\t}\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 01:37:13 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 7:07 AM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n>\\n> On Wed, Nov 10, 2021 7:29 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > I don't understand the purpose of idx_b in the above test case, why is it\\n> > required to reproduce the problem?\\n> > @@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation rel, char\\n> > ri_type, Oid indexOid,\\n> > CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self, pg_index_tuple);\\n> > InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\\n> > InvalidOid, is_internal);\\n> > + CacheInvalidateRelcache(rel);\\n> >\\n> > CatalogTupleUpdate internally calls heap_update which calls\\n> > CacheInvalidateHeapTuple(), why is that not sufficient for invalidation?\\n>\\n> I think it's because the bug happens only when changing REPLICA IDENTITY index\\n> from one(idx_a) to another one(idx_b).\\n>\\n\\nOkay, but do we need to invalidate the rel cache each time the dirty\\nflag is set? Can't we do it once outside the foreach index loop? I\\nthink we can record whether to do relcache invalidation in a new\\nboolean variable need_rel_inval or something like that. Also, it is\\nbetter if you can add a comment on the lines of: \\\"Invalidate the\\nrelcache for the table so that after we commit all sessions will\\nrefresh the table's replica identity index before attempting any\\nupdate on the table.\\\".\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 09:37:32 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 9:37 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Thu, Nov 11, 2021 at 7:07 AM houzj.fnst@fujitsu.com\\n> <houzj.fnst@fujitsu.com> wrote:\\n> >\\n> > On Wed, Nov 10, 2021 7:29 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > >\\n> > > I don't understand the purpose of idx_b in the above test case, why is it\\n> > > required to reproduce the problem?\\n> > > @@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation rel, char\\n> > > ri_type, Oid indexOid,\\n> > > CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self, pg_index_tuple);\\n> > > InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\\n> > > InvalidOid, is_internal);\\n> > > + CacheInvalidateRelcache(rel);\\n> > >\\n> > > CatalogTupleUpdate internally calls heap_update which calls\\n> > > CacheInvalidateHeapTuple(), why is that not sufficient for invalidation?\\n> >\\n> > I think it's because the bug happens only when changing REPLICA IDENTITY index\\n> > from one(idx_a) to another one(idx_b).\\n> >\\n>\\n> Okay, but do we need to invalidate the rel cache each time the dirty\\n> flag is set? Can't we do it once outside the foreach index loop? I\\n> think we can record whether to do relcache invalidation in a new\\n> boolean variable need_rel_inval or something like that. Also, it is\\n> better if you can add a comment on the lines of: \\\"Invalidate the\\n> relcache for the table so that after we commit all sessions will\\n> refresh the table's replica identity index before attempting any\\n> update on the table.\\\".\\n>\\n\\nFew comments on testcase added by the patch:\\n==================================\\n1.\\n+$node_subscriber->safe_psql('postgres',\\n+);\\n\\nIt is not clear what is the purpose served by this statement.\\n\\n2.\\n+# replication of the table without replica identity index but not primary key\\n+$node_subscriber->safe_psql('postgres',\\n+ \\\"CREATE TABLE tab_replidentity_index(a int not null, b int not null)\\\");\\n\\nI think the part of the comment \\\"but not primary key\\\" is not required.\\n\\n3. Can we move this test to subscription/t/100_bugs?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 09:59:15 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Thur, Nov 11, 2021 12:08 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> On Thu, Nov 11, 2021 at 7:07 AM houzj.fnst@fujitsu.com <houzj.fnst@fujitsu.com> wrote:\\r\\n> >\\r\\n> > On Wed, Nov 10, 2021 7:29 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> > >\\r\\n> > > I don't understand the purpose of idx_b in the above test case, why is it\\r\\n> > > required to reproduce the problem?\\r\\n> > > @@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation rel,\\r\\n> char\\r\\n> > > ri_type, Oid indexOid,\\r\\n> > > CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self,\\r\\n> pg_index_tuple);\\r\\n> > > InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\\r\\n> > > InvalidOid, is_internal);\\r\\n> > > + CacheInvalidateRelcache(rel);\\r\\n> > >\\r\\n> > > CatalogTupleUpdate internally calls heap_update which calls\\r\\n> > > CacheInvalidateHeapTuple(), why is that not sufficient for invalidation?\\r\\n> >\\r\\n> > I think it's because the bug happens only when changing REPLICA IDENTITY\\r\\n> index\\r\\n> > from one(idx_a) to another one(idx_b).\\r\\n> >\\r\\n> \\r\\n> Okay, but do we need to invalidate the rel cache each time the dirty\\r\\n> flag is set? Can't we do it once outside the foreach index loop? I\\r\\n> think we can record whether to do relcache invalidation in a new\\r\\n> boolean variable need_rel_inval or something like that. Also, it is\\r\\n> better if you can add a comment on the lines of: \\\"Invalidate the\\r\\n> relcache for the table so that after we commit all sessions will\\r\\n> refresh the table's replica identity index before attempting any\\r\\n> update on the table.\\\".\\r\\n\\r\\nI agree.\\r\\n\\r\\nAttach the new version fix patch which move the invalidation out of the loop\\r\\nand addressed the comments[1] for the testcases.\\r\\n\\r\\n[1] https://www.postgresql.org/message-id/CAA4eK1Ki-kUx52uRER3bZSC6bLc%3Diix%2BQnBxs1pZSHHYZOEF1A%40mail.gmail.com\\r\\n\\r\\nBest regards,\\r\\nHou zj\",\n \"msg_date\": \"Thu, 11 Nov 2021 09:35:32 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Thursday, November 11, 2021 5:36 PM houzj.fnst@fujitsu.com wrote:\\r\\n> On Thur, Nov 11, 2021 12:08 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> > On Thu, Nov 11, 2021 at 7:07 AM houzj.fnst@fujitsu.com\\r\\n> <houzj.fnst@fujitsu.com> wrote:\\r\\n> > >\\r\\n> > > On Wed, Nov 10, 2021 7:29 PM Amit Kapila <amit.kapila16@gmail.com>\\r\\n> wrote:\\r\\n> > > >\\r\\n> > > > I don't understand the purpose of idx_b in the above test case,\\r\\n> > > > why is it required to reproduce the problem?\\r\\n> > > > @@ -15488,6 +15488,7 @@ relation_mark_replica_identity(Relation\\r\\n> > > > rel,\\r\\n> > char\\r\\n> > > > ri_type, Oid indexOid,\\r\\n> > > > CatalogTupleUpdate(pg_index, &pg_index_tuple->t_self,\\r\\n> > pg_index_tuple);\\r\\n> > > > InvokeObjectPostAlterHookArg(IndexRelationId, thisIndexOid, 0,\\r\\n> > > > InvalidOid, is_internal);\\r\\n> > > > + CacheInvalidateRelcache(rel);\\r\\n> > > >\\r\\n> > > > CatalogTupleUpdate internally calls heap_update which calls\\r\\n> > > > CacheInvalidateHeapTuple(), why is that not sufficient for invalidation?\\r\\n> > >\\r\\n> > > I think it's because the bug happens only when changing REPLICA\\r\\n> > > IDENTITY\\r\\n> > index\\r\\n> > > from one(idx_a) to another one(idx_b).\\r\\n> > >\\r\\n> >\\r\\n> > Okay, but do we need to invalidate the rel cache each time the dirty\\r\\n> > flag is set? Can't we do it once outside the foreach index loop? I\\r\\n> > think we can record whether to do relcache invalidation in a new\\r\\n> > boolean variable need_rel_inval or something like that. Also, it is\\r\\n> > better if you can add a comment on the lines of: \\\"Invalidate the\\r\\n> > relcache for the table so that after we commit all sessions will\\r\\n> > refresh the table's replica identity index before attempting any\\r\\n> > update on the table.\\\".\\r\\n> \\r\\n> I agree.\\r\\n> \\r\\n> Attach the new version fix patch which move the invalidation out of the loop and\\r\\n> addressed the comments[1] for the testcases.\\r\\n\\r\\nAlso attach the patches for back branch and remove some unnecessary\\r\\nchanges from pgindent.\\r\\n\\r\\nBest regards,\\r\\nHou zj\",\n \"msg_date\": \"Thu, 11 Nov 2021 12:01:33 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021, at 9:01 AM, houzj.fnst@fujitsu.com wrote:\\n> Also attach the patches for back branch and remove some unnecessary\\n> changes from pgindent.\\nI reviewed your patch and I think the fix could be simplified by\\n\\nif (OidIsValid(indexOid))\\nCacheInvalidateRelcache(rel);\\n\\nIf indexOid is valid it is a REPLICA IDENTITY INDEX. A few lines above there is\\na check for a valid indexOid that makes sure the index is already marked as a\\nreplica identity; if so, it bail out. If it is not, the relation should be\\ninvalidated. Am I missing something?\\n\\nI also modified your test case to include a DELETE command, wait the initial\\ntable sync to avoid failing a subsequent test and improve some comments.\\n\\n\\n--\\nEuler Taveira\\nEDB https://www.enterprisedb.com/\",\n \"msg_date\": \"Thu, 11 Nov 2021 21:14:45 -0300\",\n \"msg_from\": \"\\\"Euler Taveira\\\" <euler@eulerto.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Friday, November 12, 2021 8:15 AM Euler Taveira <euler@eulerto.com> wrote:\\n> I reviewed your patch and I think the fix could be simplified by\\n> \\n> if (OidIsValid(indexOid))\\n> CacheInvalidateRelcache(rel);\\n> \\n> If indexOid is valid it is a REPLICA IDENTITY INDEX. A few lines above there is\\n> a check for a valid indexOid that makes sure the index is already marked as a\\n> replica identity; if so, it bail out. If it is not, the relation should be\\n> invalidated. Am I missing something?\\n\\nThanks for reviewing !\\nBut I am not sure it's better to simplify the code like\\n\\\"if (OidIsValid(indexOid)) CacheInvalidate\\\". After this change, it will\\nalso invalidate the relcache when changing REPLICA IDENTITY from\\nDEFAULT/FULL/NOTHING to USING INDEX which I think is unnecessary, because the\\ninvalidate for these cases was already handled by\\n\\\"CatalogTupleUpdate(pg_class\\\".\\n\\nSo, I still think the flag need_rel_inval is needed.\\n\\n> I also modified your test case to include a DELETE command, wait the initial\\n> table sync to avoid failing a subsequent test and improve some comments.\\nThanks!\\n\\nAttach the patch which changed the code use need_rel_inval flag.\\nAlso ran pgperltidy for the testcases. I will post patches for backbranch\\nif there are no more comments.\\n\\nBest regards,\\nHou zj\",\n \"msg_date\": \"Fri, 12 Nov 2021 02:45:57 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Friday, November 12, 2021 10:46 AM I wrote:\\n> On Friday, November 12, 2021 8:15 AM Euler Taveira <euler@eulerto.com>\\n> wrote:\\n> > I reviewed your patch and I think the fix could be simplified by\\n> >\\n> > if (OidIsValid(indexOid))\\n> > CacheInvalidateRelcache(rel);\\n> >\\n> > If indexOid is valid it is a REPLICA IDENTITY INDEX. A few lines above\\n> > there is a check for a valid indexOid that makes sure the index is\\n> > already marked as a replica identity; if so, it bail out. If it is\\n> > not, the relation should be invalidated. Am I missing something?\\n> \\n> Thanks for reviewing !\\n> But I am not sure it's better to simplify the code like \\\"if (OidIsValid(indexOid))\\n> CacheInvalidate\\\". \\n\\nOh, I got confused with the logic in relation_mark_replica_identity, sorry for that.\\nI now realize that you are right, we can just check \\\"if (OidIsValid(indexOid))\\\" here\\nto simplify the code.\\n\\nAttach the new version patch which simplify the check as pointed out by Euler.\\n\\nBest regards,\\nHou zj\",\n \"msg_date\": \"Fri, 12 Nov 2021 04:57:33 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 10:27 AM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n>\\n> On Friday, November 12, 2021 10:46 AM I wrote:\\n> > On Friday, November 12, 2021 8:15 AM Euler Taveira <euler@eulerto.com>\\n> > wrote:\\n> > > I reviewed your patch and I think the fix could be simplified by\\n> > >\\n> > > if (OidIsValid(indexOid))\\n> > > CacheInvalidateRelcache(rel);\\n> > >\\n> > > If indexOid is valid it is a REPLICA IDENTITY INDEX. A few lines above\\n> > > there is a check for a valid indexOid that makes sure the index is\\n> > > already marked as a replica identity; if so, it bail out. If it is\\n> > > not, the relation should be invalidated. Am I missing something?\\n> >\\n> > Thanks for reviewing !\\n> > But I am not sure it's better to simplify the code like \\\"if (OidIsValid(indexOid))\\n> > CacheInvalidate\\\".\\n>\\n> Oh, I got confused with the logic in relation_mark_replica_identity, sorry for that.\\n> I now realize that you are right, we can just check \\\"if (OidIsValid(indexOid))\\\" here\\n> to simplify the code.\\n>\\n\\nBut won't that generate invalidation for the rel twice in the case\\n(change Replica Identity from Nothing to some index) you mentioned in\\nthe previous email?\\n\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 10:50:14 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 10:50 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Fri, Nov 12, 2021 at 10:27 AM houzj.fnst@fujitsu.com\\n> <houzj.fnst@fujitsu.com> wrote:\\n> >\\n> > On Friday, November 12, 2021 10:46 AM I wrote:\\n> > > On Friday, November 12, 2021 8:15 AM Euler Taveira <euler@eulerto.com>\\n> > > wrote:\\n> > > > I reviewed your patch and I think the fix could be simplified by\\n> > > >\\n> > > > if (OidIsValid(indexOid))\\n> > > > CacheInvalidateRelcache(rel);\\n> > > >\\n> > > > If indexOid is valid it is a REPLICA IDENTITY INDEX. A few lines above\\n> > > > there is a check for a valid indexOid that makes sure the index is\\n> > > > already marked as a replica identity; if so, it bail out. If it is\\n> > > > not, the relation should be invalidated. Am I missing something?\\n> > >\\n> > > Thanks for reviewing !\\n> > > But I am not sure it's better to simplify the code like \\\"if (OidIsValid(indexOid))\\n> > > CacheInvalidate\\\".\\n> >\\n> > Oh, I got confused with the logic in relation_mark_replica_identity, sorry for that.\\n> > I now realize that you are right, we can just check \\\"if (OidIsValid(indexOid))\\\" here\\n> > to simplify the code.\\n> >\\n>\\n> But won't that generate invalidation for the rel twice in the case\\n> (change Replica Identity from Nothing to some index) you mentioned in\\n> the previous email?\\n>\\n\\nOh, I see the point. I think this is okay because\\nAddRelcacheInvalidationMessage doesn't allow to add duplicate rel\\ninvalidation. If that is the case I wonder why not simply register\\ninvalidation without any check in the for loop as was the case with\\nTang's original patch?\\n\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 11:02:55 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 1:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> On Fri, Nov 12, 2021 at 10:50 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> > But won't that generate invalidation for the rel twice in the case\\r\\n> > (change Replica Identity from Nothing to some index) you mentioned in\\r\\n> > the previous email?\\r\\n> >\\r\\n> \\r\\n> Oh, I see the point. I think this is okay because\\r\\n> AddRelcacheInvalidationMessage doesn't allow to add duplicate rel\\r\\n> invalidation. If that is the case I wonder why not simply register\\r\\n> invalidation without any check in the for loop as was the case with\\r\\n> Tang's original patch?\\r\\n\\r\\nOK, I also think the code in Tang's original patch is fine.\\r\\nAttach the patch which register invalidation without any check in the for loop.\\r\\n\\r\\nBest regards,\\r\\nHou zj\",\n \"msg_date\": \"Fri, 12 Nov 2021 06:10:07 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021, at 3:10 AM, houzj.fnst@fujitsu.com wrote:\\n> On Fri, Nov 12, 2021 1:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Fri, Nov 12, 2021 at 10:50 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > But won't that generate invalidation for the rel twice in the case\\n> > > (change Replica Identity from Nothing to some index) you mentioned in\\n> > > the previous email?\\n> > >\\n> > \\n> > Oh, I see the point. I think this is okay because\\n> > AddRelcacheInvalidationMessage doesn't allow to add duplicate rel\\n> > invalidation. If that is the case I wonder why not simply register\\n> > invalidation without any check in the for loop as was the case with\\n> > Tang's original patch?\\n> \\n> OK, I also think the code in Tang's original patch is fine.\\n> Attach the patch which register invalidation without any check in the for loop.\\nWFM.\\n\\n\\n--\\nEuler Taveira\\nEDB https://www.enterprisedb.com/\\n\\nOn Fri, Nov 12, 2021, at 3:10 AM, houzj.fnst@fujitsu.com wrote:On Fri, Nov 12, 2021 1:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:> On Fri, Nov 12, 2021 at 10:50 AM Amit Kapila <amit.kapila16@gmail.com> wrote:> > But won't that generate invalidation for the rel twice in the case> > (change Replica Identity from Nothing to some index) you mentioned in> > the previous email?> >> > Oh, I see the point. I think this is okay because> AddRelcacheInvalidationMessage doesn't allow to add duplicate rel> invalidation. If that is the case I wonder why not simply register> invalidation without any check in the for loop as was the case with> Tang's original patch?OK, I also think the code in Tang's original patch is fine.Attach the patch which register invalidation without any check in the for loop.WFM.--Euler TaveiraEDB https://www.enterprisedb.com/\",\n \"msg_date\": \"Fri, 12 Nov 2021 16:17:16 -0300\",\n \"msg_from\": \"\\\"Euler Taveira\\\" <euler@eulerto.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 13, 2021 at 12:47 AM Euler Taveira <euler@eulerto.com> wrote:\\n>\\n> On Fri, Nov 12, 2021, at 3:10 AM, houzj.fnst@fujitsu.com wrote:\\n>\\n> On Fri, Nov 12, 2021 1:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Fri, Nov 12, 2021 at 10:50 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > But won't that generate invalidation for the rel twice in the case\\n> > > (change Replica Identity from Nothing to some index) you mentioned in\\n> > > the previous email?\\n> > >\\n> >\\n> > Oh, I see the point. I think this is okay because\\n> > AddRelcacheInvalidationMessage doesn't allow to add duplicate rel\\n> > invalidation. If that is the case I wonder why not simply register\\n> > invalidation without any check in the for loop as was the case with\\n> > Tang's original patch?\\n>\\n> OK, I also think the code in Tang's original patch is fine.\\n> Attach the patch which register invalidation without any check in the for loop.\\n>\\n> WFM.\\n>\\n\\nThe patch looks mostly good to me. I have slightly tweaked the\\ncomments in the code (as per my previous suggestion) and test. Also, I\\nhave slightly modified the commit message. If the attached looks good\\nto you then kindly prepare patches for back-branches.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\",\n \"msg_date\": \"Sat, 13 Nov 2021 16:19:41 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 13, 2021 6:50 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> \\r\\n> The patch looks mostly good to me. I have slightly tweaked the comments in\\r\\n> the code (as per my previous suggestion) and test. Also, I have slightly\\r\\n> modified the commit message. If the attached looks good to you then kindly\\r\\n> prepare patches for back-branches.\\r\\n\\r\\nThanks for reviewing, the latest patch looks good to me.\\r\\nAttach the patches for back branch (HEAD ~ v10).\\r\\n\\r\\nBest regards,\\r\\nHou zj\",\n \"msg_date\": \"Mon, 15 Nov 2021 01:50:11 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 7:20 AM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n>\\n> On Sat, Nov 13, 2021 6:50 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > The patch looks mostly good to me. I have slightly tweaked the comments in\\n> > the code (as per my previous suggestion) and test. Also, I have slightly\\n> > modified the commit message. If the attached looks good to you then kindly\\n> > prepare patches for back-branches.\\n>\\n> Thanks for reviewing, the latest patch looks good to me.\\n> Attach the patches for back branch (HEAD ~ v10).\\n>\\n\\nPushed.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 17:27:46 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG]Invalidate relcache when setting REPLICA IDENTITY\"\n }\n]","isConversational":false}}},{"rowIdx":53069,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Greetings hackers.\n\nIt seems that PostgreSQL 14 allows using the AT TIME ZONE operator within\ngenerated column definitions; according to the docs, that means the\noperator is considered immutable. However, unless I'm mistaken, the result\nof AT TIME ZONE depends on the time zone database, which is external and\ncan change. I think that means that generated column data can become\nout-of-date upon tz database changes.\n\nSample table creation DDL:\n\nCREATE TABLE events (\n id integer PRIMARY KEY GENERATED BY DEFAULT AS IDENTITY,\n local_timestamp timestamp without time zone NOT NULL,\n utc_timestamp timestamp with time zone GENERATED ALWAYS AS\n(local_timestamp AT TIME ZONE time_zone_id) STORED,\n time_zone_id text NULL\n);\n\nFor comparison, SQL Server does consider AT TIME ZONE to be\nnon-deterministic, and therefore does not allow it in stored generated\ncolumns (it does allow it in non-stored ones).\n\nShay\n\nGreetings hackers.It seems that PostgreSQL 14 allows using the AT TIME ZONE operator within generated column definitions; according to the docs, that means the operator is considered immutable. However, unless I'm mistaken, the result of AT TIME ZONE depends on the time zone database, which is external and can change. I think that means that generated column data can become out-of-date upon tz database changes.Sample table creation DDL:CREATE TABLE events ( id integer PRIMARY KEY GENERATED BY DEFAULT AS IDENTITY, local_timestamp timestamp without time zone NOT NULL, utc_timestamp timestamp with time zone GENERATED ALWAYS AS (local_timestamp AT TIME ZONE time_zone_id) STORED, time_zone_id text NULL);For comparison, SQL Server does consider AT TIME ZONE to be non-deterministic, and therefore does not allow it in stored generated columns (it does allow it in non-stored ones).Shay","msg_date":"Wed, 10 Nov 2021 10:03:29 +0100","msg_from":"Shay Rojansky <roji@roji.org>","msg_from_op":true,"msg_subject":"Should AT TIME ZONE be volatile?"},{"msg_contents":"Greetings hackers:\n\n\nI have a Synology NAS, and there is a PostgreSQL version 11.11 on it. I tried to modify pg_hba.conf file and add a new rule for external access.\n\n\nBut there is a strange thing, the $PGDATA is /var/services/pgsql, so changing the file on $PGDATA/pg_hba.conf maybe ok, but not. See below:\n\n\nThen I tried to find the pg_hba.conf file, there is another one on /etc/postgresql/pg_hba.conf, this is actually the right file.\n\n\nMy question is that PostgreSQL usually use $PGDATA/pg_hba.conf as the access control file, is there other way to specify conf file in other place?\n\n\nBest regards,\nChris","msg_date":"Wed, 10 Nov 2021 18:32:16 +0800 (GMT+08:00)","msg_from":"chris <yuanzefuwater@126.com>","msg_from_op":false,"msg_subject":"Confused with PostgreSQL on Synology NAS"},{"msg_contents":"Hello postgresql uses hba_file configuration parameter: https://www.postgresql.org/docs/current/runtime-config-file-locations.htmlSo could be changed in postgresql.conf regards, Sergei\n","msg_date":"Wed, 10 Nov 2021 13:38:03 +0300","msg_from":"Sergei Kornilov <sk@zsrv.org>","msg_from_op":false,"msg_subject":"Re: Confused with PostgreSQL on Synology NAS"},{"msg_contents":"Wow, thanks so much, I checked it and there is a config on postgresql.conf.\n\n\nRegards,\nChris\nOn 11/10/2021 18:38，Sergei Kornilov<sk@zsrv.org> wrote：\nHello\n \npostgresql uses hba_file configuration parameter: https://www.postgresql.org/docs/current/runtime-config-file-locations.html\nSo could be changed in postgresql.conf\n \nregards, Sergei\n\n\n\n\n\n\n\n\nWow, thanks so much, I checked it and there is a config on postgresql.conf.Regards,ChrisOn 11/10/2021 18:38，Sergei Kornilov<sk@zsrv.org> wrote： \n\n\nHello postgresql uses hba_file configuration parameter: https://www.postgresql.org/docs/current/runtime-config-file-locations.htmlSo could be changed in postgresql.conf regards, Sergei","msg_date":"Wed, 10 Nov 2021 18:59:44 +0800 (GMT+08:00)","msg_from":"chris <yuanzefuwater@126.com>","msg_from_op":false,"msg_subject":"Re: Confused with PostgreSQL on Synology NAS"},{"msg_contents":"Shay Rojansky <roji@roji.org> writes:\n> It seems that PostgreSQL 14 allows using the AT TIME ZONE operator within\n> generated column definitions; according to the docs, that means the\n> operator is considered immutable. However, unless I'm mistaken, the result\n> of AT TIME ZONE depends on the time zone database, which is external and\n> can change. I think that means that generated column data can become\n> out-of-date upon tz database changes.\n\nYeah, we generally don't take such hazards into account. The poster\nchild here is that if we were strict about this, text comparisons\ncouldn't be immutable, because the underlying collation rules can\n(and do) change from time to time. That's obviously unworkable.\n\nI'm not sure how big a deal this really is for timestamps. The actual\nstored time is either UTC or local time, and those are generally pretty\nwell-defined. If you make the wrong choice of which one to store for\nyour use-case, you might be unhappy.\n\nFWIW, I believe the text search operators are also labeled as if the\nunderlying configurations won't change, which of course isn't really true.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 10 Nov 2021 09:49:35 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"> > It seems that PostgreSQL 14 allows using the AT TIME ZONE operator\nwithin\n> > generated column definitions; according to the docs, that means the\n> > operator is considered immutable. However, unless I'm mistaken, the\nresult\n> > of AT TIME ZONE depends on the time zone database, which is external and\n> > can change. I think that means that generated column data can become\n> > out-of-date upon tz database changes.\n>\n> Yeah, we generally don't take such hazards into account. The poster\n> child here is that if we were strict about this, text comparisons\n> couldn't be immutable, because the underlying collation rules can\n> (and do) change from time to time. That's obviously unworkable.\n\nThanks for the explanation Tom. I get the logic, though I think there may\nbe a difference between \"dependent on external rules which may\ntheoretically change but almost never actually do\" and \"dependent on\nsomething that really does change frequently\"... Countries really do change\ntheir daylight savings quite frequently, whereas I'm assuming collation\nrules are relatively immutable and changes are very rare.\n\n> I'm not sure how big a deal this really is for timestamps. The actual\n> stored time is either UTC or local time, and those are generally pretty\n> well-defined. If you make the wrong choice of which one to store for\n> your use-case, you might be unhappy.\n\nThe example I'm working with, is storing a user-provided local timestamp\nand time zone ID, but also having an index generated column in UTC, to be\nable to order all rows on the global timeline regardless of time zone (see this\nblog post\n<https://codeblog.jonskeet.uk/2019/03/27/storing-utc-is-not-a-silver-bullet/>\nby Jon Skeet for some context). If the time zone database changes after the\ngenerated column is computed, the UTC timestamp is out of sync with regards\nto the reality. This seems unsafe.\n\nOn the other hand, it could be argued that this should be allowed, and that\nit should be the user's responsibility to update generated columns when the\ntime zone database changes (or periodically, or whatever). Users always\nhave the option to define a trigger anyway, so we may as well make this\neasier via a generated column.\n\nIn any case, if this is the intended behavior, no problem - I was a bit\nsurprised by it, and found the difference with SQL Server interesting.\n\nShay\n\n> > It seems that PostgreSQL 14 allows using the AT TIME ZONE operator within> > generated column definitions; according to the docs, that means the> > operator is considered immutable. However, unless I'm mistaken, the result> > of AT TIME ZONE depends on the time zone database, which is external and> > can change. I think that means that generated column data can become> > out-of-date upon tz database changes.> > Yeah, we generally don't take such hazards into account. The poster> child here is that if we were strict about this, text comparisons> couldn't be immutable, because the underlying collation rules can> (and do) change from time to time. That's obviously unworkable.Thanks for the explanation Tom. I get the logic, though I think there may be a difference between \"dependent on external rules which may theoretically change but almost never actually do\" and \"dependent on something that really does change frequently\"... Countries really do change their daylight savings quite frequently, whereas I'm assuming collation rules are relatively immutable and changes are very rare.> I'm not sure how big a deal this really is for timestamps. The actual> stored time is either UTC or local time, and those are generally pretty> well-defined. If you make the wrong choice of which one to store for> your use-case, you might be unhappy.The example I'm working with, is storing a user-provided local timestamp and time zone ID, but also having an index generated column in UTC, to be able to order all rows on the global timeline regardless of time zone (see this blog post by Jon Skeet for some context). If the time zone database changes after the generated column is computed, the UTC timestamp is out of sync with regards to the reality. This seems unsafe.On the other hand, it could be argued that this should be allowed, and that it should be the user's responsibility to update generated columns when the time zone database changes (or periodically, or whatever). Users always have the option to define a trigger anyway, so we may as well make this easier via a generated column.In any case, if this is the intended behavior, no problem - I was a bit surprised by it, and found the difference with SQL Server interesting.Shay","msg_date":"Wed, 10 Nov 2021 23:25:51 +0100","msg_from":"Shay Rojansky <roji@roji.org>","msg_from_op":true,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"Shay Rojansky <roji@roji.org> writes:\n>> Yeah, we generally don't take such hazards into account. The poster\n>> child here is that if we were strict about this, text comparisons\n>> couldn't be immutable, because the underlying collation rules can\n>> (and do) change from time to time. That's obviously unworkable.\n\n> Thanks for the explanation Tom. I get the logic, though I think there may\n> be a difference between \"dependent on external rules which may\n> theoretically change but almost never actually do\" and \"dependent on\n> something that really does change frequently\"... Countries really do change\n> their daylight savings quite frequently, whereas I'm assuming collation\n> rules are relatively immutable and changes are very rare.\n\nMeh. Yeah, there are some banana republics that change their DST rules\nat the drop of a hat. More serious governments realize that there are\ncosts to that. For comparison's sake, glibc have modified their\ncollation rules significantly (enough for us to hear complaints about\nit) at least twice in the past decade. That's considerably *more*\nfrequent than DST law changes where I live.\n\n> On the other hand, it could be argued that this should be allowed, and that\n> it should be the user's responsibility to update generated columns when the\n> time zone database changes (or periodically, or whatever). Users always\n> have the option to define a trigger anyway, so we may as well make this\n> easier via a generated column.\n\nYeah, it's not clear that forbidding this would make anyone's life any\nbetter. If you want an index on the UTC equivalent of a local time,\nyou're going to have to find a way to cope with potential mapping\nchanges. But refusing to let you use a generated column doesn't\nseem to help that.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 10 Nov 2021 18:03:12 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"> Yeah, it's not clear that forbidding this would make anyone's life any\n> better. If you want an index on the UTC equivalent of a local time,\n> you're going to have to find a way to cope with potential mapping\n> changes. But refusing to let you use a generated column doesn't\n> seem to help that.\n\nMakes sense, thanks Tom.\n\n> Yeah, it's not clear that forbidding this would make anyone's life any> better. If you want an index on the UTC equivalent of a local time,> you're going to have to find a way to cope with potential mapping> changes. But refusing to let you use a generated column doesn't> seem to help that.Makes sense, thanks Tom.","msg_date":"Thu, 11 Nov 2021 12:05:22 +0100","msg_from":"Shay Rojansky <roji@roji.org>","msg_from_op":true,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Wed, Nov 10, 2021 at 6:03 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> For comparison's sake, glibc have modified their\n> collation rules significantly (enough for us to hear complaints about\n> it) at least twice in the past decade. That's considerably *more*\n> frequent than DST law changes where I live.\n\nYes. It seems to be extremely common for people to get hosed by\ncollation changes. Different major versions of RHEL ship with\ndifferent collations. Different minor versions of RHEL ship with\ndifferent collations. Tiny little changes in very end of the glibc\nversion number include collation changes. I believe that it's been\nexplicitly stated by Ulrich Drepper that you should not rely on\ncollation definitions not to change at any time, and that relying on\nthem for any sort of on-disk ordering is nuts. Which seems like an\ninsane idea, because (1) surely the only point of such definitions is\nto help you sort your data, and you probably don't want to resort it\nin a continuous loop in case somebody decided to change the collation\ndefinition under you and (2) how important can it be to continually\ntinker with the sorting rules?\n\nI'm not really convinced that ICU is better, either. I think it's more\nthat it isn't used as much.\n\nI don't have any constructive proposal for what to do about any of\nthis. It sure is frustrating, though.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 11 Nov 2021 09:21:59 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> I'm not really convinced that ICU is better, either. I think it's more\n> that it isn't used as much.\n\nWell, at least ICU has a notion of attaching versions to collations.\nHow mindful they are of bumping the version number when necessary\nremains to be seen. But the POSIX locale APIs don't even offer the\nopportunity to get it right.\n\n> I don't have any constructive proposal for what to do about any of\n> this. It sure is frustrating, though.\n\nYup. If we had reliable ways to detect changes in this sort of\nenvironment-supplied data, maybe we could do something about it\n(a la the work that's been happening on attaching collation versions\nto indexes). But personally I can't summon the motivation to work\non that, when ICU is the *only* such infrastructure that offers\nreadily program-readable versioning.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 11 Nov 2021 09:52:52 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Thu, 2021-11-11 at 09:52 -0500, Tom Lane wrote:\n> Robert Haas <robertmhaas@gmail.com> writes:\n> > I'm not really convinced that ICU is better, either. I think it's more\n> > that it isn't used as much.\n> \n> Well, at least ICU has a notion of attaching versions to collations.\n> How mindful they are of bumping the version number when necessary\n> remains to be seen. But the POSIX locale APIs don't even offer the\n> opportunity to get it right.\n\nAlso, it is much easier *not* to upgrade libicu than it is to *not*\nupgrade libc, which an essential component of the operating system.\n\n> > I don't have any constructive proposal for what to do about any of\n> > this. It sure is frustrating, though.\n> \n> Yup. If we had reliable ways to detect changes in this sort of\n> environment-supplied data, maybe we could do something about it\n> (a la the work that's been happening on attaching collation versions\n> to indexes). But personally I can't summon the motivation to work\n> on that, when ICU is the *only* such infrastructure that offers\n> readily program-readable versioning.\n\nNobody will want to hear that, but the only really good solution would\nbe for PostgreSQL to have its own built-in collations.\n\nYours,\nLaurenz Albe\n\n\n\n","msg_date":"Thu, 11 Nov 2021 17:16:55 +0100","msg_from":"Laurenz Albe <laurenz.albe@cybertec.at>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"Laurenz Albe <laurenz.albe@cybertec.at> writes:\n> On Thu, 2021-11-11 at 09:52 -0500, Tom Lane wrote:\n>> Yup. If we had reliable ways to detect changes in this sort of\n>> environment-supplied data, maybe we could do something about it\n>> (a la the work that's been happening on attaching collation versions\n>> to indexes). But personally I can't summon the motivation to work\n>> on that, when ICU is the *only* such infrastructure that offers\n>> readily program-readable versioning.\n\n> Nobody will want to hear that, but the only really good solution would\n> be for PostgreSQL to have its own built-in collations.\n\nAnd our own tzdb too? Maybe an outfit like Oracle has the resources\nand will to maintain their own copies of such data, but I can't see\nus wanting to do it.\n\ntzdb has an additional problem, which is that not updating is not an\noption: if you're affected by a DST law change, you want that update,\nand you frequently need it yesterday. We're definitely not set up\nto handle that sort of update process, which is why we recommend\n--with-system-tzdata.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 11 Nov 2021 12:07:36 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Thu, Nov 11, 2021 at 11:16 AM Laurenz Albe <laurenz.albe@cybertec.at> wrote:\n> Nobody will want to hear that, but the only really good solution would\n> be for PostgreSQL to have its own built-in collations.\n\n+1.\n\nI agree with Tom that it sounds like a lot of work. And to be honest\nit's work that I don't really feel very excited about. It would be\nnecessary to understand not only the bona fide sorting rules of every\nhuman language out there, which might actually be sort of fun at least\nfor a while, but also to decide - probably according to some\nincomprehensible standard - how Japanese katakana ought to sort in\ncomparison to, say, box-drawing characters, the Mongolian alphabet,\nand smiley-face emojis. I think it's not particularly likely that\nthere are a whole lot of documents out there that include all of those\nthings, but the comparison algorithm has to return something, and\nprobably there are people who have strong feelings about what the\nright answers are. That's a pretty unappealing thing to tackle, and I\nam not volunteering.\n\nOn the other hand, if we don't do it, I'm suspicious that things will\nnever get any better. And that would be sad.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 11 Nov 2021 12:32:59 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> I agree with Tom that it sounds like a lot of work. And to be honest\n> it's work that I don't really feel very excited about.\n\nEven if you were excited about it, would maintaining such data be\na good use of project resources? It's not like we lack other things\nwe ought to be doing. I agree that the lack of reliable versioning\ninfo is a problem, but I can't see that \"let's fork ICU and tzdb too\"\nis a good answer.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 11 Nov 2021 13:38:13 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":">Laurenz Albe <laurenz.albe@cybertec.at> writes:\n>> On Thu, 2021-11-11 at 09:52 -0500, Tom Lane wrote:\n>>> Yup. If we had reliable ways to detect changes in this sort of\n>>> environment-supplied data, maybe we could do something about it\n>>> (a la the work that's been happening on attaching collation versions\n>>> to indexes). But personally I can't summon the motivation to work\n>>> on that, when ICU is the *only* such infrastructure that offers\n>>> readily program-readable versioning.\n\n>> Nobody will want to hear that, but the only really good solution would\n>> be for PostgreSQL to have its own built-in collations.\n\n>And our own tzdb too? Maybe an outfit like Oracle has the resources\n>and will to maintain their own copies of such data, but I can't see\n>us wanting to do it.\n\n>tzdb has an additional problem, which is that not updating is not an\n>option: if you're affected by a DST law change, you want that update,\n>and you frequently need it yesterday. We're definitely not set up\n>to handle that sort of update process, which is why we recommend\n>--with-system-tzdata.\n\nWhere in the docs is this recommended? The only place I can find it is here:\nhttps://www.postgresql.org/docs/current/install-procedure.html\n\nRegards\nDaniel\n\n\n","msg_date":"Thu, 11 Nov 2021 19:07:46 +0000","msg_from":"\"Daniel Westermann (DWE)\" <daniel.westermann@dbi-services.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Thu, Nov 11, 2021 at 1:38 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Even if you were excited about it, would maintaining such data be\n> a good use of project resources? It's not like we lack other things\n> we ought to be doing. I agree that the lack of reliable versioning\n> info is a problem, but I can't see that \"let's fork ICU and tzdb too\"\n> is a good answer.\n\nYou might be right, but I think it's hard to say for certain. I don't\nthink this is one of our top 10 problems, but it's probably one of our\ntop 1000 problems, and it might be one of our top 100 problems. It's\nentirely subjective, and people are likely to disagree, but based on\nthose numbers I'd say it's not worth 2% of our resources but it might\nwell be worth 0.02% of our resources. Everybody's going to have their\nown opinion, though. I'm not sure how relevant those opinions are in\nthe end, though. The community has little power to force anybody to\nwork on anything; people work on what they want to work on, or what\nthey get paid to work on, not what somebody else in the community\ndecides is most important.\n\nAnyway, from my point of view, if some well-respected community member\nshowed up and wanted to add a new kind of collation that is provided\nby PostgreSQL itself and had some well-thought-out candidates for\ninitial integration, I don't know that it would be smart to turn that\ndown because solving the whole problem for every case might be more\nwork than anyone's willing to do. The only real issue for the project\nis if somebody makes a drive-by contribution of something that's going\nto need continuous updating. That sort of thing would be bad on\nmultiple fronts: not only do we not want to get forced into spending\nongoing maintenance effort on something like this, but we want\ncollation definitions that *actually don't change*.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 11 Nov 2021 14:09:37 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"\"Daniel Westermann (DWE)\" <daniel.westermann@dbi-services.com> writes:\n> On Thu, 2021-11-11 at 09:52 -0500, Tom Lane wrote:\n>> tzdb has an additional problem, which is that not updating is not an\n>> option: if you're affected by a DST law change, you want that update,\n>> and you frequently need it yesterday. We're definitely not set up\n>> to handle that sort of update process, which is why we recommend\n>> --with-system-tzdata.\n\n> Where in the docs is this recommended? The only place I can find it is here:\n> https://www.postgresql.org/docs/current/install-procedure.html\n\nYup, that's exactly the text I was thinking of. Maybe the recommendation\nshould be more enthusiastic --- it was written back when it was still\nrather questionable whether a platform would have an up-to-date copy of\ntzdata. (Maybe it still is, at least for the \"up-to-date\" part.)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 11 Nov 2021 14:17:59 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> ... but we want\n> collation definitions that *actually don't change*.\n\nUm ... how would that work? Unicode is a moving target. Even without\ntheir continual addition of stuff, I'm not convinced that social rules\nabout how to sort are engraved on stone tablets. The need for collation\nupdates may not be as predictable as the need for timezone updates,\nbut I doubt that we can just freeze the data forever.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 11 Nov 2021 14:23:34 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Thu, Nov 11, 2021 at 2:23 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Robert Haas <robertmhaas@gmail.com> writes:\n> > ... but we want\n> > collation definitions that *actually don't change*.\n>\n> Um ... how would that work? Unicode is a moving target. Even without\n> their continual addition of stuff, I'm not convinced that social rules\n> about how to sort are engraved on stone tablets. The need for collation\n> updates may not be as predictable as the need for timezone updates,\n> but I doubt that we can just freeze the data forever.\n\nI don't know, but I think the social rules that actually matter change\nextremely slowly. To my knowledge, the alphabet song has not changed\nsince I was in kindergarten. Now I agree that in some countries it\nprobably has ... but I doubt those events are super-common, because a\ncountry does change its definition of alphabetical order, there's a\nheck of a lot more updating to do than just reindexing your PostgreSQL\ndatabases. The signs saying A-L go to the left and M-Z go to the right\nwill need revision if we decide M comes before L. I feel like it has\nto be the case that most of the updates that are being made involve\nthings like how obscure characters compare to other obscure\ncharacters, or what to do in corner-case situations involving multiple\ndiacritical marks. I know I've seen collation changes on Macs that\nchanged the order in which en_US.UTF8 strings sorted. But it wasn't\nthat the rules about English sorting have actually changed. It was\nthat somebody somewhere decided that the algorithm should be more or\nless case-sensitive, or that we ought to ignore the amount of\nwhitespace between words instead of not ignoring it, or I don't know\nexactly, but not anything that people universally agree on. Tinkering\nwith obscure rules that actual human beings wouldn't agree on and\nprioritizing that over a stable algorithm is, IMHO, ridiculous.\n\nIf the Unicode consortium introduces a new emoji for \"annoyed\nPostgreSQL hacker,\" I really do not care whether that collates before\nor after the existing symbol for \"floral heart bullet, reversed\nrotated.\" I care much more about whether it collates the same way\nafter the next minor release as it does the day it's released. And I\nseriously doubt that I am alone in that.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 11 Nov 2021 14:41:47 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Thu, 11 Nov 2021 at 14:42, Robert Haas <robertmhaas@gmail.com> wrote:\n\n\n> diacritical marks. I know I've seen collation changes on Macs that\n> changed the order in which en_US.UTF8 strings sorted. But it wasn't\n> that the rules about English sorting have actually changed. It was\n> that somebody somewhere decided that the algorithm should be more or\n> less case-sensitive, or that we ought to ignore the amount of\n> whitespace between words instead of not ignoring it, or I don't know\n> exactly, but not anything that people universally agree on. Tinkering\n> with obscure rules that actual human beings wouldn't agree on and\n> prioritizing that over a stable algorithm is, IMHO, ridiculous.\n>\n\nYes, I thought the point here was to nail down each change as a separate\nversion. So for example maybe I'm running Universal Compare Everything\nCollation v1.2435 while your database is running Universal Compare\nEverything Collation v1.2436, with the only difference being whether e\ndiaresis circumflex comes before or after e circumflex diaresis. If I do a\nsystem upgrade I won't just silently corrupt any indexes with those\ncharacters; instead I'll be told that my collation is out of date and then\nI can decide whether to stick with the old collation or rebuild my indexes\nand upgrade.\n\nThere is however one kind of change at least that I think can be made\nsafely: adding a new character in between existing characters. That\nshouldn't affect any existing indexes.\n\nIf the Unicode consortium introduces a new emoji for \"annoyed\n> PostgreSQL hacker,\" I really do not care whether that collates before\n> or after the existing symbol for \"floral heart bullet, reversed\n> rotated.\" I care much more about whether it collates the same way\n> after the next minor release as it does the day it's released. And I\n> seriously doubt that I am alone in that.\n>\n\nOn Thu, 11 Nov 2021 at 14:42, Robert Haas <robertmhaas@gmail.com> wrote: \ndiacritical marks. I know I've seen collation changes on Macs that\nchanged the order in which en_US.UTF8 strings sorted. But it wasn't\nthat the rules about English sorting have actually changed. It was\nthat somebody somewhere decided that the algorithm should be more or\nless case-sensitive, or that we ought to ignore the amount of\nwhitespace between words instead of not ignoring it, or I don't know\nexactly, but not anything that people universally agree on. Tinkering\nwith obscure rules that actual human beings wouldn't agree on and\nprioritizing that over a stable algorithm is, IMHO, ridiculous.Yes, I thought the point here was to nail down each change as a separate version. So for example maybe I'm running Universal Compare Everything Collation v1.2435 while your database is running Universal Compare Everything Collation v1.2436, with the only difference being whether e diaresis circumflex comes before or after e circumflex diaresis. If I do a system upgrade I won't just silently corrupt any indexes with those characters; instead I'll be told that my collation is out of date and then I can decide whether to stick with the old collation or rebuild my indexes and upgrade.There is however one kind of change at least that I think can be made safely: adding a new character in between existing characters. That shouldn't affect any existing indexes.\nIf the Unicode consortium introduces a new emoji for \"annoyed\nPostgreSQL hacker,\" I really do not care whether that collates before\nor after the existing symbol for \"floral heart bullet, reversed\nrotated.\" I care much more about whether it collates the same way\nafter the next minor release as it does the day it's released. And I\nseriously doubt that I am alone in that.","msg_date":"Thu, 11 Nov 2021 15:45:48 -0500","msg_from":"Isaac Morland <isaac.morland@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Thu, Nov 11, 2021 at 3:45 PM Isaac Morland <isaac.morland@gmail.com> wrote:\n> There is however one kind of change at least that I think can be made safely: adding a new character in between existing characters. That shouldn't affect any existing indexes.\n\nOnly if you can guarantee that said character is not present already.\nI don't think we update the end of the acceptable code point range\nevery time that Unicode adds new stuff, so probably those things are\nsubject to some default rule unless and until someone installs\nsomething more specific. Therefore I doubt that even this case is\ntruly safe.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 11 Nov 2021 16:08:22 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Thu, 11 Nov 2021 at 16:08, Robert Haas <robertmhaas@gmail.com> wrote:\n\n> On Thu, Nov 11, 2021 at 3:45 PM Isaac Morland <isaac.morland@gmail.com>\n> wrote:\n> > There is however one kind of change at least that I think can be made\n> safely: adding a new character in between existing characters. That\n> shouldn't affect any existing indexes.\n>\n> Only if you can guarantee that said character is not present already.\n> I don't think we update the end of the acceptable code point range\n> every time that Unicode adds new stuff, so probably those things are\n> subject to some default rule unless and until someone installs\n> something more specific. Therefore I doubt that even this case is\n> truly safe.\n>\n\nWouldn't an existing index only have characters that were already part of\nthe collation? Attempting to use one not covered by the collation I would\nhave expected to cause an error at insert time. But definitely I agree I\nwouldn't feel confident about the safety of any change.\n\nOn Thu, 11 Nov 2021 at 16:08, Robert Haas <robertmhaas@gmail.com> wrote:On Thu, Nov 11, 2021 at 3:45 PM Isaac Morland <isaac.morland@gmail.com> wrote:\n> There is however one kind of change at least that I think can be made safely: adding a new character in between existing characters. That shouldn't affect any existing indexes.\n\nOnly if you can guarantee that said character is not present already.\nI don't think we update the end of the acceptable code point range\nevery time that Unicode adds new stuff, so probably those things are\nsubject to some default rule unless and until someone installs\nsomething more specific. Therefore I doubt that even this case is\ntruly safe.Wouldn't an existing index only have characters that were already part of the collation? Attempting to use one not covered by the collation I would have expected to cause an error at insert time. But definitely I agree I wouldn't feel confident about the safety of any change.","msg_date":"Thu, 11 Nov 2021 17:04:39 -0500","msg_from":"Isaac Morland <isaac.morland@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Thu, Nov 11, 2021 at 5:04 PM Isaac Morland <isaac.morland@gmail.com> wrote:\n> Wouldn't an existing index only have characters that were already part of the collation? Attempting to use one not covered by the collation I would have expected to cause an error at insert time. But definitely I agree I wouldn't feel confident about the safety of any change.\n\nI mean it's not like we are updating the definition of\npg_utf8_verifychar() every time they define a new code point.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 11 Nov 2021 18:08:47 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Fri, Nov 12, 2021 at 12:09 PM Robert Haas <robertmhaas@gmail.com> wrote:\n> On Thu, Nov 11, 2021 at 5:04 PM Isaac Morland <isaac.morland@gmail.com> wrote:\n> > Wouldn't an existing index only have characters that were already part of the collation? Attempting to use one not covered by the collation I would have expected to cause an error at insert time. But definitely I agree I wouldn't feel confident about the safety of any change.\n>\n> I mean it's not like we are updating the definition of\n> pg_utf8_verifychar() every time they define a new code point.\n\nRight, and there may be other systems that do this. That is, reject\ninvalid code points, because they have no sort order. You can see\nsome sign of this in the major and minor collation version numbers\nreported by Windows (though I'm not sure if this was lost with the\nrecent move to ICU): if only the minor version changes, the\ndocumentation says it means \"we only added new code points, no\nexisting code points changed\", so a sufficiently clever program\ndoesn't need to rebuild its persistent ordered data structures, if it\nnever allowed any unknown code points into the structure before. Not\nonly does PostgreSQL not have the logic for that, it also doesn't have\nthe data: the set of existing code points has to be the one used by\nthe collation provider, and the collection providers we have don't\nreject unknown code points on comparison.\n\n\n","msg_date":"Fri, 12 Nov 2021 12:31:18 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On 11.11.21 18:32, Robert Haas wrote:\n> I agree with Tom that it sounds like a lot of work. And to be honest\n> it's work that I don't really feel very excited about. It would be\n> necessary to understand not only the bona fide sorting rules of every\n> human language out there, which might actually be sort of fun at least\n> for a while, but also to decide - probably according to some\n> incomprehensible standard - how Japanese katakana ought to sort in\n> comparison to, say, box-drawing characters, the Mongolian alphabet,\n> and smiley-face emojis. I think it's not particularly likely that\n> there are a whole lot of documents out there that include all of those\n> things, but the comparison algorithm has to return something, and\n> probably there are people who have strong feelings about what the\n> right answers are. That's a pretty unappealing thing to tackle, and I\n> am not volunteering.\n> \n> On the other hand, if we don't do it, I'm suspicious that things will\n> never get any better. And that would be sad.\n\nThere are standards for sort order, and the major hiccups we had in the \npast were mostly moving from older versions of those standards to newer \nversions. So at some point this should stabilize.\n\n\n\n","msg_date":"Fri, 12 Nov 2021 14:42:30 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Fri, Nov 12, 2021 at 8:42 AM Peter Eisentraut\n<peter.eisentraut@enterprisedb.com> wrote:\n> There are standards for sort order, and the major hiccups we had in the\n> past were mostly moving from older versions of those standards to newer\n> versions. So at some point this should stabilize.\n\nOnly if they don't keep making new versions of the standards.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 12 Nov 2021 09:49:12 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Thu, Nov 11, 2021 at 09:52:52AM -0500, Tom Lane wrote:\n> Robert Haas <robertmhaas@gmail.com> writes:\n> > I'm not really convinced that ICU is better, either. I think it's more\n> > that it isn't used as much.\n> \n> Well, at least ICU has a notion of attaching versions to collations.\n> How mindful they are of bumping the version number when necessary\n> remains to be seen. But the POSIX locale APIs don't even offer the\n> opportunity to get it right.\n> \n> > I don't have any constructive proposal for what to do about any of\n> > this. It sure is frustrating, though.\n> \n> Yup. If we had reliable ways to detect changes in this sort of\n> environment-supplied data, maybe we could do something about it\n> (a la the work that's been happening on attaching collation versions\n> to indexes). But personally I can't summon the motivation to work\n\n Theoretically there are versions attached to collations already:\nthe collation in index is an oid referencing the pg_collation.\n And the pg_collation already has versions.\n\n Currently for glibc the version looks like glibc version at\ninitdb, and that doesn't seem very reliable, but that could be a\ndifferent task (to find LC_COLLATE file and put hash of the\nusuable data into version string, for example).\n\n Currently, it is questionable how to work with the different\nversions of collations -- but that could be solved e.g. via ap-\npropriate naming. Perhaps \"collation@ver\" ? But if the version\nwould contain a hash, a full version could be a bit dubious.\n And some database maintainance task could check that all the old\ncollations are available, rename them as needed, and create a set\nthe new ones.\n Automatically invalidating all the indexes, unfortunately.\n\n> on that, when ICU is the *only* such infrastructure that offers\n> readily program-readable versioning.\n> \n> \t\t\tregards, tom lane\n> \n\n\n","msg_date":"Sat, 13 Nov 2021 01:47:17 +0300","msg_from":"Ilya Anfimov <ilan@tzirechnoy.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"},{"msg_contents":"On Sat, Nov 13, 2021 at 11:47 AM Ilya Anfimov <ilan@tzirechnoy.com> wrote:\n> Currently for glibc the version looks like glibc version at\n> initdb, and that doesn't seem very reliable, but that could be a\n> different task (to find LC_COLLATE file and put hash of the\n> usuable data into version string, for example).\n\nYeah, I had a system exactly like that working (that is, a way to run\narbitrary commands to capture version strings, that could be used to\nhash your collation definition files, patches somewhere in the\narchives), but then we thought it'd be better to use glibc versions,\nand separately, to perhaps try to ask the glibc people to expose a\nversion. FreeBSD (at my request), Windows and ICU do expose versions\nin a straightforward way, and we capture those.\n\n> Currently, it is questionable how to work with the different\n> versions of collations -- but that could be solved e.g. via ap-\n> propriate naming. Perhaps \"collation@ver\" ? But if the version\n> would contain a hash, a full version could be a bit dubious.\n> And some database maintainance task could check that all the old\n> collations are available, rename them as needed, and create a set\n> the new ones.\n> Automatically invalidating all the indexes, unfortunately.\n\nWe built a system that at least detected the changes on a per-index\nlevel, but failed to ship it in release 14. See revert commit, and\nlinks back to previous commits and discussion:\n\nhttps://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=ec48314708262d8ea6cdcb83f803fc83dd89e721\n\nIt didn't invalidate indexes, but complained about each individual\nindex on first access in each session, until you REINDEXed it. We\nwill try again :-)\n\nIn the archives you can find discussions of how to make a system that\ntolerates multiple version existing at the same time as I think you're\ngetting at, like DB2 does. It's tricky, because it's code and data.\nDB2 ships with N copies of ICU in it.\n\n\n","msg_date":"Sat, 13 Nov 2021 13:09:33 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should AT TIME ZONE be volatile?"}],"string":"[\n {\n \"msg_contents\": \"Greetings hackers.\\n\\nIt seems that PostgreSQL 14 allows using the AT TIME ZONE operator within\\ngenerated column definitions; according to the docs, that means the\\noperator is considered immutable. However, unless I'm mistaken, the result\\nof AT TIME ZONE depends on the time zone database, which is external and\\ncan change. I think that means that generated column data can become\\nout-of-date upon tz database changes.\\n\\nSample table creation DDL:\\n\\nCREATE TABLE events (\\n id integer PRIMARY KEY GENERATED BY DEFAULT AS IDENTITY,\\n local_timestamp timestamp without time zone NOT NULL,\\n utc_timestamp timestamp with time zone GENERATED ALWAYS AS\\n(local_timestamp AT TIME ZONE time_zone_id) STORED,\\n time_zone_id text NULL\\n);\\n\\nFor comparison, SQL Server does consider AT TIME ZONE to be\\nnon-deterministic, and therefore does not allow it in stored generated\\ncolumns (it does allow it in non-stored ones).\\n\\nShay\\n\\nGreetings hackers.It seems that PostgreSQL 14 allows using the AT TIME ZONE operator within generated column definitions; according to the docs, that means the operator is considered immutable. However, unless I'm mistaken, the result of AT TIME ZONE depends on the time zone database, which is external and can change. I think that means that generated column data can become out-of-date upon tz database changes.Sample table creation DDL:CREATE TABLE events ( id integer PRIMARY KEY GENERATED BY DEFAULT AS IDENTITY, local_timestamp timestamp without time zone NOT NULL, utc_timestamp timestamp with time zone GENERATED ALWAYS AS (local_timestamp AT TIME ZONE time_zone_id) STORED, time_zone_id text NULL);For comparison, SQL Server does consider AT TIME ZONE to be non-deterministic, and therefore does not allow it in stored generated columns (it does allow it in non-stored ones).Shay\",\n \"msg_date\": \"Wed, 10 Nov 2021 10:03:29 +0100\",\n \"msg_from\": \"Shay Rojansky <roji@roji.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"Greetings hackers:\\n\\n\\nI have a Synology NAS, and there is a PostgreSQL version 11.11 on it. I tried to modify pg_hba.conf file and add a new rule for external access.\\n\\n\\nBut there is a strange thing, the $PGDATA is /var/services/pgsql, so changing the file on $PGDATA/pg_hba.conf maybe ok, but not. See below:\\n\\n\\nThen I tried to find the pg_hba.conf file, there is another one on /etc/postgresql/pg_hba.conf, this is actually the right file.\\n\\n\\nMy question is that PostgreSQL usually use $PGDATA/pg_hba.conf as the access control file, is there other way to specify conf file in other place?\\n\\n\\nBest regards,\\nChris\",\n \"msg_date\": \"Wed, 10 Nov 2021 18:32:16 +0800 (GMT+08:00)\",\n \"msg_from\": \"chris <yuanzefuwater@126.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Confused with PostgreSQL on Synology NAS\"\n },\n {\n \"msg_contents\": \"Hello postgresql uses hba_file configuration parameter: https://www.postgresql.org/docs/current/runtime-config-file-locations.htmlSo could be changed in postgresql.conf regards, Sergei\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 13:38:03 +0300\",\n \"msg_from\": \"Sergei Kornilov <sk@zsrv.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Confused with PostgreSQL on Synology NAS\"\n },\n {\n \"msg_contents\": \"Wow, thanks so much, I checked it and there is a config on postgresql.conf.\\n\\n\\nRegards,\\nChris\\nOn 11/10/2021 18:38，Sergei Kornilov<sk@zsrv.org> wrote：\\nHello\\n \\npostgresql uses hba_file configuration parameter: https://www.postgresql.org/docs/current/runtime-config-file-locations.html\\nSo could be changed in postgresql.conf\\n \\nregards, Sergei\\n\\n\\n\\n\\n\\n\\n\\n\\nWow, thanks so much, I checked it and there is a config on postgresql.conf.Regards,ChrisOn 11/10/2021 18:38，Sergei Kornilov<sk@zsrv.org> wrote： \\n\\n\\nHello postgresql uses hba_file configuration parameter: https://www.postgresql.org/docs/current/runtime-config-file-locations.htmlSo could be changed in postgresql.conf regards, Sergei\",\n \"msg_date\": \"Wed, 10 Nov 2021 18:59:44 +0800 (GMT+08:00)\",\n \"msg_from\": \"chris <yuanzefuwater@126.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Confused with PostgreSQL on Synology NAS\"\n },\n {\n \"msg_contents\": \"Shay Rojansky <roji@roji.org> writes:\\n> It seems that PostgreSQL 14 allows using the AT TIME ZONE operator within\\n> generated column definitions; according to the docs, that means the\\n> operator is considered immutable. However, unless I'm mistaken, the result\\n> of AT TIME ZONE depends on the time zone database, which is external and\\n> can change. I think that means that generated column data can become\\n> out-of-date upon tz database changes.\\n\\nYeah, we generally don't take such hazards into account. The poster\\nchild here is that if we were strict about this, text comparisons\\ncouldn't be immutable, because the underlying collation rules can\\n(and do) change from time to time. That's obviously unworkable.\\n\\nI'm not sure how big a deal this really is for timestamps. The actual\\nstored time is either UTC or local time, and those are generally pretty\\nwell-defined. If you make the wrong choice of which one to store for\\nyour use-case, you might be unhappy.\\n\\nFWIW, I believe the text search operators are also labeled as if the\\nunderlying configurations won't change, which of course isn't really true.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 09:49:35 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"> > It seems that PostgreSQL 14 allows using the AT TIME ZONE operator\\nwithin\\n> > generated column definitions; according to the docs, that means the\\n> > operator is considered immutable. However, unless I'm mistaken, the\\nresult\\n> > of AT TIME ZONE depends on the time zone database, which is external and\\n> > can change. I think that means that generated column data can become\\n> > out-of-date upon tz database changes.\\n>\\n> Yeah, we generally don't take such hazards into account. The poster\\n> child here is that if we were strict about this, text comparisons\\n> couldn't be immutable, because the underlying collation rules can\\n> (and do) change from time to time. That's obviously unworkable.\\n\\nThanks for the explanation Tom. I get the logic, though I think there may\\nbe a difference between \\\"dependent on external rules which may\\ntheoretically change but almost never actually do\\\" and \\\"dependent on\\nsomething that really does change frequently\\\"... Countries really do change\\ntheir daylight savings quite frequently, whereas I'm assuming collation\\nrules are relatively immutable and changes are very rare.\\n\\n> I'm not sure how big a deal this really is for timestamps. The actual\\n> stored time is either UTC or local time, and those are generally pretty\\n> well-defined. If you make the wrong choice of which one to store for\\n> your use-case, you might be unhappy.\\n\\nThe example I'm working with, is storing a user-provided local timestamp\\nand time zone ID, but also having an index generated column in UTC, to be\\nable to order all rows on the global timeline regardless of time zone (see this\\nblog post\\n<https://codeblog.jonskeet.uk/2019/03/27/storing-utc-is-not-a-silver-bullet/>\\nby Jon Skeet for some context). If the time zone database changes after the\\ngenerated column is computed, the UTC timestamp is out of sync with regards\\nto the reality. This seems unsafe.\\n\\nOn the other hand, it could be argued that this should be allowed, and that\\nit should be the user's responsibility to update generated columns when the\\ntime zone database changes (or periodically, or whatever). Users always\\nhave the option to define a trigger anyway, so we may as well make this\\neasier via a generated column.\\n\\nIn any case, if this is the intended behavior, no problem - I was a bit\\nsurprised by it, and found the difference with SQL Server interesting.\\n\\nShay\\n\\n> > It seems that PostgreSQL 14 allows using the AT TIME ZONE operator within> > generated column definitions; according to the docs, that means the> > operator is considered immutable. However, unless I'm mistaken, the result> > of AT TIME ZONE depends on the time zone database, which is external and> > can change. I think that means that generated column data can become> > out-of-date upon tz database changes.> > Yeah, we generally don't take such hazards into account. The poster> child here is that if we were strict about this, text comparisons> couldn't be immutable, because the underlying collation rules can> (and do) change from time to time. That's obviously unworkable.Thanks for the explanation Tom. I get the logic, though I think there may be a difference between \\\"dependent on external rules which may theoretically change but almost never actually do\\\" and \\\"dependent on something that really does change frequently\\\"... Countries really do change their daylight savings quite frequently, whereas I'm assuming collation rules are relatively immutable and changes are very rare.> I'm not sure how big a deal this really is for timestamps. The actual> stored time is either UTC or local time, and those are generally pretty> well-defined. If you make the wrong choice of which one to store for> your use-case, you might be unhappy.The example I'm working with, is storing a user-provided local timestamp and time zone ID, but also having an index generated column in UTC, to be able to order all rows on the global timeline regardless of time zone (see this blog post by Jon Skeet for some context). If the time zone database changes after the generated column is computed, the UTC timestamp is out of sync with regards to the reality. This seems unsafe.On the other hand, it could be argued that this should be allowed, and that it should be the user's responsibility to update generated columns when the time zone database changes (or periodically, or whatever). Users always have the option to define a trigger anyway, so we may as well make this easier via a generated column.In any case, if this is the intended behavior, no problem - I was a bit surprised by it, and found the difference with SQL Server interesting.Shay\",\n \"msg_date\": \"Wed, 10 Nov 2021 23:25:51 +0100\",\n \"msg_from\": \"Shay Rojansky <roji@roji.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"Shay Rojansky <roji@roji.org> writes:\\n>> Yeah, we generally don't take such hazards into account. The poster\\n>> child here is that if we were strict about this, text comparisons\\n>> couldn't be immutable, because the underlying collation rules can\\n>> (and do) change from time to time. That's obviously unworkable.\\n\\n> Thanks for the explanation Tom. I get the logic, though I think there may\\n> be a difference between \\\"dependent on external rules which may\\n> theoretically change but almost never actually do\\\" and \\\"dependent on\\n> something that really does change frequently\\\"... Countries really do change\\n> their daylight savings quite frequently, whereas I'm assuming collation\\n> rules are relatively immutable and changes are very rare.\\n\\nMeh. Yeah, there are some banana republics that change their DST rules\\nat the drop of a hat. More serious governments realize that there are\\ncosts to that. For comparison's sake, glibc have modified their\\ncollation rules significantly (enough for us to hear complaints about\\nit) at least twice in the past decade. That's considerably *more*\\nfrequent than DST law changes where I live.\\n\\n> On the other hand, it could be argued that this should be allowed, and that\\n> it should be the user's responsibility to update generated columns when the\\n> time zone database changes (or periodically, or whatever). Users always\\n> have the option to define a trigger anyway, so we may as well make this\\n> easier via a generated column.\\n\\nYeah, it's not clear that forbidding this would make anyone's life any\\nbetter. If you want an index on the UTC equivalent of a local time,\\nyou're going to have to find a way to cope with potential mapping\\nchanges. But refusing to let you use a generated column doesn't\\nseem to help that.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 18:03:12 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"> Yeah, it's not clear that forbidding this would make anyone's life any\\n> better. If you want an index on the UTC equivalent of a local time,\\n> you're going to have to find a way to cope with potential mapping\\n> changes. But refusing to let you use a generated column doesn't\\n> seem to help that.\\n\\nMakes sense, thanks Tom.\\n\\n> Yeah, it's not clear that forbidding this would make anyone's life any> better. If you want an index on the UTC equivalent of a local time,> you're going to have to find a way to cope with potential mapping> changes. But refusing to let you use a generated column doesn't> seem to help that.Makes sense, thanks Tom.\",\n \"msg_date\": \"Thu, 11 Nov 2021 12:05:22 +0100\",\n \"msg_from\": \"Shay Rojansky <roji@roji.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 10, 2021 at 6:03 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> For comparison's sake, glibc have modified their\\n> collation rules significantly (enough for us to hear complaints about\\n> it) at least twice in the past decade. That's considerably *more*\\n> frequent than DST law changes where I live.\\n\\nYes. It seems to be extremely common for people to get hosed by\\ncollation changes. Different major versions of RHEL ship with\\ndifferent collations. Different minor versions of RHEL ship with\\ndifferent collations. Tiny little changes in very end of the glibc\\nversion number include collation changes. I believe that it's been\\nexplicitly stated by Ulrich Drepper that you should not rely on\\ncollation definitions not to change at any time, and that relying on\\nthem for any sort of on-disk ordering is nuts. Which seems like an\\ninsane idea, because (1) surely the only point of such definitions is\\nto help you sort your data, and you probably don't want to resort it\\nin a continuous loop in case somebody decided to change the collation\\ndefinition under you and (2) how important can it be to continually\\ntinker with the sorting rules?\\n\\nI'm not really convinced that ICU is better, either. I think it's more\\nthat it isn't used as much.\\n\\nI don't have any constructive proposal for what to do about any of\\nthis. It sure is frustrating, though.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 09:21:59 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> I'm not really convinced that ICU is better, either. I think it's more\\n> that it isn't used as much.\\n\\nWell, at least ICU has a notion of attaching versions to collations.\\nHow mindful they are of bumping the version number when necessary\\nremains to be seen. But the POSIX locale APIs don't even offer the\\nopportunity to get it right.\\n\\n> I don't have any constructive proposal for what to do about any of\\n> this. It sure is frustrating, though.\\n\\nYup. If we had reliable ways to detect changes in this sort of\\nenvironment-supplied data, maybe we could do something about it\\n(a la the work that's been happening on attaching collation versions\\nto indexes). But personally I can't summon the motivation to work\\non that, when ICU is the *only* such infrastructure that offers\\nreadily program-readable versioning.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 09:52:52 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Thu, 2021-11-11 at 09:52 -0500, Tom Lane wrote:\\n> Robert Haas <robertmhaas@gmail.com> writes:\\n> > I'm not really convinced that ICU is better, either. I think it's more\\n> > that it isn't used as much.\\n> \\n> Well, at least ICU has a notion of attaching versions to collations.\\n> How mindful they are of bumping the version number when necessary\\n> remains to be seen. But the POSIX locale APIs don't even offer the\\n> opportunity to get it right.\\n\\nAlso, it is much easier *not* to upgrade libicu than it is to *not*\\nupgrade libc, which an essential component of the operating system.\\n\\n> > I don't have any constructive proposal for what to do about any of\\n> > this. It sure is frustrating, though.\\n> \\n> Yup. If we had reliable ways to detect changes in this sort of\\n> environment-supplied data, maybe we could do something about it\\n> (a la the work that's been happening on attaching collation versions\\n> to indexes). But personally I can't summon the motivation to work\\n> on that, when ICU is the *only* such infrastructure that offers\\n> readily program-readable versioning.\\n\\nNobody will want to hear that, but the only really good solution would\\nbe for PostgreSQL to have its own built-in collations.\\n\\nYours,\\nLaurenz Albe\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 17:16:55 +0100\",\n \"msg_from\": \"Laurenz Albe <laurenz.albe@cybertec.at>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"Laurenz Albe <laurenz.albe@cybertec.at> writes:\\n> On Thu, 2021-11-11 at 09:52 -0500, Tom Lane wrote:\\n>> Yup. If we had reliable ways to detect changes in this sort of\\n>> environment-supplied data, maybe we could do something about it\\n>> (a la the work that's been happening on attaching collation versions\\n>> to indexes). But personally I can't summon the motivation to work\\n>> on that, when ICU is the *only* such infrastructure that offers\\n>> readily program-readable versioning.\\n\\n> Nobody will want to hear that, but the only really good solution would\\n> be for PostgreSQL to have its own built-in collations.\\n\\nAnd our own tzdb too? Maybe an outfit like Oracle has the resources\\nand will to maintain their own copies of such data, but I can't see\\nus wanting to do it.\\n\\ntzdb has an additional problem, which is that not updating is not an\\noption: if you're affected by a DST law change, you want that update,\\nand you frequently need it yesterday. We're definitely not set up\\nto handle that sort of update process, which is why we recommend\\n--with-system-tzdata.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 12:07:36 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 11:16 AM Laurenz Albe <laurenz.albe@cybertec.at> wrote:\\n> Nobody will want to hear that, but the only really good solution would\\n> be for PostgreSQL to have its own built-in collations.\\n\\n+1.\\n\\nI agree with Tom that it sounds like a lot of work. And to be honest\\nit's work that I don't really feel very excited about. It would be\\nnecessary to understand not only the bona fide sorting rules of every\\nhuman language out there, which might actually be sort of fun at least\\nfor a while, but also to decide - probably according to some\\nincomprehensible standard - how Japanese katakana ought to sort in\\ncomparison to, say, box-drawing characters, the Mongolian alphabet,\\nand smiley-face emojis. I think it's not particularly likely that\\nthere are a whole lot of documents out there that include all of those\\nthings, but the comparison algorithm has to return something, and\\nprobably there are people who have strong feelings about what the\\nright answers are. That's a pretty unappealing thing to tackle, and I\\nam not volunteering.\\n\\nOn the other hand, if we don't do it, I'm suspicious that things will\\nnever get any better. And that would be sad.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 12:32:59 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> I agree with Tom that it sounds like a lot of work. And to be honest\\n> it's work that I don't really feel very excited about.\\n\\nEven if you were excited about it, would maintaining such data be\\na good use of project resources? It's not like we lack other things\\nwe ought to be doing. I agree that the lack of reliable versioning\\ninfo is a problem, but I can't see that \\\"let's fork ICU and tzdb too\\\"\\nis a good answer.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 13:38:13 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \">Laurenz Albe <laurenz.albe@cybertec.at> writes:\\n>> On Thu, 2021-11-11 at 09:52 -0500, Tom Lane wrote:\\n>>> Yup. If we had reliable ways to detect changes in this sort of\\n>>> environment-supplied data, maybe we could do something about it\\n>>> (a la the work that's been happening on attaching collation versions\\n>>> to indexes). But personally I can't summon the motivation to work\\n>>> on that, when ICU is the *only* such infrastructure that offers\\n>>> readily program-readable versioning.\\n\\n>> Nobody will want to hear that, but the only really good solution would\\n>> be for PostgreSQL to have its own built-in collations.\\n\\n>And our own tzdb too? Maybe an outfit like Oracle has the resources\\n>and will to maintain their own copies of such data, but I can't see\\n>us wanting to do it.\\n\\n>tzdb has an additional problem, which is that not updating is not an\\n>option: if you're affected by a DST law change, you want that update,\\n>and you frequently need it yesterday. We're definitely not set up\\n>to handle that sort of update process, which is why we recommend\\n>--with-system-tzdata.\\n\\nWhere in the docs is this recommended? The only place I can find it is here:\\nhttps://www.postgresql.org/docs/current/install-procedure.html\\n\\nRegards\\nDaniel\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 19:07:46 +0000\",\n \"msg_from\": \"\\\"Daniel Westermann (DWE)\\\" <daniel.westermann@dbi-services.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 1:38 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Even if you were excited about it, would maintaining such data be\\n> a good use of project resources? It's not like we lack other things\\n> we ought to be doing. I agree that the lack of reliable versioning\\n> info is a problem, but I can't see that \\\"let's fork ICU and tzdb too\\\"\\n> is a good answer.\\n\\nYou might be right, but I think it's hard to say for certain. I don't\\nthink this is one of our top 10 problems, but it's probably one of our\\ntop 1000 problems, and it might be one of our top 100 problems. It's\\nentirely subjective, and people are likely to disagree, but based on\\nthose numbers I'd say it's not worth 2% of our resources but it might\\nwell be worth 0.02% of our resources. Everybody's going to have their\\nown opinion, though. I'm not sure how relevant those opinions are in\\nthe end, though. The community has little power to force anybody to\\nwork on anything; people work on what they want to work on, or what\\nthey get paid to work on, not what somebody else in the community\\ndecides is most important.\\n\\nAnyway, from my point of view, if some well-respected community member\\nshowed up and wanted to add a new kind of collation that is provided\\nby PostgreSQL itself and had some well-thought-out candidates for\\ninitial integration, I don't know that it would be smart to turn that\\ndown because solving the whole problem for every case might be more\\nwork than anyone's willing to do. The only real issue for the project\\nis if somebody makes a drive-by contribution of something that's going\\nto need continuous updating. That sort of thing would be bad on\\nmultiple fronts: not only do we not want to get forced into spending\\nongoing maintenance effort on something like this, but we want\\ncollation definitions that *actually don't change*.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 14:09:37 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"\\\"Daniel Westermann (DWE)\\\" <daniel.westermann@dbi-services.com> writes:\\n> On Thu, 2021-11-11 at 09:52 -0500, Tom Lane wrote:\\n>> tzdb has an additional problem, which is that not updating is not an\\n>> option: if you're affected by a DST law change, you want that update,\\n>> and you frequently need it yesterday. We're definitely not set up\\n>> to handle that sort of update process, which is why we recommend\\n>> --with-system-tzdata.\\n\\n> Where in the docs is this recommended? The only place I can find it is here:\\n> https://www.postgresql.org/docs/current/install-procedure.html\\n\\nYup, that's exactly the text I was thinking of. Maybe the recommendation\\nshould be more enthusiastic --- it was written back when it was still\\nrather questionable whether a platform would have an up-to-date copy of\\ntzdata. (Maybe it still is, at least for the \\\"up-to-date\\\" part.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 14:17:59 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> ... but we want\\n> collation definitions that *actually don't change*.\\n\\nUm ... how would that work? Unicode is a moving target. Even without\\ntheir continual addition of stuff, I'm not convinced that social rules\\nabout how to sort are engraved on stone tablets. The need for collation\\nupdates may not be as predictable as the need for timezone updates,\\nbut I doubt that we can just freeze the data forever.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 14:23:34 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 2:23 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Robert Haas <robertmhaas@gmail.com> writes:\\n> > ... but we want\\n> > collation definitions that *actually don't change*.\\n>\\n> Um ... how would that work? Unicode is a moving target. Even without\\n> their continual addition of stuff, I'm not convinced that social rules\\n> about how to sort are engraved on stone tablets. The need for collation\\n> updates may not be as predictable as the need for timezone updates,\\n> but I doubt that we can just freeze the data forever.\\n\\nI don't know, but I think the social rules that actually matter change\\nextremely slowly. To my knowledge, the alphabet song has not changed\\nsince I was in kindergarten. Now I agree that in some countries it\\nprobably has ... but I doubt those events are super-common, because a\\ncountry does change its definition of alphabetical order, there's a\\nheck of a lot more updating to do than just reindexing your PostgreSQL\\ndatabases. The signs saying A-L go to the left and M-Z go to the right\\nwill need revision if we decide M comes before L. I feel like it has\\nto be the case that most of the updates that are being made involve\\nthings like how obscure characters compare to other obscure\\ncharacters, or what to do in corner-case situations involving multiple\\ndiacritical marks. I know I've seen collation changes on Macs that\\nchanged the order in which en_US.UTF8 strings sorted. But it wasn't\\nthat the rules about English sorting have actually changed. It was\\nthat somebody somewhere decided that the algorithm should be more or\\nless case-sensitive, or that we ought to ignore the amount of\\nwhitespace between words instead of not ignoring it, or I don't know\\nexactly, but not anything that people universally agree on. Tinkering\\nwith obscure rules that actual human beings wouldn't agree on and\\nprioritizing that over a stable algorithm is, IMHO, ridiculous.\\n\\nIf the Unicode consortium introduces a new emoji for \\\"annoyed\\nPostgreSQL hacker,\\\" I really do not care whether that collates before\\nor after the existing symbol for \\\"floral heart bullet, reversed\\nrotated.\\\" I care much more about whether it collates the same way\\nafter the next minor release as it does the day it's released. And I\\nseriously doubt that I am alone in that.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 14:41:47 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Thu, 11 Nov 2021 at 14:42, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n\\n> diacritical marks. I know I've seen collation changes on Macs that\\n> changed the order in which en_US.UTF8 strings sorted. But it wasn't\\n> that the rules about English sorting have actually changed. It was\\n> that somebody somewhere decided that the algorithm should be more or\\n> less case-sensitive, or that we ought to ignore the amount of\\n> whitespace between words instead of not ignoring it, or I don't know\\n> exactly, but not anything that people universally agree on. Tinkering\\n> with obscure rules that actual human beings wouldn't agree on and\\n> prioritizing that over a stable algorithm is, IMHO, ridiculous.\\n>\\n\\nYes, I thought the point here was to nail down each change as a separate\\nversion. So for example maybe I'm running Universal Compare Everything\\nCollation v1.2435 while your database is running Universal Compare\\nEverything Collation v1.2436, with the only difference being whether e\\ndiaresis circumflex comes before or after e circumflex diaresis. If I do a\\nsystem upgrade I won't just silently corrupt any indexes with those\\ncharacters; instead I'll be told that my collation is out of date and then\\nI can decide whether to stick with the old collation or rebuild my indexes\\nand upgrade.\\n\\nThere is however one kind of change at least that I think can be made\\nsafely: adding a new character in between existing characters. That\\nshouldn't affect any existing indexes.\\n\\nIf the Unicode consortium introduces a new emoji for \\\"annoyed\\n> PostgreSQL hacker,\\\" I really do not care whether that collates before\\n> or after the existing symbol for \\\"floral heart bullet, reversed\\n> rotated.\\\" I care much more about whether it collates the same way\\n> after the next minor release as it does the day it's released. And I\\n> seriously doubt that I am alone in that.\\n>\\n\\nOn Thu, 11 Nov 2021 at 14:42, Robert Haas <robertmhaas@gmail.com> wrote: \\ndiacritical marks. I know I've seen collation changes on Macs that\\nchanged the order in which en_US.UTF8 strings sorted. But it wasn't\\nthat the rules about English sorting have actually changed. It was\\nthat somebody somewhere decided that the algorithm should be more or\\nless case-sensitive, or that we ought to ignore the amount of\\nwhitespace between words instead of not ignoring it, or I don't know\\nexactly, but not anything that people universally agree on. Tinkering\\nwith obscure rules that actual human beings wouldn't agree on and\\nprioritizing that over a stable algorithm is, IMHO, ridiculous.Yes, I thought the point here was to nail down each change as a separate version. So for example maybe I'm running Universal Compare Everything Collation v1.2435 while your database is running Universal Compare Everything Collation v1.2436, with the only difference being whether e diaresis circumflex comes before or after e circumflex diaresis. If I do a system upgrade I won't just silently corrupt any indexes with those characters; instead I'll be told that my collation is out of date and then I can decide whether to stick with the old collation or rebuild my indexes and upgrade.There is however one kind of change at least that I think can be made safely: adding a new character in between existing characters. That shouldn't affect any existing indexes.\\nIf the Unicode consortium introduces a new emoji for \\\"annoyed\\nPostgreSQL hacker,\\\" I really do not care whether that collates before\\nor after the existing symbol for \\\"floral heart bullet, reversed\\nrotated.\\\" I care much more about whether it collates the same way\\nafter the next minor release as it does the day it's released. And I\\nseriously doubt that I am alone in that.\",\n \"msg_date\": \"Thu, 11 Nov 2021 15:45:48 -0500\",\n \"msg_from\": \"Isaac Morland <isaac.morland@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 3:45 PM Isaac Morland <isaac.morland@gmail.com> wrote:\\n> There is however one kind of change at least that I think can be made safely: adding a new character in between existing characters. That shouldn't affect any existing indexes.\\n\\nOnly if you can guarantee that said character is not present already.\\nI don't think we update the end of the acceptable code point range\\nevery time that Unicode adds new stuff, so probably those things are\\nsubject to some default rule unless and until someone installs\\nsomething more specific. Therefore I doubt that even this case is\\ntruly safe.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 16:08:22 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Thu, 11 Nov 2021 at 16:08, Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Thu, Nov 11, 2021 at 3:45 PM Isaac Morland <isaac.morland@gmail.com>\\n> wrote:\\n> > There is however one kind of change at least that I think can be made\\n> safely: adding a new character in between existing characters. That\\n> shouldn't affect any existing indexes.\\n>\\n> Only if you can guarantee that said character is not present already.\\n> I don't think we update the end of the acceptable code point range\\n> every time that Unicode adds new stuff, so probably those things are\\n> subject to some default rule unless and until someone installs\\n> something more specific. Therefore I doubt that even this case is\\n> truly safe.\\n>\\n\\nWouldn't an existing index only have characters that were already part of\\nthe collation? Attempting to use one not covered by the collation I would\\nhave expected to cause an error at insert time. But definitely I agree I\\nwouldn't feel confident about the safety of any change.\\n\\nOn Thu, 11 Nov 2021 at 16:08, Robert Haas <robertmhaas@gmail.com> wrote:On Thu, Nov 11, 2021 at 3:45 PM Isaac Morland <isaac.morland@gmail.com> wrote:\\n> There is however one kind of change at least that I think can be made safely: adding a new character in between existing characters. That shouldn't affect any existing indexes.\\n\\nOnly if you can guarantee that said character is not present already.\\nI don't think we update the end of the acceptable code point range\\nevery time that Unicode adds new stuff, so probably those things are\\nsubject to some default rule unless and until someone installs\\nsomething more specific. Therefore I doubt that even this case is\\ntruly safe.Wouldn't an existing index only have characters that were already part of the collation? Attempting to use one not covered by the collation I would have expected to cause an error at insert time. But definitely I agree I wouldn't feel confident about the safety of any change.\",\n \"msg_date\": \"Thu, 11 Nov 2021 17:04:39 -0500\",\n \"msg_from\": \"Isaac Morland <isaac.morland@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 5:04 PM Isaac Morland <isaac.morland@gmail.com> wrote:\\n> Wouldn't an existing index only have characters that were already part of the collation? Attempting to use one not covered by the collation I would have expected to cause an error at insert time. But definitely I agree I wouldn't feel confident about the safety of any change.\\n\\nI mean it's not like we are updating the definition of\\npg_utf8_verifychar() every time they define a new code point.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 18:08:47 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 12:09 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Thu, Nov 11, 2021 at 5:04 PM Isaac Morland <isaac.morland@gmail.com> wrote:\\n> > Wouldn't an existing index only have characters that were already part of the collation? Attempting to use one not covered by the collation I would have expected to cause an error at insert time. But definitely I agree I wouldn't feel confident about the safety of any change.\\n>\\n> I mean it's not like we are updating the definition of\\n> pg_utf8_verifychar() every time they define a new code point.\\n\\nRight, and there may be other systems that do this. That is, reject\\ninvalid code points, because they have no sort order. You can see\\nsome sign of this in the major and minor collation version numbers\\nreported by Windows (though I'm not sure if this was lost with the\\nrecent move to ICU): if only the minor version changes, the\\ndocumentation says it means \\\"we only added new code points, no\\nexisting code points changed\\\", so a sufficiently clever program\\ndoesn't need to rebuild its persistent ordered data structures, if it\\nnever allowed any unknown code points into the structure before. Not\\nonly does PostgreSQL not have the logic for that, it also doesn't have\\nthe data: the set of existing code points has to be the one used by\\nthe collation provider, and the collection providers we have don't\\nreject unknown code points on comparison.\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 12:31:18 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On 11.11.21 18:32, Robert Haas wrote:\\n> I agree with Tom that it sounds like a lot of work. And to be honest\\n> it's work that I don't really feel very excited about. It would be\\n> necessary to understand not only the bona fide sorting rules of every\\n> human language out there, which might actually be sort of fun at least\\n> for a while, but also to decide - probably according to some\\n> incomprehensible standard - how Japanese katakana ought to sort in\\n> comparison to, say, box-drawing characters, the Mongolian alphabet,\\n> and smiley-face emojis. I think it's not particularly likely that\\n> there are a whole lot of documents out there that include all of those\\n> things, but the comparison algorithm has to return something, and\\n> probably there are people who have strong feelings about what the\\n> right answers are. That's a pretty unappealing thing to tackle, and I\\n> am not volunteering.\\n> \\n> On the other hand, if we don't do it, I'm suspicious that things will\\n> never get any better. And that would be sad.\\n\\nThere are standards for sort order, and the major hiccups we had in the \\npast were mostly moving from older versions of those standards to newer \\nversions. So at some point this should stabilize.\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 14:42:30 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 8:42 AM Peter Eisentraut\\n<peter.eisentraut@enterprisedb.com> wrote:\\n> There are standards for sort order, and the major hiccups we had in the\\n> past were mostly moving from older versions of those standards to newer\\n> versions. So at some point this should stabilize.\\n\\nOnly if they don't keep making new versions of the standards.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 09:49:12 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 09:52:52AM -0500, Tom Lane wrote:\\n> Robert Haas <robertmhaas@gmail.com> writes:\\n> > I'm not really convinced that ICU is better, either. I think it's more\\n> > that it isn't used as much.\\n> \\n> Well, at least ICU has a notion of attaching versions to collations.\\n> How mindful they are of bumping the version number when necessary\\n> remains to be seen. But the POSIX locale APIs don't even offer the\\n> opportunity to get it right.\\n> \\n> > I don't have any constructive proposal for what to do about any of\\n> > this. It sure is frustrating, though.\\n> \\n> Yup. If we had reliable ways to detect changes in this sort of\\n> environment-supplied data, maybe we could do something about it\\n> (a la the work that's been happening on attaching collation versions\\n> to indexes). But personally I can't summon the motivation to work\\n\\n Theoretically there are versions attached to collations already:\\nthe collation in index is an oid referencing the pg_collation.\\n And the pg_collation already has versions.\\n\\n Currently for glibc the version looks like glibc version at\\ninitdb, and that doesn't seem very reliable, but that could be a\\ndifferent task (to find LC_COLLATE file and put hash of the\\nusuable data into version string, for example).\\n\\n Currently, it is questionable how to work with the different\\nversions of collations -- but that could be solved e.g. via ap-\\npropriate naming. Perhaps \\\"collation@ver\\\" ? But if the version\\nwould contain a hash, a full version could be a bit dubious.\\n And some database maintainance task could check that all the old\\ncollations are available, rename them as needed, and create a set\\nthe new ones.\\n Automatically invalidating all the indexes, unfortunately.\\n\\n> on that, when ICU is the *only* such infrastructure that offers\\n> readily program-readable versioning.\\n> \\n> \\t\\t\\tregards, tom lane\\n> \\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 01:47:17 +0300\",\n \"msg_from\": \"Ilya Anfimov <ilan@tzirechnoy.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 13, 2021 at 11:47 AM Ilya Anfimov <ilan@tzirechnoy.com> wrote:\\n> Currently for glibc the version looks like glibc version at\\n> initdb, and that doesn't seem very reliable, but that could be a\\n> different task (to find LC_COLLATE file and put hash of the\\n> usuable data into version string, for example).\\n\\nYeah, I had a system exactly like that working (that is, a way to run\\narbitrary commands to capture version strings, that could be used to\\nhash your collation definition files, patches somewhere in the\\narchives), but then we thought it'd be better to use glibc versions,\\nand separately, to perhaps try to ask the glibc people to expose a\\nversion. FreeBSD (at my request), Windows and ICU do expose versions\\nin a straightforward way, and we capture those.\\n\\n> Currently, it is questionable how to work with the different\\n> versions of collations -- but that could be solved e.g. via ap-\\n> propriate naming. Perhaps \\\"collation@ver\\\" ? But if the version\\n> would contain a hash, a full version could be a bit dubious.\\n> And some database maintainance task could check that all the old\\n> collations are available, rename them as needed, and create a set\\n> the new ones.\\n> Automatically invalidating all the indexes, unfortunately.\\n\\nWe built a system that at least detected the changes on a per-index\\nlevel, but failed to ship it in release 14. See revert commit, and\\nlinks back to previous commits and discussion:\\n\\nhttps://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=ec48314708262d8ea6cdcb83f803fc83dd89e721\\n\\nIt didn't invalidate indexes, but complained about each individual\\nindex on first access in each session, until you REINDEXed it. We\\nwill try again :-)\\n\\nIn the archives you can find discussions of how to make a system that\\ntolerates multiple version existing at the same time as I think you're\\ngetting at, like DB2 does. It's tricky, because it's code and data.\\nDB2 ships with N copies of ICU in it.\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 13:09:33 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should AT TIME ZONE be volatile?\"\n }\n]","isConversational":false}}},{"rowIdx":53070,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello,\n\nafter plain ./configure without options (including noticeable absence of \n--with-openssl) and make,\nmake -C contrib COPT=-Werror gives error with gcc-11 on REL_9_6_STABLE, \nREL_10_STABLE or REL_11_STABLE.\n\nThe warning/error is about misleading indent in \ncontrib/pgcrypto/imath.c's CLAMP macro:\n\nimath.c: In function ‘mp_int_add’:\nimath.c:133:1: error: this ‘while’ clause does not guard... \n[-Werror=misleading-indentation]\n 133 | while(uz_ > 1 && (*dz_-- == 0)) --uz_;MP_USED(z_)=uz_;}while(0)\n | ^~~~~\nimath.c:670:17: note: in expansion of macro ‘CLAMP’\n 670 | CLAMP(c);\n | ^~~~~\nIn file included from imath.c:34:\nimath.h:68:26: note: ...this statement, but the latter is misleadingly \nindented as if it were guarded by the ‘while’\n 68 | #define MP_USED(Z) ((Z)->used)\n | ^\nimath.c:133:39: note: in expansion of macro ‘MP_USED’\n 133 | while(uz_ > 1 && (*dz_-- == 0)) --uz_;MP_USED(z_)=uz_;}while(0)\n | ^~~~~~~\nimath.c:670:17: note: in expansion of macro ‘CLAMP’\n 670 | CLAMP(c);\n | ^~~~~\n\npgcrypto-fix-imath-clamp-warning.patch, attached, is a suggested minimal \nfix:\ndiff --git a/contrib/pgcrypto/imath.c b/contrib/pgcrypto/imath.c\nindex b94a51b81a4..801b843cbe3 100644\n--- a/contrib/pgcrypto/imath.c\n+++ b/contrib/pgcrypto/imath.c\n@@ -130,7 +130,8 @@ do{T *u_=(A),*v_=u_+(N)-1;while(u_<v_){T \nxch=*u_;*u_++=*v_;*v_--=xch;}}while(0)\n #else\n #define CLAMP(Z) \\\n do{mp_int z_=(Z);mp_size uz_=MP_USED(z_);mp_digit \n*dz_=MP_DIGITS(z_)+uz_-1;\\\n-while(uz_ > 1 && (*dz_-- == 0)) --uz_;MP_USED(z_)=uz_;}while(0)\n+while(uz_ > 1 && (*dz_-- == 0)) --uz_;\\\n+MP_USED(z_)=uz_;}while(0)\n #endif\n\n #undef MIN\n\nThe same patch works for all 9.6, 10 and 11. It's not needed in 12 or \nlater, they compile without warnings just fine even without --with-openssl.\n\nIn addition, 9.6 (unlike 10 and 11) gives four \"array-parameter=\" \nwarnings about contrib/pgcrypto/sha2.c:\n\nsha2.c:552:20: error: argument 1 of type ‘uint8[]’ {aka ‘unsigned \nchar[]’} with mismatched bound [-Werror=array-parameter=]\n 552 | SHA256_Final(uint8 digest[], SHA256_CTX *context)\n | ~~~~~~^~~~~~~~\nIn file included from sha2.c:44:\nsha2.h:93:30: note: previously declared as ‘uint8[32]’ {aka ‘unsigned \nchar[32]’}\n 93 | void SHA256_Final(uint8[SHA256_DIGEST_LENGTH], \nSHA256_CTX *);\n | ^~~~~~~~~~~~~~~~~~~~~~~~~~~\n\nand the same for SHA{512,384,224}_Final.\n\npgcrypto-fix-sha2-warning.patch, attached, is a suggested minimal fix \nfor that:\n void\n-SHA256_Final(uint8 digest[], SHA256_CTX *context)\n+SHA256_Final(uint8 digest[SHA256_DIGEST_LENGTH], SHA256_CTX *context)\n {\netc.\n\nPlease consider fixing those warnings.\n\n-- \nAnton Voloshin\nPostgres Professional, The Russian Postgres Company\nhttps://postgrespro.ru","msg_date":"Wed, 10 Nov 2021 17:14:51 +0700","msg_from":"Anton Voloshin <a.voloshin@postgrespro.ru>","msg_from_op":true,"msg_subject":"fix warnings in 9.6, 10, 11's contrib when compiling without openssl"},{"msg_contents":"> On 10 Nov 2021, at 11:14, Anton Voloshin <a.voloshin@postgrespro.ru> wrote:\n\n> The warning/error is about misleading indent in contrib/pgcrypto/imath.c's CLAMP macro:\n\nThis is upstream code in a stable branch formatted by pgindent at the time of\nthe release, I'm not sure the churn is worth it for fixing warnings for\nnon-standard compiler options here. I don't see any of these warnings skimming\nlogs in the buildfarm.\n\n> In addition, 9.6 (unlike 10 and 11) gives four \"array-parameter=\" warnings about contrib/pgcrypto/sha2.c:\n\n9.6 will be EOL from the release tomorrow, no more fixes are applied to that\nbranch.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Wed, 10 Nov 2021 14:56:00 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: fix warnings in 9.6, 10, 11's contrib when compiling without\n openssl"},{"msg_contents":"On Wed, Nov 10, 2021 at 02:56:00PM +0100, Daniel Gustafsson wrote:\n> This is upstream code in a stable branch formatted by pgindent at the time of\n> the release, I'm not sure the churn is worth it for fixing warnings for\n> non-standard compiler options here. I don't see any of these warnings skimming\n> logs in the buildfarm.\n\nI would suggest to send any improvements for imath directly to\nupstream, then Postgres would just feed on that when we update it:\nhttps://github.com/creachadair/imath\n\nPeter has removed this code as of db7d1a7, but there could still be a\npoint in updating imath on the stable branches where it exists if\nthere is an issue with it impacting pgcrypto without OpenSSL.\n--\nMichael","msg_date":"Thu, 11 Nov 2021 10:14:34 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: fix warnings in 9.6, 10, 11's contrib when compiling without\n openssl"},{"msg_contents":"On 11/11/2021 08:14, Michael Paquier wrote:\n> I would suggest to send any improvements for imath directly to\n> upstream, then Postgres would just feed on that when we update it:\n> https://github.com/creachadair/imath\n\nI didn't realise imath comes (occasionally) from external repo. Thank you.\n\nThe compilation warnings are only relevant for their old releases, \nbefore imath v1.27 where they've replaced macros with inline functions, see\nhttps://github.com/creachadair/imath/commit/d4760eef8\n\nSo imath in upstream doesn't require updates (especially since those \nwarnings come specifically from pgindent's work, not from the upstream \nsource). But Postgres' 10_STABLE and 11_STABLE still use old version \nwith CLAMP as a macro.\n\nNoah Misch was last to update imath from upstream in February 2019 and \nthat change became part of REL_12_STABLE and later:\n > Import changes from IMath versions (1.3, 1.29].\nhttps://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=48e24ba6b7fd3bfd156b51e8d768fd48df0d288b\n\nNoah and others, do you think it would be a good idea to update imath in \nREL_10_STABLE and REL_11_STABLE to a later upstream version to avoid \nthose warnings about misleading indent? (see beginning of this thread) \nAdmittedly, those warnings can only be seen in unusual circumstances (no \nopenssl), but still, as you have mentioned in your commit message,\n > We're better off naively tracking upstream than reactively\n > maintaining a twelve-year-old snapshot of upstream.\n\nPerhaps even update imath in all relevant stable branches, 10..14 \ninclusive, to the same upstream version?\n\nOn 11/11/2021 08:14, Michael Paquier wrote:\n> Peter has removed this code as of db7d1a7, but there could still be a\n> point in updating imath on the stable branches where it exists if\n> there is an issue with it impacting pgcrypto without OpenSSL.\n\nPerhaps that would be a good idea, see above.\n\n-- \nAnton Voloshin\nhttps://postgrespro.ru\nPostgres Professional, The Russian Postgres Company\n\n\n","msg_date":"Thu, 11 Nov 2021 20:29:52 +0700","msg_from":"Anton Voloshin <a.voloshin@postgrespro.ru>","msg_from_op":true,"msg_subject":"Re: fix warnings in 9.6, 10, 11's contrib when compiling without\n openssl"}],"string":"[\n {\n \"msg_contents\": \"Hello,\\n\\nafter plain ./configure without options (including noticeable absence of \\n--with-openssl) and make,\\nmake -C contrib COPT=-Werror gives error with gcc-11 on REL_9_6_STABLE, \\nREL_10_STABLE or REL_11_STABLE.\\n\\nThe warning/error is about misleading indent in \\ncontrib/pgcrypto/imath.c's CLAMP macro:\\n\\nimath.c: In function ‘mp_int_add’:\\nimath.c:133:1: error: this ‘while’ clause does not guard... \\n[-Werror=misleading-indentation]\\n 133 | while(uz_ > 1 && (*dz_-- == 0)) --uz_;MP_USED(z_)=uz_;}while(0)\\n | ^~~~~\\nimath.c:670:17: note: in expansion of macro ‘CLAMP’\\n 670 | CLAMP(c);\\n | ^~~~~\\nIn file included from imath.c:34:\\nimath.h:68:26: note: ...this statement, but the latter is misleadingly \\nindented as if it were guarded by the ‘while’\\n 68 | #define MP_USED(Z) ((Z)->used)\\n | ^\\nimath.c:133:39: note: in expansion of macro ‘MP_USED’\\n 133 | while(uz_ > 1 && (*dz_-- == 0)) --uz_;MP_USED(z_)=uz_;}while(0)\\n | ^~~~~~~\\nimath.c:670:17: note: in expansion of macro ‘CLAMP’\\n 670 | CLAMP(c);\\n | ^~~~~\\n\\npgcrypto-fix-imath-clamp-warning.patch, attached, is a suggested minimal \\nfix:\\ndiff --git a/contrib/pgcrypto/imath.c b/contrib/pgcrypto/imath.c\\nindex b94a51b81a4..801b843cbe3 100644\\n--- a/contrib/pgcrypto/imath.c\\n+++ b/contrib/pgcrypto/imath.c\\n@@ -130,7 +130,8 @@ do{T *u_=(A),*v_=u_+(N)-1;while(u_<v_){T \\nxch=*u_;*u_++=*v_;*v_--=xch;}}while(0)\\n #else\\n #define CLAMP(Z) \\\\\\n do{mp_int z_=(Z);mp_size uz_=MP_USED(z_);mp_digit \\n*dz_=MP_DIGITS(z_)+uz_-1;\\\\\\n-while(uz_ > 1 && (*dz_-- == 0)) --uz_;MP_USED(z_)=uz_;}while(0)\\n+while(uz_ > 1 && (*dz_-- == 0)) --uz_;\\\\\\n+MP_USED(z_)=uz_;}while(0)\\n #endif\\n\\n #undef MIN\\n\\nThe same patch works for all 9.6, 10 and 11. It's not needed in 12 or \\nlater, they compile without warnings just fine even without --with-openssl.\\n\\nIn addition, 9.6 (unlike 10 and 11) gives four \\\"array-parameter=\\\" \\nwarnings about contrib/pgcrypto/sha2.c:\\n\\nsha2.c:552:20: error: argument 1 of type ‘uint8[]’ {aka ‘unsigned \\nchar[]’} with mismatched bound [-Werror=array-parameter=]\\n 552 | SHA256_Final(uint8 digest[], SHA256_CTX *context)\\n | ~~~~~~^~~~~~~~\\nIn file included from sha2.c:44:\\nsha2.h:93:30: note: previously declared as ‘uint8[32]’ {aka ‘unsigned \\nchar[32]’}\\n 93 | void SHA256_Final(uint8[SHA256_DIGEST_LENGTH], \\nSHA256_CTX *);\\n | ^~~~~~~~~~~~~~~~~~~~~~~~~~~\\n\\nand the same for SHA{512,384,224}_Final.\\n\\npgcrypto-fix-sha2-warning.patch, attached, is a suggested minimal fix \\nfor that:\\n void\\n-SHA256_Final(uint8 digest[], SHA256_CTX *context)\\n+SHA256_Final(uint8 digest[SHA256_DIGEST_LENGTH], SHA256_CTX *context)\\n {\\netc.\\n\\nPlease consider fixing those warnings.\\n\\n-- \\nAnton Voloshin\\nPostgres Professional, The Russian Postgres Company\\nhttps://postgrespro.ru\",\n \"msg_date\": \"Wed, 10 Nov 2021 17:14:51 +0700\",\n \"msg_from\": \"Anton Voloshin <a.voloshin@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"fix warnings in 9.6, 10, 11's contrib when compiling without openssl\"\n },\n {\n \"msg_contents\": \"> On 10 Nov 2021, at 11:14, Anton Voloshin <a.voloshin@postgrespro.ru> wrote:\\n\\n> The warning/error is about misleading indent in contrib/pgcrypto/imath.c's CLAMP macro:\\n\\nThis is upstream code in a stable branch formatted by pgindent at the time of\\nthe release, I'm not sure the churn is worth it for fixing warnings for\\nnon-standard compiler options here. I don't see any of these warnings skimming\\nlogs in the buildfarm.\\n\\n> In addition, 9.6 (unlike 10 and 11) gives four \\\"array-parameter=\\\" warnings about contrib/pgcrypto/sha2.c:\\n\\n9.6 will be EOL from the release tomorrow, no more fixes are applied to that\\nbranch.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 14:56:00 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: fix warnings in 9.6, 10, 11's contrib when compiling without\\n openssl\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 10, 2021 at 02:56:00PM +0100, Daniel Gustafsson wrote:\\n> This is upstream code in a stable branch formatted by pgindent at the time of\\n> the release, I'm not sure the churn is worth it for fixing warnings for\\n> non-standard compiler options here. I don't see any of these warnings skimming\\n> logs in the buildfarm.\\n\\nI would suggest to send any improvements for imath directly to\\nupstream, then Postgres would just feed on that when we update it:\\nhttps://github.com/creachadair/imath\\n\\nPeter has removed this code as of db7d1a7, but there could still be a\\npoint in updating imath on the stable branches where it exists if\\nthere is an issue with it impacting pgcrypto without OpenSSL.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 11 Nov 2021 10:14:34 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: fix warnings in 9.6, 10, 11's contrib when compiling without\\n openssl\"\n },\n {\n \"msg_contents\": \"On 11/11/2021 08:14, Michael Paquier wrote:\\n> I would suggest to send any improvements for imath directly to\\n> upstream, then Postgres would just feed on that when we update it:\\n> https://github.com/creachadair/imath\\n\\nI didn't realise imath comes (occasionally) from external repo. Thank you.\\n\\nThe compilation warnings are only relevant for their old releases, \\nbefore imath v1.27 where they've replaced macros with inline functions, see\\nhttps://github.com/creachadair/imath/commit/d4760eef8\\n\\nSo imath in upstream doesn't require updates (especially since those \\nwarnings come specifically from pgindent's work, not from the upstream \\nsource). But Postgres' 10_STABLE and 11_STABLE still use old version \\nwith CLAMP as a macro.\\n\\nNoah Misch was last to update imath from upstream in February 2019 and \\nthat change became part of REL_12_STABLE and later:\\n > Import changes from IMath versions (1.3, 1.29].\\nhttps://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=48e24ba6b7fd3bfd156b51e8d768fd48df0d288b\\n\\nNoah and others, do you think it would be a good idea to update imath in \\nREL_10_STABLE and REL_11_STABLE to a later upstream version to avoid \\nthose warnings about misleading indent? (see beginning of this thread) \\nAdmittedly, those warnings can only be seen in unusual circumstances (no \\nopenssl), but still, as you have mentioned in your commit message,\\n > We're better off naively tracking upstream than reactively\\n > maintaining a twelve-year-old snapshot of upstream.\\n\\nPerhaps even update imath in all relevant stable branches, 10..14 \\ninclusive, to the same upstream version?\\n\\nOn 11/11/2021 08:14, Michael Paquier wrote:\\n> Peter has removed this code as of db7d1a7, but there could still be a\\n> point in updating imath on the stable branches where it exists if\\n> there is an issue with it impacting pgcrypto without OpenSSL.\\n\\nPerhaps that would be a good idea, see above.\\n\\n-- \\nAnton Voloshin\\nhttps://postgrespro.ru\\nPostgres Professional, The Russian Postgres Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 20:29:52 +0700\",\n \"msg_from\": \"Anton Voloshin <a.voloshin@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: fix warnings in 9.6, 10, 11's contrib when compiling without\\n openssl\"\n }\n]","isConversational":false}}},{"rowIdx":53071,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nAs discussed in [1], isn't it a better idea to add some of activity\nmessages [2] such as recovery, archive, backup, streaming etc. to\nserver logs at LOG level? They are currently being set into ps display\nwhich is good if the postgres is being run on a standalone box/VM\nwhere users can see the ps display, but it doesn't help much in case\nthe postgres is being run on a cloud environment where users don't\nhave access to ps display output. Moreover, the ps display is\ntransient and will not help to analyze after an issue occurs.\n\nHaving the above messages in the server logs will be useful to\nunderstand how the system is/was doing/progressing with these\n(sometimes time-intensive) operations.\n\nThoughts?\n\n[1] - https://www.postgresql.org/message-id/CALj2ACWE6D%2BbbKnA9tQ0d-pPN1n4xrsFPOmiURm2my3qt0w3OA%40mail.gmail.com\n[2] -\nsnprintf(activitymsg, sizeof(activitymsg), \"waiting for %s\",\nsnprintf(activitymsg, sizeof(activitymsg), \"recovering %s\",\nsnprintf(activitymsg, sizeof(activitymsg), \"performing %s%s%s\",\nsnprintf(activitymsg, sizeof(activitymsg), \"archiving %s\", xlog);\nsnprintf(activitymsg, sizeof(activitymsg), \"failed on %s\", xlog);\nelog(DEBUG1, \"archived write-ahead log file \\\"%s\\\"\", xlog);\nsnprintf(activitymsg, sizeof(activitymsg), \"sending backup \\\"%s\\\"\",\nsnprintf(activitymsg, sizeof(activitymsg), \"restarting at %X/%X\",\nsnprintf(activitymsg, sizeof(activitymsg), \"streaming %X/%X\",\nsnprintf(activitymsg, sizeof(activitymsg), \"streaming %X/%X\",\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Wed, 10 Nov 2021 23:08:51 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"add recovery, backup, archive, streaming etc. activity messages to\n server logs along with ps display"},{"msg_contents":"On 11/10/21, 9:43 AM, \"Bharath Rupireddy\" <bharath.rupireddyforpostgres@gmail.com> wrote:\r\n> As discussed in [1], isn't it a better idea to add some of activity\r\n> messages [2] such as recovery, archive, backup, streaming etc. to\r\n> server logs at LOG level? They are currently being set into ps display\r\n> which is good if the postgres is being run on a standalone box/VM\r\n> where users can see the ps display, but it doesn't help much in case\r\n> the postgres is being run on a cloud environment where users don't\r\n> have access to ps display output. Moreover, the ps display is\r\n> transient and will not help to analyze after an issue occurs.\r\n>\r\n> Having the above messages in the server logs will be useful to\r\n> understand how the system is/was doing/progressing with these\r\n> (sometimes time-intensive) operations.\r\n\r\nI think this would make the logs far too noisy for many servers. For\r\narchiving alone, this could cause tens of thousands more log messages\r\nper hour on a busy system. I think you can already see such\r\ninformation at a debug level, anyway.\r\n\r\nNathan\r\n\r\n","msg_date":"Wed, 10 Nov 2021 18:31:20 +0000","msg_from":"\"Bossart, Nathan\" <bossartn@amazon.com>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive,\n streaming etc. activity messages to server logs along with ps display"},{"msg_contents":"On Wed, Nov 10, 2021, at 2:38 PM, Bharath Rupireddy wrote:\n> As discussed in [1], isn't it a better idea to add some of activity\n> messages [2] such as recovery, archive, backup, streaming etc. to\n> server logs at LOG level? They are currently being set into ps display\n> which is good if the postgres is being run on a standalone box/VM\n> where users can see the ps display, but it doesn't help much in case\n> the postgres is being run on a cloud environment where users don't\n> have access to ps display output. Moreover, the ps display is\n> transient and will not help to analyze after an issue occurs.\nBesides recovery, the other activities already provide information through\nviews. I might be missing something but it seems the current views already\nprovide the information that ps displays.\n\n> Having the above messages in the server logs will be useful to\n> understand how the system is/was doing/progressing with these\n> (sometimes time-intensive) operations.\nIt could be useful for a root cause analysis, however, you are also increasing\nthe number of LOG messages in a system that doesn't/won't require such\ninformation. It is fine to add additional DEBUG messages if there isn't a\nsimilar one yet. If at least the message level were module-controlled, you\ncould modify a setting to gather more messages from a specific module.\nUnfortunately, that's not possible so we should avoid superfluous messages.\n\n\n--\nEuler Taveira\nEDB https://www.enterprisedb.com/\n\nOn Wed, Nov 10, 2021, at 2:38 PM, Bharath Rupireddy wrote:As discussed in [1], isn't it a better idea to add some of activitymessages [2] such as recovery, archive, backup, streaming etc. toserver logs at LOG level? They are currently being set into ps displaywhich is good if the postgres is being run on a standalone box/VMwhere users can see the ps display, but it doesn't help much in casethe postgres is being run on a cloud environment where users don'thave access to ps display output. Moreover, the ps display istransient and will not help to analyze after an issue occurs.Besides recovery, the other activities already provide information throughviews. I might be missing something but it seems the current views alreadyprovide the information that ps displays.Having the above messages in the server logs will be useful tounderstand how the system is/was doing/progressing with these(sometimes time-intensive) operations.It could be useful for a root cause analysis, however, you are also increasingthe number of LOG messages in a system that doesn't/won't require suchinformation. It is fine to add additional DEBUG messages if there isn't asimilar one yet. If at least the message level were module-controlled, youcould modify a setting to gather more messages from a specific module.Unfortunately, that's not possible so we should avoid superfluous messages.--Euler TaveiraEDB https://www.enterprisedb.com/","msg_date":"Wed, 10 Nov 2021 17:09:43 -0300","msg_from":"\"Euler Taveira\" <euler@eulerto.com>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive,\n streaming etc. activity messages to server\n logs along with ps display"},{"msg_contents":"On Thu, Nov 11, 2021 at 12:01 AM Bossart, Nathan <bossartn@amazon.com> wrote:\n>\n> On 11/10/21, 9:43 AM, \"Bharath Rupireddy\" <bharath.rupireddyforpostgres@gmail.com> wrote:\n> > As discussed in [1], isn't it a better idea to add some of activity\n> > messages [2] such as recovery, archive, backup, streaming etc. to\n> > server logs at LOG level? They are currently being set into ps display\n> > which is good if the postgres is being run on a standalone box/VM\n> > where users can see the ps display, but it doesn't help much in case\n> > the postgres is being run on a cloud environment where users don't\n> > have access to ps display output. Moreover, the ps display is\n> > transient and will not help to analyze after an issue occurs.\n> >\n> > Having the above messages in the server logs will be useful to\n> > understand how the system is/was doing/progressing with these\n> > (sometimes time-intensive) operations.\n>\n> I think this would make the logs far too noisy for many servers. For\n> archiving alone, this could cause tens of thousands more log messages\n> per hour on a busy system. I think you can already see such\n> information at a debug level, anyway.\n\nYou are right, having them at LOG level may blow up the log files and\ndisk storage, having them at DEBUG1 like other messages would at least\nhelp than having them as activitymsg in ps display.\n\nAnd not all of the activitymsgs are logged at DEBUGX level, at least I\ndon't see them(on a quick search).\nsnprintf(activitymsg, sizeof(activitymsg), \"waiting for %s\",\nsnprintf(activitymsg, sizeof(activitymsg), \"recovering %s\",\nsnprintf(activitymsg, sizeof(activitymsg), \"performing %s%s%s\",\nsnprintf(activitymsg, sizeof(activitymsg), \"sending backup \\\"%s\\\"\",\nsnprintf(activitymsg, sizeof(activitymsg), \"restarting at %X/%X\",\nsnprintf(activitymsg, sizeof(activitymsg), \"streaming %X/%X\",\n\nBut for the snprintf(activitymsg, sizeof(activitymsg), \"archiving %s\",\nxlog); we have elog(DEBUG1, \"archived write-ahead log file \\\"%s\\\"\",\nxlog); after the archiving command. It is also good to have a similar\ndebug message before archive command execution, we can get to know\nwhether archive command is executed or not, if yes how much time did\nit take etc.?\n\nI'm not sure whether it is okay to do the following, for every of the\nactivitymsg, before the command execution, we have an elog(DEBUG1\nmessage and after the command execution another elog(DEBUG1 message.\n\nThoughts?\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Thu, 11 Nov 2021 11:20:44 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Thu, Nov 11, 2021 at 1:40 AM Euler Taveira <euler@eulerto.com> wrote:\n>\n> On Wed, Nov 10, 2021, at 2:38 PM, Bharath Rupireddy wrote:\n>\n> As discussed in [1], isn't it a better idea to add some of activity\n> messages [2] such as recovery, archive, backup, streaming etc. to\n> server logs at LOG level? They are currently being set into ps display\n> which is good if the postgres is being run on a standalone box/VM\n> where users can see the ps display, but it doesn't help much in case\n> the postgres is being run on a cloud environment where users don't\n> have access to ps display output. Moreover, the ps display is\n> transient and will not help to analyze after an issue occurs.\n>\n> Besides recovery, the other activities already provide information through\n> views. I might be missing something but it seems the current views already\n> provide the information that ps displays.\n>\n> Having the above messages in the server logs will be useful to\n> understand how the system is/was doing/progressing with these\n> (sometimes time-intensive) operations.\n>\n> It could be useful for a root cause analysis, however, you are also increasing\n> the number of LOG messages in a system that doesn't/won't require such\n> information. It is fine to add additional DEBUG messages if there isn't a\n> similar one yet. If at least the message level were module-controlled, you\n> could modify a setting to gather more messages from a specific module.\n> Unfortunately, that's not possible so we should avoid superfluous messages.\n\nYes, at least having the messages at DEBUG1 level would help do the\nRCA after an issue occurs. Please see my response upthread at [1].\n\n[1] - https://www.postgresql.org/message-id/CALj2ACXVwnHhek9bo9wgamtW9SBinj4ohF1f1Yf3Su2o5WarWg%40mail.gmail.com\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Thu, 11 Nov 2021 11:23:09 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On 2021-Nov-10, Bossart, Nathan wrote:\n\n> On 11/10/21, 9:43 AM, \"Bharath Rupireddy\" <bharath.rupireddyforpostgres@gmail.com> wrote:\n> > As discussed in [1], isn't it a better idea to add some of activity\n> > messages [2] such as recovery, archive, backup, streaming etc. to\n> > server logs at LOG level?\n\n> I think this would make the logs far too noisy for many servers. For\n> archiving alone, this could cause tens of thousands more log messages\n> per hour on a busy system. I think you can already see such\n> information at a debug level, anyway.\n\nYeah. If we had some sort of ring buffer in which to store these\nmessages, the user could examine them through a view; they would still\nbe accessible in a running server, but they would not be written to the\nserver log.\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Thu, 11 Nov 2021 09:37:41 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On 11/10/21, 9:51 PM, \"Bharath Rupireddy\" <bharath.rupireddyforpostgres@gmail.com> wrote:\r\n> But for the snprintf(activitymsg, sizeof(activitymsg), \"archiving %s\",\r\n> xlog); we have elog(DEBUG1, \"archived write-ahead log file \\\"%s\\\"\",\r\n> xlog); after the archiving command. It is also good to have a similar\r\n> debug message before archive command execution, we can get to know\r\n> whether archive command is executed or not, if yes how much time did\r\n> it take etc.?\r\n\r\nThere is a DEBUG3 that emits \"executing archive command...\" prior to\r\nexecuting the archive command.\r\n\r\n> I'm not sure whether it is okay to do the following, for every of the\r\n> activitymsg, before the command execution, we have an elog(DEBUG1\r\n> message and after the command execution another elog(DEBUG1 message.\r\n\r\nThe general idea seems reasonable to me.\r\n\r\nNathan\r\n\r\n","msg_date":"Thu, 11 Nov 2021 15:11:49 +0000","msg_from":"\"Bossart, Nathan\" <bossartn@amazon.com>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive,\n streaming etc. activity messages to server logs along with ps display"},{"msg_contents":"On 11/11/21, 4:38 AM, \"Alvaro Herrera\" <alvherre@alvh.no-ip.org> wrote:\r\n> Yeah. If we had some sort of ring buffer in which to store these\r\n> messages, the user could examine them through a view; they would still\r\n> be accessible in a running server, but they would not be written to the\r\n> server log.\r\n\r\nThat's an interesting idea. It'd be nice to have a way to see more\r\ndebugging information without exploding the server log.\r\n\r\nNathan\r\n\r\n","msg_date":"Thu, 11 Nov 2021 15:18:16 +0000","msg_from":"\"Bossart, Nathan\" <bossartn@amazon.com>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive,\n streaming etc. activity messages to server logs along with ps display"},{"msg_contents":"On Thu, Nov 11, 2021, at 2:50 AM, Bharath Rupireddy wrote:\n> But for the snprintf(activitymsg, sizeof(activitymsg), \"archiving %s\",\n> xlog); we have elog(DEBUG1, \"archived write-ahead log file \\\"%s\\\"\",\n> xlog); after the archiving command. It is also good to have a similar\n> debug message before archive command execution, we can get to know\n> whether archive command is executed or not, if yes how much time did\n> it take etc.?\nIf you are only interested in how much time it took to archive a WAL file, you\ndon't need 2 messages (before and after); one message is sufficient. If you\ncalculate the difference and then print it as part of the existent message.\nCalculate the difference iif the log level is at least DEBUG1 (same as the log\nmessage).\n\nMessages related to system() e.g. archive and recovery, you could measure the\nexecution time in your own script/program.\n\nI like Alvaro's idea of implementing a ring buffer for this kind of activity.\nThis implementation could be applied to checkpoints, archiving, recovery,\nreplication, base backup, WAL usage, connections and others.\n\n\n--\nEuler Taveira\nEDB https://www.enterprisedb.com/\n\nOn Thu, Nov 11, 2021, at 2:50 AM, Bharath Rupireddy wrote:But for the snprintf(activitymsg, sizeof(activitymsg), \"archiving %s\",xlog); we have elog(DEBUG1, \"archived write-ahead log file \\\"%s\\\"\",xlog); after the archiving command. It is also good to have a similardebug message before archive command execution, we can get to knowwhether archive command is executed or not, if yes how much time didit take etc.?If you are only interested in how much time it took to archive a WAL file, youdon't need 2 messages (before and after); one message is sufficient. If youcalculate the difference and then print it as part of the existent message.Calculate the difference iif the log level is at least DEBUG1 (same as the logmessage).Messages related to system() e.g. archive and recovery, you could measure theexecution time in your own script/program.I like Alvaro's idea of implementing a ring buffer for this kind of activity.This implementation could be applied to checkpoints, archiving, recovery,replication, base backup, WAL usage, connections and others.--Euler TaveiraEDB https://www.enterprisedb.com/","msg_date":"Thu, 11 Nov 2021 12:46:47 -0300","msg_from":"\"Euler Taveira\" <euler@eulerto.com>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive,\n streaming etc. activity messages to server\n logs along with ps display"},{"msg_contents":"On Thu, Nov 11, 2021 at 03:18:16PM +0000, Bossart, Nathan wrote:\n> On 11/11/21, 4:38 AM, \"Alvaro Herrera\" <alvherre@alvh.no-ip.org> wrote:\n> > Yeah. If we had some sort of ring buffer in which to store these\n> > messages, the user could examine them through a view; they would still\n> > be accessible in a running server, but they would not be written to the\n> > server log.\n> \n> That's an interesting idea. It'd be nice to have a way to see more\n> debugging information without exploding the server log.\n\nUh, that's what dmesg does, which I mentioned earlier.\n\n-- \n Bruce Momjian <bruce@momjian.us> https://momjian.us\n EDB https://enterprisedb.com\n\n If only the physical world exists, free will is an illusion.\n\n\n\n","msg_date":"Thu, 11 Nov 2021 12:46:38 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Thu, Nov 11, 2021 at 6:07 PM Alvaro Herrera <alvherre@alvh.no-ip.org>\nwrote:\n\n> On 2021-Nov-10, Bossart, Nathan wrote:\n>\n> > On 11/10/21, 9:43 AM, \"Bharath Rupireddy\" <\n> bharath.rupireddyforpostgres@gmail.com> wrote:\n> > > As discussed in [1], isn't it a better idea to add some of activity\n> > > messages [2] such as recovery, archive, backup, streaming etc. to\n> > > server logs at LOG level?\n>\n> > I think this would make the logs far too noisy for many servers. For\n> > archiving alone, this could cause tens of thousands more log messages\n> > per hour on a busy system. I think you can already see such\n> > information at a debug level, anyway.\n>\n> Yeah. If we had some sort of ring buffer in which to store these\n> messages, the user could examine them through a view; they would still\n> be accessible in a running server, but they would not be written to the\n> server log.\n>\n\nThat's a good idea. How about also adding some GUC(s) to the log archive,\nrecovery related log messages just like we have for checkpoints, autovacuum\netc? Maybe something like log_archive, log_recovery etc.\n\n--\nWith Regards,\nAshutosh Sharma.\n\nOn Thu, Nov 11, 2021 at 6:07 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:On 2021-Nov-10, Bossart, Nathan wrote:\n\n> On 11/10/21, 9:43 AM, \"Bharath Rupireddy\" <bharath.rupireddyforpostgres@gmail.com> wrote:\n> > As discussed in [1], isn't it a better idea to add some of activity\n> > messages [2] such as recovery, archive, backup, streaming etc. to\n> > server logs at LOG level?\n\n> I think this would make the logs far too noisy for many servers. For\n> archiving alone, this could cause tens of thousands more log messages\n> per hour on a busy system. I think you can already see such\n> information at a debug level, anyway.\n\nYeah. If we had some sort of ring buffer in which to store these\nmessages, the user could examine them through a view; they would still\nbe accessible in a running server, but they would not be written to the\nserver log.That's a good idea. How about also adding some GUC(s) to the log archive, recovery related log messages just like we have for checkpoints, autovacuum etc? Maybe something like log_archive, log_recovery etc.--With Regards,Ashutosh Sharma.","msg_date":"Fri, 12 Nov 2021 21:10:48 +0530","msg_from":"Ashutosh Sharma <ashu.coek88@gmail.com>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Thu, Nov 11, 2021 at 8:41 PM Bossart, Nathan <bossartn@amazon.com> wrote:\n> > I'm not sure whether it is okay to do the following, for every of the\n> > activitymsg, before the command execution, we have an elog(DEBUG1\n> > message and after the command execution another elog(DEBUG1 message.\n>\n> The general idea seems reasonable to me.\n\nHere's a simple patch that I could come up with. Please have a look.\n\nRegards,\nBharath Rupireddy.","msg_date":"Sat, 13 Nov 2021 19:29:36 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"A bunch of these now execute snprintf()s unnecessarily. I think these\nshould be made conditional on message_level_is_interesting(DEBUG1) to\navoid that overhead.\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\"You're _really_ hosed if the person doing the hiring doesn't understand\nrelational systems: you end up with a whole raft of programmers, none of\nwhom has had a Date with the clue stick.\" (Andrew Sullivan)\n\n\n","msg_date":"Sat, 13 Nov 2021 12:34:37 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Sat, Nov 13, 2021 at 9:04 PM Alvaro Herrera <alvherre@alvh.no-ip.org>\nwrote:\n>\n> A bunch of these now execute snprintf()s unnecessarily. I think these\n> should be made conditional on message_level_is_interesting(DEBUG1) to\n> avoid that overhead.\n\nThanks. Attaching the v2 to avoid that by directly using the message in\nereport instead of activitymsg.\n\nRegards,\nBharath Rupireddy.","msg_date":"Sun, 14 Nov 2021 17:47:41 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On 11/14/21, 4:18 AM, \"Bharath Rupireddy\" <bharath.rupireddyforpostgres@gmail.com> wrote:\r\n> Thanks. Attaching the v2 to avoid that by directly using the message\r\n> in ereport instead of activitymsg.\r\n\r\n+ ereport(DEBUG1,\r\n+ (errmsg_internal(\"waiting for %s\", xlogfname)));\r\n\r\nAs a general comment, I think we should include more detail than what\r\ngoes in the ps title. While we probably want to keep the ps title,\r\nbrief, I don't see any issue with including a bit more context in\r\ndebug statements.\r\n\r\n+ ereport(DEBUG1,\r\n+ (errmsg_internal(\"executing archive command \\\"%s\\\" for archving %s\",\r\n+ xlogarchcmd, xlog)));\r\n\r\nI wonder if it's necessary to include the \"for archiving\" part in this\r\none. Typically, the xlogarchcmd will include it somewhere.\r\n\r\nNathan\r\n\r\n","msg_date":"Mon, 15 Nov 2021 22:36:09 +0000","msg_from":"\"Bossart, Nathan\" <bossartn@amazon.com>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive,\n streaming etc. activity messages to server logs along with ps display"},{"msg_contents":"On Mon, Nov 15, 2021 at 10:36:09PM +0000, Bossart, Nathan wrote:\n> + ereport(DEBUG1,\n> + (errmsg_internal(\"executing archive command \\\"%s\\\" for archving %s\",\n> + xlogarchcmd, xlog)));\n> \n> I wonder if it's necessary to include the \"for archiving\" part in this\n> one. Typically, the xlogarchcmd will include it somewhere.\n\nDon't see any need to add that either (s/archving/archiving/, btw).\n\n@@ -1000,6 +1004,10 @@ XLogWalRcvFlush(bool dying, TimeLineID tli)\n+ ereport(DEBUG1,\n+ (errmsg_internal(\"streaming %X/%X\",\n+ LSN_FORMAT_ARGS(LogstreamResult.Write))));\nHaving the same wording for the WAL receiver and the WAL sender could\nbe confusing when it comes to a cascading standby node, particularly\nif log_line_prefix does not include the process PID.\n\n+ ereport(DEBUG1,\n+ (errmsg_internal(\"streaming %X/%X\",\n+ LSN_FORMAT_ARGS(sentPtr))));\nAnyway, those two ones are going to make the logs much more noisy, no?\nThe same could be said about XLogFileRead(), joining the point of\nNathan's upthread. So I cannot get excited by this change.\n--\nMichael","msg_date":"Tue, 16 Nov 2021 12:13:51 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On 11/15/21, 7:14 PM, \"Michael Paquier\" <michael@paquier.xyz> wrote:\r\n> + ereport(DEBUG1,\r\n> + (errmsg_internal(\"streaming %X/%X\",\r\n> + LSN_FORMAT_ARGS(sentPtr))));\r\n> Anyway, those two ones are going to make the logs much more noisy, no?\r\n> The same could be said about XLogFileRead(), joining the point of\r\n> Nathan's upthread. So I cannot get excited by this change.\r\n\r\nYeah, this might even be too noisy for DEBUG5.\r\n\r\nNathan\r\n\r\n","msg_date":"Tue, 16 Nov 2021 16:01:54 +0000","msg_from":"\"Bossart, Nathan\" <bossartn@amazon.com>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive,\n streaming etc. activity messages to server logs along with ps display"},{"msg_contents":"On 2021-Nov-16, Bossart, Nathan wrote:\n\n> On 11/15/21, 7:14 PM, \"Michael Paquier\" <michael@paquier.xyz> wrote:\n> > + ereport(DEBUG1,\n> > + (errmsg_internal(\"streaming %X/%X\",\n> > + LSN_FORMAT_ARGS(sentPtr))));\n> > Anyway, those two ones are going to make the logs much more noisy, no?\n> > The same could be said about XLogFileRead(), joining the point of\n> > Nathan's upthread. So I cannot get excited by this change.\n> \n> Yeah, this might even be too noisy for DEBUG5.\n\nNod. And if you're at DEBUG5, the log contains so much other crap that\nit is unusable for any purposes anyway.\n\nMy opinion is that adding these things willy-nilly is not a solution to\nany actual problem. Adding a few additional log lines that are\nlow-volume at DEBUG1 might be useful, but below that (DEBUG2 etc) it's\nnot good for anything other than specific development, IMO. At least\nthis particular one for streaming replication I think we should not\ninclude.\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Tue, 16 Nov 2021 13:26:49 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Tue, Nov 16, 2021 at 01:26:49PM -0300, Alvaro Herrera wrote:\n> My opinion is that adding these things willy-nilly is not a solution to\n> any actual problem. Adding a few additional log lines that are\n> low-volume at DEBUG1 might be useful, but below that (DEBUG2 etc) it's\n> not good for anything other than specific development, IMO. At least\n> this particular one for streaming replication I think we should not\n> include.\n\nLooking at v2, I think that this leaves the additions of the DEBUG1\nentries in SendBaseBackup() and WalRcvWaitForStartPosition(), then.\nThe one in pgarch.c does not provide any additional information as the\nsegment to-be-archived should be part of the command.\n--\nMichael","msg_date":"Wed, 17 Nov 2021 11:31:16 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Wed, Nov 17, 2021 at 8:01 AM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Tue, Nov 16, 2021 at 01:26:49PM -0300, Alvaro Herrera wrote:\n> > My opinion is that adding these things willy-nilly is not a solution to\n> > any actual problem. Adding a few additional log lines that are\n> > low-volume at DEBUG1 might be useful, but below that (DEBUG2 etc) it's\n> > not good for anything other than specific development, IMO. At least\n> > this particular one for streaming replication I think we should not\n> > include.\n>\n> Looking at v2, I think that this leaves the additions of the DEBUG1\n> entries in SendBaseBackup() and WalRcvWaitForStartPosition(), then.\n> The one in pgarch.c does not provide any additional information as the\n> segment to-be-archived should be part of the command.\n\nThank you all for the inputs. Here's the patch that I've come up with.\n\nUpon thinking further, having at least the messages at LOG level [1]\nwould be helpful to know what's happening with the system while in\nrecovery. Although these messages at LOG level seem to be filling up\nthe server logs, having a good log consumption and rotation mechanism\n(I'm sure every major postgres vendor would have one) would be\nsufficient to allay that concern.\n\nThese LOG messages would help us know how much time a restore command\ntakes to fetch the WAL file and what is the current WAL file the\nserver is recovering and where is it recovering from. The customer\noften asks questions like: when will my server come up? how much time\ndoes the recovery of my server take?\n\nAs a developer or admin, one can monitor these logs and do bunch of things:\n1) see how many WAL files left to be recovered by looking at the WAL\nfiles in the archive location or pg_wal directory or from primary\n2) provide an approximate estimation of when the server will come up\nor how much more the recovery takes by looking at these previous LOG\nmessages, one can know the number of WAL files that server recovered\nover a minute and with the help of left-over WAL files calculated from\n(1).\n\n[1]\nereport(LOG,\n (errmsg(\"waiting for WAL segment \\\"%s\\\" from archive\",\n xlogfname)));\n\nereport(LOG,\n (errmsg(\"restored WAL segment \\\"%s\\\" from archive\",\n xlogfname)));\n\nereport(LOG,\n (errmsg(\"recovering WAL segment \\\"%s\\\" from source \\\"%s\\\"\",\n xlogfname, srcname)));\n\nRegards,\nBharath Rupireddy.","msg_date":"Tue, 7 Dec 2021 18:28:24 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Tue, Dec 07, 2021 at 06:28:24PM +0530, Bharath Rupireddy wrote:\n> Upon thinking further, having at least the messages at LOG level [1]\n> would be helpful to know what's happening with the system while in\n> recovery. Although these messages at LOG level seem to be filling up\n> the server logs, having a good log consumption and rotation mechanism\n> (I'm sure every major postgres vendor would have one) would be\n> sufficient to allay that concern.\n\n+ ereport(LOG,\n+ (errmsg(\"recovering WAL segment \\\"%s\\\" from source \\\"%s\\\"\",\n+ xlogfname, srcname)));\nIsn't this approach an issue for translations? It seems to me that\nterms like \"stream\" or \"archive\" had better be translated properly,\nmeaning that this needs a fully-constructed sentence.\n--\nMichael","msg_date":"Wed, 8 Dec 2021 13:35:48 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Wed, Dec 8, 2021 at 10:05 AM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Tue, Dec 07, 2021 at 06:28:24PM +0530, Bharath Rupireddy wrote:\n> > Upon thinking further, having at least the messages at LOG level [1]\n> > would be helpful to know what's happening with the system while in\n> > recovery. Although these messages at LOG level seem to be filling up\n> > the server logs, having a good log consumption and rotation mechanism\n> > (I'm sure every major postgres vendor would have one) would be\n> > sufficient to allay that concern.\n>\n> + ereport(LOG,\n> + (errmsg(\"recovering WAL segment \\\"%s\\\" from source \\\"%s\\\"\",\n> + xlogfname, srcname)));\n> Isn't this approach an issue for translations? It seems to me that\n> terms like \"stream\" or \"archive\" had better be translated properly,\n> meaning that this needs a fully-constructed sentence.\n\nThanks for taking a look at the patch. How about the attached v4?\n\nI added a CF entry - https://commitfest.postgresql.org/36/3443/\n\nRegards,\nBharath Rupireddy.","msg_date":"Wed, 8 Dec 2021 10:47:57 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Wed, Dec 08, 2021 at 10:47:57AM +0530, Bharath Rupireddy wrote:\n> Thanks for taking a look at the patch. How about the attached v4?\n> \n> I added a CF entry - https://commitfest.postgresql.org/36/3443/\n\n+ else if (source == XLOG_FROM_STREAM)\n+ ereport(LOG,\n+ (errmsg(\"recovering WAL segment \\\"%s\\\" received from primary\",\n+ xlogfname)));\nThis is incorrect when using a cascading standby. And perhaps this\ncould use a switch/case?\n\nWhile quickly testing, I got reminded that the LOG for a segment\nretrieved from the local pg_wal would generate some noise when running\nfor example the bootstrap process. Is this one really interesting\ncompared to the two others?\n--\nMichael","msg_date":"Wed, 8 Dec 2021 16:21:55 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Wed, Dec 8, 2021 at 12:52 PM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Wed, Dec 08, 2021 at 10:47:57AM +0530, Bharath Rupireddy wrote:\n> > Thanks for taking a look at the patch. How about the attached v4?\n> >\n> > I added a CF entry - https://commitfest.postgresql.org/36/3443/\n>\n> + else if (source == XLOG_FROM_STREAM)\n> + ereport(LOG,\n> + (errmsg(\"recovering WAL segment \\\"%s\\\" received from primary\",\n> + xlogfname)));\n> This is incorrect when using a cascading standby.\n\nI visualize the setup like this: primary -> standby -> cascading standby\n\nFor standby, \"received from primary\" makes sense. For cascading\nstandby too it makes sense because the standby still acts as primary\nfor cascading standby, no? And we don't distinguish any other existing\nmessages for a standby acting as a cascading standby right? The\ncomments around the enum XLOG_FROM_STREAM and usage of it still says\nfrom \"primary\".\n\nDo you want me to add \"received from stream\" in general?\n\n> And perhaps this could use a switch/case?\n\nYes, I will change it to switch-case in the next version.\n\n> While quickly testing, I got reminded that the LOG for a segment\n> retrieved from the local pg_wal would generate some noise when running\n> for example the bootstrap process. Is this one really interesting\n> compared to the two others?\n\nYes. I observed that for a simple primary-standby setup, inserting\n10million rows generated around 40 WAL files and so 40 LOG messages in\nthe standby server logs, in a typical production environment with\nheavy-write workload, the amount of LOG generation is going to be much\nmore. And we don't want to have them as DEBUG1 messages and enable\nDEBUG1 just for recovery messages that one is interested in.\n\nHow about we invent a new GUC log_recovery, with default set to false,\nsimilar to log_checkpoints? And, these messages can be like:\nereport(log_recovery ? LOG : DEBUG1, .....)? This GUC can be useful to\nadd some recovery stats in future as well. Thoughts?\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Wed, 8 Dec 2021 14:00:12 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Wed, Dec 08, 2021 at 02:00:12PM +0530, Bharath Rupireddy wrote:\n> For standby, \"received from primary\" makes sense. For cascading\n> standby too it makes sense because the standby still acts as primary\n> for cascading standby, no? And we don't distinguish any other existing\n> messages for a standby acting as a cascading standby right? The\n> comments around the enum XLOG_FROM_STREAM and usage of it still says\n> from \"primary\".\n\nQuoting the docs:\nhttps://www.postgresql.org/docs/devel/glossary.html#GLOSSARY-PRIMARY-SERVER\nAnd a standby is no such thing.\n\n> Do you want me to add \"received from stream\" in general?\n\n\"from stream\" would be fine, IMHO.\n\n> How about we invent a new GUC log_recovery, with default set to false,\n> similar to log_checkpoints? And, these messages can be like:\n> ereport(log_recovery ? LOG : DEBUG1, .....)? This GUC can be useful to\n> add some recovery stats in future as well. Thoughts?\n\nThis discussion does not justify a new GUC at this stage IMO.\n--\nMichael","msg_date":"Wed, 8 Dec 2021 17:38:14 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Wed, Dec 8, 2021 at 2:08 PM Michael Paquier <michael@paquier.xyz> wrote:\n> > Do you want me to add \"received from stream\" in general?\n>\n> \"from stream\" would be fine, IMHO.\n\nDone in the attached v5 patch.\n\n> > How about we invent a new GUC log_recovery, with default set to false,\n> > similar to log_checkpoints? And, these messages can be like:\n> > ereport(log_recovery ? LOG : DEBUG1, .....)? This GUC can be useful to\n> > add some recovery stats in future as well. Thoughts?\n>\n> This discussion does not justify a new GUC at this stage IMO.\n\nHaving \"log_recovery\" GUC gives the flexibility to the users to enable\nthe recovery logs added by this patch if they are concerned otherwise\nthe logs just be at DEBUG1 level. Setting log_min_messages level to\nDEBUG1 would definitely flood the server logs.\n\n> While quickly testing, I got reminded that the LOG for a segment\n> retrieved from the local pg_wal would generate some noise when running\n> for example the bootstrap process. Is this one really interesting\n> compared to the two others?\n\nIf we just have the archive related message at LOG level and others at\nDEBUG1 level, it is a kind of partial recovery log information that\none has in the log files, missing recovery messages from the pg_wal\ndirectory in the logs.\n\nAnother option is to just have all the messages (added in the patch)\nat DEBUG1 level, let's users deal with it.\n\nIMO, the new GUC \"log_recovery\" is way better as it can be extended to\ninclude some recovery stats info as well.\n\nThoughts?\n\nRegards,\nBharath Rupireddy.","msg_date":"Wed, 8 Dec 2021 16:57:49 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Wed, Dec 08, 2021 at 04:57:49PM +0530, Bharath Rupireddy wrote:\n> Done in the attached v5 patch.\n\nThanks for the new version. FWIW, as the information logged when\nrecovering a WAL segment from the local pg_wal could be rather\nconfusing at bootstrap phase, I would suggest to simplify things as of\nthe attached and call it a day.\n--\nMichael","msg_date":"Thu, 9 Dec 2021 11:26:38 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Thu, Dec 9, 2021 at 7:56 AM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Wed, Dec 08, 2021 at 04:57:49PM +0530, Bharath Rupireddy wrote:\n> > Done in the attached v5 patch.\n>\n> Thanks for the new version. FWIW, as the information logged when\n> recovering a WAL segment from the local pg_wal could be rather\n> confusing at bootstrap phase, I would suggest to simplify things as of\n> the attached and call it a day.\n\nThanks for the patch. It looks good to me. I have a suggestion: It's\nbetter if we specify why we are not emitting any message if the WAL\nsegment is found in pg_wal directory, something like \"we are\nintentionally not emitting any messages in case if the required WAL\nsegment is found in the pg_wal directory to avoid noise in the server\nlogs\".\n\n+ case XLOG_FROM_PG_WAL:\n+ case XLOG_FROM_ANY:\n+ break;\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Thu, 9 Dec 2021 11:28:41 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Thu, Dec 9, 2021 at 11:28 AM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n>\n> On Thu, Dec 9, 2021 at 7:56 AM Michael Paquier <michael@paquier.xyz> wrote:\n> >\n> > On Wed, Dec 08, 2021 at 04:57:49PM +0530, Bharath Rupireddy wrote:\n> > > Done in the attached v5 patch.\n> >\n> > Thanks for the new version. FWIW, as the information logged when\n> > recovering a WAL segment from the local pg_wal could be rather\n> > confusing at bootstrap phase, I would suggest to simplify things as of\n> > the attached and call it a day.\n>\n> Thanks for the patch. It looks good to me. I have a suggestion: It's\n> better if we specify why we are not emitting any message if the WAL\n> segment is found in pg_wal directory, something like \"we are\n> intentionally not emitting any messages in case if the required WAL\n> segment is found in the pg_wal directory to avoid noise in the server\n> logs\".\n>\n> + case XLOG_FROM_PG_WAL:\n> + case XLOG_FROM_ANY:\n> + break;\n\nI just want to call this out: an insertion of 10 million rows in the\nprimary generates a bunch of messages in the standby [1] within 40 sec\n(size of the standby server log grows by 5K).\n\n[1]\n2021-12-08 08:11:53.906 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000004\" from stream\n2021-12-08 08:11:55.864 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000005\" from stream\n2021-12-08 08:11:56.713 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000006\" from stream\n2021-12-08 08:11:57.495 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000007\" from stream\n2021-12-08 08:11:58.301 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000008\" from stream\n2021-12-08 08:11:59.062 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000009\" from stream\n2021-12-08 08:11:59.853 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000000A\" from stream\n2021-12-08 08:12:00.573 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000000B\" from stream\n2021-12-08 08:12:01.346 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000000C\" from stream\n2021-12-08 08:12:02.048 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000000D\" from stream\n2021-12-08 08:12:02.798 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000000E\" from stream\n2021-12-08 08:12:03.544 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000000F\" from stream\n2021-12-08 08:12:04.310 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000010\" from stream\n2021-12-08 08:12:05.096 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000011\" from stream\n2021-12-08 08:12:06.003 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000012\" from stream\n2021-12-08 08:12:06.762 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000013\" from stream\n2021-12-08 08:12:07.595 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000014\" from stream\n2021-12-08 08:12:08.431 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000015\" from stream\n2021-12-08 08:12:09.246 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000016\" from stream\n2021-12-08 08:12:10.144 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000017\" from stream\n2021-12-08 08:12:14.944 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000018\" from stream\n2021-12-08 08:12:15.784 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000019\" from stream\n2021-12-08 08:12:16.757 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000001A\" from stream\n2021-12-08 08:12:17.571 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000001B\" from stream\n2021-12-08 08:12:18.375 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000001C\" from stream\n2021-12-08 08:12:19.179 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000001D\" from stream\n2021-12-08 08:12:19.964 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000001E\" from stream\n2021-12-08 08:12:20.839 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000001F\" from stream\n2021-12-08 08:12:21.782 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000020\" from stream\n2021-12-08 08:12:24.422 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000021\" from stream\n2021-12-08 08:12:25.164 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000022\" from stream\n2021-12-08 08:12:25.989 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000023\" from stream\n2021-12-08 08:12:26.874 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000024\" from stream\n2021-12-08 08:12:27.592 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000025\" from stream\n2021-12-08 08:12:28.471 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000026\" from stream\n2021-12-08 08:12:29.158 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000027\" from stream\n2021-12-08 08:12:30.017 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000028\" from stream\n2021-12-08 08:12:30.778 UTC [1219898] LOG: recovering WAL segment\n\"000000010000000000000029\" from stream\n2021-12-08 08:12:31.859 UTC [1219898] LOG: recovering WAL segment\n\"00000001000000000000002A\" from stream\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Thu, 9 Dec 2021 11:42:30 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On 2021-Dec-09, Bharath Rupireddy wrote:\n\n> I just want to call this out: an insertion of 10 million rows in the\n> primary generates a bunch of messages in the standby [1] within 40 sec\n> (size of the standby server log grows by 5K).\n\nHmm, that does sound excessive to me in terms of log bloat increase.\nRemember the discussion about turning log_checkpoints on by default?\n\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\"Las mujeres son como hondas: mientras más resistencia tienen,\n más lejos puedes llegar con ellas\" (Jonas Nightingale, Leap of Faith)\n\n\n","msg_date":"Thu, 9 Dec 2021 09:30:34 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Thu, Dec 9, 2021 at 6:00 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>\n> On 2021-Dec-09, Bharath Rupireddy wrote:\n>\n> > I just want to call this out: an insertion of 10 million rows in the\n> > primary generates a bunch of messages in the standby [1] within 40 sec\n> > (size of the standby server log grows by 5K).\n>\n> Hmm, that does sound excessive to me in terms of log bloat increase.\n> Remember the discussion about turning log_checkpoints on by default?\n\nThe amount of LOG messages generated when the log_checkpoints GUC is\nset to on, are quite less, hardly 4 messages per-checkpoint (5min). I\nthink enabling log_checkpoints is still acceptable as most of the\nhackers agreed in another thread [1] that the advantages with it are\nmore and it doesn't seem to be bloating the server logs (in a day at\nmax 1152 messages).\n\nI'm not sure if it is worth having a GUC log_recovery if enabled the\nrecovery messages can be emitted at LOG level otherwise DEBUG1 level.\nlog_recovery GUC can also be used to collect and emit some recovery\nstats like log_checkpoints.\n\n[1] - https://www.postgresql.org/message-id/CA%2BTgmoaDFpFmNQuaWj6%2B78CPVBrF_WPT1wFHBTvio%3DtRmxzUcQ%40mail.gmail.com\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Thu, 9 Dec 2021 21:20:54 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On 2021-Dec-09, Bharath Rupireddy wrote:\n\n> On Thu, Dec 9, 2021 at 6:00 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> >\n> > On 2021-Dec-09, Bharath Rupireddy wrote:\n> >\n> > > I just want to call this out: an insertion of 10 million rows in the\n> > > primary generates a bunch of messages in the standby [1] within 40 sec\n> > > (size of the standby server log grows by 5K).\n> >\n> > Hmm, that does sound excessive to me in terms of log bloat increase.\n> > Remember the discussion about turning log_checkpoints on by default?\n> \n> The amount of LOG messages generated when the log_checkpoints GUC is\n> set to on, are quite less, hardly 4 messages per-checkpoint (5min). I\n> think enabling log_checkpoints is still acceptable as most of the\n> hackers agreed in another thread [1] that the advantages with it are\n> more and it doesn't seem to be bloating the server logs (in a day at\n> max 1152 messages).\n\nMy point is that it was argued, in that thread, that setting\nlog_checkpoints to on by default would cause too much additional log\nvolume. That argument was shot down, but in this thread we're arguing\nthat we want five kilobytes of messages in forty seconds. That sounds\nmuch less acceptable to me, so making it default behavior or\nunconditional behavior is not going to fly very high.\n\n> I'm not sure if it is worth having a GUC log_recovery if enabled the\n> recovery messages can be emitted at LOG level otherwise DEBUG1 level.\n\nMaybe some tunable like\nlog_wal_traffic = { none, medium, high }\nwhere \"none\" is current behavior of no noise, \"medium\" gets (say) once\nevery 256 segments (e.g., when going from XFF to (X+1)00), \"high\" gets\nyou one message per segment.\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Thu, 9 Dec 2021 12:58:21 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Thu, Dec 9, 2021 at 9:28 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>\n> On 2021-Dec-09, Bharath Rupireddy wrote:\n>\n> > On Thu, Dec 9, 2021 at 6:00 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> > >\n> > > On 2021-Dec-09, Bharath Rupireddy wrote:\n> > >\n> > > > I just want to call this out: an insertion of 10 million rows in the\n> > > > primary generates a bunch of messages in the standby [1] within 40 sec\n> > > > (size of the standby server log grows by 5K).\n> > >\n> > > Hmm, that does sound excessive to me in terms of log bloat increase.\n> > > Remember the discussion about turning log_checkpoints on by default?\n> >\n> > The amount of LOG messages generated when the log_checkpoints GUC is\n> > set to on, are quite less, hardly 4 messages per-checkpoint (5min). I\n> > think enabling log_checkpoints is still acceptable as most of the\n> > hackers agreed in another thread [1] that the advantages with it are\n> > more and it doesn't seem to be bloating the server logs (in a day at\n> > max 1152 messages).\n>\n> My point is that it was argued, in that thread, that setting\n> log_checkpoints to on by default would cause too much additional log\n> volume. That argument was shot down, but in this thread we're arguing\n> that we want five kilobytes of messages in forty seconds. That sounds\n> much less acceptable to me, so making it default behavior or\n> unconditional behavior is not going to fly very high.\n>\n> > I'm not sure if it is worth having a GUC log_recovery if enabled the\n> > recovery messages can be emitted at LOG level otherwise DEBUG1 level.\n>\n> Maybe some tunable like\n> log_wal_traffic = { none, medium, high }\n> where \"none\" is current behavior of no noise, \"medium\" gets (say) once\n> every 256 segments (e.g., when going from XFF to (X+1)00), \"high\" gets\n> you one message per segment.\n\nYeah, that can be an option.\n\nAnother idea could be to use the infrastructure laid out by the commit\n9ce346e [1]. With ereport_startup_progress, we can emit the LOGs(of\ncurrent recovering WAL file) for every log_startup_progress_interval\nseconds/milliseconds.\n\nOne problem is that ereport_startup_progress doesn't work on\nStandbyMode, maybe we can remove this restriction unless we have a\nmajor reason for not allowing it on the standby.\n /* Prepare to report progress of the redo phase. */\n if (!StandbyMode)\n begin_startup_progress_phase();\n\nThoughts?\n\n[1]\ncommit 9ce346eabf350a130bba46be3f8c50ba28506969\nAuthor: Robert Haas <rhaas@postgresql.org>\nDate: Mon Oct 25 11:51:57 2021 -0400\n\n Report progress of startup operations that take a long time.\n\n Discussion:\nhttps://postgr.es/m/CA+TgmoaHQrgDFOBwgY16XCoMtXxsrVGFB2jNCvb7-ubuEe1MGg@mail.gmail.com\n Discussion:\nhttps://postgr.es/m/CAMm1aWaHF7VE69572_OLQ+MgpT5RUiUDgF1x5RrtkJBLdpRj3Q@mail.gmail.com\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Sun, 12 Dec 2021 18:08:05 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Sun, Dec 12, 2021 at 06:08:05PM +0530, Bharath Rupireddy wrote:\n> Another idea could be to use the infrastructure laid out by the commit\n> 9ce346e [1]. With ereport_startup_progress, we can emit the LOGs(of\n> current recovering WAL file) for every log_startup_progress_interval\n> seconds/milliseconds.\n> \n> One problem is that ereport_startup_progress doesn't work on\n> StandbyMode, maybe we can remove this restriction unless we have a\n> major reason for not allowing it on the standby.\n> /* Prepare to report progress of the redo phase. */\n> if (!StandbyMode)\n> begin_startup_progress_phase();\n\nThe relevant conversation starts here.\nhttps://www.postgresql.org/message-id/20210814214700.GO10479%40telsasoft.com\n\nThere was a lot of confusion in that thread, though.\n\nThe understanding was that it didn't make sense for that feature to\ncontinuously log messages on a standby (every 10sec by default). That seems\nlike too much - the issue of a checkpointed logged every 5min was enough of a\nhurdle.\n\nIf you're talking about a new feature that uses the infrastructre from 9ce3,\nbut is controlled by a separate GUC like log_wal_traffic, that could be okay.\n\n-- \nJustin\n\n\n","msg_date":"Fri, 24 Dec 2021 07:49:46 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Fri, Dec 24, 2021 at 7:19 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n>\n> On Sun, Dec 12, 2021 at 06:08:05PM +0530, Bharath Rupireddy wrote:\n> > Another idea could be to use the infrastructure laid out by the commit\n> > 9ce346e [1]. With ereport_startup_progress, we can emit the LOGs(of\n> > current recovering WAL file) for every log_startup_progress_interval\n> > seconds/milliseconds.\n> >\n> > One problem is that ereport_startup_progress doesn't work on\n> > StandbyMode, maybe we can remove this restriction unless we have a\n> > major reason for not allowing it on the standby.\n> > /* Prepare to report progress of the redo phase. */\n> > if (!StandbyMode)\n> > begin_startup_progress_phase();\n>\n> The relevant conversation starts here.\n> https://www.postgresql.org/message-id/20210814214700.GO10479%40telsasoft.com\n>\n> There was a lot of confusion in that thread, though.\n>\n> The understanding was that it didn't make sense for that feature to\n> continuously log messages on a standby (every 10sec by default). That seems\n> like too much - the issue of a checkpointed logged every 5min was enough of a\n> hurdle.\n>\n> If you're talking about a new feature that uses the infrastructre from 9ce3,\n> but is controlled by a separate GUC like log_wal_traffic, that could be okay.\n\nDo you see any problems using the same GUC\nlog_startup_progress_interval and ereport_startup_progress API\nintroduced by 9ce346e? I prefer this instead of a new GUC. Thoughts?\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Fri, 24 Dec 2021 20:04:08 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Fri, Dec 24, 2021 at 08:04:08PM +0530, Bharath Rupireddy wrote:\n> On Fri, Dec 24, 2021 at 7:19 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > On Sun, Dec 12, 2021 at 06:08:05PM +0530, Bharath Rupireddy wrote:\n> > > Another idea could be to use the infrastructure laid out by the commit\n> > > 9ce346e [1]. With ereport_startup_progress, we can emit the LOGs(of\n> > > current recovering WAL file) for every log_startup_progress_interval\n> > > seconds/milliseconds.\n> > >\n> > > One problem is that ereport_startup_progress doesn't work on\n> > > StandbyMode, maybe we can remove this restriction unless we have a\n> > > major reason for not allowing it on the standby.\n> > > /* Prepare to report progress of the redo phase. */\n> > > if (!StandbyMode)\n> > > begin_startup_progress_phase();\n> >\n> > The relevant conversation starts here.\n> > https://www.postgresql.org/message-id/20210814214700.GO10479%40telsasoft.com\n> >\n> > There was a lot of confusion in that thread, though.\n> >\n> > The understanding was that it didn't make sense for that feature to\n> > continuously log messages on a standby (every 10sec by default). That seems\n> > like too much - the issue of a checkpointed logged every 5min was enough of a\n> > hurdle.\n> >\n> > If you're talking about a new feature that uses the infrastructre from 9ce3,\n> > but is controlled by a separate GUC like log_wal_traffic, that could be okay.\n> \n> Do you see any problems using the same GUC\n> log_startup_progress_interval and ereport_startup_progress API\n> introduced by 9ce346e? I prefer this instead of a new GUC. Thoughts?\n\nI referenced the thread where we avoided writing progress logs during normal\noperation of a standby, since the logs would be both useless and excessive.\n\nIf your patch were to use the same infrastructure, you'd still want to avoid\ndisturbing that behavior and writing useless logs.\n\n-- \nJustin\n\n\n","msg_date":"Sat, 25 Dec 2021 18:20:54 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Thu, Dec 9, 2021 at 9:28 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>\n> Maybe some tunable like\n> log_wal_traffic = { none, medium, high }\n> where \"none\" is current behavior of no noise, \"medium\" gets (say) once\n> every 256 segments (e.g., when going from XFF to (X+1)00), \"high\" gets\n> you one message per segment.\n\nOn Fri, Dec 24, 2021 at 7:19 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n>\n> If you're talking about a new feature that uses the infrastructre from 9ce3,\n> but is controlled by a separate GUC like log_wal_traffic, that could be okay.\n\nThanks for the suggestion. I've added a new GUC log_wal_traffic as\nsuggested above. PSA v7 patch.\n\nRegards,\nBharath Rupireddy.","msg_date":"Wed, 29 Dec 2021 06:50:53 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: add recovery, backup, archive, streaming etc. activity messages\n to server logs along with ps display"},{"msg_contents":"On Wed, Dec 29, 2021 at 6:50 AM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n>\n> On Thu, Dec 9, 2021 at 9:28 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> >\n> > Maybe some tunable like\n> > log_wal_traffic = { none, medium, high }\n> > where \"none\" is current behavior of no noise, \"medium\" gets (say) once\n> > every 256 segments (e.g., when going from XFF to (X+1)00), \"high\" gets\n> > you one message per segment.\n>\n> On Fri, Dec 24, 2021 at 7:19 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> >\n> > If you're talking about a new feature that uses the infrastructre from 9ce3,\n> > but is controlled by a separate GUC like log_wal_traffic, that could be okay.\n>\n> Thanks for the suggestion. I've added a new GUC log_wal_traffic as\n> suggested above. PSA v7 patch.\n\nHere's the rebased v8 patch, please review it.\n\nhttps://commitfest.postgresql.org/37/3443/\n\nRegards,\nBharath Rupireddy.","msg_date":"Sun, 27 Feb 2022 12:50:25 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Add WAL recovery messages with log_wal_traffic GUC (was: add\n recovery, backup, archive, streaming etc. activity messages to server logs\n along with ps display)"},{"msg_contents":"On Sun, Feb 27, 2022 at 12:50 PM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n>\n> On Wed, Dec 29, 2021 at 6:50 AM Bharath Rupireddy\n> <bharath.rupireddyforpostgres@gmail.com> wrote:\n> >\n> > On Thu, Dec 9, 2021 at 9:28 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> > >\n> > > Maybe some tunable like\n> > > log_wal_traffic = { none, medium, high }\n> > > where \"none\" is current behavior of no noise, \"medium\" gets (say) once\n> > > every 256 segments (e.g., when going from XFF to (X+1)00), \"high\" gets\n> > > you one message per segment.\n> >\n> > On Fri, Dec 24, 2021 at 7:19 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > >\n> > > If you're talking about a new feature that uses the infrastructre from 9ce3,\n> > > but is controlled by a separate GUC like log_wal_traffic, that could be okay.\n> >\n> > Thanks for the suggestion. I've added a new GUC log_wal_traffic as\n> > suggested above. PSA v7 patch.\n>\n> Here's the rebased v8 patch, please review it.\n>\n> https://commitfest.postgresql.org/37/3443/\n\nHere's the rebased v9 patch.\n\nRegards,\nBharath Rupireddy.","msg_date":"Fri, 29 Apr 2022 14:40:56 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\n recovery, backup, archive, streaming etc. activity messages to server logs\n along with ps display)"},{"msg_contents":"Did we ever consider the idea of using a new pg_stat_wal_activity_progress\nview or something like that, using the backend_progress.c functionality?\nI don't see it mentioned in the thread.\n\nIt's not an *exact* fit, since this is not about one \"command\" being\nexecuted by a \"backend\", but it seems a near enough fit, and it would\nunobtrusive enough that we don't need to worry about the progress-update\nrate or have a GUC to determine how frequently to update (with a gauge\nthat's heavily workload-dependant and is likely to make little sense to\nsome users -- consider differently-sized WAL segments, for one thing).\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Fri, 29 Apr 2022 12:41:15 +0200","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\n recovery, backup, archive, streaming etc. activity messages to server logs\n along with ps display)"},{"msg_contents":"On Fri, Apr 29, 2022 at 4:11 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>\n> Did we ever consider the idea of using a new pg_stat_wal_activity_progress\n> view or something like that, using the backend_progress.c functionality?\n> I don't see it mentioned in the thread.\n\nIMO, progress reporting works well on a running server and at the\nmoment. The WAL recovery/replay can happen even before the server\nopens up for connections and the progress report view can't be used\nfor later analysis like how much time the restoring WAL files from\narchive location took and also the WAL file names can't be reported in\nprogress reporting mechanism (only integers columns, of course if\nrequired we can add text columns to pg_stat_get_progress_info). Having\nthe recovery info in server logs might help.\n\n> It's not an *exact* fit, since this is not about one \"command\" being\n> executed by a \"backend\", but it seems a near enough fit, and it would\n> unobtrusive enough that we don't need to worry about the progress-update\n> rate or have a GUC to determine how frequently to update (with a gauge\n> that's heavily workload-dependant and is likely to make little sense to\n> some users -- consider differently-sized WAL segments, for one thing).\n\nI think reporting a long-running file processing operation (removing\nor syncing) within postgres is a generic problem (for snapshot,\nmapping, temporary (pgsql_tmp), temp relation files, old WAL file\nprocessing, WAL file processing during recovery etc.) and needs to be\nsolved in two ways: 1) logging progress into server logs (which helps\nfor analysis and report when the server isn't available for\nconnections, crash recovery), a generic GUC\nlog_file_processing_traffic = {none, medium, high} might help here\n(also proposed in [1]) and 2) pg_stat_file_processing_progress\n(extending progress reporting pg_stat_get_progress_info to have few\ntext columns for current file name and directory path).\n\nThoughts?\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Thu, 5 May 2022 13:35:12 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\n recovery, backup, archive, streaming etc. activity messages to server logs\n along with ps display)"},{"msg_contents":"On 2022-May-05, Bharath Rupireddy wrote:\n\n> On Fri, Apr 29, 2022 at 4:11 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> >\n> > Did we ever consider the idea of using a new pg_stat_wal_activity_progress\n> > view or something like that, using the backend_progress.c functionality?\n> > I don't see it mentioned in the thread.\n> \n> IMO, progress reporting works well on a running server and at the\n> moment. The WAL recovery/replay can happen even before the server\n> opens up for connections\n\nIt's definitely true that you wouldn't be able to use it for when the\nserver is not accepting connections.\n\n> and the progress report view can't be used\n> for later analysis like how much time the restoring WAL files from\n> archive location took\n\nThis is true too -- progress report doesn't store historical data, only\ncurrent status.\n\n> and also the WAL file names can't be reported in progress reporting\n> mechanism\n\nAlso true.\n\n> (only integers columns, of course if required we can add text columns\n> to pg_stat_get_progress_info).\n\nYeah, I don't think adding text columns is terribly easy, because the\nwhole point of the progress reporting infrastructure is that it can be\nupdated very cheaply as atomic operations, and if you want to transmit\ntext columns, that's no longer possible.\n\n> Having the recovery info in server logs might help.\n\nI suppose it might.\n\n> I think reporting a long-running file processing operation (removing\n> or syncing) within postgres is a generic problem (for snapshot,\n> mapping, temporary (pgsql_tmp), temp relation files, old WAL file\n> processing, WAL file processing during recovery etc.) and needs to be\n> solved\n\nI agree up to here.\n\n> in two ways: 1) logging progress into server logs (which helps\n> for analysis and report when the server isn't available for\n> connections, crash recovery), a generic GUC\n> log_file_processing_traffic = {none, medium, high} might help here\n> (also proposed in [1]) and 2) pg_stat_file_processing_progress\n> (extending progress reporting pg_stat_get_progress_info to have few\n> text columns for current file name and directory path).\n\nI think using the server log to store telemetry data is not a great fit.\nIt bloats the log files and can be so slow as to block other operations\nin the server. Server logs should normally be considered critical info\nthat's not okay to lose; telemetry tends to be of secondary importance\nand in a pinch you can drop a few messages without hurting too much.\n\nWe've done moderately okay so far with having some system views where\nsome telemetry readings can be obtained, but there several drawbacks to\nthat approach that we should at some point solve. My opinion on this is\nthat we need to bite the bullet and develop separate infrastructure for\nreporting server metrics.\n\nThat said, I'm not opposed to having a patch somewhat as posted. I just\nthink that we should look into a new mechanism going forward.\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Thu, 5 May 2022 10:37:20 +0200","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\n recovery, backup, archive, streaming etc. activity messages to server logs\n along with ps display)"},{"msg_contents":"On Thu, May 5, 2022 at 2:07 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>\n> On 2022-May-05, Bharath Rupireddy wrote:\n>\n> > On Fri, Apr 29, 2022 at 4:11 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> > >\n> > > Did we ever consider the idea of using a new pg_stat_wal_activity_progress\n> > > view or something like that, using the backend_progress.c functionality?\n> > > I don't see it mentioned in the thread.\n> >\n> > IMO, progress reporting works well on a running server and at the\n> > moment. The WAL recovery/replay can happen even before the server\n> > opens up for connections\n>\n> It's definitely true that you wouldn't be able to use it for when the\n> server is not accepting connections.\n>\n> > and the progress report view can't be used\n> > for later analysis like how much time the restoring WAL files from\n> > archive location took\n>\n> This is true too -- progress report doesn't store historical data, only\n> current status.\n>\n> > and also the WAL file names can't be reported in progress reporting\n> > mechanism\n>\n> Also true.\n>\n> > (only integers columns, of course if required we can add text columns\n> > to pg_stat_get_progress_info).\n>\n> Yeah, I don't think adding text columns is terribly easy, because the\n> whole point of the progress reporting infrastructure is that it can be\n> updated very cheaply as atomic operations, and if you want to transmit\n> text columns, that's no longer possible.\n>\n> > Having the recovery info in server logs might help.\n>\n> I suppose it might.\n>\n> > I think reporting a long-running file processing operation (removing\n> > or syncing) within postgres is a generic problem (for snapshot,\n> > mapping, temporary (pgsql_tmp), temp relation files, old WAL file\n> > processing, WAL file processing during recovery etc.) and needs to be\n> > solved\n>\n> I agree up to here.\n>\n> > in two ways: 1) logging progress into server logs (which helps\n> > for analysis and report when the server isn't available for\n> > connections, crash recovery), a generic GUC\n> > log_file_processing_traffic = {none, medium, high} might help here\n> > (also proposed in [1]) and 2) pg_stat_file_processing_progress\n> > (extending progress reporting pg_stat_get_progress_info to have few\n> > text columns for current file name and directory path).\n>\n> I think using the server log to store telemetry data is not a great fit.\n> It bloats the log files and can be so slow as to block other operations\n> in the server. Server logs should normally be considered critical info\n> that's not okay to lose; telemetry tends to be of secondary importance\n> and in a pinch you can drop a few messages without hurting too much.\n>\n> We've done moderately okay so far with having some system views where\n> some telemetry readings can be obtained, but there several drawbacks to\n> that approach that we should at some point solve. My opinion on this is\n> that we need to bite the bullet and develop separate infrastructure for\n> reporting server metrics.\n>\n> I just\n> think that we should look into a new mechanism going forward.\n\nI completely agree that we must have new ways to report important\nserver metrics in an easily consumable fashion - it could be something\nlike server computing metrics (time taken, the file being processed\netc.) important/time-taking operations such as execution of archive\ncommand, restore command, recovery, checkpoint, old WAL files\nremoval/recycling, processing of various files - snapshot, mapping,\npgsql_tmp files, temp relation files etc. and writing these metrics to\na file or a stat table on the database itself (this table can have an\nautomatic mechanism of clean up or limit on number of rows or size of\nthe table so that the clients can read those metrics and use it in\nwhichever way they want. To start with, a simple structure of the\nmetrics can be {OPERATION text, START_TIME timestamp, END_TIME\ntimestamp, few text, double and/or integer flexible columns that each\noperation can use to store metrics}.\n\nI can start a discussion in a separate thread to seek more thoughts on\nthe server metrics part.\n\n> That said, I'm not opposed to having a patch somewhat as posted.\n\nThanks. I'm thinking if we should have a generic GUC\nlog_file_processing_traffic = {none, medium, high} that might help\nreporting other time taking file processing operations such as [1].\n\n[1] https://www.postgresql.org/message-id/CALj2ACW-ELOF5JT2zPavs95wbZ0BrLPrqvSZ7Ac%2BpjxCkmXtEQ%40mail.gmail.com\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Fri, 13 May 2022 18:32:31 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\n recovery, backup, archive, streaming etc. activity messages to server logs\n along with ps display)"},{"msg_contents":"On Fri, Apr 29, 2022 at 5:11 AM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n> Here's the rebased v9 patch.\n\nThis seems like it has enormous overlap with the existing\nfunctionality that we have from log_startup_progress_interval.\n\nI think that facility is also better-designed than this one. It prints\nout a message based on elapsed time, whereas this patch prints out a\nmessage based progress through the WAL. That means that if WAL replay\nisn't actually advancing for some reason, you just won't get any log\nmessages and you don't know whether it's advancing slowly or not at\nall or the server is just hung. With that facility you can distinguish\nthose cases.\n\nAlso, if for some reason we do think that amount of WAL replayed is\nthe right metric, rather than time, why would we only allow high=1\nsegment and low=128 segments, rather than say any number of MB or GB\nthat the user would like to configure?\n\nI suggest that if log_startup_progress_interval doesn't meet your\nneeds here, we should try to understand why not and maybe enhance it,\ninstead of adding a separate facility.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 13 May 2022 09:10:52 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\n recovery, backup, archive, streaming etc. activity messages to server logs\n along with ps display)"},{"msg_contents":"On Fri, May 13, 2022 at 6:41 PM Robert Haas <robertmhaas@gmail.com> wrote:\n>\n> On Fri, Apr 29, 2022 at 5:11 AM Bharath Rupireddy\n> <bharath.rupireddyforpostgres@gmail.com> wrote:\n> > Here's the rebased v9 patch.\n>\n> This seems like it has enormous overlap with the existing\n> functionality that we have from log_startup_progress_interval.\n>\n> I think that facility is also better-designed than this one. It prints\n> out a message based on elapsed time, whereas this patch prints out a\n> message based progress through the WAL. That means that if WAL replay\n> isn't actually advancing for some reason, you just won't get any log\n> messages and you don't know whether it's advancing slowly or not at\n> all or the server is just hung. With that facility you can distinguish\n> those cases.\n>\n> Also, if for some reason we do think that amount of WAL replayed is\n> the right metric, rather than time, why would we only allow high=1\n> segment and low=128 segments, rather than say any number of MB or GB\n> that the user would like to configure?\n>\n> I suggest that if log_startup_progress_interval doesn't meet your\n> needs here, we should try to understand why not and maybe enhance it,\n> instead of adding a separate facility.\n\nThanks Robert!\n\nIn a production environment (of course with a better management of\nserver logs) one can set log_wal_traffic to \"high\" and emit the\nrequired info to answer some of the customer questions like - \"How far\nthe server is in recovery? How much time recovery of each WAL file\napproximately took? How much time will it take to recover all the WAL\nfiles? What's the rate of recovery - time per WAL file? etc.\"\n\nWhereas ereport_startup_progress facility will help to emit log\nmessages only if \"some\" operation takes longer than set\nlog_startup_progress_interval time which may not serve the above\npurpose.\n\nActually, IMO a generic GUC log_file_processing_traffic = {none,\nmedium, high} to help server emit logs for all the critical file\nprocessing operations - WAL file recovery (as proposed here in this\nthread), temp file processing during server startup or restarts\n(log_startup_progress_interval can't be used here as postmaster\ndoesn't register for any timeouts) [1], snapshot and mapping file\nprocessing during checkpoint, temp relation files, removing old WAL\nfiles and so on.\n\nThoughts?\n\n[1] https://www.postgresql.org/message-id/CALj2ACW-ELOF5JT2zPavs95wbZ0BrLPrqvSZ7Ac%2BpjxCkmXtEQ%40mail.gmail.com\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Tue, 17 May 2022 11:01:57 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\n recovery, backup, archive, streaming etc. activity messages to server logs\n along with ps display)"},{"msg_contents":"On Tue, May 17, 2022 at 1:32 AM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n> In a production environment (of course with a better management of\n> server logs) one can set log_wal_traffic to \"high\" and emit the\n> required info to answer some of the customer questions like - \"How far\n> the server is in recovery? How much time recovery of each WAL file\n> approximately took? How much time will it take to recover all the WAL\n> files? What's the rate of recovery - time per WAL file? etc.\"\n>\n> Whereas ereport_startup_progress facility will help to emit log\n> messages only if \"some\" operation takes longer than set\n> log_startup_progress_interval time which may not serve the above\n> purpose.\n\nI think you are misunderstanding what an \"operation\" is here. If you\ntake the time to test out that facility, I think you will find that it\nis quite capable of answering all of those questions.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 17 May 2022 12:24:24 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\n recovery, backup, archive, streaming etc. activity messages to server logs\n along with ps display)"},{"msg_contents":"On Fri, May 13, 2022 at 09:10:52AM -0400, Robert Haas wrote:\n> I suggest that if log_startup_progress_interval doesn't meet your\n> needs here, we should try to understand why not and maybe enhance it,\n> instead of adding a separate facility.\n\n+1. AFAICT it should be possible to make the log_startup_progress_interval\nmachinery extensible enough to be reused for several other tasks that can\ntake a while but have little visibility.\n\nSince this thread has been inactive for over 2 months, I'm marking the\ncommitfest entry as Waiting on Author.\n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Mon, 18 Jul 2022 16:01:08 -0700","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\n recovery, backup, archive, streaming etc. activity messages to server logs\n along with ps display)"},{"msg_contents":"On Fri, May 13, 2022 at 6:41 PM Robert Haas <robertmhaas@gmail.com> wrote:\n>\n> On Fri, Apr 29, 2022 at 5:11 AM Bharath Rupireddy\n> <bharath.rupireddyforpostgres@gmail.com> wrote:\n> > Here's the rebased v9 patch.\n>\n> This seems like it has enormous overlap with the existing\n> functionality that we have from log_startup_progress_interval.\n>\n> I think that facility is also better-designed than this one. It prints\n> out a message based on elapsed time, whereas this patch prints out a\n> message based progress through the WAL. That means that if WAL replay\n> isn't actually advancing for some reason, you just won't get any log\n> messages and you don't know whether it's advancing slowly or not at\n> all or the server is just hung. With that facility you can distinguish\n> those cases.\n>\n> Also, if for some reason we do think that amount of WAL replayed is\n> the right metric, rather than time, why would we only allow high=1\n> segment and low=128 segments, rather than say any number of MB or GB\n> that the user would like to configure?\n>\n> I suggest that if log_startup_progress_interval doesn't meet your\n> needs here, we should try to understand why not and maybe enhance it,\n> instead of adding a separate facility.\n\nAfter thinking for a while, I agree with Robert and others that we\ncould leverage the existing log_startup_progress_interval mechanism\nfor reporting which WAL file currently is being replayed. I added\ncurrent TLI (helps to construct the WAL file name from current LSN)\nand current WAL file source (helps to know where the WAL files was\nfetched from) to the existing \"redo in progress, elapsed time:...\"\nmessage. This very well serves the purpose of identifying the issues\nsuch as the restore command taking a lot of time (>\nlog_startup_progress_interval for instance), WAL replay rate on\nstandby or primary for long recoveries and so on.\n\nHowever, ereport_startup_progress isn't enabled on standby to not let\nit bloat the server logs. I believe the \"redo in progress, elapsed\ntime:...\" message can provide some important info/metric for standby\ntoo and there's no way for the users to enable it on standbys today.\nFor instance, users can know how well the standby fares in replaying,\nthey can figure this out, by looking at two or more such messages. If\nenabled, with default value of 10 sec for\nlog_startup_progress_interval, the standby can emit 8640 messages per\nday which is too much. I'm not sure if we are okay to change the\ndefault value of log_startup_progress_interval to say 1min or 5min so\nthat 1440 messages are emitted. In production environments, typically\nusers may or may not be interested if recovery takes just 10sec, but\nthey really are interested if it takes in the order of minutes.\nBasically, I would like to enable \"redo in progress, elapsed time:...\"\nmessage for standbys too.\n\nThoughts?\n\nPSA v10 patch with enhanced \"redo in progress, elapsed time:...\"\nmessage. Note that it's not a final patch though.\n\nRegards,\nBharath Rupireddy.","msg_date":"Tue, 26 Jul 2022 15:36:57 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\n recovery, backup, archive, streaming etc. activity messages to server logs\n along with ps display)"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nAs discussed in [1], isn't it a better idea to add some of activity\\nmessages [2] such as recovery, archive, backup, streaming etc. to\\nserver logs at LOG level? They are currently being set into ps display\\nwhich is good if the postgres is being run on a standalone box/VM\\nwhere users can see the ps display, but it doesn't help much in case\\nthe postgres is being run on a cloud environment where users don't\\nhave access to ps display output. Moreover, the ps display is\\ntransient and will not help to analyze after an issue occurs.\\n\\nHaving the above messages in the server logs will be useful to\\nunderstand how the system is/was doing/progressing with these\\n(sometimes time-intensive) operations.\\n\\nThoughts?\\n\\n[1] - https://www.postgresql.org/message-id/CALj2ACWE6D%2BbbKnA9tQ0d-pPN1n4xrsFPOmiURm2my3qt0w3OA%40mail.gmail.com\\n[2] -\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"waiting for %s\\\",\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"recovering %s\\\",\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"performing %s%s%s\\\",\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"archiving %s\\\", xlog);\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"failed on %s\\\", xlog);\\nelog(DEBUG1, \\\"archived write-ahead log file \\\\\\\"%s\\\\\\\"\\\", xlog);\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"sending backup \\\\\\\"%s\\\\\\\"\\\",\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"restarting at %X/%X\\\",\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"streaming %X/%X\\\",\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"streaming %X/%X\\\",\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 23:08:51 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"add recovery, backup, archive, streaming etc. activity messages to\\n server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On 11/10/21, 9:43 AM, \\\"Bharath Rupireddy\\\" <bharath.rupireddyforpostgres@gmail.com> wrote:\\r\\n> As discussed in [1], isn't it a better idea to add some of activity\\r\\n> messages [2] such as recovery, archive, backup, streaming etc. to\\r\\n> server logs at LOG level? They are currently being set into ps display\\r\\n> which is good if the postgres is being run on a standalone box/VM\\r\\n> where users can see the ps display, but it doesn't help much in case\\r\\n> the postgres is being run on a cloud environment where users don't\\r\\n> have access to ps display output. Moreover, the ps display is\\r\\n> transient and will not help to analyze after an issue occurs.\\r\\n>\\r\\n> Having the above messages in the server logs will be useful to\\r\\n> understand how the system is/was doing/progressing with these\\r\\n> (sometimes time-intensive) operations.\\r\\n\\r\\nI think this would make the logs far too noisy for many servers. For\\r\\narchiving alone, this could cause tens of thousands more log messages\\r\\nper hour on a busy system. I think you can already see such\\r\\ninformation at a debug level, anyway.\\r\\n\\r\\nNathan\\r\\n\\r\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 18:31:20 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive,\\n streaming etc. activity messages to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 10, 2021, at 2:38 PM, Bharath Rupireddy wrote:\\n> As discussed in [1], isn't it a better idea to add some of activity\\n> messages [2] such as recovery, archive, backup, streaming etc. to\\n> server logs at LOG level? They are currently being set into ps display\\n> which is good if the postgres is being run on a standalone box/VM\\n> where users can see the ps display, but it doesn't help much in case\\n> the postgres is being run on a cloud environment where users don't\\n> have access to ps display output. Moreover, the ps display is\\n> transient and will not help to analyze after an issue occurs.\\nBesides recovery, the other activities already provide information through\\nviews. I might be missing something but it seems the current views already\\nprovide the information that ps displays.\\n\\n> Having the above messages in the server logs will be useful to\\n> understand how the system is/was doing/progressing with these\\n> (sometimes time-intensive) operations.\\nIt could be useful for a root cause analysis, however, you are also increasing\\nthe number of LOG messages in a system that doesn't/won't require such\\ninformation. It is fine to add additional DEBUG messages if there isn't a\\nsimilar one yet. If at least the message level were module-controlled, you\\ncould modify a setting to gather more messages from a specific module.\\nUnfortunately, that's not possible so we should avoid superfluous messages.\\n\\n\\n--\\nEuler Taveira\\nEDB https://www.enterprisedb.com/\\n\\nOn Wed, Nov 10, 2021, at 2:38 PM, Bharath Rupireddy wrote:As discussed in [1], isn't it a better idea to add some of activitymessages [2] such as recovery, archive, backup, streaming etc. toserver logs at LOG level? They are currently being set into ps displaywhich is good if the postgres is being run on a standalone box/VMwhere users can see the ps display, but it doesn't help much in casethe postgres is being run on a cloud environment where users don'thave access to ps display output. Moreover, the ps display istransient and will not help to analyze after an issue occurs.Besides recovery, the other activities already provide information throughviews. I might be missing something but it seems the current views alreadyprovide the information that ps displays.Having the above messages in the server logs will be useful tounderstand how the system is/was doing/progressing with these(sometimes time-intensive) operations.It could be useful for a root cause analysis, however, you are also increasingthe number of LOG messages in a system that doesn't/won't require suchinformation. It is fine to add additional DEBUG messages if there isn't asimilar one yet. If at least the message level were module-controlled, youcould modify a setting to gather more messages from a specific module.Unfortunately, that's not possible so we should avoid superfluous messages.--Euler TaveiraEDB https://www.enterprisedb.com/\",\n \"msg_date\": \"Wed, 10 Nov 2021 17:09:43 -0300\",\n \"msg_from\": \"\\\"Euler Taveira\\\" <euler@eulerto.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive,\\n streaming etc. activity messages to server\\n logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 12:01 AM Bossart, Nathan <bossartn@amazon.com> wrote:\\n>\\n> On 11/10/21, 9:43 AM, \\\"Bharath Rupireddy\\\" <bharath.rupireddyforpostgres@gmail.com> wrote:\\n> > As discussed in [1], isn't it a better idea to add some of activity\\n> > messages [2] such as recovery, archive, backup, streaming etc. to\\n> > server logs at LOG level? They are currently being set into ps display\\n> > which is good if the postgres is being run on a standalone box/VM\\n> > where users can see the ps display, but it doesn't help much in case\\n> > the postgres is being run on a cloud environment where users don't\\n> > have access to ps display output. Moreover, the ps display is\\n> > transient and will not help to analyze after an issue occurs.\\n> >\\n> > Having the above messages in the server logs will be useful to\\n> > understand how the system is/was doing/progressing with these\\n> > (sometimes time-intensive) operations.\\n>\\n> I think this would make the logs far too noisy for many servers. For\\n> archiving alone, this could cause tens of thousands more log messages\\n> per hour on a busy system. I think you can already see such\\n> information at a debug level, anyway.\\n\\nYou are right, having them at LOG level may blow up the log files and\\ndisk storage, having them at DEBUG1 like other messages would at least\\nhelp than having them as activitymsg in ps display.\\n\\nAnd not all of the activitymsgs are logged at DEBUGX level, at least I\\ndon't see them(on a quick search).\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"waiting for %s\\\",\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"recovering %s\\\",\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"performing %s%s%s\\\",\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"sending backup \\\\\\\"%s\\\\\\\"\\\",\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"restarting at %X/%X\\\",\\nsnprintf(activitymsg, sizeof(activitymsg), \\\"streaming %X/%X\\\",\\n\\nBut for the snprintf(activitymsg, sizeof(activitymsg), \\\"archiving %s\\\",\\nxlog); we have elog(DEBUG1, \\\"archived write-ahead log file \\\\\\\"%s\\\\\\\"\\\",\\nxlog); after the archiving command. It is also good to have a similar\\ndebug message before archive command execution, we can get to know\\nwhether archive command is executed or not, if yes how much time did\\nit take etc.?\\n\\nI'm not sure whether it is okay to do the following, for every of the\\nactivitymsg, before the command execution, we have an elog(DEBUG1\\nmessage and after the command execution another elog(DEBUG1 message.\\n\\nThoughts?\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 11:20:44 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 1:40 AM Euler Taveira <euler@eulerto.com> wrote:\\n>\\n> On Wed, Nov 10, 2021, at 2:38 PM, Bharath Rupireddy wrote:\\n>\\n> As discussed in [1], isn't it a better idea to add some of activity\\n> messages [2] such as recovery, archive, backup, streaming etc. to\\n> server logs at LOG level? They are currently being set into ps display\\n> which is good if the postgres is being run on a standalone box/VM\\n> where users can see the ps display, but it doesn't help much in case\\n> the postgres is being run on a cloud environment where users don't\\n> have access to ps display output. Moreover, the ps display is\\n> transient and will not help to analyze after an issue occurs.\\n>\\n> Besides recovery, the other activities already provide information through\\n> views. I might be missing something but it seems the current views already\\n> provide the information that ps displays.\\n>\\n> Having the above messages in the server logs will be useful to\\n> understand how the system is/was doing/progressing with these\\n> (sometimes time-intensive) operations.\\n>\\n> It could be useful for a root cause analysis, however, you are also increasing\\n> the number of LOG messages in a system that doesn't/won't require such\\n> information. It is fine to add additional DEBUG messages if there isn't a\\n> similar one yet. If at least the message level were module-controlled, you\\n> could modify a setting to gather more messages from a specific module.\\n> Unfortunately, that's not possible so we should avoid superfluous messages.\\n\\nYes, at least having the messages at DEBUG1 level would help do the\\nRCA after an issue occurs. Please see my response upthread at [1].\\n\\n[1] - https://www.postgresql.org/message-id/CALj2ACXVwnHhek9bo9wgamtW9SBinj4ohF1f1Yf3Su2o5WarWg%40mail.gmail.com\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 11:23:09 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-10, Bossart, Nathan wrote:\\n\\n> On 11/10/21, 9:43 AM, \\\"Bharath Rupireddy\\\" <bharath.rupireddyforpostgres@gmail.com> wrote:\\n> > As discussed in [1], isn't it a better idea to add some of activity\\n> > messages [2] such as recovery, archive, backup, streaming etc. to\\n> > server logs at LOG level?\\n\\n> I think this would make the logs far too noisy for many servers. For\\n> archiving alone, this could cause tens of thousands more log messages\\n> per hour on a busy system. I think you can already see such\\n> information at a debug level, anyway.\\n\\nYeah. If we had some sort of ring buffer in which to store these\\nmessages, the user could examine them through a view; they would still\\nbe accessible in a running server, but they would not be written to the\\nserver log.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 09:37:41 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On 11/10/21, 9:51 PM, \\\"Bharath Rupireddy\\\" <bharath.rupireddyforpostgres@gmail.com> wrote:\\r\\n> But for the snprintf(activitymsg, sizeof(activitymsg), \\\"archiving %s\\\",\\r\\n> xlog); we have elog(DEBUG1, \\\"archived write-ahead log file \\\\\\\"%s\\\\\\\"\\\",\\r\\n> xlog); after the archiving command. It is also good to have a similar\\r\\n> debug message before archive command execution, we can get to know\\r\\n> whether archive command is executed or not, if yes how much time did\\r\\n> it take etc.?\\r\\n\\r\\nThere is a DEBUG3 that emits \\\"executing archive command...\\\" prior to\\r\\nexecuting the archive command.\\r\\n\\r\\n> I'm not sure whether it is okay to do the following, for every of the\\r\\n> activitymsg, before the command execution, we have an elog(DEBUG1\\r\\n> message and after the command execution another elog(DEBUG1 message.\\r\\n\\r\\nThe general idea seems reasonable to me.\\r\\n\\r\\nNathan\\r\\n\\r\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 15:11:49 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive,\\n streaming etc. activity messages to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On 11/11/21, 4:38 AM, \\\"Alvaro Herrera\\\" <alvherre@alvh.no-ip.org> wrote:\\r\\n> Yeah. If we had some sort of ring buffer in which to store these\\r\\n> messages, the user could examine them through a view; they would still\\r\\n> be accessible in a running server, but they would not be written to the\\r\\n> server log.\\r\\n\\r\\nThat's an interesting idea. It'd be nice to have a way to see more\\r\\ndebugging information without exploding the server log.\\r\\n\\r\\nNathan\\r\\n\\r\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 15:18:16 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive,\\n streaming etc. activity messages to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021, at 2:50 AM, Bharath Rupireddy wrote:\\n> But for the snprintf(activitymsg, sizeof(activitymsg), \\\"archiving %s\\\",\\n> xlog); we have elog(DEBUG1, \\\"archived write-ahead log file \\\\\\\"%s\\\\\\\"\\\",\\n> xlog); after the archiving command. It is also good to have a similar\\n> debug message before archive command execution, we can get to know\\n> whether archive command is executed or not, if yes how much time did\\n> it take etc.?\\nIf you are only interested in how much time it took to archive a WAL file, you\\ndon't need 2 messages (before and after); one message is sufficient. If you\\ncalculate the difference and then print it as part of the existent message.\\nCalculate the difference iif the log level is at least DEBUG1 (same as the log\\nmessage).\\n\\nMessages related to system() e.g. archive and recovery, you could measure the\\nexecution time in your own script/program.\\n\\nI like Alvaro's idea of implementing a ring buffer for this kind of activity.\\nThis implementation could be applied to checkpoints, archiving, recovery,\\nreplication, base backup, WAL usage, connections and others.\\n\\n\\n--\\nEuler Taveira\\nEDB https://www.enterprisedb.com/\\n\\nOn Thu, Nov 11, 2021, at 2:50 AM, Bharath Rupireddy wrote:But for the snprintf(activitymsg, sizeof(activitymsg), \\\"archiving %s\\\",xlog); we have elog(DEBUG1, \\\"archived write-ahead log file \\\\\\\"%s\\\\\\\"\\\",xlog); after the archiving command. It is also good to have a similardebug message before archive command execution, we can get to knowwhether archive command is executed or not, if yes how much time didit take etc.?If you are only interested in how much time it took to archive a WAL file, youdon't need 2 messages (before and after); one message is sufficient. If youcalculate the difference and then print it as part of the existent message.Calculate the difference iif the log level is at least DEBUG1 (same as the logmessage).Messages related to system() e.g. archive and recovery, you could measure theexecution time in your own script/program.I like Alvaro's idea of implementing a ring buffer for this kind of activity.This implementation could be applied to checkpoints, archiving, recovery,replication, base backup, WAL usage, connections and others.--Euler TaveiraEDB https://www.enterprisedb.com/\",\n \"msg_date\": \"Thu, 11 Nov 2021 12:46:47 -0300\",\n \"msg_from\": \"\\\"Euler Taveira\\\" <euler@eulerto.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive,\\n streaming etc. activity messages to server\\n logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 03:18:16PM +0000, Bossart, Nathan wrote:\\n> On 11/11/21, 4:38 AM, \\\"Alvaro Herrera\\\" <alvherre@alvh.no-ip.org> wrote:\\n> > Yeah. If we had some sort of ring buffer in which to store these\\n> > messages, the user could examine them through a view; they would still\\n> > be accessible in a running server, but they would not be written to the\\n> > server log.\\n> \\n> That's an interesting idea. It'd be nice to have a way to see more\\n> debugging information without exploding the server log.\\n\\nUh, that's what dmesg does, which I mentioned earlier.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> https://momjian.us\\n EDB https://enterprisedb.com\\n\\n If only the physical world exists, free will is an illusion.\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 12:46:38 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 6:07 PM Alvaro Herrera <alvherre@alvh.no-ip.org>\\nwrote:\\n\\n> On 2021-Nov-10, Bossart, Nathan wrote:\\n>\\n> > On 11/10/21, 9:43 AM, \\\"Bharath Rupireddy\\\" <\\n> bharath.rupireddyforpostgres@gmail.com> wrote:\\n> > > As discussed in [1], isn't it a better idea to add some of activity\\n> > > messages [2] such as recovery, archive, backup, streaming etc. to\\n> > > server logs at LOG level?\\n>\\n> > I think this would make the logs far too noisy for many servers. For\\n> > archiving alone, this could cause tens of thousands more log messages\\n> > per hour on a busy system. I think you can already see such\\n> > information at a debug level, anyway.\\n>\\n> Yeah. If we had some sort of ring buffer in which to store these\\n> messages, the user could examine them through a view; they would still\\n> be accessible in a running server, but they would not be written to the\\n> server log.\\n>\\n\\nThat's a good idea. How about also adding some GUC(s) to the log archive,\\nrecovery related log messages just like we have for checkpoints, autovacuum\\netc? Maybe something like log_archive, log_recovery etc.\\n\\n--\\nWith Regards,\\nAshutosh Sharma.\\n\\nOn Thu, Nov 11, 2021 at 6:07 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:On 2021-Nov-10, Bossart, Nathan wrote:\\n\\n> On 11/10/21, 9:43 AM, \\\"Bharath Rupireddy\\\" <bharath.rupireddyforpostgres@gmail.com> wrote:\\n> > As discussed in [1], isn't it a better idea to add some of activity\\n> > messages [2] such as recovery, archive, backup, streaming etc. to\\n> > server logs at LOG level?\\n\\n> I think this would make the logs far too noisy for many servers. For\\n> archiving alone, this could cause tens of thousands more log messages\\n> per hour on a busy system. I think you can already see such\\n> information at a debug level, anyway.\\n\\nYeah. If we had some sort of ring buffer in which to store these\\nmessages, the user could examine them through a view; they would still\\nbe accessible in a running server, but they would not be written to the\\nserver log.That's a good idea. How about also adding some GUC(s) to the log archive, recovery related log messages just like we have for checkpoints, autovacuum etc? Maybe something like log_archive, log_recovery etc.--With Regards,Ashutosh Sharma.\",\n \"msg_date\": \"Fri, 12 Nov 2021 21:10:48 +0530\",\n \"msg_from\": \"Ashutosh Sharma <ashu.coek88@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 11, 2021 at 8:41 PM Bossart, Nathan <bossartn@amazon.com> wrote:\\n> > I'm not sure whether it is okay to do the following, for every of the\\n> > activitymsg, before the command execution, we have an elog(DEBUG1\\n> > message and after the command execution another elog(DEBUG1 message.\\n>\\n> The general idea seems reasonable to me.\\n\\nHere's a simple patch that I could come up with. Please have a look.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Sat, 13 Nov 2021 19:29:36 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"A bunch of these now execute snprintf()s unnecessarily. I think these\\nshould be made conditional on message_level_is_interesting(DEBUG1) to\\navoid that overhead.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\\"You're _really_ hosed if the person doing the hiring doesn't understand\\nrelational systems: you end up with a whole raft of programmers, none of\\nwhom has had a Date with the clue stick.\\\" (Andrew Sullivan)\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 12:34:37 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 13, 2021 at 9:04 PM Alvaro Herrera <alvherre@alvh.no-ip.org>\\nwrote:\\n>\\n> A bunch of these now execute snprintf()s unnecessarily. I think these\\n> should be made conditional on message_level_is_interesting(DEBUG1) to\\n> avoid that overhead.\\n\\nThanks. Attaching the v2 to avoid that by directly using the message in\\nereport instead of activitymsg.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Sun, 14 Nov 2021 17:47:41 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On 11/14/21, 4:18 AM, \\\"Bharath Rupireddy\\\" <bharath.rupireddyforpostgres@gmail.com> wrote:\\r\\n> Thanks. Attaching the v2 to avoid that by directly using the message\\r\\n> in ereport instead of activitymsg.\\r\\n\\r\\n+ ereport(DEBUG1,\\r\\n+ (errmsg_internal(\\\"waiting for %s\\\", xlogfname)));\\r\\n\\r\\nAs a general comment, I think we should include more detail than what\\r\\ngoes in the ps title. While we probably want to keep the ps title,\\r\\nbrief, I don't see any issue with including a bit more context in\\r\\ndebug statements.\\r\\n\\r\\n+ ereport(DEBUG1,\\r\\n+ (errmsg_internal(\\\"executing archive command \\\\\\\"%s\\\\\\\" for archving %s\\\",\\r\\n+ xlogarchcmd, xlog)));\\r\\n\\r\\nI wonder if it's necessary to include the \\\"for archiving\\\" part in this\\r\\none. Typically, the xlogarchcmd will include it somewhere.\\r\\n\\r\\nNathan\\r\\n\\r\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 22:36:09 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive,\\n streaming etc. activity messages to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 10:36:09PM +0000, Bossart, Nathan wrote:\\n> + ereport(DEBUG1,\\n> + (errmsg_internal(\\\"executing archive command \\\\\\\"%s\\\\\\\" for archving %s\\\",\\n> + xlogarchcmd, xlog)));\\n> \\n> I wonder if it's necessary to include the \\\"for archiving\\\" part in this\\n> one. Typically, the xlogarchcmd will include it somewhere.\\n\\nDon't see any need to add that either (s/archving/archiving/, btw).\\n\\n@@ -1000,6 +1004,10 @@ XLogWalRcvFlush(bool dying, TimeLineID tli)\\n+ ereport(DEBUG1,\\n+ (errmsg_internal(\\\"streaming %X/%X\\\",\\n+ LSN_FORMAT_ARGS(LogstreamResult.Write))));\\nHaving the same wording for the WAL receiver and the WAL sender could\\nbe confusing when it comes to a cascading standby node, particularly\\nif log_line_prefix does not include the process PID.\\n\\n+ ereport(DEBUG1,\\n+ (errmsg_internal(\\\"streaming %X/%X\\\",\\n+ LSN_FORMAT_ARGS(sentPtr))));\\nAnyway, those two ones are going to make the logs much more noisy, no?\\nThe same could be said about XLogFileRead(), joining the point of\\nNathan's upthread. So I cannot get excited by this change.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 16 Nov 2021 12:13:51 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On 11/15/21, 7:14 PM, \\\"Michael Paquier\\\" <michael@paquier.xyz> wrote:\\r\\n> + ereport(DEBUG1,\\r\\n> + (errmsg_internal(\\\"streaming %X/%X\\\",\\r\\n> + LSN_FORMAT_ARGS(sentPtr))));\\r\\n> Anyway, those two ones are going to make the logs much more noisy, no?\\r\\n> The same could be said about XLogFileRead(), joining the point of\\r\\n> Nathan's upthread. So I cannot get excited by this change.\\r\\n\\r\\nYeah, this might even be too noisy for DEBUG5.\\r\\n\\r\\nNathan\\r\\n\\r\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 16:01:54 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive,\\n streaming etc. activity messages to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-16, Bossart, Nathan wrote:\\n\\n> On 11/15/21, 7:14 PM, \\\"Michael Paquier\\\" <michael@paquier.xyz> wrote:\\n> > + ereport(DEBUG1,\\n> > + (errmsg_internal(\\\"streaming %X/%X\\\",\\n> > + LSN_FORMAT_ARGS(sentPtr))));\\n> > Anyway, those two ones are going to make the logs much more noisy, no?\\n> > The same could be said about XLogFileRead(), joining the point of\\n> > Nathan's upthread. So I cannot get excited by this change.\\n> \\n> Yeah, this might even be too noisy for DEBUG5.\\n\\nNod. And if you're at DEBUG5, the log contains so much other crap that\\nit is unusable for any purposes anyway.\\n\\nMy opinion is that adding these things willy-nilly is not a solution to\\nany actual problem. Adding a few additional log lines that are\\nlow-volume at DEBUG1 might be useful, but below that (DEBUG2 etc) it's\\nnot good for anything other than specific development, IMO. At least\\nthis particular one for streaming replication I think we should not\\ninclude.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 13:26:49 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 01:26:49PM -0300, Alvaro Herrera wrote:\\n> My opinion is that adding these things willy-nilly is not a solution to\\n> any actual problem. Adding a few additional log lines that are\\n> low-volume at DEBUG1 might be useful, but below that (DEBUG2 etc) it's\\n> not good for anything other than specific development, IMO. At least\\n> this particular one for streaming replication I think we should not\\n> include.\\n\\nLooking at v2, I think that this leaves the additions of the DEBUG1\\nentries in SendBaseBackup() and WalRcvWaitForStartPosition(), then.\\nThe one in pgarch.c does not provide any additional information as the\\nsegment to-be-archived should be part of the command.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 17 Nov 2021 11:31:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 8:01 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Tue, Nov 16, 2021 at 01:26:49PM -0300, Alvaro Herrera wrote:\\n> > My opinion is that adding these things willy-nilly is not a solution to\\n> > any actual problem. Adding a few additional log lines that are\\n> > low-volume at DEBUG1 might be useful, but below that (DEBUG2 etc) it's\\n> > not good for anything other than specific development, IMO. At least\\n> > this particular one for streaming replication I think we should not\\n> > include.\\n>\\n> Looking at v2, I think that this leaves the additions of the DEBUG1\\n> entries in SendBaseBackup() and WalRcvWaitForStartPosition(), then.\\n> The one in pgarch.c does not provide any additional information as the\\n> segment to-be-archived should be part of the command.\\n\\nThank you all for the inputs. Here's the patch that I've come up with.\\n\\nUpon thinking further, having at least the messages at LOG level [1]\\nwould be helpful to know what's happening with the system while in\\nrecovery. Although these messages at LOG level seem to be filling up\\nthe server logs, having a good log consumption and rotation mechanism\\n(I'm sure every major postgres vendor would have one) would be\\nsufficient to allay that concern.\\n\\nThese LOG messages would help us know how much time a restore command\\ntakes to fetch the WAL file and what is the current WAL file the\\nserver is recovering and where is it recovering from. The customer\\noften asks questions like: when will my server come up? how much time\\ndoes the recovery of my server take?\\n\\nAs a developer or admin, one can monitor these logs and do bunch of things:\\n1) see how many WAL files left to be recovered by looking at the WAL\\nfiles in the archive location or pg_wal directory or from primary\\n2) provide an approximate estimation of when the server will come up\\nor how much more the recovery takes by looking at these previous LOG\\nmessages, one can know the number of WAL files that server recovered\\nover a minute and with the help of left-over WAL files calculated from\\n(1).\\n\\n[1]\\nereport(LOG,\\n (errmsg(\\\"waiting for WAL segment \\\\\\\"%s\\\\\\\" from archive\\\",\\n xlogfname)));\\n\\nereport(LOG,\\n (errmsg(\\\"restored WAL segment \\\\\\\"%s\\\\\\\" from archive\\\",\\n xlogfname)));\\n\\nereport(LOG,\\n (errmsg(\\\"recovering WAL segment \\\\\\\"%s\\\\\\\" from source \\\\\\\"%s\\\\\\\"\\\",\\n xlogfname, srcname)));\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Tue, 7 Dec 2021 18:28:24 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 07, 2021 at 06:28:24PM +0530, Bharath Rupireddy wrote:\\n> Upon thinking further, having at least the messages at LOG level [1]\\n> would be helpful to know what's happening with the system while in\\n> recovery. Although these messages at LOG level seem to be filling up\\n> the server logs, having a good log consumption and rotation mechanism\\n> (I'm sure every major postgres vendor would have one) would be\\n> sufficient to allay that concern.\\n\\n+ ereport(LOG,\\n+ (errmsg(\\\"recovering WAL segment \\\\\\\"%s\\\\\\\" from source \\\\\\\"%s\\\\\\\"\\\",\\n+ xlogfname, srcname)));\\nIsn't this approach an issue for translations? It seems to me that\\nterms like \\\"stream\\\" or \\\"archive\\\" had better be translated properly,\\nmeaning that this needs a fully-constructed sentence.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 8 Dec 2021 13:35:48 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 8, 2021 at 10:05 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Tue, Dec 07, 2021 at 06:28:24PM +0530, Bharath Rupireddy wrote:\\n> > Upon thinking further, having at least the messages at LOG level [1]\\n> > would be helpful to know what's happening with the system while in\\n> > recovery. Although these messages at LOG level seem to be filling up\\n> > the server logs, having a good log consumption and rotation mechanism\\n> > (I'm sure every major postgres vendor would have one) would be\\n> > sufficient to allay that concern.\\n>\\n> + ereport(LOG,\\n> + (errmsg(\\\"recovering WAL segment \\\\\\\"%s\\\\\\\" from source \\\\\\\"%s\\\\\\\"\\\",\\n> + xlogfname, srcname)));\\n> Isn't this approach an issue for translations? It seems to me that\\n> terms like \\\"stream\\\" or \\\"archive\\\" had better be translated properly,\\n> meaning that this needs a fully-constructed sentence.\\n\\nThanks for taking a look at the patch. How about the attached v4?\\n\\nI added a CF entry - https://commitfest.postgresql.org/36/3443/\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Wed, 8 Dec 2021 10:47:57 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 08, 2021 at 10:47:57AM +0530, Bharath Rupireddy wrote:\\n> Thanks for taking a look at the patch. How about the attached v4?\\n> \\n> I added a CF entry - https://commitfest.postgresql.org/36/3443/\\n\\n+ else if (source == XLOG_FROM_STREAM)\\n+ ereport(LOG,\\n+ (errmsg(\\\"recovering WAL segment \\\\\\\"%s\\\\\\\" received from primary\\\",\\n+ xlogfname)));\\nThis is incorrect when using a cascading standby. And perhaps this\\ncould use a switch/case?\\n\\nWhile quickly testing, I got reminded that the LOG for a segment\\nretrieved from the local pg_wal would generate some noise when running\\nfor example the bootstrap process. Is this one really interesting\\ncompared to the two others?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 8 Dec 2021 16:21:55 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 8, 2021 at 12:52 PM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Wed, Dec 08, 2021 at 10:47:57AM +0530, Bharath Rupireddy wrote:\\n> > Thanks for taking a look at the patch. How about the attached v4?\\n> >\\n> > I added a CF entry - https://commitfest.postgresql.org/36/3443/\\n>\\n> + else if (source == XLOG_FROM_STREAM)\\n> + ereport(LOG,\\n> + (errmsg(\\\"recovering WAL segment \\\\\\\"%s\\\\\\\" received from primary\\\",\\n> + xlogfname)));\\n> This is incorrect when using a cascading standby.\\n\\nI visualize the setup like this: primary -> standby -> cascading standby\\n\\nFor standby, \\\"received from primary\\\" makes sense. For cascading\\nstandby too it makes sense because the standby still acts as primary\\nfor cascading standby, no? And we don't distinguish any other existing\\nmessages for a standby acting as a cascading standby right? The\\ncomments around the enum XLOG_FROM_STREAM and usage of it still says\\nfrom \\\"primary\\\".\\n\\nDo you want me to add \\\"received from stream\\\" in general?\\n\\n> And perhaps this could use a switch/case?\\n\\nYes, I will change it to switch-case in the next version.\\n\\n> While quickly testing, I got reminded that the LOG for a segment\\n> retrieved from the local pg_wal would generate some noise when running\\n> for example the bootstrap process. Is this one really interesting\\n> compared to the two others?\\n\\nYes. I observed that for a simple primary-standby setup, inserting\\n10million rows generated around 40 WAL files and so 40 LOG messages in\\nthe standby server logs, in a typical production environment with\\nheavy-write workload, the amount of LOG generation is going to be much\\nmore. And we don't want to have them as DEBUG1 messages and enable\\nDEBUG1 just for recovery messages that one is interested in.\\n\\nHow about we invent a new GUC log_recovery, with default set to false,\\nsimilar to log_checkpoints? And, these messages can be like:\\nereport(log_recovery ? LOG : DEBUG1, .....)? This GUC can be useful to\\nadd some recovery stats in future as well. Thoughts?\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Wed, 8 Dec 2021 14:00:12 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 08, 2021 at 02:00:12PM +0530, Bharath Rupireddy wrote:\\n> For standby, \\\"received from primary\\\" makes sense. For cascading\\n> standby too it makes sense because the standby still acts as primary\\n> for cascading standby, no? And we don't distinguish any other existing\\n> messages for a standby acting as a cascading standby right? The\\n> comments around the enum XLOG_FROM_STREAM and usage of it still says\\n> from \\\"primary\\\".\\n\\nQuoting the docs:\\nhttps://www.postgresql.org/docs/devel/glossary.html#GLOSSARY-PRIMARY-SERVER\\nAnd a standby is no such thing.\\n\\n> Do you want me to add \\\"received from stream\\\" in general?\\n\\n\\\"from stream\\\" would be fine, IMHO.\\n\\n> How about we invent a new GUC log_recovery, with default set to false,\\n> similar to log_checkpoints? And, these messages can be like:\\n> ereport(log_recovery ? LOG : DEBUG1, .....)? This GUC can be useful to\\n> add some recovery stats in future as well. Thoughts?\\n\\nThis discussion does not justify a new GUC at this stage IMO.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 8 Dec 2021 17:38:14 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 8, 2021 at 2:08 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> > Do you want me to add \\\"received from stream\\\" in general?\\n>\\n> \\\"from stream\\\" would be fine, IMHO.\\n\\nDone in the attached v5 patch.\\n\\n> > How about we invent a new GUC log_recovery, with default set to false,\\n> > similar to log_checkpoints? And, these messages can be like:\\n> > ereport(log_recovery ? LOG : DEBUG1, .....)? This GUC can be useful to\\n> > add some recovery stats in future as well. Thoughts?\\n>\\n> This discussion does not justify a new GUC at this stage IMO.\\n\\nHaving \\\"log_recovery\\\" GUC gives the flexibility to the users to enable\\nthe recovery logs added by this patch if they are concerned otherwise\\nthe logs just be at DEBUG1 level. Setting log_min_messages level to\\nDEBUG1 would definitely flood the server logs.\\n\\n> While quickly testing, I got reminded that the LOG for a segment\\n> retrieved from the local pg_wal would generate some noise when running\\n> for example the bootstrap process. Is this one really interesting\\n> compared to the two others?\\n\\nIf we just have the archive related message at LOG level and others at\\nDEBUG1 level, it is a kind of partial recovery log information that\\none has in the log files, missing recovery messages from the pg_wal\\ndirectory in the logs.\\n\\nAnother option is to just have all the messages (added in the patch)\\nat DEBUG1 level, let's users deal with it.\\n\\nIMO, the new GUC \\\"log_recovery\\\" is way better as it can be extended to\\ninclude some recovery stats info as well.\\n\\nThoughts?\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Wed, 8 Dec 2021 16:57:49 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 08, 2021 at 04:57:49PM +0530, Bharath Rupireddy wrote:\\n> Done in the attached v5 patch.\\n\\nThanks for the new version. FWIW, as the information logged when\\nrecovering a WAL segment from the local pg_wal could be rather\\nconfusing at bootstrap phase, I would suggest to simplify things as of\\nthe attached and call it a day.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 9 Dec 2021 11:26:38 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 9, 2021 at 7:56 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Wed, Dec 08, 2021 at 04:57:49PM +0530, Bharath Rupireddy wrote:\\n> > Done in the attached v5 patch.\\n>\\n> Thanks for the new version. FWIW, as the information logged when\\n> recovering a WAL segment from the local pg_wal could be rather\\n> confusing at bootstrap phase, I would suggest to simplify things as of\\n> the attached and call it a day.\\n\\nThanks for the patch. It looks good to me. I have a suggestion: It's\\nbetter if we specify why we are not emitting any message if the WAL\\nsegment is found in pg_wal directory, something like \\\"we are\\nintentionally not emitting any messages in case if the required WAL\\nsegment is found in the pg_wal directory to avoid noise in the server\\nlogs\\\".\\n\\n+ case XLOG_FROM_PG_WAL:\\n+ case XLOG_FROM_ANY:\\n+ break;\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Dec 2021 11:28:41 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 9, 2021 at 11:28 AM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n>\\n> On Thu, Dec 9, 2021 at 7:56 AM Michael Paquier <michael@paquier.xyz> wrote:\\n> >\\n> > On Wed, Dec 08, 2021 at 04:57:49PM +0530, Bharath Rupireddy wrote:\\n> > > Done in the attached v5 patch.\\n> >\\n> > Thanks for the new version. FWIW, as the information logged when\\n> > recovering a WAL segment from the local pg_wal could be rather\\n> > confusing at bootstrap phase, I would suggest to simplify things as of\\n> > the attached and call it a day.\\n>\\n> Thanks for the patch. It looks good to me. I have a suggestion: It's\\n> better if we specify why we are not emitting any message if the WAL\\n> segment is found in pg_wal directory, something like \\\"we are\\n> intentionally not emitting any messages in case if the required WAL\\n> segment is found in the pg_wal directory to avoid noise in the server\\n> logs\\\".\\n>\\n> + case XLOG_FROM_PG_WAL:\\n> + case XLOG_FROM_ANY:\\n> + break;\\n\\nI just want to call this out: an insertion of 10 million rows in the\\nprimary generates a bunch of messages in the standby [1] within 40 sec\\n(size of the standby server log grows by 5K).\\n\\n[1]\\n2021-12-08 08:11:53.906 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000004\\\" from stream\\n2021-12-08 08:11:55.864 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000005\\\" from stream\\n2021-12-08 08:11:56.713 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000006\\\" from stream\\n2021-12-08 08:11:57.495 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000007\\\" from stream\\n2021-12-08 08:11:58.301 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000008\\\" from stream\\n2021-12-08 08:11:59.062 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000009\\\" from stream\\n2021-12-08 08:11:59.853 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000000A\\\" from stream\\n2021-12-08 08:12:00.573 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000000B\\\" from stream\\n2021-12-08 08:12:01.346 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000000C\\\" from stream\\n2021-12-08 08:12:02.048 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000000D\\\" from stream\\n2021-12-08 08:12:02.798 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000000E\\\" from stream\\n2021-12-08 08:12:03.544 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000000F\\\" from stream\\n2021-12-08 08:12:04.310 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000010\\\" from stream\\n2021-12-08 08:12:05.096 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000011\\\" from stream\\n2021-12-08 08:12:06.003 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000012\\\" from stream\\n2021-12-08 08:12:06.762 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000013\\\" from stream\\n2021-12-08 08:12:07.595 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000014\\\" from stream\\n2021-12-08 08:12:08.431 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000015\\\" from stream\\n2021-12-08 08:12:09.246 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000016\\\" from stream\\n2021-12-08 08:12:10.144 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000017\\\" from stream\\n2021-12-08 08:12:14.944 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000018\\\" from stream\\n2021-12-08 08:12:15.784 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000019\\\" from stream\\n2021-12-08 08:12:16.757 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000001A\\\" from stream\\n2021-12-08 08:12:17.571 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000001B\\\" from stream\\n2021-12-08 08:12:18.375 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000001C\\\" from stream\\n2021-12-08 08:12:19.179 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000001D\\\" from stream\\n2021-12-08 08:12:19.964 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000001E\\\" from stream\\n2021-12-08 08:12:20.839 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000001F\\\" from stream\\n2021-12-08 08:12:21.782 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000020\\\" from stream\\n2021-12-08 08:12:24.422 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000021\\\" from stream\\n2021-12-08 08:12:25.164 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000022\\\" from stream\\n2021-12-08 08:12:25.989 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000023\\\" from stream\\n2021-12-08 08:12:26.874 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000024\\\" from stream\\n2021-12-08 08:12:27.592 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000025\\\" from stream\\n2021-12-08 08:12:28.471 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000026\\\" from stream\\n2021-12-08 08:12:29.158 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000027\\\" from stream\\n2021-12-08 08:12:30.017 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000028\\\" from stream\\n2021-12-08 08:12:30.778 UTC [1219898] LOG: recovering WAL segment\\n\\\"000000010000000000000029\\\" from stream\\n2021-12-08 08:12:31.859 UTC [1219898] LOG: recovering WAL segment\\n\\\"00000001000000000000002A\\\" from stream\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Dec 2021 11:42:30 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On 2021-Dec-09, Bharath Rupireddy wrote:\\n\\n> I just want to call this out: an insertion of 10 million rows in the\\n> primary generates a bunch of messages in the standby [1] within 40 sec\\n> (size of the standby server log grows by 5K).\\n\\nHmm, that does sound excessive to me in terms of log bloat increase.\\nRemember the discussion about turning log_checkpoints on by default?\\n\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\\"Las mujeres son como hondas: mientras más resistencia tienen,\\n más lejos puedes llegar con ellas\\\" (Jonas Nightingale, Leap of Faith)\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Dec 2021 09:30:34 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 9, 2021 at 6:00 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>\\n> On 2021-Dec-09, Bharath Rupireddy wrote:\\n>\\n> > I just want to call this out: an insertion of 10 million rows in the\\n> > primary generates a bunch of messages in the standby [1] within 40 sec\\n> > (size of the standby server log grows by 5K).\\n>\\n> Hmm, that does sound excessive to me in terms of log bloat increase.\\n> Remember the discussion about turning log_checkpoints on by default?\\n\\nThe amount of LOG messages generated when the log_checkpoints GUC is\\nset to on, are quite less, hardly 4 messages per-checkpoint (5min). I\\nthink enabling log_checkpoints is still acceptable as most of the\\nhackers agreed in another thread [1] that the advantages with it are\\nmore and it doesn't seem to be bloating the server logs (in a day at\\nmax 1152 messages).\\n\\nI'm not sure if it is worth having a GUC log_recovery if enabled the\\nrecovery messages can be emitted at LOG level otherwise DEBUG1 level.\\nlog_recovery GUC can also be used to collect and emit some recovery\\nstats like log_checkpoints.\\n\\n[1] - https://www.postgresql.org/message-id/CA%2BTgmoaDFpFmNQuaWj6%2B78CPVBrF_WPT1wFHBTvio%3DtRmxzUcQ%40mail.gmail.com\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Dec 2021 21:20:54 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On 2021-Dec-09, Bharath Rupireddy wrote:\\n\\n> On Thu, Dec 9, 2021 at 6:00 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> >\\n> > On 2021-Dec-09, Bharath Rupireddy wrote:\\n> >\\n> > > I just want to call this out: an insertion of 10 million rows in the\\n> > > primary generates a bunch of messages in the standby [1] within 40 sec\\n> > > (size of the standby server log grows by 5K).\\n> >\\n> > Hmm, that does sound excessive to me in terms of log bloat increase.\\n> > Remember the discussion about turning log_checkpoints on by default?\\n> \\n> The amount of LOG messages generated when the log_checkpoints GUC is\\n> set to on, are quite less, hardly 4 messages per-checkpoint (5min). I\\n> think enabling log_checkpoints is still acceptable as most of the\\n> hackers agreed in another thread [1] that the advantages with it are\\n> more and it doesn't seem to be bloating the server logs (in a day at\\n> max 1152 messages).\\n\\nMy point is that it was argued, in that thread, that setting\\nlog_checkpoints to on by default would cause too much additional log\\nvolume. That argument was shot down, but in this thread we're arguing\\nthat we want five kilobytes of messages in forty seconds. That sounds\\nmuch less acceptable to me, so making it default behavior or\\nunconditional behavior is not going to fly very high.\\n\\n> I'm not sure if it is worth having a GUC log_recovery if enabled the\\n> recovery messages can be emitted at LOG level otherwise DEBUG1 level.\\n\\nMaybe some tunable like\\nlog_wal_traffic = { none, medium, high }\\nwhere \\\"none\\\" is current behavior of no noise, \\\"medium\\\" gets (say) once\\nevery 256 segments (e.g., when going from XFF to (X+1)00), \\\"high\\\" gets\\nyou one message per segment.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Dec 2021 12:58:21 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 9, 2021 at 9:28 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>\\n> On 2021-Dec-09, Bharath Rupireddy wrote:\\n>\\n> > On Thu, Dec 9, 2021 at 6:00 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> > >\\n> > > On 2021-Dec-09, Bharath Rupireddy wrote:\\n> > >\\n> > > > I just want to call this out: an insertion of 10 million rows in the\\n> > > > primary generates a bunch of messages in the standby [1] within 40 sec\\n> > > > (size of the standby server log grows by 5K).\\n> > >\\n> > > Hmm, that does sound excessive to me in terms of log bloat increase.\\n> > > Remember the discussion about turning log_checkpoints on by default?\\n> >\\n> > The amount of LOG messages generated when the log_checkpoints GUC is\\n> > set to on, are quite less, hardly 4 messages per-checkpoint (5min). I\\n> > think enabling log_checkpoints is still acceptable as most of the\\n> > hackers agreed in another thread [1] that the advantages with it are\\n> > more and it doesn't seem to be bloating the server logs (in a day at\\n> > max 1152 messages).\\n>\\n> My point is that it was argued, in that thread, that setting\\n> log_checkpoints to on by default would cause too much additional log\\n> volume. That argument was shot down, but in this thread we're arguing\\n> that we want five kilobytes of messages in forty seconds. That sounds\\n> much less acceptable to me, so making it default behavior or\\n> unconditional behavior is not going to fly very high.\\n>\\n> > I'm not sure if it is worth having a GUC log_recovery if enabled the\\n> > recovery messages can be emitted at LOG level otherwise DEBUG1 level.\\n>\\n> Maybe some tunable like\\n> log_wal_traffic = { none, medium, high }\\n> where \\\"none\\\" is current behavior of no noise, \\\"medium\\\" gets (say) once\\n> every 256 segments (e.g., when going from XFF to (X+1)00), \\\"high\\\" gets\\n> you one message per segment.\\n\\nYeah, that can be an option.\\n\\nAnother idea could be to use the infrastructure laid out by the commit\\n9ce346e [1]. With ereport_startup_progress, we can emit the LOGs(of\\ncurrent recovering WAL file) for every log_startup_progress_interval\\nseconds/milliseconds.\\n\\nOne problem is that ereport_startup_progress doesn't work on\\nStandbyMode, maybe we can remove this restriction unless we have a\\nmajor reason for not allowing it on the standby.\\n /* Prepare to report progress of the redo phase. */\\n if (!StandbyMode)\\n begin_startup_progress_phase();\\n\\nThoughts?\\n\\n[1]\\ncommit 9ce346eabf350a130bba46be3f8c50ba28506969\\nAuthor: Robert Haas <rhaas@postgresql.org>\\nDate: Mon Oct 25 11:51:57 2021 -0400\\n\\n Report progress of startup operations that take a long time.\\n\\n Discussion:\\nhttps://postgr.es/m/CA+TgmoaHQrgDFOBwgY16XCoMtXxsrVGFB2jNCvb7-ubuEe1MGg@mail.gmail.com\\n Discussion:\\nhttps://postgr.es/m/CAMm1aWaHF7VE69572_OLQ+MgpT5RUiUDgF1x5RrtkJBLdpRj3Q@mail.gmail.com\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Sun, 12 Dec 2021 18:08:05 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Sun, Dec 12, 2021 at 06:08:05PM +0530, Bharath Rupireddy wrote:\\n> Another idea could be to use the infrastructure laid out by the commit\\n> 9ce346e [1]. With ereport_startup_progress, we can emit the LOGs(of\\n> current recovering WAL file) for every log_startup_progress_interval\\n> seconds/milliseconds.\\n> \\n> One problem is that ereport_startup_progress doesn't work on\\n> StandbyMode, maybe we can remove this restriction unless we have a\\n> major reason for not allowing it on the standby.\\n> /* Prepare to report progress of the redo phase. */\\n> if (!StandbyMode)\\n> begin_startup_progress_phase();\\n\\nThe relevant conversation starts here.\\nhttps://www.postgresql.org/message-id/20210814214700.GO10479%40telsasoft.com\\n\\nThere was a lot of confusion in that thread, though.\\n\\nThe understanding was that it didn't make sense for that feature to\\ncontinuously log messages on a standby (every 10sec by default). That seems\\nlike too much - the issue of a checkpointed logged every 5min was enough of a\\nhurdle.\\n\\nIf you're talking about a new feature that uses the infrastructre from 9ce3,\\nbut is controlled by a separate GUC like log_wal_traffic, that could be okay.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Fri, 24 Dec 2021 07:49:46 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 24, 2021 at 7:19 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>\\n> On Sun, Dec 12, 2021 at 06:08:05PM +0530, Bharath Rupireddy wrote:\\n> > Another idea could be to use the infrastructure laid out by the commit\\n> > 9ce346e [1]. With ereport_startup_progress, we can emit the LOGs(of\\n> > current recovering WAL file) for every log_startup_progress_interval\\n> > seconds/milliseconds.\\n> >\\n> > One problem is that ereport_startup_progress doesn't work on\\n> > StandbyMode, maybe we can remove this restriction unless we have a\\n> > major reason for not allowing it on the standby.\\n> > /* Prepare to report progress of the redo phase. */\\n> > if (!StandbyMode)\\n> > begin_startup_progress_phase();\\n>\\n> The relevant conversation starts here.\\n> https://www.postgresql.org/message-id/20210814214700.GO10479%40telsasoft.com\\n>\\n> There was a lot of confusion in that thread, though.\\n>\\n> The understanding was that it didn't make sense for that feature to\\n> continuously log messages on a standby (every 10sec by default). That seems\\n> like too much - the issue of a checkpointed logged every 5min was enough of a\\n> hurdle.\\n>\\n> If you're talking about a new feature that uses the infrastructre from 9ce3,\\n> but is controlled by a separate GUC like log_wal_traffic, that could be okay.\\n\\nDo you see any problems using the same GUC\\nlog_startup_progress_interval and ereport_startup_progress API\\nintroduced by 9ce346e? I prefer this instead of a new GUC. Thoughts?\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Fri, 24 Dec 2021 20:04:08 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 24, 2021 at 08:04:08PM +0530, Bharath Rupireddy wrote:\\n> On Fri, Dec 24, 2021 at 7:19 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > On Sun, Dec 12, 2021 at 06:08:05PM +0530, Bharath Rupireddy wrote:\\n> > > Another idea could be to use the infrastructure laid out by the commit\\n> > > 9ce346e [1]. With ereport_startup_progress, we can emit the LOGs(of\\n> > > current recovering WAL file) for every log_startup_progress_interval\\n> > > seconds/milliseconds.\\n> > >\\n> > > One problem is that ereport_startup_progress doesn't work on\\n> > > StandbyMode, maybe we can remove this restriction unless we have a\\n> > > major reason for not allowing it on the standby.\\n> > > /* Prepare to report progress of the redo phase. */\\n> > > if (!StandbyMode)\\n> > > begin_startup_progress_phase();\\n> >\\n> > The relevant conversation starts here.\\n> > https://www.postgresql.org/message-id/20210814214700.GO10479%40telsasoft.com\\n> >\\n> > There was a lot of confusion in that thread, though.\\n> >\\n> > The understanding was that it didn't make sense for that feature to\\n> > continuously log messages on a standby (every 10sec by default). That seems\\n> > like too much - the issue of a checkpointed logged every 5min was enough of a\\n> > hurdle.\\n> >\\n> > If you're talking about a new feature that uses the infrastructre from 9ce3,\\n> > but is controlled by a separate GUC like log_wal_traffic, that could be okay.\\n> \\n> Do you see any problems using the same GUC\\n> log_startup_progress_interval and ereport_startup_progress API\\n> introduced by 9ce346e? I prefer this instead of a new GUC. Thoughts?\\n\\nI referenced the thread where we avoided writing progress logs during normal\\noperation of a standby, since the logs would be both useless and excessive.\\n\\nIf your patch were to use the same infrastructure, you'd still want to avoid\\ndisturbing that behavior and writing useless logs.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sat, 25 Dec 2021 18:20:54 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 9, 2021 at 9:28 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>\\n> Maybe some tunable like\\n> log_wal_traffic = { none, medium, high }\\n> where \\\"none\\\" is current behavior of no noise, \\\"medium\\\" gets (say) once\\n> every 256 segments (e.g., when going from XFF to (X+1)00), \\\"high\\\" gets\\n> you one message per segment.\\n\\nOn Fri, Dec 24, 2021 at 7:19 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>\\n> If you're talking about a new feature that uses the infrastructre from 9ce3,\\n> but is controlled by a separate GUC like log_wal_traffic, that could be okay.\\n\\nThanks for the suggestion. I've added a new GUC log_wal_traffic as\\nsuggested above. PSA v7 patch.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Wed, 29 Dec 2021 06:50:53 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: add recovery, backup, archive, streaming etc. activity messages\\n to server logs along with ps display\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 29, 2021 at 6:50 AM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n>\\n> On Thu, Dec 9, 2021 at 9:28 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> >\\n> > Maybe some tunable like\\n> > log_wal_traffic = { none, medium, high }\\n> > where \\\"none\\\" is current behavior of no noise, \\\"medium\\\" gets (say) once\\n> > every 256 segments (e.g., when going from XFF to (X+1)00), \\\"high\\\" gets\\n> > you one message per segment.\\n>\\n> On Fri, Dec 24, 2021 at 7:19 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> >\\n> > If you're talking about a new feature that uses the infrastructre from 9ce3,\\n> > but is controlled by a separate GUC like log_wal_traffic, that could be okay.\\n>\\n> Thanks for the suggestion. I've added a new GUC log_wal_traffic as\\n> suggested above. PSA v7 patch.\\n\\nHere's the rebased v8 patch, please review it.\\n\\nhttps://commitfest.postgresql.org/37/3443/\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Sun, 27 Feb 2022 12:50:25 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Add WAL recovery messages with log_wal_traffic GUC (was: add\\n recovery, backup, archive, streaming etc. activity messages to server logs\\n along with ps display)\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 12:50 PM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n>\\n> On Wed, Dec 29, 2021 at 6:50 AM Bharath Rupireddy\\n> <bharath.rupireddyforpostgres@gmail.com> wrote:\\n> >\\n> > On Thu, Dec 9, 2021 at 9:28 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> > >\\n> > > Maybe some tunable like\\n> > > log_wal_traffic = { none, medium, high }\\n> > > where \\\"none\\\" is current behavior of no noise, \\\"medium\\\" gets (say) once\\n> > > every 256 segments (e.g., when going from XFF to (X+1)00), \\\"high\\\" gets\\n> > > you one message per segment.\\n> >\\n> > On Fri, Dec 24, 2021 at 7:19 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > >\\n> > > If you're talking about a new feature that uses the infrastructre from 9ce3,\\n> > > but is controlled by a separate GUC like log_wal_traffic, that could be okay.\\n> >\\n> > Thanks for the suggestion. I've added a new GUC log_wal_traffic as\\n> > suggested above. PSA v7 patch.\\n>\\n> Here's the rebased v8 patch, please review it.\\n>\\n> https://commitfest.postgresql.org/37/3443/\\n\\nHere's the rebased v9 patch.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Fri, 29 Apr 2022 14:40:56 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\\n recovery, backup, archive, streaming etc. activity messages to server logs\\n along with ps display)\"\n },\n {\n \"msg_contents\": \"Did we ever consider the idea of using a new pg_stat_wal_activity_progress\\nview or something like that, using the backend_progress.c functionality?\\nI don't see it mentioned in the thread.\\n\\nIt's not an *exact* fit, since this is not about one \\\"command\\\" being\\nexecuted by a \\\"backend\\\", but it seems a near enough fit, and it would\\nunobtrusive enough that we don't need to worry about the progress-update\\nrate or have a GUC to determine how frequently to update (with a gauge\\nthat's heavily workload-dependant and is likely to make little sense to\\nsome users -- consider differently-sized WAL segments, for one thing).\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Fri, 29 Apr 2022 12:41:15 +0200\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\\n recovery, backup, archive, streaming etc. activity messages to server logs\\n along with ps display)\"\n },\n {\n \"msg_contents\": \"On Fri, Apr 29, 2022 at 4:11 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>\\n> Did we ever consider the idea of using a new pg_stat_wal_activity_progress\\n> view or something like that, using the backend_progress.c functionality?\\n> I don't see it mentioned in the thread.\\n\\nIMO, progress reporting works well on a running server and at the\\nmoment. The WAL recovery/replay can happen even before the server\\nopens up for connections and the progress report view can't be used\\nfor later analysis like how much time the restoring WAL files from\\narchive location took and also the WAL file names can't be reported in\\nprogress reporting mechanism (only integers columns, of course if\\nrequired we can add text columns to pg_stat_get_progress_info). Having\\nthe recovery info in server logs might help.\\n\\n> It's not an *exact* fit, since this is not about one \\\"command\\\" being\\n> executed by a \\\"backend\\\", but it seems a near enough fit, and it would\\n> unobtrusive enough that we don't need to worry about the progress-update\\n> rate or have a GUC to determine how frequently to update (with a gauge\\n> that's heavily workload-dependant and is likely to make little sense to\\n> some users -- consider differently-sized WAL segments, for one thing).\\n\\nI think reporting a long-running file processing operation (removing\\nor syncing) within postgres is a generic problem (for snapshot,\\nmapping, temporary (pgsql_tmp), temp relation files, old WAL file\\nprocessing, WAL file processing during recovery etc.) and needs to be\\nsolved in two ways: 1) logging progress into server logs (which helps\\nfor analysis and report when the server isn't available for\\nconnections, crash recovery), a generic GUC\\nlog_file_processing_traffic = {none, medium, high} might help here\\n(also proposed in [1]) and 2) pg_stat_file_processing_progress\\n(extending progress reporting pg_stat_get_progress_info to have few\\ntext columns for current file name and directory path).\\n\\nThoughts?\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Thu, 5 May 2022 13:35:12 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\\n recovery, backup, archive, streaming etc. activity messages to server logs\\n along with ps display)\"\n },\n {\n \"msg_contents\": \"On 2022-May-05, Bharath Rupireddy wrote:\\n\\n> On Fri, Apr 29, 2022 at 4:11 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> >\\n> > Did we ever consider the idea of using a new pg_stat_wal_activity_progress\\n> > view or something like that, using the backend_progress.c functionality?\\n> > I don't see it mentioned in the thread.\\n> \\n> IMO, progress reporting works well on a running server and at the\\n> moment. The WAL recovery/replay can happen even before the server\\n> opens up for connections\\n\\nIt's definitely true that you wouldn't be able to use it for when the\\nserver is not accepting connections.\\n\\n> and the progress report view can't be used\\n> for later analysis like how much time the restoring WAL files from\\n> archive location took\\n\\nThis is true too -- progress report doesn't store historical data, only\\ncurrent status.\\n\\n> and also the WAL file names can't be reported in progress reporting\\n> mechanism\\n\\nAlso true.\\n\\n> (only integers columns, of course if required we can add text columns\\n> to pg_stat_get_progress_info).\\n\\nYeah, I don't think adding text columns is terribly easy, because the\\nwhole point of the progress reporting infrastructure is that it can be\\nupdated very cheaply as atomic operations, and if you want to transmit\\ntext columns, that's no longer possible.\\n\\n> Having the recovery info in server logs might help.\\n\\nI suppose it might.\\n\\n> I think reporting a long-running file processing operation (removing\\n> or syncing) within postgres is a generic problem (for snapshot,\\n> mapping, temporary (pgsql_tmp), temp relation files, old WAL file\\n> processing, WAL file processing during recovery etc.) and needs to be\\n> solved\\n\\nI agree up to here.\\n\\n> in two ways: 1) logging progress into server logs (which helps\\n> for analysis and report when the server isn't available for\\n> connections, crash recovery), a generic GUC\\n> log_file_processing_traffic = {none, medium, high} might help here\\n> (also proposed in [1]) and 2) pg_stat_file_processing_progress\\n> (extending progress reporting pg_stat_get_progress_info to have few\\n> text columns for current file name and directory path).\\n\\nI think using the server log to store telemetry data is not a great fit.\\nIt bloats the log files and can be so slow as to block other operations\\nin the server. Server logs should normally be considered critical info\\nthat's not okay to lose; telemetry tends to be of secondary importance\\nand in a pinch you can drop a few messages without hurting too much.\\n\\nWe've done moderately okay so far with having some system views where\\nsome telemetry readings can be obtained, but there several drawbacks to\\nthat approach that we should at some point solve. My opinion on this is\\nthat we need to bite the bullet and develop separate infrastructure for\\nreporting server metrics.\\n\\nThat said, I'm not opposed to having a patch somewhat as posted. I just\\nthink that we should look into a new mechanism going forward.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 5 May 2022 10:37:20 +0200\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\\n recovery, backup, archive, streaming etc. activity messages to server logs\\n along with ps display)\"\n },\n {\n \"msg_contents\": \"On Thu, May 5, 2022 at 2:07 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>\\n> On 2022-May-05, Bharath Rupireddy wrote:\\n>\\n> > On Fri, Apr 29, 2022 at 4:11 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> > >\\n> > > Did we ever consider the idea of using a new pg_stat_wal_activity_progress\\n> > > view or something like that, using the backend_progress.c functionality?\\n> > > I don't see it mentioned in the thread.\\n> >\\n> > IMO, progress reporting works well on a running server and at the\\n> > moment. The WAL recovery/replay can happen even before the server\\n> > opens up for connections\\n>\\n> It's definitely true that you wouldn't be able to use it for when the\\n> server is not accepting connections.\\n>\\n> > and the progress report view can't be used\\n> > for later analysis like how much time the restoring WAL files from\\n> > archive location took\\n>\\n> This is true too -- progress report doesn't store historical data, only\\n> current status.\\n>\\n> > and also the WAL file names can't be reported in progress reporting\\n> > mechanism\\n>\\n> Also true.\\n>\\n> > (only integers columns, of course if required we can add text columns\\n> > to pg_stat_get_progress_info).\\n>\\n> Yeah, I don't think adding text columns is terribly easy, because the\\n> whole point of the progress reporting infrastructure is that it can be\\n> updated very cheaply as atomic operations, and if you want to transmit\\n> text columns, that's no longer possible.\\n>\\n> > Having the recovery info in server logs might help.\\n>\\n> I suppose it might.\\n>\\n> > I think reporting a long-running file processing operation (removing\\n> > or syncing) within postgres is a generic problem (for snapshot,\\n> > mapping, temporary (pgsql_tmp), temp relation files, old WAL file\\n> > processing, WAL file processing during recovery etc.) and needs to be\\n> > solved\\n>\\n> I agree up to here.\\n>\\n> > in two ways: 1) logging progress into server logs (which helps\\n> > for analysis and report when the server isn't available for\\n> > connections, crash recovery), a generic GUC\\n> > log_file_processing_traffic = {none, medium, high} might help here\\n> > (also proposed in [1]) and 2) pg_stat_file_processing_progress\\n> > (extending progress reporting pg_stat_get_progress_info to have few\\n> > text columns for current file name and directory path).\\n>\\n> I think using the server log to store telemetry data is not a great fit.\\n> It bloats the log files and can be so slow as to block other operations\\n> in the server. Server logs should normally be considered critical info\\n> that's not okay to lose; telemetry tends to be of secondary importance\\n> and in a pinch you can drop a few messages without hurting too much.\\n>\\n> We've done moderately okay so far with having some system views where\\n> some telemetry readings can be obtained, but there several drawbacks to\\n> that approach that we should at some point solve. My opinion on this is\\n> that we need to bite the bullet and develop separate infrastructure for\\n> reporting server metrics.\\n>\\n> I just\\n> think that we should look into a new mechanism going forward.\\n\\nI completely agree that we must have new ways to report important\\nserver metrics in an easily consumable fashion - it could be something\\nlike server computing metrics (time taken, the file being processed\\netc.) important/time-taking operations such as execution of archive\\ncommand, restore command, recovery, checkpoint, old WAL files\\nremoval/recycling, processing of various files - snapshot, mapping,\\npgsql_tmp files, temp relation files etc. and writing these metrics to\\na file or a stat table on the database itself (this table can have an\\nautomatic mechanism of clean up or limit on number of rows or size of\\nthe table so that the clients can read those metrics and use it in\\nwhichever way they want. To start with, a simple structure of the\\nmetrics can be {OPERATION text, START_TIME timestamp, END_TIME\\ntimestamp, few text, double and/or integer flexible columns that each\\noperation can use to store metrics}.\\n\\nI can start a discussion in a separate thread to seek more thoughts on\\nthe server metrics part.\\n\\n> That said, I'm not opposed to having a patch somewhat as posted.\\n\\nThanks. I'm thinking if we should have a generic GUC\\nlog_file_processing_traffic = {none, medium, high} that might help\\nreporting other time taking file processing operations such as [1].\\n\\n[1] https://www.postgresql.org/message-id/CALj2ACW-ELOF5JT2zPavs95wbZ0BrLPrqvSZ7Ac%2BpjxCkmXtEQ%40mail.gmail.com\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Fri, 13 May 2022 18:32:31 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\\n recovery, backup, archive, streaming etc. activity messages to server logs\\n along with ps display)\"\n },\n {\n \"msg_contents\": \"On Fri, Apr 29, 2022 at 5:11 AM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n> Here's the rebased v9 patch.\\n\\nThis seems like it has enormous overlap with the existing\\nfunctionality that we have from log_startup_progress_interval.\\n\\nI think that facility is also better-designed than this one. It prints\\nout a message based on elapsed time, whereas this patch prints out a\\nmessage based progress through the WAL. That means that if WAL replay\\nisn't actually advancing for some reason, you just won't get any log\\nmessages and you don't know whether it's advancing slowly or not at\\nall or the server is just hung. With that facility you can distinguish\\nthose cases.\\n\\nAlso, if for some reason we do think that amount of WAL replayed is\\nthe right metric, rather than time, why would we only allow high=1\\nsegment and low=128 segments, rather than say any number of MB or GB\\nthat the user would like to configure?\\n\\nI suggest that if log_startup_progress_interval doesn't meet your\\nneeds here, we should try to understand why not and maybe enhance it,\\ninstead of adding a separate facility.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 13 May 2022 09:10:52 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\\n recovery, backup, archive, streaming etc. activity messages to server logs\\n along with ps display)\"\n },\n {\n \"msg_contents\": \"On Fri, May 13, 2022 at 6:41 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> On Fri, Apr 29, 2022 at 5:11 AM Bharath Rupireddy\\n> <bharath.rupireddyforpostgres@gmail.com> wrote:\\n> > Here's the rebased v9 patch.\\n>\\n> This seems like it has enormous overlap with the existing\\n> functionality that we have from log_startup_progress_interval.\\n>\\n> I think that facility is also better-designed than this one. It prints\\n> out a message based on elapsed time, whereas this patch prints out a\\n> message based progress through the WAL. That means that if WAL replay\\n> isn't actually advancing for some reason, you just won't get any log\\n> messages and you don't know whether it's advancing slowly or not at\\n> all or the server is just hung. With that facility you can distinguish\\n> those cases.\\n>\\n> Also, if for some reason we do think that amount of WAL replayed is\\n> the right metric, rather than time, why would we only allow high=1\\n> segment and low=128 segments, rather than say any number of MB or GB\\n> that the user would like to configure?\\n>\\n> I suggest that if log_startup_progress_interval doesn't meet your\\n> needs here, we should try to understand why not and maybe enhance it,\\n> instead of adding a separate facility.\\n\\nThanks Robert!\\n\\nIn a production environment (of course with a better management of\\nserver logs) one can set log_wal_traffic to \\\"high\\\" and emit the\\nrequired info to answer some of the customer questions like - \\\"How far\\nthe server is in recovery? How much time recovery of each WAL file\\napproximately took? How much time will it take to recover all the WAL\\nfiles? What's the rate of recovery - time per WAL file? etc.\\\"\\n\\nWhereas ereport_startup_progress facility will help to emit log\\nmessages only if \\\"some\\\" operation takes longer than set\\nlog_startup_progress_interval time which may not serve the above\\npurpose.\\n\\nActually, IMO a generic GUC log_file_processing_traffic = {none,\\nmedium, high} to help server emit logs for all the critical file\\nprocessing operations - WAL file recovery (as proposed here in this\\nthread), temp file processing during server startup or restarts\\n(log_startup_progress_interval can't be used here as postmaster\\ndoesn't register for any timeouts) [1], snapshot and mapping file\\nprocessing during checkpoint, temp relation files, removing old WAL\\nfiles and so on.\\n\\nThoughts?\\n\\n[1] https://www.postgresql.org/message-id/CALj2ACW-ELOF5JT2zPavs95wbZ0BrLPrqvSZ7Ac%2BpjxCkmXtEQ%40mail.gmail.com\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Tue, 17 May 2022 11:01:57 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\\n recovery, backup, archive, streaming etc. activity messages to server logs\\n along with ps display)\"\n },\n {\n \"msg_contents\": \"On Tue, May 17, 2022 at 1:32 AM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n> In a production environment (of course with a better management of\\n> server logs) one can set log_wal_traffic to \\\"high\\\" and emit the\\n> required info to answer some of the customer questions like - \\\"How far\\n> the server is in recovery? How much time recovery of each WAL file\\n> approximately took? How much time will it take to recover all the WAL\\n> files? What's the rate of recovery - time per WAL file? etc.\\\"\\n>\\n> Whereas ereport_startup_progress facility will help to emit log\\n> messages only if \\\"some\\\" operation takes longer than set\\n> log_startup_progress_interval time which may not serve the above\\n> purpose.\\n\\nI think you are misunderstanding what an \\\"operation\\\" is here. If you\\ntake the time to test out that facility, I think you will find that it\\nis quite capable of answering all of those questions.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 17 May 2022 12:24:24 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\\n recovery, backup, archive, streaming etc. activity messages to server logs\\n along with ps display)\"\n },\n {\n \"msg_contents\": \"On Fri, May 13, 2022 at 09:10:52AM -0400, Robert Haas wrote:\\n> I suggest that if log_startup_progress_interval doesn't meet your\\n> needs here, we should try to understand why not and maybe enhance it,\\n> instead of adding a separate facility.\\n\\n+1. AFAICT it should be possible to make the log_startup_progress_interval\\nmachinery extensible enough to be reused for several other tasks that can\\ntake a while but have little visibility.\\n\\nSince this thread has been inactive for over 2 months, I'm marking the\\ncommitfest entry as Waiting on Author.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 18 Jul 2022 16:01:08 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\\n recovery, backup, archive, streaming etc. activity messages to server logs\\n along with ps display)\"\n },\n {\n \"msg_contents\": \"On Fri, May 13, 2022 at 6:41 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> On Fri, Apr 29, 2022 at 5:11 AM Bharath Rupireddy\\n> <bharath.rupireddyforpostgres@gmail.com> wrote:\\n> > Here's the rebased v9 patch.\\n>\\n> This seems like it has enormous overlap with the existing\\n> functionality that we have from log_startup_progress_interval.\\n>\\n> I think that facility is also better-designed than this one. It prints\\n> out a message based on elapsed time, whereas this patch prints out a\\n> message based progress through the WAL. That means that if WAL replay\\n> isn't actually advancing for some reason, you just won't get any log\\n> messages and you don't know whether it's advancing slowly or not at\\n> all or the server is just hung. With that facility you can distinguish\\n> those cases.\\n>\\n> Also, if for some reason we do think that amount of WAL replayed is\\n> the right metric, rather than time, why would we only allow high=1\\n> segment and low=128 segments, rather than say any number of MB or GB\\n> that the user would like to configure?\\n>\\n> I suggest that if log_startup_progress_interval doesn't meet your\\n> needs here, we should try to understand why not and maybe enhance it,\\n> instead of adding a separate facility.\\n\\nAfter thinking for a while, I agree with Robert and others that we\\ncould leverage the existing log_startup_progress_interval mechanism\\nfor reporting which WAL file currently is being replayed. I added\\ncurrent TLI (helps to construct the WAL file name from current LSN)\\nand current WAL file source (helps to know where the WAL files was\\nfetched from) to the existing \\\"redo in progress, elapsed time:...\\\"\\nmessage. This very well serves the purpose of identifying the issues\\nsuch as the restore command taking a lot of time (>\\nlog_startup_progress_interval for instance), WAL replay rate on\\nstandby or primary for long recoveries and so on.\\n\\nHowever, ereport_startup_progress isn't enabled on standby to not let\\nit bloat the server logs. I believe the \\\"redo in progress, elapsed\\ntime:...\\\" message can provide some important info/metric for standby\\ntoo and there's no way for the users to enable it on standbys today.\\nFor instance, users can know how well the standby fares in replaying,\\nthey can figure this out, by looking at two or more such messages. If\\nenabled, with default value of 10 sec for\\nlog_startup_progress_interval, the standby can emit 8640 messages per\\nday which is too much. I'm not sure if we are okay to change the\\ndefault value of log_startup_progress_interval to say 1min or 5min so\\nthat 1440 messages are emitted. In production environments, typically\\nusers may or may not be interested if recovery takes just 10sec, but\\nthey really are interested if it takes in the order of minutes.\\nBasically, I would like to enable \\\"redo in progress, elapsed time:...\\\"\\nmessage for standbys too.\\n\\nThoughts?\\n\\nPSA v10 patch with enhanced \\\"redo in progress, elapsed time:...\\\"\\nmessage. Note that it's not a final patch though.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Tue, 26 Jul 2022 15:36:57 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add WAL recovery messages with log_wal_traffic GUC (was: add\\n recovery, backup, archive, streaming etc. activity messages to server logs\\n along with ps display)\"\n }\n]","isConversational":false}}},{"rowIdx":53072,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"postgres=# SELECT log_time , database, user_name, error_severity sev, left(message,99) FROM postgres_log_2021_11_10_0800 WHERE log_time BETWEEN '2021-11-10 08:57' AND '2021-11-10 08:58' AND database IS NULL;\n log_time | database | user_name | sev | left \n----------------------------+----------+-----------+-----+-----------------------------------------------------------------------------------------------------\n 2021-11-10 08:57:17.041-04 | | | LOG | checkpoint starting: time\n 2021-11-10 08:57:55.494-04 | | | LOG | server process (PID 31345) was terminated by signal 6: Aborted\n 2021-11-10 08:57:55.494-04 | | | LOG | terminating any other active server processes\n 2021-11-10 08:57:55.524-04 | | | LOG | all server processes terminated; reinitializing\n 2021-11-10 08:57:55.719-04 | | | LOG | database system was interrupted; last known up at 2021-11-10 08:56:38 AST\n 2021-11-10 08:57:56.16-04 | | | LOG | database system was not properly shut down; automatic recovery in progress\n 2021-11-10 08:57:56.24-04 | | | LOG | redo starts at 1B4E/F019B9F8\n 2021-11-10 08:57:56.271-04 | | | LOG | invalid record length at 1B4E/F0245B58: wanted 24, got 0\n 2021-11-10 08:57:56.272-04 | | | LOG | redo done at 1B4E/F0245B10 system usage: CPU: user: 0.00 s, system: 0.00 s, elapsed: 0.03 s\n 2021-11-10 08:57:56.412-04 | | | LOG | checkpoint starting: end-of-recovery immediate\n 2021-11-10 08:57:56.773-04 | | | LOG | checkpoint complete: wrote 116 buffers (0.2%); 0 WAL file(s) added, 0 removed, 0 recycled; write=0.\n 2021-11-10 08:57:56.856-04 | | | LOG | database system is ready to accept connections\n\nI'm surprised that not a single message was logged at PANIC, even though it's\ndefined to mean:\n\n| PANIC \tReports an error that caused all database sessions to abort.\nhttps://www.postgresql.org/docs/current/runtime-config-logging.html#RUNTIME-CONFIG-SEVERITY-LEVELS\n\nIs it intended ?\n\n-- \nJustin\n\n\n","msg_date":"Wed, 10 Nov 2021 14:06:12 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":true,"msg_subject":"SIGABRT causes messages at LOG but not PANIC"},{"msg_contents":"Justin Pryzby <pryzby@telsasoft.com> writes:\n> I'm surprised that not a single message was logged at PANIC, even though it's\n> defined to mean:\n\nThe backend you killed didn't get a chance to log anything.\n\nOr to put it another way: a PANIC ereport is the trigger for a database\nrestart, not a response to some other event that has triggered one.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 10 Nov 2021 15:20:13 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: SIGABRT causes messages at LOG but not PANIC"}],"string":"[\n {\n \"msg_contents\": \"postgres=# SELECT log_time , database, user_name, error_severity sev, left(message,99) FROM postgres_log_2021_11_10_0800 WHERE log_time BETWEEN '2021-11-10 08:57' AND '2021-11-10 08:58' AND database IS NULL;\\n log_time | database | user_name | sev | left \\n----------------------------+----------+-----------+-----+-----------------------------------------------------------------------------------------------------\\n 2021-11-10 08:57:17.041-04 | | | LOG | checkpoint starting: time\\n 2021-11-10 08:57:55.494-04 | | | LOG | server process (PID 31345) was terminated by signal 6: Aborted\\n 2021-11-10 08:57:55.494-04 | | | LOG | terminating any other active server processes\\n 2021-11-10 08:57:55.524-04 | | | LOG | all server processes terminated; reinitializing\\n 2021-11-10 08:57:55.719-04 | | | LOG | database system was interrupted; last known up at 2021-11-10 08:56:38 AST\\n 2021-11-10 08:57:56.16-04 | | | LOG | database system was not properly shut down; automatic recovery in progress\\n 2021-11-10 08:57:56.24-04 | | | LOG | redo starts at 1B4E/F019B9F8\\n 2021-11-10 08:57:56.271-04 | | | LOG | invalid record length at 1B4E/F0245B58: wanted 24, got 0\\n 2021-11-10 08:57:56.272-04 | | | LOG | redo done at 1B4E/F0245B10 system usage: CPU: user: 0.00 s, system: 0.00 s, elapsed: 0.03 s\\n 2021-11-10 08:57:56.412-04 | | | LOG | checkpoint starting: end-of-recovery immediate\\n 2021-11-10 08:57:56.773-04 | | | LOG | checkpoint complete: wrote 116 buffers (0.2%); 0 WAL file(s) added, 0 removed, 0 recycled; write=0.\\n 2021-11-10 08:57:56.856-04 | | | LOG | database system is ready to accept connections\\n\\nI'm surprised that not a single message was logged at PANIC, even though it's\\ndefined to mean:\\n\\n| PANIC \\tReports an error that caused all database sessions to abort.\\nhttps://www.postgresql.org/docs/current/runtime-config-logging.html#RUNTIME-CONFIG-SEVERITY-LEVELS\\n\\nIs it intended ?\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 14:06:12 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"SIGABRT causes messages at LOG but not PANIC\"\n },\n {\n \"msg_contents\": \"Justin Pryzby <pryzby@telsasoft.com> writes:\\n> I'm surprised that not a single message was logged at PANIC, even though it's\\n> defined to mean:\\n\\nThe backend you killed didn't get a chance to log anything.\\n\\nOr to put it another way: a PANIC ereport is the trigger for a database\\nrestart, not a response to some other event that has triggered one.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 10 Nov 2021 15:20:13 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: SIGABRT causes messages at LOG but not PANIC\"\n }\n]","isConversational":false}}},{"rowIdx":53073,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello I have such case: create table test (id int not null, status text);insert into test select i, 'foo' from generate_series(1,1000000) i;update test set status = 'bar' where id <= 10;create index test_id on test (id );create index test_status_partial on test (status) where status = 'bar';analyze test ;explain (analyze) select * from test where id = any('{}'); Gives query plan: QUERY PLAN -------------------------------------------------------------------------------------------------------------------------- Index Scan using test_status_partial on test (cost=0.12..4.14 rows=1 width=8) (actual time=0.024..0.025 rows=0 loops=1) Filter: (id = ANY ('{}'::integer[])) Rows Removed by Filter: 10 Planning Time: 0.327 ms Execution Time: 0.048 ms I don't understand why the planner chose such an unrelated partial index. I expected \"One-Time Filter: false\" here or use of test_id index. I agree, a strange condition, the application should avoid such condition, but why use such an index? Initially was spotted on 13.3 production system (slow query due too much Rows Removed by Filter), then I checked this behaviour on 14.0 and fresh HEAD (db9f287711ac49d9799f93f664d6d101ff8f5891) regards, Sergei","msg_date":"Thu, 11 Nov 2021 14:00:56 +0300","msg_from":"Sergei Kornilov <sk@zsrv.org>","msg_from_op":true,"msg_subject":"unexpected plan with id = any('{}') condition"},{"msg_contents":"Sergei Kornilov <sk@zsrv.org> writes:\n> [ unreadable HTML mess ]\n\nFor the archives, the test case here looks like\n\ncreate table test (id int not null, status text);\ninsert into test select i, 'foo' from generate_series(1,1000000) i;\nupdate test set status = 'bar' where id <= 10;\ncreate index test_id on test (id );\ncreate index test_status_partial on test (status) where status = 'bar';\nanalyze test ;\nexplain select * from test where id = any('{}');\n\nand the question is why you get\n\n QUERY PLAN \n--------------------------------------------------------------------------------\n Index Scan using test_status_partial on test (cost=0.12..4.14 rows=1 width=8)\n Filter: (id = ANY ('{}'::integer[]))\n(2 rows)\n\nwhen that partial index seems unrelated to the query conditions.\n\nThe answer is that predicate_implied_by() returns TRUE, because\nit can see that the WHERE condition is equivalent to constant FALSE,\nand FALSE implies anything. So the partial index is accepted as\nproven by the WHERE condition, and then it looks a trifle cheaper\nto scan than the complete index, which gets an estimate like\n\n Index Scan using test_id on test (cost=0.42..4.44 rows=1 width=8)\n Index Cond: (id = ANY ('{}'::integer[]))\n\nI don't see anything wrong here. Maybe one could wish that the\nplanner reduced the WHERE to constant-FALSE outright, which would\nyield an even cheaper plan; but I'm not convinced we want to expend\nextra cycles looking for such cases.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 11 Nov 2021 12:32:04 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: unexpected plan with id = any('{}') condition"},{"msg_contents":"Hello\n\nThank you for the explanation!\n\n> unreadable HTML mess\n\nouch, sorry. \"Nobody uses plain text mail, we dropped this thing in the interface” said yandex team. (I know that some members of the Yandex team read mailing lists, could you ping your colleagues?)\n\n> but I'm not convinced we want to expend\n> extra cycles looking for such cases.\n\nAgree. This is an application bug and needs to be fixed there. There is no point in slowing down all queries for the sake of improving an inherently erroneous query condition.\n\nregards, Sergei\n\n\n","msg_date":"Thu, 11 Nov 2021 21:15:54 +0300","msg_from":"Sergei Kornilov <sk@zsrv.org>","msg_from_op":true,"msg_subject":"Re:unexpected plan with id = any('{}') condition"},{"msg_contents":"Sergei Kornilov <sk@zsrv.org> writes:\n>> unreadable HTML mess\n\n> ouch, sorry. \"Nobody uses plain text mail, we dropped this thing in the interface” said yandex team.\n\nI'll confess to being a troglodyte, but it's unreadable in the archives\ntoo [1], so there was some misconfiguration somewhere.\n\n\t\t\tregards, tom lane\n\n[1] https://www.postgresql.org/message-id/1910161636627509%40mail.yandex.ru\n\n\n","msg_date":"Thu, 11 Nov 2021 13:22:59 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: unexpected plan with id = any('{}') condition"}],"string":"[\n {\n \"msg_contents\": \"Hello I have such case: create table test (id int not null, status text);insert into test select i, 'foo' from generate_series(1,1000000) i;update test set status = 'bar' where id <= 10;create index test_id on test (id );create index test_status_partial on test (status) where status = 'bar';analyze test ;explain (analyze) select * from test where id = any('{}'); Gives query plan: QUERY PLAN -------------------------------------------------------------------------------------------------------------------------- Index Scan using test_status_partial on test (cost=0.12..4.14 rows=1 width=8) (actual time=0.024..0.025 rows=0 loops=1) Filter: (id = ANY ('{}'::integer[])) Rows Removed by Filter: 10 Planning Time: 0.327 ms Execution Time: 0.048 ms I don't understand why the planner chose such an unrelated partial index. I expected \\\"One-Time Filter: false\\\" here or use of test_id index. I agree, a strange condition, the application should avoid such condition, but why use such an index? Initially was spotted on 13.3 production system (slow query due too much Rows Removed by Filter), then I checked this behaviour on 14.0 and fresh HEAD (db9f287711ac49d9799f93f664d6d101ff8f5891) regards, Sergei\",\n \"msg_date\": \"Thu, 11 Nov 2021 14:00:56 +0300\",\n \"msg_from\": \"Sergei Kornilov <sk@zsrv.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"unexpected plan with id = any('{}') condition\"\n },\n {\n \"msg_contents\": \"Sergei Kornilov <sk@zsrv.org> writes:\\n> [ unreadable HTML mess ]\\n\\nFor the archives, the test case here looks like\\n\\ncreate table test (id int not null, status text);\\ninsert into test select i, 'foo' from generate_series(1,1000000) i;\\nupdate test set status = 'bar' where id <= 10;\\ncreate index test_id on test (id );\\ncreate index test_status_partial on test (status) where status = 'bar';\\nanalyze test ;\\nexplain select * from test where id = any('{}');\\n\\nand the question is why you get\\n\\n QUERY PLAN \\n--------------------------------------------------------------------------------\\n Index Scan using test_status_partial on test (cost=0.12..4.14 rows=1 width=8)\\n Filter: (id = ANY ('{}'::integer[]))\\n(2 rows)\\n\\nwhen that partial index seems unrelated to the query conditions.\\n\\nThe answer is that predicate_implied_by() returns TRUE, because\\nit can see that the WHERE condition is equivalent to constant FALSE,\\nand FALSE implies anything. So the partial index is accepted as\\nproven by the WHERE condition, and then it looks a trifle cheaper\\nto scan than the complete index, which gets an estimate like\\n\\n Index Scan using test_id on test (cost=0.42..4.44 rows=1 width=8)\\n Index Cond: (id = ANY ('{}'::integer[]))\\n\\nI don't see anything wrong here. Maybe one could wish that the\\nplanner reduced the WHERE to constant-FALSE outright, which would\\nyield an even cheaper plan; but I'm not convinced we want to expend\\nextra cycles looking for such cases.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 12:32:04 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: unexpected plan with id = any('{}') condition\"\n },\n {\n \"msg_contents\": \"Hello\\n\\nThank you for the explanation!\\n\\n> unreadable HTML mess\\n\\nouch, sorry. \\\"Nobody uses plain text mail, we dropped this thing in the interface” said yandex team. (I know that some members of the Yandex team read mailing lists, could you ping your colleagues?)\\n\\n> but I'm not convinced we want to expend\\n> extra cycles looking for such cases.\\n\\nAgree. This is an application bug and needs to be fixed there. There is no point in slowing down all queries for the sake of improving an inherently erroneous query condition.\\n\\nregards, Sergei\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 21:15:54 +0300\",\n \"msg_from\": \"Sergei Kornilov <sk@zsrv.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re:unexpected plan with id = any('{}') condition\"\n },\n {\n \"msg_contents\": \"Sergei Kornilov <sk@zsrv.org> writes:\\n>> unreadable HTML mess\\n\\n> ouch, sorry. \\\"Nobody uses plain text mail, we dropped this thing in the interface” said yandex team.\\n\\nI'll confess to being a troglodyte, but it's unreadable in the archives\\ntoo [1], so there was some misconfiguration somewhere.\\n\\n\\t\\t\\tregards, tom lane\\n\\n[1] https://www.postgresql.org/message-id/1910161636627509%40mail.yandex.ru\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 13:22:59 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: unexpected plan with id = any('{}') condition\"\n }\n]","isConversational":false}}},{"rowIdx":53074,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"tadarida just reported this [1]:\n\n2021-11-11 14:40:40.040 CET [29620:388] 023_cic_2pc.pl ERROR: could not read two-phase state from WAL at 0/1716C68\n\nThat error message is remarkably unhelpful, and I think unnecessarily so.\nWhy the heck isn't XlogReadTwoPhaseData passing on the error string\nthat it just got from XLogReadRecord?\n\nNow, XLogReadRecord's API spec does say it might return NULL errormsg\nif \"the page read callback already reported the error\". But there's\nno sign of that in tadarida's log. Anyway, I have in mind to do\n\n ereport(ERROR,\n (errcode_for_file_access(),\n- errmsg(\"could not read two-phase state from WAL at %X/%X\",\n- LSN_FORMAT_ARGS(lsn))));\n+ errmsg(\"could not read two-phase state from WAL at %X/%X: %s\",\n+ LSN_FORMAT_ARGS(lsn),\n+ errormsg ? errormsg : \"(no details available)\")));\n\nBut first, I'm going to look around for other XLogReadRecord\ncallers that might be equally slipshod.\n\n(Identifying the actual underlying bug, if there is one, is a different\nproblem. This test being so new, I'd not be surprised if it is showing\nus a real-but-intermittent issue. But perhaps fixing this error report\nwill yield more info.)\n\n\t\t\tregards, tom lane\n\n[1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-11%2013%3A29%3A58\n\n\n","msg_date":"Thu, 11 Nov 2021 11:06:24 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Strange error in new 003_cic_2pc.pl test"},{"msg_contents":"\n\n> tadarida just reported this [1]:\n> \n> 2021-11-11 14:40:40.040 CET [29620:388] 023_cic_2pc.pl ERROR: could not read two-phase state from WAL at 0/1716C68\n> \n> That error message is remarkably unhelpful, and I think unnecessarily so.\n> Why the heck isn't XlogReadTwoPhaseData passing on the error string\n> that it just got from XLogReadRecord?\n> \n> Now, XLogReadRecord's API spec does say it might return NULL errormsg\n> if \"the page read callback already reported the error\". But there's\n> no sign of that in tadarida's log. Anyway, I have in mind to do\n> \n> ereport(ERROR,\n> (errcode_for_file_access(),\n> - errmsg(\"could not read two-phase state from WAL at %X/%X\",\n> - LSN_FORMAT_ARGS(lsn))));\n> + errmsg(\"could not read two-phase state from WAL at %X/%X: %s\",\n> + LSN_FORMAT_ARGS(lsn),\n> + errormsg ? errormsg : \"(no details available)\")));\n> \n\nNoah recently proposed the same [0]. But responders like me started to complain about other places without error reporting around xlog reading.\n\nProbably I had to be more clear on the matter. Error logging certainly worth doing, at least this.\n\nThanks!\n\n\nBest regards, Andrey Borodin.\n\n[0] https://www.postgresql.org/message-id/flat/20211107013157.GB790288%40rfd.leadboat.com\n\n\n","msg_date":"Thu, 11 Nov 2021 21:20:28 +0500","msg_from":"Andrey Borodin <x4mmm@yandex-team.ru>","msg_from_op":false,"msg_subject":"Re: Strange error in new 003_cic_2pc.pl test"},{"msg_contents":"Andrey Borodin <x4mmm@yandex-team.ru> writes:\n>> That error message is remarkably unhelpful, and I think unnecessarily so.\n>> Why the heck isn't XlogReadTwoPhaseData passing on the error string\n>> that it just got from XLogReadRecord?\n\n> Noah recently proposed the same [0]. But responders like me started to complain about other places without error reporting around xlog reading.\n\nAh, I missed that thread. Now that we've seen another instance of the\nsame failure, I agree it's urgent to get something committed. Noah's\npatch matches the style of some other callers, so let's go with that.\n\n(And yeah, why are some of the other callers ignoring errormsg? But\nthat can wait.)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 11 Nov 2021 11:52:49 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Strange error in new 003_cic_2pc.pl test"}],"string":"[\n {\n \"msg_contents\": \"tadarida just reported this [1]:\\n\\n2021-11-11 14:40:40.040 CET [29620:388] 023_cic_2pc.pl ERROR: could not read two-phase state from WAL at 0/1716C68\\n\\nThat error message is remarkably unhelpful, and I think unnecessarily so.\\nWhy the heck isn't XlogReadTwoPhaseData passing on the error string\\nthat it just got from XLogReadRecord?\\n\\nNow, XLogReadRecord's API spec does say it might return NULL errormsg\\nif \\\"the page read callback already reported the error\\\". But there's\\nno sign of that in tadarida's log. Anyway, I have in mind to do\\n\\n ereport(ERROR,\\n (errcode_for_file_access(),\\n- errmsg(\\\"could not read two-phase state from WAL at %X/%X\\\",\\n- LSN_FORMAT_ARGS(lsn))));\\n+ errmsg(\\\"could not read two-phase state from WAL at %X/%X: %s\\\",\\n+ LSN_FORMAT_ARGS(lsn),\\n+ errormsg ? errormsg : \\\"(no details available)\\\")));\\n\\nBut first, I'm going to look around for other XLogReadRecord\\ncallers that might be equally slipshod.\\n\\n(Identifying the actual underlying bug, if there is one, is a different\\nproblem. This test being so new, I'd not be surprised if it is showing\\nus a real-but-intermittent issue. But perhaps fixing this error report\\nwill yield more info.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n[1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=tadarida&dt=2021-11-11%2013%3A29%3A58\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 11:06:24 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Strange error in new 003_cic_2pc.pl test\"\n },\n {\n \"msg_contents\": \"\\n\\n> tadarida just reported this [1]:\\n> \\n> 2021-11-11 14:40:40.040 CET [29620:388] 023_cic_2pc.pl ERROR: could not read two-phase state from WAL at 0/1716C68\\n> \\n> That error message is remarkably unhelpful, and I think unnecessarily so.\\n> Why the heck isn't XlogReadTwoPhaseData passing on the error string\\n> that it just got from XLogReadRecord?\\n> \\n> Now, XLogReadRecord's API spec does say it might return NULL errormsg\\n> if \\\"the page read callback already reported the error\\\". But there's\\n> no sign of that in tadarida's log. Anyway, I have in mind to do\\n> \\n> ereport(ERROR,\\n> (errcode_for_file_access(),\\n> - errmsg(\\\"could not read two-phase state from WAL at %X/%X\\\",\\n> - LSN_FORMAT_ARGS(lsn))));\\n> + errmsg(\\\"could not read two-phase state from WAL at %X/%X: %s\\\",\\n> + LSN_FORMAT_ARGS(lsn),\\n> + errormsg ? errormsg : \\\"(no details available)\\\")));\\n> \\n\\nNoah recently proposed the same [0]. But responders like me started to complain about other places without error reporting around xlog reading.\\n\\nProbably I had to be more clear on the matter. Error logging certainly worth doing, at least this.\\n\\nThanks!\\n\\n\\nBest regards, Andrey Borodin.\\n\\n[0] https://www.postgresql.org/message-id/flat/20211107013157.GB790288%40rfd.leadboat.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 21:20:28 +0500\",\n \"msg_from\": \"Andrey Borodin <x4mmm@yandex-team.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Strange error in new 003_cic_2pc.pl test\"\n },\n {\n \"msg_contents\": \"Andrey Borodin <x4mmm@yandex-team.ru> writes:\\n>> That error message is remarkably unhelpful, and I think unnecessarily so.\\n>> Why the heck isn't XlogReadTwoPhaseData passing on the error string\\n>> that it just got from XLogReadRecord?\\n\\n> Noah recently proposed the same [0]. But responders like me started to complain about other places without error reporting around xlog reading.\\n\\nAh, I missed that thread. Now that we've seen another instance of the\\nsame failure, I agree it's urgent to get something committed. Noah's\\npatch matches the style of some other callers, so let's go with that.\\n\\n(And yeah, why are some of the other callers ignoring errormsg? But\\nthat can wait.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 11:52:49 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Strange error in new 003_cic_2pc.pl test\"\n }\n]","isConversational":false}}},{"rowIdx":53075,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI have noticed that $SUBJECT, and I think it's bad, and I'd like to\npropose some patches to fix it. 0001 is hopefully uncontroversial, but\n0002 might need some discussion as I'm not quite sure what approach to\ntake.\n\nThe reason why RecoveryInProgress() has critical side effects is that\nit calls InitXLOGAccess(). It turns out that the only critical thing\nthat InitXLOGAccess() does is call InitXLogInsert().[1] If a backend\ndoesn't call InitXLogInsert(), max_rdatas won't be initialized, and\nthe first attempt to insert WAL will probably fail with something like\n\"ERROR: too much WAL data\". But there's no real reason why\nInitXLogInsert() needs to happen only after recovery is finished. The\nwork that it does is just setting up data structures in backend-local\nmemory, and not much is lost if they are set up and not used. So, in\n0001, I propose to move the call to InitXLogInsert() from\nInitXLOGAccess() to BaseInit(). That means every backend will do it\nthe first time it starts up. It also means that CreateCheckPoint()\nwill no longer need a special case call to InitXLogInsert(), which\nseems like a good thing.\n\nNow, what about the other stuff that InitXLOGAccess() does? As far as\nI can tell, \"wal_segment_size = ControlFile->xlog_seg_size\" is always\nredundant, because LocalProcessControlFile() has always been called,\nand it does the same thing. The postmaster does that call, and in\nnon-EXEC_BACKEND builds, the global variable setting will be inherited\nby all child processes. In EXEC_BACKEND builds, every child process\nwill go through SubPostmasterMain() which also calls\nLocalProcessControlFile() very early in the startup sequence.\nInitXLOGAccess() also sets RedoRecPtr and doPageWrites ... but these\nare non-critical values. If you set them to 0 and false respectively\nand then call XLogInsert(), GetFullPagesWritesInfo() will return 0 and\nfalse, XLogRecordAssemble() will use those values to make wrong\ndecisions, and then XLogInsertRecord() will notice that the wrong\nvalues were used and will update them and cause XLogInsert() to loop\naround. Similarly, XLogCheckpointNeeded() relies on the caller to set\nRedoRecPtr appropriately, but all callers recheck after updating\nRedoRecPtr, so it's fine if the initial value is zero. The only other\ninteresting case is XLogCheckBufferNeedsBackup(), which calls\nGetFullPageWriteInfo(), but if somehow it gets the wrong answer, the\nonly thing that can happen is we might come the wrong conclusion about\nwhether a heap-update WAL record should attempt to compress by looking\nfor a common prefix and suffix between the old and new tuples. That's\nnon-critical, and if it happens or doesn't happen at most once per\nbackend, fine.\n\nNevertheless, what I did in 0002 is remove InitXLOGAccess() but move\nthe initialization of RedoRecPtr and doPageWrites into\nGetFullPageWritesInfo(), only in the case where RedoRecPtr hasn't been\nset yet. This has one property that I really really like, which is\nthat makes the code more understandable. There is no suggestion that\nthe correctness of WAL insertion somehow depends on a\nRecoveryInProgress() call which may or may not have occurred at some\narbitrarily distant time in the past. Initializing the value at the\npoint of first use is just way clearer than initializing it as a\nside-effect of some other function that's not even that tightly\nconnected. However, it does have the effect of introducing a branch\ninto GetFullPageWritesInfo(), which I'm not 100% sure is cheap enough\nnot to matter. I think the obvious alternative is to just not\ninitialize RedoRecPtr and doPageWrites at all and document in the\ncomments that the first time each backend does something with them\nit's going to end up retrying once; perhaps that is preferable. Going\nthe other way, we could get rid of the global variables altogether and\nhave GetFullPageWrites() read the data from shared memory. However,\nunless 8-byte atomicity is guaranteed, that requires a spinlock, so it\nseems likely unappealing.\n\n(In case the explanation above is unclear to casual readers, here's a\nbit more detail: RedoRecPtr is the redo pointer from what we believe\nto be the most recent checkpoint record, and doFullPageWrites is\nwhether full_page_writes are required at the moment. The latter can be\ntrue either because full_page_writes=on or because a base backup is in\nprogress. When assembling an XLOG record, we have to decide whether to\ninclude full-page images in that or not, and we do that based on the\nvalue of these variables: if full page writes are required in general\nand if the page LSN precedes RedoRecPtr, we include a full page image.\nIf it turns out that our RedoRecPtr was out of date - i.e. there's a\nnewer checkpoint we didn't know about - then we go back and\nre-assemble the record to make sure all required full-page images are\nincluded.)\n\nThoughts?\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n[1] Isn't it great that these two similarly-named functions capitalize\n\"xlog\" differently? Wa-hoo.","msg_date":"Thu, 11 Nov 2021 12:15:03 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"RecoveryInProgress() has critical side effects"},{"msg_contents":"On 11/11/21, 9:20 AM, \"Robert Haas\" <robertmhaas@gmail.com> wrote:\r\n> memory, and not much is lost if they are set up and not used. So, in\r\n> 0001, I propose to move the call to InitXLogInsert() from\r\n> InitXLOGAccess() to BaseInit(). That means every backend will do it\r\n> the first time it starts up. It also means that CreateCheckPoint()\r\n> will no longer need a special case call to InitXLogInsert(), which\r\n> seems like a good thing.\r\n\r\n0001 looks reasonable to me.\r\n\r\n> connected. However, it does have the effect of introducing a branch\r\n> into GetFullPageWritesInfo(), which I'm not 100% sure is cheap enough\r\n> not to matter. I think the obvious alternative is to just not\r\n> initialize RedoRecPtr and doPageWrites at all and document in the\r\n> comments that the first time each backend does something with them\r\n> it's going to end up retrying once; perhaps that is preferable. Going\r\n\r\nUnless there's a demonstrable performance benefit from adding the\r\nbranch, my preference is to leave it out. I could be off-base, but it\r\nseems possible that future changes might end up depending on any side\r\neffects from this new branch, which is exactly what you're trying to\r\nfix. Plus, always using the retry path is a good way to test that it\r\nworks as expected.\r\n\r\nNathan\r\n\r\n","msg_date":"Fri, 12 Nov 2021 15:23:30 +0000","msg_from":"\"Bossart, Nathan\" <bossartn@amazon.com>","msg_from_op":false,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On Fri, Nov 12, 2021 at 10:23 AM Bossart, Nathan <bossartn@amazon.com> wrote:\n> 0001 looks reasonable to me.\n\nCool.\n\n> Unless there's a demonstrable performance benefit from adding the\n> branch, my preference is to leave it out. I could be off-base, but it\n> seems possible that future changes might end up depending on any side\n> effects from this new branch, which is exactly what you're trying to\n> fix. Plus, always using the retry path is a good way to test that it\n> works as expected.\n\nHere's a new version that does it that way. Any other opinions?\n\nThe best thing I could come up with for a test case for this was to\ntry repeatedly making a new connection and running \"SELECT\ntxid_current()\", which will cause just one WAL record to be generated.\nUnfortunately that path has overhead from a lot of other causes so I'm\nnot sure the results are very meaningful, but here they are:\n\nv1: 0.378 ms\nv2: 0.391 ms\ncommon base commit (10eae82b2): 0.376 ms\n\nmethodology:\nfor i in `seq 1 1000`; do psql -c '\\timing' -c 'select\ntxid_current();'; done | grep '^Time:' | awk '{total+=$2} END {print\ntotal/NR}'\nrun twice, discarding the first result, since the very first\nconnection attempt after starting a new server process is notably\nslower\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com","msg_date":"Mon, 15 Nov 2021 16:29:28 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On 11/15/21, 1:30 PM, \"Robert Haas\" <robertmhaas@gmail.com> wrote:\r\n> Here's a new version that does it that way. Any other opinions?\r\n\r\nLGTM\r\n\r\n> The best thing I could come up with for a test case for this was to\r\n> try repeatedly making a new connection and running \"SELECT\r\n> txid_current()\", which will cause just one WAL record to be generated.\r\n> Unfortunately that path has overhead from a lot of other causes so I'm\r\n> not sure the results are very meaningful, but here they are:\r\n>\r\n> v1: 0.378 ms\r\n> v2: 0.391 ms\r\n> common base commit (10eae82b2): 0.376 ms\r\n\r\nI'm personally not too worried about a ~4% regression in this\r\nparticular benchmark...\r\n\r\nNathan\r\n\r\n","msg_date":"Mon, 15 Nov 2021 22:09:17 +0000","msg_from":"\"Bossart, Nathan\" <bossartn@amazon.com>","msg_from_op":false,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On Mon, Nov 15, 2021 at 10:09:17PM +0000, Bossart, Nathan wrote:\n> On 11/15/21, 1:30 PM, \"Robert Haas\" <robertmhaas@gmail.com> wrote:\n>> Here's a new version that does it that way. Any other opinions?\n> \n> LGTM\n\nPatch 0001 means that the startup process would set up the structures\nto be able to build WAL records before running any kind of recovery\naction rather than doing it when it really needs it. That's fine by\nme.\n\nIs patch 0002 actually right regarding the handling of doPageWrites?\nOnce applied, we finish by setting it when the startup process starts\nand not anymore at the end of recovery based on its the state of\nInsert, but this could have changed while in recovery when replaying\none or more XLOG_FPW_CHANGE records.\n\n> I'm personally not too worried about a ~4% regression in this\n> particular benchmark...\n\nThis is not a hot code path, that should be fine.\n--\nMichael","msg_date":"Tue, 16 Nov 2021 11:34:40 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On Mon, Nov 15, 2021 at 9:34 PM Michael Paquier <michael@paquier.xyz> wrote:\n> Patch 0001 means that the startup process would set up the structures\n> to be able to build WAL records before running any kind of recovery\n> action rather than doing it when it really needs it. That's fine by\n> me.\n\nGreat, thanks. I think I'll go ahead and commit this one, then.\n\n> Is patch 0002 actually right regarding the handling of doPageWrites?\n> Once applied, we finish by setting it when the startup process starts\n> and not anymore at the end of recovery based on its the state of\n> Insert, but this could have changed while in recovery when replaying\n> one or more XLOG_FPW_CHANGE records.\n\nMaybe I'm not understanding you properly here, but it seems like you\nmight be forgetting that this is a local variable and thus every\nbackend is going to have something different. In the startup process,\nit will be initialized by StartupXLOG(); in other processes, it's\ncurrently initialized by RecoveryInProgress(), but with this patch it\nwouldn't be. Either way, it's then updated by future calls to\nXLogInsertRecord() as required. XLOG_FPW_CHANGE records might affect\nthe new value that gets set the next time XLogInsertRecord(), but they\ndon't directly affect doPageWrites.\n\n> > I'm personally not too worried about a ~4% regression in this\n> > particular benchmark...\n>\n> This is not a hot code path, that should be fine.\n\nOK. I'll wait a while and see if anyone else wants to weigh in.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 16 Nov 2021 09:41:58 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"Hi,\n\nOn 2021-11-11 12:15:03 -0500, Robert Haas wrote:\n> The reason why RecoveryInProgress() has critical side effects is that\n> it calls InitXLOGAccess(). It turns out that the only critical thing\n> that InitXLOGAccess() does is call InitXLogInsert().[1] If a backend\n> doesn't call InitXLogInsert(), max_rdatas won't be initialized, and\n> the first attempt to insert WAL will probably fail with something like\n> \"ERROR: too much WAL data\". But there's no real reason why\n> InitXLogInsert() needs to happen only after recovery is finished.\n\nI think we should consider allocating that memory in postmaster, on !windows\nat least. Or, perhaps even better, to initially use some static variable, and\nonly use a separate memory context when XLogEnsureRecordSpace(). Reserving\nall that memory in every process just needlessly increases our per-connection\nmemory usage, for no good reason.\n\n\n> The\n> work that it does is just setting up data structures in backend-local\n> memory, and not much is lost if they are set up and not used. So, in\n> 0001, I propose to move the call to InitXLogInsert() from\n> InitXLOGAccess() to BaseInit(). That means every backend will do it\n> the first time it starts up. It also means that CreateCheckPoint()\n> will no longer need a special case call to InitXLogInsert(), which\n> seems like a good thing.\n\nYes. Seems making BaseInit() being executed at a halfway consistent point is\nstarting to have some benefits at least ;)\n\n\n> Nevertheless, what I did in 0002 is remove InitXLOGAccess() but move\n> the initialization of RedoRecPtr and doPageWrites into\n> GetFullPageWritesInfo(), only in the case where RedoRecPtr hasn't been\n> set yet. This has one property that I really really like, which is\n> that makes the code more understandable. There is no suggestion that\n> the correctness of WAL insertion somehow depends on a\n> RecoveryInProgress() call which may or may not have occurred at some\n> arbitrarily distant time in the past. Initializing the value at the\n> point of first use is just way clearer than initializing it as a\n> side-effect of some other function that's not even that tightly\n> connected. However, it does have the effect of introducing a branch\n> into GetFullPageWritesInfo(), which I'm not 100% sure is cheap enough\n> not to matter.\n\nHm. Compared to the other costs of the XLogInsert do/while loop it probably\ndoesn't matter. One nearly-always-false branch is vastly cheaper than going\nthrough the loop unnecessarily. Sure, the unnecessarily iteration saved will\nonly be the first (for now, it might be worth to refresh the values more\noften), but there's enough branches in the body of the loop that I can't\nreally see it mattering.\n\nMaybe kinda conceivably the cost of that branch could be an argument if\nGetFullPageWritesInfo() were inline in XLogInsert(). But it isn't.\n\n\n> I think the obvious alternative is to just not\n> initialize RedoRecPtr and doPageWrites at all and document in the\n> comments that the first time each backend does something with them\n> it's going to end up retrying once; perhaps that is preferable. Going\n> the other way, we could get rid of the global variables altogether and\n> have GetFullPageWrites() read the data from shared memory. However,\n> unless 8-byte atomicity is guaranteed, that requires a spinlock, so it\n> seems likely unappealing.\n\nI think it'd be fine to just make them 8byte atomics, which'd be lock-free on\nrelevant platforms (at least once the aarch64 thing is fixed, there's a\npatch).\n\nXLogCtlInsert already takes care to ensure that RedoRecPtr/fullPageWrites are\non a cacheline that's not constantly modified.\n\n\nIf we really wanted to optimize the no-8-byte-single-copy-atomicity path, we\ncould introduce a counter counting how many times RedoRecPtr/fullPageWrites\nhave changed. But it seems unlikely to be worth it.\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 16 Nov 2021 10:51:16 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"Hi,\n\nOn 2021-11-15 16:29:28 -0500, Robert Haas wrote:\n> v1: 0.378 ms\n> v2: 0.391 ms\n> common base commit (10eae82b2): 0.376 ms\n\nIs this with assertions enabled? Because on my workstation (which likely is\nslower on a single-core basis than your laptop) I get around half of this with\nan optimized build (and a non-optimized build rarely is worth using as a\nmeasuring stick). It could also be that your psqlrc does something\nheavyweight?\n\n\n> methodology:\n> for i in `seq 1 1000`; do psql -c '\\timing' -c 'select\n> txid_current();'; done | grep '^Time:' | awk '{total+=$2} END {print\n> total/NR}'\n> run twice, discarding the first result, since the very first\n> connection attempt after starting a new server process is notably\n> slower\n\nHm. I think this might included a bunch of convoluting factors that make it\nhard to judge the actual size of the performance difference. It'll\ne.g. include a fair bit of syscache initialization overhead that won't change\nbetween the two approaches. This could be addressed by doing something like -c\n\"SELECT 'txid_current'::regproc;\" first.\n\nAlso, the psql invocations itself use up a fair bit of time, even if you\nignored them from the result with the \\timing trick. A pgbench -C doing 1k\nSELECT 1;s is about ~1.5s for me, whereas psql is ~5.7s.\n\nJust to size up the order of magnitude of overhead of the connection\nestablishment and syscache initialization, I ran a pgbench with a script of\n\nSELECT 1;\nSELECT 1;\nSELECT 'txid_current()'::regprocedure;\nSELECT 'txid_current()'::regprocedure;\nSELECT txid_current();\nSELECT txid_current();\n\nand ran that with pgbench -n -f /tmp/txid.sql -C -t1000 -P1 --report-latencies\nand got\nstatement latencies in milliseconds:\n 0.125 SELECT 1;\n 0.024 SELECT 1;\n 0.095 SELECT 'txid_current()'::regprocedure;\n 0.025 SELECT 'txid_current()'::regprocedure;\n 0.033 SELECT txid_current();\n 0.024 SELECT txid_current();\n\nwhich nicely shows how much of the overhead is not related to the actual\ntxid_current() invocation.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 16 Nov 2021 11:27:29 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On Tue, Nov 16, 2021 at 2:27 PM Andres Freund <andres@anarazel.de> wrote:\n> Is this with assertions enabled? Because on my workstation (which likely is\n> slower on a single-core basis than your laptop) I get around half of this with\n> an optimized build (and a non-optimized build rarely is worth using as a\n> measuring stick). It could also be that your psqlrc does something\n> heavyweight?\n\nAssertions disabled. ~/.psqlrc does not exist.\n\n> Hm. I think this might included a bunch of convoluting factors that make it\n> hard to judge the actual size of the performance difference.\n\nYes, I think so, too.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 16 Nov 2021 15:19:19 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"Hi,\n\nOn 2021-11-16 15:19:19 -0500, Robert Haas wrote:\n> > Hm. I think this might included a bunch of convoluting factors that make it\n> > hard to judge the actual size of the performance difference.\n> \n> Yes, I think so, too.\n\nFWIW I ran that pgench thing I presented upthread, and I didn't see any\nmeaningful and repeatable performance difference 354a1f8d220, ad26ee28250 and\n0002 applied ontop of ad26ee28250. The run-to-run variance is way higher than\nthe difference between the changes.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 16 Nov 2021 12:42:23 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On Tue, Nov 16, 2021 at 3:42 PM Andres Freund <andres@anarazel.de> wrote:\n> On 2021-11-16 15:19:19 -0500, Robert Haas wrote:\n> > > Hm. I think this might included a bunch of convoluting factors that make it\n> > > hard to judge the actual size of the performance difference.\n> >\n> > Yes, I think so, too.\n>\n> FWIW I ran that pgench thing I presented upthread, and I didn't see any\n> meaningful and repeatable performance difference 354a1f8d220, ad26ee28250 and\n> 0002 applied ontop of ad26ee28250. The run-to-run variance is way higher than\n> the difference between the changes.\n\nThanks. I suspected that the results I was seeing were not meaningful,\nbut it's hard to be sure when the results seem to be repeatable\nlocally.\n\nI'm still not entirely clear on whether you prefer v1-0002, v2-0002,\nor something else.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 16 Nov 2021 16:30:27 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On 2021-11-16 16:30:27 -0500, Robert Haas wrote:\n> I'm still not entirely clear on whether you prefer v1-0002, v2-0002,\n> or something else.\n\nI think it basically doesn't matter much. It's such a small piece of the cost\ncompared to either the cost of a single insert or the ratio between\nRedoRecPtr/FPW changes and the number of inserted records.\n\nI guess v2-0002 is mildly simpler, so I very weakly lean towards that.\n\n\n","msg_date":"Tue, 16 Nov 2021 14:04:17 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On 17/11/2021 00:04, Andres Freund wrote:\n> On 2021-11-16 16:30:27 -0500, Robert Haas wrote:\n>> I'm still not entirely clear on whether you prefer v1-0002, v2-0002,\n>> or something else.\n> \n> I think it basically doesn't matter much. It's such a small piece of the cost\n> compared to either the cost of a single insert or the ratio between\n> RedoRecPtr/FPW changes and the number of inserted records.\n> \n> I guess v2-0002 is mildly simpler, so I very weakly lean towards that.\n\nI was going to argue for v1, because I was worried about generating \nfull-page writes unnecessarily. But looking closer, it doesn't do that. \nIt will initially think that no full page images are needed, and perform \nthe extra loop in XLogInsert() to try again with full page images. \nSimilarly in heap_update(), it will fail in the direction of sometimes \ndoing the \"prefix-suffix compression\" unnecessarily, which means a \nlittle bit of extra CPU work, but it won't affect the WAL record that \ngets written.\n\nSo either way is fine, really. That's a bit subtle, though, a comment \nsomewhere would be good.\n\nBut here's yet another idea: We could initialize RedoRecPtr and \ndoPageWrites in InitXLogInsert(). It would seem like a pretty natural \nplace for it.\n\n- Heikki\n\nPS. typo in patch v2: s/prepard/prepared/\n\n\n","msg_date":"Sat, 20 Nov 2021 22:22:30 +0200","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":false,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On Sat, Nov 20, 2021 at 3:22 PM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\n> But here's yet another idea: We could initialize RedoRecPtr and\n> doPageWrites in InitXLogInsert(). It would seem like a pretty natural\n> place for it.\n\nI considered that, but it seemed like an abstraction violation to me,\nbecause that code is part of xloginsert.c, which is mostly concerned\nwith assembly of write-ahead log records, not the associated\nshared-memory state. Also, I don't think there's any guarantee that\nthe state in shared memory is initialized at the time we call that\nfunction, so we might just be copying uninitialized memory into other\nuninitialized memory.\n\n> PS. typo in patch v2: s/prepard/prepared/\n\nThanks, fixed.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com","msg_date":"Wed, 1 Dec 2021 10:38:27 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"There are a couple typos. \"uninitalized,\" in one place.\n\nAlso,\n\n+ * If this is either a bootstrap process nor a standalone backend, start\n\nIs \"either .. nor\" correct? It looks very wrong to me. I think you\nmeant \"either .. or\".\n\nOverall, this patch seems mostly about documenting some undesirable\nexisting behavior. Is part of your plan to get such things fixed/\nreplaced entirely?\n\n-- \nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\n\"We're here to devour each other alive\" (Hobbes)\n\n\n","msg_date":"Wed, 1 Dec 2021 13:38:25 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On Tue, Nov 16, 2021 at 09:41:58AM -0500, Robert Haas wrote:\n> Maybe I'm not understanding you properly here, but it seems like you\n> might be forgetting that this is a local variable and thus every\n> backend is going to have something different. In the startup process,\n> it will be initialized by StartupXLOG(); in other processes, it's\n> currently initialized by RecoveryInProgress(), but with this patch it\n> wouldn't be. Either way, it's then updated by future calls to\n> XLogInsertRecord() as required. XLOG_FPW_CHANGE records might affect\n> the new value that gets set the next time XLogInsertRecord(), but they\n> don't directly affect doPageWrites.\n\nMy main worry here is that this changes slightly the definition of\ndoPageWrites across stable branches at the end of recovery as there\ncould be records generated there. Note that GetFullPageWriteInfo()\ngets called in XLogInsert(), while Insert->fullPageWrites gets updated\nbefore CleanupAfterArchiveRecovery(). And it may influence\nthe value of doPageWrites in the startup process.\n--\nMichael","msg_date":"Sun, 5 Dec 2021 09:44:16 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On Sat, Dec 4, 2021 at 7:44 PM Michael Paquier <michael@paquier.xyz> wrote:\n> My main worry here is that this changes slightly the definition of\n> doPageWrites across stable branches at the end of recovery as there\n> could be records generated there. Note that GetFullPageWriteInfo()\n> gets called in XLogInsert(), while Insert->fullPageWrites gets updated\n> before CleanupAfterArchiveRecovery(). And it may influence\n> the value of doPageWrites in the startup process.\n\nBut ... so what? All the code that uses it retries if the value that\nwas tentatively used turns out to be wrong.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 7 Dec 2021 17:55:02 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: RecoveryInProgress() has critical side effects"},{"msg_contents":"On Tue, Dec 7, 2021 at 5:55 PM Robert Haas <robertmhaas@gmail.com> wrote:\n> On Sat, Dec 4, 2021 at 7:44 PM Michael Paquier <michael@paquier.xyz> wrote:\n> > My main worry here is that this changes slightly the definition of\n> > doPageWrites across stable branches at the end of recovery as there\n> > could be records generated there. Note that GetFullPageWriteInfo()\n> > gets called in XLogInsert(), while Insert->fullPageWrites gets updated\n> > before CleanupAfterArchiveRecovery(). And it may influence\n> > the value of doPageWrites in the startup process.\n>\n> But ... so what? All the code that uses it retries if the value that\n> was tentatively used turns out to be wrong.\n\nDespite the fact that this question didn't get further discussion, and\nthe fact that nobody seems quite sure what the best way of proceeding\nhere is, my interpretation of the comments made so far is that nobody\nthinks that what we have now is superior to either of the proposed\nalternatives, and there's a weak preference for v2 over v1. So I\ncommitted that one.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 13 Dec 2021 10:10:43 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":true,"msg_subject":"Re: RecoveryInProgress() has critical side effects"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI have noticed that $SUBJECT, and I think it's bad, and I'd like to\\npropose some patches to fix it. 0001 is hopefully uncontroversial, but\\n0002 might need some discussion as I'm not quite sure what approach to\\ntake.\\n\\nThe reason why RecoveryInProgress() has critical side effects is that\\nit calls InitXLOGAccess(). It turns out that the only critical thing\\nthat InitXLOGAccess() does is call InitXLogInsert().[1] If a backend\\ndoesn't call InitXLogInsert(), max_rdatas won't be initialized, and\\nthe first attempt to insert WAL will probably fail with something like\\n\\\"ERROR: too much WAL data\\\". But there's no real reason why\\nInitXLogInsert() needs to happen only after recovery is finished. The\\nwork that it does is just setting up data structures in backend-local\\nmemory, and not much is lost if they are set up and not used. So, in\\n0001, I propose to move the call to InitXLogInsert() from\\nInitXLOGAccess() to BaseInit(). That means every backend will do it\\nthe first time it starts up. It also means that CreateCheckPoint()\\nwill no longer need a special case call to InitXLogInsert(), which\\nseems like a good thing.\\n\\nNow, what about the other stuff that InitXLOGAccess() does? As far as\\nI can tell, \\\"wal_segment_size = ControlFile->xlog_seg_size\\\" is always\\nredundant, because LocalProcessControlFile() has always been called,\\nand it does the same thing. The postmaster does that call, and in\\nnon-EXEC_BACKEND builds, the global variable setting will be inherited\\nby all child processes. In EXEC_BACKEND builds, every child process\\nwill go through SubPostmasterMain() which also calls\\nLocalProcessControlFile() very early in the startup sequence.\\nInitXLOGAccess() also sets RedoRecPtr and doPageWrites ... but these\\nare non-critical values. If you set them to 0 and false respectively\\nand then call XLogInsert(), GetFullPagesWritesInfo() will return 0 and\\nfalse, XLogRecordAssemble() will use those values to make wrong\\ndecisions, and then XLogInsertRecord() will notice that the wrong\\nvalues were used and will update them and cause XLogInsert() to loop\\naround. Similarly, XLogCheckpointNeeded() relies on the caller to set\\nRedoRecPtr appropriately, but all callers recheck after updating\\nRedoRecPtr, so it's fine if the initial value is zero. The only other\\ninteresting case is XLogCheckBufferNeedsBackup(), which calls\\nGetFullPageWriteInfo(), but if somehow it gets the wrong answer, the\\nonly thing that can happen is we might come the wrong conclusion about\\nwhether a heap-update WAL record should attempt to compress by looking\\nfor a common prefix and suffix between the old and new tuples. That's\\nnon-critical, and if it happens or doesn't happen at most once per\\nbackend, fine.\\n\\nNevertheless, what I did in 0002 is remove InitXLOGAccess() but move\\nthe initialization of RedoRecPtr and doPageWrites into\\nGetFullPageWritesInfo(), only in the case where RedoRecPtr hasn't been\\nset yet. This has one property that I really really like, which is\\nthat makes the code more understandable. There is no suggestion that\\nthe correctness of WAL insertion somehow depends on a\\nRecoveryInProgress() call which may or may not have occurred at some\\narbitrarily distant time in the past. Initializing the value at the\\npoint of first use is just way clearer than initializing it as a\\nside-effect of some other function that's not even that tightly\\nconnected. However, it does have the effect of introducing a branch\\ninto GetFullPageWritesInfo(), which I'm not 100% sure is cheap enough\\nnot to matter. I think the obvious alternative is to just not\\ninitialize RedoRecPtr and doPageWrites at all and document in the\\ncomments that the first time each backend does something with them\\nit's going to end up retrying once; perhaps that is preferable. Going\\nthe other way, we could get rid of the global variables altogether and\\nhave GetFullPageWrites() read the data from shared memory. However,\\nunless 8-byte atomicity is guaranteed, that requires a spinlock, so it\\nseems likely unappealing.\\n\\n(In case the explanation above is unclear to casual readers, here's a\\nbit more detail: RedoRecPtr is the redo pointer from what we believe\\nto be the most recent checkpoint record, and doFullPageWrites is\\nwhether full_page_writes are required at the moment. The latter can be\\ntrue either because full_page_writes=on or because a base backup is in\\nprogress. When assembling an XLOG record, we have to decide whether to\\ninclude full-page images in that or not, and we do that based on the\\nvalue of these variables: if full page writes are required in general\\nand if the page LSN precedes RedoRecPtr, we include a full page image.\\nIf it turns out that our RedoRecPtr was out of date - i.e. there's a\\nnewer checkpoint we didn't know about - then we go back and\\nre-assemble the record to make sure all required full-page images are\\nincluded.)\\n\\nThoughts?\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n[1] Isn't it great that these two similarly-named functions capitalize\\n\\\"xlog\\\" differently? Wa-hoo.\",\n \"msg_date\": \"Thu, 11 Nov 2021 12:15:03 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On 11/11/21, 9:20 AM, \\\"Robert Haas\\\" <robertmhaas@gmail.com> wrote:\\r\\n> memory, and not much is lost if they are set up and not used. So, in\\r\\n> 0001, I propose to move the call to InitXLogInsert() from\\r\\n> InitXLOGAccess() to BaseInit(). That means every backend will do it\\r\\n> the first time it starts up. It also means that CreateCheckPoint()\\r\\n> will no longer need a special case call to InitXLogInsert(), which\\r\\n> seems like a good thing.\\r\\n\\r\\n0001 looks reasonable to me.\\r\\n\\r\\n> connected. However, it does have the effect of introducing a branch\\r\\n> into GetFullPageWritesInfo(), which I'm not 100% sure is cheap enough\\r\\n> not to matter. I think the obvious alternative is to just not\\r\\n> initialize RedoRecPtr and doPageWrites at all and document in the\\r\\n> comments that the first time each backend does something with them\\r\\n> it's going to end up retrying once; perhaps that is preferable. Going\\r\\n\\r\\nUnless there's a demonstrable performance benefit from adding the\\r\\nbranch, my preference is to leave it out. I could be off-base, but it\\r\\nseems possible that future changes might end up depending on any side\\r\\neffects from this new branch, which is exactly what you're trying to\\r\\nfix. Plus, always using the retry path is a good way to test that it\\r\\nworks as expected.\\r\\n\\r\\nNathan\\r\\n\\r\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 15:23:30 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 10:23 AM Bossart, Nathan <bossartn@amazon.com> wrote:\\n> 0001 looks reasonable to me.\\n\\nCool.\\n\\n> Unless there's a demonstrable performance benefit from adding the\\n> branch, my preference is to leave it out. I could be off-base, but it\\n> seems possible that future changes might end up depending on any side\\n> effects from this new branch, which is exactly what you're trying to\\n> fix. Plus, always using the retry path is a good way to test that it\\n> works as expected.\\n\\nHere's a new version that does it that way. Any other opinions?\\n\\nThe best thing I could come up with for a test case for this was to\\ntry repeatedly making a new connection and running \\\"SELECT\\ntxid_current()\\\", which will cause just one WAL record to be generated.\\nUnfortunately that path has overhead from a lot of other causes so I'm\\nnot sure the results are very meaningful, but here they are:\\n\\nv1: 0.378 ms\\nv2: 0.391 ms\\ncommon base commit (10eae82b2): 0.376 ms\\n\\nmethodology:\\nfor i in `seq 1 1000`; do psql -c '\\\\timing' -c 'select\\ntxid_current();'; done | grep '^Time:' | awk '{total+=$2} END {print\\ntotal/NR}'\\nrun twice, discarding the first result, since the very first\\nconnection attempt after starting a new server process is notably\\nslower\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Mon, 15 Nov 2021 16:29:28 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On 11/15/21, 1:30 PM, \\\"Robert Haas\\\" <robertmhaas@gmail.com> wrote:\\r\\n> Here's a new version that does it that way. Any other opinions?\\r\\n\\r\\nLGTM\\r\\n\\r\\n> The best thing I could come up with for a test case for this was to\\r\\n> try repeatedly making a new connection and running \\\"SELECT\\r\\n> txid_current()\\\", which will cause just one WAL record to be generated.\\r\\n> Unfortunately that path has overhead from a lot of other causes so I'm\\r\\n> not sure the results are very meaningful, but here they are:\\r\\n>\\r\\n> v1: 0.378 ms\\r\\n> v2: 0.391 ms\\r\\n> common base commit (10eae82b2): 0.376 ms\\r\\n\\r\\nI'm personally not too worried about a ~4% regression in this\\r\\nparticular benchmark...\\r\\n\\r\\nNathan\\r\\n\\r\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 22:09:17 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 10:09:17PM +0000, Bossart, Nathan wrote:\\n> On 11/15/21, 1:30 PM, \\\"Robert Haas\\\" <robertmhaas@gmail.com> wrote:\\n>> Here's a new version that does it that way. Any other opinions?\\n> \\n> LGTM\\n\\nPatch 0001 means that the startup process would set up the structures\\nto be able to build WAL records before running any kind of recovery\\naction rather than doing it when it really needs it. That's fine by\\nme.\\n\\nIs patch 0002 actually right regarding the handling of doPageWrites?\\nOnce applied, we finish by setting it when the startup process starts\\nand not anymore at the end of recovery based on its the state of\\nInsert, but this could have changed while in recovery when replaying\\none or more XLOG_FPW_CHANGE records.\\n\\n> I'm personally not too worried about a ~4% regression in this\\n> particular benchmark...\\n\\nThis is not a hot code path, that should be fine.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 16 Nov 2021 11:34:40 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 9:34 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> Patch 0001 means that the startup process would set up the structures\\n> to be able to build WAL records before running any kind of recovery\\n> action rather than doing it when it really needs it. That's fine by\\n> me.\\n\\nGreat, thanks. I think I'll go ahead and commit this one, then.\\n\\n> Is patch 0002 actually right regarding the handling of doPageWrites?\\n> Once applied, we finish by setting it when the startup process starts\\n> and not anymore at the end of recovery based on its the state of\\n> Insert, but this could have changed while in recovery when replaying\\n> one or more XLOG_FPW_CHANGE records.\\n\\nMaybe I'm not understanding you properly here, but it seems like you\\nmight be forgetting that this is a local variable and thus every\\nbackend is going to have something different. In the startup process,\\nit will be initialized by StartupXLOG(); in other processes, it's\\ncurrently initialized by RecoveryInProgress(), but with this patch it\\nwouldn't be. Either way, it's then updated by future calls to\\nXLogInsertRecord() as required. XLOG_FPW_CHANGE records might affect\\nthe new value that gets set the next time XLogInsertRecord(), but they\\ndon't directly affect doPageWrites.\\n\\n> > I'm personally not too worried about a ~4% regression in this\\n> > particular benchmark...\\n>\\n> This is not a hot code path, that should be fine.\\n\\nOK. I'll wait a while and see if anyone else wants to weigh in.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 09:41:58 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2021-11-11 12:15:03 -0500, Robert Haas wrote:\\n> The reason why RecoveryInProgress() has critical side effects is that\\n> it calls InitXLOGAccess(). It turns out that the only critical thing\\n> that InitXLOGAccess() does is call InitXLogInsert().[1] If a backend\\n> doesn't call InitXLogInsert(), max_rdatas won't be initialized, and\\n> the first attempt to insert WAL will probably fail with something like\\n> \\\"ERROR: too much WAL data\\\". But there's no real reason why\\n> InitXLogInsert() needs to happen only after recovery is finished.\\n\\nI think we should consider allocating that memory in postmaster, on !windows\\nat least. Or, perhaps even better, to initially use some static variable, and\\nonly use a separate memory context when XLogEnsureRecordSpace(). Reserving\\nall that memory in every process just needlessly increases our per-connection\\nmemory usage, for no good reason.\\n\\n\\n> The\\n> work that it does is just setting up data structures in backend-local\\n> memory, and not much is lost if they are set up and not used. So, in\\n> 0001, I propose to move the call to InitXLogInsert() from\\n> InitXLOGAccess() to BaseInit(). That means every backend will do it\\n> the first time it starts up. It also means that CreateCheckPoint()\\n> will no longer need a special case call to InitXLogInsert(), which\\n> seems like a good thing.\\n\\nYes. Seems making BaseInit() being executed at a halfway consistent point is\\nstarting to have some benefits at least ;)\\n\\n\\n> Nevertheless, what I did in 0002 is remove InitXLOGAccess() but move\\n> the initialization of RedoRecPtr and doPageWrites into\\n> GetFullPageWritesInfo(), only in the case where RedoRecPtr hasn't been\\n> set yet. This has one property that I really really like, which is\\n> that makes the code more understandable. There is no suggestion that\\n> the correctness of WAL insertion somehow depends on a\\n> RecoveryInProgress() call which may or may not have occurred at some\\n> arbitrarily distant time in the past. Initializing the value at the\\n> point of first use is just way clearer than initializing it as a\\n> side-effect of some other function that's not even that tightly\\n> connected. However, it does have the effect of introducing a branch\\n> into GetFullPageWritesInfo(), which I'm not 100% sure is cheap enough\\n> not to matter.\\n\\nHm. Compared to the other costs of the XLogInsert do/while loop it probably\\ndoesn't matter. One nearly-always-false branch is vastly cheaper than going\\nthrough the loop unnecessarily. Sure, the unnecessarily iteration saved will\\nonly be the first (for now, it might be worth to refresh the values more\\noften), but there's enough branches in the body of the loop that I can't\\nreally see it mattering.\\n\\nMaybe kinda conceivably the cost of that branch could be an argument if\\nGetFullPageWritesInfo() were inline in XLogInsert(). But it isn't.\\n\\n\\n> I think the obvious alternative is to just not\\n> initialize RedoRecPtr and doPageWrites at all and document in the\\n> comments that the first time each backend does something with them\\n> it's going to end up retrying once; perhaps that is preferable. Going\\n> the other way, we could get rid of the global variables altogether and\\n> have GetFullPageWrites() read the data from shared memory. However,\\n> unless 8-byte atomicity is guaranteed, that requires a spinlock, so it\\n> seems likely unappealing.\\n\\nI think it'd be fine to just make them 8byte atomics, which'd be lock-free on\\nrelevant platforms (at least once the aarch64 thing is fixed, there's a\\npatch).\\n\\nXLogCtlInsert already takes care to ensure that RedoRecPtr/fullPageWrites are\\non a cacheline that's not constantly modified.\\n\\n\\nIf we really wanted to optimize the no-8-byte-single-copy-atomicity path, we\\ncould introduce a counter counting how many times RedoRecPtr/fullPageWrites\\nhave changed. But it seems unlikely to be worth it.\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 10:51:16 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2021-11-15 16:29:28 -0500, Robert Haas wrote:\\n> v1: 0.378 ms\\n> v2: 0.391 ms\\n> common base commit (10eae82b2): 0.376 ms\\n\\nIs this with assertions enabled? Because on my workstation (which likely is\\nslower on a single-core basis than your laptop) I get around half of this with\\nan optimized build (and a non-optimized build rarely is worth using as a\\nmeasuring stick). It could also be that your psqlrc does something\\nheavyweight?\\n\\n\\n> methodology:\\n> for i in `seq 1 1000`; do psql -c '\\\\timing' -c 'select\\n> txid_current();'; done | grep '^Time:' | awk '{total+=$2} END {print\\n> total/NR}'\\n> run twice, discarding the first result, since the very first\\n> connection attempt after starting a new server process is notably\\n> slower\\n\\nHm. I think this might included a bunch of convoluting factors that make it\\nhard to judge the actual size of the performance difference. It'll\\ne.g. include a fair bit of syscache initialization overhead that won't change\\nbetween the two approaches. This could be addressed by doing something like -c\\n\\\"SELECT 'txid_current'::regproc;\\\" first.\\n\\nAlso, the psql invocations itself use up a fair bit of time, even if you\\nignored them from the result with the \\\\timing trick. A pgbench -C doing 1k\\nSELECT 1;s is about ~1.5s for me, whereas psql is ~5.7s.\\n\\nJust to size up the order of magnitude of overhead of the connection\\nestablishment and syscache initialization, I ran a pgbench with a script of\\n\\nSELECT 1;\\nSELECT 1;\\nSELECT 'txid_current()'::regprocedure;\\nSELECT 'txid_current()'::regprocedure;\\nSELECT txid_current();\\nSELECT txid_current();\\n\\nand ran that with pgbench -n -f /tmp/txid.sql -C -t1000 -P1 --report-latencies\\nand got\\nstatement latencies in milliseconds:\\n 0.125 SELECT 1;\\n 0.024 SELECT 1;\\n 0.095 SELECT 'txid_current()'::regprocedure;\\n 0.025 SELECT 'txid_current()'::regprocedure;\\n 0.033 SELECT txid_current();\\n 0.024 SELECT txid_current();\\n\\nwhich nicely shows how much of the overhead is not related to the actual\\ntxid_current() invocation.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 11:27:29 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 2:27 PM Andres Freund <andres@anarazel.de> wrote:\\n> Is this with assertions enabled? Because on my workstation (which likely is\\n> slower on a single-core basis than your laptop) I get around half of this with\\n> an optimized build (and a non-optimized build rarely is worth using as a\\n> measuring stick). It could also be that your psqlrc does something\\n> heavyweight?\\n\\nAssertions disabled. ~/.psqlrc does not exist.\\n\\n> Hm. I think this might included a bunch of convoluting factors that make it\\n> hard to judge the actual size of the performance difference.\\n\\nYes, I think so, too.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 15:19:19 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2021-11-16 15:19:19 -0500, Robert Haas wrote:\\n> > Hm. I think this might included a bunch of convoluting factors that make it\\n> > hard to judge the actual size of the performance difference.\\n> \\n> Yes, I think so, too.\\n\\nFWIW I ran that pgench thing I presented upthread, and I didn't see any\\nmeaningful and repeatable performance difference 354a1f8d220, ad26ee28250 and\\n0002 applied ontop of ad26ee28250. The run-to-run variance is way higher than\\nthe difference between the changes.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 12:42:23 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 3:42 PM Andres Freund <andres@anarazel.de> wrote:\\n> On 2021-11-16 15:19:19 -0500, Robert Haas wrote:\\n> > > Hm. I think this might included a bunch of convoluting factors that make it\\n> > > hard to judge the actual size of the performance difference.\\n> >\\n> > Yes, I think so, too.\\n>\\n> FWIW I ran that pgench thing I presented upthread, and I didn't see any\\n> meaningful and repeatable performance difference 354a1f8d220, ad26ee28250 and\\n> 0002 applied ontop of ad26ee28250. The run-to-run variance is way higher than\\n> the difference between the changes.\\n\\nThanks. I suspected that the results I was seeing were not meaningful,\\nbut it's hard to be sure when the results seem to be repeatable\\nlocally.\\n\\nI'm still not entirely clear on whether you prefer v1-0002, v2-0002,\\nor something else.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 16:30:27 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On 2021-11-16 16:30:27 -0500, Robert Haas wrote:\\n> I'm still not entirely clear on whether you prefer v1-0002, v2-0002,\\n> or something else.\\n\\nI think it basically doesn't matter much. It's such a small piece of the cost\\ncompared to either the cost of a single insert or the ratio between\\nRedoRecPtr/FPW changes and the number of inserted records.\\n\\nI guess v2-0002 is mildly simpler, so I very weakly lean towards that.\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 14:04:17 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On 17/11/2021 00:04, Andres Freund wrote:\\n> On 2021-11-16 16:30:27 -0500, Robert Haas wrote:\\n>> I'm still not entirely clear on whether you prefer v1-0002, v2-0002,\\n>> or something else.\\n> \\n> I think it basically doesn't matter much. It's such a small piece of the cost\\n> compared to either the cost of a single insert or the ratio between\\n> RedoRecPtr/FPW changes and the number of inserted records.\\n> \\n> I guess v2-0002 is mildly simpler, so I very weakly lean towards that.\\n\\nI was going to argue for v1, because I was worried about generating \\nfull-page writes unnecessarily. But looking closer, it doesn't do that. \\nIt will initially think that no full page images are needed, and perform \\nthe extra loop in XLogInsert() to try again with full page images. \\nSimilarly in heap_update(), it will fail in the direction of sometimes \\ndoing the \\\"prefix-suffix compression\\\" unnecessarily, which means a \\nlittle bit of extra CPU work, but it won't affect the WAL record that \\ngets written.\\n\\nSo either way is fine, really. That's a bit subtle, though, a comment \\nsomewhere would be good.\\n\\nBut here's yet another idea: We could initialize RedoRecPtr and \\ndoPageWrites in InitXLogInsert(). It would seem like a pretty natural \\nplace for it.\\n\\n- Heikki\\n\\nPS. typo in patch v2: s/prepard/prepared/\\n\\n\\n\",\n \"msg_date\": \"Sat, 20 Nov 2021 22:22:30 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 20, 2021 at 3:22 PM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n> But here's yet another idea: We could initialize RedoRecPtr and\\n> doPageWrites in InitXLogInsert(). It would seem like a pretty natural\\n> place for it.\\n\\nI considered that, but it seemed like an abstraction violation to me,\\nbecause that code is part of xloginsert.c, which is mostly concerned\\nwith assembly of write-ahead log records, not the associated\\nshared-memory state. Also, I don't think there's any guarantee that\\nthe state in shared memory is initialized at the time we call that\\nfunction, so we might just be copying uninitialized memory into other\\nuninitialized memory.\\n\\n> PS. typo in patch v2: s/prepard/prepared/\\n\\nThanks, fixed.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Wed, 1 Dec 2021 10:38:27 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"There are a couple typos. \\\"uninitalized,\\\" in one place.\\n\\nAlso,\\n\\n+ * If this is either a bootstrap process nor a standalone backend, start\\n\\nIs \\\"either .. nor\\\" correct? It looks very wrong to me. I think you\\nmeant \\\"either .. or\\\".\\n\\nOverall, this patch seems mostly about documenting some undesirable\\nexisting behavior. Is part of your plan to get such things fixed/\\nreplaced entirely?\\n\\n-- \\nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\\n\\\"We're here to devour each other alive\\\" (Hobbes)\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Dec 2021 13:38:25 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 09:41:58AM -0500, Robert Haas wrote:\\n> Maybe I'm not understanding you properly here, but it seems like you\\n> might be forgetting that this is a local variable and thus every\\n> backend is going to have something different. In the startup process,\\n> it will be initialized by StartupXLOG(); in other processes, it's\\n> currently initialized by RecoveryInProgress(), but with this patch it\\n> wouldn't be. Either way, it's then updated by future calls to\\n> XLogInsertRecord() as required. XLOG_FPW_CHANGE records might affect\\n> the new value that gets set the next time XLogInsertRecord(), but they\\n> don't directly affect doPageWrites.\\n\\nMy main worry here is that this changes slightly the definition of\\ndoPageWrites across stable branches at the end of recovery as there\\ncould be records generated there. Note that GetFullPageWriteInfo()\\ngets called in XLogInsert(), while Insert->fullPageWrites gets updated\\nbefore CleanupAfterArchiveRecovery(). And it may influence\\nthe value of doPageWrites in the startup process.\\n--\\nMichael\",\n \"msg_date\": \"Sun, 5 Dec 2021 09:44:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On Sat, Dec 4, 2021 at 7:44 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> My main worry here is that this changes slightly the definition of\\n> doPageWrites across stable branches at the end of recovery as there\\n> could be records generated there. Note that GetFullPageWriteInfo()\\n> gets called in XLogInsert(), while Insert->fullPageWrites gets updated\\n> before CleanupAfterArchiveRecovery(). And it may influence\\n> the value of doPageWrites in the startup process.\\n\\nBut ... so what? All the code that uses it retries if the value that\\nwas tentatively used turns out to be wrong.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 7 Dec 2021 17:55:02 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 7, 2021 at 5:55 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Sat, Dec 4, 2021 at 7:44 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> > My main worry here is that this changes slightly the definition of\\n> > doPageWrites across stable branches at the end of recovery as there\\n> > could be records generated there. Note that GetFullPageWriteInfo()\\n> > gets called in XLogInsert(), while Insert->fullPageWrites gets updated\\n> > before CleanupAfterArchiveRecovery(). And it may influence\\n> > the value of doPageWrites in the startup process.\\n>\\n> But ... so what? All the code that uses it retries if the value that\\n> was tentatively used turns out to be wrong.\\n\\nDespite the fact that this question didn't get further discussion, and\\nthe fact that nobody seems quite sure what the best way of proceeding\\nhere is, my interpretation of the comments made so far is that nobody\\nthinks that what we have now is superior to either of the proposed\\nalternatives, and there's a weak preference for v2 over v1. So I\\ncommitted that one.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 13 Dec 2021 10:10:43 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RecoveryInProgress() has critical side effects\"\n }\n]","isConversational":false}}},{"rowIdx":53076,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello\n\nRecently I noticed extremely suspicious behavior from queries under JIT.\nA query (details for which sadly I cannot disclose) had an EXPLAIN\nlooking like this:\n\n JIT:\n Functions: 743\n Options: Inlining true, Optimization true, Expressions true, Deforming true\n Timing: Generation 47.013 ms, Inlining 20.277 ms, Optimization 3269.381 ms, Emission 2392.115 ms, Total 5728.786 ms\n Execution Time: 5735.552 ms\n\nDisabling JIT, the query takes a few tens of milliseconds, so of course\nthe typical recommendation is to disable JIT. But I noted that the\naverage time to optimize, per function, is 4.7ms. I didn't know what to\nmake of this -- is it a reasonable to to take to optimize a single\nfunction?\n\nI put that thought aside. But then I spotted a couple of minor things\nin the query that could be made better; so I did some of that, and\nnoticed that the number of functions was reduced from 743 to 671, and\nthe JIT details in the explain looked like this:\n\n Planning Time: 15.060 ms\n JIT:\n Functions: 671\n Options: Inlining false, Optimization true, Expressions true, Deforming true\n Timing: Generation 75.205 ms, Inlining 0.000 ms, Optimization 1387.741 ms, Emission 1110.807 ms, Total 2573.753 ms\n Execution Time: 2587.776 ms\n\nBut what really surprised me is that the the average time to optimize\nper function is now 2.06ms ... less than half of the previous\nmeasurement. It emits 10% less functions than before, but the time to\nboth optimize and emit is reduced by 50%. How does that make sense?\n\nI don't know how to approach an investigation into this, but I thought\nthat maybe we're doing quadratic effort in the optimization / emission\nphase or something like that.\n\n-- \nÁlvaro Herrera\n\n\n","msg_date":"Thu, 11 Nov 2021 17:02:09 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":true,"msg_subject":"JIT doing duplicative optimization?"},{"msg_contents":"On 2021-Nov-11, Alvaro Herrera wrote:\n\n> But what really surprised me is that the the average time to optimize\n> per function is now 2.06ms ... less than half of the previous\n> measurement. It emits 10% less functions than before, but the time to\n> both optimize and emit is reduced by 50%. How does that make sense?\n\nAh, here's a query that illustrates what I'm on about. I found this\nquery[1] in a blog post[2].\n\n```\nWITH RECURSIVE typeinfo_tree(\n oid, ns, name, kind, basetype, elemtype, elemdelim,\n range_subtype, attrtypoids, attrnames, depth)\nAS (\n SELECT\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n ti.elemtype, ti.elemdelim, ti.range_subtype,\n ti.attrtypoids, ti.attrnames, 0\n FROM\n (\n SELECT\n t.oid AS oid,\n ns.nspname AS ns,\n t.typname AS name,\n t.typtype AS kind,\n (CASE WHEN t.typtype = 'd' THEN\n (WITH RECURSIVE typebases(oid, depth) AS (\n SELECT\n t2.typbasetype AS oid,\n 0 AS depth\n FROM\n pg_type t2\n WHERE\n t2.oid = t.oid\n UNION ALL\n SELECT\n t2.typbasetype AS oid,\n tb.depth + 1 AS depth\n FROM\n pg_type t2,\n typebases tb\n WHERE\n tb.oid = t2.oid\n AND t2.typbasetype != 0\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n ELSE NULL\n END) AS basetype,\n t.typelem AS elemtype,\n elem_t.typdelim AS elemdelim,\n range_t.rngsubtype AS range_subtype,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.atttypid ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrtypoids,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.attname::text ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrnames\n FROM\n pg_catalog.pg_type AS t\n INNER JOIN pg_catalog.pg_namespace ns ON (\n ns.oid = t.typnamespace)\n LEFT JOIN pg_type elem_t ON (\n t.typlen = -1 AND\n t.typelem != 0 AND\n t.typelem = elem_t.oid\n )\n LEFT JOIN pg_range range_t ON (\n t.oid = range_t.rngtypid\n )\n ) AS ti\n WHERE\n ti.oid = any(ARRAY[16,17]::oid[])\n\n UNION ALL\n\n SELECT\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n ti.elemtype, ti.elemdelim, ti.range_subtype,\n ti.attrtypoids, ti.attrnames, tt.depth + 1\n FROM\n (\n SELECT\n t.oid AS oid,\n ns.nspname AS ns,\n t.typname AS name,\n t.typtype AS kind,\n (CASE WHEN t.typtype = 'd' THEN\n (WITH RECURSIVE typebases(oid, depth) AS (\n SELECT\n t2.typbasetype AS oid,\n 0 AS depth\n FROM\n pg_type t2\n WHERE\n t2.oid = t.oid\n UNION ALL\n SELECT\n t2.typbasetype AS oid,\n tb.depth + 1 AS depth\n FROM\n pg_type t2,\n typebases tb\n WHERE\n tb.oid = t2.oid\n AND t2.typbasetype != 0\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n ELSE NULL\n END) AS basetype,\n t.typelem AS elemtype,\n elem_t.typdelim AS elemdelim,\n range_t.rngsubtype AS range_subtype,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.atttypid ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrtypoids,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.attname::text ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrnames\n FROM\n pg_catalog.pg_type AS t\n INNER JOIN pg_catalog.pg_namespace ns ON (\n ns.oid = t.typnamespace)\n LEFT JOIN pg_type elem_t ON (\n t.typlen = -1 AND\n t.typelem != 0 AND\n t.typelem = elem_t.oid\n )\n LEFT JOIN pg_range range_t ON (\n t.oid = range_t.rngtypid\n )\n ) ti,\n typeinfo_tree tt\n WHERE\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\n)\nSELECT DISTINCT\n *,\n basetype::regtype::text AS basetype_name,\n elemtype::regtype::text AS elemtype_name,\n range_subtype::regtype::text AS range_subtype_name\nFROM\n typeinfo_tree\nORDER BY\n depth DESC;\n```\n\nI did an EXPLAIN ANALYZE and at the bottom you see this:\n\n Planning Time: 2.606 ms\n JIT:\n Functions: 148\n Options: Inlining true, Optimization true, Expressions true, Deforming true\n Timing: Generation 19.670 ms, Inlining 19.224 ms, Optimization 435.153 ms, Emission 282.216 ms, Total 756.263 ms\n Execution Time: 757.643 ms\n\nAverage time to optimize, per function 435.153/148 = 2.940ms;\naverage time to emit per function 282.216/148 = 1.906ms\n\nNow let's change the query by making the first innermost recursive CTE into\nnon-recursive; just delete the RECURSIVE keyword and everything after the UNION\nALL. You now get this query.\n\n```\nexplain analyze\nWITH RECURSIVE typeinfo_tree(\n oid, ns, name, kind, basetype, elemtype, elemdelim,\n range_subtype, attrtypoids, attrnames, depth)\nAS (\n SELECT\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n ti.elemtype, ti.elemdelim, ti.range_subtype,\n ti.attrtypoids, ti.attrnames, 0\n FROM\n (\n SELECT\n t.oid AS oid,\n ns.nspname AS ns,\n t.typname AS name,\n t.typtype AS kind,\n (CASE WHEN t.typtype = 'd' THEN\n (WITH typebases(oid, depth) AS (\n SELECT\n t2.typbasetype AS oid,\n 0 AS depth\n FROM\n pg_type t2\n WHERE\n t2.oid = t.oid\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n ELSE NULL\n END) AS basetype,\n t.typelem AS elemtype,\n elem_t.typdelim AS elemdelim,\n range_t.rngsubtype AS range_subtype,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.atttypid ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrtypoids,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.attname::text ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrnames\n FROM\n pg_catalog.pg_type AS t\n INNER JOIN pg_catalog.pg_namespace ns ON (\n ns.oid = t.typnamespace)\n LEFT JOIN pg_type elem_t ON (\n t.typlen = -1 AND\n t.typelem != 0 AND\n t.typelem = elem_t.oid\n )\n LEFT JOIN pg_range range_t ON (\n t.oid = range_t.rngtypid\n )\n ) AS ti\n WHERE\n ti.oid = any(ARRAY[16,17]::oid[])\n\n UNION ALL\n\n SELECT\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n ti.elemtype, ti.elemdelim, ti.range_subtype,\n ti.attrtypoids, ti.attrnames, tt.depth + 1\n FROM\n (\n SELECT\n t.oid AS oid,\n ns.nspname AS ns,\n t.typname AS name,\n t.typtype AS kind,\n (CASE WHEN t.typtype = 'd' THEN\n (WITH RECURSIVE typebases(oid, depth) AS (\n SELECT\n t2.typbasetype AS oid,\n 0 AS depth\n FROM\n pg_type t2\n WHERE\n t2.oid = t.oid\n UNION ALL\n SELECT\n t2.typbasetype AS oid,\n tb.depth + 1 AS depth\n FROM\n pg_type t2,\n typebases tb\n WHERE\n tb.oid = t2.oid\n AND t2.typbasetype != 0\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n ELSE NULL\n END) AS basetype,\n t.typelem AS elemtype,\n elem_t.typdelim AS elemdelim,\n range_t.rngsubtype AS range_subtype,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.atttypid ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrtypoids,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.attname::text ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrnames\n FROM\n pg_catalog.pg_type AS t\n INNER JOIN pg_catalog.pg_namespace ns ON (\n ns.oid = t.typnamespace)\n LEFT JOIN pg_type elem_t ON (\n t.typlen = -1 AND\n t.typelem != 0 AND\n t.typelem = elem_t.oid\n )\n LEFT JOIN pg_range range_t ON (\n t.oid = range_t.rngtypid\n )\n ) ti,\n typeinfo_tree tt\n WHERE\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\n)\n\nSELECT DISTINCT\n *,\n basetype::regtype::text AS basetype_name,\n elemtype::regtype::text AS elemtype_name,\n range_subtype::regtype::text AS range_subtype_name\nFROM\n typeinfo_tree\nORDER BY\n depth DESC\n``` \n\nAt the bottom of the explain you get this.\n\n Planning Time: 2.508 ms\n JIT:\n Functions: 137\n Options: Inlining true, Optimization true, Expressions true, Deforming true\n Timing: Generation 10.346 ms, Inlining 10.210 ms, Optimization 374.103 ms, Emission 254.557 ms, Total 649.216 ms\n Execution Time: 650.168 ms\n\nAverage time to optimize, per function: 374.103/137 = 2.730ms\nAverage time to emit, per function 254.557 / 137 = 1.858ms\n\n\nNow do it one more time, with the second innermost recursive CTE. You get this\nquery.\n\n```\nexplain analyze\nWITH RECURSIVE typeinfo_tree(\n oid, ns, name, kind, basetype, elemtype, elemdelim,\n range_subtype, attrtypoids, attrnames, depth)\nAS (\n SELECT\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n ti.elemtype, ti.elemdelim, ti.range_subtype,\n ti.attrtypoids, ti.attrnames, 0\n FROM\n (\n SELECT\n t.oid AS oid,\n ns.nspname AS ns,\n t.typname AS name,\n t.typtype AS kind,\n (CASE WHEN t.typtype = 'd' THEN\n (WITH typebases(oid, depth) AS (\n SELECT\n t2.typbasetype AS oid,\n 0 AS depth\n FROM\n pg_type t2\n WHERE\n t2.oid = t.oid\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n ELSE NULL\n END) AS basetype,\n t.typelem AS elemtype,\n elem_t.typdelim AS elemdelim,\n range_t.rngsubtype AS range_subtype,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.atttypid ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrtypoids,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.attname::text ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrnames\n FROM\n pg_catalog.pg_type AS t\n INNER JOIN pg_catalog.pg_namespace ns ON (\n ns.oid = t.typnamespace)\n LEFT JOIN pg_type elem_t ON (\n t.typlen = -1 AND\n t.typelem != 0 AND\n t.typelem = elem_t.oid\n )\n LEFT JOIN pg_range range_t ON (\n t.oid = range_t.rngtypid\n )\n ) AS ti\n WHERE\n ti.oid = any(ARRAY[16,17]::oid[])\n\n UNION ALL\n\n SELECT\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n ti.elemtype, ti.elemdelim, ti.range_subtype,\n ti.attrtypoids, ti.attrnames, tt.depth + 1\n FROM\n (\n SELECT\n t.oid AS oid,\n ns.nspname AS ns,\n t.typname AS name,\n t.typtype AS kind,\n (CASE WHEN t.typtype = 'd' THEN\n (WITH typebases(oid, depth) AS (\n SELECT\n t2.typbasetype AS oid,\n 0 AS depth\n FROM\n pg_type t2\n WHERE\n t2.oid = t.oid\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n ELSE NULL\n END) AS basetype,\n t.typelem AS elemtype,\n elem_t.typdelim AS elemdelim,\n range_t.rngsubtype AS range_subtype,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.atttypid ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrtypoids,\n (CASE WHEN t.typtype = 'c' THEN\n (SELECT\n array_agg(ia.attname::text ORDER BY ia.attnum)\n FROM\n pg_attribute ia\n INNER JOIN pg_class c\n ON (ia.attrelid = c.oid)\n WHERE\n ia.attnum > 0 AND NOT ia.attisdropped\n AND c.reltype = t.oid)\n ELSE NULL\n END) AS attrnames\n FROM\n pg_catalog.pg_type AS t\n INNER JOIN pg_catalog.pg_namespace ns ON (\n ns.oid = t.typnamespace)\n LEFT JOIN pg_type elem_t ON (\n t.typlen = -1 AND\n t.typelem != 0 AND\n t.typelem = elem_t.oid\n )\n LEFT JOIN pg_range range_t ON (\n t.oid = range_t.rngtypid\n )\n ) ti,\n typeinfo_tree tt\n WHERE\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\n)\n\nSELECT DISTINCT\n *,\n basetype::regtype::text AS basetype_name,\n elemtype::regtype::text AS elemtype_name,\n range_subtype::regtype::text AS range_subtype_name\nFROM\n typeinfo_tree\nORDER BY\n depth DESC;\n```\n\nAnd the JIT numbers at the bottom look like this:\n\n Planning Time: 2.041 ms\n JIT:\n Functions: 126\n Options: Inlining false, Optimization true, Expressions true, Deforming true\n Timing: Generation 10.002 ms, Inlining 0.000 ms, Optimization 229.128 ms, Emission 167.338 ms, Total 406.469 ms\n Execution Time: 407.315 ms\n\nAverage time to optimize per function 229.128 / 126 = 1.181ms\nAverage time to emit per function 167.338 / 126 = 1.328ms\n\n\nIn summary,\n\nQuery 1, 148 functions JIT-compiled.\nAverage time to optimize, per function 435.153/148 = 2.940ms;\naverage time to emit per function 282.216/148 = 1.906ms\n\nQuery 2, 137 functions JIT-compiled.\nAverage time to optimize, per function: 374.103/137 = 2.730ms\nAverage time to emit, per function 254.557 / 137 = 1.858ms\n\nQuery 3, 126 functions JIT-compiled.\nAverage time to optimize per function 229.128 / 126 = 1.181ms\nAverage time to emit per function 167.338 / 126 = 1.328ms\n\n???? Something looks very wrong here.\n\n[1] https://gist.github.com/saicitus/251ba20b211e9e73285af35e61b19580\n[2] https://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2a\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\"Java is clearly an example of money oriented programming\" (A. Stepanov)\n\n\n","msg_date":"Sun, 14 Nov 2021 14:06:46 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":true,"msg_subject":"Re: JIT doing duplicative optimization?"},{"msg_contents":"On Sun, Nov 14, 2021 at 9:07 AM Alvaro Herrera <alvherre@alvh.no-ip.org>\nwrote:\n\n> On 2021-Nov-11, Alvaro Herrera wrote:\n>\n> > But what really surprised me is that the the average time to optimize\n> > per function is now 2.06ms ... less than half of the previous\n> > measurement. It emits 10% less functions than before, but the time to\n> > both optimize and emit is reduced by 50%. How does that make sense?\n>\n> Ah, here's a query that illustrates what I'm on about. I found this\n> query[1] in a blog post[2].\n>\n> ```\n> WITH RECURSIVE typeinfo_tree(\n> oid, ns, name, kind, basetype, elemtype, elemdelim,\n> range_subtype, attrtypoids, attrnames, depth)\n> AS (\n> SELECT\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\n> ti.attrtypoids, ti.attrnames, 0\n> FROM\n> (\n> SELECT\n> t.oid AS oid,\n> ns.nspname AS ns,\n> t.typname AS name,\n> t.typtype AS kind,\n> (CASE WHEN t.typtype = 'd' THEN\n> (WITH RECURSIVE typebases(oid, depth) AS (\n> SELECT\n> t2.typbasetype AS oid,\n> 0 AS depth\n> FROM\n> pg_type t2\n> WHERE\n> t2.oid = t.oid\n> UNION ALL\n> SELECT\n> t2.typbasetype AS oid,\n> tb.depth + 1 AS depth\n> FROM\n> pg_type t2,\n> typebases tb\n> WHERE\n> tb.oid = t2.oid\n> AND t2.typbasetype != 0\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n> ELSE NULL\n> END) AS basetype,\n> t.typelem AS elemtype,\n> elem_t.typdelim AS elemdelim,\n> range_t.rngsubtype AS range_subtype,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.atttypid ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrtypoids,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.attname::text ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrnames\n> FROM\n> pg_catalog.pg_type AS t\n> INNER JOIN pg_catalog.pg_namespace ns ON (\n> ns.oid = t.typnamespace)\n> LEFT JOIN pg_type elem_t ON (\n> t.typlen = -1 AND\n> t.typelem != 0 AND\n> t.typelem = elem_t.oid\n> )\n> LEFT JOIN pg_range range_t ON (\n> t.oid = range_t.rngtypid\n> )\n> ) AS ti\n> WHERE\n> ti.oid = any(ARRAY[16,17]::oid[])\n>\n> UNION ALL\n>\n> SELECT\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\n> ti.attrtypoids, ti.attrnames, tt.depth + 1\n> FROM\n> (\n> SELECT\n> t.oid AS oid,\n> ns.nspname AS ns,\n> t.typname AS name,\n> t.typtype AS kind,\n> (CASE WHEN t.typtype = 'd' THEN\n> (WITH RECURSIVE typebases(oid, depth) AS (\n> SELECT\n> t2.typbasetype AS oid,\n> 0 AS depth\n> FROM\n> pg_type t2\n> WHERE\n> t2.oid = t.oid\n> UNION ALL\n> SELECT\n> t2.typbasetype AS oid,\n> tb.depth + 1 AS depth\n> FROM\n> pg_type t2,\n> typebases tb\n> WHERE\n> tb.oid = t2.oid\n> AND t2.typbasetype != 0\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n> ELSE NULL\n> END) AS basetype,\n> t.typelem AS elemtype,\n> elem_t.typdelim AS elemdelim,\n> range_t.rngsubtype AS range_subtype,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.atttypid ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrtypoids,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.attname::text ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrnames\n> FROM\n> pg_catalog.pg_type AS t\n> INNER JOIN pg_catalog.pg_namespace ns ON (\n> ns.oid = t.typnamespace)\n> LEFT JOIN pg_type elem_t ON (\n> t.typlen = -1 AND\n> t.typelem != 0 AND\n> t.typelem = elem_t.oid\n> )\n> LEFT JOIN pg_range range_t ON (\n> t.oid = range_t.rngtypid\n> )\n> ) ti,\n> typeinfo_tree tt\n> WHERE\n> (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\n> OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\n> OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\n> )\n> SELECT DISTINCT\n> *,\n> basetype::regtype::text AS basetype_name,\n> elemtype::regtype::text AS elemtype_name,\n> range_subtype::regtype::text AS range_subtype_name\n> FROM\n> typeinfo_tree\n> ORDER BY\n> depth DESC;\n> ```\n>\n> I did an EXPLAIN ANALYZE and at the bottom you see this:\n>\n> Planning Time: 2.606 ms\n> JIT:\n> Functions: 148\n> Options: Inlining true, Optimization true, Expressions true, Deforming\n> true\n> Timing: Generation 19.670 ms, Inlining 19.224 ms, Optimization 435.153\n> ms, Emission 282.216 ms, Total 756.263 ms\n> Execution Time: 757.643 ms\n>\n> Average time to optimize, per function 435.153/148 = 2.940ms;\n> average time to emit per function 282.216/148 = 1.906ms\n>\n> Now let's change the query by making the first innermost recursive CTE into\n> non-recursive; just delete the RECURSIVE keyword and everything after the\n> UNION\n> ALL. You now get this query.\n>\n> ```\n> explain analyze\n> WITH RECURSIVE typeinfo_tree(\n> oid, ns, name, kind, basetype, elemtype, elemdelim,\n> range_subtype, attrtypoids, attrnames, depth)\n> AS (\n> SELECT\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\n> ti.attrtypoids, ti.attrnames, 0\n> FROM\n> (\n> SELECT\n> t.oid AS oid,\n> ns.nspname AS ns,\n> t.typname AS name,\n> t.typtype AS kind,\n> (CASE WHEN t.typtype = 'd' THEN\n> (WITH typebases(oid, depth) AS (\n> SELECT\n> t2.typbasetype AS oid,\n> 0 AS depth\n> FROM\n> pg_type t2\n> WHERE\n> t2.oid = t.oid\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n> ELSE NULL\n> END) AS basetype,\n> t.typelem AS elemtype,\n> elem_t.typdelim AS elemdelim,\n> range_t.rngsubtype AS range_subtype,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.atttypid ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrtypoids,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.attname::text ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrnames\n> FROM\n> pg_catalog.pg_type AS t\n> INNER JOIN pg_catalog.pg_namespace ns ON (\n> ns.oid = t.typnamespace)\n> LEFT JOIN pg_type elem_t ON (\n> t.typlen = -1 AND\n> t.typelem != 0 AND\n> t.typelem = elem_t.oid\n> )\n> LEFT JOIN pg_range range_t ON (\n> t.oid = range_t.rngtypid\n> )\n> ) AS ti\n> WHERE\n> ti.oid = any(ARRAY[16,17]::oid[])\n>\n> UNION ALL\n>\n> SELECT\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\n> ti.attrtypoids, ti.attrnames, tt.depth + 1\n> FROM\n> (\n> SELECT\n> t.oid AS oid,\n> ns.nspname AS ns,\n> t.typname AS name,\n> t.typtype AS kind,\n> (CASE WHEN t.typtype = 'd' THEN\n> (WITH RECURSIVE typebases(oid, depth) AS (\n> SELECT\n> t2.typbasetype AS oid,\n> 0 AS depth\n> FROM\n> pg_type t2\n> WHERE\n> t2.oid = t.oid\n> UNION ALL\n> SELECT\n> t2.typbasetype AS oid,\n> tb.depth + 1 AS depth\n> FROM\n> pg_type t2,\n> typebases tb\n> WHERE\n> tb.oid = t2.oid\n> AND t2.typbasetype != 0\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n> ELSE NULL\n> END) AS basetype,\n> t.typelem AS elemtype,\n> elem_t.typdelim AS elemdelim,\n> range_t.rngsubtype AS range_subtype,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.atttypid ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrtypoids,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.attname::text ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrnames\n> FROM\n> pg_catalog.pg_type AS t\n> INNER JOIN pg_catalog.pg_namespace ns ON (\n> ns.oid = t.typnamespace)\n> LEFT JOIN pg_type elem_t ON (\n> t.typlen = -1 AND\n> t.typelem != 0 AND\n> t.typelem = elem_t.oid\n> )\n> LEFT JOIN pg_range range_t ON (\n> t.oid = range_t.rngtypid\n> )\n> ) ti,\n> typeinfo_tree tt\n> WHERE\n> (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\n> OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\n> OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\n> )\n>\n> SELECT DISTINCT\n> *,\n> basetype::regtype::text AS basetype_name,\n> elemtype::regtype::text AS elemtype_name,\n> range_subtype::regtype::text AS range_subtype_name\n> FROM\n> typeinfo_tree\n> ORDER BY\n> depth DESC\n> ```\n>\n> At the bottom of the explain you get this.\n>\n> Planning Time: 2.508 ms\n> JIT:\n> Functions: 137\n> Options: Inlining true, Optimization true, Expressions true, Deforming\n> true\n> Timing: Generation 10.346 ms, Inlining 10.210 ms, Optimization 374.103\n> ms, Emission 254.557 ms, Total 649.216 ms\n> Execution Time: 650.168 ms\n>\n> Average time to optimize, per function: 374.103/137 = 2.730ms\n> Average time to emit, per function 254.557 / 137 = 1.858ms\n>\n>\n> Now do it one more time, with the second innermost recursive CTE. You get\n> this\n> query.\n>\n> ```\n> explain analyze\n> WITH RECURSIVE typeinfo_tree(\n> oid, ns, name, kind, basetype, elemtype, elemdelim,\n> range_subtype, attrtypoids, attrnames, depth)\n> AS (\n> SELECT\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\n> ti.attrtypoids, ti.attrnames, 0\n> FROM\n> (\n> SELECT\n> t.oid AS oid,\n> ns.nspname AS ns,\n> t.typname AS name,\n> t.typtype AS kind,\n> (CASE WHEN t.typtype = 'd' THEN\n> (WITH typebases(oid, depth) AS (\n> SELECT\n> t2.typbasetype AS oid,\n> 0 AS depth\n> FROM\n> pg_type t2\n> WHERE\n> t2.oid = t.oid\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n> ELSE NULL\n> END) AS basetype,\n> t.typelem AS elemtype,\n> elem_t.typdelim AS elemdelim,\n> range_t.rngsubtype AS range_subtype,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.atttypid ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrtypoids,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.attname::text ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrnames\n> FROM\n> pg_catalog.pg_type AS t\n> INNER JOIN pg_catalog.pg_namespace ns ON (\n> ns.oid = t.typnamespace)\n> LEFT JOIN pg_type elem_t ON (\n> t.typlen = -1 AND\n> t.typelem != 0 AND\n> t.typelem = elem_t.oid\n> )\n> LEFT JOIN pg_range range_t ON (\n> t.oid = range_t.rngtypid\n> )\n> ) AS ti\n> WHERE\n> ti.oid = any(ARRAY[16,17]::oid[])\n>\n> UNION ALL\n>\n> SELECT\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\n> ti.attrtypoids, ti.attrnames, tt.depth + 1\n> FROM\n> (\n> SELECT\n> t.oid AS oid,\n> ns.nspname AS ns,\n> t.typname AS name,\n> t.typtype AS kind,\n> (CASE WHEN t.typtype = 'd' THEN\n> (WITH typebases(oid, depth) AS (\n> SELECT\n> t2.typbasetype AS oid,\n> 0 AS depth\n> FROM\n> pg_type t2\n> WHERE\n> t2.oid = t.oid\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\n> ELSE NULL\n> END) AS basetype,\n> t.typelem AS elemtype,\n> elem_t.typdelim AS elemdelim,\n> range_t.rngsubtype AS range_subtype,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.atttypid ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrtypoids,\n> (CASE WHEN t.typtype = 'c' THEN\n> (SELECT\n> array_agg(ia.attname::text ORDER BY ia.attnum)\n> FROM\n> pg_attribute ia\n> INNER JOIN pg_class c\n> ON (ia.attrelid = c.oid)\n> WHERE\n> ia.attnum > 0 AND NOT ia.attisdropped\n> AND c.reltype = t.oid)\n> ELSE NULL\n> END) AS attrnames\n> FROM\n> pg_catalog.pg_type AS t\n> INNER JOIN pg_catalog.pg_namespace ns ON (\n> ns.oid = t.typnamespace)\n> LEFT JOIN pg_type elem_t ON (\n> t.typlen = -1 AND\n> t.typelem != 0 AND\n> t.typelem = elem_t.oid\n> )\n> LEFT JOIN pg_range range_t ON (\n> t.oid = range_t.rngtypid\n> )\n> ) ti,\n> typeinfo_tree tt\n> WHERE\n> (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\n> OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\n> OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\n> )\n>\n> SELECT DISTINCT\n> *,\n> basetype::regtype::text AS basetype_name,\n> elemtype::regtype::text AS elemtype_name,\n> range_subtype::regtype::text AS range_subtype_name\n> FROM\n> typeinfo_tree\n> ORDER BY\n> depth DESC;\n> ```\n>\n> And the JIT numbers at the bottom look like this:\n>\n> Planning Time: 2.041 ms\n> JIT:\n> Functions: 126\n> Options: Inlining false, Optimization true, Expressions true, Deforming\n> true\n> Timing: Generation 10.002 ms, Inlining 0.000 ms, Optimization 229.128\n> ms, Emission 167.338 ms, Total 406.469 ms\n> Execution Time: 407.315 ms\n>\n> Average time to optimize per function 229.128 / 126 = 1.181ms\n> Average time to emit per function 167.338 / 126 = 1.328ms\n>\n>\n> In summary,\n>\n> Query 1, 148 functions JIT-compiled.\n> Average time to optimize, per function 435.153/148 = 2.940ms;\n> average time to emit per function 282.216/148 = 1.906ms\n>\n> Query 2, 137 functions JIT-compiled.\n> Average time to optimize, per function: 374.103/137 = 2.730ms\n> Average time to emit, per function 254.557 / 137 = 1.858ms\n>\n> Query 3, 126 functions JIT-compiled.\n> Average time to optimize per function 229.128 / 126 = 1.181ms\n> Average time to emit per function 167.338 / 126 = 1.328ms\n>\n> ???? Something looks very wrong here.\n>\n> [1] https://gist.github.com/saicitus/251ba20b211e9e73285af35e61b19580\n> [2]\n> https://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2a\n>\n> --\n> Álvaro Herrera PostgreSQL Developer —\n> https://www.EnterpriseDB.com/\n> \"Java is clearly an example of money oriented programming\" (A. Stepanov)\n>\n>\n> Hi,\nFor the 3rd query, I noticed some difference in options:\n\n Options: Inlining false, Optimization true, Expressions true, Deforming\ntrue\n\n Inlining was true for the first two queries.\n\nFYI\n\nOn Sun, Nov 14, 2021 at 9:07 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:On 2021-Nov-11, Alvaro Herrera wrote:\n\r\n> But what really surprised me is that the the average time to optimize\r\n> per function is now 2.06ms ... less than half of the previous\r\n> measurement. It emits 10% less functions than before, but the time to\r\n> both optimize and emit is reduced by 50%. How does that make sense?\n\r\nAh, here's a query that illustrates what I'm on about. I found this\r\nquery[1] in a blog post[2].\n\r\n```\r\nWITH RECURSIVE typeinfo_tree(\r\n oid, ns, name, kind, basetype, elemtype, elemdelim,\r\n range_subtype, attrtypoids, attrnames, depth)\r\nAS (\r\n SELECT\r\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\r\n ti.elemtype, ti.elemdelim, ti.range_subtype,\r\n ti.attrtypoids, ti.attrnames, 0\r\n FROM\r\n (\r\n SELECT\r\n t.oid AS oid,\r\n ns.nspname AS ns,\r\n t.typname AS name,\r\n t.typtype AS kind,\r\n (CASE WHEN t.typtype = 'd' THEN\r\n (WITH RECURSIVE typebases(oid, depth) AS (\r\n SELECT\r\n t2.typbasetype AS oid,\r\n 0 AS depth\r\n FROM\r\n pg_type t2\r\n WHERE\r\n t2.oid = t.oid\r\n UNION ALL\r\n SELECT\r\n t2.typbasetype AS oid,\r\n tb.depth + 1 AS depth\r\n FROM\r\n pg_type t2,\r\n typebases tb\r\n WHERE\r\n tb.oid = t2.oid\r\n AND t2.typbasetype != 0\r\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\r\n ELSE NULL\r\n END) AS basetype,\r\n t.typelem AS elemtype,\r\n elem_t.typdelim AS elemdelim,\r\n range_t.rngsubtype AS range_subtype,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.atttypid ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrtypoids,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.attname::text ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrnames\r\n FROM\r\n pg_catalog.pg_type AS t\r\n INNER JOIN pg_catalog.pg_namespace ns ON (\r\n ns.oid = t.typnamespace)\r\n LEFT JOIN pg_type elem_t ON (\r\n t.typlen = -1 AND\r\n t.typelem != 0 AND\r\n t.typelem = elem_t.oid\r\n )\r\n LEFT JOIN pg_range range_t ON (\r\n t.oid = range_t.rngtypid\r\n )\r\n ) AS ti\r\n WHERE\r\n ti.oid = any(ARRAY[16,17]::oid[])\n\r\n UNION ALL\n\r\n SELECT\r\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\r\n ti.elemtype, ti.elemdelim, ti.range_subtype,\r\n ti.attrtypoids, ti.attrnames, tt.depth + 1\r\n FROM\r\n (\r\n SELECT\r\n t.oid AS oid,\r\n ns.nspname AS ns,\r\n t.typname AS name,\r\n t.typtype AS kind,\r\n (CASE WHEN t.typtype = 'd' THEN\r\n (WITH RECURSIVE typebases(oid, depth) AS (\r\n SELECT\r\n t2.typbasetype AS oid,\r\n 0 AS depth\r\n FROM\r\n pg_type t2\r\n WHERE\r\n t2.oid = t.oid\r\n UNION ALL\r\n SELECT\r\n t2.typbasetype AS oid,\r\n tb.depth + 1 AS depth\r\n FROM\r\n pg_type t2,\r\n typebases tb\r\n WHERE\r\n tb.oid = t2.oid\r\n AND t2.typbasetype != 0\r\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\r\n ELSE NULL\r\n END) AS basetype,\r\n t.typelem AS elemtype,\r\n elem_t.typdelim AS elemdelim,\r\n range_t.rngsubtype AS range_subtype,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.atttypid ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrtypoids,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.attname::text ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrnames\r\n FROM\r\n pg_catalog.pg_type AS t\r\n INNER JOIN pg_catalog.pg_namespace ns ON (\r\n ns.oid = t.typnamespace)\r\n LEFT JOIN pg_type elem_t ON (\r\n t.typlen = -1 AND\r\n t.typelem != 0 AND\r\n t.typelem = elem_t.oid\r\n )\r\n LEFT JOIN pg_range range_t ON (\r\n t.oid = range_t.rngtypid\r\n )\r\n ) ti,\r\n typeinfo_tree tt\r\n WHERE\r\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\r\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\r\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\r\n)\r\nSELECT DISTINCT\r\n *,\r\n basetype::regtype::text AS basetype_name,\r\n elemtype::regtype::text AS elemtype_name,\r\n range_subtype::regtype::text AS range_subtype_name\r\nFROM\r\n typeinfo_tree\r\nORDER BY\r\n depth DESC;\r\n```\n\r\nI did an EXPLAIN ANALYZE and at the bottom you see this:\n\r\n Planning Time: 2.606 ms\r\n JIT:\r\n Functions: 148\r\n Options: Inlining true, Optimization true, Expressions true, Deforming true\r\n Timing: Generation 19.670 ms, Inlining 19.224 ms, Optimization 435.153 ms, Emission 282.216 ms, Total 756.263 ms\r\n Execution Time: 757.643 ms\n\r\nAverage time to optimize, per function 435.153/148 = 2.940ms;\r\naverage time to emit per function 282.216/148 = 1.906ms\n\r\nNow let's change the query by making the first innermost recursive CTE into\r\nnon-recursive; just delete the RECURSIVE keyword and everything after the UNION\r\nALL. You now get this query.\n\r\n```\r\nexplain analyze\r\nWITH RECURSIVE typeinfo_tree(\r\n oid, ns, name, kind, basetype, elemtype, elemdelim,\r\n range_subtype, attrtypoids, attrnames, depth)\r\nAS (\r\n SELECT\r\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\r\n ti.elemtype, ti.elemdelim, ti.range_subtype,\r\n ti.attrtypoids, ti.attrnames, 0\r\n FROM\r\n (\r\n SELECT\r\n t.oid AS oid,\r\n ns.nspname AS ns,\r\n t.typname AS name,\r\n t.typtype AS kind,\r\n (CASE WHEN t.typtype = 'd' THEN\r\n (WITH typebases(oid, depth) AS (\r\n SELECT\r\n t2.typbasetype AS oid,\r\n 0 AS depth\r\n FROM\r\n pg_type t2\r\n WHERE\r\n t2.oid = t.oid\r\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\r\n ELSE NULL\r\n END) AS basetype,\r\n t.typelem AS elemtype,\r\n elem_t.typdelim AS elemdelim,\r\n range_t.rngsubtype AS range_subtype,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.atttypid ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrtypoids,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.attname::text ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrnames\r\n FROM\r\n pg_catalog.pg_type AS t\r\n INNER JOIN pg_catalog.pg_namespace ns ON (\r\n ns.oid = t.typnamespace)\r\n LEFT JOIN pg_type elem_t ON (\r\n t.typlen = -1 AND\r\n t.typelem != 0 AND\r\n t.typelem = elem_t.oid\r\n )\r\n LEFT JOIN pg_range range_t ON (\r\n t.oid = range_t.rngtypid\r\n )\r\n ) AS ti\r\n WHERE\r\n ti.oid = any(ARRAY[16,17]::oid[])\n\r\n UNION ALL\n\r\n SELECT\r\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\r\n ti.elemtype, ti.elemdelim, ti.range_subtype,\r\n ti.attrtypoids, ti.attrnames, tt.depth + 1\r\n FROM\r\n (\r\n SELECT\r\n t.oid AS oid,\r\n ns.nspname AS ns,\r\n t.typname AS name,\r\n t.typtype AS kind,\r\n (CASE WHEN t.typtype = 'd' THEN\r\n (WITH RECURSIVE typebases(oid, depth) AS (\r\n SELECT\r\n t2.typbasetype AS oid,\r\n 0 AS depth\r\n FROM\r\n pg_type t2\r\n WHERE\r\n t2.oid = t.oid\r\n UNION ALL\r\n SELECT\r\n t2.typbasetype AS oid,\r\n tb.depth + 1 AS depth\r\n FROM\r\n pg_type t2,\r\n typebases tb\r\n WHERE\r\n tb.oid = t2.oid\r\n AND t2.typbasetype != 0\r\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\r\n ELSE NULL\r\n END) AS basetype,\r\n t.typelem AS elemtype,\r\n elem_t.typdelim AS elemdelim,\r\n range_t.rngsubtype AS range_subtype,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.atttypid ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrtypoids,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.attname::text ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrnames\r\n FROM\r\n pg_catalog.pg_type AS t\r\n INNER JOIN pg_catalog.pg_namespace ns ON (\r\n ns.oid = t.typnamespace)\r\n LEFT JOIN pg_type elem_t ON (\r\n t.typlen = -1 AND\r\n t.typelem != 0 AND\r\n t.typelem = elem_t.oid\r\n )\r\n LEFT JOIN pg_range range_t ON (\r\n t.oid = range_t.rngtypid\r\n )\r\n ) ti,\r\n typeinfo_tree tt\r\n WHERE\r\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\r\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\r\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\r\n)\n\r\nSELECT DISTINCT\r\n *,\r\n basetype::regtype::text AS basetype_name,\r\n elemtype::regtype::text AS elemtype_name,\r\n range_subtype::regtype::text AS range_subtype_name\r\nFROM\r\n typeinfo_tree\r\nORDER BY\r\n depth DESC\r\n``` \n\r\nAt the bottom of the explain you get this.\n\r\n Planning Time: 2.508 ms\r\n JIT:\r\n Functions: 137\r\n Options: Inlining true, Optimization true, Expressions true, Deforming true\r\n Timing: Generation 10.346 ms, Inlining 10.210 ms, Optimization 374.103 ms, Emission 254.557 ms, Total 649.216 ms\r\n Execution Time: 650.168 ms\n\r\nAverage time to optimize, per function: 374.103/137 = 2.730ms\r\nAverage time to emit, per function 254.557 / 137 = 1.858ms\n\n\r\nNow do it one more time, with the second innermost recursive CTE. You get this\r\nquery.\n\r\n```\r\nexplain analyze\r\nWITH RECURSIVE typeinfo_tree(\r\n oid, ns, name, kind, basetype, elemtype, elemdelim,\r\n range_subtype, attrtypoids, attrnames, depth)\r\nAS (\r\n SELECT\r\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\r\n ti.elemtype, ti.elemdelim, ti.range_subtype,\r\n ti.attrtypoids, ti.attrnames, 0\r\n FROM\r\n (\r\n SELECT\r\n t.oid AS oid,\r\n ns.nspname AS ns,\r\n t.typname AS name,\r\n t.typtype AS kind,\r\n (CASE WHEN t.typtype = 'd' THEN\r\n (WITH typebases(oid, depth) AS (\r\n SELECT\r\n t2.typbasetype AS oid,\r\n 0 AS depth\r\n FROM\r\n pg_type t2\r\n WHERE\r\n t2.oid = t.oid\r\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\r\n ELSE NULL\r\n END) AS basetype,\r\n t.typelem AS elemtype,\r\n elem_t.typdelim AS elemdelim,\r\n range_t.rngsubtype AS range_subtype,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.atttypid ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrtypoids,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.attname::text ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrnames\r\n FROM\r\n pg_catalog.pg_type AS t\r\n INNER JOIN pg_catalog.pg_namespace ns ON (\r\n ns.oid = t.typnamespace)\r\n LEFT JOIN pg_type elem_t ON (\r\n t.typlen = -1 AND\r\n t.typelem != 0 AND\r\n t.typelem = elem_t.oid\r\n )\r\n LEFT JOIN pg_range range_t ON (\r\n t.oid = range_t.rngtypid\r\n )\r\n ) AS ti\r\n WHERE\r\n ti.oid = any(ARRAY[16,17]::oid[])\n\r\n UNION ALL\n\r\n SELECT\r\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\r\n ti.elemtype, ti.elemdelim, ti.range_subtype,\r\n ti.attrtypoids, ti.attrnames, tt.depth + 1\r\n FROM\r\n (\r\n SELECT\r\n t.oid AS oid,\r\n ns.nspname AS ns,\r\n t.typname AS name,\r\n t.typtype AS kind,\r\n (CASE WHEN t.typtype = 'd' THEN\r\n (WITH typebases(oid, depth) AS (\r\n SELECT\r\n t2.typbasetype AS oid,\r\n 0 AS depth\r\n FROM\r\n pg_type t2\r\n WHERE\r\n t2.oid = t.oid\r\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\r\n ELSE NULL\r\n END) AS basetype,\r\n t.typelem AS elemtype,\r\n elem_t.typdelim AS elemdelim,\r\n range_t.rngsubtype AS range_subtype,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.atttypid ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrtypoids,\r\n (CASE WHEN t.typtype = 'c' THEN\r\n (SELECT\r\n array_agg(ia.attname::text ORDER BY ia.attnum)\r\n FROM\r\n pg_attribute ia\r\n INNER JOIN pg_class c\r\n ON (ia.attrelid = c.oid)\r\n WHERE\r\n ia.attnum > 0 AND NOT ia.attisdropped\r\n AND c.reltype = t.oid)\r\n ELSE NULL\r\n END) AS attrnames\r\n FROM\r\n pg_catalog.pg_type AS t\r\n INNER JOIN pg_catalog.pg_namespace ns ON (\r\n ns.oid = t.typnamespace)\r\n LEFT JOIN pg_type elem_t ON (\r\n t.typlen = -1 AND\r\n t.typelem != 0 AND\r\n t.typelem = elem_t.oid\r\n )\r\n LEFT JOIN pg_range range_t ON (\r\n t.oid = range_t.rngtypid\r\n )\r\n ) ti,\r\n typeinfo_tree tt\r\n WHERE\r\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\r\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\r\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\r\n)\n\r\nSELECT DISTINCT\r\n *,\r\n basetype::regtype::text AS basetype_name,\r\n elemtype::regtype::text AS elemtype_name,\r\n range_subtype::regtype::text AS range_subtype_name\r\nFROM\r\n typeinfo_tree\r\nORDER BY\r\n depth DESC;\r\n```\n\r\nAnd the JIT numbers at the bottom look like this:\n\r\n Planning Time: 2.041 ms\r\n JIT:\r\n Functions: 126\r\n Options: Inlining false, Optimization true, Expressions true, Deforming true\r\n Timing: Generation 10.002 ms, Inlining 0.000 ms, Optimization 229.128 ms, Emission 167.338 ms, Total 406.469 ms\r\n Execution Time: 407.315 ms\n\r\nAverage time to optimize per function 229.128 / 126 = 1.181ms\r\nAverage time to emit per function 167.338 / 126 = 1.328ms\n\n\r\nIn summary,\n\r\nQuery 1, 148 functions JIT-compiled.\r\nAverage time to optimize, per function 435.153/148 = 2.940ms;\r\naverage time to emit per function 282.216/148 = 1.906ms\n\r\nQuery 2, 137 functions JIT-compiled.\r\nAverage time to optimize, per function: 374.103/137 = 2.730ms\r\nAverage time to emit, per function 254.557 / 137 = 1.858ms\n\r\nQuery 3, 126 functions JIT-compiled.\r\nAverage time to optimize per function 229.128 / 126 = 1.181ms\r\nAverage time to emit per function 167.338 / 126 = 1.328ms\n\r\n???? Something looks very wrong here.\n\r\n[1] https://gist.github.com/saicitus/251ba20b211e9e73285af35e61b19580\r\n[2] https://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2a\n\r\n-- \r\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\r\n\"Java is clearly an example of money oriented programming\" (A. Stepanov)\n\nHi,For the 3rd query, I noticed some difference in options: Options: Inlining false, Optimization true, Expressions true, Deforming true Inlining was true for the first two queries.FYI","msg_date":"Sun, 14 Nov 2021 09:22:04 -0800","msg_from":"Zhihong Yu <zyu@yugabyte.com>","msg_from_op":false,"msg_subject":"Re: JIT doing duplicative optimization?"},{"msg_contents":"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> On 2021-Nov-11, Alvaro Herrera wrote:\n>> But what really surprised me is that the the average time to optimize\n>> per function is now 2.06ms ... less than half of the previous\n>> measurement. It emits 10% less functions than before, but the time to\n>> both optimize and emit is reduced by 50%. How does that make sense?\n\n> Ah, here's a query that illustrates what I'm on about. I found this\n> query[1] in a blog post[2].\n> ...\n\n> Query 1, 148 functions JIT-compiled.\n> Average time to optimize, per function 435.153/148 = 2.940ms;\n> average time to emit per function 282.216/148 = 1.906ms\n\n> Query 2, 137 functions JIT-compiled.\n> Average time to optimize, per function: 374.103/137 = 2.730ms\n> Average time to emit, per function 254.557 / 137 = 1.858ms\n\n> Query 3, 126 functions JIT-compiled.\n> Average time to optimize per function 229.128 / 126 = 1.181ms\n> Average time to emit per function 167.338 / 126 = 1.328ms\n\nYeah, in combination with your other measurement, it sure does look like\nthere's something worse-than-linear going on here. The alternative is to\nassume that the individual functions are more complex in one query than\nthe other, and that seems like a bit of a stretch.\n\nYou could probably generate some queries with lots and lots of expressions\nto characterize this better. If it is O(N^2), it should not be hard to\ndrive the cost up to the point where the guilty bit of code would stand\nout in a perf trace.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 14 Nov 2021 13:30:35 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: JIT doing duplicative optimization?"},{"msg_contents":"I wrote:\n> You could probably generate some queries with lots and lots of expressions\n> to characterize this better. If it is O(N^2), it should not be hard to\n> drive the cost up to the point where the guilty bit of code would stand\n> out in a perf trace.\n\nI experimented with that, using a few different-size queries generated\nlike this:\n\nprint \"explain analyze\\n\";\nfor (my $i = 1; $i < 100; $i++) {\n\tprint \" select sum(f1+$i) from base union all\\n\";\n}\nprint \"select sum(f1+0) from base;\\n\";\n\non a table made like\n\ncreate table base as select generate_series(1,10000000) f1;\n\nWhat I got, after setting max_parallel_workers_per_gather = 0,\nwas\n\n10 subqueries:\n\n Planning Time: 0.260 ms\n JIT:\n Functions: 30\n Options: Inlining true, Optimization true, Expressions true, Deforming true\n Timing: Generation 4.651 ms, Inlining 8.870 ms, Optimization 152.937 ms, Emis\nsion 95.046 ms, Total 261.504 ms\n Execution Time: 15258.249 ms\n\n100 subqueries:\n\n Planning Time: 2.231 ms\n JIT:\n Functions: 300\n Options: Inlining true, Optimization true, Expressions true, Deforming true\n Timing: Generation 44.163 ms, Inlining 9.934 ms, Optimization 1448.971 ms, Em\nission 928.438 ms, Total 2431.506 ms\n Execution Time: 154815.515 ms\n\n1000 subqueries:\n\n Planning Time: 29.480 ms\n JIT:\n Functions: 3000\n Options: Inlining true, Optimization true, Expressions true, Deforming true\n Timing: Generation 444.479 ms, Inlining 25.688 ms, Optimization 14989.696 ms,\n Emission 9891.993 ms, Total 25351.856 ms\n Execution Time: 1522011.367 ms\n\nSo the overhead looks pretty linear, or even a shade sublinear for the\n\"inlining\" bit, *as long as only one process is involved*. However,\nI noted that if I didn't force that, the JIT overhead went up because\nthe planner wanted to use more workers and each worker has to do its own\ncompilations. So perhaps the apparent nonlinearity in your examples comes\nfrom that?\n\nBTW, I realized while working on this that I have little idea what the\n\"Functions:\" count is. Nor does our documentation explain that (or any\nother of these numbers), at least not anywhere I could find. That seems\nlike a pretty serious documentation fail. If these numbers aren't\nimportant enough to explain in the docs, why are we printing them at all?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 14 Nov 2021 18:46:34 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: JIT doing duplicative optimization?"},{"msg_contents":"Hi,\n\nOn 2021-11-11 17:02:09 -0300, Alvaro Herrera wrote:\n> Recently I noticed extremely suspicious behavior from queries under JIT.\n> A query (details for which sadly I cannot disclose) had an EXPLAIN\n> looking like this:\n> \n> JIT:\n> Functions: 743\n> Options: Inlining true, Optimization true, Expressions true, Deforming true\n> Timing: Generation 47.013 ms, Inlining 20.277 ms, Optimization 3269.381 ms, Emission 2392.115 ms, Total 5728.786 ms\n> Execution Time: 5735.552 ms\n> \n> Disabling JIT, the query takes a few tens of milliseconds, so of course\n> the typical recommendation is to disable JIT. But I noted that the\n> average time to optimize, per function, is 4.7ms. I didn't know what to\n> make of this -- is it a reasonable to to take to optimize a single\n> function?\n\nNo, that's pretty insane. Any chance you share the generated bitcode for this\n- it'll have some details about the query / table structure though. Perhaps\njust the contained function names?\n\n\n> I put that thought aside. But then I spotted a couple of minor things\n> in the query that could be made better; so I did some of that, and\n> noticed that the number of functions was reduced from 743 to 671, and\n> the JIT details in the explain looked like this:\n\n> But what really surprised me is that the the average time to optimize\n> per function is now 2.06ms ... less than half of the previous\n> measurement. It emits 10% less functions than before, but the time to\n> both optimize and emit is reduced by 50%. How does that make sense?\n\nWithout knowing the concrete query I don't know if it it's that. But\n\n\n> I don't know how to approach an investigation into this, but I thought\n> that maybe we're doing quadratic effort in the optimization / emission\n> phase or something like that.\n\nYea. One way to investigate it is to enable jit_dump_bitcode and then use\nllvm's 'opt' tool to see details about optimization times.\n\nopt -time-passes -O3 -o /dev/null < /srv/dev/pgdev-dev/3990041.77.bc |less\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 15 Nov 2021 09:13:46 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: JIT doing duplicative optimization?"},{"msg_contents":"Hi,\n\nOn 2021-11-14 18:46:34 -0500, Tom Lane wrote:\n> BTW, I realized while working on this that I have little idea what the\n> \"Functions:\" count is. Nor does our documentation explain that (or any\n> other of these numbers), at least not anywhere I could find. That seems\n> like a pretty serious documentation fail.\n\nIt's literally the number of functions created in the JIT code. That could be\nfunctions for evaluating one expression, or functions for deforming tuples of\none type of tupledesc. It's the simplest measure of \"amount of JITed stuff\" I\ncould find. I'd proposed a more detailed explain option where we'd optionally\nprint details about what exactly is JITed where a while back, but there were\nsome explain output format challenges...\n\n\n> If these numbers aren't important enough to explain in the docs, why are we\n> printing them at all?\n\nAfaict we don't have detailed docs for EXPLAIN output in general?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 15 Nov 2021 09:18:27 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: JIT doing duplicative optimization?"}],"string":"[\n {\n \"msg_contents\": \"Hello\\n\\nRecently I noticed extremely suspicious behavior from queries under JIT.\\nA query (details for which sadly I cannot disclose) had an EXPLAIN\\nlooking like this:\\n\\n JIT:\\n Functions: 743\\n Options: Inlining true, Optimization true, Expressions true, Deforming true\\n Timing: Generation 47.013 ms, Inlining 20.277 ms, Optimization 3269.381 ms, Emission 2392.115 ms, Total 5728.786 ms\\n Execution Time: 5735.552 ms\\n\\nDisabling JIT, the query takes a few tens of milliseconds, so of course\\nthe typical recommendation is to disable JIT. But I noted that the\\naverage time to optimize, per function, is 4.7ms. I didn't know what to\\nmake of this -- is it a reasonable to to take to optimize a single\\nfunction?\\n\\nI put that thought aside. But then I spotted a couple of minor things\\nin the query that could be made better; so I did some of that, and\\nnoticed that the number of functions was reduced from 743 to 671, and\\nthe JIT details in the explain looked like this:\\n\\n Planning Time: 15.060 ms\\n JIT:\\n Functions: 671\\n Options: Inlining false, Optimization true, Expressions true, Deforming true\\n Timing: Generation 75.205 ms, Inlining 0.000 ms, Optimization 1387.741 ms, Emission 1110.807 ms, Total 2573.753 ms\\n Execution Time: 2587.776 ms\\n\\nBut what really surprised me is that the the average time to optimize\\nper function is now 2.06ms ... less than half of the previous\\nmeasurement. It emits 10% less functions than before, but the time to\\nboth optimize and emit is reduced by 50%. How does that make sense?\\n\\nI don't know how to approach an investigation into this, but I thought\\nthat maybe we're doing quadratic effort in the optimization / emission\\nphase or something like that.\\n\\n-- \\nÁlvaro Herrera\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 17:02:09 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"JIT doing duplicative optimization?\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-11, Alvaro Herrera wrote:\\n\\n> But what really surprised me is that the the average time to optimize\\n> per function is now 2.06ms ... less than half of the previous\\n> measurement. It emits 10% less functions than before, but the time to\\n> both optimize and emit is reduced by 50%. How does that make sense?\\n\\nAh, here's a query that illustrates what I'm on about. I found this\\nquery[1] in a blog post[2].\\n\\n```\\nWITH RECURSIVE typeinfo_tree(\\n oid, ns, name, kind, basetype, elemtype, elemdelim,\\n range_subtype, attrtypoids, attrnames, depth)\\nAS (\\n SELECT\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\n ti.attrtypoids, ti.attrnames, 0\\n FROM\\n (\\n SELECT\\n t.oid AS oid,\\n ns.nspname AS ns,\\n t.typname AS name,\\n t.typtype AS kind,\\n (CASE WHEN t.typtype = 'd' THEN\\n (WITH RECURSIVE typebases(oid, depth) AS (\\n SELECT\\n t2.typbasetype AS oid,\\n 0 AS depth\\n FROM\\n pg_type t2\\n WHERE\\n t2.oid = t.oid\\n UNION ALL\\n SELECT\\n t2.typbasetype AS oid,\\n tb.depth + 1 AS depth\\n FROM\\n pg_type t2,\\n typebases tb\\n WHERE\\n tb.oid = t2.oid\\n AND t2.typbasetype != 0\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n ELSE NULL\\n END) AS basetype,\\n t.typelem AS elemtype,\\n elem_t.typdelim AS elemdelim,\\n range_t.rngsubtype AS range_subtype,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrtypoids,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrnames\\n FROM\\n pg_catalog.pg_type AS t\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\n ns.oid = t.typnamespace)\\n LEFT JOIN pg_type elem_t ON (\\n t.typlen = -1 AND\\n t.typelem != 0 AND\\n t.typelem = elem_t.oid\\n )\\n LEFT JOIN pg_range range_t ON (\\n t.oid = range_t.rngtypid\\n )\\n ) AS ti\\n WHERE\\n ti.oid = any(ARRAY[16,17]::oid[])\\n\\n UNION ALL\\n\\n SELECT\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\n ti.attrtypoids, ti.attrnames, tt.depth + 1\\n FROM\\n (\\n SELECT\\n t.oid AS oid,\\n ns.nspname AS ns,\\n t.typname AS name,\\n t.typtype AS kind,\\n (CASE WHEN t.typtype = 'd' THEN\\n (WITH RECURSIVE typebases(oid, depth) AS (\\n SELECT\\n t2.typbasetype AS oid,\\n 0 AS depth\\n FROM\\n pg_type t2\\n WHERE\\n t2.oid = t.oid\\n UNION ALL\\n SELECT\\n t2.typbasetype AS oid,\\n tb.depth + 1 AS depth\\n FROM\\n pg_type t2,\\n typebases tb\\n WHERE\\n tb.oid = t2.oid\\n AND t2.typbasetype != 0\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n ELSE NULL\\n END) AS basetype,\\n t.typelem AS elemtype,\\n elem_t.typdelim AS elemdelim,\\n range_t.rngsubtype AS range_subtype,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrtypoids,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrnames\\n FROM\\n pg_catalog.pg_type AS t\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\n ns.oid = t.typnamespace)\\n LEFT JOIN pg_type elem_t ON (\\n t.typlen = -1 AND\\n t.typelem != 0 AND\\n t.typelem = elem_t.oid\\n )\\n LEFT JOIN pg_range range_t ON (\\n t.oid = range_t.rngtypid\\n )\\n ) ti,\\n typeinfo_tree tt\\n WHERE\\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\\n)\\nSELECT DISTINCT\\n *,\\n basetype::regtype::text AS basetype_name,\\n elemtype::regtype::text AS elemtype_name,\\n range_subtype::regtype::text AS range_subtype_name\\nFROM\\n typeinfo_tree\\nORDER BY\\n depth DESC;\\n```\\n\\nI did an EXPLAIN ANALYZE and at the bottom you see this:\\n\\n Planning Time: 2.606 ms\\n JIT:\\n Functions: 148\\n Options: Inlining true, Optimization true, Expressions true, Deforming true\\n Timing: Generation 19.670 ms, Inlining 19.224 ms, Optimization 435.153 ms, Emission 282.216 ms, Total 756.263 ms\\n Execution Time: 757.643 ms\\n\\nAverage time to optimize, per function 435.153/148 = 2.940ms;\\naverage time to emit per function 282.216/148 = 1.906ms\\n\\nNow let's change the query by making the first innermost recursive CTE into\\nnon-recursive; just delete the RECURSIVE keyword and everything after the UNION\\nALL. You now get this query.\\n\\n```\\nexplain analyze\\nWITH RECURSIVE typeinfo_tree(\\n oid, ns, name, kind, basetype, elemtype, elemdelim,\\n range_subtype, attrtypoids, attrnames, depth)\\nAS (\\n SELECT\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\n ti.attrtypoids, ti.attrnames, 0\\n FROM\\n (\\n SELECT\\n t.oid AS oid,\\n ns.nspname AS ns,\\n t.typname AS name,\\n t.typtype AS kind,\\n (CASE WHEN t.typtype = 'd' THEN\\n (WITH typebases(oid, depth) AS (\\n SELECT\\n t2.typbasetype AS oid,\\n 0 AS depth\\n FROM\\n pg_type t2\\n WHERE\\n t2.oid = t.oid\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n ELSE NULL\\n END) AS basetype,\\n t.typelem AS elemtype,\\n elem_t.typdelim AS elemdelim,\\n range_t.rngsubtype AS range_subtype,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrtypoids,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrnames\\n FROM\\n pg_catalog.pg_type AS t\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\n ns.oid = t.typnamespace)\\n LEFT JOIN pg_type elem_t ON (\\n t.typlen = -1 AND\\n t.typelem != 0 AND\\n t.typelem = elem_t.oid\\n )\\n LEFT JOIN pg_range range_t ON (\\n t.oid = range_t.rngtypid\\n )\\n ) AS ti\\n WHERE\\n ti.oid = any(ARRAY[16,17]::oid[])\\n\\n UNION ALL\\n\\n SELECT\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\n ti.attrtypoids, ti.attrnames, tt.depth + 1\\n FROM\\n (\\n SELECT\\n t.oid AS oid,\\n ns.nspname AS ns,\\n t.typname AS name,\\n t.typtype AS kind,\\n (CASE WHEN t.typtype = 'd' THEN\\n (WITH RECURSIVE typebases(oid, depth) AS (\\n SELECT\\n t2.typbasetype AS oid,\\n 0 AS depth\\n FROM\\n pg_type t2\\n WHERE\\n t2.oid = t.oid\\n UNION ALL\\n SELECT\\n t2.typbasetype AS oid,\\n tb.depth + 1 AS depth\\n FROM\\n pg_type t2,\\n typebases tb\\n WHERE\\n tb.oid = t2.oid\\n AND t2.typbasetype != 0\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n ELSE NULL\\n END) AS basetype,\\n t.typelem AS elemtype,\\n elem_t.typdelim AS elemdelim,\\n range_t.rngsubtype AS range_subtype,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrtypoids,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrnames\\n FROM\\n pg_catalog.pg_type AS t\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\n ns.oid = t.typnamespace)\\n LEFT JOIN pg_type elem_t ON (\\n t.typlen = -1 AND\\n t.typelem != 0 AND\\n t.typelem = elem_t.oid\\n )\\n LEFT JOIN pg_range range_t ON (\\n t.oid = range_t.rngtypid\\n )\\n ) ti,\\n typeinfo_tree tt\\n WHERE\\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\\n)\\n\\nSELECT DISTINCT\\n *,\\n basetype::regtype::text AS basetype_name,\\n elemtype::regtype::text AS elemtype_name,\\n range_subtype::regtype::text AS range_subtype_name\\nFROM\\n typeinfo_tree\\nORDER BY\\n depth DESC\\n``` \\n\\nAt the bottom of the explain you get this.\\n\\n Planning Time: 2.508 ms\\n JIT:\\n Functions: 137\\n Options: Inlining true, Optimization true, Expressions true, Deforming true\\n Timing: Generation 10.346 ms, Inlining 10.210 ms, Optimization 374.103 ms, Emission 254.557 ms, Total 649.216 ms\\n Execution Time: 650.168 ms\\n\\nAverage time to optimize, per function: 374.103/137 = 2.730ms\\nAverage time to emit, per function 254.557 / 137 = 1.858ms\\n\\n\\nNow do it one more time, with the second innermost recursive CTE. You get this\\nquery.\\n\\n```\\nexplain analyze\\nWITH RECURSIVE typeinfo_tree(\\n oid, ns, name, kind, basetype, elemtype, elemdelim,\\n range_subtype, attrtypoids, attrnames, depth)\\nAS (\\n SELECT\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\n ti.attrtypoids, ti.attrnames, 0\\n FROM\\n (\\n SELECT\\n t.oid AS oid,\\n ns.nspname AS ns,\\n t.typname AS name,\\n t.typtype AS kind,\\n (CASE WHEN t.typtype = 'd' THEN\\n (WITH typebases(oid, depth) AS (\\n SELECT\\n t2.typbasetype AS oid,\\n 0 AS depth\\n FROM\\n pg_type t2\\n WHERE\\n t2.oid = t.oid\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n ELSE NULL\\n END) AS basetype,\\n t.typelem AS elemtype,\\n elem_t.typdelim AS elemdelim,\\n range_t.rngsubtype AS range_subtype,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrtypoids,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrnames\\n FROM\\n pg_catalog.pg_type AS t\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\n ns.oid = t.typnamespace)\\n LEFT JOIN pg_type elem_t ON (\\n t.typlen = -1 AND\\n t.typelem != 0 AND\\n t.typelem = elem_t.oid\\n )\\n LEFT JOIN pg_range range_t ON (\\n t.oid = range_t.rngtypid\\n )\\n ) AS ti\\n WHERE\\n ti.oid = any(ARRAY[16,17]::oid[])\\n\\n UNION ALL\\n\\n SELECT\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\n ti.attrtypoids, ti.attrnames, tt.depth + 1\\n FROM\\n (\\n SELECT\\n t.oid AS oid,\\n ns.nspname AS ns,\\n t.typname AS name,\\n t.typtype AS kind,\\n (CASE WHEN t.typtype = 'd' THEN\\n (WITH typebases(oid, depth) AS (\\n SELECT\\n t2.typbasetype AS oid,\\n 0 AS depth\\n FROM\\n pg_type t2\\n WHERE\\n t2.oid = t.oid\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n ELSE NULL\\n END) AS basetype,\\n t.typelem AS elemtype,\\n elem_t.typdelim AS elemdelim,\\n range_t.rngsubtype AS range_subtype,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrtypoids,\\n (CASE WHEN t.typtype = 'c' THEN\\n (SELECT\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\n FROM\\n pg_attribute ia\\n INNER JOIN pg_class c\\n ON (ia.attrelid = c.oid)\\n WHERE\\n ia.attnum > 0 AND NOT ia.attisdropped\\n AND c.reltype = t.oid)\\n ELSE NULL\\n END) AS attrnames\\n FROM\\n pg_catalog.pg_type AS t\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\n ns.oid = t.typnamespace)\\n LEFT JOIN pg_type elem_t ON (\\n t.typlen = -1 AND\\n t.typelem != 0 AND\\n t.typelem = elem_t.oid\\n )\\n LEFT JOIN pg_range range_t ON (\\n t.oid = range_t.rngtypid\\n )\\n ) ti,\\n typeinfo_tree tt\\n WHERE\\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\\n)\\n\\nSELECT DISTINCT\\n *,\\n basetype::regtype::text AS basetype_name,\\n elemtype::regtype::text AS elemtype_name,\\n range_subtype::regtype::text AS range_subtype_name\\nFROM\\n typeinfo_tree\\nORDER BY\\n depth DESC;\\n```\\n\\nAnd the JIT numbers at the bottom look like this:\\n\\n Planning Time: 2.041 ms\\n JIT:\\n Functions: 126\\n Options: Inlining false, Optimization true, Expressions true, Deforming true\\n Timing: Generation 10.002 ms, Inlining 0.000 ms, Optimization 229.128 ms, Emission 167.338 ms, Total 406.469 ms\\n Execution Time: 407.315 ms\\n\\nAverage time to optimize per function 229.128 / 126 = 1.181ms\\nAverage time to emit per function 167.338 / 126 = 1.328ms\\n\\n\\nIn summary,\\n\\nQuery 1, 148 functions JIT-compiled.\\nAverage time to optimize, per function 435.153/148 = 2.940ms;\\naverage time to emit per function 282.216/148 = 1.906ms\\n\\nQuery 2, 137 functions JIT-compiled.\\nAverage time to optimize, per function: 374.103/137 = 2.730ms\\nAverage time to emit, per function 254.557 / 137 = 1.858ms\\n\\nQuery 3, 126 functions JIT-compiled.\\nAverage time to optimize per function 229.128 / 126 = 1.181ms\\nAverage time to emit per function 167.338 / 126 = 1.328ms\\n\\n???? Something looks very wrong here.\\n\\n[1] https://gist.github.com/saicitus/251ba20b211e9e73285af35e61b19580\\n[2] https://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2a\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\\"Java is clearly an example of money oriented programming\\\" (A. Stepanov)\\n\\n\\n\",\n \"msg_date\": \"Sun, 14 Nov 2021 14:06:46 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: JIT doing duplicative optimization?\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 14, 2021 at 9:07 AM Alvaro Herrera <alvherre@alvh.no-ip.org>\\nwrote:\\n\\n> On 2021-Nov-11, Alvaro Herrera wrote:\\n>\\n> > But what really surprised me is that the the average time to optimize\\n> > per function is now 2.06ms ... less than half of the previous\\n> > measurement. It emits 10% less functions than before, but the time to\\n> > both optimize and emit is reduced by 50%. How does that make sense?\\n>\\n> Ah, here's a query that illustrates what I'm on about. I found this\\n> query[1] in a blog post[2].\\n>\\n> ```\\n> WITH RECURSIVE typeinfo_tree(\\n> oid, ns, name, kind, basetype, elemtype, elemdelim,\\n> range_subtype, attrtypoids, attrnames, depth)\\n> AS (\\n> SELECT\\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\\n> ti.attrtypoids, ti.attrnames, 0\\n> FROM\\n> (\\n> SELECT\\n> t.oid AS oid,\\n> ns.nspname AS ns,\\n> t.typname AS name,\\n> t.typtype AS kind,\\n> (CASE WHEN t.typtype = 'd' THEN\\n> (WITH RECURSIVE typebases(oid, depth) AS (\\n> SELECT\\n> t2.typbasetype AS oid,\\n> 0 AS depth\\n> FROM\\n> pg_type t2\\n> WHERE\\n> t2.oid = t.oid\\n> UNION ALL\\n> SELECT\\n> t2.typbasetype AS oid,\\n> tb.depth + 1 AS depth\\n> FROM\\n> pg_type t2,\\n> typebases tb\\n> WHERE\\n> tb.oid = t2.oid\\n> AND t2.typbasetype != 0\\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n> ELSE NULL\\n> END) AS basetype,\\n> t.typelem AS elemtype,\\n> elem_t.typdelim AS elemdelim,\\n> range_t.rngsubtype AS range_subtype,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.atttypid ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrtypoids,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.attname::text ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrnames\\n> FROM\\n> pg_catalog.pg_type AS t\\n> INNER JOIN pg_catalog.pg_namespace ns ON (\\n> ns.oid = t.typnamespace)\\n> LEFT JOIN pg_type elem_t ON (\\n> t.typlen = -1 AND\\n> t.typelem != 0 AND\\n> t.typelem = elem_t.oid\\n> )\\n> LEFT JOIN pg_range range_t ON (\\n> t.oid = range_t.rngtypid\\n> )\\n> ) AS ti\\n> WHERE\\n> ti.oid = any(ARRAY[16,17]::oid[])\\n>\\n> UNION ALL\\n>\\n> SELECT\\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\\n> ti.attrtypoids, ti.attrnames, tt.depth + 1\\n> FROM\\n> (\\n> SELECT\\n> t.oid AS oid,\\n> ns.nspname AS ns,\\n> t.typname AS name,\\n> t.typtype AS kind,\\n> (CASE WHEN t.typtype = 'd' THEN\\n> (WITH RECURSIVE typebases(oid, depth) AS (\\n> SELECT\\n> t2.typbasetype AS oid,\\n> 0 AS depth\\n> FROM\\n> pg_type t2\\n> WHERE\\n> t2.oid = t.oid\\n> UNION ALL\\n> SELECT\\n> t2.typbasetype AS oid,\\n> tb.depth + 1 AS depth\\n> FROM\\n> pg_type t2,\\n> typebases tb\\n> WHERE\\n> tb.oid = t2.oid\\n> AND t2.typbasetype != 0\\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n> ELSE NULL\\n> END) AS basetype,\\n> t.typelem AS elemtype,\\n> elem_t.typdelim AS elemdelim,\\n> range_t.rngsubtype AS range_subtype,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.atttypid ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrtypoids,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.attname::text ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrnames\\n> FROM\\n> pg_catalog.pg_type AS t\\n> INNER JOIN pg_catalog.pg_namespace ns ON (\\n> ns.oid = t.typnamespace)\\n> LEFT JOIN pg_type elem_t ON (\\n> t.typlen = -1 AND\\n> t.typelem != 0 AND\\n> t.typelem = elem_t.oid\\n> )\\n> LEFT JOIN pg_range range_t ON (\\n> t.oid = range_t.rngtypid\\n> )\\n> ) ti,\\n> typeinfo_tree tt\\n> WHERE\\n> (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\\n> OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\\n> OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\\n> )\\n> SELECT DISTINCT\\n> *,\\n> basetype::regtype::text AS basetype_name,\\n> elemtype::regtype::text AS elemtype_name,\\n> range_subtype::regtype::text AS range_subtype_name\\n> FROM\\n> typeinfo_tree\\n> ORDER BY\\n> depth DESC;\\n> ```\\n>\\n> I did an EXPLAIN ANALYZE and at the bottom you see this:\\n>\\n> Planning Time: 2.606 ms\\n> JIT:\\n> Functions: 148\\n> Options: Inlining true, Optimization true, Expressions true, Deforming\\n> true\\n> Timing: Generation 19.670 ms, Inlining 19.224 ms, Optimization 435.153\\n> ms, Emission 282.216 ms, Total 756.263 ms\\n> Execution Time: 757.643 ms\\n>\\n> Average time to optimize, per function 435.153/148 = 2.940ms;\\n> average time to emit per function 282.216/148 = 1.906ms\\n>\\n> Now let's change the query by making the first innermost recursive CTE into\\n> non-recursive; just delete the RECURSIVE keyword and everything after the\\n> UNION\\n> ALL. You now get this query.\\n>\\n> ```\\n> explain analyze\\n> WITH RECURSIVE typeinfo_tree(\\n> oid, ns, name, kind, basetype, elemtype, elemdelim,\\n> range_subtype, attrtypoids, attrnames, depth)\\n> AS (\\n> SELECT\\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\\n> ti.attrtypoids, ti.attrnames, 0\\n> FROM\\n> (\\n> SELECT\\n> t.oid AS oid,\\n> ns.nspname AS ns,\\n> t.typname AS name,\\n> t.typtype AS kind,\\n> (CASE WHEN t.typtype = 'd' THEN\\n> (WITH typebases(oid, depth) AS (\\n> SELECT\\n> t2.typbasetype AS oid,\\n> 0 AS depth\\n> FROM\\n> pg_type t2\\n> WHERE\\n> t2.oid = t.oid\\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n> ELSE NULL\\n> END) AS basetype,\\n> t.typelem AS elemtype,\\n> elem_t.typdelim AS elemdelim,\\n> range_t.rngsubtype AS range_subtype,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.atttypid ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrtypoids,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.attname::text ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrnames\\n> FROM\\n> pg_catalog.pg_type AS t\\n> INNER JOIN pg_catalog.pg_namespace ns ON (\\n> ns.oid = t.typnamespace)\\n> LEFT JOIN pg_type elem_t ON (\\n> t.typlen = -1 AND\\n> t.typelem != 0 AND\\n> t.typelem = elem_t.oid\\n> )\\n> LEFT JOIN pg_range range_t ON (\\n> t.oid = range_t.rngtypid\\n> )\\n> ) AS ti\\n> WHERE\\n> ti.oid = any(ARRAY[16,17]::oid[])\\n>\\n> UNION ALL\\n>\\n> SELECT\\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\\n> ti.attrtypoids, ti.attrnames, tt.depth + 1\\n> FROM\\n> (\\n> SELECT\\n> t.oid AS oid,\\n> ns.nspname AS ns,\\n> t.typname AS name,\\n> t.typtype AS kind,\\n> (CASE WHEN t.typtype = 'd' THEN\\n> (WITH RECURSIVE typebases(oid, depth) AS (\\n> SELECT\\n> t2.typbasetype AS oid,\\n> 0 AS depth\\n> FROM\\n> pg_type t2\\n> WHERE\\n> t2.oid = t.oid\\n> UNION ALL\\n> SELECT\\n> t2.typbasetype AS oid,\\n> tb.depth + 1 AS depth\\n> FROM\\n> pg_type t2,\\n> typebases tb\\n> WHERE\\n> tb.oid = t2.oid\\n> AND t2.typbasetype != 0\\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n> ELSE NULL\\n> END) AS basetype,\\n> t.typelem AS elemtype,\\n> elem_t.typdelim AS elemdelim,\\n> range_t.rngsubtype AS range_subtype,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.atttypid ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrtypoids,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.attname::text ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrnames\\n> FROM\\n> pg_catalog.pg_type AS t\\n> INNER JOIN pg_catalog.pg_namespace ns ON (\\n> ns.oid = t.typnamespace)\\n> LEFT JOIN pg_type elem_t ON (\\n> t.typlen = -1 AND\\n> t.typelem != 0 AND\\n> t.typelem = elem_t.oid\\n> )\\n> LEFT JOIN pg_range range_t ON (\\n> t.oid = range_t.rngtypid\\n> )\\n> ) ti,\\n> typeinfo_tree tt\\n> WHERE\\n> (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\\n> OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\\n> OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\\n> )\\n>\\n> SELECT DISTINCT\\n> *,\\n> basetype::regtype::text AS basetype_name,\\n> elemtype::regtype::text AS elemtype_name,\\n> range_subtype::regtype::text AS range_subtype_name\\n> FROM\\n> typeinfo_tree\\n> ORDER BY\\n> depth DESC\\n> ```\\n>\\n> At the bottom of the explain you get this.\\n>\\n> Planning Time: 2.508 ms\\n> JIT:\\n> Functions: 137\\n> Options: Inlining true, Optimization true, Expressions true, Deforming\\n> true\\n> Timing: Generation 10.346 ms, Inlining 10.210 ms, Optimization 374.103\\n> ms, Emission 254.557 ms, Total 649.216 ms\\n> Execution Time: 650.168 ms\\n>\\n> Average time to optimize, per function: 374.103/137 = 2.730ms\\n> Average time to emit, per function 254.557 / 137 = 1.858ms\\n>\\n>\\n> Now do it one more time, with the second innermost recursive CTE. You get\\n> this\\n> query.\\n>\\n> ```\\n> explain analyze\\n> WITH RECURSIVE typeinfo_tree(\\n> oid, ns, name, kind, basetype, elemtype, elemdelim,\\n> range_subtype, attrtypoids, attrnames, depth)\\n> AS (\\n> SELECT\\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\\n> ti.attrtypoids, ti.attrnames, 0\\n> FROM\\n> (\\n> SELECT\\n> t.oid AS oid,\\n> ns.nspname AS ns,\\n> t.typname AS name,\\n> t.typtype AS kind,\\n> (CASE WHEN t.typtype = 'd' THEN\\n> (WITH typebases(oid, depth) AS (\\n> SELECT\\n> t2.typbasetype AS oid,\\n> 0 AS depth\\n> FROM\\n> pg_type t2\\n> WHERE\\n> t2.oid = t.oid\\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n> ELSE NULL\\n> END) AS basetype,\\n> t.typelem AS elemtype,\\n> elem_t.typdelim AS elemdelim,\\n> range_t.rngsubtype AS range_subtype,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.atttypid ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrtypoids,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.attname::text ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrnames\\n> FROM\\n> pg_catalog.pg_type AS t\\n> INNER JOIN pg_catalog.pg_namespace ns ON (\\n> ns.oid = t.typnamespace)\\n> LEFT JOIN pg_type elem_t ON (\\n> t.typlen = -1 AND\\n> t.typelem != 0 AND\\n> t.typelem = elem_t.oid\\n> )\\n> LEFT JOIN pg_range range_t ON (\\n> t.oid = range_t.rngtypid\\n> )\\n> ) AS ti\\n> WHERE\\n> ti.oid = any(ARRAY[16,17]::oid[])\\n>\\n> UNION ALL\\n>\\n> SELECT\\n> ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\n> ti.elemtype, ti.elemdelim, ti.range_subtype,\\n> ti.attrtypoids, ti.attrnames, tt.depth + 1\\n> FROM\\n> (\\n> SELECT\\n> t.oid AS oid,\\n> ns.nspname AS ns,\\n> t.typname AS name,\\n> t.typtype AS kind,\\n> (CASE WHEN t.typtype = 'd' THEN\\n> (WITH typebases(oid, depth) AS (\\n> SELECT\\n> t2.typbasetype AS oid,\\n> 0 AS depth\\n> FROM\\n> pg_type t2\\n> WHERE\\n> t2.oid = t.oid\\n> ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\n> ELSE NULL\\n> END) AS basetype,\\n> t.typelem AS elemtype,\\n> elem_t.typdelim AS elemdelim,\\n> range_t.rngsubtype AS range_subtype,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.atttypid ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrtypoids,\\n> (CASE WHEN t.typtype = 'c' THEN\\n> (SELECT\\n> array_agg(ia.attname::text ORDER BY ia.attnum)\\n> FROM\\n> pg_attribute ia\\n> INNER JOIN pg_class c\\n> ON (ia.attrelid = c.oid)\\n> WHERE\\n> ia.attnum > 0 AND NOT ia.attisdropped\\n> AND c.reltype = t.oid)\\n> ELSE NULL\\n> END) AS attrnames\\n> FROM\\n> pg_catalog.pg_type AS t\\n> INNER JOIN pg_catalog.pg_namespace ns ON (\\n> ns.oid = t.typnamespace)\\n> LEFT JOIN pg_type elem_t ON (\\n> t.typlen = -1 AND\\n> t.typelem != 0 AND\\n> t.typelem = elem_t.oid\\n> )\\n> LEFT JOIN pg_range range_t ON (\\n> t.oid = range_t.rngtypid\\n> )\\n> ) ti,\\n> typeinfo_tree tt\\n> WHERE\\n> (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\\n> OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\\n> OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\\n> )\\n>\\n> SELECT DISTINCT\\n> *,\\n> basetype::regtype::text AS basetype_name,\\n> elemtype::regtype::text AS elemtype_name,\\n> range_subtype::regtype::text AS range_subtype_name\\n> FROM\\n> typeinfo_tree\\n> ORDER BY\\n> depth DESC;\\n> ```\\n>\\n> And the JIT numbers at the bottom look like this:\\n>\\n> Planning Time: 2.041 ms\\n> JIT:\\n> Functions: 126\\n> Options: Inlining false, Optimization true, Expressions true, Deforming\\n> true\\n> Timing: Generation 10.002 ms, Inlining 0.000 ms, Optimization 229.128\\n> ms, Emission 167.338 ms, Total 406.469 ms\\n> Execution Time: 407.315 ms\\n>\\n> Average time to optimize per function 229.128 / 126 = 1.181ms\\n> Average time to emit per function 167.338 / 126 = 1.328ms\\n>\\n>\\n> In summary,\\n>\\n> Query 1, 148 functions JIT-compiled.\\n> Average time to optimize, per function 435.153/148 = 2.940ms;\\n> average time to emit per function 282.216/148 = 1.906ms\\n>\\n> Query 2, 137 functions JIT-compiled.\\n> Average time to optimize, per function: 374.103/137 = 2.730ms\\n> Average time to emit, per function 254.557 / 137 = 1.858ms\\n>\\n> Query 3, 126 functions JIT-compiled.\\n> Average time to optimize per function 229.128 / 126 = 1.181ms\\n> Average time to emit per function 167.338 / 126 = 1.328ms\\n>\\n> ???? Something looks very wrong here.\\n>\\n> [1] https://gist.github.com/saicitus/251ba20b211e9e73285af35e61b19580\\n> [2]\\n> https://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2a\\n>\\n> --\\n> Álvaro Herrera PostgreSQL Developer —\\n> https://www.EnterpriseDB.com/\\n> \\\"Java is clearly an example of money oriented programming\\\" (A. Stepanov)\\n>\\n>\\n> Hi,\\nFor the 3rd query, I noticed some difference in options:\\n\\n Options: Inlining false, Optimization true, Expressions true, Deforming\\ntrue\\n\\n Inlining was true for the first two queries.\\n\\nFYI\\n\\nOn Sun, Nov 14, 2021 at 9:07 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:On 2021-Nov-11, Alvaro Herrera wrote:\\n\\r\\n> But what really surprised me is that the the average time to optimize\\r\\n> per function is now 2.06ms ... less than half of the previous\\r\\n> measurement. It emits 10% less functions than before, but the time to\\r\\n> both optimize and emit is reduced by 50%. How does that make sense?\\n\\r\\nAh, here's a query that illustrates what I'm on about. I found this\\r\\nquery[1] in a blog post[2].\\n\\r\\n```\\r\\nWITH RECURSIVE typeinfo_tree(\\r\\n oid, ns, name, kind, basetype, elemtype, elemdelim,\\r\\n range_subtype, attrtypoids, attrnames, depth)\\r\\nAS (\\r\\n SELECT\\r\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\r\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\r\\n ti.attrtypoids, ti.attrnames, 0\\r\\n FROM\\r\\n (\\r\\n SELECT\\r\\n t.oid AS oid,\\r\\n ns.nspname AS ns,\\r\\n t.typname AS name,\\r\\n t.typtype AS kind,\\r\\n (CASE WHEN t.typtype = 'd' THEN\\r\\n (WITH RECURSIVE typebases(oid, depth) AS (\\r\\n SELECT\\r\\n t2.typbasetype AS oid,\\r\\n 0 AS depth\\r\\n FROM\\r\\n pg_type t2\\r\\n WHERE\\r\\n t2.oid = t.oid\\r\\n UNION ALL\\r\\n SELECT\\r\\n t2.typbasetype AS oid,\\r\\n tb.depth + 1 AS depth\\r\\n FROM\\r\\n pg_type t2,\\r\\n typebases tb\\r\\n WHERE\\r\\n tb.oid = t2.oid\\r\\n AND t2.typbasetype != 0\\r\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\r\\n ELSE NULL\\r\\n END) AS basetype,\\r\\n t.typelem AS elemtype,\\r\\n elem_t.typdelim AS elemdelim,\\r\\n range_t.rngsubtype AS range_subtype,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrtypoids,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrnames\\r\\n FROM\\r\\n pg_catalog.pg_type AS t\\r\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\r\\n ns.oid = t.typnamespace)\\r\\n LEFT JOIN pg_type elem_t ON (\\r\\n t.typlen = -1 AND\\r\\n t.typelem != 0 AND\\r\\n t.typelem = elem_t.oid\\r\\n )\\r\\n LEFT JOIN pg_range range_t ON (\\r\\n t.oid = range_t.rngtypid\\r\\n )\\r\\n ) AS ti\\r\\n WHERE\\r\\n ti.oid = any(ARRAY[16,17]::oid[])\\n\\r\\n UNION ALL\\n\\r\\n SELECT\\r\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\r\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\r\\n ti.attrtypoids, ti.attrnames, tt.depth + 1\\r\\n FROM\\r\\n (\\r\\n SELECT\\r\\n t.oid AS oid,\\r\\n ns.nspname AS ns,\\r\\n t.typname AS name,\\r\\n t.typtype AS kind,\\r\\n (CASE WHEN t.typtype = 'd' THEN\\r\\n (WITH RECURSIVE typebases(oid, depth) AS (\\r\\n SELECT\\r\\n t2.typbasetype AS oid,\\r\\n 0 AS depth\\r\\n FROM\\r\\n pg_type t2\\r\\n WHERE\\r\\n t2.oid = t.oid\\r\\n UNION ALL\\r\\n SELECT\\r\\n t2.typbasetype AS oid,\\r\\n tb.depth + 1 AS depth\\r\\n FROM\\r\\n pg_type t2,\\r\\n typebases tb\\r\\n WHERE\\r\\n tb.oid = t2.oid\\r\\n AND t2.typbasetype != 0\\r\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\r\\n ELSE NULL\\r\\n END) AS basetype,\\r\\n t.typelem AS elemtype,\\r\\n elem_t.typdelim AS elemdelim,\\r\\n range_t.rngsubtype AS range_subtype,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrtypoids,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrnames\\r\\n FROM\\r\\n pg_catalog.pg_type AS t\\r\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\r\\n ns.oid = t.typnamespace)\\r\\n LEFT JOIN pg_type elem_t ON (\\r\\n t.typlen = -1 AND\\r\\n t.typelem != 0 AND\\r\\n t.typelem = elem_t.oid\\r\\n )\\r\\n LEFT JOIN pg_range range_t ON (\\r\\n t.oid = range_t.rngtypid\\r\\n )\\r\\n ) ti,\\r\\n typeinfo_tree tt\\r\\n WHERE\\r\\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\\r\\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\\r\\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\\r\\n)\\r\\nSELECT DISTINCT\\r\\n *,\\r\\n basetype::regtype::text AS basetype_name,\\r\\n elemtype::regtype::text AS elemtype_name,\\r\\n range_subtype::regtype::text AS range_subtype_name\\r\\nFROM\\r\\n typeinfo_tree\\r\\nORDER BY\\r\\n depth DESC;\\r\\n```\\n\\r\\nI did an EXPLAIN ANALYZE and at the bottom you see this:\\n\\r\\n Planning Time: 2.606 ms\\r\\n JIT:\\r\\n Functions: 148\\r\\n Options: Inlining true, Optimization true, Expressions true, Deforming true\\r\\n Timing: Generation 19.670 ms, Inlining 19.224 ms, Optimization 435.153 ms, Emission 282.216 ms, Total 756.263 ms\\r\\n Execution Time: 757.643 ms\\n\\r\\nAverage time to optimize, per function 435.153/148 = 2.940ms;\\r\\naverage time to emit per function 282.216/148 = 1.906ms\\n\\r\\nNow let's change the query by making the first innermost recursive CTE into\\r\\nnon-recursive; just delete the RECURSIVE keyword and everything after the UNION\\r\\nALL. You now get this query.\\n\\r\\n```\\r\\nexplain analyze\\r\\nWITH RECURSIVE typeinfo_tree(\\r\\n oid, ns, name, kind, basetype, elemtype, elemdelim,\\r\\n range_subtype, attrtypoids, attrnames, depth)\\r\\nAS (\\r\\n SELECT\\r\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\r\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\r\\n ti.attrtypoids, ti.attrnames, 0\\r\\n FROM\\r\\n (\\r\\n SELECT\\r\\n t.oid AS oid,\\r\\n ns.nspname AS ns,\\r\\n t.typname AS name,\\r\\n t.typtype AS kind,\\r\\n (CASE WHEN t.typtype = 'd' THEN\\r\\n (WITH typebases(oid, depth) AS (\\r\\n SELECT\\r\\n t2.typbasetype AS oid,\\r\\n 0 AS depth\\r\\n FROM\\r\\n pg_type t2\\r\\n WHERE\\r\\n t2.oid = t.oid\\r\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\r\\n ELSE NULL\\r\\n END) AS basetype,\\r\\n t.typelem AS elemtype,\\r\\n elem_t.typdelim AS elemdelim,\\r\\n range_t.rngsubtype AS range_subtype,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrtypoids,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrnames\\r\\n FROM\\r\\n pg_catalog.pg_type AS t\\r\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\r\\n ns.oid = t.typnamespace)\\r\\n LEFT JOIN pg_type elem_t ON (\\r\\n t.typlen = -1 AND\\r\\n t.typelem != 0 AND\\r\\n t.typelem = elem_t.oid\\r\\n )\\r\\n LEFT JOIN pg_range range_t ON (\\r\\n t.oid = range_t.rngtypid\\r\\n )\\r\\n ) AS ti\\r\\n WHERE\\r\\n ti.oid = any(ARRAY[16,17]::oid[])\\n\\r\\n UNION ALL\\n\\r\\n SELECT\\r\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\r\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\r\\n ti.attrtypoids, ti.attrnames, tt.depth + 1\\r\\n FROM\\r\\n (\\r\\n SELECT\\r\\n t.oid AS oid,\\r\\n ns.nspname AS ns,\\r\\n t.typname AS name,\\r\\n t.typtype AS kind,\\r\\n (CASE WHEN t.typtype = 'd' THEN\\r\\n (WITH RECURSIVE typebases(oid, depth) AS (\\r\\n SELECT\\r\\n t2.typbasetype AS oid,\\r\\n 0 AS depth\\r\\n FROM\\r\\n pg_type t2\\r\\n WHERE\\r\\n t2.oid = t.oid\\r\\n UNION ALL\\r\\n SELECT\\r\\n t2.typbasetype AS oid,\\r\\n tb.depth + 1 AS depth\\r\\n FROM\\r\\n pg_type t2,\\r\\n typebases tb\\r\\n WHERE\\r\\n tb.oid = t2.oid\\r\\n AND t2.typbasetype != 0\\r\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\r\\n ELSE NULL\\r\\n END) AS basetype,\\r\\n t.typelem AS elemtype,\\r\\n elem_t.typdelim AS elemdelim,\\r\\n range_t.rngsubtype AS range_subtype,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrtypoids,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrnames\\r\\n FROM\\r\\n pg_catalog.pg_type AS t\\r\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\r\\n ns.oid = t.typnamespace)\\r\\n LEFT JOIN pg_type elem_t ON (\\r\\n t.typlen = -1 AND\\r\\n t.typelem != 0 AND\\r\\n t.typelem = elem_t.oid\\r\\n )\\r\\n LEFT JOIN pg_range range_t ON (\\r\\n t.oid = range_t.rngtypid\\r\\n )\\r\\n ) ti,\\r\\n typeinfo_tree tt\\r\\n WHERE\\r\\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\\r\\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\\r\\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\\r\\n)\\n\\r\\nSELECT DISTINCT\\r\\n *,\\r\\n basetype::regtype::text AS basetype_name,\\r\\n elemtype::regtype::text AS elemtype_name,\\r\\n range_subtype::regtype::text AS range_subtype_name\\r\\nFROM\\r\\n typeinfo_tree\\r\\nORDER BY\\r\\n depth DESC\\r\\n``` \\n\\r\\nAt the bottom of the explain you get this.\\n\\r\\n Planning Time: 2.508 ms\\r\\n JIT:\\r\\n Functions: 137\\r\\n Options: Inlining true, Optimization true, Expressions true, Deforming true\\r\\n Timing: Generation 10.346 ms, Inlining 10.210 ms, Optimization 374.103 ms, Emission 254.557 ms, Total 649.216 ms\\r\\n Execution Time: 650.168 ms\\n\\r\\nAverage time to optimize, per function: 374.103/137 = 2.730ms\\r\\nAverage time to emit, per function 254.557 / 137 = 1.858ms\\n\\n\\r\\nNow do it one more time, with the second innermost recursive CTE. You get this\\r\\nquery.\\n\\r\\n```\\r\\nexplain analyze\\r\\nWITH RECURSIVE typeinfo_tree(\\r\\n oid, ns, name, kind, basetype, elemtype, elemdelim,\\r\\n range_subtype, attrtypoids, attrnames, depth)\\r\\nAS (\\r\\n SELECT\\r\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\r\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\r\\n ti.attrtypoids, ti.attrnames, 0\\r\\n FROM\\r\\n (\\r\\n SELECT\\r\\n t.oid AS oid,\\r\\n ns.nspname AS ns,\\r\\n t.typname AS name,\\r\\n t.typtype AS kind,\\r\\n (CASE WHEN t.typtype = 'd' THEN\\r\\n (WITH typebases(oid, depth) AS (\\r\\n SELECT\\r\\n t2.typbasetype AS oid,\\r\\n 0 AS depth\\r\\n FROM\\r\\n pg_type t2\\r\\n WHERE\\r\\n t2.oid = t.oid\\r\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\r\\n ELSE NULL\\r\\n END) AS basetype,\\r\\n t.typelem AS elemtype,\\r\\n elem_t.typdelim AS elemdelim,\\r\\n range_t.rngsubtype AS range_subtype,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrtypoids,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrnames\\r\\n FROM\\r\\n pg_catalog.pg_type AS t\\r\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\r\\n ns.oid = t.typnamespace)\\r\\n LEFT JOIN pg_type elem_t ON (\\r\\n t.typlen = -1 AND\\r\\n t.typelem != 0 AND\\r\\n t.typelem = elem_t.oid\\r\\n )\\r\\n LEFT JOIN pg_range range_t ON (\\r\\n t.oid = range_t.rngtypid\\r\\n )\\r\\n ) AS ti\\r\\n WHERE\\r\\n ti.oid = any(ARRAY[16,17]::oid[])\\n\\r\\n UNION ALL\\n\\r\\n SELECT\\r\\n ti.oid, ti.ns, ti.name, ti.kind, ti.basetype,\\r\\n ti.elemtype, ti.elemdelim, ti.range_subtype,\\r\\n ti.attrtypoids, ti.attrnames, tt.depth + 1\\r\\n FROM\\r\\n (\\r\\n SELECT\\r\\n t.oid AS oid,\\r\\n ns.nspname AS ns,\\r\\n t.typname AS name,\\r\\n t.typtype AS kind,\\r\\n (CASE WHEN t.typtype = 'd' THEN\\r\\n (WITH typebases(oid, depth) AS (\\r\\n SELECT\\r\\n t2.typbasetype AS oid,\\r\\n 0 AS depth\\r\\n FROM\\r\\n pg_type t2\\r\\n WHERE\\r\\n t2.oid = t.oid\\r\\n ) SELECT oid FROM typebases ORDER BY depth DESC LIMIT 1)\\r\\n ELSE NULL\\r\\n END) AS basetype,\\r\\n t.typelem AS elemtype,\\r\\n elem_t.typdelim AS elemdelim,\\r\\n range_t.rngsubtype AS range_subtype,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.atttypid ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrtypoids,\\r\\n (CASE WHEN t.typtype = 'c' THEN\\r\\n (SELECT\\r\\n array_agg(ia.attname::text ORDER BY ia.attnum)\\r\\n FROM\\r\\n pg_attribute ia\\r\\n INNER JOIN pg_class c\\r\\n ON (ia.attrelid = c.oid)\\r\\n WHERE\\r\\n ia.attnum > 0 AND NOT ia.attisdropped\\r\\n AND c.reltype = t.oid)\\r\\n ELSE NULL\\r\\n END) AS attrnames\\r\\n FROM\\r\\n pg_catalog.pg_type AS t\\r\\n INNER JOIN pg_catalog.pg_namespace ns ON (\\r\\n ns.oid = t.typnamespace)\\r\\n LEFT JOIN pg_type elem_t ON (\\r\\n t.typlen = -1 AND\\r\\n t.typelem != 0 AND\\r\\n t.typelem = elem_t.oid\\r\\n )\\r\\n LEFT JOIN pg_range range_t ON (\\r\\n t.oid = range_t.rngtypid\\r\\n )\\r\\n ) ti,\\r\\n typeinfo_tree tt\\r\\n WHERE\\r\\n (tt.elemtype IS NOT NULL AND ti.oid = tt.elemtype)\\r\\n OR (tt.attrtypoids IS NOT NULL AND ti.oid = any(tt.attrtypoids))\\r\\n OR (tt.range_subtype IS NOT NULL AND ti.oid = tt.range_subtype)\\r\\n)\\n\\r\\nSELECT DISTINCT\\r\\n *,\\r\\n basetype::regtype::text AS basetype_name,\\r\\n elemtype::regtype::text AS elemtype_name,\\r\\n range_subtype::regtype::text AS range_subtype_name\\r\\nFROM\\r\\n typeinfo_tree\\r\\nORDER BY\\r\\n depth DESC;\\r\\n```\\n\\r\\nAnd the JIT numbers at the bottom look like this:\\n\\r\\n Planning Time: 2.041 ms\\r\\n JIT:\\r\\n Functions: 126\\r\\n Options: Inlining false, Optimization true, Expressions true, Deforming true\\r\\n Timing: Generation 10.002 ms, Inlining 0.000 ms, Optimization 229.128 ms, Emission 167.338 ms, Total 406.469 ms\\r\\n Execution Time: 407.315 ms\\n\\r\\nAverage time to optimize per function 229.128 / 126 = 1.181ms\\r\\nAverage time to emit per function 167.338 / 126 = 1.328ms\\n\\n\\r\\nIn summary,\\n\\r\\nQuery 1, 148 functions JIT-compiled.\\r\\nAverage time to optimize, per function 435.153/148 = 2.940ms;\\r\\naverage time to emit per function 282.216/148 = 1.906ms\\n\\r\\nQuery 2, 137 functions JIT-compiled.\\r\\nAverage time to optimize, per function: 374.103/137 = 2.730ms\\r\\nAverage time to emit, per function 254.557 / 137 = 1.858ms\\n\\r\\nQuery 3, 126 functions JIT-compiled.\\r\\nAverage time to optimize per function 229.128 / 126 = 1.181ms\\r\\nAverage time to emit per function 167.338 / 126 = 1.328ms\\n\\r\\n???? Something looks very wrong here.\\n\\r\\n[1] https://gist.github.com/saicitus/251ba20b211e9e73285af35e61b19580\\r\\n[2] https://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2a\\n\\r\\n-- \\r\\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\r\\n\\\"Java is clearly an example of money oriented programming\\\" (A. Stepanov)\\n\\nHi,For the 3rd query, I noticed some difference in options: Options: Inlining false, Optimization true, Expressions true, Deforming true Inlining was true for the first two queries.FYI\",\n \"msg_date\": \"Sun, 14 Nov 2021 09:22:04 -0800\",\n \"msg_from\": \"Zhihong Yu <zyu@yugabyte.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: JIT doing duplicative optimization?\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> On 2021-Nov-11, Alvaro Herrera wrote:\\n>> But what really surprised me is that the the average time to optimize\\n>> per function is now 2.06ms ... less than half of the previous\\n>> measurement. It emits 10% less functions than before, but the time to\\n>> both optimize and emit is reduced by 50%. How does that make sense?\\n\\n> Ah, here's a query that illustrates what I'm on about. I found this\\n> query[1] in a blog post[2].\\n> ...\\n\\n> Query 1, 148 functions JIT-compiled.\\n> Average time to optimize, per function 435.153/148 = 2.940ms;\\n> average time to emit per function 282.216/148 = 1.906ms\\n\\n> Query 2, 137 functions JIT-compiled.\\n> Average time to optimize, per function: 374.103/137 = 2.730ms\\n> Average time to emit, per function 254.557 / 137 = 1.858ms\\n\\n> Query 3, 126 functions JIT-compiled.\\n> Average time to optimize per function 229.128 / 126 = 1.181ms\\n> Average time to emit per function 167.338 / 126 = 1.328ms\\n\\nYeah, in combination with your other measurement, it sure does look like\\nthere's something worse-than-linear going on here. The alternative is to\\nassume that the individual functions are more complex in one query than\\nthe other, and that seems like a bit of a stretch.\\n\\nYou could probably generate some queries with lots and lots of expressions\\nto characterize this better. If it is O(N^2), it should not be hard to\\ndrive the cost up to the point where the guilty bit of code would stand\\nout in a perf trace.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 14 Nov 2021 13:30:35 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: JIT doing duplicative optimization?\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> You could probably generate some queries with lots and lots of expressions\\n> to characterize this better. If it is O(N^2), it should not be hard to\\n> drive the cost up to the point where the guilty bit of code would stand\\n> out in a perf trace.\\n\\nI experimented with that, using a few different-size queries generated\\nlike this:\\n\\nprint \\\"explain analyze\\\\n\\\";\\nfor (my $i = 1; $i < 100; $i++) {\\n\\tprint \\\" select sum(f1+$i) from base union all\\\\n\\\";\\n}\\nprint \\\"select sum(f1+0) from base;\\\\n\\\";\\n\\non a table made like\\n\\ncreate table base as select generate_series(1,10000000) f1;\\n\\nWhat I got, after setting max_parallel_workers_per_gather = 0,\\nwas\\n\\n10 subqueries:\\n\\n Planning Time: 0.260 ms\\n JIT:\\n Functions: 30\\n Options: Inlining true, Optimization true, Expressions true, Deforming true\\n Timing: Generation 4.651 ms, Inlining 8.870 ms, Optimization 152.937 ms, Emis\\nsion 95.046 ms, Total 261.504 ms\\n Execution Time: 15258.249 ms\\n\\n100 subqueries:\\n\\n Planning Time: 2.231 ms\\n JIT:\\n Functions: 300\\n Options: Inlining true, Optimization true, Expressions true, Deforming true\\n Timing: Generation 44.163 ms, Inlining 9.934 ms, Optimization 1448.971 ms, Em\\nission 928.438 ms, Total 2431.506 ms\\n Execution Time: 154815.515 ms\\n\\n1000 subqueries:\\n\\n Planning Time: 29.480 ms\\n JIT:\\n Functions: 3000\\n Options: Inlining true, Optimization true, Expressions true, Deforming true\\n Timing: Generation 444.479 ms, Inlining 25.688 ms, Optimization 14989.696 ms,\\n Emission 9891.993 ms, Total 25351.856 ms\\n Execution Time: 1522011.367 ms\\n\\nSo the overhead looks pretty linear, or even a shade sublinear for the\\n\\\"inlining\\\" bit, *as long as only one process is involved*. However,\\nI noted that if I didn't force that, the JIT overhead went up because\\nthe planner wanted to use more workers and each worker has to do its own\\ncompilations. So perhaps the apparent nonlinearity in your examples comes\\nfrom that?\\n\\nBTW, I realized while working on this that I have little idea what the\\n\\\"Functions:\\\" count is. Nor does our documentation explain that (or any\\nother of these numbers), at least not anywhere I could find. That seems\\nlike a pretty serious documentation fail. If these numbers aren't\\nimportant enough to explain in the docs, why are we printing them at all?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 14 Nov 2021 18:46:34 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: JIT doing duplicative optimization?\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2021-11-11 17:02:09 -0300, Alvaro Herrera wrote:\\n> Recently I noticed extremely suspicious behavior from queries under JIT.\\n> A query (details for which sadly I cannot disclose) had an EXPLAIN\\n> looking like this:\\n> \\n> JIT:\\n> Functions: 743\\n> Options: Inlining true, Optimization true, Expressions true, Deforming true\\n> Timing: Generation 47.013 ms, Inlining 20.277 ms, Optimization 3269.381 ms, Emission 2392.115 ms, Total 5728.786 ms\\n> Execution Time: 5735.552 ms\\n> \\n> Disabling JIT, the query takes a few tens of milliseconds, so of course\\n> the typical recommendation is to disable JIT. But I noted that the\\n> average time to optimize, per function, is 4.7ms. I didn't know what to\\n> make of this -- is it a reasonable to to take to optimize a single\\n> function?\\n\\nNo, that's pretty insane. Any chance you share the generated bitcode for this\\n- it'll have some details about the query / table structure though. Perhaps\\njust the contained function names?\\n\\n\\n> I put that thought aside. But then I spotted a couple of minor things\\n> in the query that could be made better; so I did some of that, and\\n> noticed that the number of functions was reduced from 743 to 671, and\\n> the JIT details in the explain looked like this:\\n\\n> But what really surprised me is that the the average time to optimize\\n> per function is now 2.06ms ... less than half of the previous\\n> measurement. It emits 10% less functions than before, but the time to\\n> both optimize and emit is reduced by 50%. How does that make sense?\\n\\nWithout knowing the concrete query I don't know if it it's that. But\\n\\n\\n> I don't know how to approach an investigation into this, but I thought\\n> that maybe we're doing quadratic effort in the optimization / emission\\n> phase or something like that.\\n\\nYea. One way to investigate it is to enable jit_dump_bitcode and then use\\nllvm's 'opt' tool to see details about optimization times.\\n\\nopt -time-passes -O3 -o /dev/null < /srv/dev/pgdev-dev/3990041.77.bc |less\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 09:13:46 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: JIT doing duplicative optimization?\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2021-11-14 18:46:34 -0500, Tom Lane wrote:\\n> BTW, I realized while working on this that I have little idea what the\\n> \\\"Functions:\\\" count is. Nor does our documentation explain that (or any\\n> other of these numbers), at least not anywhere I could find. That seems\\n> like a pretty serious documentation fail.\\n\\nIt's literally the number of functions created in the JIT code. That could be\\nfunctions for evaluating one expression, or functions for deforming tuples of\\none type of tupledesc. It's the simplest measure of \\\"amount of JITed stuff\\\" I\\ncould find. I'd proposed a more detailed explain option where we'd optionally\\nprint details about what exactly is JITed where a while back, but there were\\nsome explain output format challenges...\\n\\n\\n> If these numbers aren't important enough to explain in the docs, why are we\\n> printing them at all?\\n\\nAfaict we don't have detailed docs for EXPLAIN output in general?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 09:18:27 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: JIT doing duplicative optimization?\"\n }\n]","isConversational":false}}},{"rowIdx":53077,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"\nperldoc Test::More says:\n\n Key feature milestones include:\n\n subtests\n Subtests were released in Test::More 0.94, which came with Perl\n 5.12.0. Subtests did not implicitly call \"done_testing()\" until\n 0.96; the first Perl with that fix was Perl 5.14.0 with 0.98.\n\n \"done_testing()\"\n This was released in Test::More 0.88 and first shipped with Perl in\n 5.10.1 as part of Test::More 0.92.\n\nWe actually test for a version >= 0.87 both in PostgreSQL::Test::Utils\nand in configure.ac. That seems wrong, since we need done_testing(). \nBut I'd really like it if we could shift the goalposts a bit and require\n0.96. subtests would be really nice to have available. Unfortunately I\ndon't have any idea what versions of Test::More are actually in use. \nPerhaps we could add a line to one of the TAP tests that would spit out\nthe version on the log?\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Thu, 11 Nov 2021 15:04:25 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":true,"msg_subject":"Test::More version"},{"msg_contents":"On 2021-Nov-11, Andrew Dunstan wrote:\n\n> But I'd really like it if we could shift the goalposts a bit and require\n> 0.96. subtests would be really nice to have available.\n\nAgreed, subtests would be nice.\n\n> Perhaps we could add a line to one of the TAP tests that would spit out\n> the version on the log?\n\nMaybe have it spit out the version of *all* the modules we require\n(which aren't all that many), not just Test::More?\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Thu, 11 Nov 2021 17:31:59 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Test::More version"},{"msg_contents":"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> On 2021-Nov-11, Andrew Dunstan wrote:\n>> Perhaps we could add a line to one of the TAP tests that would spit out\n>> the version on the log?\n\n> Maybe have it spit out the version of *all* the modules we require\n> (which aren't all that many), not just Test::More?\n\nYeah ... configure is already checking those versions, so maybe we could\nmake it print them out along the way? I'd been thinking of doing exactly\nthis just for documentation purposes, so if there's a concrete need for\nit, let's get it done.\n\nHaving said that, I don't think I believe that changelog entry,\nbecause prairiedog is doing just fine with\n\n$ perl -MTest::More -e 'print \"$Test::More::VERSION\\n\"'\n0.8701\n\nI have no objection to updating prairiedog to whatever new minimum\nversion we settle on. But if there are other buildfarm animals with\nold versions, that might be a reason not to change it.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 11 Nov 2021 15:44:27 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Test::More version"},{"msg_contents":"On 11.11.21 21:04, Andrew Dunstan wrote:\n> But I'd really like it if we could shift the goalposts a bit and require\n> 0.96. subtests would be really nice to have available. Unfortunately I\n> don't have any idea what versions of Test::More are actually in use.\n\nMy initial patch for TAP tests used subtests, and IIRC that was taken \nout because the Perl that comes with CentOS 6 is too old.\n\n\n","msg_date":"Fri, 12 Nov 2021 14:48:54 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Test::More version"},{"msg_contents":"On 11/11/21 15:44, Tom Lane wrote:\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n>> On 2021-Nov-11, Andrew Dunstan wrote:\n>>> Perhaps we could add a line to one of the TAP tests that would spit out\n>>> the version on the log?\n>> Maybe have it spit out the version of *all* the modules we require\n>> (which aren't all that many), not just Test::More?\n> Yeah ... configure is already checking those versions, so maybe we could\n> make it print them out along the way? I'd been thinking of doing exactly\n> this just for documentation purposes, so if there's a concrete need for\n> it, let's get it done.\n>\n\nThe attached seems to do the trick:\n\n\nchecking for perl module IPC::Run 0.79... 20200505.0\nchecking for perl module Test::More 0.87... 1.302183\nchecking for perl module Time::HiRes 1.52... 1.9764\n\n\n>\n> I have no objection to updating prairiedog to whatever new minimum\n> version we settle on. But if there are other buildfarm animals with\n> old versions, that might be a reason not to change it.\n>\n> \t\t\t\n\n\n\nYeah, let's see what's actually in use and then we can decide.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Fri, 12 Nov 2021 10:00:14 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":true,"msg_subject":"Re: Test::More version"},{"msg_contents":"Andrew Dunstan <andrew@dunslane.net> writes:\n> On 11/11/21 15:44, Tom Lane wrote:\n>> Yeah ... configure is already checking those versions, so maybe we could\n>> make it print them out along the way? I'd been thinking of doing exactly\n>> this just for documentation purposes, so if there's a concrete need for\n>> it, let's get it done.\n\n> The attached seems to do the trick:\n\nLGTM (by eyeball, didn't test)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 12 Nov 2021 10:05:22 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Test::More version"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> On 11.11.21 21:04, Andrew Dunstan wrote:\n>> But I'd really like it if we could shift the goalposts a bit and require\n>> 0.96. subtests would be really nice to have available. Unfortunately I\n>> don't have any idea what versions of Test::More are actually in use.\n\n> My initial patch for TAP tests used subtests, and IIRC that was taken \n> out because the Perl that comes with CentOS 6 is too old.\n\nI retired my RHEL6 installation a year ago, but I still have backups\nat hand (...digs...) Yeah, its Test/More.pm says $VERSION = '0.92';\nso your recollection is accurate.\n\nRHEL6 has been EOL for awhile, and that specific platform isn't in\nthe buildfarm, so it needn't constrain our decision. But it may\nwell be that some of the older buildfarm members have similarly\nancient Test::More. I await results from Andrew's patch.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 12 Nov 2021 10:30:14 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Test::More version"},{"msg_contents":"\nOn 11/12/21 10:05, Tom Lane wrote:\n> Andrew Dunstan <andrew@dunslane.net> writes:\n>> On 11/11/21 15:44, Tom Lane wrote:\n>>> Yeah ... configure is already checking those versions, so maybe we could\n>>> make it print them out along the way? I'd been thinking of doing exactly\n>>> this just for documentation purposes, so if there's a concrete need for\n>>> it, let's get it done.\n>> The attached seems to do the trick:\n> LGTM (by eyeball, didn't test)\n\n\n\nThanks. pushed a slightly more robust version.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Fri, 12 Nov 2021 10:42:14 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":true,"msg_subject":"Re: Test::More version"},{"msg_contents":"Andrew Dunstan <andrew@dunslane.net> writes:\n> Thanks. pushed a slightly more robust version.\n\nI think we have enough info now to draw conclusions. There are several\nbuildfarm animals that are running TAP tests but haven't reported since\nthis went in. However, I don't think any of them will yield surprises:\n\n frogfish | 2021-11-10 02:24:27 | checking for perl module Test::More 0.87... ok\n mandrill | 2021-11-12 04:45:42 | checking for perl module Test::More 0.87... ok\n mereswine | 2021-11-12 09:48:49 | checking for perl module Test::More 0.87... ok\n sungazer | 2021-11-10 20:27:12 | checking for perl module Test::More 0.87... ok\n tern | 2021-11-09 05:27:38 | checking for perl module Test::More 0.87... ok\n topminnow | 2021-11-08 05:03:17 | checking for perl module Test::More 0.87... ok\n\nmereswine is running pretty up-to-date Debian, so that's not going to\nbe an old version. mandrill, sungazer, and tern share the same host\nas hornet, which has reported using 1.302183. I cheated and looked at\nfrogfish/topminnow's host, and it has Test::More 0.98:\n\ntgl@erpro8-fsf1:~$ perl -MTest::More -e 'print \"$Test::More::VERSION\\n\"';\n0.98\n\nSo this should be a pretty complete summary of relevant old Test::More\nversions:\n\n prairiedog | 2021-11-13 12:38:32 | checking for perl module Test::More 0.87... 0.8701\n lapwing | 2021-11-13 04:40:12 | checking for perl module Test::More 0.87... 0.98\n prion | 2021-11-13 03:53:16 | checking for perl module Test::More 0.87... 0.98\n snapper | 2021-11-13 17:27:29 | checking for perl module Test::More 0.87... 0.98\n curculio | 2021-11-13 05:15:54 | checking for perl module Test::More 0.87... 1.001002\n butterflyfish | 2021-11-13 06:00:29 | checking for perl module Test::More 0.87... 1.001014\n\nThere are a couple more running 1.001014, and everything else is\n1.302something.\n\nSo, discounting prairiedog's intentionally trailing-edge installation,\nthe oldest stuff in the buildfarm is 0.98, of which there are five\ninstances belonging to four different owners.\n\nBased on this, I'm inclined to think we should select 0.98 as the\nminimum version. Anything later would inconvenience some people.\nOTOH, it doesn't look like there's any value in promising compatibility\nwith 0.96 instead, especially since I see a couple of subplan-related\nbug fixes in 0.97 and 0.98 in Test::Simple's changelog.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 15 Nov 2021 12:43:39 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Test::More version"},{"msg_contents":"Tom Lane <tgl@sss.pgh.pa.us> writes:\n\n> Based on this, I'm inclined to think we should select 0.98 as the\n> minimum version. Anything later would inconvenience some people.\n> OTOH, it doesn't look like there's any value in promising compatibility\n> with 0.96 instead, especially since I see a couple of subplan-related\n> bug fixes in 0.97 and 0.98 in Test::Simple's changelog.\n\nFWIW, here's a list of Test::More versions shipped with various Perl\nreleases. Since v5.12.0, the version does not change in minor releases.\n\n v5.8.3 0.47\n v5.8.7 0.54\n v5.8.8 0.62\n v5.8.9 0.80\n v5.10.0 0.72\n v5.10.1 0.92\n v5.12.0 0.94\n v5.14.0 0.98\n v5.16.0 0.98\n v5.18.0 0.98\n v5.20.0 1.001002\n v5.22.0 1.001014\n v5.24.0 1.001014\n v5.26.0 1.302073\n v5.28.0 1.302133\n v5.30.0 1.302162\n v5.32.0 1.302175\n v5.34.0 1.302183\n\n- ilmari\n\n\n","msg_date":"Mon, 15 Nov 2021 18:42:53 +0000","msg_from":"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org>","msg_from_op":false,"msg_subject":"Re: Test::More version"},{"msg_contents":"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org> writes:\n> Tom Lane <tgl@sss.pgh.pa.us> writes:\n>> Based on this, I'm inclined to think we should select 0.98 as the\n>> minimum version. Anything later would inconvenience some people.\n>> OTOH, it doesn't look like there's any value in promising compatibility\n>> with 0.96 instead, especially since I see a couple of subplan-related\n>> bug fixes in 0.97 and 0.98 in Test::Simple's changelog.\n\n> FWIW, here's a list of Test::More versions shipped with various Perl\n> releases. Since v5.12.0, the version does not change in minor releases.\n> ...\n> v5.12.0 0.94\n> v5.14.0 0.98\n\nAh-hah, that's a good way to think about it too. There's surely\nlittle reason in worrying about a module version that did not ship\nin any Perl release, because what 99.9% of people will have is what\nshipped with their Perl release.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 15 Nov 2021 14:14:36 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Test::More version"},{"msg_contents":"Tom Lane <tgl@sss.pgh.pa.us> writes:\n\n> =?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org> writes:\n\nYour MUA seems a bit MIME-deficient…\n\n>> Tom Lane <tgl@sss.pgh.pa.us> writes:\n>>> Based on this, I'm inclined to think we should select 0.98 as the\n>>> minimum version. Anything later would inconvenience some people.\n>>> OTOH, it doesn't look like there's any value in promising compatibility\n>>> with 0.96 instead, especially since I see a couple of subplan-related\n>>> bug fixes in 0.97 and 0.98 in Test::Simple's changelog.\n>\n>> FWIW, here's a list of Test::More versions shipped with various Perl\n>> releases. Since v5.12.0, the version does not change in minor releases.\n>> ...\n>> v5.12.0 0.94\n>> v5.14.0 0.98\n>\n> Ah-hah, that's a good way to think about it too. There's surely\n> little reason in worrying about a module version that did not ship\n> in any Perl release, because what 99.9% of people will have is what\n> shipped with their Perl release.\n\nOr newer, if they ever install anything from CPAN, since many modules\ndepend on a newer version. Some OSes even package newer versions\nseparately in addition to the core version, which install earlier in the\ninclude path. E.g. Debian has the following versions:\n\nDebian | Perl | core | packaged\n8 | 5.20.2 | 1.001002 | 1.001008\n9 | 5.24.1 | 1.001014 | 1.302075\n10 | 5.28.1 | 1.302133 | 1.302162\n11 | 5.32.1 | 1.302175 | 1.302183\n\n- ilmari\n\n\n","msg_date":"Mon, 15 Nov 2021 20:19:43 +0000","msg_from":"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org>","msg_from_op":false,"msg_subject":"Re: Test::More version"},{"msg_contents":"\nOn 11/15/21 14:14, Tom Lane wrote:\n> =?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org> writes:\n>> Tom Lane <tgl@sss.pgh.pa.us> writes:\n>>> Based on this, I'm inclined to think we should select 0.98 as the\n>>> minimum version. Anything later would inconvenience some people.\n>>> OTOH, it doesn't look like there's any value in promising compatibility\n>>> with 0.96 instead, especially since I see a couple of subplan-related\n>>> bug fixes in 0.97 and 0.98 in Test::Simple's changelog.\n>> FWIW, here's a list of Test::More versions shipped with various Perl\n>> releases. Since v5.12.0, the version does not change in minor releases.\n>> ...\n>> v5.12.0 0.94\n>> v5.14.0 0.98\n> Ah-hah, that's a good way to think about it too. There's surely\n> little reason in worrying about a module version that did not ship\n> in any Perl release, because what 99.9% of people will have is what\n> shipped with their Perl release.\n>\n> \t\t\t\n\n\nAnd newer versions of Test::Simple/Test::More still work with very old\nversions of perl. jacana runs prove with perl 5.8.8 but has Test::More\n1.001014\n\nSo I agree 0.98 seems like a perfectly reasonable target. You should be\nable to drop a replacement on prairiedog quite simply. For jacana I just\nunpacked it and pointed PERL5LIB at it.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Tue, 16 Nov 2021 11:05:43 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":true,"msg_subject":"Re: Test::More version"},{"msg_contents":"\nOn 11/15/21 12:43, Tom Lane wrote:\n>\n> So, discounting prairiedog's intentionally trailing-edge installation,\n> the oldest stuff in the buildfarm is 0.98, of which there are five\n> instances belonging to four different owners.\n>\n> Based on this, I'm inclined to think we should select 0.98 as the\n> minimum version. Anything later would inconvenience some people.\n> OTOH, it doesn't look like there's any value in promising compatibility\n> with 0.96 instead, especially since I see a couple of subplan-related\n> bug fixes in 0.97 and 0.98 in Test::Simple's changelog.\n>\n> \t\n\n\n\nYeah, so I think at this stage we're just waiting for you to update\nprairiedog and we can make this change.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Sat, 20 Nov 2021 09:08:27 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":true,"msg_subject":"Re: Test::More version"},{"msg_contents":"Andrew Dunstan <andrew@dunslane.net> writes:\n> Yeah, so I think at this stage we're just waiting for you to update\n> prairiedog and we can make this change.\n\nOh! I was intentionally waiting to do that, with the idea of verifying\nthat the version-detection test works ;-). I'm prepared to do it as\nsoon as you push an update and we see prairiedog go red.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 20 Nov 2021 11:14:14 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Test::More version"},{"msg_contents":"\nOn 11/20/21 11:14, Tom Lane wrote:\n> Andrew Dunstan <andrew@dunslane.net> writes:\n>> Yeah, so I think at this stage we're just waiting for you to update\n>> prairiedog and we can make this change.\n> Oh! I was intentionally waiting to do that, with the idea of verifying\n> that the version-detection test works ;-). I'm prepared to do it as\n> soon as you push an update and we see prairiedog go red.\n\n\nAh, ok. Pushed.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Sat, 20 Nov 2021 17:58:07 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":true,"msg_subject":"Re: Test::More version"}],"string":"[\n {\n \"msg_contents\": \"\\nperldoc Test::More says:\\n\\n Key feature milestones include:\\n\\n subtests\\n Subtests were released in Test::More 0.94, which came with Perl\\n 5.12.0. Subtests did not implicitly call \\\"done_testing()\\\" until\\n 0.96; the first Perl with that fix was Perl 5.14.0 with 0.98.\\n\\n \\\"done_testing()\\\"\\n This was released in Test::More 0.88 and first shipped with Perl in\\n 5.10.1 as part of Test::More 0.92.\\n\\nWe actually test for a version >= 0.87 both in PostgreSQL::Test::Utils\\nand in configure.ac. That seems wrong, since we need done_testing(). \\nBut I'd really like it if we could shift the goalposts a bit and require\\n0.96. subtests would be really nice to have available. Unfortunately I\\ndon't have any idea what versions of Test::More are actually in use. \\nPerhaps we could add a line to one of the TAP tests that would spit out\\nthe version on the log?\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 15:04:25 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Test::More version\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-11, Andrew Dunstan wrote:\\n\\n> But I'd really like it if we could shift the goalposts a bit and require\\n> 0.96. subtests would be really nice to have available.\\n\\nAgreed, subtests would be nice.\\n\\n> Perhaps we could add a line to one of the TAP tests that would spit out\\n> the version on the log?\\n\\nMaybe have it spit out the version of *all* the modules we require\\n(which aren't all that many), not just Test::More?\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 17:31:59 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> On 2021-Nov-11, Andrew Dunstan wrote:\\n>> Perhaps we could add a line to one of the TAP tests that would spit out\\n>> the version on the log?\\n\\n> Maybe have it spit out the version of *all* the modules we require\\n> (which aren't all that many), not just Test::More?\\n\\nYeah ... configure is already checking those versions, so maybe we could\\nmake it print them out along the way? I'd been thinking of doing exactly\\nthis just for documentation purposes, so if there's a concrete need for\\nit, let's get it done.\\n\\nHaving said that, I don't think I believe that changelog entry,\\nbecause prairiedog is doing just fine with\\n\\n$ perl -MTest::More -e 'print \\\"$Test::More::VERSION\\\\n\\\"'\\n0.8701\\n\\nI have no objection to updating prairiedog to whatever new minimum\\nversion we settle on. But if there are other buildfarm animals with\\nold versions, that might be a reason not to change it.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 15:44:27 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"On 11.11.21 21:04, Andrew Dunstan wrote:\\n> But I'd really like it if we could shift the goalposts a bit and require\\n> 0.96. subtests would be really nice to have available. Unfortunately I\\n> don't have any idea what versions of Test::More are actually in use.\\n\\nMy initial patch for TAP tests used subtests, and IIRC that was taken \\nout because the Perl that comes with CentOS 6 is too old.\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 14:48:54 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"On 11/11/21 15:44, Tom Lane wrote:\\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n>> On 2021-Nov-11, Andrew Dunstan wrote:\\n>>> Perhaps we could add a line to one of the TAP tests that would spit out\\n>>> the version on the log?\\n>> Maybe have it spit out the version of *all* the modules we require\\n>> (which aren't all that many), not just Test::More?\\n> Yeah ... configure is already checking those versions, so maybe we could\\n> make it print them out along the way? I'd been thinking of doing exactly\\n> this just for documentation purposes, so if there's a concrete need for\\n> it, let's get it done.\\n>\\n\\nThe attached seems to do the trick:\\n\\n\\nchecking for perl module IPC::Run 0.79... 20200505.0\\nchecking for perl module Test::More 0.87... 1.302183\\nchecking for perl module Time::HiRes 1.52... 1.9764\\n\\n\\n>\\n> I have no objection to updating prairiedog to whatever new minimum\\n> version we settle on. But if there are other buildfarm animals with\\n> old versions, that might be a reason not to change it.\\n>\\n> \\t\\t\\t\\n\\n\\n\\nYeah, let's see what's actually in use and then we can decide.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Fri, 12 Nov 2021 10:00:14 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew@dunslane.net> writes:\\n> On 11/11/21 15:44, Tom Lane wrote:\\n>> Yeah ... configure is already checking those versions, so maybe we could\\n>> make it print them out along the way? I'd been thinking of doing exactly\\n>> this just for documentation purposes, so if there's a concrete need for\\n>> it, let's get it done.\\n\\n> The attached seems to do the trick:\\n\\nLGTM (by eyeball, didn't test)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 10:05:22 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> On 11.11.21 21:04, Andrew Dunstan wrote:\\n>> But I'd really like it if we could shift the goalposts a bit and require\\n>> 0.96. subtests would be really nice to have available. Unfortunately I\\n>> don't have any idea what versions of Test::More are actually in use.\\n\\n> My initial patch for TAP tests used subtests, and IIRC that was taken \\n> out because the Perl that comes with CentOS 6 is too old.\\n\\nI retired my RHEL6 installation a year ago, but I still have backups\\nat hand (...digs...) Yeah, its Test/More.pm says $VERSION = '0.92';\\nso your recollection is accurate.\\n\\nRHEL6 has been EOL for awhile, and that specific platform isn't in\\nthe buildfarm, so it needn't constrain our decision. But it may\\nwell be that some of the older buildfarm members have similarly\\nancient Test::More. I await results from Andrew's patch.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 10:30:14 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"\\nOn 11/12/21 10:05, Tom Lane wrote:\\n> Andrew Dunstan <andrew@dunslane.net> writes:\\n>> On 11/11/21 15:44, Tom Lane wrote:\\n>>> Yeah ... configure is already checking those versions, so maybe we could\\n>>> make it print them out along the way? I'd been thinking of doing exactly\\n>>> this just for documentation purposes, so if there's a concrete need for\\n>>> it, let's get it done.\\n>> The attached seems to do the trick:\\n> LGTM (by eyeball, didn't test)\\n\\n\\n\\nThanks. pushed a slightly more robust version.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 10:42:14 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew@dunslane.net> writes:\\n> Thanks. pushed a slightly more robust version.\\n\\nI think we have enough info now to draw conclusions. There are several\\nbuildfarm animals that are running TAP tests but haven't reported since\\nthis went in. However, I don't think any of them will yield surprises:\\n\\n frogfish | 2021-11-10 02:24:27 | checking for perl module Test::More 0.87... ok\\n mandrill | 2021-11-12 04:45:42 | checking for perl module Test::More 0.87... ok\\n mereswine | 2021-11-12 09:48:49 | checking for perl module Test::More 0.87... ok\\n sungazer | 2021-11-10 20:27:12 | checking for perl module Test::More 0.87... ok\\n tern | 2021-11-09 05:27:38 | checking for perl module Test::More 0.87... ok\\n topminnow | 2021-11-08 05:03:17 | checking for perl module Test::More 0.87... ok\\n\\nmereswine is running pretty up-to-date Debian, so that's not going to\\nbe an old version. mandrill, sungazer, and tern share the same host\\nas hornet, which has reported using 1.302183. I cheated and looked at\\nfrogfish/topminnow's host, and it has Test::More 0.98:\\n\\ntgl@erpro8-fsf1:~$ perl -MTest::More -e 'print \\\"$Test::More::VERSION\\\\n\\\"';\\n0.98\\n\\nSo this should be a pretty complete summary of relevant old Test::More\\nversions:\\n\\n prairiedog | 2021-11-13 12:38:32 | checking for perl module Test::More 0.87... 0.8701\\n lapwing | 2021-11-13 04:40:12 | checking for perl module Test::More 0.87... 0.98\\n prion | 2021-11-13 03:53:16 | checking for perl module Test::More 0.87... 0.98\\n snapper | 2021-11-13 17:27:29 | checking for perl module Test::More 0.87... 0.98\\n curculio | 2021-11-13 05:15:54 | checking for perl module Test::More 0.87... 1.001002\\n butterflyfish | 2021-11-13 06:00:29 | checking for perl module Test::More 0.87... 1.001014\\n\\nThere are a couple more running 1.001014, and everything else is\\n1.302something.\\n\\nSo, discounting prairiedog's intentionally trailing-edge installation,\\nthe oldest stuff in the buildfarm is 0.98, of which there are five\\ninstances belonging to four different owners.\\n\\nBased on this, I'm inclined to think we should select 0.98 as the\\nminimum version. Anything later would inconvenience some people.\\nOTOH, it doesn't look like there's any value in promising compatibility\\nwith 0.96 instead, especially since I see a couple of subplan-related\\nbug fixes in 0.97 and 0.98 in Test::Simple's changelog.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 12:43:39 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"Tom Lane <tgl@sss.pgh.pa.us> writes:\\n\\n> Based on this, I'm inclined to think we should select 0.98 as the\\n> minimum version. Anything later would inconvenience some people.\\n> OTOH, it doesn't look like there's any value in promising compatibility\\n> with 0.96 instead, especially since I see a couple of subplan-related\\n> bug fixes in 0.97 and 0.98 in Test::Simple's changelog.\\n\\nFWIW, here's a list of Test::More versions shipped with various Perl\\nreleases. Since v5.12.0, the version does not change in minor releases.\\n\\n v5.8.3 0.47\\n v5.8.7 0.54\\n v5.8.8 0.62\\n v5.8.9 0.80\\n v5.10.0 0.72\\n v5.10.1 0.92\\n v5.12.0 0.94\\n v5.14.0 0.98\\n v5.16.0 0.98\\n v5.18.0 0.98\\n v5.20.0 1.001002\\n v5.22.0 1.001014\\n v5.24.0 1.001014\\n v5.26.0 1.302073\\n v5.28.0 1.302133\\n v5.30.0 1.302162\\n v5.32.0 1.302175\\n v5.34.0 1.302183\\n\\n- ilmari\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 18:42:53 +0000\",\n \"msg_from\": \"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org> writes:\\n> Tom Lane <tgl@sss.pgh.pa.us> writes:\\n>> Based on this, I'm inclined to think we should select 0.98 as the\\n>> minimum version. Anything later would inconvenience some people.\\n>> OTOH, it doesn't look like there's any value in promising compatibility\\n>> with 0.96 instead, especially since I see a couple of subplan-related\\n>> bug fixes in 0.97 and 0.98 in Test::Simple's changelog.\\n\\n> FWIW, here's a list of Test::More versions shipped with various Perl\\n> releases. Since v5.12.0, the version does not change in minor releases.\\n> ...\\n> v5.12.0 0.94\\n> v5.14.0 0.98\\n\\nAh-hah, that's a good way to think about it too. There's surely\\nlittle reason in worrying about a module version that did not ship\\nin any Perl release, because what 99.9% of people will have is what\\nshipped with their Perl release.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 14:14:36 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"Tom Lane <tgl@sss.pgh.pa.us> writes:\\n\\n> =?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org> writes:\\n\\nYour MUA seems a bit MIME-deficient…\\n\\n>> Tom Lane <tgl@sss.pgh.pa.us> writes:\\n>>> Based on this, I'm inclined to think we should select 0.98 as the\\n>>> minimum version. Anything later would inconvenience some people.\\n>>> OTOH, it doesn't look like there's any value in promising compatibility\\n>>> with 0.96 instead, especially since I see a couple of subplan-related\\n>>> bug fixes in 0.97 and 0.98 in Test::Simple's changelog.\\n>\\n>> FWIW, here's a list of Test::More versions shipped with various Perl\\n>> releases. Since v5.12.0, the version does not change in minor releases.\\n>> ...\\n>> v5.12.0 0.94\\n>> v5.14.0 0.98\\n>\\n> Ah-hah, that's a good way to think about it too. There's surely\\n> little reason in worrying about a module version that did not ship\\n> in any Perl release, because what 99.9% of people will have is what\\n> shipped with their Perl release.\\n\\nOr newer, if they ever install anything from CPAN, since many modules\\ndepend on a newer version. Some OSes even package newer versions\\nseparately in addition to the core version, which install earlier in the\\ninclude path. E.g. Debian has the following versions:\\n\\nDebian | Perl | core | packaged\\n8 | 5.20.2 | 1.001002 | 1.001008\\n9 | 5.24.1 | 1.001014 | 1.302075\\n10 | 5.28.1 | 1.302133 | 1.302162\\n11 | 5.32.1 | 1.302175 | 1.302183\\n\\n- ilmari\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 20:19:43 +0000\",\n \"msg_from\": \"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"\\nOn 11/15/21 14:14, Tom Lane wrote:\\n> =?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org> writes:\\n>> Tom Lane <tgl@sss.pgh.pa.us> writes:\\n>>> Based on this, I'm inclined to think we should select 0.98 as the\\n>>> minimum version. Anything later would inconvenience some people.\\n>>> OTOH, it doesn't look like there's any value in promising compatibility\\n>>> with 0.96 instead, especially since I see a couple of subplan-related\\n>>> bug fixes in 0.97 and 0.98 in Test::Simple's changelog.\\n>> FWIW, here's a list of Test::More versions shipped with various Perl\\n>> releases. Since v5.12.0, the version does not change in minor releases.\\n>> ...\\n>> v5.12.0 0.94\\n>> v5.14.0 0.98\\n> Ah-hah, that's a good way to think about it too. There's surely\\n> little reason in worrying about a module version that did not ship\\n> in any Perl release, because what 99.9% of people will have is what\\n> shipped with their Perl release.\\n>\\n> \\t\\t\\t\\n\\n\\nAnd newer versions of Test::Simple/Test::More still work with very old\\nversions of perl. jacana runs prove with perl 5.8.8 but has Test::More\\n1.001014\\n\\nSo I agree 0.98 seems like a perfectly reasonable target. You should be\\nable to drop a replacement on prairiedog quite simply. For jacana I just\\nunpacked it and pointed PERL5LIB at it.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 11:05:43 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"\\nOn 11/15/21 12:43, Tom Lane wrote:\\n>\\n> So, discounting prairiedog's intentionally trailing-edge installation,\\n> the oldest stuff in the buildfarm is 0.98, of which there are five\\n> instances belonging to four different owners.\\n>\\n> Based on this, I'm inclined to think we should select 0.98 as the\\n> minimum version. Anything later would inconvenience some people.\\n> OTOH, it doesn't look like there's any value in promising compatibility\\n> with 0.96 instead, especially since I see a couple of subplan-related\\n> bug fixes in 0.97 and 0.98 in Test::Simple's changelog.\\n>\\n> \\t\\n\\n\\n\\nYeah, so I think at this stage we're just waiting for you to update\\nprairiedog and we can make this change.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 20 Nov 2021 09:08:27 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew@dunslane.net> writes:\\n> Yeah, so I think at this stage we're just waiting for you to update\\n> prairiedog and we can make this change.\\n\\nOh! I was intentionally waiting to do that, with the idea of verifying\\nthat the version-detection test works ;-). I'm prepared to do it as\\nsoon as you push an update and we see prairiedog go red.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 20 Nov 2021 11:14:14 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Test::More version\"\n },\n {\n \"msg_contents\": \"\\nOn 11/20/21 11:14, Tom Lane wrote:\\n> Andrew Dunstan <andrew@dunslane.net> writes:\\n>> Yeah, so I think at this stage we're just waiting for you to update\\n>> prairiedog and we can make this change.\\n> Oh! I was intentionally waiting to do that, with the idea of verifying\\n> that the version-detection test works ;-). I'm prepared to do it as\\n> soon as you push an update and we see prairiedog go red.\\n\\n\\nAh, ok. Pushed.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 20 Nov 2021 17:58:07 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Test::More version\"\n }\n]","isConversational":false}}},{"rowIdx":53078,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Welcome to the third installment of the CF 2021-11 patch triage, looking at the\npatches which have been in four CF's and parts of the ones with three CF's. As\nper usual I might very well have misunderstood something (or many things), but\nthere seem to be a fair few ones which should be closeable in this CF here.\n\n\n2962: Fix DROP TABLESPACE on Windows with ProcSignalBarrier?\n============================================================\nThis fixes an old bug which can prevent DROP TABLESPACE from working on\nWindows. There has been a fair bit of review and testing, by among other\nmyself, and I feel this patch is in good enough shape to be considered for\nReady for Committer. There is a goalpost movement request in the thread for\nDROP DATABASE support, but I think we can get this in in two commits to move us\nforward. What are your thoughts on this one Thomas?\n\n\n2979: Support tab completion for upper character inputs in psql\n===============================================================\nThere is broad concensus that this patch fixes a less than ideal behavior in\npsql, and the patch has seen a lot of review and rework. It currently needs a\nreview of the latest version, but skimming it and judging by the thread I'd say\nit's close to Ready for Committer (if not already).\n\n\n2942: increase size of pg_commit_ts buffers\n===========================================\nProduction use has validated this patch as fixing a performance regression, and\nthe patch has support from another committer. There is mention of patch 2627\n(More scalable multixacts buffers and locking) in the thread as an alternative\napproach, that buffers potentially should be configurable instead. Since that\npatch seems like further out from landing, should we go ahead with this to have\nat least a partial solution for 15? Do have any plans to go ahead with this\npatch Álvaro or are you awaiting results on the configurable approach?\n\n\n2992: Allow batched insert during cross-partition updates\n=========================================================\nThis is a follow-up to 927f453a94106 to support batched inserts of internal\nINSERT commands which are used in cross-partition updates for local/remote\npartitions. The patch was after a lot of review initially marked Ready for\nCommitter but went to WOA after failing to apply. Once updated it went to Need\nReview which IMO is misleading given it's status and a skim so I'm moving it to\nRFC again. Georgios, having reviewed it earlier do you any thoughts on the\nlatest version?\n\n\n2970: Set default transactions to read-only at servers start in pg_upgrade\n==========================================================================\nGoalpost movement saw this patch go from read-only transactions to stopping\nbgworkers from running at all during pg_upgrade. There is agreement on the\napproach, but discussion on the exact implementation (which has since trailed\noff). I have on my TODO to bring this over the finish line.\n\n\n2948: Simplify some RI checks to reduce SPI overhead\n====================================================\nMarked ready for committer it seems that raised concerns have been addressed.\nPeter, being listed as reviewer do you have any thoughts on this?\n\n\n2957: Identify missing publications from publisher while create/alter\nsubscription.\n=====================================================================\nPatch is being actively reviewed, has been kept up to date, and was marked RFC\nback in May. Hopefully we can close this soon then.\n\n\n2968: Minimal logical decoding on standbys (take 6)\n===================================================\nThe latest revision of the patch has had feedback from Robert (and Andres)\nwhich needs to be addressed, and the patch no longer applies (after the recent\nThisTimeLineID changes perhaps?). I'm moving this patch to WoA for now.\n\n\n2999: Catalog version access\n============================\nThe initial patch in this thread was dismissed for not providing enough\ninformation to solve the stated problem, but there has since been a patch\nposted which performs ReadControlFile as recommended in the thread. Vik, can\nyou confirm if this patch solves your original problem?\n\n\n2996: Speed up verifying UTF-8\n==============================\nAn IMO interesting approach to speeding up input validation with promising\nresults posted. The discussion seem to be quite confident with the patch, but\nthe latest version takes out a portion of the previous patch which was deemed\noverly complicated. Heikki, are you up for reviewing this version?\n\n\n2932: Partial foreign key updates in referential integrity triggers\n===================================================================\nA recently posted update from Peter indicates that this is close to being\ncommittable, so fingers crossed.\n\n\n3025: postgres_fdw: suppress explicit casts in text:text comparisons\n====================================================================\nCurrently undergoing active review and receiving updates with a new version\nposted today by Tom. Hopefully it can be closed in this CF.\n\n\n2947: Full support for index LP_DEAD hint bits on standby\n=========================================================\nThis has potential to be a nice performance boost for a common usecase for\nscalable production setups. The patch changed course half-way through but has\nsince been marked Ready for Committer. Any takers on this?\n\n\n3120: Rewriting the test of pg_upgrade as a TAP test - take three\n=================================================================\nCurrently this is stuck on a discussion around how to deal with test.sh and the\neffects on the buildfarm. It seems like the questions raised have been\nanswered, but we need to settle on how to continue.\n\n\n3230: Column filtering in logical replication\n=============================================\nThis fails badly to apply, which is likely because there has been a lot of\ndiscussion on the grammar making a new patch not really an option until\nsettled. That discussion ended in late September so hopefully there is a new\nversion during this CF, else we'll have to close this with RwF awaiting a new\npatch in a later CF.\n\n\n3223: Delegating superuser tasks to new security roles\n======================================================\nThere is a lot of ongoing discussion on this one, but skimming the thread it's\nnot really clear (to me) where it leaves the patch in question.\n\n\n3138: Support for NSS as a libpq TLS backend\n============================================\nI'm clearly fairly biased here but I hope that at 48 revisions in the patch is\nin good enough shape to be close to RFC. Joshua Brindle has recently picked up\nreviewing it so hope we can reach somewhere for 15 with this work.\n\n\n3035: Removing unused trailing linepointers from heap pages\n===========================================================\nAn alternative to this was committed for 14, which led the author to change the\npatch into a follow-up to that commit. The merit and value of this second\npatch is debated, with seemingly good arguments both for and against. There\nwas an ask for testcases to provide evidence of correctness back in August\nwhich has gone unanswered, so unless this moves during this CF I think we\nshould mark it RWF awaiting a new version of the patch.\n\n\n3058: Eliminating \"Permission denied\" errors on stat() (on Windows)\n===================================================================\nThis was committed back in July, and then promptly reverted as there were\nissues with mingw. There has since been a patch posted addressing the problem\nwhich needs review. Michael, having looked at the earlier version, do you feel\nup for taking a look at the reworked one?\n\n\n3172: Consistent use of SSL/TLS in docs\n=======================================\nThe proposal to use SSL/TLS instead of just SSL in the docs came about after a\nsmall discussion in the NSS thread, but it wasn't to anyones liking so the\nreworked patch just adds notes that SSL and TLS are used interchangeably in the\ndocs. This version is a pretty small addition that I think we should go ahead\nwith, but more eyes are needed to validate that.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Thu, 11 Nov 2021 23:12:23 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Commitfest 2021-11 Patch Triage - Part 3"},{"msg_contents":"\n\n> On Nov 11, 2021, at 2:12 PM, Daniel Gustafsson <daniel@yesql.se> wrote:\n> \n> 3223: Delegating superuser tasks to new security roles\n> ======================================================\n> There is a lot of ongoing discussion on this one, but skimming the thread it's\n> not really clear (to me) where it leaves the patch in question.\n\nThis set of patches got a fair amount of in-person off-list discussion at the\nPGConf NYC with folks who had concerns about some or all parts of the patch.\nMost of the known contentious design details were resolved, and I've been\nupdating the patches to repost this week. The subscriptions patch and guc\npatch should be close to ready. The CREATEROLE patch may take a bit longer.\n\nConversation Highlights:\n\n - The logical replication subscription patch needs to consider RLS.\n (Jeff Davis)\n\n This is implemented but I still need to update the documentation before\n posting.\n\n - The patch for granting SET VALUE and/or ALTER SYSTEM on guc settings should\n be extended to allow revoking privilege to SET VALUE on userset gucs.\n (Joshua Brindle, Stephen Frost).\n\n This is implemented for core by means of converting PGC_USERSET variables\n to PGC_SUSET, with a corresponding \"GRANT SET VALUE ON .. TO public\" added\n to the new catalog/setting_privileges.sql file. By making it an explicit\n grant, it is revokable without the need for any special logic. This would allow, for\n example, REVOKE SET VALUE ON work_mem FROM public; followed by explicit\n grants to whichever roles you want to allow to set work_mem.\n\n I still need to adjust gucs defined in some contrib modules, update\n documentation, and verify pg_upgrade from older versions before posting. \n\n External projects are not harmed by this implementation. The only issue\n for them is that SET VALUE cannot be revoked from any PGC_USERSET gucs they\n define until they update their module to use PGC_SUSET plus a GRANT\n somewhere in their extension creation or upgrade sql script.\n\n - The CREATEROLE patch can go forward with roles having owners, but other\n spec compliant behavior for GRANT/REVOKE should be implemented.\n (Stephen Frost)\n\n The bug wherein the grantor field in pg_auth_members may point to a dropped\n role should be fixed. Probably this field should simply be dropped. We\n don't use it anywhere, and on upgrade we can't know if existing values are\n correct or bogus. Values that do not refer to any existing role are\n obviously bogus, but if a role Oid was reassigned, a grantor entry may be\n pointing wrongly at the new role, and there is no way to know that.\n\n Stephen indicated that, for all privileges, only the role which GRANTed a\n privilege should be able to REVOKE it. This differs from the current\n implementation and is likely controversial. Robert Haas, at least,\n described that idea as a non-starter. I don't plan to include that in this\n patch, as it doesn't seem necessary to combine the two features even if we\n eventually intend to do both. (And certainly unhelpful to include it if\n the idea is rejected.)\n\n Whether the owner of a role is implicitly a member (or admin) of the owned\n role is still disputed. Stephen Frost would rather the owner has the right\n to grant itself into the role, or grant itself ADMIN on the role, but not\n be a member nor an admin by default. Robert Haas preferred that the owner\n is both a member and an admin by default, making role ownership analogous\n to how superuser works. I lean slightly in Stephen's favor, as any role\n \"foo\" with CREATEROLE could always run:\n\n CREATE ROLE bar [ ROLE foo | ADMIN foo];\n \n and thereby give themselves the privilege they want in the same command\n that creates the role. Going with Robert's proposal, a creator who does\n not want to have privileges in the role would need to run:\n\n CREATE ROLE bar;\n REVOKE [ADMIN OPTION FOR ] bar FROM foo;\n \n which requires two separate commands, and creates a short window inbetween\n where a (possibly undesireable) configuration exists.\n\nJeff Davis and Andrew Dunstan have volunteered as committers.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Mon, 13 Dec 2021 11:20:38 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 3"},{"msg_contents":"On Mon, 2021-12-13 at 11:20 -0800, Mark Dilger wrote:\n> - The logical replication subscription patch needs to consider RLS.\n> (Jeff Davis)\n> \n> This is implemented but I still need to update the documentation\n> before\n> posting.\n\nWe also discussed how much of the insert path we want to include in the\napply worker. The immediate need is for the patch to support RLS, but\nit raised the question: \"why not just use the entire ExecInsert()\npath?\".\n\nThat's not a blocking issue for this patch, but something to consider\nthat will avoid problems if we want to support updatable views or\nforeign tables as a target of a subscription.\n\nRegards,\n\tJeff Davis\n\n\n\n\n","msg_date":"Mon, 13 Dec 2021 22:18:00 -0800","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 3"},{"msg_contents":"\n\n> On Dec 13, 2021, at 10:18 PM, Jeff Davis <pgsql@j-davis.com> wrote:\n> \n>> This is implemented but I still need to update the documentation\n>> before\n>> posting.\n> \n> We also discussed how much of the insert path we want to include in the\n> apply worker. The immediate need is for the patch to support RLS, but\n> it raised the question: \"why not just use the entire ExecInsert()\n> path?\".\n\nI went a different direction with this. The need is to prevent non-superuser subscription owners from *circumventing* RLS. For this patch set, I'm just checking whether RLS should be enforced against the subscription owner, and if so, raising an ERROR, same as for a privilege violation. This should be sufficient in practice, as the table owner, roles with bypassrls privilege, and superusers should still be able to replicate into an RLS enabled table.\n\nThe problem with actually *supporting* RLS is that the current logical replication implementation is a mix of different practical (rather than theoretical) design choices. After working to get RLS working as part of that, I became concerned that there may be subtle differences between how I was making RLS work in logical replication vs. how it works for regular inserts/updates/deletes. Rather than create a mess that would need to be supported going forward, I bailed out and just forbade it.\n\nAs far as completeness and a clean implementation (but not performance) are concerned, using ExecInsert is quite appealing. I think that would require reworking the logical replication protocol to send more than one row at a time, so that the overhead of such a choice is not paid *per row*. That seems quite out of scope for this release cycle, though.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Tue, 14 Dec 2021 07:41:25 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 3"},{"msg_contents":"On Tue, 2021-12-14 at 07:41 -0800, Mark Dilger wrote:\n> I went a different direction with this. The need is to prevent non-\n> superuser subscription owners from *circumventing* RLS. For this\n> patch set, I'm just checking whether RLS should be enforced against\n> the subscription owner, and if so, raising an ERROR, same as for a\n> privilege violation. This should be sufficient in practice, as the\n> table owner, roles with bypassrls privilege, and superusers should\n> still be able to replicate into an RLS enabled table.\n\nI agree with this. Let's consider subscription targets involving RLS a\nseparate feature that is just blocked for now (except for bypassrls\nroles and superusers).\n\n> As far as completeness and a clean implementation (but not\n> performance) are concerned, using ExecInsert is quite appealing. I\n> think that would require reworking the logical replication protocol\n> to send more than one row at a time, so that the overhead of such a\n> choice is not paid *per row*. That seems quite out of scope for this\n> release cycle, though.\n\nAre you sure overhead is a problem? INSERT INTO ... SELECT goes through\nthat path. And the existing logical replication insert path does some\nextra stuff, like opening the table for each row, so it's not clear to\nme that logical replication is much faster.\n\nAlso, the current patch does per-row ACL checks. Did you consider\nmaking the checks a part of logicalrep_rel_open()? That already has a\npath to consider relcache invalidation, so it would also be a good\nplace to check if the table has RLS enabled.\n\nRegards,\n\tJeff Davis\n\n\n\n\n","msg_date":"Tue, 14 Dec 2021 10:58:28 -0800","msg_from":"Jeff Davis <pgsql@j-davis.com>","msg_from_op":false,"msg_subject":"Re: Commitfest 2021-11 Patch Triage - Part 3"}],"string":"[\n {\n \"msg_contents\": \"Welcome to the third installment of the CF 2021-11 patch triage, looking at the\\npatches which have been in four CF's and parts of the ones with three CF's. As\\nper usual I might very well have misunderstood something (or many things), but\\nthere seem to be a fair few ones which should be closeable in this CF here.\\n\\n\\n2962: Fix DROP TABLESPACE on Windows with ProcSignalBarrier?\\n============================================================\\nThis fixes an old bug which can prevent DROP TABLESPACE from working on\\nWindows. There has been a fair bit of review and testing, by among other\\nmyself, and I feel this patch is in good enough shape to be considered for\\nReady for Committer. There is a goalpost movement request in the thread for\\nDROP DATABASE support, but I think we can get this in in two commits to move us\\nforward. What are your thoughts on this one Thomas?\\n\\n\\n2979: Support tab completion for upper character inputs in psql\\n===============================================================\\nThere is broad concensus that this patch fixes a less than ideal behavior in\\npsql, and the patch has seen a lot of review and rework. It currently needs a\\nreview of the latest version, but skimming it and judging by the thread I'd say\\nit's close to Ready for Committer (if not already).\\n\\n\\n2942: increase size of pg_commit_ts buffers\\n===========================================\\nProduction use has validated this patch as fixing a performance regression, and\\nthe patch has support from another committer. There is mention of patch 2627\\n(More scalable multixacts buffers and locking) in the thread as an alternative\\napproach, that buffers potentially should be configurable instead. Since that\\npatch seems like further out from landing, should we go ahead with this to have\\nat least a partial solution for 15? Do have any plans to go ahead with this\\npatch Álvaro or are you awaiting results on the configurable approach?\\n\\n\\n2992: Allow batched insert during cross-partition updates\\n=========================================================\\nThis is a follow-up to 927f453a94106 to support batched inserts of internal\\nINSERT commands which are used in cross-partition updates for local/remote\\npartitions. The patch was after a lot of review initially marked Ready for\\nCommitter but went to WOA after failing to apply. Once updated it went to Need\\nReview which IMO is misleading given it's status and a skim so I'm moving it to\\nRFC again. Georgios, having reviewed it earlier do you any thoughts on the\\nlatest version?\\n\\n\\n2970: Set default transactions to read-only at servers start in pg_upgrade\\n==========================================================================\\nGoalpost movement saw this patch go from read-only transactions to stopping\\nbgworkers from running at all during pg_upgrade. There is agreement on the\\napproach, but discussion on the exact implementation (which has since trailed\\noff). I have on my TODO to bring this over the finish line.\\n\\n\\n2948: Simplify some RI checks to reduce SPI overhead\\n====================================================\\nMarked ready for committer it seems that raised concerns have been addressed.\\nPeter, being listed as reviewer do you have any thoughts on this?\\n\\n\\n2957: Identify missing publications from publisher while create/alter\\nsubscription.\\n=====================================================================\\nPatch is being actively reviewed, has been kept up to date, and was marked RFC\\nback in May. Hopefully we can close this soon then.\\n\\n\\n2968: Minimal logical decoding on standbys (take 6)\\n===================================================\\nThe latest revision of the patch has had feedback from Robert (and Andres)\\nwhich needs to be addressed, and the patch no longer applies (after the recent\\nThisTimeLineID changes perhaps?). I'm moving this patch to WoA for now.\\n\\n\\n2999: Catalog version access\\n============================\\nThe initial patch in this thread was dismissed for not providing enough\\ninformation to solve the stated problem, but there has since been a patch\\nposted which performs ReadControlFile as recommended in the thread. Vik, can\\nyou confirm if this patch solves your original problem?\\n\\n\\n2996: Speed up verifying UTF-8\\n==============================\\nAn IMO interesting approach to speeding up input validation with promising\\nresults posted. The discussion seem to be quite confident with the patch, but\\nthe latest version takes out a portion of the previous patch which was deemed\\noverly complicated. Heikki, are you up for reviewing this version?\\n\\n\\n2932: Partial foreign key updates in referential integrity triggers\\n===================================================================\\nA recently posted update from Peter indicates that this is close to being\\ncommittable, so fingers crossed.\\n\\n\\n3025: postgres_fdw: suppress explicit casts in text:text comparisons\\n====================================================================\\nCurrently undergoing active review and receiving updates with a new version\\nposted today by Tom. Hopefully it can be closed in this CF.\\n\\n\\n2947: Full support for index LP_DEAD hint bits on standby\\n=========================================================\\nThis has potential to be a nice performance boost for a common usecase for\\nscalable production setups. The patch changed course half-way through but has\\nsince been marked Ready for Committer. Any takers on this?\\n\\n\\n3120: Rewriting the test of pg_upgrade as a TAP test - take three\\n=================================================================\\nCurrently this is stuck on a discussion around how to deal with test.sh and the\\neffects on the buildfarm. It seems like the questions raised have been\\nanswered, but we need to settle on how to continue.\\n\\n\\n3230: Column filtering in logical replication\\n=============================================\\nThis fails badly to apply, which is likely because there has been a lot of\\ndiscussion on the grammar making a new patch not really an option until\\nsettled. That discussion ended in late September so hopefully there is a new\\nversion during this CF, else we'll have to close this with RwF awaiting a new\\npatch in a later CF.\\n\\n\\n3223: Delegating superuser tasks to new security roles\\n======================================================\\nThere is a lot of ongoing discussion on this one, but skimming the thread it's\\nnot really clear (to me) where it leaves the patch in question.\\n\\n\\n3138: Support for NSS as a libpq TLS backend\\n============================================\\nI'm clearly fairly biased here but I hope that at 48 revisions in the patch is\\nin good enough shape to be close to RFC. Joshua Brindle has recently picked up\\nreviewing it so hope we can reach somewhere for 15 with this work.\\n\\n\\n3035: Removing unused trailing linepointers from heap pages\\n===========================================================\\nAn alternative to this was committed for 14, which led the author to change the\\npatch into a follow-up to that commit. The merit and value of this second\\npatch is debated, with seemingly good arguments both for and against. There\\nwas an ask for testcases to provide evidence of correctness back in August\\nwhich has gone unanswered, so unless this moves during this CF I think we\\nshould mark it RWF awaiting a new version of the patch.\\n\\n\\n3058: Eliminating \\\"Permission denied\\\" errors on stat() (on Windows)\\n===================================================================\\nThis was committed back in July, and then promptly reverted as there were\\nissues with mingw. There has since been a patch posted addressing the problem\\nwhich needs review. Michael, having looked at the earlier version, do you feel\\nup for taking a look at the reworked one?\\n\\n\\n3172: Consistent use of SSL/TLS in docs\\n=======================================\\nThe proposal to use SSL/TLS instead of just SSL in the docs came about after a\\nsmall discussion in the NSS thread, but it wasn't to anyones liking so the\\nreworked patch just adds notes that SSL and TLS are used interchangeably in the\\ndocs. This version is a pretty small addition that I think we should go ahead\\nwith, but more eyes are needed to validate that.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Nov 2021 23:12:23 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Commitfest 2021-11 Patch Triage - Part 3\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 11, 2021, at 2:12 PM, Daniel Gustafsson <daniel@yesql.se> wrote:\\n> \\n> 3223: Delegating superuser tasks to new security roles\\n> ======================================================\\n> There is a lot of ongoing discussion on this one, but skimming the thread it's\\n> not really clear (to me) where it leaves the patch in question.\\n\\nThis set of patches got a fair amount of in-person off-list discussion at the\\nPGConf NYC with folks who had concerns about some or all parts of the patch.\\nMost of the known contentious design details were resolved, and I've been\\nupdating the patches to repost this week. The subscriptions patch and guc\\npatch should be close to ready. The CREATEROLE patch may take a bit longer.\\n\\nConversation Highlights:\\n\\n - The logical replication subscription patch needs to consider RLS.\\n (Jeff Davis)\\n\\n This is implemented but I still need to update the documentation before\\n posting.\\n\\n - The patch for granting SET VALUE and/or ALTER SYSTEM on guc settings should\\n be extended to allow revoking privilege to SET VALUE on userset gucs.\\n (Joshua Brindle, Stephen Frost).\\n\\n This is implemented for core by means of converting PGC_USERSET variables\\n to PGC_SUSET, with a corresponding \\\"GRANT SET VALUE ON .. TO public\\\" added\\n to the new catalog/setting_privileges.sql file. By making it an explicit\\n grant, it is revokable without the need for any special logic. This would allow, for\\n example, REVOKE SET VALUE ON work_mem FROM public; followed by explicit\\n grants to whichever roles you want to allow to set work_mem.\\n\\n I still need to adjust gucs defined in some contrib modules, update\\n documentation, and verify pg_upgrade from older versions before posting. \\n\\n External projects are not harmed by this implementation. The only issue\\n for them is that SET VALUE cannot be revoked from any PGC_USERSET gucs they\\n define until they update their module to use PGC_SUSET plus a GRANT\\n somewhere in their extension creation or upgrade sql script.\\n\\n - The CREATEROLE patch can go forward with roles having owners, but other\\n spec compliant behavior for GRANT/REVOKE should be implemented.\\n (Stephen Frost)\\n\\n The bug wherein the grantor field in pg_auth_members may point to a dropped\\n role should be fixed. Probably this field should simply be dropped. We\\n don't use it anywhere, and on upgrade we can't know if existing values are\\n correct or bogus. Values that do not refer to any existing role are\\n obviously bogus, but if a role Oid was reassigned, a grantor entry may be\\n pointing wrongly at the new role, and there is no way to know that.\\n\\n Stephen indicated that, for all privileges, only the role which GRANTed a\\n privilege should be able to REVOKE it. This differs from the current\\n implementation and is likely controversial. Robert Haas, at least,\\n described that idea as a non-starter. I don't plan to include that in this\\n patch, as it doesn't seem necessary to combine the two features even if we\\n eventually intend to do both. (And certainly unhelpful to include it if\\n the idea is rejected.)\\n\\n Whether the owner of a role is implicitly a member (or admin) of the owned\\n role is still disputed. Stephen Frost would rather the owner has the right\\n to grant itself into the role, or grant itself ADMIN on the role, but not\\n be a member nor an admin by default. Robert Haas preferred that the owner\\n is both a member and an admin by default, making role ownership analogous\\n to how superuser works. I lean slightly in Stephen's favor, as any role\\n \\\"foo\\\" with CREATEROLE could always run:\\n\\n CREATE ROLE bar [ ROLE foo | ADMIN foo];\\n \\n and thereby give themselves the privilege they want in the same command\\n that creates the role. Going with Robert's proposal, a creator who does\\n not want to have privileges in the role would need to run:\\n\\n CREATE ROLE bar;\\n REVOKE [ADMIN OPTION FOR ] bar FROM foo;\\n \\n which requires two separate commands, and creates a short window inbetween\\n where a (possibly undesireable) configuration exists.\\n\\nJeff Davis and Andrew Dunstan have volunteered as committers.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 13 Dec 2021 11:20:38 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 3\"\n },\n {\n \"msg_contents\": \"On Mon, 2021-12-13 at 11:20 -0800, Mark Dilger wrote:\\n> - The logical replication subscription patch needs to consider RLS.\\n> (Jeff Davis)\\n> \\n> This is implemented but I still need to update the documentation\\n> before\\n> posting.\\n\\nWe also discussed how much of the insert path we want to include in the\\napply worker. The immediate need is for the patch to support RLS, but\\nit raised the question: \\\"why not just use the entire ExecInsert()\\npath?\\\".\\n\\nThat's not a blocking issue for this patch, but something to consider\\nthat will avoid problems if we want to support updatable views or\\nforeign tables as a target of a subscription.\\n\\nRegards,\\n\\tJeff Davis\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 13 Dec 2021 22:18:00 -0800\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 3\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Dec 13, 2021, at 10:18 PM, Jeff Davis <pgsql@j-davis.com> wrote:\\n> \\n>> This is implemented but I still need to update the documentation\\n>> before\\n>> posting.\\n> \\n> We also discussed how much of the insert path we want to include in the\\n> apply worker. The immediate need is for the patch to support RLS, but\\n> it raised the question: \\\"why not just use the entire ExecInsert()\\n> path?\\\".\\n\\nI went a different direction with this. The need is to prevent non-superuser subscription owners from *circumventing* RLS. For this patch set, I'm just checking whether RLS should be enforced against the subscription owner, and if so, raising an ERROR, same as for a privilege violation. This should be sufficient in practice, as the table owner, roles with bypassrls privilege, and superusers should still be able to replicate into an RLS enabled table.\\n\\nThe problem with actually *supporting* RLS is that the current logical replication implementation is a mix of different practical (rather than theoretical) design choices. After working to get RLS working as part of that, I became concerned that there may be subtle differences between how I was making RLS work in logical replication vs. how it works for regular inserts/updates/deletes. Rather than create a mess that would need to be supported going forward, I bailed out and just forbade it.\\n\\nAs far as completeness and a clean implementation (but not performance) are concerned, using ExecInsert is quite appealing. I think that would require reworking the logical replication protocol to send more than one row at a time, so that the overhead of such a choice is not paid *per row*. That seems quite out of scope for this release cycle, though.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 14 Dec 2021 07:41:25 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 3\"\n },\n {\n \"msg_contents\": \"On Tue, 2021-12-14 at 07:41 -0800, Mark Dilger wrote:\\n> I went a different direction with this. The need is to prevent non-\\n> superuser subscription owners from *circumventing* RLS. For this\\n> patch set, I'm just checking whether RLS should be enforced against\\n> the subscription owner, and if so, raising an ERROR, same as for a\\n> privilege violation. This should be sufficient in practice, as the\\n> table owner, roles with bypassrls privilege, and superusers should\\n> still be able to replicate into an RLS enabled table.\\n\\nI agree with this. Let's consider subscription targets involving RLS a\\nseparate feature that is just blocked for now (except for bypassrls\\nroles and superusers).\\n\\n> As far as completeness and a clean implementation (but not\\n> performance) are concerned, using ExecInsert is quite appealing. I\\n> think that would require reworking the logical replication protocol\\n> to send more than one row at a time, so that the overhead of such a\\n> choice is not paid *per row*. That seems quite out of scope for this\\n> release cycle, though.\\n\\nAre you sure overhead is a problem? INSERT INTO ... SELECT goes through\\nthat path. And the existing logical replication insert path does some\\nextra stuff, like opening the table for each row, so it's not clear to\\nme that logical replication is much faster.\\n\\nAlso, the current patch does per-row ACL checks. Did you consider\\nmaking the checks a part of logicalrep_rel_open()? That already has a\\npath to consider relcache invalidation, so it would also be a good\\nplace to check if the table has RLS enabled.\\n\\nRegards,\\n\\tJeff Davis\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 14 Dec 2021 10:58:28 -0800\",\n \"msg_from\": \"Jeff Davis <pgsql@j-davis.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest 2021-11 Patch Triage - Part 3\"\n }\n]","isConversational":false}}},{"rowIdx":53079,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nCurrently, the server shuts down with a FATAL error (added by commit\n[1]) when the recovery target isn't reached. This can cause a server\navailability problem, especially in case of disaster recovery (geo\nrestores) where the primary was down and the user is doing a PITR on a\nserver lying in another region where it had missed to receive few of\nthe last WAL files required to reach the recovery target. In this\ncase, users might want the server to be available rather than a no\nserver. With the commit [1], there's no way to achieve what users\nwanted.\n\nThere can be many reasons for the last few WAL files not reaching the\ntarget server where the user is performing the PITR. The primary may\nhave been down before archiving the last few WAL files to the archive\nlocations, or archive command fails for whatever reasons or network\nlatency from primary to archive location and archive location to the\ntarget server, or recovery command on the target server fails or users\nmay have chosen some wrong/futuristic recovery targets etc. If the\nPITR fails with FATAL error and we may ask them to restart the server,\nbut imagine the wastage of compute resources - if there are a 1 TB of\nWAL files to be replayed and just last 16MB WAL file is missing,\neverything has to be replayed from the beginning.\n\nHere's a proposal(and a patch) to have a GUC so that users can choose\neither to emit a warning and promote or shutdown with FATAL error (as\ndefault) when recovery target isn't reached. In reality, users can\nchoose to shutdown with FATAL error, if strict consistency is the\nnecessity, otherwise they can choose to get promoted, if availability\nis preferred. There is some discussion around this idea in [2].\n\nThoughts?\n\n[1] - commit dc788668bb269b10a108e87d14fefd1b9301b793\nAuthor: Peter Eisentraut <peter@eisentraut.org>\nDate: Wed Jan 29 15:43:32 2020 +0100\n\n Fail if recovery target is not reached\n\n Before, if a recovery target is configured, but the archive ended\n before the target was reached, recovery would end and the server would\n promote without further notice. That was deemed to be pretty wrong.\n With this change, if the recovery target is not reached, it is a fatal\n error.\n\n Based-on-patch-by: Leif Gunnar Erlandsen <leif@lako.no>\n Reviewed-by: Kyotaro Horiguchi <horikyota.ntt@gmail.com>\n Discussion:\nhttps://www.postgresql.org/message-id/flat/993736dd3f1713ec1f63fc3b653839f5@lako.no\n\n[2] - https://www.postgresql.org/message-id/b334d61396e6b0657a63dc38e16d429703fe9b96.camel%40j-davis.com\n\nRegards,\nBharath Rupireddy.","msg_date":"Fri, 12 Nov 2021 15:44:00 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Allow users to choose what happens when recovery target is not\n reached"},{"msg_contents":"On Fri, Nov 12, 2021 at 6:14 PM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n>\n> Currently, the server shuts down with a FATAL error (added by commit\n> [1]) when the recovery target isn't reached. This can cause a server\n> availability problem, especially in case of disaster recovery (geo\n> restores) where the primary was down and the user is doing a PITR on a\n> server lying in another region where it had missed to receive few of\n> the last WAL files required to reach the recovery target. In this\n> case, users might want the server to be available rather than a no\n> server. With the commit [1], there's no way to achieve what users\n> wanted.\n\nif users don't mind if the recovery target is reached or not isn't it\nbetter to simply don't specify a target and let the recovery go as far\nas possible?\n\n\n","msg_date":"Fri, 12 Nov 2021 18:39:31 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Allow users to choose what happens when recovery target is not\n reached"},{"msg_contents":"On Fri, Nov 12, 2021 at 4:09 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n>\n> On Fri, Nov 12, 2021 at 6:14 PM Bharath Rupireddy\n> <bharath.rupireddyforpostgres@gmail.com> wrote:\n> >\n> > Currently, the server shuts down with a FATAL error (added by commit\n> > [1]) when the recovery target isn't reached. This can cause a server\n> > availability problem, especially in case of disaster recovery (geo\n> > restores) where the primary was down and the user is doing a PITR on a\n> > server lying in another region where it had missed to receive few of\n> > the last WAL files required to reach the recovery target. In this\n> > case, users might want the server to be available rather than a no\n> > server. With the commit [1], there's no way to achieve what users\n> > wanted.\n>\n> if users don't mind if the recovery target is reached or not isn't it\n> better to simply don't specify a target and let the recovery go as far\n> as possible?\n\nUsers will always be optimistic and set a recovery target and try to\nreach it, but somehow the few of the WAL files haven't arrived (for\nwhatever the reasons) the PITR target server, imagine if their primary\nisn't available too, then with the proposal I made, they can choose to\nhave at least an available target server rather than a FATALly failed\none.\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Sat, 13 Nov 2021 08:29:52 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Allow users to choose what happens when recovery target is not\n reached"},{"msg_contents":"On Sat, Nov 13, 2021 at 11:00 AM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n>\n> Users will always be optimistic and set a recovery target and try to\n> reach it, but somehow the few of the WAL files haven't arrived (for\n> whatever the reasons) the PITR target server, imagine if their primary\n> isn't available too, then with the proposal I made, they can choose to\n> have at least an available target server rather than a FATALly failed\n> one.\n\nIf your primary server isn't available, why would you want a recovery\ntarget in the first place? I just don't understand in which case\nsomeone would want to setup a recovery target and wouldn't care if the\nrecovery wasn't reached, especially if it can be off by GB / days of\ndata.\n\nIt seems like it could have the opposite effect of what you want most\nof the time. What if for some reason the restore_command is flawed,\nand you end up starting your server because it couldn't restore WAL\nthat are actually available? You would have to restart from scratch\nand waste more time than if you didn't use this.\n\nIt look like what you actually want is some kind of a target window,\nbut the window you currently propose is a hardcoded (consistency,\ngiven target], and it seems too dangerous to be useful.\n\n\n","msg_date":"Sat, 13 Nov 2021 12:15:11 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Allow users to choose what happens when recovery target is not\n reached"},{"msg_contents":"On Sat, Nov 13, 2021 at 9:45 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\n>\n> On Sat, Nov 13, 2021 at 11:00 AM Bharath Rupireddy\n> <bharath.rupireddyforpostgres@gmail.com> wrote:\n> >\n> > Users will always be optimistic and set a recovery target and try to\n> > reach it, but somehow the few of the WAL files haven't arrived (for\n> > whatever the reasons) the PITR target server, imagine if their primary\n> > isn't available too, then with the proposal I made, they can choose to\n> > have at least an available target server rather than a FATALly failed\n> > one.\n>\n> If your primary server isn't available, why would you want a recovery\n> target in the first place? I just don't understand in which case\n> someone would want to setup a recovery target and wouldn't care if the\n> recovery wasn't reached, especially if it can be off by GB / days of\n> data.\n>\n> It seems like it could have the opposite effect of what you want most\n> of the time. What if for some reason the restore_command is flawed,\n> and you end up starting your server because it couldn't restore WAL\n> that are actually available? You would have to restart from scratch\n> and waste more time than if you didn't use this.\n\nFirstly, the proposed patch adds no new behaviour as such, it just\ngives the ability that is existing today on v12 and below (prior to\ncommit dc78866 which went into v13 and later).\n\nI think performing PITR is the user's wish - whether the primary is\navailable or not, it is completely the user's choice. The user might\nstart the PITR, when the primary is available, thinking that it sends\nall the WAL files required for achieving recovery target. But imagine\na disaster happens and the primary server crashes, say the recovery\nhas replayed a huge bunch of WAL records (a TB may be), and the\nprimary failed without sending the last one or few WAL files, should\nthe PITR target server be failing this case after replaying a huge\nbunch of WAL records? The user might want the target server to be\navailable instead of FATALly shutting down. This is the exact problem\nthe proposed patch is trying to solve.\n\nWith the GUC proposed, the user can choose what to do in these\nscenarios. The user will be fully aware what she needs when she choose\nto set the new GUC recovery_end_before_target_action to 'promote'\ninstead of default 'shutdown'.\n\n> It look like what you actually want is some kind of a target window,\n> but the window you currently propose is a hardcoded (consistency,\n> given target], and it seems too dangerous to be useful.\n\nAs I said earlier, the behaviour is not too dangerous as it is not\nsomething new that the patch is proposing, it exists today in v12 and\nbelow. In fact, it gives a way out of a \"dangerous situation\" if the\nuser ever gets stuck in it without wasting recovery cycles and compute\nresources, by quickly getting the database to be available(of course,\nthe responsibility lies with the user to deal with the missing WAL\nfiles).\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Sat, 13 Nov 2021 18:45:49 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Allow users to choose what happens when recovery target is not\n reached"},{"msg_contents":"On Sat, Nov 13, 2021, at 10:15 AM, Bharath Rupireddy wrote:\n> Firstly, the proposed patch adds no new behaviour as such, it just\n> gives the ability that is existing today on v12 and below (prior to\n> commit dc78866 which went into v13 and later).\nIt reintroduces an awkward behavior [1].\n\n> I think performing PITR is the user's wish - whether the primary is\n> available or not, it is completely the user's choice. The user might\n> start the PITR, when the primary is available, thinking that it sends\n> all the WAL files required for achieving recovery target. But imagine\n> a disaster happens and the primary server crashes, say the recovery\n> has replayed a huge bunch of WAL records (a TB may be), and the\n> primary failed without sending the last one or few WAL files, should\n> the PITR target server be failing this case after replaying a huge\n> bunch of WAL records? The user might want the target server to be\n> available instead of FATALly shutting down. This is the exact problem\n> the proposed patch is trying to solve.\nAre you archiving on the primary server? You are risking your customer's\nbusiness suggesting such setup. You should store the WAL files on your backup\nserver.\n\nIt seems your setup has a flaw. You set a recovery target but accept a scenario\nthat is not what you initially asked for. If it is a real PITR, it is awkward\nlike Peter [1] said. You could validate your recovery settings checking the\ntimestamp of the last WAL file as a rough approximation of the maximum recovery\ntarget time. The other option is to run pg_waldump to obtain the last commit\ntimestamp.\n\nIf you care about your customer's data, you won't use such option. Otherwise, I\nrepeat the Julien's question [2]: isn't it better to simply don't specify a target\nand let the recovery go as far as possible?\n\n> As I said earlier, the behaviour is not too dangerous as it is not\n> something new that the patch is proposing, it exists today in v12 and\n> below. In fact, it gives a way out of a \"dangerous situation\" if the\n> user ever gets stuck in it without wasting recovery cycles and compute\n> resources, by quickly getting the database to be available(of course,\n> the responsibility lies with the user to deal with the missing WAL\n> files).\nYour proposal seems that the user is shooting in the dark. If a FATAL message\nwas got it means the user missed the target. Even after that the user accepts\nthe situation, remove the target parameters and start the server again. I think\npromote or even pause might lead to incorrect expectations (if the user doesn't\ncarefully inspect the log messages).\n\nA disadvantage of this proposal is that if you have it set to 'promote', start\nthe recovery and the server gets promoted before reaching the target. While\ninspecting your server configuration, you realized that you are pointing to the\nincorrect archive or the WAL files were not available in time (due to timing\nissues). You have no option but start from scratch.\n\n[1] https://postgr.es/m/234a0c50-1160-86c2-4e4b-35e9684f1799%402ndquadrant.com\n[2] https://postgr.es/m/CAOBaU_ZDkyoQvEsYT0-p1Hb0m_nGtQJ4tTGm2-Ay6v%3DTCjmsWg%40mail.gmail.com\n\n\n--\nEuler Taveira\nEDB https://www.enterprisedb.com/\n\nOn Sat, Nov 13, 2021, at 10:15 AM, Bharath Rupireddy wrote:Firstly, the proposed patch adds no new behaviour as such, it justgives the ability that is existing today on v12 and below (prior tocommit dc78866 which went into v13 and later).It reintroduces an awkward behavior [1].I think performing PITR is the user's wish - whether the primary isavailable or not, it is completely the user's choice. The user mightstart the PITR, when the primary is available, thinking that it sendsall the WAL files required for achieving recovery target. But imaginea disaster happens and the primary server crashes, say the recoveryhas replayed a huge bunch of WAL records (a TB may be), and theprimary failed without sending the last one or few WAL files, shouldthe PITR target server be failing this case after replaying a hugebunch of WAL records? The user might want the target server to beavailable instead of FATALly shutting down. This is the exact problemthe proposed patch is trying to solve.Are you archiving on the primary server? You are risking your customer'sbusiness suggesting such setup. You should store the WAL files on your backupserver.It seems your setup has a flaw. You set a recovery target but accept a scenariothat is not what you initially asked for. If it is a real PITR, it is awkwardlike Peter [1] said. You could validate your recovery settings checking thetimestamp of the last WAL file as a rough approximation of the maximum recoverytarget time. The other option is to run pg_waldump to obtain the last committimestamp.If you care about your customer's data, you won't use such option. Otherwise, Irepeat the Julien's question [2]: isn't it better to simply don't specify a targetand let the recovery go as far as possible?As I said earlier, the behaviour is not too dangerous as it is notsomething new that the patch is proposing, it exists today in v12 andbelow. In fact, it gives a way out of a \"dangerous situation\" if theuser ever gets stuck in it without wasting recovery cycles and computeresources, by quickly getting the database to be available(of course,the responsibility lies with the user to deal with the missing WALfiles).Your proposal seems that the user is shooting in the dark. If a FATAL messagewas got it means the user missed the target. Even after that the user acceptsthe situation, remove the target parameters and start the server again. I thinkpromote or even pause might lead to incorrect expectations (if the user doesn'tcarefully inspect the log messages).A disadvantage of this proposal is that if you have it set to 'promote', startthe recovery and the server gets promoted before reaching the target. Whileinspecting your server configuration, you realized that you are pointing to theincorrect archive or the WAL files were not available in time (due to timingissues). You have no option but start from scratch.[1] https://postgr.es/m/234a0c50-1160-86c2-4e4b-35e9684f1799%402ndquadrant.com[2] https://postgr.es/m/CAOBaU_ZDkyoQvEsYT0-p1Hb0m_nGtQJ4tTGm2-Ay6v%3DTCjmsWg%40mail.gmail.com--Euler TaveiraEDB https://www.enterprisedb.com/","msg_date":"Sat, 13 Nov 2021 12:45:31 -0300","msg_from":"\"Euler Taveira\" <euler@eulerto.com>","msg_from_op":false,"msg_subject":"Re: Allow users to choose what happens when recovery target is not\n reached"},{"msg_contents":"On Sat, Nov 13, 2021 at 6:45 PM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n>\n> Firstly, the proposed patch adds no new behaviour as such, it just\n> gives the ability that is existing today on v12 and below (prior to\n> commit dc78866 which went into v13 and later).\n>\n> I think performing PITR is the user's wish - whether the primary is\n> available or not, it is completely the user's choice. The user might\n> start the PITR, when the primary is available, thinking that it sends\n> all the WAL files required for achieving recovery target. But imagine\n> a disaster happens and the primary server crashes, say the recovery\n> has replayed a huge bunch of WAL records (a TB may be), and the\n> primary failed without sending the last one or few WAL files, should\n> the PITR target server be failing this case after replaying a huge\n> bunch of WAL records? The user might want the target server to be\n> available instead of FATALly shutting down. This is the exact problem\n> the proposed patch is trying to solve.\n>\n> With the GUC proposed, the user can choose what to do in these\n> scenarios. The user will be fully aware what she needs when she choose\n> to set the new GUC recovery_end_before_target_action to 'promote'\n> instead of default 'shutdown'.\n\nHi Hackers, with a recent bug report [1] in pgsql-bugs, I'm checking\nif the proposal here in this thread interests anyone.\n\n[1] https://www.postgresql.org/message-id/CALj2ACVnCsNyJTG_75%2B5Us2evfsLYz5CEhmCV4qH%3DVPa0kWOvw%40mail.gmail.com\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Sat, 29 Jan 2022 20:35:51 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Allow users to choose what happens when recovery target is not\n reached"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nCurrently, the server shuts down with a FATAL error (added by commit\\n[1]) when the recovery target isn't reached. This can cause a server\\navailability problem, especially in case of disaster recovery (geo\\nrestores) where the primary was down and the user is doing a PITR on a\\nserver lying in another region where it had missed to receive few of\\nthe last WAL files required to reach the recovery target. In this\\ncase, users might want the server to be available rather than a no\\nserver. With the commit [1], there's no way to achieve what users\\nwanted.\\n\\nThere can be many reasons for the last few WAL files not reaching the\\ntarget server where the user is performing the PITR. The primary may\\nhave been down before archiving the last few WAL files to the archive\\nlocations, or archive command fails for whatever reasons or network\\nlatency from primary to archive location and archive location to the\\ntarget server, or recovery command on the target server fails or users\\nmay have chosen some wrong/futuristic recovery targets etc. If the\\nPITR fails with FATAL error and we may ask them to restart the server,\\nbut imagine the wastage of compute resources - if there are a 1 TB of\\nWAL files to be replayed and just last 16MB WAL file is missing,\\neverything has to be replayed from the beginning.\\n\\nHere's a proposal(and a patch) to have a GUC so that users can choose\\neither to emit a warning and promote or shutdown with FATAL error (as\\ndefault) when recovery target isn't reached. In reality, users can\\nchoose to shutdown with FATAL error, if strict consistency is the\\nnecessity, otherwise they can choose to get promoted, if availability\\nis preferred. There is some discussion around this idea in [2].\\n\\nThoughts?\\n\\n[1] - commit dc788668bb269b10a108e87d14fefd1b9301b793\\nAuthor: Peter Eisentraut <peter@eisentraut.org>\\nDate: Wed Jan 29 15:43:32 2020 +0100\\n\\n Fail if recovery target is not reached\\n\\n Before, if a recovery target is configured, but the archive ended\\n before the target was reached, recovery would end and the server would\\n promote without further notice. That was deemed to be pretty wrong.\\n With this change, if the recovery target is not reached, it is a fatal\\n error.\\n\\n Based-on-patch-by: Leif Gunnar Erlandsen <leif@lako.no>\\n Reviewed-by: Kyotaro Horiguchi <horikyota.ntt@gmail.com>\\n Discussion:\\nhttps://www.postgresql.org/message-id/flat/993736dd3f1713ec1f63fc3b653839f5@lako.no\\n\\n[2] - https://www.postgresql.org/message-id/b334d61396e6b0657a63dc38e16d429703fe9b96.camel%40j-davis.com\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Fri, 12 Nov 2021 15:44:00 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Allow users to choose what happens when recovery target is not\\n reached\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 6:14 PM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n>\\n> Currently, the server shuts down with a FATAL error (added by commit\\n> [1]) when the recovery target isn't reached. This can cause a server\\n> availability problem, especially in case of disaster recovery (geo\\n> restores) where the primary was down and the user is doing a PITR on a\\n> server lying in another region where it had missed to receive few of\\n> the last WAL files required to reach the recovery target. In this\\n> case, users might want the server to be available rather than a no\\n> server. With the commit [1], there's no way to achieve what users\\n> wanted.\\n\\nif users don't mind if the recovery target is reached or not isn't it\\nbetter to simply don't specify a target and let the recovery go as far\\nas possible?\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 18:39:31 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow users to choose what happens when recovery target is not\\n reached\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 4:09 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n>\\n> On Fri, Nov 12, 2021 at 6:14 PM Bharath Rupireddy\\n> <bharath.rupireddyforpostgres@gmail.com> wrote:\\n> >\\n> > Currently, the server shuts down with a FATAL error (added by commit\\n> > [1]) when the recovery target isn't reached. This can cause a server\\n> > availability problem, especially in case of disaster recovery (geo\\n> > restores) where the primary was down and the user is doing a PITR on a\\n> > server lying in another region where it had missed to receive few of\\n> > the last WAL files required to reach the recovery target. In this\\n> > case, users might want the server to be available rather than a no\\n> > server. With the commit [1], there's no way to achieve what users\\n> > wanted.\\n>\\n> if users don't mind if the recovery target is reached or not isn't it\\n> better to simply don't specify a target and let the recovery go as far\\n> as possible?\\n\\nUsers will always be optimistic and set a recovery target and try to\\nreach it, but somehow the few of the WAL files haven't arrived (for\\nwhatever the reasons) the PITR target server, imagine if their primary\\nisn't available too, then with the proposal I made, they can choose to\\nhave at least an available target server rather than a FATALly failed\\none.\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 08:29:52 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow users to choose what happens when recovery target is not\\n reached\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 13, 2021 at 11:00 AM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n>\\n> Users will always be optimistic and set a recovery target and try to\\n> reach it, but somehow the few of the WAL files haven't arrived (for\\n> whatever the reasons) the PITR target server, imagine if their primary\\n> isn't available too, then with the proposal I made, they can choose to\\n> have at least an available target server rather than a FATALly failed\\n> one.\\n\\nIf your primary server isn't available, why would you want a recovery\\ntarget in the first place? I just don't understand in which case\\nsomeone would want to setup a recovery target and wouldn't care if the\\nrecovery wasn't reached, especially if it can be off by GB / days of\\ndata.\\n\\nIt seems like it could have the opposite effect of what you want most\\nof the time. What if for some reason the restore_command is flawed,\\nand you end up starting your server because it couldn't restore WAL\\nthat are actually available? You would have to restart from scratch\\nand waste more time than if you didn't use this.\\n\\nIt look like what you actually want is some kind of a target window,\\nbut the window you currently propose is a hardcoded (consistency,\\ngiven target], and it seems too dangerous to be useful.\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 12:15:11 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow users to choose what happens when recovery target is not\\n reached\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 13, 2021 at 9:45 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n>\\n> On Sat, Nov 13, 2021 at 11:00 AM Bharath Rupireddy\\n> <bharath.rupireddyforpostgres@gmail.com> wrote:\\n> >\\n> > Users will always be optimistic and set a recovery target and try to\\n> > reach it, but somehow the few of the WAL files haven't arrived (for\\n> > whatever the reasons) the PITR target server, imagine if their primary\\n> > isn't available too, then with the proposal I made, they can choose to\\n> > have at least an available target server rather than a FATALly failed\\n> > one.\\n>\\n> If your primary server isn't available, why would you want a recovery\\n> target in the first place? I just don't understand in which case\\n> someone would want to setup a recovery target and wouldn't care if the\\n> recovery wasn't reached, especially if it can be off by GB / days of\\n> data.\\n>\\n> It seems like it could have the opposite effect of what you want most\\n> of the time. What if for some reason the restore_command is flawed,\\n> and you end up starting your server because it couldn't restore WAL\\n> that are actually available? You would have to restart from scratch\\n> and waste more time than if you didn't use this.\\n\\nFirstly, the proposed patch adds no new behaviour as such, it just\\ngives the ability that is existing today on v12 and below (prior to\\ncommit dc78866 which went into v13 and later).\\n\\nI think performing PITR is the user's wish - whether the primary is\\navailable or not, it is completely the user's choice. The user might\\nstart the PITR, when the primary is available, thinking that it sends\\nall the WAL files required for achieving recovery target. But imagine\\na disaster happens and the primary server crashes, say the recovery\\nhas replayed a huge bunch of WAL records (a TB may be), and the\\nprimary failed without sending the last one or few WAL files, should\\nthe PITR target server be failing this case after replaying a huge\\nbunch of WAL records? The user might want the target server to be\\navailable instead of FATALly shutting down. This is the exact problem\\nthe proposed patch is trying to solve.\\n\\nWith the GUC proposed, the user can choose what to do in these\\nscenarios. The user will be fully aware what she needs when she choose\\nto set the new GUC recovery_end_before_target_action to 'promote'\\ninstead of default 'shutdown'.\\n\\n> It look like what you actually want is some kind of a target window,\\n> but the window you currently propose is a hardcoded (consistency,\\n> given target], and it seems too dangerous to be useful.\\n\\nAs I said earlier, the behaviour is not too dangerous as it is not\\nsomething new that the patch is proposing, it exists today in v12 and\\nbelow. In fact, it gives a way out of a \\\"dangerous situation\\\" if the\\nuser ever gets stuck in it without wasting recovery cycles and compute\\nresources, by quickly getting the database to be available(of course,\\nthe responsibility lies with the user to deal with the missing WAL\\nfiles).\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 18:45:49 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow users to choose what happens when recovery target is not\\n reached\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 13, 2021, at 10:15 AM, Bharath Rupireddy wrote:\\n> Firstly, the proposed patch adds no new behaviour as such, it just\\n> gives the ability that is existing today on v12 and below (prior to\\n> commit dc78866 which went into v13 and later).\\nIt reintroduces an awkward behavior [1].\\n\\n> I think performing PITR is the user's wish - whether the primary is\\n> available or not, it is completely the user's choice. The user might\\n> start the PITR, when the primary is available, thinking that it sends\\n> all the WAL files required for achieving recovery target. But imagine\\n> a disaster happens and the primary server crashes, say the recovery\\n> has replayed a huge bunch of WAL records (a TB may be), and the\\n> primary failed without sending the last one or few WAL files, should\\n> the PITR target server be failing this case after replaying a huge\\n> bunch of WAL records? The user might want the target server to be\\n> available instead of FATALly shutting down. This is the exact problem\\n> the proposed patch is trying to solve.\\nAre you archiving on the primary server? You are risking your customer's\\nbusiness suggesting such setup. You should store the WAL files on your backup\\nserver.\\n\\nIt seems your setup has a flaw. You set a recovery target but accept a scenario\\nthat is not what you initially asked for. If it is a real PITR, it is awkward\\nlike Peter [1] said. You could validate your recovery settings checking the\\ntimestamp of the last WAL file as a rough approximation of the maximum recovery\\ntarget time. The other option is to run pg_waldump to obtain the last commit\\ntimestamp.\\n\\nIf you care about your customer's data, you won't use such option. Otherwise, I\\nrepeat the Julien's question [2]: isn't it better to simply don't specify a target\\nand let the recovery go as far as possible?\\n\\n> As I said earlier, the behaviour is not too dangerous as it is not\\n> something new that the patch is proposing, it exists today in v12 and\\n> below. In fact, it gives a way out of a \\\"dangerous situation\\\" if the\\n> user ever gets stuck in it without wasting recovery cycles and compute\\n> resources, by quickly getting the database to be available(of course,\\n> the responsibility lies with the user to deal with the missing WAL\\n> files).\\nYour proposal seems that the user is shooting in the dark. If a FATAL message\\nwas got it means the user missed the target. Even after that the user accepts\\nthe situation, remove the target parameters and start the server again. I think\\npromote or even pause might lead to incorrect expectations (if the user doesn't\\ncarefully inspect the log messages).\\n\\nA disadvantage of this proposal is that if you have it set to 'promote', start\\nthe recovery and the server gets promoted before reaching the target. While\\ninspecting your server configuration, you realized that you are pointing to the\\nincorrect archive or the WAL files were not available in time (due to timing\\nissues). You have no option but start from scratch.\\n\\n[1] https://postgr.es/m/234a0c50-1160-86c2-4e4b-35e9684f1799%402ndquadrant.com\\n[2] https://postgr.es/m/CAOBaU_ZDkyoQvEsYT0-p1Hb0m_nGtQJ4tTGm2-Ay6v%3DTCjmsWg%40mail.gmail.com\\n\\n\\n--\\nEuler Taveira\\nEDB https://www.enterprisedb.com/\\n\\nOn Sat, Nov 13, 2021, at 10:15 AM, Bharath Rupireddy wrote:Firstly, the proposed patch adds no new behaviour as such, it justgives the ability that is existing today on v12 and below (prior tocommit dc78866 which went into v13 and later).It reintroduces an awkward behavior [1].I think performing PITR is the user's wish - whether the primary isavailable or not, it is completely the user's choice. The user mightstart the PITR, when the primary is available, thinking that it sendsall the WAL files required for achieving recovery target. But imaginea disaster happens and the primary server crashes, say the recoveryhas replayed a huge bunch of WAL records (a TB may be), and theprimary failed without sending the last one or few WAL files, shouldthe PITR target server be failing this case after replaying a hugebunch of WAL records? The user might want the target server to beavailable instead of FATALly shutting down. This is the exact problemthe proposed patch is trying to solve.Are you archiving on the primary server? You are risking your customer'sbusiness suggesting such setup. You should store the WAL files on your backupserver.It seems your setup has a flaw. You set a recovery target but accept a scenariothat is not what you initially asked for. If it is a real PITR, it is awkwardlike Peter [1] said. You could validate your recovery settings checking thetimestamp of the last WAL file as a rough approximation of the maximum recoverytarget time. The other option is to run pg_waldump to obtain the last committimestamp.If you care about your customer's data, you won't use such option. Otherwise, Irepeat the Julien's question [2]: isn't it better to simply don't specify a targetand let the recovery go as far as possible?As I said earlier, the behaviour is not too dangerous as it is notsomething new that the patch is proposing, it exists today in v12 andbelow. In fact, it gives a way out of a \\\"dangerous situation\\\" if theuser ever gets stuck in it without wasting recovery cycles and computeresources, by quickly getting the database to be available(of course,the responsibility lies with the user to deal with the missing WALfiles).Your proposal seems that the user is shooting in the dark. If a FATAL messagewas got it means the user missed the target. Even after that the user acceptsthe situation, remove the target parameters and start the server again. I thinkpromote or even pause might lead to incorrect expectations (if the user doesn'tcarefully inspect the log messages).A disadvantage of this proposal is that if you have it set to 'promote', startthe recovery and the server gets promoted before reaching the target. Whileinspecting your server configuration, you realized that you are pointing to theincorrect archive or the WAL files were not available in time (due to timingissues). You have no option but start from scratch.[1] https://postgr.es/m/234a0c50-1160-86c2-4e4b-35e9684f1799%402ndquadrant.com[2] https://postgr.es/m/CAOBaU_ZDkyoQvEsYT0-p1Hb0m_nGtQJ4tTGm2-Ay6v%3DTCjmsWg%40mail.gmail.com--Euler TaveiraEDB https://www.enterprisedb.com/\",\n \"msg_date\": \"Sat, 13 Nov 2021 12:45:31 -0300\",\n \"msg_from\": \"\\\"Euler Taveira\\\" <euler@eulerto.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow users to choose what happens when recovery target is not\\n reached\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 13, 2021 at 6:45 PM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n>\\n> Firstly, the proposed patch adds no new behaviour as such, it just\\n> gives the ability that is existing today on v12 and below (prior to\\n> commit dc78866 which went into v13 and later).\\n>\\n> I think performing PITR is the user's wish - whether the primary is\\n> available or not, it is completely the user's choice. The user might\\n> start the PITR, when the primary is available, thinking that it sends\\n> all the WAL files required for achieving recovery target. But imagine\\n> a disaster happens and the primary server crashes, say the recovery\\n> has replayed a huge bunch of WAL records (a TB may be), and the\\n> primary failed without sending the last one or few WAL files, should\\n> the PITR target server be failing this case after replaying a huge\\n> bunch of WAL records? The user might want the target server to be\\n> available instead of FATALly shutting down. This is the exact problem\\n> the proposed patch is trying to solve.\\n>\\n> With the GUC proposed, the user can choose what to do in these\\n> scenarios. The user will be fully aware what she needs when she choose\\n> to set the new GUC recovery_end_before_target_action to 'promote'\\n> instead of default 'shutdown'.\\n\\nHi Hackers, with a recent bug report [1] in pgsql-bugs, I'm checking\\nif the proposal here in this thread interests anyone.\\n\\n[1] https://www.postgresql.org/message-id/CALj2ACVnCsNyJTG_75%2B5Us2evfsLYz5CEhmCV4qH%3DVPa0kWOvw%40mail.gmail.com\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Sat, 29 Jan 2022 20:35:51 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow users to choose what happens when recovery target is not\\n reached\"\n }\n]","isConversational":false}}},{"rowIdx":53080,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"While reading through and working with the hundreds of patches in the CF app a\nsmall feature/process request struck me: it would be really helpful if the\npatch had a brief abstract outlining what it aims to add or fix (or summary,\ndescription or something else; not sure what to call it). Basically a\ntwo-sentence or so version of the email posting the patch to -hackers.\n\nThe \"title\" field isn't always explanatory enough, and diving into a long\nthread to piece together clues can be quite daunting. I'm especially thinking\nabout new contributors who would like to contribute but who have a hard time\nfinding something in the CF app (which is a complaint I've heard from multiple\nsuch contributors).\n\nIt should be a free-form textfield, and only apply to new submissions to avoid\na need to backfill. The field should, IMHO, be mandatory. The implementation\ncan be left to discussion on -www iff this list agrees that it would be a good\nthing.\n\nThoughts?\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Fri, 12 Nov 2021 13:51:28 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Patch abstracts in the Commitfest app"},{"msg_contents":"On Fri, Nov 12, 2021 at 8:51 PM Daniel Gustafsson <daniel@yesql.se> wrote:\n>\n> While reading through and working with the hundreds of patches in the CF app a\n> small feature/process request struck me: it would be really helpful if the\n> patch had a brief abstract outlining what it aims to add or fix (or summary,\n> description or something else; not sure what to call it). Basically a\n> two-sentence or so version of the email posting the patch to -hackers.\n>\n> The \"title\" field isn't always explanatory enough, and diving into a long\n> thread to piece together clues can be quite daunting. I'm especially thinking\n> about new contributors who would like to contribute but who have a hard time\n> finding something in the CF app (which is a complaint I've heard from multiple\n> such contributors).\n>\n> It should be a free-form textfield, and only apply to new submissions to avoid\n> a need to backfill. The field should, IMHO, be mandatory. The implementation\n> can be left to discussion on -www iff this list agrees that it would be a good\n> thing.\n\n+1 for the idea. Not sure if that should be mandatory though, as some\nsimple entries should be self explanatory. Really this is mainly\nhelpful for long and lengthy threads, when people start to ask a\nsummary of the situation, so it should be editable and maybe\nautomatically generates an email on the thread when that happens.\n\n\n","msg_date":"Fri, 12 Nov 2021 20:59:57 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Patch abstracts in the Commitfest app"},{"msg_contents":"On Fri, Nov 12, 2021 at 01:51:28PM +0100, Daniel Gustafsson wrote:\n> While reading through and working with the hundreds of patches in the CF app a\n> small feature/process request struck me: it would be really helpful if the\n> patch had a brief abstract outlining what it aims to add or fix (or summary,\n> description or something else; not sure what to call it). Basically a\n> two-sentence or so version of the email posting the patch to -hackers.\n\nThis seems fine ; that purpose is partially served (and duplicated) by the\npatch commit messages (if used).\n\nActually, it's already possible to add an \"annotation\" to a given message-id.\nI've just added this here: https://commitfest.postgresql.org/35/2800/\n\nI think this that feature is rarely used, but I can see the value in using it\nmore, for long threads.\n\nIf you really wanted to make an \"abstract\" not associated with any message-id,\nI think it would be implemented simply by removing the restriction that an\nmessage-id must be specified and must match an existing message in the thead.\n\n-- \nJustin\n\n\n","msg_date":"Fri, 12 Nov 2021 08:24:01 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Patch abstracts in the Commitfest app"},{"msg_contents":"> On 12 Nov 2021, at 15:24, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> \n> On Fri, Nov 12, 2021 at 01:51:28PM +0100, Daniel Gustafsson wrote:\n>> While reading through and working with the hundreds of patches in the CF app a\n>> small feature/process request struck me: it would be really helpful if the\n>> patch had a brief abstract outlining what it aims to add or fix (or summary,\n>> description or something else; not sure what to call it). Basically a\n>> two-sentence or so version of the email posting the patch to -hackers.\n> \n> This seems fine ; that purpose is partially served (and duplicated) by the\n> patch commit messages (if used).\n\nThat's the problem, many patches are in diff format and don't have commit\nmessages at all. There are also many entries with patchsets containing\nmultiple patches and thus commitmessages but the app will only link to one.\n\n> Actually, it's already possible to add an \"annotation\" to a given message-id.\n> I've just added this here: https://commitfest.postgresql.org/35/2800/\n> \n> I think this that feature is rarely used, but I can see the value in using it\n> more, for long threads.\n\nI consider this feature very valuable, but I don't see it as a replacement for\na text field Abstract in the patch entry.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Fri, 12 Nov 2021 15:36:43 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Re: Patch abstracts in the Commitfest app"},{"msg_contents":"On 11/12/21, 4:53 AM, \"Daniel Gustafsson\" <daniel@yesql.se> wrote:\r\n> While reading through and working with the hundreds of patches in the CF app a\r\n> small feature/process request struck me: it would be really helpful if the\r\n> patch had a brief abstract outlining what it aims to add or fix (or summary,\r\n> description or something else; not sure what to call it). Basically a\r\n> two-sentence or so version of the email posting the patch to -hackers.\r\n\r\n+1\r\n\r\nNathan\r\n\r\n","msg_date":"Fri, 12 Nov 2021 14:50:11 +0000","msg_from":"\"Bossart, Nathan\" <bossartn@amazon.com>","msg_from_op":false,"msg_subject":"Re: Patch abstracts in the Commitfest app"},{"msg_contents":"> On Fri, Nov 12, 2021 at 03:36:43PM +0100, Daniel Gustafsson wrote:\n> > On 12 Nov 2021, at 15:24, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> >\n> > On Fri, Nov 12, 2021 at 01:51:28PM +0100, Daniel Gustafsson wrote:\n> >> While reading through and working with the hundreds of patches in the CF app a\n> >> small feature/process request struck me: it would be really helpful if the\n> >> patch had a brief abstract outlining what it aims to add or fix (or summary,\n> >> description or something else; not sure what to call it). Basically a\n> >> two-sentence or so version of the email posting the patch to -hackers.\n> >\n> > This seems fine ; that purpose is partially served (and duplicated) by the\n> > patch commit messages (if used).\n>\n> That's the problem, many patches are in diff format and don't have commit\n> messages at all. There are also many entries with patchsets containing\n> multiple patches and thus commitmessages but the app will only link to one.\n\nProbably encouraging to use cover letters, generated for such patch\nseries and linked by the CF app, would be a good idea?\n\n\n","msg_date":"Fri, 12 Nov 2021 15:51:33 +0100","msg_from":"Dmitry Dolgov <9erthalion6@gmail.com>","msg_from_op":false,"msg_subject":"Re: Patch abstracts in the Commitfest app"},{"msg_contents":"Julien Rouhaud <rjuju123@gmail.com> writes:\n> On Fri, Nov 12, 2021 at 8:51 PM Daniel Gustafsson <daniel@yesql.se> wrote:\n>> It should be a free-form textfield, and only apply to new submissions to avoid\n>> a need to backfill. The field should, IMHO, be mandatory. The implementation\n>> can be left to discussion on -www iff this list agrees that it would be a good\n>> thing.\n\n> +1 for the idea. Not sure if that should be mandatory though, as some\n> simple entries should be self explanatory.\n\nYeah, there are plenty of entries for which the title seems adequate.\nNo objection to making room for an additional optional description,\nthough.\n\n> Really this is mainly\n> helpful for long and lengthy threads, when people start to ask a\n> summary of the situation, so it should be editable and maybe\n> automatically generates an email on the thread when that happens.\n\nEditable certainly, but I don't think we want any auto-generated mail.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 12 Nov 2021 10:03:16 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Patch abstracts in the Commitfest app"},{"msg_contents":"On 11/12/21 7:51 AM, Daniel Gustafsson wrote:\n> \n> It should be a free-form textfield, and only apply to new submissions to avoid\n> a need to backfill. The field should, IMHO, be mandatory. The implementation\n> can be left to discussion on -www iff this list agrees that it would be a good\n> thing.\n\n+1.\n\nThe patch title is often not enough to tell what the patch does, and \npatches also change over time. The abstract may also be helpful to the \ncommitter when generating a commit message.\n\nRegards,\n-- \n-David\ndavid@pgmasters.net\n\n\n","msg_date":"Fri, 12 Nov 2021 10:31:35 -0500","msg_from":"David Steele <david@pgmasters.net>","msg_from_op":false,"msg_subject":"Re: Patch abstracts in the Commitfest app"}],"string":"[\n {\n \"msg_contents\": \"While reading through and working with the hundreds of patches in the CF app a\\nsmall feature/process request struck me: it would be really helpful if the\\npatch had a brief abstract outlining what it aims to add or fix (or summary,\\ndescription or something else; not sure what to call it). Basically a\\ntwo-sentence or so version of the email posting the patch to -hackers.\\n\\nThe \\\"title\\\" field isn't always explanatory enough, and diving into a long\\nthread to piece together clues can be quite daunting. I'm especially thinking\\nabout new contributors who would like to contribute but who have a hard time\\nfinding something in the CF app (which is a complaint I've heard from multiple\\nsuch contributors).\\n\\nIt should be a free-form textfield, and only apply to new submissions to avoid\\na need to backfill. The field should, IMHO, be mandatory. The implementation\\ncan be left to discussion on -www iff this list agrees that it would be a good\\nthing.\\n\\nThoughts?\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 13:51:28 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Patch abstracts in the Commitfest app\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 8:51 PM Daniel Gustafsson <daniel@yesql.se> wrote:\\n>\\n> While reading through and working with the hundreds of patches in the CF app a\\n> small feature/process request struck me: it would be really helpful if the\\n> patch had a brief abstract outlining what it aims to add or fix (or summary,\\n> description or something else; not sure what to call it). Basically a\\n> two-sentence or so version of the email posting the patch to -hackers.\\n>\\n> The \\\"title\\\" field isn't always explanatory enough, and diving into a long\\n> thread to piece together clues can be quite daunting. I'm especially thinking\\n> about new contributors who would like to contribute but who have a hard time\\n> finding something in the CF app (which is a complaint I've heard from multiple\\n> such contributors).\\n>\\n> It should be a free-form textfield, and only apply to new submissions to avoid\\n> a need to backfill. The field should, IMHO, be mandatory. The implementation\\n> can be left to discussion on -www iff this list agrees that it would be a good\\n> thing.\\n\\n+1 for the idea. Not sure if that should be mandatory though, as some\\nsimple entries should be self explanatory. Really this is mainly\\nhelpful for long and lengthy threads, when people start to ask a\\nsummary of the situation, so it should be editable and maybe\\nautomatically generates an email on the thread when that happens.\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 20:59:57 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch abstracts in the Commitfest app\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 01:51:28PM +0100, Daniel Gustafsson wrote:\\n> While reading through and working with the hundreds of patches in the CF app a\\n> small feature/process request struck me: it would be really helpful if the\\n> patch had a brief abstract outlining what it aims to add or fix (or summary,\\n> description or something else; not sure what to call it). Basically a\\n> two-sentence or so version of the email posting the patch to -hackers.\\n\\nThis seems fine ; that purpose is partially served (and duplicated) by the\\npatch commit messages (if used).\\n\\nActually, it's already possible to add an \\\"annotation\\\" to a given message-id.\\nI've just added this here: https://commitfest.postgresql.org/35/2800/\\n\\nI think this that feature is rarely used, but I can see the value in using it\\nmore, for long threads.\\n\\nIf you really wanted to make an \\\"abstract\\\" not associated with any message-id,\\nI think it would be implemented simply by removing the restriction that an\\nmessage-id must be specified and must match an existing message in the thead.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 08:24:01 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch abstracts in the Commitfest app\"\n },\n {\n \"msg_contents\": \"> On 12 Nov 2021, at 15:24, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> \\n> On Fri, Nov 12, 2021 at 01:51:28PM +0100, Daniel Gustafsson wrote:\\n>> While reading through and working with the hundreds of patches in the CF app a\\n>> small feature/process request struck me: it would be really helpful if the\\n>> patch had a brief abstract outlining what it aims to add or fix (or summary,\\n>> description or something else; not sure what to call it). Basically a\\n>> two-sentence or so version of the email posting the patch to -hackers.\\n> \\n> This seems fine ; that purpose is partially served (and duplicated) by the\\n> patch commit messages (if used).\\n\\nThat's the problem, many patches are in diff format and don't have commit\\nmessages at all. There are also many entries with patchsets containing\\nmultiple patches and thus commitmessages but the app will only link to one.\\n\\n> Actually, it's already possible to add an \\\"annotation\\\" to a given message-id.\\n> I've just added this here: https://commitfest.postgresql.org/35/2800/\\n> \\n> I think this that feature is rarely used, but I can see the value in using it\\n> more, for long threads.\\n\\nI consider this feature very valuable, but I don't see it as a replacement for\\na text field Abstract in the patch entry.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 15:36:43 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Patch abstracts in the Commitfest app\"\n },\n {\n \"msg_contents\": \"On 11/12/21, 4:53 AM, \\\"Daniel Gustafsson\\\" <daniel@yesql.se> wrote:\\r\\n> While reading through and working with the hundreds of patches in the CF app a\\r\\n> small feature/process request struck me: it would be really helpful if the\\r\\n> patch had a brief abstract outlining what it aims to add or fix (or summary,\\r\\n> description or something else; not sure what to call it). Basically a\\r\\n> two-sentence or so version of the email posting the patch to -hackers.\\r\\n\\r\\n+1\\r\\n\\r\\nNathan\\r\\n\\r\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 14:50:11 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch abstracts in the Commitfest app\"\n },\n {\n \"msg_contents\": \"> On Fri, Nov 12, 2021 at 03:36:43PM +0100, Daniel Gustafsson wrote:\\n> > On 12 Nov 2021, at 15:24, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> >\\n> > On Fri, Nov 12, 2021 at 01:51:28PM +0100, Daniel Gustafsson wrote:\\n> >> While reading through and working with the hundreds of patches in the CF app a\\n> >> small feature/process request struck me: it would be really helpful if the\\n> >> patch had a brief abstract outlining what it aims to add or fix (or summary,\\n> >> description or something else; not sure what to call it). Basically a\\n> >> two-sentence or so version of the email posting the patch to -hackers.\\n> >\\n> > This seems fine ; that purpose is partially served (and duplicated) by the\\n> > patch commit messages (if used).\\n>\\n> That's the problem, many patches are in diff format and don't have commit\\n> messages at all. There are also many entries with patchsets containing\\n> multiple patches and thus commitmessages but the app will only link to one.\\n\\nProbably encouraging to use cover letters, generated for such patch\\nseries and linked by the CF app, would be a good idea?\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 15:51:33 +0100\",\n \"msg_from\": \"Dmitry Dolgov <9erthalion6@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch abstracts in the Commitfest app\"\n },\n {\n \"msg_contents\": \"Julien Rouhaud <rjuju123@gmail.com> writes:\\n> On Fri, Nov 12, 2021 at 8:51 PM Daniel Gustafsson <daniel@yesql.se> wrote:\\n>> It should be a free-form textfield, and only apply to new submissions to avoid\\n>> a need to backfill. The field should, IMHO, be mandatory. The implementation\\n>> can be left to discussion on -www iff this list agrees that it would be a good\\n>> thing.\\n\\n> +1 for the idea. Not sure if that should be mandatory though, as some\\n> simple entries should be self explanatory.\\n\\nYeah, there are plenty of entries for which the title seems adequate.\\nNo objection to making room for an additional optional description,\\nthough.\\n\\n> Really this is mainly\\n> helpful for long and lengthy threads, when people start to ask a\\n> summary of the situation, so it should be editable and maybe\\n> automatically generates an email on the thread when that happens.\\n\\nEditable certainly, but I don't think we want any auto-generated mail.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 10:03:16 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch abstracts in the Commitfest app\"\n },\n {\n \"msg_contents\": \"On 11/12/21 7:51 AM, Daniel Gustafsson wrote:\\n> \\n> It should be a free-form textfield, and only apply to new submissions to avoid\\n> a need to backfill. The field should, IMHO, be mandatory. The implementation\\n> can be left to discussion on -www iff this list agrees that it would be a good\\n> thing.\\n\\n+1.\\n\\nThe patch title is often not enough to tell what the patch does, and \\npatches also change over time. The abstract may also be helpful to the \\ncommitter when generating a commit message.\\n\\nRegards,\\n-- \\n-David\\ndavid@pgmasters.net\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 10:31:35 -0500\",\n \"msg_from\": \"David Steele <david@pgmasters.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch abstracts in the Commitfest app\"\n }\n]","isConversational":false}}},{"rowIdx":53081,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"The combination of these two statements in a transaction hits an\nAssert in heapam.c at line 4770 on REL_14_STABLE\n\nBEGIN;\nSELECT * FROM queue LIMIT 1 FOR UPDATE SKIP LOCKED;\n...\nUPDATE queue SET status = 'UPDATED' WHERE id = :id;\nCOMMIT;\n\npgbench reliably finds this, running from inside a PL/pgSQL function.\n\nAlvaro suggests we just ignore the Assert, which on testing appears to\nbe the right approach. Patch attached, which solves it for me.\n\nThere is no formal test that does lock then update, so I have\nattempted to create one, but this has not successfully reproduced it\n(attached anyway), but this code is different from the pgbench test.\n\n-- \nSimon Riggs http://www.EnterpriseDB.com/","msg_date":"Fri, 12 Nov 2021 16:55:17 +0000","msg_from":"Simon Riggs <simon.riggs@enterprisedb.com>","msg_from_op":true,"msg_subject":"SKIP LOCKED assert triggered"},{"msg_contents":"On 11/12/21, 8:56 AM, \"Simon Riggs\" <simon.riggs@enterprisedb.com> wrote:\r\n> The combination of these two statements in a transaction hits an\r\n> Assert in heapam.c at line 4770 on REL_14_STABLE\r\n\r\nI've been unable to reproduce this. Do you have any tips for how to\r\ndo so? Does there need to be some sort of concurrent workload?\r\n\r\nNathan\r\n\r\n","msg_date":"Wed, 1 Dec 2021 19:33:05 +0000","msg_from":"\"Bossart, Nathan\" <bossartn@amazon.com>","msg_from_op":false,"msg_subject":"Re: SKIP LOCKED assert triggered"},{"msg_contents":"On Wed, 1 Dec 2021 at 14:33, Bossart, Nathan <bossartn@amazon.com> wrote:\n>\n> On 11/12/21, 8:56 AM, \"Simon Riggs\" <simon.riggs@enterprisedb.com> wrote:\n> > The combination of these two statements in a transaction hits an\n> > Assert in heapam.c at line 4770 on REL_14_STABLE\n>\n> I've been unable to reproduce this. Do you have any tips for how to\n> do so? Does there need to be some sort of concurrent workload?\n\nThat path is only ever taken when there are multiple sessions, and as\nI said, pgbench finds this reliably. I guess I didn't say \"use -c 2\"\n\n-- \nSimon Riggs http://www.EnterpriseDB.com/\n\n\n","msg_date":"Wed, 1 Dec 2021 19:38:22 +0000","msg_from":"Simon Riggs <simon.riggs@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: SKIP LOCKED assert triggered"},{"msg_contents":"On 2021-Dec-01, Simon Riggs wrote:\n\n> On Wed, 1 Dec 2021 at 14:33, Bossart, Nathan <bossartn@amazon.com> wrote:\n> >\n> > On 11/12/21, 8:56 AM, \"Simon Riggs\" <simon.riggs@enterprisedb.com> wrote:\n> > > The combination of these two statements in a transaction hits an\n> > > Assert in heapam.c at line 4770 on REL_14_STABLE\n> >\n> > I've been unable to reproduce this. Do you have any tips for how to\n> > do so? Does there need to be some sort of concurrent workload?\n> \n> That path is only ever taken when there are multiple sessions, and as\n> I said, pgbench finds this reliably. I guess I didn't say \"use -c 2\"\n\nSimon had sent me the pgbench scripts earlier, so I attach them here.\nI don't actually get a crash with -c2 or -c3, but I do get almost\nimmediate crashes with -c4 and above. If I run it under \"rr\", it\ndoesn't occur either. I suspect the rr framework kills concurrency in\nsome way that hides the problem. I didn't find a way to reproduce it\nwith isolationtester. (If somebody wants to play with a debugger, I\nfind that it's much easier to reproduce by adding a short sleep after\nthe UpdateXmaxHintBits() call in line 4735; but that sleep occurs in a\nsession *other* than the one that dies. And under rr I still don't see\na crash with a sleep there; in fact the sleep doesn't seem to occur at\nall, which is weird.)\n\nThe patch does fix the crasher under pgbench, and I think it makes sense\nthat you can get WouldBlock and yet have the tuple marked with\nXMAX_INVALID: if transaction A is writing the tuple, and transaction B\nis acquiring the tuple lock, then transaction C also tries to acquire\nthe tuple lock but that returns nay (because of B), then transaction A\ncompletes, then transaction B could set the XMAX_INVALID flag in time\nfor C to have a seizure in its way out. So patching the assertion to\nallow the case is correct.\n\n\nWhat I don't understand is why hasn't this been reported already: this\nbug is pretty old. My only explanation is that nobody runs sufficiently-\nconcurrent load with SKIP LOCKED in assert-enabled builds.\n\n[1] https://www.postgresql.org/message-id/flat/CADLWmXUvd5Z%2BcFczi6Zj1WcTrXzipgP-wj0pZOWSaRUy%3DF0omQ%40mail.gmail.com\n\n-- \nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/","msg_date":"Mon, 3 Jan 2022 19:27:42 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: SKIP LOCKED assert triggered"},{"msg_contents":"On 2022-Jan-03, Alvaro Herrera wrote:\n\n> What I don't understand is why hasn't this been reported already: this\n> bug is pretty old. My only explanation is that nobody runs sufficiently-\n> concurrent load with SKIP LOCKED in assert-enabled builds.\n\nPushed, thanks Simon for reporting this problem.\n\n\n> [1] https://www.postgresql.org/message-id/flat/CADLWmXUvd5Z%2BcFczi6Zj1WcTrXzipgP-wj0pZOWSaRUy%3DF0omQ%40mail.gmail.com\n\nHeh, I deleted a paragraph from my previous email and forgot to remove\nthe footnote that it referenced.\n\n-- \nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\n\"I love the Postgres community. It's all about doing things _properly_. :-)\"\n(David Garamond)\n\n\n","msg_date":"Tue, 4 Jan 2022 13:03:38 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: SKIP LOCKED assert triggered"},{"msg_contents":"On 2022-Jan-04, Alvaro Herrera wrote:\n\n> On 2022-Jan-03, Alvaro Herrera wrote:\n> \n> > What I don't understand is why hasn't this been reported already: this\n> > bug is pretty old. My only explanation is that nobody runs sufficiently-\n> > concurrent load with SKIP LOCKED in assert-enabled builds.\n> \n> Pushed, thanks Simon for reporting this problem.\n\nWow, what an embarrassing problem this fix has.\n\n-- \nÁlvaro Herrera 39°49'30\"S 73°17'W — https://www.EnterpriseDB.com/\n\"Pensar que el espectro que vemos es ilusorio no lo despoja de espanto,\nsólo le suma el nuevo terror de la locura\" (Perelandra, C.S. Lewis)\n\n\n","msg_date":"Tue, 4 Jan 2022 13:13:23 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: SKIP LOCKED assert triggered"},{"msg_contents":"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> Pushed, thanks Simon for reporting this problem.\n\nUmm ...\n\n Assert(TM_WouldBlock || !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));\n\nAFAICS, this assertion condition is constant-true,\nbecause TM_WouldBlock is a nonzero constant. Perhaps you meant\n\n Assert(result == TM_WouldBlock || !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));\n\n?\n\nI'd be inclined to format it more like the adjacent Assert, too.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 04 Jan 2022 11:15:30 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: SKIP LOCKED assert triggered"},{"msg_contents":"On Tue, Jan 04, 2022 at 11:15:30AM -0500, Tom Lane wrote:\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> > Pushed, thanks Simon for reporting this problem.\n> \n> Umm ...\n> \n> Assert(TM_WouldBlock || !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));\n> \n> AFAICS, this assertion condition is constant-true,\n\nThe new cfbot was failing like this\n\nhttps://cirrus-ci.com/github/postgresql-cfbot/postgresql/commitfest/36/3423\nhttps://cirrus-ci.com/build/5839382107127808\n[22:52:27.978] heapam.c:4754:24: error: converting the enum constant to a boolean [-Werror,-Wint-in-bool-context]\n\n\n","msg_date":"Tue, 4 Jan 2022 10:55:59 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: SKIP LOCKED assert triggered"},{"msg_contents":"On Tue, 4 Jan 2022 at 16:15, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> > Pushed, thanks Simon for reporting this problem.\n\nAnd causing another; my bad, apologies.\n\n> Umm ...\n>\n> Assert(TM_WouldBlock || !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));\n>\n> AFAICS, this assertion condition is constant-true,\n> because TM_WouldBlock is a nonzero constant. Perhaps you meant\n>\n> Assert(result == TM_WouldBlock || !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));\n\nYes, I think that's what I meant.\n\n-- \nSimon Riggs http://www.EnterpriseDB.com/\n\n\n","msg_date":"Tue, 4 Jan 2022 20:09:37 +0000","msg_from":"Simon Riggs <simon.riggs@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: SKIP LOCKED assert triggered"}],"string":"[\n {\n \"msg_contents\": \"The combination of these two statements in a transaction hits an\\nAssert in heapam.c at line 4770 on REL_14_STABLE\\n\\nBEGIN;\\nSELECT * FROM queue LIMIT 1 FOR UPDATE SKIP LOCKED;\\n...\\nUPDATE queue SET status = 'UPDATED' WHERE id = :id;\\nCOMMIT;\\n\\npgbench reliably finds this, running from inside a PL/pgSQL function.\\n\\nAlvaro suggests we just ignore the Assert, which on testing appears to\\nbe the right approach. Patch attached, which solves it for me.\\n\\nThere is no formal test that does lock then update, so I have\\nattempted to create one, but this has not successfully reproduced it\\n(attached anyway), but this code is different from the pgbench test.\\n\\n-- \\nSimon Riggs http://www.EnterpriseDB.com/\",\n \"msg_date\": \"Fri, 12 Nov 2021 16:55:17 +0000\",\n \"msg_from\": \"Simon Riggs <simon.riggs@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"SKIP LOCKED assert triggered\"\n },\n {\n \"msg_contents\": \"On 11/12/21, 8:56 AM, \\\"Simon Riggs\\\" <simon.riggs@enterprisedb.com> wrote:\\r\\n> The combination of these two statements in a transaction hits an\\r\\n> Assert in heapam.c at line 4770 on REL_14_STABLE\\r\\n\\r\\nI've been unable to reproduce this. Do you have any tips for how to\\r\\ndo so? Does there need to be some sort of concurrent workload?\\r\\n\\r\\nNathan\\r\\n\\r\\n\",\n \"msg_date\": \"Wed, 1 Dec 2021 19:33:05 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: SKIP LOCKED assert triggered\"\n },\n {\n \"msg_contents\": \"On Wed, 1 Dec 2021 at 14:33, Bossart, Nathan <bossartn@amazon.com> wrote:\\n>\\n> On 11/12/21, 8:56 AM, \\\"Simon Riggs\\\" <simon.riggs@enterprisedb.com> wrote:\\n> > The combination of these two statements in a transaction hits an\\n> > Assert in heapam.c at line 4770 on REL_14_STABLE\\n>\\n> I've been unable to reproduce this. Do you have any tips for how to\\n> do so? Does there need to be some sort of concurrent workload?\\n\\nThat path is only ever taken when there are multiple sessions, and as\\nI said, pgbench finds this reliably. I guess I didn't say \\\"use -c 2\\\"\\n\\n-- \\nSimon Riggs http://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Dec 2021 19:38:22 +0000\",\n \"msg_from\": \"Simon Riggs <simon.riggs@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: SKIP LOCKED assert triggered\"\n },\n {\n \"msg_contents\": \"On 2021-Dec-01, Simon Riggs wrote:\\n\\n> On Wed, 1 Dec 2021 at 14:33, Bossart, Nathan <bossartn@amazon.com> wrote:\\n> >\\n> > On 11/12/21, 8:56 AM, \\\"Simon Riggs\\\" <simon.riggs@enterprisedb.com> wrote:\\n> > > The combination of these two statements in a transaction hits an\\n> > > Assert in heapam.c at line 4770 on REL_14_STABLE\\n> >\\n> > I've been unable to reproduce this. Do you have any tips for how to\\n> > do so? Does there need to be some sort of concurrent workload?\\n> \\n> That path is only ever taken when there are multiple sessions, and as\\n> I said, pgbench finds this reliably. I guess I didn't say \\\"use -c 2\\\"\\n\\nSimon had sent me the pgbench scripts earlier, so I attach them here.\\nI don't actually get a crash with -c2 or -c3, but I do get almost\\nimmediate crashes with -c4 and above. If I run it under \\\"rr\\\", it\\ndoesn't occur either. I suspect the rr framework kills concurrency in\\nsome way that hides the problem. I didn't find a way to reproduce it\\nwith isolationtester. (If somebody wants to play with a debugger, I\\nfind that it's much easier to reproduce by adding a short sleep after\\nthe UpdateXmaxHintBits() call in line 4735; but that sleep occurs in a\\nsession *other* than the one that dies. And under rr I still don't see\\na crash with a sleep there; in fact the sleep doesn't seem to occur at\\nall, which is weird.)\\n\\nThe patch does fix the crasher under pgbench, and I think it makes sense\\nthat you can get WouldBlock and yet have the tuple marked with\\nXMAX_INVALID: if transaction A is writing the tuple, and transaction B\\nis acquiring the tuple lock, then transaction C also tries to acquire\\nthe tuple lock but that returns nay (because of B), then transaction A\\ncompletes, then transaction B could set the XMAX_INVALID flag in time\\nfor C to have a seizure in its way out. So patching the assertion to\\nallow the case is correct.\\n\\n\\nWhat I don't understand is why hasn't this been reported already: this\\nbug is pretty old. My only explanation is that nobody runs sufficiently-\\nconcurrent load with SKIP LOCKED in assert-enabled builds.\\n\\n[1] https://www.postgresql.org/message-id/flat/CADLWmXUvd5Z%2BcFczi6Zj1WcTrXzipgP-wj0pZOWSaRUy%3DF0omQ%40mail.gmail.com\\n\\n-- \\nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\",\n \"msg_date\": \"Mon, 3 Jan 2022 19:27:42 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: SKIP LOCKED assert triggered\"\n },\n {\n \"msg_contents\": \"On 2022-Jan-03, Alvaro Herrera wrote:\\n\\n> What I don't understand is why hasn't this been reported already: this\\n> bug is pretty old. My only explanation is that nobody runs sufficiently-\\n> concurrent load with SKIP LOCKED in assert-enabled builds.\\n\\nPushed, thanks Simon for reporting this problem.\\n\\n\\n> [1] https://www.postgresql.org/message-id/flat/CADLWmXUvd5Z%2BcFczi6Zj1WcTrXzipgP-wj0pZOWSaRUy%3DF0omQ%40mail.gmail.com\\n\\nHeh, I deleted a paragraph from my previous email and forgot to remove\\nthe footnote that it referenced.\\n\\n-- \\nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\\n\\\"I love the Postgres community. It's all about doing things _properly_. :-)\\\"\\n(David Garamond)\\n\\n\\n\",\n \"msg_date\": \"Tue, 4 Jan 2022 13:03:38 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: SKIP LOCKED assert triggered\"\n },\n {\n \"msg_contents\": \"On 2022-Jan-04, Alvaro Herrera wrote:\\n\\n> On 2022-Jan-03, Alvaro Herrera wrote:\\n> \\n> > What I don't understand is why hasn't this been reported already: this\\n> > bug is pretty old. My only explanation is that nobody runs sufficiently-\\n> > concurrent load with SKIP LOCKED in assert-enabled builds.\\n> \\n> Pushed, thanks Simon for reporting this problem.\\n\\nWow, what an embarrassing problem this fix has.\\n\\n-- \\nÁlvaro Herrera 39°49'30\\\"S 73°17'W — https://www.EnterpriseDB.com/\\n\\\"Pensar que el espectro que vemos es ilusorio no lo despoja de espanto,\\nsólo le suma el nuevo terror de la locura\\\" (Perelandra, C.S. Lewis)\\n\\n\\n\",\n \"msg_date\": \"Tue, 4 Jan 2022 13:13:23 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: SKIP LOCKED assert triggered\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> Pushed, thanks Simon for reporting this problem.\\n\\nUmm ...\\n\\n Assert(TM_WouldBlock || !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));\\n\\nAFAICS, this assertion condition is constant-true,\\nbecause TM_WouldBlock is a nonzero constant. Perhaps you meant\\n\\n Assert(result == TM_WouldBlock || !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));\\n\\n?\\n\\nI'd be inclined to format it more like the adjacent Assert, too.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 04 Jan 2022 11:15:30 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: SKIP LOCKED assert triggered\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 04, 2022 at 11:15:30AM -0500, Tom Lane wrote:\\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> > Pushed, thanks Simon for reporting this problem.\\n> \\n> Umm ...\\n> \\n> Assert(TM_WouldBlock || !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));\\n> \\n> AFAICS, this assertion condition is constant-true,\\n\\nThe new cfbot was failing like this\\n\\nhttps://cirrus-ci.com/github/postgresql-cfbot/postgresql/commitfest/36/3423\\nhttps://cirrus-ci.com/build/5839382107127808\\n[22:52:27.978] heapam.c:4754:24: error: converting the enum constant to a boolean [-Werror,-Wint-in-bool-context]\\n\\n\\n\",\n \"msg_date\": \"Tue, 4 Jan 2022 10:55:59 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: SKIP LOCKED assert triggered\"\n },\n {\n \"msg_contents\": \"On Tue, 4 Jan 2022 at 16:15, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> > Pushed, thanks Simon for reporting this problem.\\n\\nAnd causing another; my bad, apologies.\\n\\n> Umm ...\\n>\\n> Assert(TM_WouldBlock || !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));\\n>\\n> AFAICS, this assertion condition is constant-true,\\n> because TM_WouldBlock is a nonzero constant. Perhaps you meant\\n>\\n> Assert(result == TM_WouldBlock || !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));\\n\\nYes, I think that's what I meant.\\n\\n-- \\nSimon Riggs http://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Tue, 4 Jan 2022 20:09:37 +0000\",\n \"msg_from\": \"Simon Riggs <simon.riggs@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: SKIP LOCKED assert triggered\"\n }\n]","isConversational":false}}},{"rowIdx":53082,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"The following code appears at the end of heap_page_prune():\n\n /*\n * If requested, report the number of tuples reclaimed to pgstats. This is\n * ndeleted minus ndead, because we don't want to count a now-DEAD root\n * item as a deletion for this purpose.\n */\n if (report_stats && ndeleted > prstate.ndead)\n pgstat_update_heap_dead_tuples(relation, ndeleted - prstate.ndead);\n\nIn general, heap_page_prune() has to know the exact definition of\nndeleted (which is its return value) -- and the definition is kind of\nsubtle. The definition is really \"the number of tuples that had\nstorage before pruning, but not after pruning, because pruning\nconsidered them DEAD and removed the tuple itself (though not\nnecessarily its line pointer, which could have just been marked\nLP_DEAD rather than LP_UNUSED)\". As far as I know, the handling for\nall that at the end of heap_prune_chain() (at the point that we decide\nsome tuples from a HOT chain are considered DEAD) is correct. It\nhandles all the subtleties in a way that's consistent.\n\nI can't see why we should need to compensate like this (by subtracting\nprstate.ndead) at all, because, as I said, heap_page_prune() *already*\ndoes everything for us. In particular, if the now-DEAD root item\nstarted out as an LP_NORMAL tuple with storage (a root item that is a\nplain heap tuple, or just a simple tuple on its own), then we already\nmake sure to count it in ndeleted here, at the end of\nheap_prune_chain():\n\n /*\n * If the root entry had been a normal tuple, we are deleting it, so\n * count it in the result. But changing a redirect (even to DEAD\n * state) doesn't count.\n */\n if (ItemIdIsNormal(rootlp))\n ndeleted++;\n\nSo we're always going to count tuples that had storage but then got\npruned away in ndeleted -- we *do* want to count root items when\nthey're LP_NORMAL. We also avoid counting root items when that is\ninappropriate: when they started out as an LP_REDIRECT, but became\nLP_DEAD. It would be inappropriate in the case of an LP_REDIRECT root\nitem specifically, because it doesn't start out with tuple storage\nanyway (heap-only tuples from the same HOT chain are another matter).\nSo why compensate when calling pgstat_update_heap_dead_tuples() at\nall?\n\nLet's assume that somehow I have it wrong. Even then, why should we\ncompensate like this for the stats collector, but not for VACUUM?\nThere's certainly no corresponding code in vacuumlazy.c that does a\nsimilar transformation with ndeleted.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Fri, 12 Nov 2021 10:45:52 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"Is heap_page_prune() stats collector accounting wrong?"},{"msg_contents":"On Fri, Nov 12, 2021 at 10:45 AM Peter Geoghegan <pg@bowt.ie> wrote:\n> Let's assume that somehow I have it wrong. Even then, why should we\n> compensate like this for the stats collector, but not for VACUUM?\n> There's certainly no corresponding code in vacuumlazy.c that does a\n> similar transformation with ndeleted.\n\nI think that I figured it out myself, but it's very confusing. Could\ndefinitely do with a better explanatory comment. Here's what I think\nis actually going on here:\n\nWe compensate here precisely because we are not running in VACUUM (it\nhas to be an opportunistic prune in practice).\n\nIf we're running in VACUUM, then we are justified in ignoring\nnewly-pruned LP_DEAD items, because 1.) They're going to be turned\ninto LP_UNUSED items in the same VACUUM anyway (barring edge-cases),\nand 2.) Newly-pruned LP_DEAD items are really no different to existing\nLP_DEAD items that VACUUM found on the page before its own pruning\noperation even began -- if VACUUM wants to count them or report on\nthem at all, then it had better count them itself, after pruning\n(which it only began to do in Postgres 14).\n\nIf we're not running in VACUUM, and have to make a statistics\ncollector call, then we don't want to forget about DEAD tuples that\nwere pruned-away (i.e. no longer have tuple storage) when they still\nhave an LP_DEAD stub item. There is obviously no justification for\njust ignoring LP_DEAD items there, because we don't know when VACUUM\nis going to run next (since we are not VACUUM).\n\nDoes that sound correct?\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Fri, 12 Nov 2021 11:29:22 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"Re: Is heap_page_prune() stats collector accounting wrong?"},{"msg_contents":"On Fri, Nov 12, 2021 at 11:29 AM Peter Geoghegan <pg@bowt.ie> wrote:\n> We compensate here precisely because we are not running in VACUUM (it\n> has to be an opportunistic prune in practice).\n\n> If we're not running in VACUUM, and have to make a statistics\n> collector call, then we don't want to forget about DEAD tuples that\n> were pruned-away (i.e. no longer have tuple storage) when they still\n> have an LP_DEAD stub item. There is obviously no justification for\n> just ignoring LP_DEAD items there, because we don't know when VACUUM\n> is going to run next (since we are not VACUUM).\n\nAttached patch clears this up by adding some comments. It also moves\nthe call to pgstat_update_heap_dead_tuples() from heap_page_prune() to\nheap_page_prune_opt(), which feels like a better place for it to me.\n\n-- \nPeter Geoghegan","msg_date":"Fri, 12 Nov 2021 12:42:24 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":true,"msg_subject":"Re: Is heap_page_prune() stats collector accounting wrong?"}],"string":"[\n {\n \"msg_contents\": \"The following code appears at the end of heap_page_prune():\\n\\n /*\\n * If requested, report the number of tuples reclaimed to pgstats. This is\\n * ndeleted minus ndead, because we don't want to count a now-DEAD root\\n * item as a deletion for this purpose.\\n */\\n if (report_stats && ndeleted > prstate.ndead)\\n pgstat_update_heap_dead_tuples(relation, ndeleted - prstate.ndead);\\n\\nIn general, heap_page_prune() has to know the exact definition of\\nndeleted (which is its return value) -- and the definition is kind of\\nsubtle. The definition is really \\\"the number of tuples that had\\nstorage before pruning, but not after pruning, because pruning\\nconsidered them DEAD and removed the tuple itself (though not\\nnecessarily its line pointer, which could have just been marked\\nLP_DEAD rather than LP_UNUSED)\\\". As far as I know, the handling for\\nall that at the end of heap_prune_chain() (at the point that we decide\\nsome tuples from a HOT chain are considered DEAD) is correct. It\\nhandles all the subtleties in a way that's consistent.\\n\\nI can't see why we should need to compensate like this (by subtracting\\nprstate.ndead) at all, because, as I said, heap_page_prune() *already*\\ndoes everything for us. In particular, if the now-DEAD root item\\nstarted out as an LP_NORMAL tuple with storage (a root item that is a\\nplain heap tuple, or just a simple tuple on its own), then we already\\nmake sure to count it in ndeleted here, at the end of\\nheap_prune_chain():\\n\\n /*\\n * If the root entry had been a normal tuple, we are deleting it, so\\n * count it in the result. But changing a redirect (even to DEAD\\n * state) doesn't count.\\n */\\n if (ItemIdIsNormal(rootlp))\\n ndeleted++;\\n\\nSo we're always going to count tuples that had storage but then got\\npruned away in ndeleted -- we *do* want to count root items when\\nthey're LP_NORMAL. We also avoid counting root items when that is\\ninappropriate: when they started out as an LP_REDIRECT, but became\\nLP_DEAD. It would be inappropriate in the case of an LP_REDIRECT root\\nitem specifically, because it doesn't start out with tuple storage\\nanyway (heap-only tuples from the same HOT chain are another matter).\\nSo why compensate when calling pgstat_update_heap_dead_tuples() at\\nall?\\n\\nLet's assume that somehow I have it wrong. Even then, why should we\\ncompensate like this for the stats collector, but not for VACUUM?\\nThere's certainly no corresponding code in vacuumlazy.c that does a\\nsimilar transformation with ndeleted.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 10:45:52 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Is heap_page_prune() stats collector accounting wrong?\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 10:45 AM Peter Geoghegan <pg@bowt.ie> wrote:\\n> Let's assume that somehow I have it wrong. Even then, why should we\\n> compensate like this for the stats collector, but not for VACUUM?\\n> There's certainly no corresponding code in vacuumlazy.c that does a\\n> similar transformation with ndeleted.\\n\\nI think that I figured it out myself, but it's very confusing. Could\\ndefinitely do with a better explanatory comment. Here's what I think\\nis actually going on here:\\n\\nWe compensate here precisely because we are not running in VACUUM (it\\nhas to be an opportunistic prune in practice).\\n\\nIf we're running in VACUUM, then we are justified in ignoring\\nnewly-pruned LP_DEAD items, because 1.) They're going to be turned\\ninto LP_UNUSED items in the same VACUUM anyway (barring edge-cases),\\nand 2.) Newly-pruned LP_DEAD items are really no different to existing\\nLP_DEAD items that VACUUM found on the page before its own pruning\\noperation even began -- if VACUUM wants to count them or report on\\nthem at all, then it had better count them itself, after pruning\\n(which it only began to do in Postgres 14).\\n\\nIf we're not running in VACUUM, and have to make a statistics\\ncollector call, then we don't want to forget about DEAD tuples that\\nwere pruned-away (i.e. no longer have tuple storage) when they still\\nhave an LP_DEAD stub item. There is obviously no justification for\\njust ignoring LP_DEAD items there, because we don't know when VACUUM\\nis going to run next (since we are not VACUUM).\\n\\nDoes that sound correct?\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 11:29:22 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Is heap_page_prune() stats collector accounting wrong?\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 11:29 AM Peter Geoghegan <pg@bowt.ie> wrote:\\n> We compensate here precisely because we are not running in VACUUM (it\\n> has to be an opportunistic prune in practice).\\n\\n> If we're not running in VACUUM, and have to make a statistics\\n> collector call, then we don't want to forget about DEAD tuples that\\n> were pruned-away (i.e. no longer have tuple storage) when they still\\n> have an LP_DEAD stub item. There is obviously no justification for\\n> just ignoring LP_DEAD items there, because we don't know when VACUUM\\n> is going to run next (since we are not VACUUM).\\n\\nAttached patch clears this up by adding some comments. It also moves\\nthe call to pgstat_update_heap_dead_tuples() from heap_page_prune() to\\nheap_page_prune_opt(), which feels like a better place for it to me.\\n\\n-- \\nPeter Geoghegan\",\n \"msg_date\": \"Fri, 12 Nov 2021 12:42:24 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Is heap_page_prune() stats collector accounting wrong?\"\n }\n]","isConversational":false}}},{"rowIdx":53083,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"I wasn't aware of $SUBJECT ... were you?\n\nHere's a demonstration:\n\ndrop table if exists pk, fk, pk1, pk2;\ncreate table pk (f1 int primary key) partition by list(f1);\ncreate table pk1 partition of pk for values in (1);\ncreate table pk2 partition of pk for values in (2);\ninsert into pk values(1);\ninsert into pk values(2);\ncreate table fk (f1 int references pk);\ninsert into fk values(1);\ninsert into fk values(2);\n\nIn session 1, next do\n\nregression=# begin isolation level serializable;\nBEGIN\nregression=*# select * from unrelated_table; -- to set the xact snapshot\n...\n\nNow in session 2, do\n\nregression=# delete from fk where f1=2;\nDELETE 1\nregression=# alter table pk detach partition pk2; \nALTER TABLE\n\nBack at session 1, we now see what's not only a serializability\nviolation, but a not-even-self-consistent view of the database:\n\nregression=*# select * from fk;\n f1 \n----\n 1\n 2\n(2 rows)\n\nregression=*# select * from pk;\n f1 \n----\n 1\n(1 row)\n\nThis is slightly ameliorated by the fact that if session 1 has\nalready touched either pk or fk, locking considerations will\nblock the DETACH. But only slightly.\n\n(Dropping a partition altogether has the same issue, of course.)\n\nAFAICS, the only real way to fix this is to acquire lock on\nthe target partition and then wait out any snapshots that are\nolder than the lock, just in case those transactions would look\nat the partitioned table later. I'm not sure if we want to go\nthere, but if we don't, we at least have to document this gotcha.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 12 Nov 2021 15:27:42 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"ALTER TABLE DETACH PARTITION violates serializability"},{"msg_contents":"On 2021-Nov-12, Tom Lane wrote:\n\n> I wasn't aware of $SUBJECT ... were you?\n\nYeah, I remember pointing out that DETACH and DROP and not fully correct\nfor serializability, but I can't find any thread where I said it in so\nmany words. At the time I had no ideas on how to fix it; the idea of\nwaiting for snapshots to go away didn't occur to anybody, or at least it\ndidn't reach me.\n\n> AFAICS, the only real way to fix this is to acquire lock on\n> the target partition and then wait out any snapshots that are\n> older than the lock, just in case those transactions would look\n> at the partitioned table later. I'm not sure if we want to go\n> there, but if we don't, we at least have to document this gotcha.\n\nALTER TABLE DETACH PARTITION CONCURRENTLY already has a\nwait-for-snapshots phase. It doesn't fix this problem, because it\ndoesn't wait for all snapshots, just the ones holding any lock on the\nparent table. I'm not sure how upset would people be if we made it wait\non *all* snapshots, and worse, to make it in all cases and not just\nCONCURRENTLY.\n\nI understand that the behavior is not fully correct, but given the way\nmost people are going to use this (which is that they're no longer\nterribly interested in the data of the partition being detached/dropped)\nand the severity of the penalties if we implement a full solution, I\nlean towards documenting it rather than fixing it.\n\nAnother option might be to play with the trade-offs in the CONCURRENTLY\nmode vs. the regular one. If we make the CONCURRENTLY mode wait for all\nsnapshots, we would only be making worse a mode that's already\ndocumented to potentially take a long time. So people who want safety\ncan use that mode, and the others are aware of the risk.\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\"I am amazed at [the pgsql-sql] mailing list for the wonderful support, and\nlack of hesitasion in answering a lost soul's question, I just wished the rest\nof the mailing list could be like this.\" (Fotis)\n (http://archives.postgresql.org/pgsql-sql/2006-06/msg00265.php)\n\n\n","msg_date":"Fri, 12 Nov 2021 20:07:04 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: ALTER TABLE DETACH PARTITION violates serializability"},{"msg_contents":"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> I understand that the behavior is not fully correct, but given the way\n> most people are going to use this (which is that they're no longer\n> terribly interested in the data of the partition being detached/dropped)\n> and the severity of the penalties if we implement a full solution, I\n> lean towards documenting it rather than fixing it.\n\n> Another option might be to play with the trade-offs in the CONCURRENTLY\n> mode vs. the regular one. If we make the CONCURRENTLY mode wait for all\n> snapshots, we would only be making worse a mode that's already\n> documented to potentially take a long time. So people who want safety\n> can use that mode, and the others are aware of the risk.\n\nMeh ... \"wrong by default\" doesn't seem like it fits the Postgres ethos.\nHow about adding an option UNSAFE (orthogonal to CONCURRENTLY) that\nactivates the current behavior, and without it we wait for everything?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 12 Nov 2021 18:21:46 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: ALTER TABLE DETACH PARTITION violates serializability"},{"msg_contents":"On Fri, Nov 12, 2021 at 3:27 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> I wasn't aware of $SUBJECT ... were you?\n\nI wasn't aware of this particular example, but I didn't spend much\ntime worrying about it, because DDL interleaved with DML generally\nisn't serializable anyway. If you issue \"SELECT * FROM\nsome_previously_unacessed_table\" in a transaction, it will find\nsome_table even if the XID that created that table is not visible to\nyour snapshot. And conversely if that table has been dropped by some\nXID that is not visible to your snapshot, you will get an error that\nthe table doesn't exist. There's also ATRewriteTable(), which is not\nMVCC-aware, and thus if someone applies that to a table before you\nfirst access it, you may see rows you shouldn't see and fail to see\nrows which you should see. I don't see those kinds of cases as being\nall that different from this one.\n\nMy judgement is that outwaiting other XIDs would make far more people\nunhappy than it would make happy. If we want to have that as an\noptional behavior, I guess that would be OK, but I sure wouldn't make\nit the default. I feel like I went to quite a bit of effort to keep\nthe lock levels as low as reasonably possible here and to allow as\nmuch concurrency as I could, and I bet there are still a LOT more\npeople who are unhappy about things that take more locks than they'd\nlike than there are people who are sad about these kinds of MVCC\nissues.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 12 Nov 2021 19:33:55 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: ALTER TABLE DETACH PARTITION violates serializability"},{"msg_contents":"On 2021-Nov-12, Tom Lane wrote:\n\n> Meh ... \"wrong by default\" doesn't seem like it fits the Postgres ethos.\n\nI certainly agree with this, yeah.\n\n> How about adding an option UNSAFE (orthogonal to CONCURRENTLY) that\n> activates the current behavior, and without it we wait for everything?\n\nSomething like that might be the best compromise. I don't like having\nto grab AEL on the parent table (which the non-concurrent version does)\nfor as long as every old snapshot is gone, so changing the default\nbehavior without any escape hatch doesn't sound very pallatable.\n\nI think the best behavior is the CONCURRENTLY option with your\nadditional wait phase -- if only it worked in a transaction block, it\nwould be perfect. Users doing DDL as strong as DETACH should be willing\nto wait potentially long time for it to finish, but only if it doesn't\ndisrupt concurrent load involving other partitions.\n\nI don't feel a need to add UNSAFE as an option to CONCURRENTLY, though.\nUsers are supposed to be already aware that it takes a long time, and it\ndoesn't block queries using the remaining partitions.\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Sat, 13 Nov 2021 12:29:55 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: ALTER TABLE DETACH PARTITION violates serializability"},{"msg_contents":"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> On 2021-Nov-12, Tom Lane wrote:\n>> How about adding an option UNSAFE (orthogonal to CONCURRENTLY) that\n>> activates the current behavior, and without it we wait for everything?\n\n> Something like that might be the best compromise. I don't like having\n> to grab AEL on the parent table (which the non-concurrent version does)\n> for as long as every old snapshot is gone, so changing the default\n> behavior without any escape hatch doesn't sound very pallatable.\n\n> I think the best behavior is the CONCURRENTLY option with your\n> additional wait phase -- if only it worked in a transaction block, it\n> would be perfect. Users doing DDL as strong as DETACH should be willing\n> to wait potentially long time for it to finish, but only if it doesn't\n> disrupt concurrent load involving other partitions.\n\nI grant that a lot of people will be less worried about the transactional\nsemantics problem than about how strong a lock they need. I'd be willing\nto settle for \"DETACH CONCURRENTLY is safe and plain DETACH isn't\", as\nlong as that's very prominently documented.\n\nHowever ... it appears to me that focusing on the lock strength for\nthe bare DETACH is misleading. When I was experimenting with this\nstuff yesterday, I observed that if I had any foreign keys on the\npartitioned table, DETACH would insist on AEL both on the partition\n*and* on the table that's the other end of the FK constraint.\nI suppose that's needed to get rid of triggers and pg_constraint\nentries. But it seems to put a large hole in the argument that\nyou're going to be able to avoid strong locks and corresponding\nwaits/blockages when you do a DETACH.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 13 Nov 2021 11:28:16 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: ALTER TABLE DETACH PARTITION violates serializability"}],"string":"[\n {\n \"msg_contents\": \"I wasn't aware of $SUBJECT ... were you?\\n\\nHere's a demonstration:\\n\\ndrop table if exists pk, fk, pk1, pk2;\\ncreate table pk (f1 int primary key) partition by list(f1);\\ncreate table pk1 partition of pk for values in (1);\\ncreate table pk2 partition of pk for values in (2);\\ninsert into pk values(1);\\ninsert into pk values(2);\\ncreate table fk (f1 int references pk);\\ninsert into fk values(1);\\ninsert into fk values(2);\\n\\nIn session 1, next do\\n\\nregression=# begin isolation level serializable;\\nBEGIN\\nregression=*# select * from unrelated_table; -- to set the xact snapshot\\n...\\n\\nNow in session 2, do\\n\\nregression=# delete from fk where f1=2;\\nDELETE 1\\nregression=# alter table pk detach partition pk2; \\nALTER TABLE\\n\\nBack at session 1, we now see what's not only a serializability\\nviolation, but a not-even-self-consistent view of the database:\\n\\nregression=*# select * from fk;\\n f1 \\n----\\n 1\\n 2\\n(2 rows)\\n\\nregression=*# select * from pk;\\n f1 \\n----\\n 1\\n(1 row)\\n\\nThis is slightly ameliorated by the fact that if session 1 has\\nalready touched either pk or fk, locking considerations will\\nblock the DETACH. But only slightly.\\n\\n(Dropping a partition altogether has the same issue, of course.)\\n\\nAFAICS, the only real way to fix this is to acquire lock on\\nthe target partition and then wait out any snapshots that are\\nolder than the lock, just in case those transactions would look\\nat the partitioned table later. I'm not sure if we want to go\\nthere, but if we don't, we at least have to document this gotcha.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 15:27:42 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"ALTER TABLE DETACH PARTITION violates serializability\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-12, Tom Lane wrote:\\n\\n> I wasn't aware of $SUBJECT ... were you?\\n\\nYeah, I remember pointing out that DETACH and DROP and not fully correct\\nfor serializability, but I can't find any thread where I said it in so\\nmany words. At the time I had no ideas on how to fix it; the idea of\\nwaiting for snapshots to go away didn't occur to anybody, or at least it\\ndidn't reach me.\\n\\n> AFAICS, the only real way to fix this is to acquire lock on\\n> the target partition and then wait out any snapshots that are\\n> older than the lock, just in case those transactions would look\\n> at the partitioned table later. I'm not sure if we want to go\\n> there, but if we don't, we at least have to document this gotcha.\\n\\nALTER TABLE DETACH PARTITION CONCURRENTLY already has a\\nwait-for-snapshots phase. It doesn't fix this problem, because it\\ndoesn't wait for all snapshots, just the ones holding any lock on the\\nparent table. I'm not sure how upset would people be if we made it wait\\non *all* snapshots, and worse, to make it in all cases and not just\\nCONCURRENTLY.\\n\\nI understand that the behavior is not fully correct, but given the way\\nmost people are going to use this (which is that they're no longer\\nterribly interested in the data of the partition being detached/dropped)\\nand the severity of the penalties if we implement a full solution, I\\nlean towards documenting it rather than fixing it.\\n\\nAnother option might be to play with the trade-offs in the CONCURRENTLY\\nmode vs. the regular one. If we make the CONCURRENTLY mode wait for all\\nsnapshots, we would only be making worse a mode that's already\\ndocumented to potentially take a long time. So people who want safety\\ncan use that mode, and the others are aware of the risk.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\\"I am amazed at [the pgsql-sql] mailing list for the wonderful support, and\\nlack of hesitasion in answering a lost soul's question, I just wished the rest\\nof the mailing list could be like this.\\\" (Fotis)\\n (http://archives.postgresql.org/pgsql-sql/2006-06/msg00265.php)\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 20:07:04 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ALTER TABLE DETACH PARTITION violates serializability\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> I understand that the behavior is not fully correct, but given the way\\n> most people are going to use this (which is that they're no longer\\n> terribly interested in the data of the partition being detached/dropped)\\n> and the severity of the penalties if we implement a full solution, I\\n> lean towards documenting it rather than fixing it.\\n\\n> Another option might be to play with the trade-offs in the CONCURRENTLY\\n> mode vs. the regular one. If we make the CONCURRENTLY mode wait for all\\n> snapshots, we would only be making worse a mode that's already\\n> documented to potentially take a long time. So people who want safety\\n> can use that mode, and the others are aware of the risk.\\n\\nMeh ... \\\"wrong by default\\\" doesn't seem like it fits the Postgres ethos.\\nHow about adding an option UNSAFE (orthogonal to CONCURRENTLY) that\\nactivates the current behavior, and without it we wait for everything?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 18:21:46 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ALTER TABLE DETACH PARTITION violates serializability\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 3:27 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> I wasn't aware of $SUBJECT ... were you?\\n\\nI wasn't aware of this particular example, but I didn't spend much\\ntime worrying about it, because DDL interleaved with DML generally\\nisn't serializable anyway. If you issue \\\"SELECT * FROM\\nsome_previously_unacessed_table\\\" in a transaction, it will find\\nsome_table even if the XID that created that table is not visible to\\nyour snapshot. And conversely if that table has been dropped by some\\nXID that is not visible to your snapshot, you will get an error that\\nthe table doesn't exist. There's also ATRewriteTable(), which is not\\nMVCC-aware, and thus if someone applies that to a table before you\\nfirst access it, you may see rows you shouldn't see and fail to see\\nrows which you should see. I don't see those kinds of cases as being\\nall that different from this one.\\n\\nMy judgement is that outwaiting other XIDs would make far more people\\nunhappy than it would make happy. If we want to have that as an\\noptional behavior, I guess that would be OK, but I sure wouldn't make\\nit the default. I feel like I went to quite a bit of effort to keep\\nthe lock levels as low as reasonably possible here and to allow as\\nmuch concurrency as I could, and I bet there are still a LOT more\\npeople who are unhappy about things that take more locks than they'd\\nlike than there are people who are sad about these kinds of MVCC\\nissues.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Nov 2021 19:33:55 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ALTER TABLE DETACH PARTITION violates serializability\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-12, Tom Lane wrote:\\n\\n> Meh ... \\\"wrong by default\\\" doesn't seem like it fits the Postgres ethos.\\n\\nI certainly agree with this, yeah.\\n\\n> How about adding an option UNSAFE (orthogonal to CONCURRENTLY) that\\n> activates the current behavior, and without it we wait for everything?\\n\\nSomething like that might be the best compromise. I don't like having\\nto grab AEL on the parent table (which the non-concurrent version does)\\nfor as long as every old snapshot is gone, so changing the default\\nbehavior without any escape hatch doesn't sound very pallatable.\\n\\nI think the best behavior is the CONCURRENTLY option with your\\nadditional wait phase -- if only it worked in a transaction block, it\\nwould be perfect. Users doing DDL as strong as DETACH should be willing\\nto wait potentially long time for it to finish, but only if it doesn't\\ndisrupt concurrent load involving other partitions.\\n\\nI don't feel a need to add UNSAFE as an option to CONCURRENTLY, though.\\nUsers are supposed to be already aware that it takes a long time, and it\\ndoesn't block queries using the remaining partitions.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 12:29:55 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ALTER TABLE DETACH PARTITION violates serializability\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> On 2021-Nov-12, Tom Lane wrote:\\n>> How about adding an option UNSAFE (orthogonal to CONCURRENTLY) that\\n>> activates the current behavior, and without it we wait for everything?\\n\\n> Something like that might be the best compromise. I don't like having\\n> to grab AEL on the parent table (which the non-concurrent version does)\\n> for as long as every old snapshot is gone, so changing the default\\n> behavior without any escape hatch doesn't sound very pallatable.\\n\\n> I think the best behavior is the CONCURRENTLY option with your\\n> additional wait phase -- if only it worked in a transaction block, it\\n> would be perfect. Users doing DDL as strong as DETACH should be willing\\n> to wait potentially long time for it to finish, but only if it doesn't\\n> disrupt concurrent load involving other partitions.\\n\\nI grant that a lot of people will be less worried about the transactional\\nsemantics problem than about how strong a lock they need. I'd be willing\\nto settle for \\\"DETACH CONCURRENTLY is safe and plain DETACH isn't\\\", as\\nlong as that's very prominently documented.\\n\\nHowever ... it appears to me that focusing on the lock strength for\\nthe bare DETACH is misleading. When I was experimenting with this\\nstuff yesterday, I observed that if I had any foreign keys on the\\npartitioned table, DETACH would insist on AEL both on the partition\\n*and* on the table that's the other end of the FK constraint.\\nI suppose that's needed to get rid of triggers and pg_constraint\\nentries. But it seems to put a large hole in the argument that\\nyou're going to be able to avoid strong locks and corresponding\\nwaits/blockages when you do a DETACH.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 11:28:16 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ALTER TABLE DETACH PARTITION violates serializability\"\n }\n]","isConversational":false}}},{"rowIdx":53084,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Re-reading the thread [1] (cannot answer there – don't have those emails in\nmy box anymore), I see that there was strong support for enabling BUFFERS\nin EXPLAIN ANALYZE by default. And there were patches. Commitfest entry [2]\nwas marked Rejected because there were questions to the implementation\nbased on GUCs.\n\nAttached is a simple patch enabling BUFFERS by default, without involving\nGUCs.\n\nWhy it is important?\n\nIn many cases, people forget about the BUFFERS option in EXPLAIN ANALYZE\nand share execution plans without it – sending it via regular communication\nchannels for work or publishing to visualization systems. Meanwhile, the\noption has a lower overhead compared to TIMING (enabled by default for\nEXPLAIN ANALYZE) and it is extremely useful for query\noptimization. This patch doesn't enable BUFFERS for EXPLAIN executed\nwithout ANALYZE.\n\nOpen questions:\n\n1. Should BUFFERS be enabled for regular (w/o ANALYZE) EXPLAIN? Now it may\nmake sense because of the buffer numbers the planner uses. This patch\ndoesn't do that, but it definitely may make sense because it can help\npeople understand why planning time is so big, in some cases.\n\n2. How to adjust documentation? Should EXPLAIN ANALYZE examples be adjusted\nto use BUFFERS OFF (easier change) or show some example buffer numbers –\nlike it is done for timing and cost numbers? I tend to think that the\nlatter makes more sense.\n\n3. How to adjust regression tests? Of course, now many tests fail. Same\nquestion as for documentation. Excluding buffer, numbers would be an easier\nfix, of course – but at least some tests should\ncheck the behavior of BUFFERS (such as both default options – with and\nwithout ANALYZE).\nOn any given platform, the buffer numbers are pretty stable, so we could\nrely on it, but I'm not sure about all possible options being tested and\nwould appreciate advice here (of course, if the patch makes it thru the\ndiscussion in general).\n\n\n## Links\n[1]\nhttps://www.postgresql.org/message-id/flat/b3197ba8-225f-f53c-326d-5b1756c77c3e%40postgresfriends.org\n[2] https://commitfest.postgresql.org/28/2567/","msg_date":"Fri, 12 Nov 2021 14:58:07 -0800","msg_from":"Nikolay Samokhvalov <samokhvalov@gmail.com>","msg_from_op":true,"msg_subject":"BUFFERS enabled by default in EXPLAIN (ANALYZE)"},{"msg_contents":"\n\nOn 11/12/21 23:58, Nikolay Samokhvalov wrote:\n> Re-reading the thread [1] (cannot answer there – don't have those emails \n> in my box anymore),\n\nYou can download the message as mbox and import it into your client \n(pretty much any client supports that, I think).\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Sat, 13 Nov 2021 01:18:31 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)"},{"msg_contents":"On 11/12/21 11:58 PM, Nikolay Samokhvalov wrote:\n> Re-reading the thread [1] (cannot answer there – don't have those emails in\n> my box anymore), I see that there was strong support for enabling BUFFERS\n> in EXPLAIN ANALYZE by default. And there were patches. Commitfest entry [2]\n> was marked Rejected because there were questions to the implementation\n> based on GUCs.\n> \n> Attached is a simple patch enabling BUFFERS by default, without involving\n> GUCs.\n\nI obviously agree with this (although I still wish we had gucs as we\nkeep adding more and more options to EXPLAIN).\n\nThe patch looks good to me, too.\n\n+1\n-- \nVik Fearing\n\n\n","msg_date":"Sat, 13 Nov 2021 02:34:05 +0100","msg_from":"Vik Fearing <vik@postgresfriends.org>","msg_from_op":false,"msg_subject":"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)"},{"msg_contents":"On Sat, Nov 13, 2021 at 8:18 AM Tomas Vondra\n<tomas.vondra@enterprisedb.com> wrote:\n>\n> On 11/12/21 23:58, Nikolay Samokhvalov wrote:\n> > Re-reading the thread [1] (cannot answer there – don't have those emails\n> > in my box anymore),\n>\n> You can download the message as mbox and import it into your client\n> (pretty much any client supports that, I think).\n\nThere is also a \"resend mail\" link available for each mail in the\narchive which should be even more straightforward.\n\n\n","msg_date":"Sat, 13 Nov 2021 12:17:58 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)"},{"msg_contents":"On Fri, Nov 12, 2021 at 02:58:07PM -0800, Nikolay Samokhvalov wrote:\n> Re-reading the thread [1] (cannot answer there – don't have those emails in\n> my box anymore), I see that there was strong support for enabling BUFFERS\n> in EXPLAIN ANALYZE by default. And there were patches. Commitfest entry [2]\n> was marked Rejected because there were questions to the implementation\n> based on GUCs.\n\nI think the only reason this isn't already the default is because of (3):\n\n> 3. How to adjust regression tests? Of course, now many tests fail. Same\n> question as for documentation. Excluding buffer, numbers would be an easier\n> fix, of course – but at least some tests should\n> check the behavior of BUFFERS (such as both default options – with and\n> without ANALYZE).\n\nSome time ago, I had a few relevant patches:\n1) add explain(REGRESS) which is shorthand for (BUFFERS OFF, TIMING OFF, COSTS OFF, SUMMARY OFF)\n2) add explain(MACHINE) which elides machine-specific output from explain;\n for example, Heap Fetches, sort spaceUsed, hash nbuckets, and tidbitmap stuff.\n\nhttps://www.postgresql.org/message-id/flat/20200306213310.GM684@telsasoft.com\n\n> 1. Should BUFFERS be enabled for regular (w/o ANALYZE) EXPLAIN? Now it may\n> make sense because of the buffer numbers the planner uses. This patch\n> doesn't do that, but it definitely may make sense because it can help\n> people understand why planning time is so big, in some cases.\n\nI think it *should* be enabled for planning, since that makes the default\neasier to understand and document, and it makes a user's use of \"explain\"\neasier.\n\nHowever, that makes this patch series more complicated: explain(ANALYZE) is\nrare in the regression tests, but this would affect the output of bare\n\"explain\", which affects almost every test.\n\nI think the answer to that is to add a GUC (as Tom suggested in an old thread)\nlike explain_regress=on, which causes explain to omit these details. This\nwould be instead of my explain(REGRESS), and would change the defaults of\nvarious params in a central place, to avoid the need to update many regression\ntests.\n\nIn the future, this might also handle the stuff that my \"explain(MACHINE)\"\nattempted to handle.\n\nI've rebased my patches like this. I think the explain(REGRESS) should be\nre-written as a GUC which is set by regress.c. I'm interested to hear from a\nreviewer if this is worth pursing like this.\n\n-- \nJustin","msg_date":"Mon, 15 Nov 2021 13:09:54 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)"},{"msg_contents":"I think it *should* be enabled for planning, since that makes the default\n> easier to understand and document, and it makes a user's use of \"explain\"\n> easier.\n\n\nI’d be keen to see BUFFERS off by default with EXPLAIN, and on by default\nwith EXPLAIN ANALYZE.\n\nThe SUMMARY flag was implemented that way, which I think has been easy\nenough for folks to understand and document. In fact, I think the only\nBUFFERS information goes in the “summary” section for EXPLAIN (BUFFERS), so\nmaybe it makes perfect sense? If it would be great if that made\nimplementation easier, too.\n\nIn any case, thank you all, I’m so glad that this is being discussed again.\nIt’d be so good to start seeing buffers in more plans.\n\n—\nMichael\n\nI think it *should* be enabled for planning, since that makes the default\neasier to understand and document, and it makes a user's use of \"explain\"\neasier.I’d be keen to see BUFFERS off by default with EXPLAIN, and on by default with EXPLAIN ANALYZE.The SUMMARY flag was implemented that way, which I think has been easy enough for folks to understand and document. In fact, I think the only BUFFERS information goes in the “summary” section for EXPLAIN (BUFFERS), so maybe it makes perfect sense? If it would be great if that made implementation easier, too.In any case, thank you all, I’m so glad that this is being discussed again. It’d be so good to start seeing buffers in more plans.—Michael","msg_date":"Wed, 24 Nov 2021 18:35:45 +0000","msg_from":"Michael Christofides <michael@pgmustard.com>","msg_from_op":false,"msg_subject":"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)"},{"msg_contents":"On Mon, Nov 15, 2021 at 01:09:54PM -0600, Justin Pryzby wrote:\n> Some time ago, I had a few relevant patches:\n> 1) add explain(REGRESS) which is shorthand for (BUFFERS OFF, TIMING OFF, COSTS OFF, SUMMARY OFF)\n> 2) add explain(MACHINE) which elides machine-specific output from explain;\n> for example, Heap Fetches, sort spaceUsed, hash nbuckets, and tidbitmap stuff.\n> \n> https://www.postgresql.org/message-id/flat/20200306213310.GM684@telsasoft.com\n\nThe attached patch series now looks like this (some minor patches are not\nincluded in this list):\n\n 1. add GUC explain_regress, which disables TIMING, SUMMARY, COSTS;\n 2. enable explain(BUFFERS) by default (but disabled by explain_regress);\n 3. Add explain(MACHINE) - which is disabled by explain_regress.\n This elides various machine-specific output like Memory and Disk use.\n Maybe it should be called something else like \"QUIET\" or \"VERBOSE_MINUS_1\"\n or ??\n\nThe regression tests now automatically run with explain_regress=on, which is\nshorthand for TIMING OFF, SUMMARY OFF, COSTS OFF, and then BUFFERS OFF.\n\nThere's a further option called explain(MACHINE) which can be disabled to hide\nportions of the output that are unstable, like Memory/Disk/Batches/\nHeap Fetches/Heap Blocks. This allows \"explain analyze\" to be used more easily\nin regression tests, and simplifies some existing tests that currently use\nplpgsql functions to filter the output. But it doesn't handle all the\nvariations from parallel workers.\n\n(3) is optional, but simplifies some regression tests. The patch series could\nbe rephrased with (3) first.\n\nUnfortunately, \"COSTS OFF\" breaks postgres_fdw remote_estimate. If specifying\n\"COSTS ON\" in postgres_fdw.c is considered to be a poor fix , then I suppose\nthis patch series could do as suggested and enable buffers by default only when\nANALYZE is specified. Then postgres_fdw is not affected, and the\nexplain_regress GUC is optional: instead, we'd need to specify BUFFERS OFF in\n~100 regression tests which use EXPLAIN ANALYZE. (3) still seems useful on its\nown.","msg_date":"Wed, 1 Dec 2021 18:58:20 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)"},{"msg_contents":"Hi,\n\nOn Wed, Dec 01, 2021 at 06:58:20PM -0600, Justin Pryzby wrote:\n> \n> The attached patch series now looks like this (some minor patches are not\n> included in this list):\n\nThis version of the patchset doesn't apply anymore:\n\nhttp://cfbot.cputube.org/patch_36_3409.log\n=== Applying patches on top of PostgreSQL commit ID 4483b2cf29bfe8091b721756928ccbe31c5c8e14 ===\n=== applying patch ./0003-Make-explain-default-to-BUFFERS-TRUE.patch\npatching file doc/src/sgml/config.sgml\nHunk #1 FAILED at 7615.\n1 out of 1 hunk FAILED -- saving rejects to file doc/src/sgml/config.sgml.rej\n\nBy the way, could you add versionning to the patchsets you send, e.g. git\nformat-patch -vXXX. It would make it easier to know which version are used or\ndiscussed.\n\nI'm switching the CF entry to Waiting on Author.\n\n\n","msg_date":"Tue, 18 Jan 2022 14:53:09 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)"},{"msg_contents":"I'm renaming this thread for better visibility, since buffers is a small,\noptional part of the patches I sent.\n\nI made a CF entry here.\nhttps://commitfest.postgresql.org/36/3409/\n\nOn Wed, Dec 01, 2021 at 06:58:20PM -0600, Justin Pryzby wrote:\n> On Mon, Nov 15, 2021 at 01:09:54PM -0600, Justin Pryzby wrote:\n> > Some time ago, I had a few relevant patches:\n> > 1) add explain(REGRESS) which is shorthand for (BUFFERS OFF, TIMING OFF, COSTS OFF, SUMMARY OFF)\n> > 2) add explain(MACHINE) which elides machine-specific output from explain;\n> > for example, Heap Fetches, sort spaceUsed, hash nbuckets, and tidbitmap stuff.\n> > \n> > https://www.postgresql.org/message-id/flat/20200306213310.GM684@telsasoft.com\n> \n> The attached patch series now looks like this (some minor patches are not\n> included in this list):\n> \n> 1. add GUC explain_regress, which disables TIMING, SUMMARY, COSTS;\n> 2. enable explain(BUFFERS) by default (but disabled by explain_regress);\n> 3. Add explain(MACHINE) - which is disabled by explain_regress.\n> This elides various machine-specific output like Memory and Disk use.\n> Maybe it should be called something else like \"QUIET\" or \"VERBOSE_MINUS_1\"\n> or ??\n> \n> The regression tests now automatically run with explain_regress=on, which is\n> shorthand for TIMING OFF, SUMMARY OFF, COSTS OFF, and then BUFFERS OFF.\n> \n> There's a further option called explain(MACHINE) which can be disabled to hide\n> portions of the output that are unstable, like Memory/Disk/Batches/\n> Heap Fetches/Heap Blocks. This allows \"explain analyze\" to be used more easily\n> in regression tests, and simplifies some existing tests that currently use\n> plpgsql functions to filter the output. But it doesn't handle all the\n> variations from parallel workers.\n> \n> (3) is optional, but simplifies some regression tests. The patch series could\n> be rephrased with (3) first.\n> \n> Unfortunately, \"COSTS OFF\" breaks postgres_fdw remote_estimate. If specifying\n> \"COSTS ON\" in postgres_fdw.c is considered to be a poor fix , then I suppose\n> this patch series could do as suggested and enable buffers by default only when\n> ANALYZE is specified. Then postgres_fdw is not affected, and the\n> explain_regress GUC is optional: instead, we'd need to specify BUFFERS OFF in\n> ~100 regression tests which use EXPLAIN ANALYZE. (3) still seems useful on its\n> own.","msg_date":"Mon, 24 Jan 2022 11:35:48 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"explain_regress, explain(MACHINE), and default to explain(BUFFERS)\n (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"},{"msg_contents":"Rebased over ebf6c5249b7db525e59563fb149642665c88f747.\nIt looks like that patch handles only query_id, and this patch also tries to\nhandle a bunch of other stuff.\n\nIf it's helpful, feel free to kick this patch to a future CF.","msg_date":"Sat, 26 Feb 2022 15:07:20 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE), and default to\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"},{"msg_contents":"On Sat, Feb 26, 2022 at 03:07:20PM -0600, Justin Pryzby wrote:\n> Rebased over ebf6c5249b7db525e59563fb149642665c88f747.\n> It looks like that patch handles only query_id, and this patch also tries to\n> handle a bunch of other stuff.\n> \n> If it's helpful, feel free to kick this patch to a future CF.\n\nRebased over MERGE.","msg_date":"Wed, 30 Mar 2022 14:21:05 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE), and default to\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"},{"msg_contents":"@cfbot: rebased","msg_date":"Thu, 7 Jul 2022 01:49:04 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE), and default to\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"},{"msg_contents":"On Mon, Jan 24, 2022 at 9:54 AM Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n> I'm renaming this thread for better visibility, since buffers is a small,\n> optional part of the patches I sent.\n>\n> I made a CF entry here.\n> https://commitfest.postgresql.org/36/3409/\n>\n> On Wed, Dec 01, 2021 at 06:58:20PM -0600, Justin Pryzby wrote:\n> > On Mon, Nov 15, 2021 at 01:09:54PM -0600, Justin Pryzby wrote:\n> > > Some time ago, I had a few relevant patches:\n> > > 1) add explain(REGRESS) which is shorthand for (BUFFERS OFF, TIMING\n> OFF, COSTS OFF, SUMMARY OFF)\n> > > 2) add explain(MACHINE) which elides machine-specific output from\n> explain;\n> > > for example, Heap Fetches, sort spaceUsed, hash nbuckets, and\n> tidbitmap stuff.\n> > >\n> > >\n> https://www.postgresql.org/message-id/flat/20200306213310.GM684@telsasoft.com\n> >\n> > The attached patch series now looks like this (some minor patches are not\n> > included in this list):\n> >\n> > 1. add GUC explain_regress, which disables TIMING, SUMMARY, COSTS;\n>\n\n\n> > 2. enable explain(BUFFERS) by default (but disabled by explain_regress);\n>\n\n+1\n\n\n> > 3. Add explain(MACHINE) - which is disabled by explain_regress.\n> > This elides various machine-specific output like Memory and Disk use.\n> > Maybe it should be called something else like \"QUIET\" or\n> \"VERBOSE_MINUS_1\"\n> > or ??\n>\n\nINCIDENTALS ?\n\nAside from being 4 syllables and, for me at least, a pain to remember how\nto spell, it seems to fit.\n\nRUNTIME is probably easier, and we just need to point out the difficulty in\nnaming things since actual rows is still shown, and timings have their own\nindependent option.\n\n>\n> > The regression tests now automatically run with explain_regress=on,\n> which is\n> > shorthand for TIMING OFF, SUMMARY OFF, COSTS OFF, and then BUFFERS OFF.\n> >\n> > There's a further option called explain(MACHINE) which can be disabled\n> to hide\n> > portions of the output that are unstable, like Memory/Disk/Batches/\n> > Heap Fetches/Heap Blocks. This allows \"explain analyze\" to be used more\n> easily\n> > in regression tests, and simplifies some existing tests that currently\n> use\n> > plpgsql functions to filter the output. But it doesn't handle all the\n> > variations from parallel workers.\n> >\n> > (3) is optional, but simplifies some regression tests. The patch series\n> could\n> > be rephrased with (3) first.\n>\n\nI like the idea of encapsulating the logic about what to show into the\ngeneration code instead of a post-processing routine.\n\nI don't see any particular ordering of the three changes being most clear.\n\n>\n> > Unfortunately, \"COSTS OFF\" breaks postgres_fdw remote_estimate. If\n> specifying\n> > \"COSTS ON\" in postgres_fdw.c is considered to be a poor fix\n\n\nGiven that we have version tests scattered throughout the postgres_fdw.c\ncode I'd say protect it with version >= 9.0 and call it good.\n\nBut I'm assuming that we are somehow also sending over the GUC to the\nremote server (which will be v16+) but I am unsure why that would be the\ncase. I'm done code reading for now so I'm leaving this an open\nrequest/question.\n\n, then I suppose\n> > this patch series could do as suggested and enable buffers by default\n> only when\n> > ANALYZE is specified. Then postgres_fdw is not affected, and the\n> > explain_regress GUC is optional: instead, we'd need to specify BUFFERS\n> OFF in\n> > ~100 regression tests which use EXPLAIN ANALYZE.\n\nI'm not following the transition from the prior sentences about COSTS to\nthis one regarding BUFFERS.\n\n (3) still seems useful on its\n> > own.\n>\n\nIt is an orthogonal feature with a different section of our user base being\ncatered to, so I'd agree by default.\n\nPatch Review Comments:\n\n\nThe following change in the machine commit seems out-of-place; are we\nfixing a bug here?\n\nexplain.c:\n /* Show buffer usage in planning */\n- if (bufusage)\n+ if (bufusage && es->buffers)\n\n\n\nTypo (trailing comment // without comment):\n\nExplainPropertyFloat(\"Actual Rows\", NULL, rows, 0, es); //\n\n\nI did a pretty thorough look-through of the locations of the various\nchanges and everything seems consistent with the established code in this\narea (took me a bit to get to the level of comprehension though). Whether\nthere may be something missing I'm less able to judge.\n\n From reading the other main thread I'm not finding this particular choice\nof GUC usage to be a problem anyone has raised and it does seem the\ncleanest solution, both for us and for anyone out there with a similar\ntesting framework.\n\nThe machine output option covers an obvious need since we've already\nwritten ad-hoc parsing code to deal with the problem.\n\nAnd, as someone who sees first-hand the kinds of conversations that occur\nsurrounding beginner's questions, and getting more into performance\nquestions specifically, I'm sympathetic with the desire to have BUFFERS\nsomething that is output by default. Given existing buy-in for that idea\nI'd hope that at minimum the \"update 100 .sql and expected output files,\nthen change the default\" commit happens even if the rest take some further\ndiscussion. Though with even just a bit of attention I don't see any real\nproblems getting the whole thing resolved one way or another.\n\nDavid J.\n\nOn Mon, Jan 24, 2022 at 9:54 AM Justin Pryzby <pryzby@telsasoft.com> wrote:I'm renaming this thread for better visibility, since buffers is a small,\noptional part of the patches I sent.\n\nI made a CF entry here.\nhttps://commitfest.postgresql.org/36/3409/\n\nOn Wed, Dec 01, 2021 at 06:58:20PM -0600, Justin Pryzby wrote:\n> On Mon, Nov 15, 2021 at 01:09:54PM -0600, Justin Pryzby wrote:\n> > Some time ago, I had a few relevant patches:\n> > 1) add explain(REGRESS) which is shorthand for (BUFFERS OFF, TIMING OFF, COSTS OFF, SUMMARY OFF)\n> > 2) add explain(MACHINE) which elides machine-specific output from explain;\n> > for example, Heap Fetches, sort spaceUsed, hash nbuckets, and tidbitmap stuff.\n> > \n> > https://www.postgresql.org/message-id/flat/20200306213310.GM684@telsasoft.com\n> \n> The attached patch series now looks like this (some minor patches are not\n> included in this list):\n> \n> 1. add GUC explain_regress, which disables TIMING, SUMMARY, COSTS; \n> 2. enable explain(BUFFERS) by default (but disabled by explain_regress);+1 \n> 3. Add explain(MACHINE) - which is disabled by explain_regress.\n> This elides various machine-specific output like Memory and Disk use.\n> Maybe it should be called something else like \"QUIET\" or \"VERBOSE_MINUS_1\"\n> or ??INCIDENTALS ?Aside from being 4 syllables and, for me at least, a pain to remember how to spell, it seems to fit.RUNTIME is probably easier, and we just need to point out the difficulty in naming things since actual rows is still shown, and timings have their own independent option.\n> \n> The regression tests now automatically run with explain_regress=on, which is\n> shorthand for TIMING OFF, SUMMARY OFF, COSTS OFF, and then BUFFERS OFF.\n> \n> There's a further option called explain(MACHINE) which can be disabled to hide\n> portions of the output that are unstable, like Memory/Disk/Batches/\n> Heap Fetches/Heap Blocks. This allows \"explain analyze\" to be used more easily\n> in regression tests, and simplifies some existing tests that currently use\n> plpgsql functions to filter the output. But it doesn't handle all the\n> variations from parallel workers.\n> \n> (3) is optional, but simplifies some regression tests. The patch series could\n> be rephrased with (3) first.I like the idea of encapsulating the logic about what to show into the generation code instead of a post-processing routine. I don't see any particular ordering of the three changes being most clear.\n> \n> Unfortunately, \"COSTS OFF\" breaks postgres_fdw remote_estimate. If specifying\n> \"COSTS ON\" in postgres_fdw.c is considered to be a poor fix Given that we have version tests scattered throughout the postgres_fdw.c code I'd say protect it with version >= 9.0 and call it good.But I'm assuming that we are somehow also sending over the GUC to the remote server (which will be v16+) but I am unsure why that would be the case. I'm done code reading for now so I'm leaving this an open request/question., then I suppose\n> this patch series could do as suggested and enable buffers by default only when\n> ANALYZE is specified. Then postgres_fdw is not affected, and the\n> explain_regress GUC is optional: instead, we'd need to specify BUFFERS OFF in\n> ~100 regression tests which use EXPLAIN ANALYZE.I'm not following the transition from the prior sentences about COSTS to this one regarding BUFFERS. (3) still seems useful on its\n> own.It is an orthogonal feature with a different section of our user base being catered to, so I'd agree by default.Patch Review Comments:The following change in the machine commit seems out-of-place; are we fixing a bug here?explain.c: /* Show buffer usage in planning */-\tif (bufusage)+\tif (bufusage && es->buffers)Typo (trailing comment // without comment):ExplainPropertyFloat(\"Actual Rows\", NULL, rows, 0, es); //I did a pretty thorough look-through of the locations of the various changes and everything seems consistent with the established code in this area (took me a bit to get to the level of comprehension though). Whether there may be something missing I'm less able to judge.From reading the other main thread I'm not finding this particular choice of GUC usage to be a problem anyone has raised and it does seem the cleanest solution, both for us and for anyone out there with a similar testing framework.The machine output option covers an obvious need since we've already written ad-hoc parsing code to deal with the problem.And, as someone who sees first-hand the kinds of conversations that occur surrounding beginner's questions, and getting more into performance questions specifically, I'm sympathetic with the desire to have BUFFERS something that is output by default. Given existing buy-in for that idea I'd hope that at minimum the \"update 100 .sql and expected output files, then change the default\" commit happens even if the rest take some further discussion. Though with even just a bit of attention I don't see any real problems getting the whole thing resolved one way or another.David J.","msg_date":"Tue, 26 Jul 2022 15:38:53 -0700","msg_from":"\"David G. Johnston\" <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE), and default to\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"},{"msg_contents":"On Tue, Jul 26, 2022 at 03:38:53PM -0700, David G. Johnston wrote:\n> On Mon, Jan 24, 2022 at 9:54 AM Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n> > > Unfortunately, \"COSTS OFF\" breaks postgres_fdw remote_estimate. If specifying\n> > > \"COSTS ON\" in postgres_fdw.c is considered to be a poor fix , then I suppose\n> > > this patch series could do as suggested and enable buffers by default only when\n> > > ANALYZE is specified. Then postgres_fdw is not affected, and the\n> > > explain_regress GUC is optional: instead, we'd need to specify BUFFERS OFF in\n> > > ~100 regression tests which use EXPLAIN ANALYZE.\n> \n> I'm not following the transition from the prior sentences about COSTS to\n> this one regarding BUFFERS.\n\nIn this patch, SET explain_regress=on changes to defaults to\nexplain(COSTS OFF), so postgres_fdw now explicitly uses COSTS ON. The\npatch also changes default to (BUFFERS OFF), which is what allows\nenabling buffers by default. I think I was just saying that the\nenabling buffers by default depends on somehow disabling it in\nregression tests (preferably not by adding 100 instances of \"BUFFERS\nOFF\").\n\n> The following change in the machine commit seems out-of-place; are we\n> fixing a bug here?\n> \n> explain.c:\n> /* Show buffer usage in planning */\n> - if (bufusage)\n> + if (bufusage && es->buffers)\n\nI think that was an extraneous change, maybe from a conflict while\nre-ordering the patches. Removed it and rebased. Thanks for looking.\n\n-- \nJustin","msg_date":"Mon, 5 Sep 2022 06:20:31 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE), and default to\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"},{"msg_contents":"Rebased.\n\nBTW, I think it may be that the GUC should be marked PGDLLIMPORT ?","msg_date":"Thu, 20 Oct 2022 21:09:34 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE), and default to\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"},{"msg_contents":"On Thu, 2022-10-20 at 21:09 -0500, Justin Pryzby wrote:\n> Rebased.\n\nI had a look at the patch set.\n\nIt applies and builds cleanly and passes the regression tests.\n\n0001: Add GUC: explain_regress\n\n I like the idea of the \"explain_regress\" GUC. That should simplify\n the regression tests.\n\n --- a/src/test/regress/pg_regress.c\n +++ b/src/test/regress/pg_regress.c\n @@ -625,7 +625,7 @@ initialize_environment(void)\n * user's ability to set other variables through that.\n */\n {\n - const char *my_pgoptions = \"-c intervalstyle=postgres_verbose\";\n + const char *my_pgoptions = \"-c intervalstyle=postgres_verbose -c explain_regress=on\";\n const char *old_pgoptions = getenv(\"PGOPTIONS\");\n char *new_pgoptions;\n \n @@ -2288,6 +2288,7 @@ regression_main(int argc, char *argv[],\n fputs(\"log_lock_waits = on\\n\", pg_conf);\n fputs(\"log_temp_files = 128kB\\n\", pg_conf);\n fputs(\"max_prepared_transactions = 2\\n\", pg_conf);\n + // fputs(\"explain_regress = yes\\n\", pg_conf);\n \n for (sl = temp_configs; sl != NULL; sl = sl->next)\n {\n\n This code comment is a leftover and should go.\n\n0002: exercise explain_regress\n\n This is the promised simplification. Looks good.\n\n0003: Make explain default to BUFFERS TRUE\n\n Yes, please!\n This patch is independent from the previous patches.\n I'd like this to be applied, even if the GUC is rejected.\n\n (I understand that that would cause more changes in the regression\n test output if we changed that without introducing \"explain_regress\".)\n\n The patch changes the default value of \"auto_explain.log_buffers\"\n to \"on\", which makes sense. However, the documentation in\n doc/src/sgml/auto-explain.sgml should be updated to reflect that.\n\n --- a/doc/src/sgml/perform.sgml\n +++ b/doc/src/sgml/perform.sgml\n @@ -731,8 +731,8 @@ EXPLAIN ANALYZE SELECT * FROM polygon_tbl WHERE f1 @> polygon '(0.5,2.0)';\n </para>\n \n <para>\n - <command>EXPLAIN</command> has a <literal>BUFFERS</literal> option that can be used with\n - <literal>ANALYZE</literal> to get even more run time statistics:\n + <command>EXPLAIN ANALYZE</command> has a <literal>BUFFERS</literal> option which\n + provides even more run time statistics:\n \n <screen>\n EXPLAIN (ANALYZE, BUFFERS) SELECT * FROM tenk1 WHERE unique1 < 100 AND unique2 > 9000;\n\n This is not enough. The patch would have to update all the examples that use EXPLAIN ANALYZE.\n I see two options:\n\n 1. Change the output of all the examples and move this explanation to the first example\n with EXPLAIN ANALYZE:\n\n The numbers in the <literal>Buffers:</literal> line help to identify which parts\n of the query are the most I/O-intensive.\n\n 2. Change all examples that currently do *not* use BUFFERS to EXPLAIN (BUFFERS OFF).\n Then you could change the example that shows BUFFERS to something like\n\n If you do not suppress the <literal>BUFFERS</literal> option that can be used with\n <command>EXPLAIN (ANALYZE)</command>, you get even more run time statistics:\n\n0004, 0005, 0006, 0007: EXPLAIN (MACHINE)\n\n I think it is confusing that these are included in this patch set.\n EXPLAIN (MACHINE OFF) is similar to \"explain_regress = on\", only it goes even further:\n no query ID, no Heap Fetches, no Sort details, ...\n Why not add this functionality to the GUC?\n\n 0005 suppresses \"rows removed by filter\", but how is that machine dependent?\n\n> BTW, I think it may be that the GUC should be marked PGDLLIMPORT ?\n\nI think it is project policy to apply this mark wherever it is needed. Do you think\nthat third-party extensions might need to use this in C code?\n\nYours,\nLaurenz Albe\n\n\n","msg_date":"Tue, 25 Oct 2022 15:49:14 +0200","msg_from":"Laurenz Albe <laurenz.albe@cybertec.at>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE), and default to\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"},{"msg_contents":"On Tue, Oct 25, 2022 at 03:49:14PM +0200, Laurenz Albe wrote:\n> On Thu, 2022-10-20 at 21:09 -0500, Justin Pryzby wrote:\n> > Rebased.\n> \n> I had a look at the patch set.\n\nThanks for looking\n\n> @@ -2288,6 +2288,7 @@ regression_main(int argc, char *argv[],\n> fputs(\"log_lock_waits = on\\n\", pg_conf);\n> fputs(\"log_temp_files = 128kB\\n\", pg_conf);\n> fputs(\"max_prepared_transactions = 2\\n\", pg_conf);\n> + // fputs(\"explain_regress = yes\\n\", pg_conf);\n> \n> for (sl = temp_configs; sl != NULL; sl = sl->next)\n> {\n> \n> This code comment is a leftover and should go.\n\nNo - I was wondering whether the GUC should be specified by the\nenvironment or by the file; I think it's better by file. Then it also\nneeds to be in Cluster.pm.\n\n> 0004, 0005, 0006, 0007: EXPLAIN (MACHINE)\n> \n> I think it is confusing that these are included in this patch set.\n> EXPLAIN (MACHINE OFF) is similar to \"explain_regress = on\", only it goes even further:\n> no query ID, no Heap Fetches, no Sort details, ...\n> Why not add this functionality to the GUC?\n\nGood question, and one I've asked myself.\n\nexplain_regress affects pre-existing uses of explain in the regression\ntests, aiming to simplify current (or maybe only future) uses. And\nis obviously used to allow enabling BUFFERS by default.\n\nexplain(MACHINE off) is mostly about allowing \"explain ANALYZE\" to be\nused in regression tests. Which, right now, is rare. Maybe I shouldn't\ninclude those patches until the earlier patches progress (or, maybe we\nshould just defer their discussion).\n\n> 0005 suppresses \"rows removed by filter\", but how is that machine dependent?\n\nIt can vary with the number of parallel workers (see 13e8b2ee8), which\nmay not be \"machine dependent\" but the point of that patch is to allow\npredictable output of EXPLAIN ANALYZE. Maybe it needs a better name (or\nmaybe it should be folded into the first patch).\n\nI'm not actually sure if this patch should try to address parallel\nworkers at all, or (if so) if what it's doing now is adequate.\n\n> > BTW, I think it may be that the GUC should be marked PGDLLIMPORT ?\n> \n> I think it is project policy to apply this mark wherever it is needed. Do you think\n> that third-party extensions might need to use this in C code?\n\nSince v15 (8ec569479), the policy is:\n| On Windows, export all the server's global variables using PGDLLIMPORT markers (Robert Haas)\n\nI'm convinced now that's what's intended here.\n\n> This is not enough. The patch would have to update all the examples\n> that use EXPLAIN ANALYZE.\n\nFair enough. I've left a comment to handle this later if the patch\ngains any traction and 001 is progressed.\n\n-- \nJustin","msg_date":"Fri, 28 Oct 2022 17:59:04 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE), and default to\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"},{"msg_contents":"Thanks for the updated patch set!\n\nOn Fri, 2022-10-28 at 17:59 -0500, Justin Pryzby wrote:\n> \n> > 0004, 0005, 0006, 0007: EXPLAIN (MACHINE)\n> > \n> > I think it is confusing that these are included in this patch set.\n> > EXPLAIN (MACHINE OFF) is similar to \"explain_regress = on\", only it goes even further:\n> > no query ID, no Heap Fetches, no Sort details, ...\n> > Why not add this functionality to the GUC?\n> \n> Good question, and one I've asked myself.\n> \n> explain_regress affects pre-existing uses of explain in the regression\n> tests, aiming to simplify current (or maybe only future) uses. And\n> is obviously used to allow enabling BUFFERS by default.\n> \n> explain(MACHINE off) is mostly about allowing \"explain ANALYZE\" to be\n> used in regression tests. Which, right now, is rare. Maybe I shouldn't\n> include those patches until the earlier patches progress (or, maybe we\n> should just defer their discussion).\n\nEssentially, \"explain_regress\" is to simplify the current regression tests,\nand MACHINE OFF is to simplify future regression tests. That does not seem\nlike a meaningful distinction to me. Both are only useful for the regression\ntests, and I see no need for two different knobs.\n\nMy opinion is now like this:\n\n+1 on enabling BUFFERS by default for EXPLAIN (ANALYZE)\n\n+0.2 on \"explain_regress\"\n\n-1 on EXPLAIN (MACHINE) in addition to \"explain_regress\"\n\n-0.1 on adding the MACHINE OFF functionality to \"explain_regress\"\n\n\"explain_regress\" reduces the EXPLAIN options you need for regression tests.\nThis is somewhat useful, but not a big win. Also, it will make backpatching\nregression tests slightly harder for the next 5 years.\n\nHence the +0.2 for \"explain_regress\".\n\nFor me, an important question is: do we really want more EXPLAIN (ANALYZE)\nin the regression tests? It will probably slow the tests down, and I wonder\nif there is a lot of added value, if we have to remove all the machine-specific\noutput anyway.\n\nHence the -0.1.\n\n> > 0005 suppresses \"rows removed by filter\", but how is that machine dependent?\n> \n> It can vary with the number of parallel workers (see 13e8b2ee8), which\n> may not be \"machine dependent\" but the point of that patch is to allow\n> predictable output of EXPLAIN ANALYZE. Maybe it needs a better name (or\n> maybe it should be folded into the first patch).\n\nNow it makes sense to me. I just didn't think of that.\n\n> I'm not actually sure if this patch should try to address parallel\n> workers at all, or (if so) if what it's doing now is adequate.\n> \n> > > BTW, I think it may be that the GUC should be marked PGDLLIMPORT ?\n> > \n> > I think it is project policy to apply this mark wherever it is needed. Do you think\n> > that third-party extensions might need to use this in C code?\n> \n> Since v15 (8ec569479), the policy is:\n> > On Windows, export all the server's global variables using PGDLLIMPORT markers (Robert Haas)\n> \n> I'm convinced now that's what's intended here.\n\nYou convinced me too.\n\n> > This is not enough. The patch would have to update all the examples\n> > that use EXPLAIN ANALYZE.\n> \n> Fair enough. I've left a comment to handle this later if the patch\n> gains any traction and 001 is progressed.\n\nI agree with that.\n\nI would really like to see 0003 go in, but it currently depends on 0001, which\nI am not so sure about.\nI understand that you did that so that \"explain_regress\" can turn off BUFFERS\nand there is no extra churn in the regression tests.\n\nStill, it would be a shame if resistance against \"explain_regress\" would\nbe a show-stopper for 0003.\n\nIf I could get my way, I'd want two separate patches: first, one to turn\nBUFFERS on, and second one for \"explain_regress\" with its current functionality\non top of that.\n\nYours,\nLaurenz Albe\n\n\n","msg_date":"Fri, 04 Nov 2022 11:46:03 +0100","msg_from":"Laurenz Albe <laurenz.albe@cybertec.at>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE), and default to\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"},{"msg_contents":"On Fri, Nov 04, 2022 at 11:46:03AM +0100, Laurenz Albe wrote:\n\n> \"explain_regress\" reduces the EXPLAIN options you need for regression tests.\n> This is somewhat useful, but not a big win. Also, it will make backpatching\n> regression tests slightly harder for the next 5 years.\n\nBut it doesn't cause backpatching to be harder; if a patch is meant to\nbe backpatched, its author can still choose to write the old way, with\n(COSTS OFF, ...)\n\n> For me, an important question is: do we really want more EXPLAIN (ANALYZE)\n> in the regression tests? It will probably slow the tests down, and I wonder\n> if there is a lot of added value, if we have to remove all the machine-specific\n> output anyway.\n\nI don't intend to encourage putting lots of \"explain analyze\" in the\ntests. Rather, this makes that a reasonable possibility rather than\nsomething that people are discouraged from doing by its difficulty.\nUsing \"explain analyze\" still needs to be evaluated (for speed and\notherwise), same as always.\n\nIn any case, I suggested to defer review of patches 004 and beyond,\nwhich are optional, and it distracts from the essential discussion about\n001.\n\n> I would really like to see 0003 go in, but it currently depends on 0001, which\n> I am not so sure about.\n> I understand that you did that so that \"explain_regress\" can turn off BUFFERS\n> and there is no extra churn in the regression tests.\n> \n> Still, it would be a shame if resistance against \"explain_regress\" would\n> be a show-stopper for 0003.\n> \n> If I could get my way, I'd want two separate patches: first, one to turn\n> BUFFERS on, and second one for \"explain_regress\" with its current functionality\n> on top of that.\n\nYou set the patch to \"waiting on author\", which indicates that there's\nno need for further input or review. But, I think that's precisely\nwhat's needed - without input from more people, what could I do to\nprogress the patch ? I don't think it's reasonable to put 001 first and\nchange thousands (actually, 1338) of regression results. If nobody\nwants to discuss 001, then this patch series won't progress.\n\n-- \nJustin\n\n\n","msg_date":"Sat, 5 Nov 2022 08:44:25 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE), and default to\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"},{"msg_contents":"Justin Pryzby <pryzby@telsasoft.com> writes:\n> You set the patch to \"waiting on author\", which indicates that there's\n> no need for further input or review. But, I think that's precisely\n> what's needed - without input from more people, what could I do to\n> progress the patch ? I don't think it's reasonable to put 001 first and\n> change thousands (actually, 1338) of regression results. If nobody\n> wants to discuss 001, then this patch series won't progress.\n\nWell ...\n\n1. 0001 invents a new GUC but provides no documentation for it.\nThat certainly isn't committable, and it's discouraging the\ndiscussion you seek because people have to read the whole patch\nin detail to understand what is being proposed.\n\n2. The same complaint for 0004, which labors under the additional\nproblem that MACHINE is one of the worst option names I've ever\nseen proposed around here. It conveys *nothing* about what it does\nor is good for. The fact that it's got no relationship to the GUC\nname, and yet they're intimately connected, doesn't improve my\nopinion of it.\n\n3. I'm not really on board with the entire approach. Making\nEXPLAIN work significantly differently for developers and test\ncritters than it does for everyone else seems likely to promote\nconfusion and hide bugs. I don't think getting rid of the need\nfor filter functions in test scripts is worth that.\n\n4. I don't see how the current patches could pass under \"make\ninstallcheck\" --- it looks to me like it only takes care of\napplying the GUC change in installations built by pg_regress\nor Cluster.pm. To make this actually work, you'd have to\nadd \"explain_regress = on\" to the ALTER DATABASE SET commands\nissued for regression databases. That'd substantially increase\nthe problem of \"it works differently than what users see\", at\nleast for me --- I do a lot of stuff interactively in the\nregression database.\n\n5. The change in postgres_fdw.c in 0001 concerns me too.\nYeah, it will fix things if an updated postgres_fdw talks to\nan updated server, but what about an older postgres_fdw?\nDon't really want to see that case break; communication with\nservers of different major versions is one of postgres_fdw's\nmain goals.\n\nAs a side note, 0001 also creates a hard break in postgres_fdw's\nability to work with pre-9.0 remote servers, which won't accept\n\"EXPLAIN (COSTS)\". Maybe we don't care about that anymore, given\nthe policy we've adopted elsewhere that we won't test or support\ninteroperability with versions before 9.2. But we should either\nmake the command spelling conditional on remote version, or\nofficially move that goalpost.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 05 Nov 2022 10:43:07 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE),\n and default to explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN\n (ANALYZE))"},{"msg_contents":"On Sat, Nov 05, 2022 at 10:43:07AM -0400, Tom Lane wrote:\n> Justin Pryzby <pryzby@telsasoft.com> writes:\n> > You set the patch to \"waiting on author\", which indicates that there's\n> > no need for further input or review. But, I think that's precisely\n> > what's needed - without input from more people, what could I do to\n> > progress the patch ? I don't think it's reasonable to put 001 first and\n> > change thousands (actually, 1338) of regression results. If nobody\n> > wants to discuss 001, then this patch series won't progress.\n> \n> Well ...\n> \n> 1. 0001 invents a new GUC but provides no documentation for it.\n> That certainly isn't committable, and it's discouraging the\n> discussion you seek because people have to read the whole patch\n> in detail to understand what is being proposed.\n> \n> 2. The same complaint for 0004, which labors under the additional\n\nI suggested that the discussion be limited to the early patches (004 is\nan optional, possible idea that I had, but it's evidently still causing\nconfusion).\n\n> 3. I'm not really on board with the entire approach. Making\n> EXPLAIN work significantly differently for developers and test\n> critters than it does for everyone else seems likely to promote\n> confusion and hide bugs. I don't think getting rid of the need\n> for filter functions in test scripts is worth that.\n\nI'm open to suggestions how else to move towards the goal.\n\nNote that there's a few other things which have vaguely-similar\nbehavior:\n\nHIDE_TABLEAM=on\nHIDE_TOAST_COMPRESSION=on\ncompute_query_id = regress\n\nAnother possibility is to make a new \"explain\" format, like\n| explain(FORMAT REGRESS, ...).\n\n...but I don't see how that's would be less objectionable than what I've\nwritten.\n\nThe patch record should probably be closed until someone proposes\nanother way to implement what's necessary to enable explain(BUFFERS) by\ndefault.\n\n-- \nJustin\n\n\n","msg_date":"Thu, 1 Dec 2022 12:08:56 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: explain_regress, explain(MACHINE), and default to\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))"}],"string":"[\n {\n \"msg_contents\": \"Re-reading the thread [1] (cannot answer there – don't have those emails in\\nmy box anymore), I see that there was strong support for enabling BUFFERS\\nin EXPLAIN ANALYZE by default. And there were patches. Commitfest entry [2]\\nwas marked Rejected because there were questions to the implementation\\nbased on GUCs.\\n\\nAttached is a simple patch enabling BUFFERS by default, without involving\\nGUCs.\\n\\nWhy it is important?\\n\\nIn many cases, people forget about the BUFFERS option in EXPLAIN ANALYZE\\nand share execution plans without it – sending it via regular communication\\nchannels for work or publishing to visualization systems. Meanwhile, the\\noption has a lower overhead compared to TIMING (enabled by default for\\nEXPLAIN ANALYZE) and it is extremely useful for query\\noptimization. This patch doesn't enable BUFFERS for EXPLAIN executed\\nwithout ANALYZE.\\n\\nOpen questions:\\n\\n1. Should BUFFERS be enabled for regular (w/o ANALYZE) EXPLAIN? Now it may\\nmake sense because of the buffer numbers the planner uses. This patch\\ndoesn't do that, but it definitely may make sense because it can help\\npeople understand why planning time is so big, in some cases.\\n\\n2. How to adjust documentation? Should EXPLAIN ANALYZE examples be adjusted\\nto use BUFFERS OFF (easier change) or show some example buffer numbers –\\nlike it is done for timing and cost numbers? I tend to think that the\\nlatter makes more sense.\\n\\n3. How to adjust regression tests? Of course, now many tests fail. Same\\nquestion as for documentation. Excluding buffer, numbers would be an easier\\nfix, of course – but at least some tests should\\ncheck the behavior of BUFFERS (such as both default options – with and\\nwithout ANALYZE).\\nOn any given platform, the buffer numbers are pretty stable, so we could\\nrely on it, but I'm not sure about all possible options being tested and\\nwould appreciate advice here (of course, if the patch makes it thru the\\ndiscussion in general).\\n\\n\\n## Links\\n[1]\\nhttps://www.postgresql.org/message-id/flat/b3197ba8-225f-f53c-326d-5b1756c77c3e%40postgresfriends.org\\n[2] https://commitfest.postgresql.org/28/2567/\",\n \"msg_date\": \"Fri, 12 Nov 2021 14:58:07 -0800\",\n \"msg_from\": \"Nikolay Samokhvalov <samokhvalov@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"BUFFERS enabled by default in EXPLAIN (ANALYZE)\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 11/12/21 23:58, Nikolay Samokhvalov wrote:\\n> Re-reading the thread [1] (cannot answer there – don't have those emails \\n> in my box anymore),\\n\\nYou can download the message as mbox and import it into your client \\n(pretty much any client supports that, I think).\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 01:18:31 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)\"\n },\n {\n \"msg_contents\": \"On 11/12/21 11:58 PM, Nikolay Samokhvalov wrote:\\n> Re-reading the thread [1] (cannot answer there – don't have those emails in\\n> my box anymore), I see that there was strong support for enabling BUFFERS\\n> in EXPLAIN ANALYZE by default. And there were patches. Commitfest entry [2]\\n> was marked Rejected because there were questions to the implementation\\n> based on GUCs.\\n> \\n> Attached is a simple patch enabling BUFFERS by default, without involving\\n> GUCs.\\n\\nI obviously agree with this (although I still wish we had gucs as we\\nkeep adding more and more options to EXPLAIN).\\n\\nThe patch looks good to me, too.\\n\\n+1\\n-- \\nVik Fearing\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 02:34:05 +0100\",\n \"msg_from\": \"Vik Fearing <vik@postgresfriends.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 13, 2021 at 8:18 AM Tomas Vondra\\n<tomas.vondra@enterprisedb.com> wrote:\\n>\\n> On 11/12/21 23:58, Nikolay Samokhvalov wrote:\\n> > Re-reading the thread [1] (cannot answer there – don't have those emails\\n> > in my box anymore),\\n>\\n> You can download the message as mbox and import it into your client\\n> (pretty much any client supports that, I think).\\n\\nThere is also a \\\"resend mail\\\" link available for each mail in the\\narchive which should be even more straightforward.\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 12:17:58 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 12, 2021 at 02:58:07PM -0800, Nikolay Samokhvalov wrote:\\n> Re-reading the thread [1] (cannot answer there – don't have those emails in\\n> my box anymore), I see that there was strong support for enabling BUFFERS\\n> in EXPLAIN ANALYZE by default. And there were patches. Commitfest entry [2]\\n> was marked Rejected because there were questions to the implementation\\n> based on GUCs.\\n\\nI think the only reason this isn't already the default is because of (3):\\n\\n> 3. How to adjust regression tests? Of course, now many tests fail. Same\\n> question as for documentation. Excluding buffer, numbers would be an easier\\n> fix, of course – but at least some tests should\\n> check the behavior of BUFFERS (such as both default options – with and\\n> without ANALYZE).\\n\\nSome time ago, I had a few relevant patches:\\n1) add explain(REGRESS) which is shorthand for (BUFFERS OFF, TIMING OFF, COSTS OFF, SUMMARY OFF)\\n2) add explain(MACHINE) which elides machine-specific output from explain;\\n for example, Heap Fetches, sort spaceUsed, hash nbuckets, and tidbitmap stuff.\\n\\nhttps://www.postgresql.org/message-id/flat/20200306213310.GM684@telsasoft.com\\n\\n> 1. Should BUFFERS be enabled for regular (w/o ANALYZE) EXPLAIN? Now it may\\n> make sense because of the buffer numbers the planner uses. This patch\\n> doesn't do that, but it definitely may make sense because it can help\\n> people understand why planning time is so big, in some cases.\\n\\nI think it *should* be enabled for planning, since that makes the default\\neasier to understand and document, and it makes a user's use of \\\"explain\\\"\\neasier.\\n\\nHowever, that makes this patch series more complicated: explain(ANALYZE) is\\nrare in the regression tests, but this would affect the output of bare\\n\\\"explain\\\", which affects almost every test.\\n\\nI think the answer to that is to add a GUC (as Tom suggested in an old thread)\\nlike explain_regress=on, which causes explain to omit these details. This\\nwould be instead of my explain(REGRESS), and would change the defaults of\\nvarious params in a central place, to avoid the need to update many regression\\ntests.\\n\\nIn the future, this might also handle the stuff that my \\\"explain(MACHINE)\\\"\\nattempted to handle.\\n\\nI've rebased my patches like this. I think the explain(REGRESS) should be\\nre-written as a GUC which is set by regress.c. I'm interested to hear from a\\nreviewer if this is worth pursing like this.\\n\\n-- \\nJustin\",\n \"msg_date\": \"Mon, 15 Nov 2021 13:09:54 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)\"\n },\n {\n \"msg_contents\": \"I think it *should* be enabled for planning, since that makes the default\\n> easier to understand and document, and it makes a user's use of \\\"explain\\\"\\n> easier.\\n\\n\\nI’d be keen to see BUFFERS off by default with EXPLAIN, and on by default\\nwith EXPLAIN ANALYZE.\\n\\nThe SUMMARY flag was implemented that way, which I think has been easy\\nenough for folks to understand and document. In fact, I think the only\\nBUFFERS information goes in the “summary” section for EXPLAIN (BUFFERS), so\\nmaybe it makes perfect sense? If it would be great if that made\\nimplementation easier, too.\\n\\nIn any case, thank you all, I’m so glad that this is being discussed again.\\nIt’d be so good to start seeing buffers in more plans.\\n\\n—\\nMichael\\n\\nI think it *should* be enabled for planning, since that makes the default\\neasier to understand and document, and it makes a user's use of \\\"explain\\\"\\neasier.I’d be keen to see BUFFERS off by default with EXPLAIN, and on by default with EXPLAIN ANALYZE.The SUMMARY flag was implemented that way, which I think has been easy enough for folks to understand and document. In fact, I think the only BUFFERS information goes in the “summary” section for EXPLAIN (BUFFERS), so maybe it makes perfect sense? If it would be great if that made implementation easier, too.In any case, thank you all, I’m so glad that this is being discussed again. It’d be so good to start seeing buffers in more plans.—Michael\",\n \"msg_date\": \"Wed, 24 Nov 2021 18:35:45 +0000\",\n \"msg_from\": \"Michael Christofides <michael@pgmustard.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 01:09:54PM -0600, Justin Pryzby wrote:\\n> Some time ago, I had a few relevant patches:\\n> 1) add explain(REGRESS) which is shorthand for (BUFFERS OFF, TIMING OFF, COSTS OFF, SUMMARY OFF)\\n> 2) add explain(MACHINE) which elides machine-specific output from explain;\\n> for example, Heap Fetches, sort spaceUsed, hash nbuckets, and tidbitmap stuff.\\n> \\n> https://www.postgresql.org/message-id/flat/20200306213310.GM684@telsasoft.com\\n\\nThe attached patch series now looks like this (some minor patches are not\\nincluded in this list):\\n\\n 1. add GUC explain_regress, which disables TIMING, SUMMARY, COSTS;\\n 2. enable explain(BUFFERS) by default (but disabled by explain_regress);\\n 3. Add explain(MACHINE) - which is disabled by explain_regress.\\n This elides various machine-specific output like Memory and Disk use.\\n Maybe it should be called something else like \\\"QUIET\\\" or \\\"VERBOSE_MINUS_1\\\"\\n or ??\\n\\nThe regression tests now automatically run with explain_regress=on, which is\\nshorthand for TIMING OFF, SUMMARY OFF, COSTS OFF, and then BUFFERS OFF.\\n\\nThere's a further option called explain(MACHINE) which can be disabled to hide\\nportions of the output that are unstable, like Memory/Disk/Batches/\\nHeap Fetches/Heap Blocks. This allows \\\"explain analyze\\\" to be used more easily\\nin regression tests, and simplifies some existing tests that currently use\\nplpgsql functions to filter the output. But it doesn't handle all the\\nvariations from parallel workers.\\n\\n(3) is optional, but simplifies some regression tests. The patch series could\\nbe rephrased with (3) first.\\n\\nUnfortunately, \\\"COSTS OFF\\\" breaks postgres_fdw remote_estimate. If specifying\\n\\\"COSTS ON\\\" in postgres_fdw.c is considered to be a poor fix , then I suppose\\nthis patch series could do as suggested and enable buffers by default only when\\nANALYZE is specified. Then postgres_fdw is not affected, and the\\nexplain_regress GUC is optional: instead, we'd need to specify BUFFERS OFF in\\n~100 regression tests which use EXPLAIN ANALYZE. (3) still seems useful on its\\nown.\",\n \"msg_date\": \"Wed, 1 Dec 2021 18:58:20 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Wed, Dec 01, 2021 at 06:58:20PM -0600, Justin Pryzby wrote:\\n> \\n> The attached patch series now looks like this (some minor patches are not\\n> included in this list):\\n\\nThis version of the patchset doesn't apply anymore:\\n\\nhttp://cfbot.cputube.org/patch_36_3409.log\\n=== Applying patches on top of PostgreSQL commit ID 4483b2cf29bfe8091b721756928ccbe31c5c8e14 ===\\n=== applying patch ./0003-Make-explain-default-to-BUFFERS-TRUE.patch\\npatching file doc/src/sgml/config.sgml\\nHunk #1 FAILED at 7615.\\n1 out of 1 hunk FAILED -- saving rejects to file doc/src/sgml/config.sgml.rej\\n\\nBy the way, could you add versionning to the patchsets you send, e.g. git\\nformat-patch -vXXX. It would make it easier to know which version are used or\\ndiscussed.\\n\\nI'm switching the CF entry to Waiting on Author.\\n\\n\\n\",\n \"msg_date\": \"Tue, 18 Jan 2022 14:53:09 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: BUFFERS enabled by default in EXPLAIN (ANALYZE)\"\n },\n {\n \"msg_contents\": \"I'm renaming this thread for better visibility, since buffers is a small,\\noptional part of the patches I sent.\\n\\nI made a CF entry here.\\nhttps://commitfest.postgresql.org/36/3409/\\n\\nOn Wed, Dec 01, 2021 at 06:58:20PM -0600, Justin Pryzby wrote:\\n> On Mon, Nov 15, 2021 at 01:09:54PM -0600, Justin Pryzby wrote:\\n> > Some time ago, I had a few relevant patches:\\n> > 1) add explain(REGRESS) which is shorthand for (BUFFERS OFF, TIMING OFF, COSTS OFF, SUMMARY OFF)\\n> > 2) add explain(MACHINE) which elides machine-specific output from explain;\\n> > for example, Heap Fetches, sort spaceUsed, hash nbuckets, and tidbitmap stuff.\\n> > \\n> > https://www.postgresql.org/message-id/flat/20200306213310.GM684@telsasoft.com\\n> \\n> The attached patch series now looks like this (some minor patches are not\\n> included in this list):\\n> \\n> 1. add GUC explain_regress, which disables TIMING, SUMMARY, COSTS;\\n> 2. enable explain(BUFFERS) by default (but disabled by explain_regress);\\n> 3. Add explain(MACHINE) - which is disabled by explain_regress.\\n> This elides various machine-specific output like Memory and Disk use.\\n> Maybe it should be called something else like \\\"QUIET\\\" or \\\"VERBOSE_MINUS_1\\\"\\n> or ??\\n> \\n> The regression tests now automatically run with explain_regress=on, which is\\n> shorthand for TIMING OFF, SUMMARY OFF, COSTS OFF, and then BUFFERS OFF.\\n> \\n> There's a further option called explain(MACHINE) which can be disabled to hide\\n> portions of the output that are unstable, like Memory/Disk/Batches/\\n> Heap Fetches/Heap Blocks. This allows \\\"explain analyze\\\" to be used more easily\\n> in regression tests, and simplifies some existing tests that currently use\\n> plpgsql functions to filter the output. But it doesn't handle all the\\n> variations from parallel workers.\\n> \\n> (3) is optional, but simplifies some regression tests. The patch series could\\n> be rephrased with (3) first.\\n> \\n> Unfortunately, \\\"COSTS OFF\\\" breaks postgres_fdw remote_estimate. If specifying\\n> \\\"COSTS ON\\\" in postgres_fdw.c is considered to be a poor fix , then I suppose\\n> this patch series could do as suggested and enable buffers by default only when\\n> ANALYZE is specified. Then postgres_fdw is not affected, and the\\n> explain_regress GUC is optional: instead, we'd need to specify BUFFERS OFF in\\n> ~100 regression tests which use EXPLAIN ANALYZE. (3) still seems useful on its\\n> own.\",\n \"msg_date\": \"Mon, 24 Jan 2022 11:35:48 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"explain_regress, explain(MACHINE), and default to explain(BUFFERS)\\n (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n },\n {\n \"msg_contents\": \"Rebased over ebf6c5249b7db525e59563fb149642665c88f747.\\nIt looks like that patch handles only query_id, and this patch also tries to\\nhandle a bunch of other stuff.\\n\\nIf it's helpful, feel free to kick this patch to a future CF.\",\n \"msg_date\": \"Sat, 26 Feb 2022 15:07:20 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE), and default to\\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 03:07:20PM -0600, Justin Pryzby wrote:\\n> Rebased over ebf6c5249b7db525e59563fb149642665c88f747.\\n> It looks like that patch handles only query_id, and this patch also tries to\\n> handle a bunch of other stuff.\\n> \\n> If it's helpful, feel free to kick this patch to a future CF.\\n\\nRebased over MERGE.\",\n \"msg_date\": \"Wed, 30 Mar 2022 14:21:05 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE), and default to\\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n },\n {\n \"msg_contents\": \"@cfbot: rebased\",\n \"msg_date\": \"Thu, 7 Jul 2022 01:49:04 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE), and default to\\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 24, 2022 at 9:54 AM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n> I'm renaming this thread for better visibility, since buffers is a small,\\n> optional part of the patches I sent.\\n>\\n> I made a CF entry here.\\n> https://commitfest.postgresql.org/36/3409/\\n>\\n> On Wed, Dec 01, 2021 at 06:58:20PM -0600, Justin Pryzby wrote:\\n> > On Mon, Nov 15, 2021 at 01:09:54PM -0600, Justin Pryzby wrote:\\n> > > Some time ago, I had a few relevant patches:\\n> > > 1) add explain(REGRESS) which is shorthand for (BUFFERS OFF, TIMING\\n> OFF, COSTS OFF, SUMMARY OFF)\\n> > > 2) add explain(MACHINE) which elides machine-specific output from\\n> explain;\\n> > > for example, Heap Fetches, sort spaceUsed, hash nbuckets, and\\n> tidbitmap stuff.\\n> > >\\n> > >\\n> https://www.postgresql.org/message-id/flat/20200306213310.GM684@telsasoft.com\\n> >\\n> > The attached patch series now looks like this (some minor patches are not\\n> > included in this list):\\n> >\\n> > 1. add GUC explain_regress, which disables TIMING, SUMMARY, COSTS;\\n>\\n\\n\\n> > 2. enable explain(BUFFERS) by default (but disabled by explain_regress);\\n>\\n\\n+1\\n\\n\\n> > 3. Add explain(MACHINE) - which is disabled by explain_regress.\\n> > This elides various machine-specific output like Memory and Disk use.\\n> > Maybe it should be called something else like \\\"QUIET\\\" or\\n> \\\"VERBOSE_MINUS_1\\\"\\n> > or ??\\n>\\n\\nINCIDENTALS ?\\n\\nAside from being 4 syllables and, for me at least, a pain to remember how\\nto spell, it seems to fit.\\n\\nRUNTIME is probably easier, and we just need to point out the difficulty in\\nnaming things since actual rows is still shown, and timings have their own\\nindependent option.\\n\\n>\\n> > The regression tests now automatically run with explain_regress=on,\\n> which is\\n> > shorthand for TIMING OFF, SUMMARY OFF, COSTS OFF, and then BUFFERS OFF.\\n> >\\n> > There's a further option called explain(MACHINE) which can be disabled\\n> to hide\\n> > portions of the output that are unstable, like Memory/Disk/Batches/\\n> > Heap Fetches/Heap Blocks. This allows \\\"explain analyze\\\" to be used more\\n> easily\\n> > in regression tests, and simplifies some existing tests that currently\\n> use\\n> > plpgsql functions to filter the output. But it doesn't handle all the\\n> > variations from parallel workers.\\n> >\\n> > (3) is optional, but simplifies some regression tests. The patch series\\n> could\\n> > be rephrased with (3) first.\\n>\\n\\nI like the idea of encapsulating the logic about what to show into the\\ngeneration code instead of a post-processing routine.\\n\\nI don't see any particular ordering of the three changes being most clear.\\n\\n>\\n> > Unfortunately, \\\"COSTS OFF\\\" breaks postgres_fdw remote_estimate. If\\n> specifying\\n> > \\\"COSTS ON\\\" in postgres_fdw.c is considered to be a poor fix\\n\\n\\nGiven that we have version tests scattered throughout the postgres_fdw.c\\ncode I'd say protect it with version >= 9.0 and call it good.\\n\\nBut I'm assuming that we are somehow also sending over the GUC to the\\nremote server (which will be v16+) but I am unsure why that would be the\\ncase. I'm done code reading for now so I'm leaving this an open\\nrequest/question.\\n\\n, then I suppose\\n> > this patch series could do as suggested and enable buffers by default\\n> only when\\n> > ANALYZE is specified. Then postgres_fdw is not affected, and the\\n> > explain_regress GUC is optional: instead, we'd need to specify BUFFERS\\n> OFF in\\n> > ~100 regression tests which use EXPLAIN ANALYZE.\\n\\nI'm not following the transition from the prior sentences about COSTS to\\nthis one regarding BUFFERS.\\n\\n (3) still seems useful on its\\n> > own.\\n>\\n\\nIt is an orthogonal feature with a different section of our user base being\\ncatered to, so I'd agree by default.\\n\\nPatch Review Comments:\\n\\n\\nThe following change in the machine commit seems out-of-place; are we\\nfixing a bug here?\\n\\nexplain.c:\\n /* Show buffer usage in planning */\\n- if (bufusage)\\n+ if (bufusage && es->buffers)\\n\\n\\n\\nTypo (trailing comment // without comment):\\n\\nExplainPropertyFloat(\\\"Actual Rows\\\", NULL, rows, 0, es); //\\n\\n\\nI did a pretty thorough look-through of the locations of the various\\nchanges and everything seems consistent with the established code in this\\narea (took me a bit to get to the level of comprehension though). Whether\\nthere may be something missing I'm less able to judge.\\n\\n From reading the other main thread I'm not finding this particular choice\\nof GUC usage to be a problem anyone has raised and it does seem the\\ncleanest solution, both for us and for anyone out there with a similar\\ntesting framework.\\n\\nThe machine output option covers an obvious need since we've already\\nwritten ad-hoc parsing code to deal with the problem.\\n\\nAnd, as someone who sees first-hand the kinds of conversations that occur\\nsurrounding beginner's questions, and getting more into performance\\nquestions specifically, I'm sympathetic with the desire to have BUFFERS\\nsomething that is output by default. Given existing buy-in for that idea\\nI'd hope that at minimum the \\\"update 100 .sql and expected output files,\\nthen change the default\\\" commit happens even if the rest take some further\\ndiscussion. Though with even just a bit of attention I don't see any real\\nproblems getting the whole thing resolved one way or another.\\n\\nDavid J.\\n\\nOn Mon, Jan 24, 2022 at 9:54 AM Justin Pryzby <pryzby@telsasoft.com> wrote:I'm renaming this thread for better visibility, since buffers is a small,\\noptional part of the patches I sent.\\n\\nI made a CF entry here.\\nhttps://commitfest.postgresql.org/36/3409/\\n\\nOn Wed, Dec 01, 2021 at 06:58:20PM -0600, Justin Pryzby wrote:\\n> On Mon, Nov 15, 2021 at 01:09:54PM -0600, Justin Pryzby wrote:\\n> > Some time ago, I had a few relevant patches:\\n> > 1) add explain(REGRESS) which is shorthand for (BUFFERS OFF, TIMING OFF, COSTS OFF, SUMMARY OFF)\\n> > 2) add explain(MACHINE) which elides machine-specific output from explain;\\n> > for example, Heap Fetches, sort spaceUsed, hash nbuckets, and tidbitmap stuff.\\n> > \\n> > https://www.postgresql.org/message-id/flat/20200306213310.GM684@telsasoft.com\\n> \\n> The attached patch series now looks like this (some minor patches are not\\n> included in this list):\\n> \\n> 1. add GUC explain_regress, which disables TIMING, SUMMARY, COSTS; \\n> 2. enable explain(BUFFERS) by default (but disabled by explain_regress);+1 \\n> 3. Add explain(MACHINE) - which is disabled by explain_regress.\\n> This elides various machine-specific output like Memory and Disk use.\\n> Maybe it should be called something else like \\\"QUIET\\\" or \\\"VERBOSE_MINUS_1\\\"\\n> or ??INCIDENTALS ?Aside from being 4 syllables and, for me at least, a pain to remember how to spell, it seems to fit.RUNTIME is probably easier, and we just need to point out the difficulty in naming things since actual rows is still shown, and timings have their own independent option.\\n> \\n> The regression tests now automatically run with explain_regress=on, which is\\n> shorthand for TIMING OFF, SUMMARY OFF, COSTS OFF, and then BUFFERS OFF.\\n> \\n> There's a further option called explain(MACHINE) which can be disabled to hide\\n> portions of the output that are unstable, like Memory/Disk/Batches/\\n> Heap Fetches/Heap Blocks. This allows \\\"explain analyze\\\" to be used more easily\\n> in regression tests, and simplifies some existing tests that currently use\\n> plpgsql functions to filter the output. But it doesn't handle all the\\n> variations from parallel workers.\\n> \\n> (3) is optional, but simplifies some regression tests. The patch series could\\n> be rephrased with (3) first.I like the idea of encapsulating the logic about what to show into the generation code instead of a post-processing routine. I don't see any particular ordering of the three changes being most clear.\\n> \\n> Unfortunately, \\\"COSTS OFF\\\" breaks postgres_fdw remote_estimate. If specifying\\n> \\\"COSTS ON\\\" in postgres_fdw.c is considered to be a poor fix Given that we have version tests scattered throughout the postgres_fdw.c code I'd say protect it with version >= 9.0 and call it good.But I'm assuming that we are somehow also sending over the GUC to the remote server (which will be v16+) but I am unsure why that would be the case. I'm done code reading for now so I'm leaving this an open request/question., then I suppose\\n> this patch series could do as suggested and enable buffers by default only when\\n> ANALYZE is specified. Then postgres_fdw is not affected, and the\\n> explain_regress GUC is optional: instead, we'd need to specify BUFFERS OFF in\\n> ~100 regression tests which use EXPLAIN ANALYZE.I'm not following the transition from the prior sentences about COSTS to this one regarding BUFFERS. (3) still seems useful on its\\n> own.It is an orthogonal feature with a different section of our user base being catered to, so I'd agree by default.Patch Review Comments:The following change in the machine commit seems out-of-place; are we fixing a bug here?explain.c: /* Show buffer usage in planning */-\\tif (bufusage)+\\tif (bufusage && es->buffers)Typo (trailing comment // without comment):ExplainPropertyFloat(\\\"Actual Rows\\\", NULL, rows, 0, es); //I did a pretty thorough look-through of the locations of the various changes and everything seems consistent with the established code in this area (took me a bit to get to the level of comprehension though). Whether there may be something missing I'm less able to judge.From reading the other main thread I'm not finding this particular choice of GUC usage to be a problem anyone has raised and it does seem the cleanest solution, both for us and for anyone out there with a similar testing framework.The machine output option covers an obvious need since we've already written ad-hoc parsing code to deal with the problem.And, as someone who sees first-hand the kinds of conversations that occur surrounding beginner's questions, and getting more into performance questions specifically, I'm sympathetic with the desire to have BUFFERS something that is output by default. Given existing buy-in for that idea I'd hope that at minimum the \\\"update 100 .sql and expected output files, then change the default\\\" commit happens even if the rest take some further discussion. Though with even just a bit of attention I don't see any real problems getting the whole thing resolved one way or another.David J.\",\n \"msg_date\": \"Tue, 26 Jul 2022 15:38:53 -0700\",\n \"msg_from\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE), and default to\\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 03:38:53PM -0700, David G. Johnston wrote:\\n> On Mon, Jan 24, 2022 at 9:54 AM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n> > > Unfortunately, \\\"COSTS OFF\\\" breaks postgres_fdw remote_estimate. If specifying\\n> > > \\\"COSTS ON\\\" in postgres_fdw.c is considered to be a poor fix , then I suppose\\n> > > this patch series could do as suggested and enable buffers by default only when\\n> > > ANALYZE is specified. Then postgres_fdw is not affected, and the\\n> > > explain_regress GUC is optional: instead, we'd need to specify BUFFERS OFF in\\n> > > ~100 regression tests which use EXPLAIN ANALYZE.\\n> \\n> I'm not following the transition from the prior sentences about COSTS to\\n> this one regarding BUFFERS.\\n\\nIn this patch, SET explain_regress=on changes to defaults to\\nexplain(COSTS OFF), so postgres_fdw now explicitly uses COSTS ON. The\\npatch also changes default to (BUFFERS OFF), which is what allows\\nenabling buffers by default. I think I was just saying that the\\nenabling buffers by default depends on somehow disabling it in\\nregression tests (preferably not by adding 100 instances of \\\"BUFFERS\\nOFF\\\").\\n\\n> The following change in the machine commit seems out-of-place; are we\\n> fixing a bug here?\\n> \\n> explain.c:\\n> /* Show buffer usage in planning */\\n> - if (bufusage)\\n> + if (bufusage && es->buffers)\\n\\nI think that was an extraneous change, maybe from a conflict while\\nre-ordering the patches. Removed it and rebased. Thanks for looking.\\n\\n-- \\nJustin\",\n \"msg_date\": \"Mon, 5 Sep 2022 06:20:31 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE), and default to\\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n },\n {\n \"msg_contents\": \"Rebased.\\n\\nBTW, I think it may be that the GUC should be marked PGDLLIMPORT ?\",\n \"msg_date\": \"Thu, 20 Oct 2022 21:09:34 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE), and default to\\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n },\n {\n \"msg_contents\": \"On Thu, 2022-10-20 at 21:09 -0500, Justin Pryzby wrote:\\n> Rebased.\\n\\nI had a look at the patch set.\\n\\nIt applies and builds cleanly and passes the regression tests.\\n\\n0001: Add GUC: explain_regress\\n\\n I like the idea of the \\\"explain_regress\\\" GUC. That should simplify\\n the regression tests.\\n\\n --- a/src/test/regress/pg_regress.c\\n +++ b/src/test/regress/pg_regress.c\\n @@ -625,7 +625,7 @@ initialize_environment(void)\\n * user's ability to set other variables through that.\\n */\\n {\\n - const char *my_pgoptions = \\\"-c intervalstyle=postgres_verbose\\\";\\n + const char *my_pgoptions = \\\"-c intervalstyle=postgres_verbose -c explain_regress=on\\\";\\n const char *old_pgoptions = getenv(\\\"PGOPTIONS\\\");\\n char *new_pgoptions;\\n \\n @@ -2288,6 +2288,7 @@ regression_main(int argc, char *argv[],\\n fputs(\\\"log_lock_waits = on\\\\n\\\", pg_conf);\\n fputs(\\\"log_temp_files = 128kB\\\\n\\\", pg_conf);\\n fputs(\\\"max_prepared_transactions = 2\\\\n\\\", pg_conf);\\n + // fputs(\\\"explain_regress = yes\\\\n\\\", pg_conf);\\n \\n for (sl = temp_configs; sl != NULL; sl = sl->next)\\n {\\n\\n This code comment is a leftover and should go.\\n\\n0002: exercise explain_regress\\n\\n This is the promised simplification. Looks good.\\n\\n0003: Make explain default to BUFFERS TRUE\\n\\n Yes, please!\\n This patch is independent from the previous patches.\\n I'd like this to be applied, even if the GUC is rejected.\\n\\n (I understand that that would cause more changes in the regression\\n test output if we changed that without introducing \\\"explain_regress\\\".)\\n\\n The patch changes the default value of \\\"auto_explain.log_buffers\\\"\\n to \\\"on\\\", which makes sense. However, the documentation in\\n doc/src/sgml/auto-explain.sgml should be updated to reflect that.\\n\\n --- a/doc/src/sgml/perform.sgml\\n +++ b/doc/src/sgml/perform.sgml\\n @@ -731,8 +731,8 @@ EXPLAIN ANALYZE SELECT * FROM polygon_tbl WHERE f1 @> polygon '(0.5,2.0)';\\n </para>\\n \\n <para>\\n - <command>EXPLAIN</command> has a <literal>BUFFERS</literal> option that can be used with\\n - <literal>ANALYZE</literal> to get even more run time statistics:\\n + <command>EXPLAIN ANALYZE</command> has a <literal>BUFFERS</literal> option which\\n + provides even more run time statistics:\\n \\n <screen>\\n EXPLAIN (ANALYZE, BUFFERS) SELECT * FROM tenk1 WHERE unique1 < 100 AND unique2 > 9000;\\n\\n This is not enough. The patch would have to update all the examples that use EXPLAIN ANALYZE.\\n I see two options:\\n\\n 1. Change the output of all the examples and move this explanation to the first example\\n with EXPLAIN ANALYZE:\\n\\n The numbers in the <literal>Buffers:</literal> line help to identify which parts\\n of the query are the most I/O-intensive.\\n\\n 2. Change all examples that currently do *not* use BUFFERS to EXPLAIN (BUFFERS OFF).\\n Then you could change the example that shows BUFFERS to something like\\n\\n If you do not suppress the <literal>BUFFERS</literal> option that can be used with\\n <command>EXPLAIN (ANALYZE)</command>, you get even more run time statistics:\\n\\n0004, 0005, 0006, 0007: EXPLAIN (MACHINE)\\n\\n I think it is confusing that these are included in this patch set.\\n EXPLAIN (MACHINE OFF) is similar to \\\"explain_regress = on\\\", only it goes even further:\\n no query ID, no Heap Fetches, no Sort details, ...\\n Why not add this functionality to the GUC?\\n\\n 0005 suppresses \\\"rows removed by filter\\\", but how is that machine dependent?\\n\\n> BTW, I think it may be that the GUC should be marked PGDLLIMPORT ?\\n\\nI think it is project policy to apply this mark wherever it is needed. Do you think\\nthat third-party extensions might need to use this in C code?\\n\\nYours,\\nLaurenz Albe\\n\\n\\n\",\n \"msg_date\": \"Tue, 25 Oct 2022 15:49:14 +0200\",\n \"msg_from\": \"Laurenz Albe <laurenz.albe@cybertec.at>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE), and default to\\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 25, 2022 at 03:49:14PM +0200, Laurenz Albe wrote:\\n> On Thu, 2022-10-20 at 21:09 -0500, Justin Pryzby wrote:\\n> > Rebased.\\n> \\n> I had a look at the patch set.\\n\\nThanks for looking\\n\\n> @@ -2288,6 +2288,7 @@ regression_main(int argc, char *argv[],\\n> fputs(\\\"log_lock_waits = on\\\\n\\\", pg_conf);\\n> fputs(\\\"log_temp_files = 128kB\\\\n\\\", pg_conf);\\n> fputs(\\\"max_prepared_transactions = 2\\\\n\\\", pg_conf);\\n> + // fputs(\\\"explain_regress = yes\\\\n\\\", pg_conf);\\n> \\n> for (sl = temp_configs; sl != NULL; sl = sl->next)\\n> {\\n> \\n> This code comment is a leftover and should go.\\n\\nNo - I was wondering whether the GUC should be specified by the\\nenvironment or by the file; I think it's better by file. Then it also\\nneeds to be in Cluster.pm.\\n\\n> 0004, 0005, 0006, 0007: EXPLAIN (MACHINE)\\n> \\n> I think it is confusing that these are included in this patch set.\\n> EXPLAIN (MACHINE OFF) is similar to \\\"explain_regress = on\\\", only it goes even further:\\n> no query ID, no Heap Fetches, no Sort details, ...\\n> Why not add this functionality to the GUC?\\n\\nGood question, and one I've asked myself.\\n\\nexplain_regress affects pre-existing uses of explain in the regression\\ntests, aiming to simplify current (or maybe only future) uses. And\\nis obviously used to allow enabling BUFFERS by default.\\n\\nexplain(MACHINE off) is mostly about allowing \\\"explain ANALYZE\\\" to be\\nused in regression tests. Which, right now, is rare. Maybe I shouldn't\\ninclude those patches until the earlier patches progress (or, maybe we\\nshould just defer their discussion).\\n\\n> 0005 suppresses \\\"rows removed by filter\\\", but how is that machine dependent?\\n\\nIt can vary with the number of parallel workers (see 13e8b2ee8), which\\nmay not be \\\"machine dependent\\\" but the point of that patch is to allow\\npredictable output of EXPLAIN ANALYZE. Maybe it needs a better name (or\\nmaybe it should be folded into the first patch).\\n\\nI'm not actually sure if this patch should try to address parallel\\nworkers at all, or (if so) if what it's doing now is adequate.\\n\\n> > BTW, I think it may be that the GUC should be marked PGDLLIMPORT ?\\n> \\n> I think it is project policy to apply this mark wherever it is needed. Do you think\\n> that third-party extensions might need to use this in C code?\\n\\nSince v15 (8ec569479), the policy is:\\n| On Windows, export all the server's global variables using PGDLLIMPORT markers (Robert Haas)\\n\\nI'm convinced now that's what's intended here.\\n\\n> This is not enough. The patch would have to update all the examples\\n> that use EXPLAIN ANALYZE.\\n\\nFair enough. I've left a comment to handle this later if the patch\\ngains any traction and 001 is progressed.\\n\\n-- \\nJustin\",\n \"msg_date\": \"Fri, 28 Oct 2022 17:59:04 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE), and default to\\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n },\n {\n \"msg_contents\": \"Thanks for the updated patch set!\\n\\nOn Fri, 2022-10-28 at 17:59 -0500, Justin Pryzby wrote:\\n> \\n> > 0004, 0005, 0006, 0007: EXPLAIN (MACHINE)\\n> > \\n> > I think it is confusing that these are included in this patch set.\\n> > EXPLAIN (MACHINE OFF) is similar to \\\"explain_regress = on\\\", only it goes even further:\\n> > no query ID, no Heap Fetches, no Sort details, ...\\n> > Why not add this functionality to the GUC?\\n> \\n> Good question, and one I've asked myself.\\n> \\n> explain_regress affects pre-existing uses of explain in the regression\\n> tests, aiming to simplify current (or maybe only future) uses. And\\n> is obviously used to allow enabling BUFFERS by default.\\n> \\n> explain(MACHINE off) is mostly about allowing \\\"explain ANALYZE\\\" to be\\n> used in regression tests. Which, right now, is rare. Maybe I shouldn't\\n> include those patches until the earlier patches progress (or, maybe we\\n> should just defer their discussion).\\n\\nEssentially, \\\"explain_regress\\\" is to simplify the current regression tests,\\nand MACHINE OFF is to simplify future regression tests. That does not seem\\nlike a meaningful distinction to me. Both are only useful for the regression\\ntests, and I see no need for two different knobs.\\n\\nMy opinion is now like this:\\n\\n+1 on enabling BUFFERS by default for EXPLAIN (ANALYZE)\\n\\n+0.2 on \\\"explain_regress\\\"\\n\\n-1 on EXPLAIN (MACHINE) in addition to \\\"explain_regress\\\"\\n\\n-0.1 on adding the MACHINE OFF functionality to \\\"explain_regress\\\"\\n\\n\\\"explain_regress\\\" reduces the EXPLAIN options you need for regression tests.\\nThis is somewhat useful, but not a big win. Also, it will make backpatching\\nregression tests slightly harder for the next 5 years.\\n\\nHence the +0.2 for \\\"explain_regress\\\".\\n\\nFor me, an important question is: do we really want more EXPLAIN (ANALYZE)\\nin the regression tests? It will probably slow the tests down, and I wonder\\nif there is a lot of added value, if we have to remove all the machine-specific\\noutput anyway.\\n\\nHence the -0.1.\\n\\n> > 0005 suppresses \\\"rows removed by filter\\\", but how is that machine dependent?\\n> \\n> It can vary with the number of parallel workers (see 13e8b2ee8), which\\n> may not be \\\"machine dependent\\\" but the point of that patch is to allow\\n> predictable output of EXPLAIN ANALYZE. Maybe it needs a better name (or\\n> maybe it should be folded into the first patch).\\n\\nNow it makes sense to me. I just didn't think of that.\\n\\n> I'm not actually sure if this patch should try to address parallel\\n> workers at all, or (if so) if what it's doing now is adequate.\\n> \\n> > > BTW, I think it may be that the GUC should be marked PGDLLIMPORT ?\\n> > \\n> > I think it is project policy to apply this mark wherever it is needed. Do you think\\n> > that third-party extensions might need to use this in C code?\\n> \\n> Since v15 (8ec569479), the policy is:\\n> > On Windows, export all the server's global variables using PGDLLIMPORT markers (Robert Haas)\\n> \\n> I'm convinced now that's what's intended here.\\n\\nYou convinced me too.\\n\\n> > This is not enough. The patch would have to update all the examples\\n> > that use EXPLAIN ANALYZE.\\n> \\n> Fair enough. I've left a comment to handle this later if the patch\\n> gains any traction and 001 is progressed.\\n\\nI agree with that.\\n\\nI would really like to see 0003 go in, but it currently depends on 0001, which\\nI am not so sure about.\\nI understand that you did that so that \\\"explain_regress\\\" can turn off BUFFERS\\nand there is no extra churn in the regression tests.\\n\\nStill, it would be a shame if resistance against \\\"explain_regress\\\" would\\nbe a show-stopper for 0003.\\n\\nIf I could get my way, I'd want two separate patches: first, one to turn\\nBUFFERS on, and second one for \\\"explain_regress\\\" with its current functionality\\non top of that.\\n\\nYours,\\nLaurenz Albe\\n\\n\\n\",\n \"msg_date\": \"Fri, 04 Nov 2022 11:46:03 +0100\",\n \"msg_from\": \"Laurenz Albe <laurenz.albe@cybertec.at>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE), and default to\\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 04, 2022 at 11:46:03AM +0100, Laurenz Albe wrote:\\n\\n> \\\"explain_regress\\\" reduces the EXPLAIN options you need for regression tests.\\n> This is somewhat useful, but not a big win. Also, it will make backpatching\\n> regression tests slightly harder for the next 5 years.\\n\\nBut it doesn't cause backpatching to be harder; if a patch is meant to\\nbe backpatched, its author can still choose to write the old way, with\\n(COSTS OFF, ...)\\n\\n> For me, an important question is: do we really want more EXPLAIN (ANALYZE)\\n> in the regression tests? It will probably slow the tests down, and I wonder\\n> if there is a lot of added value, if we have to remove all the machine-specific\\n> output anyway.\\n\\nI don't intend to encourage putting lots of \\\"explain analyze\\\" in the\\ntests. Rather, this makes that a reasonable possibility rather than\\nsomething that people are discouraged from doing by its difficulty.\\nUsing \\\"explain analyze\\\" still needs to be evaluated (for speed and\\notherwise), same as always.\\n\\nIn any case, I suggested to defer review of patches 004 and beyond,\\nwhich are optional, and it distracts from the essential discussion about\\n001.\\n\\n> I would really like to see 0003 go in, but it currently depends on 0001, which\\n> I am not so sure about.\\n> I understand that you did that so that \\\"explain_regress\\\" can turn off BUFFERS\\n> and there is no extra churn in the regression tests.\\n> \\n> Still, it would be a shame if resistance against \\\"explain_regress\\\" would\\n> be a show-stopper for 0003.\\n> \\n> If I could get my way, I'd want two separate patches: first, one to turn\\n> BUFFERS on, and second one for \\\"explain_regress\\\" with its current functionality\\n> on top of that.\\n\\nYou set the patch to \\\"waiting on author\\\", which indicates that there's\\nno need for further input or review. But, I think that's precisely\\nwhat's needed - without input from more people, what could I do to\\nprogress the patch ? I don't think it's reasonable to put 001 first and\\nchange thousands (actually, 1338) of regression results. If nobody\\nwants to discuss 001, then this patch series won't progress.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sat, 5 Nov 2022 08:44:25 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE), and default to\\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n },\n {\n \"msg_contents\": \"Justin Pryzby <pryzby@telsasoft.com> writes:\\n> You set the patch to \\\"waiting on author\\\", which indicates that there's\\n> no need for further input or review. But, I think that's precisely\\n> what's needed - without input from more people, what could I do to\\n> progress the patch ? I don't think it's reasonable to put 001 first and\\n> change thousands (actually, 1338) of regression results. If nobody\\n> wants to discuss 001, then this patch series won't progress.\\n\\nWell ...\\n\\n1. 0001 invents a new GUC but provides no documentation for it.\\nThat certainly isn't committable, and it's discouraging the\\ndiscussion you seek because people have to read the whole patch\\nin detail to understand what is being proposed.\\n\\n2. The same complaint for 0004, which labors under the additional\\nproblem that MACHINE is one of the worst option names I've ever\\nseen proposed around here. It conveys *nothing* about what it does\\nor is good for. The fact that it's got no relationship to the GUC\\nname, and yet they're intimately connected, doesn't improve my\\nopinion of it.\\n\\n3. I'm not really on board with the entire approach. Making\\nEXPLAIN work significantly differently for developers and test\\ncritters than it does for everyone else seems likely to promote\\nconfusion and hide bugs. I don't think getting rid of the need\\nfor filter functions in test scripts is worth that.\\n\\n4. I don't see how the current patches could pass under \\\"make\\ninstallcheck\\\" --- it looks to me like it only takes care of\\napplying the GUC change in installations built by pg_regress\\nor Cluster.pm. To make this actually work, you'd have to\\nadd \\\"explain_regress = on\\\" to the ALTER DATABASE SET commands\\nissued for regression databases. That'd substantially increase\\nthe problem of \\\"it works differently than what users see\\\", at\\nleast for me --- I do a lot of stuff interactively in the\\nregression database.\\n\\n5. The change in postgres_fdw.c in 0001 concerns me too.\\nYeah, it will fix things if an updated postgres_fdw talks to\\nan updated server, but what about an older postgres_fdw?\\nDon't really want to see that case break; communication with\\nservers of different major versions is one of postgres_fdw's\\nmain goals.\\n\\nAs a side note, 0001 also creates a hard break in postgres_fdw's\\nability to work with pre-9.0 remote servers, which won't accept\\n\\\"EXPLAIN (COSTS)\\\". Maybe we don't care about that anymore, given\\nthe policy we've adopted elsewhere that we won't test or support\\ninteroperability with versions before 9.2. But we should either\\nmake the command spelling conditional on remote version, or\\nofficially move that goalpost.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 05 Nov 2022 10:43:07 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE),\\n and default to explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN\\n (ANALYZE))\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 05, 2022 at 10:43:07AM -0400, Tom Lane wrote:\\n> Justin Pryzby <pryzby@telsasoft.com> writes:\\n> > You set the patch to \\\"waiting on author\\\", which indicates that there's\\n> > no need for further input or review. But, I think that's precisely\\n> > what's needed - without input from more people, what could I do to\\n> > progress the patch ? I don't think it's reasonable to put 001 first and\\n> > change thousands (actually, 1338) of regression results. If nobody\\n> > wants to discuss 001, then this patch series won't progress.\\n> \\n> Well ...\\n> \\n> 1. 0001 invents a new GUC but provides no documentation for it.\\n> That certainly isn't committable, and it's discouraging the\\n> discussion you seek because people have to read the whole patch\\n> in detail to understand what is being proposed.\\n> \\n> 2. The same complaint for 0004, which labors under the additional\\n\\nI suggested that the discussion be limited to the early patches (004 is\\nan optional, possible idea that I had, but it's evidently still causing\\nconfusion).\\n\\n> 3. I'm not really on board with the entire approach. Making\\n> EXPLAIN work significantly differently for developers and test\\n> critters than it does for everyone else seems likely to promote\\n> confusion and hide bugs. I don't think getting rid of the need\\n> for filter functions in test scripts is worth that.\\n\\nI'm open to suggestions how else to move towards the goal.\\n\\nNote that there's a few other things which have vaguely-similar\\nbehavior:\\n\\nHIDE_TABLEAM=on\\nHIDE_TOAST_COMPRESSION=on\\ncompute_query_id = regress\\n\\nAnother possibility is to make a new \\\"explain\\\" format, like\\n| explain(FORMAT REGRESS, ...).\\n\\n...but I don't see how that's would be less objectionable than what I've\\nwritten.\\n\\nThe patch record should probably be closed until someone proposes\\nanother way to implement what's necessary to enable explain(BUFFERS) by\\ndefault.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Thu, 1 Dec 2022 12:08:56 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: explain_regress, explain(MACHINE), and default to\\n explain(BUFFERS) (was: BUFFERS enabled by default in EXPLAIN (ANALYZE))\"\n }\n]","isConversational":false}}},{"rowIdx":53085,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"\nHi, hackers\n\nWhen I try to create table that has a varchar(n) data type, I find an\ninconsistent error message for it.\n\npostgres=# CREATE TABLE tbl (s varchar(2147483647));\nERROR: length for type varchar cannot exceed 10485760\nLINE 1: CREATE TABLE tbl (s varchar(2147483647));\n ^\n\npostgres=# CREATE TABLE tbl (s varchar(2147483648));\nERROR: syntax error at or near \"2147483648\"\nLINE 1: CREATE TABLE tbl (s varchar(2147483648));\n ^\n\nI find that in gram.y the varchar has an integer parameter which\nmeans its value don't exceed 2147483647.\n\nThe first error message is reported by anychar_typmodin(), and the later\nis reported by gram.y. IMO, the syntax error for varchar(n) is more\nconfused.\n\nAny thoughts?\n\n-- \nRegrads,\nJapin Li.\nChengDu WenWu Information Technology Co.,Ltd.\n\n\n","msg_date":"Sat, 13 Nov 2021 11:04:12 +0800","msg_from":"Japin Li <japinli@hotmail.com>","msg_from_op":true,"msg_subject":"Inconsistent error message for varchar(n)"},{"msg_contents":"Japin Li <japinli@hotmail.com> writes:\n> postgres=# CREATE TABLE tbl (s varchar(2147483647));\n> ERROR: length for type varchar cannot exceed 10485760\n> LINE 1: CREATE TABLE tbl (s varchar(2147483647));\n> ^\n\n> postgres=# CREATE TABLE tbl (s varchar(2147483648));\n> ERROR: syntax error at or near \"2147483648\"\n> LINE 1: CREATE TABLE tbl (s varchar(2147483648));\n> ^\n\nI'm having a very hard time getting excited about that. We could maybe\nswitch the grammar production to use generic expr_list syntax for the\ntypmod, like GenericType does. But that would just result in this:\n\nregression=# CREATE TABLE tbl (s \"varchar\"(2147483648));\nERROR: value \"2147483648\" is out of range for type integer\nLINE 1: CREATE TABLE tbl (s \"varchar\"(2147483648));\n ^\n\nwhich doesn't seem any less confusing for a novice who doesn't know\nthat typmods are constrained to be integers.\n\nThere might be something to be said for switching all the hard-wired\ntype productions to use opt_type_modifiers and pushing the knowledge\nthat's in, eg, opt_float out to per-type typmodin routines. But any\nbenefit would be in reduction of the grammar size, and I'm dubious\nthat it'd be worth the trouble. I suspect that overall, the resulting\nerror messages would be slightly worse not better --- note for example\nthe poorer placement of the error cursor above. A related example is\n\nregression=# CREATE TABLE tbl (s varchar(2,3));\nERROR: syntax error at or near \",\"\nLINE 1: CREATE TABLE tbl (s varchar(2,3));\n ^\nregression=# CREATE TABLE tbl (s \"varchar\"(2,3));\nERROR: invalid type modifier\nLINE 1: CREATE TABLE tbl (s \"varchar\"(2,3));\n ^\n\nThat's explained by the comment in anychar_typmodin:\n\n\t * we're not too tense about good error message here because grammar\n\t * shouldn't allow wrong number of modifiers for CHAR\n\nand we could surely improve that message, but anychar_typmodin can't give\na really on-point error cursor.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 13 Nov 2021 10:42:25 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Inconsistent error message for varchar(n)"},{"msg_contents":"\nOn Sat, 13 Nov 2021 at 23:42, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Japin Li <japinli@hotmail.com> writes:\n>> postgres=# CREATE TABLE tbl (s varchar(2147483647));\n>> ERROR: length for type varchar cannot exceed 10485760\n>> LINE 1: CREATE TABLE tbl (s varchar(2147483647));\n>> ^\n>\n>> postgres=# CREATE TABLE tbl (s varchar(2147483648));\n>> ERROR: syntax error at or near \"2147483648\"\n>> LINE 1: CREATE TABLE tbl (s varchar(2147483648));\n>> ^\n>\n> I'm having a very hard time getting excited about that. We could maybe\n> switch the grammar production to use generic expr_list syntax for the\n> typmod, like GenericType does. But that would just result in this:\n>\n> regression=# CREATE TABLE tbl (s \"varchar\"(2147483648));\n> ERROR: value \"2147483648\" is out of range for type integer\n> LINE 1: CREATE TABLE tbl (s \"varchar\"(2147483648));\n> ^\n>\n> which doesn't seem any less confusing for a novice who doesn't know\n> that typmods are constrained to be integers.\n>\n> There might be something to be said for switching all the hard-wired\n> type productions to use opt_type_modifiers and pushing the knowledge\n> that's in, eg, opt_float out to per-type typmodin routines. But any\n> benefit would be in reduction of the grammar size, and I'm dubious\n> that it'd be worth the trouble. I suspect that overall, the resulting\n> error messages would be slightly worse not better --- note for example\n> the poorer placement of the error cursor above. A related example is\n>\n> regression=# CREATE TABLE tbl (s varchar(2,3));\n> ERROR: syntax error at or near \",\"\n> LINE 1: CREATE TABLE tbl (s varchar(2,3));\n> ^\n> regression=# CREATE TABLE tbl (s \"varchar\"(2,3));\n> ERROR: invalid type modifier\n> LINE 1: CREATE TABLE tbl (s \"varchar\"(2,3));\n> ^\n>\n> That's explained by the comment in anychar_typmodin:\n>\n> \t * we're not too tense about good error message here because grammar\n> \t * shouldn't allow wrong number of modifiers for CHAR\n>\n> and we could surely improve that message, but anychar_typmodin can't give\n> a really on-point error cursor.\n>\n\nOh! I didn't consider this situation. Since the max size of varchar cannot\nexceed 10485760, however, I cannot find this in documentation [1]. Is there\nsomething I missed? Should we mention this in the documentation?\n\n[1] https://www.postgresql.org/docs/devel/datatype-character.html\n\n-- \nRegrads,\nJapin Li.\nChengDu WenWu Information Technology Co.,Ltd.\n\n\n","msg_date":"Sun, 14 Nov 2021 10:33:19 +0800","msg_from":"Japin Li <japinli@hotmail.com>","msg_from_op":true,"msg_subject":"Re: Inconsistent error message for varchar(n)"},{"msg_contents":"[ Please trim quotes appropriately when replying. Nobody wants to\n read the whole history of the thread to get to your comment. ]\n\nJapin Li <japinli@hotmail.com> writes:\n> Oh! I didn't consider this situation. Since the max size of varchar cannot\n> exceed 10485760, however, I cannot find this in documentation [1]. Is there\n> something I missed? Should we mention this in the documentation?\n> [1] https://www.postgresql.org/docs/devel/datatype-character.html\n\nI dunno, that section doesn't really get into implementation limits.\nFor comparison, it doesn't bring up the point that string values are\nconstrained to 1GB; that's dealt with elsewhere. Since the limit on\ntypmod is substantially more than that, I'm not sure there's much point\nin mentioning it specifically. Maybe there's a case for mentioning the\n1GB limit here, though.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 14 Nov 2021 14:00:27 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Inconsistent error message for varchar(n)"},{"msg_contents":"I wrote:\n> For comparison, it doesn't bring up the point that string values are\n> constrained to 1GB; that's dealt with elsewhere. Since the limit on\n> typmod is substantially more than that, I'm not sure there's much point\n> in mentioning it specifically.\n\nOh, wait, I was not counting the zeroes in that number :-(\n\nTracking it a bit further, the actual typmod limit is set by this:\n\n/*\n * MaxAttrSize is a somewhat arbitrary upper limit on the declared size of\n * data fields of char(n) and similar types. It need not have anything\n * directly to do with the *actual* upper limit of varlena values, which\n * is currently 1Gb (see TOAST structures in postgres.h). I've set it\n * at 10Mb which seems like a reasonable number --- tgl 8/6/00.\n */\n#define MaxAttrSize\t\t(10 * 1024 * 1024)\n\nSo maybe that's something we *should* document, though we'd have to\nexplain that the limit on text and unconstrained varchar is different.\n\n\t\t\tregards, tom lane\n\n(From the writing style, I suspect the \"tgl\" here is me not Tom Lockhart.\nI'm too lazy to dig in the git history to confirm it though.)\n\n\n","msg_date":"Sun, 14 Nov 2021 14:11:30 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Inconsistent error message for varchar(n)"},{"msg_contents":"Tom Lane <tgl@sss.pgh.pa.us> writes:\n\n> Tracking it a bit further, the actual typmod limit is set by this:\n>\n> /*\n> * MaxAttrSize is a somewhat arbitrary upper limit on the declared size of\n> * data fields of char(n) and similar types. It need not have anything\n> * directly to do with the *actual* upper limit of varlena values, which\n> * is currently 1Gb (see TOAST structures in postgres.h). I've set it\n> * at 10Mb which seems like a reasonable number --- tgl 8/6/00.\n> */\n> #define MaxAttrSize\t\t(10 * 1024 * 1024)\n>\n> So maybe that's something we *should* document, though we'd have to\n> explain that the limit on text and unconstrained varchar is different.\n>\n> \t\t\tregards, tom lane\n>\n> (From the writing style, I suspect the \"tgl\" here is me not Tom Lockhart.\n> I'm too lazy to dig in the git history to confirm it though.)\n\nI was bored, and found this:\n\ncommit 022417740094620880488dd9b04fbb96ff11694b\nAuthor: Tom Lane <tgl@sss.pgh.pa.us>\nDate: 2000-08-07 20:16:13 +0000\n\n TOAST mop-up work: update comments for tuple-size-related symbols such\n as MaxHeapAttributeNumber. Increase MaxAttrSize to something more\n reasonable (given what it's used for, namely checking char(n) declarations,\n I didn't make it the full 1G that it could theoretically be --- 10Mb\n seemed a more reasonable number). Improve calculation of MaxTupleSize.\n\nwhich added the above comment and changed MaxAttrSize:\n\n-#define MaxAttrSize (MaxTupleSize - MAXALIGN(sizeof(HeapTupleHeaderData)))\n+#define MaxAttrSize (10 * 1024 * 1024)\n \n- ilmari\n\n\n","msg_date":"Sun, 14 Nov 2021 20:42:33 +0000","msg_from":"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org>","msg_from_op":false,"msg_subject":"Re: Inconsistent error message for varchar(n)"},{"msg_contents":"On Sun, Nov 14, 2021 at 10:33:19AM +0800, Japin Li wrote:\n> \n> On Sat, 13 Nov 2021 at 23:42, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > Japin Li <japinli@hotmail.com> writes:\n> >> postgres=# CREATE TABLE tbl (s varchar(2147483647));\n> >> ERROR: length for type varchar cannot exceed 10485760\n> >> LINE 1: CREATE TABLE tbl (s varchar(2147483647));\n> >> ^\n> >\n> >> postgres=# CREATE TABLE tbl (s varchar(2147483648));\n> >> ERROR: syntax error at or near \"2147483648\"\n> >> LINE 1: CREATE TABLE tbl (s varchar(2147483648));\n> >> ^\n> >\n> > I'm having a very hard time getting excited about that. We could maybe\n> > switch the grammar production to use generic expr_list syntax for the\n> > typmod, like GenericType does. But that would just result in this:\n> >\n> > regression=# CREATE TABLE tbl (s \"varchar\"(2147483648));\n> > ERROR: value \"2147483648\" is out of range for type integer\n> > LINE 1: CREATE TABLE tbl (s \"varchar\"(2147483648));\n> > ^\n> >\n> > which doesn't seem any less confusing for a novice who doesn't know\n> > that typmods are constrained to be integers.\n> >\n> > There might be something to be said for switching all the hard-wired\n> > type productions to use opt_type_modifiers and pushing the knowledge\n> > that's in, eg, opt_float out to per-type typmodin routines. But any\n> > benefit would be in reduction of the grammar size, and I'm dubious\n> > that it'd be worth the trouble. I suspect that overall, the resulting\n> > error messages would be slightly worse not better --- note for example\n> > the poorer placement of the error cursor above. A related example is\n> >\n> > regression=# CREATE TABLE tbl (s varchar(2,3));\n> > ERROR: syntax error at or near \",\"\n> > LINE 1: CREATE TABLE tbl (s varchar(2,3));\n> > ^\n> > regression=# CREATE TABLE tbl (s \"varchar\"(2,3));\n> > ERROR: invalid type modifier\n> > LINE 1: CREATE TABLE tbl (s \"varchar\"(2,3));\n> > ^\n> >\n> > That's explained by the comment in anychar_typmodin:\n> >\n> > \t * we're not too tense about good error message here because grammar\n> > \t * shouldn't allow wrong number of modifiers for CHAR\n> >\n> > and we could surely improve that message, but anychar_typmodin can't give\n> > a really on-point error cursor.\n> >\n> \n> Oh! I didn't consider this situation. Since the max size of varchar cannot\n> exceed 10485760, however, I cannot find this in documentation [1]. Is there\n> something I missed? Should we mention this in the documentation?\n> \n> [1] https://www.postgresql.org/docs/devel/datatype-character.html\n\nSorry for my long delay in reviewing this issue. You are correct this\nshould be documented --- patch attached.\n\n-- \n Bruce Momjian <bruce@momjian.us> https://momjian.us\n EDB https://enterprisedb.com\n\n Indecision is a decision. Inaction is an action. Mark Batterson","msg_date":"Tue, 16 Aug 2022 21:56:17 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Inconsistent error message for varchar(n)"},{"msg_contents":"8On Tue, Aug 16, 2022 at 09:56:17PM -0400, Bruce Momjian wrote:\n> On Sun, Nov 14, 2021 at 10:33:19AM +0800, Japin Li wrote:\n> > Oh! I didn't consider this situation. Since the max size of varchar cannot\n> > exceed 10485760, however, I cannot find this in documentation [1]. Is there\n> > something I missed? Should we mention this in the documentation?\n> > \n> > [1] https://www.postgresql.org/docs/devel/datatype-character.html\n> \n> Sorry for my long delay in reviewing this issue. You are correct this\n> should be documented --- patch attached.\n\nPatch applied back to PG 10. Thanks for the report.\n\n-- \n Bruce Momjian <bruce@momjian.us> https://momjian.us\n EDB https://enterprisedb.com\n\n Indecision is a decision. Inaction is an action. Mark Batterson\n\n\n\n","msg_date":"Wed, 31 Aug 2022 19:45:34 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Inconsistent error message for varchar(n)"}],"string":"[\n {\n \"msg_contents\": \"\\nHi, hackers\\n\\nWhen I try to create table that has a varchar(n) data type, I find an\\ninconsistent error message for it.\\n\\npostgres=# CREATE TABLE tbl (s varchar(2147483647));\\nERROR: length for type varchar cannot exceed 10485760\\nLINE 1: CREATE TABLE tbl (s varchar(2147483647));\\n ^\\n\\npostgres=# CREATE TABLE tbl (s varchar(2147483648));\\nERROR: syntax error at or near \\\"2147483648\\\"\\nLINE 1: CREATE TABLE tbl (s varchar(2147483648));\\n ^\\n\\nI find that in gram.y the varchar has an integer parameter which\\nmeans its value don't exceed 2147483647.\\n\\nThe first error message is reported by anychar_typmodin(), and the later\\nis reported by gram.y. IMO, the syntax error for varchar(n) is more\\nconfused.\\n\\nAny thoughts?\\n\\n-- \\nRegrads,\\nJapin Li.\\nChengDu WenWu Information Technology Co.,Ltd.\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 11:04:12 +0800\",\n \"msg_from\": \"Japin Li <japinli@hotmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Inconsistent error message for varchar(n)\"\n },\n {\n \"msg_contents\": \"Japin Li <japinli@hotmail.com> writes:\\n> postgres=# CREATE TABLE tbl (s varchar(2147483647));\\n> ERROR: length for type varchar cannot exceed 10485760\\n> LINE 1: CREATE TABLE tbl (s varchar(2147483647));\\n> ^\\n\\n> postgres=# CREATE TABLE tbl (s varchar(2147483648));\\n> ERROR: syntax error at or near \\\"2147483648\\\"\\n> LINE 1: CREATE TABLE tbl (s varchar(2147483648));\\n> ^\\n\\nI'm having a very hard time getting excited about that. We could maybe\\nswitch the grammar production to use generic expr_list syntax for the\\ntypmod, like GenericType does. But that would just result in this:\\n\\nregression=# CREATE TABLE tbl (s \\\"varchar\\\"(2147483648));\\nERROR: value \\\"2147483648\\\" is out of range for type integer\\nLINE 1: CREATE TABLE tbl (s \\\"varchar\\\"(2147483648));\\n ^\\n\\nwhich doesn't seem any less confusing for a novice who doesn't know\\nthat typmods are constrained to be integers.\\n\\nThere might be something to be said for switching all the hard-wired\\ntype productions to use opt_type_modifiers and pushing the knowledge\\nthat's in, eg, opt_float out to per-type typmodin routines. But any\\nbenefit would be in reduction of the grammar size, and I'm dubious\\nthat it'd be worth the trouble. I suspect that overall, the resulting\\nerror messages would be slightly worse not better --- note for example\\nthe poorer placement of the error cursor above. A related example is\\n\\nregression=# CREATE TABLE tbl (s varchar(2,3));\\nERROR: syntax error at or near \\\",\\\"\\nLINE 1: CREATE TABLE tbl (s varchar(2,3));\\n ^\\nregression=# CREATE TABLE tbl (s \\\"varchar\\\"(2,3));\\nERROR: invalid type modifier\\nLINE 1: CREATE TABLE tbl (s \\\"varchar\\\"(2,3));\\n ^\\n\\nThat's explained by the comment in anychar_typmodin:\\n\\n\\t * we're not too tense about good error message here because grammar\\n\\t * shouldn't allow wrong number of modifiers for CHAR\\n\\nand we could surely improve that message, but anychar_typmodin can't give\\na really on-point error cursor.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 10:42:25 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Inconsistent error message for varchar(n)\"\n },\n {\n \"msg_contents\": \"\\nOn Sat, 13 Nov 2021 at 23:42, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Japin Li <japinli@hotmail.com> writes:\\n>> postgres=# CREATE TABLE tbl (s varchar(2147483647));\\n>> ERROR: length for type varchar cannot exceed 10485760\\n>> LINE 1: CREATE TABLE tbl (s varchar(2147483647));\\n>> ^\\n>\\n>> postgres=# CREATE TABLE tbl (s varchar(2147483648));\\n>> ERROR: syntax error at or near \\\"2147483648\\\"\\n>> LINE 1: CREATE TABLE tbl (s varchar(2147483648));\\n>> ^\\n>\\n> I'm having a very hard time getting excited about that. We could maybe\\n> switch the grammar production to use generic expr_list syntax for the\\n> typmod, like GenericType does. But that would just result in this:\\n>\\n> regression=# CREATE TABLE tbl (s \\\"varchar\\\"(2147483648));\\n> ERROR: value \\\"2147483648\\\" is out of range for type integer\\n> LINE 1: CREATE TABLE tbl (s \\\"varchar\\\"(2147483648));\\n> ^\\n>\\n> which doesn't seem any less confusing for a novice who doesn't know\\n> that typmods are constrained to be integers.\\n>\\n> There might be something to be said for switching all the hard-wired\\n> type productions to use opt_type_modifiers and pushing the knowledge\\n> that's in, eg, opt_float out to per-type typmodin routines. But any\\n> benefit would be in reduction of the grammar size, and I'm dubious\\n> that it'd be worth the trouble. I suspect that overall, the resulting\\n> error messages would be slightly worse not better --- note for example\\n> the poorer placement of the error cursor above. A related example is\\n>\\n> regression=# CREATE TABLE tbl (s varchar(2,3));\\n> ERROR: syntax error at or near \\\",\\\"\\n> LINE 1: CREATE TABLE tbl (s varchar(2,3));\\n> ^\\n> regression=# CREATE TABLE tbl (s \\\"varchar\\\"(2,3));\\n> ERROR: invalid type modifier\\n> LINE 1: CREATE TABLE tbl (s \\\"varchar\\\"(2,3));\\n> ^\\n>\\n> That's explained by the comment in anychar_typmodin:\\n>\\n> \\t * we're not too tense about good error message here because grammar\\n> \\t * shouldn't allow wrong number of modifiers for CHAR\\n>\\n> and we could surely improve that message, but anychar_typmodin can't give\\n> a really on-point error cursor.\\n>\\n\\nOh! I didn't consider this situation. Since the max size of varchar cannot\\nexceed 10485760, however, I cannot find this in documentation [1]. Is there\\nsomething I missed? Should we mention this in the documentation?\\n\\n[1] https://www.postgresql.org/docs/devel/datatype-character.html\\n\\n-- \\nRegrads,\\nJapin Li.\\nChengDu WenWu Information Technology Co.,Ltd.\\n\\n\\n\",\n \"msg_date\": \"Sun, 14 Nov 2021 10:33:19 +0800\",\n \"msg_from\": \"Japin Li <japinli@hotmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Inconsistent error message for varchar(n)\"\n },\n {\n \"msg_contents\": \"[ Please trim quotes appropriately when replying. Nobody wants to\\n read the whole history of the thread to get to your comment. ]\\n\\nJapin Li <japinli@hotmail.com> writes:\\n> Oh! I didn't consider this situation. Since the max size of varchar cannot\\n> exceed 10485760, however, I cannot find this in documentation [1]. Is there\\n> something I missed? Should we mention this in the documentation?\\n> [1] https://www.postgresql.org/docs/devel/datatype-character.html\\n\\nI dunno, that section doesn't really get into implementation limits.\\nFor comparison, it doesn't bring up the point that string values are\\nconstrained to 1GB; that's dealt with elsewhere. Since the limit on\\ntypmod is substantially more than that, I'm not sure there's much point\\nin mentioning it specifically. Maybe there's a case for mentioning the\\n1GB limit here, though.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 14 Nov 2021 14:00:27 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Inconsistent error message for varchar(n)\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> For comparison, it doesn't bring up the point that string values are\\n> constrained to 1GB; that's dealt with elsewhere. Since the limit on\\n> typmod is substantially more than that, I'm not sure there's much point\\n> in mentioning it specifically.\\n\\nOh, wait, I was not counting the zeroes in that number :-(\\n\\nTracking it a bit further, the actual typmod limit is set by this:\\n\\n/*\\n * MaxAttrSize is a somewhat arbitrary upper limit on the declared size of\\n * data fields of char(n) and similar types. It need not have anything\\n * directly to do with the *actual* upper limit of varlena values, which\\n * is currently 1Gb (see TOAST structures in postgres.h). I've set it\\n * at 10Mb which seems like a reasonable number --- tgl 8/6/00.\\n */\\n#define MaxAttrSize\\t\\t(10 * 1024 * 1024)\\n\\nSo maybe that's something we *should* document, though we'd have to\\nexplain that the limit on text and unconstrained varchar is different.\\n\\n\\t\\t\\tregards, tom lane\\n\\n(From the writing style, I suspect the \\\"tgl\\\" here is me not Tom Lockhart.\\nI'm too lazy to dig in the git history to confirm it though.)\\n\\n\\n\",\n \"msg_date\": \"Sun, 14 Nov 2021 14:11:30 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Inconsistent error message for varchar(n)\"\n },\n {\n \"msg_contents\": \"Tom Lane <tgl@sss.pgh.pa.us> writes:\\n\\n> Tracking it a bit further, the actual typmod limit is set by this:\\n>\\n> /*\\n> * MaxAttrSize is a somewhat arbitrary upper limit on the declared size of\\n> * data fields of char(n) and similar types. It need not have anything\\n> * directly to do with the *actual* upper limit of varlena values, which\\n> * is currently 1Gb (see TOAST structures in postgres.h). I've set it\\n> * at 10Mb which seems like a reasonable number --- tgl 8/6/00.\\n> */\\n> #define MaxAttrSize\\t\\t(10 * 1024 * 1024)\\n>\\n> So maybe that's something we *should* document, though we'd have to\\n> explain that the limit on text and unconstrained varchar is different.\\n>\\n> \\t\\t\\tregards, tom lane\\n>\\n> (From the writing style, I suspect the \\\"tgl\\\" here is me not Tom Lockhart.\\n> I'm too lazy to dig in the git history to confirm it though.)\\n\\nI was bored, and found this:\\n\\ncommit 022417740094620880488dd9b04fbb96ff11694b\\nAuthor: Tom Lane <tgl@sss.pgh.pa.us>\\nDate: 2000-08-07 20:16:13 +0000\\n\\n TOAST mop-up work: update comments for tuple-size-related symbols such\\n as MaxHeapAttributeNumber. Increase MaxAttrSize to something more\\n reasonable (given what it's used for, namely checking char(n) declarations,\\n I didn't make it the full 1G that it could theoretically be --- 10Mb\\n seemed a more reasonable number). Improve calculation of MaxTupleSize.\\n\\nwhich added the above comment and changed MaxAttrSize:\\n\\n-#define MaxAttrSize (MaxTupleSize - MAXALIGN(sizeof(HeapTupleHeaderData)))\\n+#define MaxAttrSize (10 * 1024 * 1024)\\n \\n- ilmari\\n\\n\\n\",\n \"msg_date\": \"Sun, 14 Nov 2021 20:42:33 +0000\",\n \"msg_from\": \"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Inconsistent error message for varchar(n)\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 14, 2021 at 10:33:19AM +0800, Japin Li wrote:\\n> \\n> On Sat, 13 Nov 2021 at 23:42, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > Japin Li <japinli@hotmail.com> writes:\\n> >> postgres=# CREATE TABLE tbl (s varchar(2147483647));\\n> >> ERROR: length for type varchar cannot exceed 10485760\\n> >> LINE 1: CREATE TABLE tbl (s varchar(2147483647));\\n> >> ^\\n> >\\n> >> postgres=# CREATE TABLE tbl (s varchar(2147483648));\\n> >> ERROR: syntax error at or near \\\"2147483648\\\"\\n> >> LINE 1: CREATE TABLE tbl (s varchar(2147483648));\\n> >> ^\\n> >\\n> > I'm having a very hard time getting excited about that. We could maybe\\n> > switch the grammar production to use generic expr_list syntax for the\\n> > typmod, like GenericType does. But that would just result in this:\\n> >\\n> > regression=# CREATE TABLE tbl (s \\\"varchar\\\"(2147483648));\\n> > ERROR: value \\\"2147483648\\\" is out of range for type integer\\n> > LINE 1: CREATE TABLE tbl (s \\\"varchar\\\"(2147483648));\\n> > ^\\n> >\\n> > which doesn't seem any less confusing for a novice who doesn't know\\n> > that typmods are constrained to be integers.\\n> >\\n> > There might be something to be said for switching all the hard-wired\\n> > type productions to use opt_type_modifiers and pushing the knowledge\\n> > that's in, eg, opt_float out to per-type typmodin routines. But any\\n> > benefit would be in reduction of the grammar size, and I'm dubious\\n> > that it'd be worth the trouble. I suspect that overall, the resulting\\n> > error messages would be slightly worse not better --- note for example\\n> > the poorer placement of the error cursor above. A related example is\\n> >\\n> > regression=# CREATE TABLE tbl (s varchar(2,3));\\n> > ERROR: syntax error at or near \\\",\\\"\\n> > LINE 1: CREATE TABLE tbl (s varchar(2,3));\\n> > ^\\n> > regression=# CREATE TABLE tbl (s \\\"varchar\\\"(2,3));\\n> > ERROR: invalid type modifier\\n> > LINE 1: CREATE TABLE tbl (s \\\"varchar\\\"(2,3));\\n> > ^\\n> >\\n> > That's explained by the comment in anychar_typmodin:\\n> >\\n> > \\t * we're not too tense about good error message here because grammar\\n> > \\t * shouldn't allow wrong number of modifiers for CHAR\\n> >\\n> > and we could surely improve that message, but anychar_typmodin can't give\\n> > a really on-point error cursor.\\n> >\\n> \\n> Oh! I didn't consider this situation. Since the max size of varchar cannot\\n> exceed 10485760, however, I cannot find this in documentation [1]. Is there\\n> something I missed? Should we mention this in the documentation?\\n> \\n> [1] https://www.postgresql.org/docs/devel/datatype-character.html\\n\\nSorry for my long delay in reviewing this issue. You are correct this\\nshould be documented --- patch attached.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> https://momjian.us\\n EDB https://enterprisedb.com\\n\\n Indecision is a decision. Inaction is an action. Mark Batterson\",\n \"msg_date\": \"Tue, 16 Aug 2022 21:56:17 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Inconsistent error message for varchar(n)\"\n },\n {\n \"msg_contents\": \"8On Tue, Aug 16, 2022 at 09:56:17PM -0400, Bruce Momjian wrote:\\n> On Sun, Nov 14, 2021 at 10:33:19AM +0800, Japin Li wrote:\\n> > Oh! I didn't consider this situation. Since the max size of varchar cannot\\n> > exceed 10485760, however, I cannot find this in documentation [1]. Is there\\n> > something I missed? Should we mention this in the documentation?\\n> > \\n> > [1] https://www.postgresql.org/docs/devel/datatype-character.html\\n> \\n> Sorry for my long delay in reviewing this issue. You are correct this\\n> should be documented --- patch attached.\\n\\nPatch applied back to PG 10. Thanks for the report.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> https://momjian.us\\n EDB https://enterprisedb.com\\n\\n Indecision is a decision. Inaction is an action. Mark Batterson\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 31 Aug 2022 19:45:34 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Inconsistent error message for varchar(n)\"\n }\n]","isConversational":false}}},{"rowIdx":53086,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"\nHi, hackers\n\nWhen I try to insert an Unicode \"\\u0000\", there is an error $subject.\n\npostgres=# CREATE TABLE tbl (s varchar(10));\nCREATE TABLE\npostgres=# INSERT INTO tbl VALUES (E'\\u0000');\nERROR: invalid Unicode escape value at or near \"\\u0000\"\nLINE 1: INSERT INTO tbl VALUES (E'\\u0000');\n ^\n\n\"\\u0000\" is valid unicode [1], why not we cannot insert it?\n\n[1] https://www.unicode.org/charts/PDF/U0000.pdf\n\n-- \nRegrads,\nJapin Li.\nChengDu WenWu Information Technology Co.,Ltd.\n\n\n","msg_date":"Sat, 13 Nov 2021 11:31:57 +0800","msg_from":"Japin Li <japinli@hotmail.com>","msg_from_op":true,"msg_subject":"Invalid Unicode escape value at or near \"\\u0000\""},{"msg_contents":"On Sat, Nov 13, 2021 at 4:32 PM Japin Li <japinli@hotmail.com> wrote:\n> When I try to insert an Unicode \"\\u0000\", there is an error $subject.\n>\n> postgres=# CREATE TABLE tbl (s varchar(10));\n> CREATE TABLE\n> postgres=# INSERT INTO tbl VALUES (E'\\u0000');\n> ERROR: invalid Unicode escape value at or near \"\\u0000\"\n> LINE 1: INSERT INTO tbl VALUES (E'\\u0000');\n> ^\n>\n> \"\\u0000\" is valid unicode [1], why not we cannot insert it?\n\nYes, it is a valid codepoint, but unfortunately PostgreSQL can't\nsupport it because it sometimes deals in null terminated string, even\nthough internally it does track string data and length separately. We\nhave to do that to use libc facilities like strcoll_r(), and probably\nmany other things.\n\n\n","msg_date":"Sat, 13 Nov 2021 18:40:41 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Invalid Unicode escape value at or near \"\\u0000\""},{"msg_contents":"\nOn 11/13/21 00:40, Thomas Munro wrote:\n> On Sat, Nov 13, 2021 at 4:32 PM Japin Li <japinli@hotmail.com> wrote:\n>> When I try to insert an Unicode \"\\u0000\", there is an error $subject.\n>>\n>> postgres=# CREATE TABLE tbl (s varchar(10));\n>> CREATE TABLE\n>> postgres=# INSERT INTO tbl VALUES (E'\\u0000');\n>> ERROR: invalid Unicode escape value at or near \"\\u0000\"\n>> LINE 1: INSERT INTO tbl VALUES (E'\\u0000');\n>> ^\n>>\n>> \"\\u0000\" is valid unicode [1], why not we cannot insert it?\n> Yes, it is a valid codepoint, but unfortunately PostgreSQL can't\n> support it because it sometimes deals in null terminated string, even\n> though internally it does track string data and length separately. We\n> have to do that to use libc facilities like strcoll_r(), and probably\n> many other things.\n>\n>\n\nAnd it's documented at\n<https://www.postgresql.org/docs/current/datatype-character.html>:\n\n\n The characters that can be stored in any of these data types are\n determined by the database character set, which is selected when the\n database is created. Regardless of the specific character set, the\n character with code zero (sometimes called NUL) cannot be stored.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Sat, 13 Nov 2021 08:52:01 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Invalid Unicode escape value at or near \"\\u0000\""},{"msg_contents":"\nOn Sat, 13 Nov 2021 at 21:52, Andrew Dunstan <andrew@dunslane.net> wrote:\n> On 11/13/21 00:40, Thomas Munro wrote:\n>> On Sat, Nov 13, 2021 at 4:32 PM Japin Li <japinli@hotmail.com> wrote:\n>>> When I try to insert an Unicode \"\\u0000\", there is an error $subject.\n>>>\n>>> postgres=# CREATE TABLE tbl (s varchar(10));\n>>> CREATE TABLE\n>>> postgres=# INSERT INTO tbl VALUES (E'\\u0000');\n>>> ERROR: invalid Unicode escape value at or near \"\\u0000\"\n>>> LINE 1: INSERT INTO tbl VALUES (E'\\u0000');\n>>> ^\n>>>\n>>> \"\\u0000\" is valid unicode [1], why not we cannot insert it?\n>> Yes, it is a valid codepoint, but unfortunately PostgreSQL can't\n>> support it because it sometimes deals in null terminated string, even\n>> though internally it does track string data and length separately. We\n>> have to do that to use libc facilities like strcoll_r(), and probably\n>> many other things.\n>>\n>>\n>\n> And it's documented at\n> <https://www.postgresql.org/docs/current/datatype-character.html>:\n>\n>\n> The characters that can be stored in any of these data types are\n> determined by the database character set, which is selected when the\n> database is created. Regardless of the specific character set, the\n> character with code zero (sometimes called NUL) cannot be stored.\n>\n>\n\nThanks Thomas and Andrew. Sorry for my ignore reading of the documentation.\n\n-- \nRegrads,\nJapin Li.\nChengDu WenWu Information Technology Co.,Ltd.\n\n\n","msg_date":"Sat, 13 Nov 2021 22:23:31 +0800","msg_from":"Japin Li <japinli@hotmail.com>","msg_from_op":true,"msg_subject":"Re: Invalid Unicode escape value at or near \"\\u0000\""},{"msg_contents":"On 13.11.21 06:40, Thomas Munro wrote:\n> Yes, it is a valid codepoint, but unfortunately PostgreSQL can't\n> support it because it sometimes deals in null terminated string, even\n> though internally it does track string data and length separately. We\n> have to do that to use libc facilities like strcoll_r(), and probably\n> many other things.\n\nBy the way, I think the server-side issues around this are fixable. The \nreal problem in my mind is how to present such values in client APIs \nsuch as libpq.\n\n\n","msg_date":"Mon, 15 Nov 2021 09:06:34 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Invalid Unicode escape value at or near \"\\u0000\""}],"string":"[\n {\n \"msg_contents\": \"\\nHi, hackers\\n\\nWhen I try to insert an Unicode \\\"\\\\u0000\\\", there is an error $subject.\\n\\npostgres=# CREATE TABLE tbl (s varchar(10));\\nCREATE TABLE\\npostgres=# INSERT INTO tbl VALUES (E'\\\\u0000');\\nERROR: invalid Unicode escape value at or near \\\"\\\\u0000\\\"\\nLINE 1: INSERT INTO tbl VALUES (E'\\\\u0000');\\n ^\\n\\n\\\"\\\\u0000\\\" is valid unicode [1], why not we cannot insert it?\\n\\n[1] https://www.unicode.org/charts/PDF/U0000.pdf\\n\\n-- \\nRegrads,\\nJapin Li.\\nChengDu WenWu Information Technology Co.,Ltd.\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 11:31:57 +0800\",\n \"msg_from\": \"Japin Li <japinli@hotmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Invalid Unicode escape value at or near \\\"\\\\u0000\\\"\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 13, 2021 at 4:32 PM Japin Li <japinli@hotmail.com> wrote:\\n> When I try to insert an Unicode \\\"\\\\u0000\\\", there is an error $subject.\\n>\\n> postgres=# CREATE TABLE tbl (s varchar(10));\\n> CREATE TABLE\\n> postgres=# INSERT INTO tbl VALUES (E'\\\\u0000');\\n> ERROR: invalid Unicode escape value at or near \\\"\\\\u0000\\\"\\n> LINE 1: INSERT INTO tbl VALUES (E'\\\\u0000');\\n> ^\\n>\\n> \\\"\\\\u0000\\\" is valid unicode [1], why not we cannot insert it?\\n\\nYes, it is a valid codepoint, but unfortunately PostgreSQL can't\\nsupport it because it sometimes deals in null terminated string, even\\nthough internally it does track string data and length separately. We\\nhave to do that to use libc facilities like strcoll_r(), and probably\\nmany other things.\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 18:40:41 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Invalid Unicode escape value at or near \\\"\\\\u0000\\\"\"\n },\n {\n \"msg_contents\": \"\\nOn 11/13/21 00:40, Thomas Munro wrote:\\n> On Sat, Nov 13, 2021 at 4:32 PM Japin Li <japinli@hotmail.com> wrote:\\n>> When I try to insert an Unicode \\\"\\\\u0000\\\", there is an error $subject.\\n>>\\n>> postgres=# CREATE TABLE tbl (s varchar(10));\\n>> CREATE TABLE\\n>> postgres=# INSERT INTO tbl VALUES (E'\\\\u0000');\\n>> ERROR: invalid Unicode escape value at or near \\\"\\\\u0000\\\"\\n>> LINE 1: INSERT INTO tbl VALUES (E'\\\\u0000');\\n>> ^\\n>>\\n>> \\\"\\\\u0000\\\" is valid unicode [1], why not we cannot insert it?\\n> Yes, it is a valid codepoint, but unfortunately PostgreSQL can't\\n> support it because it sometimes deals in null terminated string, even\\n> though internally it does track string data and length separately. We\\n> have to do that to use libc facilities like strcoll_r(), and probably\\n> many other things.\\n>\\n>\\n\\nAnd it's documented at\\n<https://www.postgresql.org/docs/current/datatype-character.html>:\\n\\n\\n The characters that can be stored in any of these data types are\\n determined by the database character set, which is selected when the\\n database is created. Regardless of the specific character set, the\\n character with code zero (sometimes called NUL) cannot be stored.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 08:52:01 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Invalid Unicode escape value at or near \\\"\\\\u0000\\\"\"\n },\n {\n \"msg_contents\": \"\\nOn Sat, 13 Nov 2021 at 21:52, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> On 11/13/21 00:40, Thomas Munro wrote:\\n>> On Sat, Nov 13, 2021 at 4:32 PM Japin Li <japinli@hotmail.com> wrote:\\n>>> When I try to insert an Unicode \\\"\\\\u0000\\\", there is an error $subject.\\n>>>\\n>>> postgres=# CREATE TABLE tbl (s varchar(10));\\n>>> CREATE TABLE\\n>>> postgres=# INSERT INTO tbl VALUES (E'\\\\u0000');\\n>>> ERROR: invalid Unicode escape value at or near \\\"\\\\u0000\\\"\\n>>> LINE 1: INSERT INTO tbl VALUES (E'\\\\u0000');\\n>>> ^\\n>>>\\n>>> \\\"\\\\u0000\\\" is valid unicode [1], why not we cannot insert it?\\n>> Yes, it is a valid codepoint, but unfortunately PostgreSQL can't\\n>> support it because it sometimes deals in null terminated string, even\\n>> though internally it does track string data and length separately. We\\n>> have to do that to use libc facilities like strcoll_r(), and probably\\n>> many other things.\\n>>\\n>>\\n>\\n> And it's documented at\\n> <https://www.postgresql.org/docs/current/datatype-character.html>:\\n>\\n>\\n> The characters that can be stored in any of these data types are\\n> determined by the database character set, which is selected when the\\n> database is created. Regardless of the specific character set, the\\n> character with code zero (sometimes called NUL) cannot be stored.\\n>\\n>\\n\\nThanks Thomas and Andrew. Sorry for my ignore reading of the documentation.\\n\\n-- \\nRegrads,\\nJapin Li.\\nChengDu WenWu Information Technology Co.,Ltd.\\n\\n\\n\",\n \"msg_date\": \"Sat, 13 Nov 2021 22:23:31 +0800\",\n \"msg_from\": \"Japin Li <japinli@hotmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Invalid Unicode escape value at or near \\\"\\\\u0000\\\"\"\n },\n {\n \"msg_contents\": \"On 13.11.21 06:40, Thomas Munro wrote:\\n> Yes, it is a valid codepoint, but unfortunately PostgreSQL can't\\n> support it because it sometimes deals in null terminated string, even\\n> though internally it does track string data and length separately. We\\n> have to do that to use libc facilities like strcoll_r(), and probably\\n> many other things.\\n\\nBy the way, I think the server-side issues around this are fixable. The \\nreal problem in my mind is how to present such values in client APIs \\nsuch as libpq.\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 09:06:34 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Invalid Unicode escape value at or near \\\"\\\\u0000\\\"\"\n }\n]","isConversational":false}}},{"rowIdx":53087,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nMy name is Deng Kaisheng, a graduate student in National University of Sinapore (NUS) now major in computer science. I've read the introduction of your organization and I'm quite interested in what you're doing.\n\nI want to do something and contribute to the community, and I wonder if there is something I can do. I major in computer science since undergraduate with a solid foundation in CS, and I'm also confident in my fast-learning skills. I have programming skills in C++, Python, Java, Objective-C, etc., and database skills in MySQL and Cassandra.\n\nI wonder if your organization is going to take part in the GSoC 2022, if possible, I would like to contribute to your project in GSOC 2022.\n\nIf you are interested in knowing more about me, please see my attached resume. I am looking forward to your reply!\n\nBest regards,\n\nKaisheng","msg_date":"Sat, 13 Nov 2021 08:41:34 +0000","msg_from":"Deng Kaisheng <kaisheng.deng@u.nus.edu>","msg_from_op":true,"msg_subject":"Anything I can contribute?"},{"msg_contents":"Hi Kaisheng,\n\nGlad that you are interested in contributing to postgres. There are\nmany ways one can contribute such as authoring a patch, reviewing\nothers' patches, documentation, performance testing etc. If you would\nlike to start with reviewing the pn-going features and their patches,\nthe list is here in the commitfest site -\nhttps://commitfest.postgresql.org/. The on-going commitfest is\nhttps://commitfest.postgresql.org/35/ and the up-coming one is\nhttps://commitfest.postgresql.org/36/. Feel free to pick any feature\nthat interests you, start reviewing the patches and post your comments\nin the hackers list.\n\nRegards,\nBharath Rupireddy.\n\nOn Tue, Nov 16, 2021 at 4:35 AM Deng Kaisheng <kaisheng.deng@u.nus.edu> wrote:\n>\n> Hi,\n>\n> My name is Deng Kaisheng, a graduate student in National University of Sinapore (NUS) now major in computer science. I've read the introduction of your organization and I'm quite interested in what you're doing.\n>\n> I want to do something and contribute to the community, and I wonder if there is something I can do. I major in computer science since undergraduate with a solid foundation in CS, and I'm also confident in my fast-learning skills. I have programming skills in C++, Python, Java, Objective-C, etc., and database skills in MySQL and Cassandra.\n>\n> I wonder if your organization is going to take part in the GSoC 2022, if possible, I would like to contribute to your project in GSOC 2022.\n>\n> If you are interested in knowing more about me, please see my attached resume. I am looking forward to your reply!\n>\n> Best regards,\n>\n> Kaisheng\n>\n>\n\n\n","msg_date":"Tue, 16 Nov 2021 08:12:57 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":false,"msg_subject":"Re: Anything I can contribute?"},{"msg_contents":"Hi Deng,\n\n> I want to do something and contribute to the community, and I wonder if\nthere is something I can do.\n\nI wrote a complete article on this topic not a long time ago [1].\nHopefully, no one will object if I link it here.\n\nGood luck and looking forward to your contribution!\n\n[1]:\nhttps://blog.timescale.com/blog/how-and-why-to-become-a-postgresql-contributor/\n\n-- \nBest regards,\nAleksander Alekseev\n\nHi Deng,> I want to do something and contribute to the community, and I wonder if there is something I can do.I wrote a complete article on this topic not a long time ago [1]. Hopefully, no one will object if I link it here.Good luck and looking forward to your contribution![1]: https://blog.timescale.com/blog/how-and-why-to-become-a-postgresql-contributor/-- Best regards,Aleksander Alekseev","msg_date":"Wed, 17 Nov 2021 11:05:41 +0300","msg_from":"Aleksander Alekseev <aleksander@timescale.com>","msg_from_op":false,"msg_subject":"Re: Anything I can contribute?"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nMy name is Deng Kaisheng, a graduate student in National University of Sinapore (NUS) now major in computer science. I've read the introduction of your organization and I'm quite interested in what you're doing.\\n\\nI want to do something and contribute to the community, and I wonder if there is something I can do. I major in computer science since undergraduate with a solid foundation in CS, and I'm also confident in my fast-learning skills. I have programming skills in C++, Python, Java, Objective-C, etc., and database skills in MySQL and Cassandra.\\n\\nI wonder if your organization is going to take part in the GSoC 2022, if possible, I would like to contribute to your project in GSOC 2022.\\n\\nIf you are interested in knowing more about me, please see my attached resume. I am looking forward to your reply!\\n\\nBest regards,\\n\\nKaisheng\",\n \"msg_date\": \"Sat, 13 Nov 2021 08:41:34 +0000\",\n \"msg_from\": \"Deng Kaisheng <kaisheng.deng@u.nus.edu>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Anything I can contribute?\"\n },\n {\n \"msg_contents\": \"Hi Kaisheng,\\n\\nGlad that you are interested in contributing to postgres. There are\\nmany ways one can contribute such as authoring a patch, reviewing\\nothers' patches, documentation, performance testing etc. If you would\\nlike to start with reviewing the pn-going features and their patches,\\nthe list is here in the commitfest site -\\nhttps://commitfest.postgresql.org/. The on-going commitfest is\\nhttps://commitfest.postgresql.org/35/ and the up-coming one is\\nhttps://commitfest.postgresql.org/36/. Feel free to pick any feature\\nthat interests you, start reviewing the patches and post your comments\\nin the hackers list.\\n\\nRegards,\\nBharath Rupireddy.\\n\\nOn Tue, Nov 16, 2021 at 4:35 AM Deng Kaisheng <kaisheng.deng@u.nus.edu> wrote:\\n>\\n> Hi,\\n>\\n> My name is Deng Kaisheng, a graduate student in National University of Sinapore (NUS) now major in computer science. I've read the introduction of your organization and I'm quite interested in what you're doing.\\n>\\n> I want to do something and contribute to the community, and I wonder if there is something I can do. I major in computer science since undergraduate with a solid foundation in CS, and I'm also confident in my fast-learning skills. I have programming skills in C++, Python, Java, Objective-C, etc., and database skills in MySQL and Cassandra.\\n>\\n> I wonder if your organization is going to take part in the GSoC 2022, if possible, I would like to contribute to your project in GSOC 2022.\\n>\\n> If you are interested in knowing more about me, please see my attached resume. I am looking forward to your reply!\\n>\\n> Best regards,\\n>\\n> Kaisheng\\n>\\n>\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 08:12:57 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Anything I can contribute?\"\n },\n {\n \"msg_contents\": \"Hi Deng,\\n\\n> I want to do something and contribute to the community, and I wonder if\\nthere is something I can do.\\n\\nI wrote a complete article on this topic not a long time ago [1].\\nHopefully, no one will object if I link it here.\\n\\nGood luck and looking forward to your contribution!\\n\\n[1]:\\nhttps://blog.timescale.com/blog/how-and-why-to-become-a-postgresql-contributor/\\n\\n-- \\nBest regards,\\nAleksander Alekseev\\n\\nHi Deng,> I want to do something and contribute to the community, and I wonder if there is something I can do.I wrote a complete article on this topic not a long time ago [1]. Hopefully, no one will object if I link it here.Good luck and looking forward to your contribution![1]: https://blog.timescale.com/blog/how-and-why-to-become-a-postgresql-contributor/-- Best regards,Aleksander Alekseev\",\n \"msg_date\": \"Wed, 17 Nov 2021 11:05:41 +0300\",\n \"msg_from\": \"Aleksander Alekseev <aleksander@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Anything I can contribute?\"\n }\n]","isConversational":false}}},{"rowIdx":53088,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi hackers,\n\nIf wrong GUCs of auth_delay, pg_trgm, postgres_fdw and sepgsql are \ndescribed in postgresql.conf, a warning is not emitted unlike \npg_stat_statements, auto_explain and pg_prewarm.\nSo, I improved it by adding EmitWarningsOnPlaceholders.\nAn example output is shown below.\n---\n2021-11-14 18:18:16.486 JST [487067] WARNING: unrecognized \nconfiguration parameter \"auth_delay.xxx\"\n---\n\nWhat do you think?\n\n-- \nRegards,\n\n--\nShinya Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION","msg_date":"Sun, 14 Nov 2021 19:03:59 +0900","msg_from":"Shinya Kato <Shinya11.Kato@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"> On 14 Nov 2021, at 11:03, Shinya Kato <Shinya11.Kato@oss.nttdata.com> wrote:\n\n> If wrong GUCs of auth_delay, pg_trgm, postgres_fdw and sepgsql are described in postgresql.conf, a warning is not emitted unlike pg_stat_statements, auto_explain and pg_prewarm.\n> So, I improved it by adding EmitWarningsOnPlaceholders.\n> An example output is shown below.\n> ---\n> 2021-11-14 18:18:16.486 JST [487067] WARNING: unrecognized configuration parameter \"auth_delay.xxx\"\n> ---\n> \n> What do you think?\n\nSeems reasonable on a quick skim, commit da2c1b8a2 did a similar roundup back\nin 2009 but at the time most of these didn't exist (pg_trgm did but didn't have\ncustom option back then). There is one additional callsite defining custom\nvariables in src/pl/tcl/pltcl.c which probably should get this treatment as\nwell, it would align it with the pl/perl counterpart.\n\nI'll have a closer look and test tomorrow.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Sun, 14 Nov 2021 20:50:56 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"On 2021-11-15 04:50, Daniel Gustafsson wrote:\n> Seems reasonable on a quick skim, commit da2c1b8a2 did a similar \n> roundup back\n> in 2009 but at the time most of these didn't exist (pg_trgm did but \n> didn't have\n> custom option back then). There is one additional callsite defining \n> custom\n> variables in src/pl/tcl/pltcl.c which probably should get this \n> treatment as\n> well, it would align it with the pl/perl counterpart.\n> \n> I'll have a closer look and test tomorrow.\nThank you for the review!\nI have missed src/pl/tcl/pltcl.c, so I created the new patch.\n\n-- \nRegards,\n\n--\nShinya Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION","msg_date":"Mon, 15 Nov 2021 10:03:22 +0900","msg_from":"Shinya Kato <Shinya11.Kato@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"On Mon, Nov 15, 2021 at 6:33 AM Shinya Kato\n<Shinya11.Kato@oss.nttdata.com> wrote:\n>\n> On 2021-11-15 04:50, Daniel Gustafsson wrote:\n> > Seems reasonable on a quick skim, commit da2c1b8a2 did a similar\n> > roundup back\n> > in 2009 but at the time most of these didn't exist (pg_trgm did but\n> > didn't have\n> > custom option back then). There is one additional callsite defining\n> > custom\n> > variables in src/pl/tcl/pltcl.c which probably should get this\n> > treatment as\n> > well, it would align it with the pl/perl counterpart.\n> >\n> > I'll have a closer look and test tomorrow.\n> Thank you for the review!\n> I have missed src/pl/tcl/pltcl.c, so I created the new patch.\n\nDo we need to add them in the following too?\n\ndelay_execution/delay_execution.c\nssl_passphrase_func.c\nworker_spi.c\n\nEspecially, worker_spi.c is important as it works as a template for\nthe bg worker code.\n\nI'm not sure if we have \"how to create an extension/a bg worker?\" in\nthe core docs, if we have, it is a good idea to make note of using\nEmitWarningsOnPlaceholders so that it will not be missed in the future\nmodules.\n\nI observed an odd behaviour:\n1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\n2) With EmitWarningsOnPlaceholders(\"postgres_fdw\"); in postgres_fdw\ncontrib module, I created the extension, have seen the following\nwarning:\n2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\nconfiguration parameter \"postgres_fdw.XXX\"\n3) I further did, \"alter system set\npostgres_fdw.XXX='I_further_messed_up_conf_file';\" and also \"select\npg_reload_conf();\", it silently accepts.\n\npostgres=# create extension postgres_fdw ;\nWARNING: unrecognized configuration parameter \"postgres_fdw.XXX\"\nCREATE EXTENSION\npostgres=# alter system set postgres_fdw.XXX='I_further_messed_up_conf_file';\nALTER SYSTEM\npostgres=# select pg_reload_conf();\n pg_reload_conf\n----------------\n t\n(1 row)\n\nMy point is, why should the \"create extension\" succeed with a WARNING\nwhen a wrong parameter related to it is set in the postgresql.conf\nfile so that we don't see further problems as shown in (3). I think\nEmitWarningsOnPlaceholders should emit an error so that the extension\nwill not get created in the first place if a wrong configuration\nparameter is set in the postgresql.conf file. Many times, users will\nnot have access to server logs so it is good to fail the \"create\nextension\" statement.\n\nThoughts?\n\npostgres=# create extension postgres_fdw ;\nERROR: unrecognized configuration parameter \"postgres_fdw.XXX\"\npostgres=#\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Mon, 15 Nov 2021 11:44:55 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"On 2021-11-15 15:14, Bharath Rupireddy wrote:\n\n> Do we need to add them in the following too?\n> \n> delay_execution/delay_execution.c\n> ssl_passphrase_func.c\n> worker_spi.c\n> \n> Especially, worker_spi.c is important as it works as a template for\n> the bg worker code.\n\nThank you for pointing this out! I added it.\n\n> I'm not sure if we have \"how to create an extension/a bg worker?\" in\n> the core docs, if we have, it is a good idea to make note of using\n> EmitWarningsOnPlaceholders so that it will not be missed in the future\n> modules.\n\nI cannot find any such docs, so I add a comment to \nsrc/backend/utils/misc/guc.c.\n\n> I observed an odd behaviour:\n> 1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\n> 2) With EmitWarningsOnPlaceholders(\"postgres_fdw\"); in postgres_fdw\n> contrib module, I created the extension, have seen the following\n> warning:\n> 2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\n> configuration parameter \"postgres_fdw.XXX\"\n> 3) I further did, \"alter system set\n> postgres_fdw.XXX='I_further_messed_up_conf_file';\" and also \"select\n> pg_reload_conf();\", it silently accepts.\n> \n> postgres=# create extension postgres_fdw ;\n> WARNING: unrecognized configuration parameter \"postgres_fdw.XXX\"\n> CREATE EXTENSION\n> postgres=# alter system set \n> postgres_fdw.XXX='I_further_messed_up_conf_file';\n> ALTER SYSTEM\n> postgres=# select pg_reload_conf();\n> pg_reload_conf\n> ----------------\n> t\n> (1 row)\n> \n> My point is, why should the \"create extension\" succeed with a WARNING\n> when a wrong parameter related to it is set in the postgresql.conf\n> file so that we don't see further problems as shown in (3). I think\n> EmitWarningsOnPlaceholders should emit an error so that the extension\n> will not get created in the first place if a wrong configuration\n> parameter is set in the postgresql.conf file. Many times, users will\n> not have access to server logs so it is good to fail the \"create\n> extension\" statement.\n> \n> Thoughts?\n> \n> postgres=# create extension postgres_fdw ;\n> ERROR: unrecognized configuration parameter \"postgres_fdw.XXX\"\n> postgres=#\n\nI confirmed it too, and I think that's definitely a problem.\nI also thought it would be a good idea to emit an ERROR as well as when \nan invalid normal GUC is set.\nI didn't change the function name because it would affect many third \nparty extensions.\n\nI plan to change to emit an error when an invalid custom GUC is set in \nthe SET or ALTER SYSTEM SET commands, but I haven't tackled this yet.\nThe patch as of now is attached.\n\n\n-- \nRegards,\n\n--\nShinya Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION","msg_date":"Tue, 16 Nov 2021 18:20:33 +0900","msg_from":"Shinya Kato <Shinya11.Kato@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"On Tue, Nov 16, 2021 at 2:50 PM Shinya Kato\n<Shinya11.Kato@oss.nttdata.com> wrote:\n> > I'm not sure if we have \"how to create an extension/a bg worker?\" in\n> > the core docs, if we have, it is a good idea to make note of using\n> > EmitWarningsOnPlaceholders so that it will not be missed in the future\n> > modules.\n>\n> I cannot find any such docs, so I add a comment to\n> src/backend/utils/misc/guc.c.\n>\n> > I observed an odd behaviour:\n> > 1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\n> > 2) With EmitWarningsOnPlaceholders(\"postgres_fdw\"); in postgres_fdw\n> > contrib module, I created the extension, have seen the following\n> > warning:\n> > 2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\n> > configuration parameter \"postgres_fdw.XXX\"\n> > 3) I further did, \"alter system set\n> > postgres_fdw.XXX='I_further_messed_up_conf_file';\" and also \"select\n> > pg_reload_conf();\", it silently accepts.\n> >\n> > postgres=# create extension postgres_fdw ;\n> > WARNING: unrecognized configuration parameter \"postgres_fdw.XXX\"\n> > CREATE EXTENSION\n> > postgres=# alter system set\n> > postgres_fdw.XXX='I_further_messed_up_conf_file';\n> > ALTER SYSTEM\n> > postgres=# select pg_reload_conf();\n> > pg_reload_conf\n> > ----------------\n> > t\n> > (1 row)\n> >\n> > My point is, why should the \"create extension\" succeed with a WARNING\n> > when a wrong parameter related to it is set in the postgresql.conf\n> > file so that we don't see further problems as shown in (3). I think\n> > EmitWarningsOnPlaceholders should emit an error so that the extension\n> > will not get created in the first place if a wrong configuration\n> > parameter is set in the postgresql.conf file. Many times, users will\n> > not have access to server logs so it is good to fail the \"create\n> > extension\" statement.\n> >\n> > Thoughts?\n> >\n> > postgres=# create extension postgres_fdw ;\n> > ERROR: unrecognized configuration parameter \"postgres_fdw.XXX\"\n> > postgres=#\n>\n> I confirmed it too, and I think that's definitely a problem.\n> I also thought it would be a good idea to emit an ERROR as well as when\n> an invalid normal GUC is set.\n> I didn't change the function name because it would affect many third\n> party extensions.\n\nFor backward compatibility you can retain the function\nEmitWarningsOnPlaceholders as-is and have another similar function\nthat emits an error:\n\nIn guc.c, have something like below:\nstatic void\ncheck_conf_params(const char *className, bool emit_error)\n{\n /* have the existing code of EmitWarningsOnPlaceholders here */\n ereport(emit_error ? ERROR : WARNING,\n (errcode(ERRCODE_UNDEFINED_OBJECT),\n errmsg(\"unrecognized configuration parameter \\\"%s\\\"\",\n var->name)));\n}\n\nvoid\nEmitErrorOnPlaceholders(const char *className)\n{\n check_conf_params(className, true);\n}\n\nvoid\nEmitWarningsOnPlaceholders(const char *className)\n{\n check_conf_params(className, false);\n}\n\nAnd change all the core extensions to use EmitErrorOnPlaceholders.\nThis way you don't break the backward compatibility of the outside\nextensions, if they want they get to choose which behaviour they want\nfor their extension.\n\nActually, I didn't quite like the function name\nEmitWarningsOnPlaceholders or EmitErrorOnPlaceholders, it could have\nbeen something else. But let's not go that direction of changing the\nfunction name for backward compatibility.\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Tue, 16 Nov 2021 15:55:13 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"Thank you for the review and sorry for the late reply.\n\nOn 2021-11-16 19:25, Bharath Rupireddy wrote:\n>> > I observed an odd behaviour:\n>> > 1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\n>> > 2) With EmitWarningsOnPlaceholders(\"postgres_fdw\"); in postgres_fdw\n>> > contrib module, I created the extension, have seen the following\n>> > warning:\n>> > 2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\n>> > configuration parameter \"postgres_fdw.XXX\"\n>> > 3) I further did, \"alter system set\n>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\" and also \"select\n>> > pg_reload_conf();\", it silently accepts.\n>> >\n>> > postgres=# create extension postgres_fdw ;\n>> > WARNING: unrecognized configuration parameter \"postgres_fdw.XXX\"\n>> > CREATE EXTENSION\n>> > postgres=# alter system set\n>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\n>> > ALTER SYSTEM\n>> > postgres=# select pg_reload_conf();\n>> > pg_reload_conf\n>> > ----------------\n>> > t\n>> > (1 row)\n\nI have made changes to achieve the above.\nHowever, it also gives me an error when\n---\npostgres=# SET abc.a TO on;\nSET\npostgres=# SET abc.b TO on;\n2021-12-16 16:22:20.351 JST [2489236] ERROR: unrecognized configuration \nparameter \"abc.a\"\n2021-12-16 16:22:20.351 JST [2489236] STATEMENT: SET abc.b TO on;\nERROR: unrecognized configuration parameter \"abc.a\"\n---\n\nI know that some people do not think this is good.\nTherefore, can I leave this problem alone or is there another better \nway?\n\n\n> For backward compatibility you can retain the function\n> EmitWarningsOnPlaceholders as-is and have another similar function\n> that emits an error:\n> \n> In guc.c, have something like below:\n> static void\n> check_conf_params(const char *className, bool emit_error)\n> {\n> /* have the existing code of EmitWarningsOnPlaceholders here */\n> ereport(emit_error ? ERROR : WARNING,\n> (errcode(ERRCODE_UNDEFINED_OBJECT),\n> errmsg(\"unrecognized configuration parameter \n> \\\"%s\\\"\",\n> var->name)));\n> }\n> \n> void\n> EmitErrorOnPlaceholders(const char *className)\n> {\n> check_conf_params(className, true);\n> }\n> \n> void\n> EmitWarningsOnPlaceholders(const char *className)\n> {\n> check_conf_params(className, false);\n> }\n> \n\nThank you for your advise.\nAccording to [1], we used the same function name, but the warning level \nwas INFO.\nTherefore, I think it is OK to use the same function name.\n\n[1] \nhttps://www.postgresql.org/message-id/flat/200901051634.n05GYNr06169%40momjian.us#1d045374f014494e4b40a4862a000723\n\n-- \nRegards,\n\n--\nShinya Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION","msg_date":"Thu, 16 Dec 2021 16:31:03 +0900","msg_from":"Shinya Kato <Shinya11.Kato@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"\n\nOn 2021/12/16 16:31, Shinya Kato wrote:\n> Thank you for the review and sorry for the late reply.\n> \n> On 2021-11-16 19:25, Bharath Rupireddy wrote:\n>>> > I observed an odd behaviour:\n>>> > 1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\n>>> > 2) With EmitWarningsOnPlaceholders(\"postgres_fdw\"); in postgres_fdw\n>>> > contrib module, I created the extension, have seen the following\n>>> > warning:\n>>> > 2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\n>>> > configuration parameter \"postgres_fdw.XXX\"\n>>> > 3) I further did, \"alter system set\n>>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\" and also \"select\n>>> > pg_reload_conf();\", it silently accepts.\n>>> >\n>>> > postgres=# create extension postgres_fdw ;\n>>> > WARNING: unrecognized configuration parameter \"postgres_fdw.XXX\"\n>>> > CREATE EXTENSION\n>>> > postgres=# alter system set\n>>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\n>>> > ALTER SYSTEM\n>>> > postgres=# select pg_reload_conf();\n>>> > pg_reload_conf\n>>> > ----------------\n>>> > t\n>>> > (1 row)\n> \n> I have made changes to achieve the above.\n\nIMO this behavior change is not good. For example, because it seems to break at least the following case. I think that these are the valid steps, but with the patch, the server fails to start up at the step #2 because pg_trgm's custom parameters are treated as invalid ones.\n\n1. Add the following two pg_trgm parameters to postgresql.conf\n - pg_trgm.similarity_threshold\n - pg_trgm.strict_word_similarity_threshold\n\n2. Start the server\n\n3. Run \"CREATE EXTENSION pg_trgm\" and then use pg_trgm\n\n\nWhen I compiled PostgreSQL with the patch, I got the following\ncompilation failure.\n\nguc.c:5453:4: error: implicit declaration of function 'EmitErrorsOnPlaceholders' is invalid in C99 [-Werror,-Wimplicit-function-declaration]\n EmitErrorsOnPlaceholders(placeholder);\n ^\n\n\n-\t\t\tereport(WARNING,\n+\t\t\tereport(ERROR,\n\nI'm still not sure why you were thinking that ERROR is more proper here.\n\nRegards,\n\n-- \nFujii Masao\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Fri, 17 Dec 2021 01:55:19 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"On 2021-12-17 01:55, Fujii Masao wrote:\n> On 2021/12/16 16:31, Shinya Kato wrote:\n>> Thank you for the review and sorry for the late reply.\n>> \n>> On 2021-11-16 19:25, Bharath Rupireddy wrote:\n>>>> > I observed an odd behaviour:\n>>>> > 1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\n>>>> > 2) With EmitWarningsOnPlaceholders(\"postgres_fdw\"); in postgres_fdw\n>>>> > contrib module, I created the extension, have seen the following\n>>>> > warning:\n>>>> > 2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\n>>>> > configuration parameter \"postgres_fdw.XXX\"\n>>>> > 3) I further did, \"alter system set\n>>>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\" and also \"select\n>>>> > pg_reload_conf();\", it silently accepts.\n>>>> >\n>>>> > postgres=# create extension postgres_fdw ;\n>>>> > WARNING: unrecognized configuration parameter \"postgres_fdw.XXX\"\n>>>> > CREATE EXTENSION\n>>>> > postgres=# alter system set\n>>>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\n>>>> > ALTER SYSTEM\n>>>> > postgres=# select pg_reload_conf();\n>>>> > pg_reload_conf\n>>>> > ----------------\n>>>> > t\n>>>> > (1 row)\n>> \n>> I have made changes to achieve the above.\n> \n> IMO this behavior change is not good. For example, because it seems to\n> break at least the following case. I think that these are the valid\n> steps, but with the patch, the server fails to start up at the step #2\n> because pg_trgm's custom parameters are treated as invalid ones.\n> \n> 1. Add the following two pg_trgm parameters to postgresql.conf\n> - pg_trgm.similarity_threshold\n> - pg_trgm.strict_word_similarity_threshold\n> \n> 2. Start the server\n> \n> 3. Run \"CREATE EXTENSION pg_trgm\" and then use pg_trgm\n\nIt can happen because the placeholder \"pg_trgm\" is not already \nregistered.\nI think too this behavior is not good.\nI need to consider an another implementation or to allow undefined GUCs \nto be set.\n\n\n> When I compiled PostgreSQL with the patch, I got the following\n> compilation failure.\n> \n> guc.c:5453:4: error: implicit declaration of function\n> 'EmitErrorsOnPlaceholders' is invalid in C99\n> [-Werror,-Wimplicit-function-declaration]\n> EmitErrorsOnPlaceholders(placeholder);\n> ^\n> \n> \n> -\t\t\tereport(WARNING,\n> +\t\t\tereport(ERROR,\n\nThanks! There was a correction omission, so I fixed it.\n\n\n> I'm still not sure why you were thinking that ERROR is more proper \n> here.\n\nSince I get an ERROR when I set the wrong normal GUCs, I thought I \nshould also get an ERROR when I set the wrong extension's GUCs.\nHowever, there are likely to be harmful effects, and I may need to think \nagain.\n\n\nFor now, I'v attached the patch that fixed the compilation error.\n\n-- \nRegards,\n\n--\nShinya Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION","msg_date":"Fri, 17 Dec 2021 11:25:22 +0900","msg_from":"Shinya Kato <Shinya11.Kato@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"\n\nOn 2021/12/17 11:25, Shinya Kato wrote:\n> For now, I'v attached the patch that fixed the compilation error.\n\nThanks for updating the patch! I confirmed that the compilation was completed successfully with this new patch. But the regression test failed. You can see that Patch Tester [1] also reported the failure of regression test for your patch.\n\n[1]\nhttp://commitfest.cputube.org/\n\nRegards,\n\n-- \nFujii Masao\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Fri, 17 Dec 2021 14:00:27 +0900","msg_from":"Fujii Masao <masao.fujii@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"On 17.12.21 03:25, Shinya Kato wrote:\n> For now, I'v attached the patch that fixed the compilation error.\n\nI think it would be good if you could split the uncontroversial new \nEmitErrorsOnPlaceholders() calls into a separate patch. And please add \nan explanation what exactly the rest of the patch changes.\n\n\n","msg_date":"Fri, 17 Dec 2021 07:42:58 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"On 2021-12-17 15:42, Peter Eisentraut wrote:\n> On 17.12.21 03:25, Shinya Kato wrote:\n>> For now, I'v attached the patch that fixed the compilation error.\n> \n> I think it would be good if you could split the uncontroversial new\n> EmitErrorsOnPlaceholders() calls into a separate patch. And please\n> add an explanation what exactly the rest of the patch changes.\n\nThank you for the comment!\nI splitted the patch.\n\n- v6-01-Add-EmitWarningsOnPlaceholders.patch\nWe should use EmitWarningsOnPlaceholders when we use \nDefineCustomXXXVariable.\nI don't think there is any room for debate.\n\n- v6-0002-Change-error-level-of-EmitWarningsOnPlaceholders.patch\nThis is a patch to change the error level of EmitWarningsOnPlaceholders \nfrom WARNING to ERROR.\nI think it's OK to emit ERROR as well as when the wrong GUC is set for \nnon-extensions, or since it does not behave abnormally, it can be left \nas WARNING.\nThought?\n\n- v6-0003-Emit-error-when-invalid-extensions-GUCs-are-set.patch\nThis is a patch to emit error when invalid extension's GUCs are set.\nNo test changes have been made, so the regression test will fail.\nI have created a patch, but I don't think this is necessary because of \nthe previous discussion.\n\n-- \nRegards,\n\n--\nShinya Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION","msg_date":"Mon, 20 Dec 2021 21:05:23 +0900","msg_from":"Shinya Kato <Shinya11.Kato@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"At Mon, 20 Dec 2021 21:05:23 +0900, Shinya Kato <Shinya11.Kato@oss.nttdata.com> wrote in \n> - v6-01-Add-EmitWarningsOnPlaceholders.patch\n> We should use EmitWarningsOnPlaceholders when we use\n> DefineCustomXXXVariable.\n> I don't think there is any room for debate.\n\nUnfortunately, pltcl.c defines variables both in pltcl and pltclu but\nthe patch adds the warning only for pltclu.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Tue, 21 Dec 2021 10:08:34 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"At Fri, 17 Dec 2021 11:25:22 +0900, Shinya Kato <Shinya11.Kato@oss.nttdata.com> wrote in \n> On 2021-12-17 01:55, Fujii Masao wrote:\n> > I'm still not sure why you were thinking that ERROR is more proper\n> > here.\n> \n> Since I get an ERROR when I set the wrong normal GUCs, I thought I\n> should also get an ERROR when I set the wrong extension's GUCs.\n> However, there are likely to be harmful effects, and I may need to\n> think again.\n\nThe GUC placeholder mechanism is evidently designed so that server\nallows to define variables unknown to the server before loading\nextension modules. ERRORing out that variables at the timing is\ncontradicting the objective for the placeholder mechanism.\n\nAddition to that EmitWarningsOnPlaceholders()'s objective is to warn\nfor variables that shouldn't be of a *known* namespace. However, the\npatch is intending to check out variable definitions of all namespaces\nthat are seen in configuration file, but it is undeterminable at that\ntime whether each of the namespaces is either just wrong or unknown\nyet. And the latter is, as mentioned above, what we are intending to\nallow.\n\nAs the concusion, I think we cannot rule-out \"wrong\" namespaces unless\nwe find any means to distinguish whether a namespace is wrong or\nnot-yet-defined before loading extension modules.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Tue, 21 Dec 2021 10:50:07 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"Kyotaro Horiguchi <horikyota.ntt@gmail.com> writes:\n> At Mon, 20 Dec 2021 21:05:23 +0900, Shinya Kato <Shinya11.Kato@oss.nttdata.com> wrote in \n>> We should use EmitWarningsOnPlaceholders when we use\n>> DefineCustomXXXVariable.\n>> I don't think there is any room for debate.\n\n> Unfortunately, pltcl.c defines variables both in pltcl and pltclu but\n> the patch adds the warning only for pltclu.\n\nRight. The rest of 0001 looks fine, so pushed with that correction.\n\nI concur that 0002 is unacceptable. The result of it will be that\na freshly-loaded extension will fail to hook into the system as it\nshould, because it will fail to complete its initialization.\nThat will cause user frustration and bug reports.\n\n(BTW, I wonder if EmitWarningsOnPlaceholders should be using LOG\nrather than WARNING when it's running in the postmaster. Use of\nWARNING is moderately likely to result in nothing getting printed\nat all.)\n\n0003 looks to me like it was superseded by 75d22069e. I do not\nparticularly like that patch though; it seems both inefficient\nand ill-designed. 0003 is adding a check in an equally bizarre\nplace. Why isn't add_placeholder_variable the place to deal\nwith this? Also, once we know that a prefix is reserved,\nwhy don't we throw an error not just a warning for attempts to\nset some unknown variable under that prefix? We don't allow\nyou to set unknown non-prefixed variables.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 21 Dec 2021 12:23:25 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"I wrote:\n> 0003 looks to me like it was superseded by 75d22069e. I do not\n> particularly like that patch though; it seems both inefficient\n> and ill-designed. 0003 is adding a check in an equally bizarre\n> place. Why isn't add_placeholder_variable the place to deal\n> with this? Also, once we know that a prefix is reserved,\n> why don't we throw an error not just a warning for attempts to\n> set some unknown variable under that prefix? We don't allow\n> you to set unknown non-prefixed variables.\n\nConcretely, I think we should do the attached, which removes much of\n75d22069e and does the check at the point of placeholder creation.\n(After looking closer, add_placeholder_variable's caller is the\nfunction that is responsible for rejecting bad names, not\nadd_placeholder_variable itself.)\n\nThis also fixes an indisputable bug in 75d22069e, namely that it\ndid prefix comparisons incorrectly and would throw warnings for\ncases it shouldn't. Reserving \"plpgsql\" shouldn't have the effect\nof reserving \"pl\" as well.\n\nI'm tempted to propose that we also rename EmitWarningsOnPlaceholders\nto something like MarkGUCPrefixReserved, to more clearly reflect\nwhat it does now. (We could provide the old name as a macro alias\nto avoid breaking extensions needlessly.)\n\n\t\t\tregards, tom lane","msg_date":"Tue, 21 Dec 2021 13:33:49 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"On 2021-12-22 02:23, Tom Lane wrote:\n> Kyotaro Horiguchi <horikyota.ntt@gmail.com> writes:\n>> At Mon, 20 Dec 2021 21:05:23 +0900, Shinya Kato \n>> <Shinya11.Kato@oss.nttdata.com> wrote in\n>>> We should use EmitWarningsOnPlaceholders when we use\n>>> DefineCustomXXXVariable.\n>>> I don't think there is any room for debate.\n> \n>> Unfortunately, pltcl.c defines variables both in pltcl and pltclu but\n>> the patch adds the warning only for pltclu.\n> \n> Right. The rest of 0001 looks fine, so pushed with that correction.\n> \n> I concur that 0002 is unacceptable. The result of it will be that\n> a freshly-loaded extension will fail to hook into the system as it\n> should, because it will fail to complete its initialization.\n> That will cause user frustration and bug reports.\n> \n> (BTW, I wonder if EmitWarningsOnPlaceholders should be using LOG\n> rather than WARNING when it's running in the postmaster. Use of\n> WARNING is moderately likely to result in nothing getting printed\n> at all.)\n> \n> 0003 looks to me like it was superseded by 75d22069e. I do not\n> particularly like that patch though; it seems both inefficient\n> and ill-designed. 0003 is adding a check in an equally bizarre\n> place. Why isn't add_placeholder_variable the place to deal\n> with this? Also, once we know that a prefix is reserved,\n> why don't we throw an error not just a warning for attempts to\n> set some unknown variable under that prefix? We don't allow\n> you to set unknown non-prefixed variables.\n> \n> \t\t\tregards, tom lane\n\nThank you for pushing!\nThank you all for the reviews, I think I will take down 0002 and 0003.\n\n-- \nRegards,\n\n--\nShinya Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Wed, 22 Dec 2021 16:50:34 +0900","msg_from":"Shinya Kato <Shinya11.Kato@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"On 2021/12/22 3:33, Tom Lane wrote:\n> I wrote:\n>> 0003 looks to me like it was superseded by 75d22069e. I do not\n>> particularly like that patch though; it seems both inefficient\n>> and ill-designed. 0003 is adding a check in an equally bizarre\n>> place. Why isn't add_placeholder_variable the place to deal\n>> with this? Also, once we know that a prefix is reserved,\n>> why don't we throw an error not just a warning for attempts to\n>> set some unknown variable under that prefix? We don't allow\n>> you to set unknown non-prefixed variables.\n> Concretely, I think we should do the attached, which removes much of\n> 75d22069e and does the check at the point of placeholder creation.\n> (After looking closer, add_placeholder_variable's caller is the\n> function that is responsible for rejecting bad names, not\n> add_placeholder_variable itself.)\n>\n> This also fixes an indisputable bug in 75d22069e, namely that it\n> did prefix comparisons incorrectly and would throw warnings for\n> cases it shouldn't. Reserving \"plpgsql\" shouldn't have the effect\n> of reserving \"pl\" as well.\n\nThank you for creating the patch.\nThis is exactly what I wanted to create. It looks good to me.\n\n\n> I'm tempted to propose that we also rename EmitWarningsOnPlaceholders\n> to something like MarkGUCPrefixReserved, to more clearly reflect\n> what it does now. (We could provide the old name as a macro alias\n> to avoid breaking extensions needlessly.)\n\n+1\n\n\n-- \nRegards,\n\n--\nShinya Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n\n\n","msg_date":"Fri, 24 Dec 2021 12:29:47 +0900","msg_from":"Shinya Kato <Shinya11.Kato@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"I wrote:\n> Concretely, I think we should do the attached, which removes much of\n> 75d22069e and does the check at the point of placeholder creation.\n\nI pushed that, and along the way moved the test case to be beside\nthe existing tests concerning custom GUC names, rather than appended\nat the end of guc.sql as it had been. That turns out to make the\nbuildfarm very unhappy. I had not thought to test that change with\n\"force_parallel_mode = regress\"; but with that on, we can see that\nthe warning (or now error) is reported each time a parallel worker\nis launched, if the parent process contains a bogus placeholder.\nSo that accidentally unveiled another deficiency in the design of\nthe original patch --- we surely don't want that to happen.\n\nAs a stopgap to turn the farm green again, I am going to revert\n75d22069e as well as my followup patches. If we don't want to\ngive up on that idea altogether, we have to find some way to\nsuppress the chatter from parallel workers. I wonder whether\nit would be appropriate to go further than we have, and actively\ndelete placeholders that turn out to be within an extension's\nreserved namespace. The core issue here is that workers don't\nnecessarily set GUCs and load extensions in the same order that\ntheir parent did, so if we leave any invalid placeholders behind\nafter reserving an extension's prefix, we're risking issues\nduring worker start.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 27 Dec 2021 15:48:43 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"I wrote:\n> As a stopgap to turn the farm green again, I am going to revert\n> 75d22069e as well as my followup patches. If we don't want to\n> give up on that idea altogether, we have to find some way to\n> suppress the chatter from parallel workers. I wonder whether\n> it would be appropriate to go further than we have, and actively\n> delete placeholders that turn out to be within an extension's\n> reserved namespace. The core issue here is that workers don't\n> necessarily set GUCs and load extensions in the same order that\n> their parent did, so if we leave any invalid placeholders behind\n> after reserving an extension's prefix, we're risking issues\n> during worker start.\n\nHere's a delta patch (meant to be applied after reverting cab5b9ab2)\nthat does things like that. It fixes the parallel-mode problem ...\nso do we want to tighten things up this much?\n\n\t\t\tregards, tom lane","msg_date":"Tue, 28 Dec 2021 12:06:55 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"Il giorno ven 18 feb 2022 alle ore 10:58 Tom Lane <tgl@sss.pgh.pa.us> ha\nscritto:\n\n> I wrote:\n> > As a stopgap to turn the farm green again, I am going to revert\n> > 75d22069e as well as my followup patches. If we don't want to\n> > give up on that idea altogether, we have to find some way to\n> > suppress the chatter from parallel workers. I wonder whether\n> > it would be appropriate to go further than we have, and actively\n> > delete placeholders that turn out to be within an extension's\n> > reserved namespace. The core issue here is that workers don't\n> > necessarily set GUCs and load extensions in the same order that\n> > their parent did, so if we leave any invalid placeholders behind\n> > after reserving an extension's prefix, we're risking issues\n> > during worker start.\n>\n> Here's a delta patch (meant to be applied after reverting cab5b9ab2)\n> that does things like that. It fixes the parallel-mode problem ...\n> so do we want to tighten things up this much?\n>\n> regards, tom lan\n>\n\nHello,\n\nThank you for taking care of the bug I introduced with 75d22069e,\nI didn't notice this thread until now.\n\nFor what it's worth, I agree with throwing an ERROR if the placeholder is\nunrecognized. Initially, I didn't want to change too much the liberty of\nsetting any\nplaceholder, but mainly to not go unnoticed.\n\nIn my humble opinion, I still think that this should go on and disallow\nbogus placeholders as we do for postgres unrecognized settings.\n\nI tested your delta patch with and without parallel mode, and, naturally,\nit behaves correctly.\n\nMy 2 cents.\n\n+1\n\nCheers,\nFlorin Irion\n\nIl giorno ven 18 feb 2022 alle ore 10:58 Tom Lane <tgl@sss.pgh.pa.us> ha scritto:I wrote:\n> As a stopgap to turn the farm green again, I am going to revert\n> 75d22069e as well as my followup patches. If we don't want to\n> give up on that idea altogether, we have to find some way to\n> suppress the chatter from parallel workers. I wonder whether\n> it would be appropriate to go further than we have, and actively\n> delete placeholders that turn out to be within an extension's\n> reserved namespace. The core issue here is that workers don't\n> necessarily set GUCs and load extensions in the same order that\n> their parent did, so if we leave any invalid placeholders behind\n> after reserving an extension's prefix, we're risking issues\n> during worker start.\n\nHere's a delta patch (meant to be applied after reverting cab5b9ab2)\nthat does things like that. It fixes the parallel-mode problem ...\nso do we want to tighten things up this much?\n\n regards, tom lanHello, Thank you for taking care of the bug I introduced with 75d22069e,I didn't notice this thread until now.For what it's worth, I agree with throwing an ERROR if the placeholder isunrecognized. Initially, I didn't want to change too much the liberty of setting anyplaceholder, but mainly to not go unnoticed.In my humble opinion, I still think that this should go on and disallow bogus placeholders as we do for postgres unrecognized settings. I tested your delta patch with and without parallel mode, and, naturally, it behaves correctly.My 2 cents.+1Cheers,Florin Irion","msg_date":"Fri, 18 Feb 2022 12:47:23 +0100","msg_from":"Florin Irion <irionr@gmail.com>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"Florin Irion <irionr@gmail.com> writes:\n> Il giorno ven 18 feb 2022 alle ore 10:58 Tom Lane <tgl@sss.pgh.pa.us> ha\n> scritto:\n>> Here's a delta patch (meant to be applied after reverting cab5b9ab2)\n>> that does things like that. It fixes the parallel-mode problem ...\n>> so do we want to tighten things up this much?\n\n> Thank you for taking care of the bug I introduced with 75d22069e,\n> I didn't notice this thread until now.\n\nYeah, this thread kinda disappeared under the rug during the Christmas\nholidays, but we need to decide whether we want to push forward or\nleave things at the status quo.\n\n> For what it's worth, I agree with throwing an ERROR if the placeholder is\n> unrecognized. Initially, I didn't want to change too much the liberty of\n> setting any placeholder, but mainly to not go unnoticed.\n\nI also think that this is probably a good change to make.\n\nOne situation where somebody would be unhappy is if they're using GUC\nvariables as plain custom variables despite them being within the\nnamespace of an extension they're using. But that seems to me to be\n(a) far-fetched and (b) mighty dangerous, since some new version of the\nextension might suddenly start reacting to those variables in ways you\npresumably didn't expect.\n\nPerhaps a more plausible use-case is \"I need to work with both versions\nX and Y of extension E, and Y has setting e.foo while X doesn't --- but\nI can just set e.foo unconditionally since X will ignore it\". With this\npatch, that could still work, but you'd have to be certain to apply the\nsetting before loading E.\n\nI don't really see any other use-cases favoring the current behavior.\nOn balance, eliminating possible mistakes seems like enough of a\nbenefit to justify disallowing these cases.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 18 Feb 2022 17:18:21 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"I wrote:\n> Florin Irion <irionr@gmail.com> writes:\n>> For what it's worth, I agree with throwing an ERROR if the placeholder is\n>> unrecognized. Initially, I didn't want to change too much the liberty of\n>> setting any placeholder, but mainly to not go unnoticed.\n\n> I also think that this is probably a good change to make.\n\nHearing no objections, done.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 21 Feb 2022 14:12:04 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"},{"msg_contents":"Il lun 21 feb 2022, 20:12 Tom Lane <tgl@sss.pgh.pa.us> ha scritto:\n\n> I wrote:\n> > Florin Irion <irionr@gmail.com> writes:\n> >> For what it's worth, I agree with throwing an ERROR if the placeholder\n> is\n> >> unrecognized. Initially, I didn't want to change too much the liberty of\n> >> setting any placeholder, but mainly to not go unnoticed.\n>\n> > I also think that this is probably a good change to make.\n>\n> Hearing no objections, done.\n>\n> regards, tom lane\n>\n\nThank you for taking care of this.\n\nCheers,\nFlorin\n\n>\n\nIl lun 21 feb 2022, 20:12 Tom Lane <tgl@sss.pgh.pa.us> ha scritto:I wrote:\n> Florin Irion <irionr@gmail.com> writes:\n>> For what it's worth, I agree with throwing an ERROR if the placeholder is\n>> unrecognized. Initially, I didn't want to change too much the liberty of\n>> setting any placeholder, but mainly to not go unnoticed.\n\n> I also think that this is probably a good change to make.\n\nHearing no objections, done.\n\n regards, tom laneThank you for taking care of this. Cheers, Florin","msg_date":"Mon, 21 Feb 2022 22:22:55 +0100","msg_from":"Florin Irion <irionr@gmail.com>","msg_from_op":false,"msg_subject":"Re: Emit a warning if the extension's GUC is set incorrectly"}],"string":"[\n {\n \"msg_contents\": \"Hi hackers,\\n\\nIf wrong GUCs of auth_delay, pg_trgm, postgres_fdw and sepgsql are \\ndescribed in postgresql.conf, a warning is not emitted unlike \\npg_stat_statements, auto_explain and pg_prewarm.\\nSo, I improved it by adding EmitWarningsOnPlaceholders.\\nAn example output is shown below.\\n---\\n2021-11-14 18:18:16.486 JST [487067] WARNING: unrecognized \\nconfiguration parameter \\\"auth_delay.xxx\\\"\\n---\\n\\nWhat do you think?\\n\\n-- \\nRegards,\\n\\n--\\nShinya Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Sun, 14 Nov 2021 19:03:59 +0900\",\n \"msg_from\": \"Shinya Kato <Shinya11.Kato@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"> On 14 Nov 2021, at 11:03, Shinya Kato <Shinya11.Kato@oss.nttdata.com> wrote:\\n\\n> If wrong GUCs of auth_delay, pg_trgm, postgres_fdw and sepgsql are described in postgresql.conf, a warning is not emitted unlike pg_stat_statements, auto_explain and pg_prewarm.\\n> So, I improved it by adding EmitWarningsOnPlaceholders.\\n> An example output is shown below.\\n> ---\\n> 2021-11-14 18:18:16.486 JST [487067] WARNING: unrecognized configuration parameter \\\"auth_delay.xxx\\\"\\n> ---\\n> \\n> What do you think?\\n\\nSeems reasonable on a quick skim, commit da2c1b8a2 did a similar roundup back\\nin 2009 but at the time most of these didn't exist (pg_trgm did but didn't have\\ncustom option back then). There is one additional callsite defining custom\\nvariables in src/pl/tcl/pltcl.c which probably should get this treatment as\\nwell, it would align it with the pl/perl counterpart.\\n\\nI'll have a closer look and test tomorrow.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Sun, 14 Nov 2021 20:50:56 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"On 2021-11-15 04:50, Daniel Gustafsson wrote:\\n> Seems reasonable on a quick skim, commit da2c1b8a2 did a similar \\n> roundup back\\n> in 2009 but at the time most of these didn't exist (pg_trgm did but \\n> didn't have\\n> custom option back then). There is one additional callsite defining \\n> custom\\n> variables in src/pl/tcl/pltcl.c which probably should get this \\n> treatment as\\n> well, it would align it with the pl/perl counterpart.\\n> \\n> I'll have a closer look and test tomorrow.\\nThank you for the review!\\nI have missed src/pl/tcl/pltcl.c, so I created the new patch.\\n\\n-- \\nRegards,\\n\\n--\\nShinya Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Mon, 15 Nov 2021 10:03:22 +0900\",\n \"msg_from\": \"Shinya Kato <Shinya11.Kato@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 6:33 AM Shinya Kato\\n<Shinya11.Kato@oss.nttdata.com> wrote:\\n>\\n> On 2021-11-15 04:50, Daniel Gustafsson wrote:\\n> > Seems reasonable on a quick skim, commit da2c1b8a2 did a similar\\n> > roundup back\\n> > in 2009 but at the time most of these didn't exist (pg_trgm did but\\n> > didn't have\\n> > custom option back then). There is one additional callsite defining\\n> > custom\\n> > variables in src/pl/tcl/pltcl.c which probably should get this\\n> > treatment as\\n> > well, it would align it with the pl/perl counterpart.\\n> >\\n> > I'll have a closer look and test tomorrow.\\n> Thank you for the review!\\n> I have missed src/pl/tcl/pltcl.c, so I created the new patch.\\n\\nDo we need to add them in the following too?\\n\\ndelay_execution/delay_execution.c\\nssl_passphrase_func.c\\nworker_spi.c\\n\\nEspecially, worker_spi.c is important as it works as a template for\\nthe bg worker code.\\n\\nI'm not sure if we have \\\"how to create an extension/a bg worker?\\\" in\\nthe core docs, if we have, it is a good idea to make note of using\\nEmitWarningsOnPlaceholders so that it will not be missed in the future\\nmodules.\\n\\nI observed an odd behaviour:\\n1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\\n2) With EmitWarningsOnPlaceholders(\\\"postgres_fdw\\\"); in postgres_fdw\\ncontrib module, I created the extension, have seen the following\\nwarning:\\n2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\\nconfiguration parameter \\\"postgres_fdw.XXX\\\"\\n3) I further did, \\\"alter system set\\npostgres_fdw.XXX='I_further_messed_up_conf_file';\\\" and also \\\"select\\npg_reload_conf();\\\", it silently accepts.\\n\\npostgres=# create extension postgres_fdw ;\\nWARNING: unrecognized configuration parameter \\\"postgres_fdw.XXX\\\"\\nCREATE EXTENSION\\npostgres=# alter system set postgres_fdw.XXX='I_further_messed_up_conf_file';\\nALTER SYSTEM\\npostgres=# select pg_reload_conf();\\n pg_reload_conf\\n----------------\\n t\\n(1 row)\\n\\nMy point is, why should the \\\"create extension\\\" succeed with a WARNING\\nwhen a wrong parameter related to it is set in the postgresql.conf\\nfile so that we don't see further problems as shown in (3). I think\\nEmitWarningsOnPlaceholders should emit an error so that the extension\\nwill not get created in the first place if a wrong configuration\\nparameter is set in the postgresql.conf file. Many times, users will\\nnot have access to server logs so it is good to fail the \\\"create\\nextension\\\" statement.\\n\\nThoughts?\\n\\npostgres=# create extension postgres_fdw ;\\nERROR: unrecognized configuration parameter \\\"postgres_fdw.XXX\\\"\\npostgres=#\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 11:44:55 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"On 2021-11-15 15:14, Bharath Rupireddy wrote:\\n\\n> Do we need to add them in the following too?\\n> \\n> delay_execution/delay_execution.c\\n> ssl_passphrase_func.c\\n> worker_spi.c\\n> \\n> Especially, worker_spi.c is important as it works as a template for\\n> the bg worker code.\\n\\nThank you for pointing this out! I added it.\\n\\n> I'm not sure if we have \\\"how to create an extension/a bg worker?\\\" in\\n> the core docs, if we have, it is a good idea to make note of using\\n> EmitWarningsOnPlaceholders so that it will not be missed in the future\\n> modules.\\n\\nI cannot find any such docs, so I add a comment to \\nsrc/backend/utils/misc/guc.c.\\n\\n> I observed an odd behaviour:\\n> 1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\\n> 2) With EmitWarningsOnPlaceholders(\\\"postgres_fdw\\\"); in postgres_fdw\\n> contrib module, I created the extension, have seen the following\\n> warning:\\n> 2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\\n> configuration parameter \\\"postgres_fdw.XXX\\\"\\n> 3) I further did, \\\"alter system set\\n> postgres_fdw.XXX='I_further_messed_up_conf_file';\\\" and also \\\"select\\n> pg_reload_conf();\\\", it silently accepts.\\n> \\n> postgres=# create extension postgres_fdw ;\\n> WARNING: unrecognized configuration parameter \\\"postgres_fdw.XXX\\\"\\n> CREATE EXTENSION\\n> postgres=# alter system set \\n> postgres_fdw.XXX='I_further_messed_up_conf_file';\\n> ALTER SYSTEM\\n> postgres=# select pg_reload_conf();\\n> pg_reload_conf\\n> ----------------\\n> t\\n> (1 row)\\n> \\n> My point is, why should the \\\"create extension\\\" succeed with a WARNING\\n> when a wrong parameter related to it is set in the postgresql.conf\\n> file so that we don't see further problems as shown in (3). I think\\n> EmitWarningsOnPlaceholders should emit an error so that the extension\\n> will not get created in the first place if a wrong configuration\\n> parameter is set in the postgresql.conf file. Many times, users will\\n> not have access to server logs so it is good to fail the \\\"create\\n> extension\\\" statement.\\n> \\n> Thoughts?\\n> \\n> postgres=# create extension postgres_fdw ;\\n> ERROR: unrecognized configuration parameter \\\"postgres_fdw.XXX\\\"\\n> postgres=#\\n\\nI confirmed it too, and I think that's definitely a problem.\\nI also thought it would be a good idea to emit an ERROR as well as when \\nan invalid normal GUC is set.\\nI didn't change the function name because it would affect many third \\nparty extensions.\\n\\nI plan to change to emit an error when an invalid custom GUC is set in \\nthe SET or ALTER SYSTEM SET commands, but I haven't tackled this yet.\\nThe patch as of now is attached.\\n\\n\\n-- \\nRegards,\\n\\n--\\nShinya Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Tue, 16 Nov 2021 18:20:33 +0900\",\n \"msg_from\": \"Shinya Kato <Shinya11.Kato@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 2:50 PM Shinya Kato\\n<Shinya11.Kato@oss.nttdata.com> wrote:\\n> > I'm not sure if we have \\\"how to create an extension/a bg worker?\\\" in\\n> > the core docs, if we have, it is a good idea to make note of using\\n> > EmitWarningsOnPlaceholders so that it will not be missed in the future\\n> > modules.\\n>\\n> I cannot find any such docs, so I add a comment to\\n> src/backend/utils/misc/guc.c.\\n>\\n> > I observed an odd behaviour:\\n> > 1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\\n> > 2) With EmitWarningsOnPlaceholders(\\\"postgres_fdw\\\"); in postgres_fdw\\n> > contrib module, I created the extension, have seen the following\\n> > warning:\\n> > 2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\\n> > configuration parameter \\\"postgres_fdw.XXX\\\"\\n> > 3) I further did, \\\"alter system set\\n> > postgres_fdw.XXX='I_further_messed_up_conf_file';\\\" and also \\\"select\\n> > pg_reload_conf();\\\", it silently accepts.\\n> >\\n> > postgres=# create extension postgres_fdw ;\\n> > WARNING: unrecognized configuration parameter \\\"postgres_fdw.XXX\\\"\\n> > CREATE EXTENSION\\n> > postgres=# alter system set\\n> > postgres_fdw.XXX='I_further_messed_up_conf_file';\\n> > ALTER SYSTEM\\n> > postgres=# select pg_reload_conf();\\n> > pg_reload_conf\\n> > ----------------\\n> > t\\n> > (1 row)\\n> >\\n> > My point is, why should the \\\"create extension\\\" succeed with a WARNING\\n> > when a wrong parameter related to it is set in the postgresql.conf\\n> > file so that we don't see further problems as shown in (3). I think\\n> > EmitWarningsOnPlaceholders should emit an error so that the extension\\n> > will not get created in the first place if a wrong configuration\\n> > parameter is set in the postgresql.conf file. Many times, users will\\n> > not have access to server logs so it is good to fail the \\\"create\\n> > extension\\\" statement.\\n> >\\n> > Thoughts?\\n> >\\n> > postgres=# create extension postgres_fdw ;\\n> > ERROR: unrecognized configuration parameter \\\"postgres_fdw.XXX\\\"\\n> > postgres=#\\n>\\n> I confirmed it too, and I think that's definitely a problem.\\n> I also thought it would be a good idea to emit an ERROR as well as when\\n> an invalid normal GUC is set.\\n> I didn't change the function name because it would affect many third\\n> party extensions.\\n\\nFor backward compatibility you can retain the function\\nEmitWarningsOnPlaceholders as-is and have another similar function\\nthat emits an error:\\n\\nIn guc.c, have something like below:\\nstatic void\\ncheck_conf_params(const char *className, bool emit_error)\\n{\\n /* have the existing code of EmitWarningsOnPlaceholders here */\\n ereport(emit_error ? ERROR : WARNING,\\n (errcode(ERRCODE_UNDEFINED_OBJECT),\\n errmsg(\\\"unrecognized configuration parameter \\\\\\\"%s\\\\\\\"\\\",\\n var->name)));\\n}\\n\\nvoid\\nEmitErrorOnPlaceholders(const char *className)\\n{\\n check_conf_params(className, true);\\n}\\n\\nvoid\\nEmitWarningsOnPlaceholders(const char *className)\\n{\\n check_conf_params(className, false);\\n}\\n\\nAnd change all the core extensions to use EmitErrorOnPlaceholders.\\nThis way you don't break the backward compatibility of the outside\\nextensions, if they want they get to choose which behaviour they want\\nfor their extension.\\n\\nActually, I didn't quite like the function name\\nEmitWarningsOnPlaceholders or EmitErrorOnPlaceholders, it could have\\nbeen something else. But let's not go that direction of changing the\\nfunction name for backward compatibility.\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 15:55:13 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"Thank you for the review and sorry for the late reply.\\n\\nOn 2021-11-16 19:25, Bharath Rupireddy wrote:\\n>> > I observed an odd behaviour:\\n>> > 1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\\n>> > 2) With EmitWarningsOnPlaceholders(\\\"postgres_fdw\\\"); in postgres_fdw\\n>> > contrib module, I created the extension, have seen the following\\n>> > warning:\\n>> > 2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\\n>> > configuration parameter \\\"postgres_fdw.XXX\\\"\\n>> > 3) I further did, \\\"alter system set\\n>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\\\" and also \\\"select\\n>> > pg_reload_conf();\\\", it silently accepts.\\n>> >\\n>> > postgres=# create extension postgres_fdw ;\\n>> > WARNING: unrecognized configuration parameter \\\"postgres_fdw.XXX\\\"\\n>> > CREATE EXTENSION\\n>> > postgres=# alter system set\\n>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\\n>> > ALTER SYSTEM\\n>> > postgres=# select pg_reload_conf();\\n>> > pg_reload_conf\\n>> > ----------------\\n>> > t\\n>> > (1 row)\\n\\nI have made changes to achieve the above.\\nHowever, it also gives me an error when\\n---\\npostgres=# SET abc.a TO on;\\nSET\\npostgres=# SET abc.b TO on;\\n2021-12-16 16:22:20.351 JST [2489236] ERROR: unrecognized configuration \\nparameter \\\"abc.a\\\"\\n2021-12-16 16:22:20.351 JST [2489236] STATEMENT: SET abc.b TO on;\\nERROR: unrecognized configuration parameter \\\"abc.a\\\"\\n---\\n\\nI know that some people do not think this is good.\\nTherefore, can I leave this problem alone or is there another better \\nway?\\n\\n\\n> For backward compatibility you can retain the function\\n> EmitWarningsOnPlaceholders as-is and have another similar function\\n> that emits an error:\\n> \\n> In guc.c, have something like below:\\n> static void\\n> check_conf_params(const char *className, bool emit_error)\\n> {\\n> /* have the existing code of EmitWarningsOnPlaceholders here */\\n> ereport(emit_error ? ERROR : WARNING,\\n> (errcode(ERRCODE_UNDEFINED_OBJECT),\\n> errmsg(\\\"unrecognized configuration parameter \\n> \\\\\\\"%s\\\\\\\"\\\",\\n> var->name)));\\n> }\\n> \\n> void\\n> EmitErrorOnPlaceholders(const char *className)\\n> {\\n> check_conf_params(className, true);\\n> }\\n> \\n> void\\n> EmitWarningsOnPlaceholders(const char *className)\\n> {\\n> check_conf_params(className, false);\\n> }\\n> \\n\\nThank you for your advise.\\nAccording to [1], we used the same function name, but the warning level \\nwas INFO.\\nTherefore, I think it is OK to use the same function name.\\n\\n[1] \\nhttps://www.postgresql.org/message-id/flat/200901051634.n05GYNr06169%40momjian.us#1d045374f014494e4b40a4862a000723\\n\\n-- \\nRegards,\\n\\n--\\nShinya Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Thu, 16 Dec 2021 16:31:03 +0900\",\n \"msg_from\": \"Shinya Kato <Shinya11.Kato@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2021/12/16 16:31, Shinya Kato wrote:\\n> Thank you for the review and sorry for the late reply.\\n> \\n> On 2021-11-16 19:25, Bharath Rupireddy wrote:\\n>>> > I observed an odd behaviour:\\n>>> > 1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\\n>>> > 2) With EmitWarningsOnPlaceholders(\\\"postgres_fdw\\\"); in postgres_fdw\\n>>> > contrib module, I created the extension, have seen the following\\n>>> > warning:\\n>>> > 2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\\n>>> > configuration parameter \\\"postgres_fdw.XXX\\\"\\n>>> > 3) I further did, \\\"alter system set\\n>>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\\\" and also \\\"select\\n>>> > pg_reload_conf();\\\", it silently accepts.\\n>>> >\\n>>> > postgres=# create extension postgres_fdw ;\\n>>> > WARNING: unrecognized configuration parameter \\\"postgres_fdw.XXX\\\"\\n>>> > CREATE EXTENSION\\n>>> > postgres=# alter system set\\n>>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\\n>>> > ALTER SYSTEM\\n>>> > postgres=# select pg_reload_conf();\\n>>> > pg_reload_conf\\n>>> > ----------------\\n>>> > t\\n>>> > (1 row)\\n> \\n> I have made changes to achieve the above.\\n\\nIMO this behavior change is not good. For example, because it seems to break at least the following case. I think that these are the valid steps, but with the patch, the server fails to start up at the step #2 because pg_trgm's custom parameters are treated as invalid ones.\\n\\n1. Add the following two pg_trgm parameters to postgresql.conf\\n - pg_trgm.similarity_threshold\\n - pg_trgm.strict_word_similarity_threshold\\n\\n2. Start the server\\n\\n3. Run \\\"CREATE EXTENSION pg_trgm\\\" and then use pg_trgm\\n\\n\\nWhen I compiled PostgreSQL with the patch, I got the following\\ncompilation failure.\\n\\nguc.c:5453:4: error: implicit declaration of function 'EmitErrorsOnPlaceholders' is invalid in C99 [-Werror,-Wimplicit-function-declaration]\\n EmitErrorsOnPlaceholders(placeholder);\\n ^\\n\\n\\n-\\t\\t\\tereport(WARNING,\\n+\\t\\t\\tereport(ERROR,\\n\\nI'm still not sure why you were thinking that ERROR is more proper here.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Fri, 17 Dec 2021 01:55:19 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"On 2021-12-17 01:55, Fujii Masao wrote:\\n> On 2021/12/16 16:31, Shinya Kato wrote:\\n>> Thank you for the review and sorry for the late reply.\\n>> \\n>> On 2021-11-16 19:25, Bharath Rupireddy wrote:\\n>>>> > I observed an odd behaviour:\\n>>>> > 1) I set postgres_fdw.XXX = 'I_messed_up_conf_file' in postgresql.conf\\n>>>> > 2) With EmitWarningsOnPlaceholders(\\\"postgres_fdw\\\"); in postgres_fdw\\n>>>> > contrib module, I created the extension, have seen the following\\n>>>> > warning:\\n>>>> > 2021-11-15 06:02:31.198 UTC [2018111] WARNING: unrecognized\\n>>>> > configuration parameter \\\"postgres_fdw.XXX\\\"\\n>>>> > 3) I further did, \\\"alter system set\\n>>>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\\\" and also \\\"select\\n>>>> > pg_reload_conf();\\\", it silently accepts.\\n>>>> >\\n>>>> > postgres=# create extension postgres_fdw ;\\n>>>> > WARNING: unrecognized configuration parameter \\\"postgres_fdw.XXX\\\"\\n>>>> > CREATE EXTENSION\\n>>>> > postgres=# alter system set\\n>>>> > postgres_fdw.XXX='I_further_messed_up_conf_file';\\n>>>> > ALTER SYSTEM\\n>>>> > postgres=# select pg_reload_conf();\\n>>>> > pg_reload_conf\\n>>>> > ----------------\\n>>>> > t\\n>>>> > (1 row)\\n>> \\n>> I have made changes to achieve the above.\\n> \\n> IMO this behavior change is not good. For example, because it seems to\\n> break at least the following case. I think that these are the valid\\n> steps, but with the patch, the server fails to start up at the step #2\\n> because pg_trgm's custom parameters are treated as invalid ones.\\n> \\n> 1. Add the following two pg_trgm parameters to postgresql.conf\\n> - pg_trgm.similarity_threshold\\n> - pg_trgm.strict_word_similarity_threshold\\n> \\n> 2. Start the server\\n> \\n> 3. Run \\\"CREATE EXTENSION pg_trgm\\\" and then use pg_trgm\\n\\nIt can happen because the placeholder \\\"pg_trgm\\\" is not already \\nregistered.\\nI think too this behavior is not good.\\nI need to consider an another implementation or to allow undefined GUCs \\nto be set.\\n\\n\\n> When I compiled PostgreSQL with the patch, I got the following\\n> compilation failure.\\n> \\n> guc.c:5453:4: error: implicit declaration of function\\n> 'EmitErrorsOnPlaceholders' is invalid in C99\\n> [-Werror,-Wimplicit-function-declaration]\\n> EmitErrorsOnPlaceholders(placeholder);\\n> ^\\n> \\n> \\n> -\\t\\t\\tereport(WARNING,\\n> +\\t\\t\\tereport(ERROR,\\n\\nThanks! There was a correction omission, so I fixed it.\\n\\n\\n> I'm still not sure why you were thinking that ERROR is more proper \\n> here.\\n\\nSince I get an ERROR when I set the wrong normal GUCs, I thought I \\nshould also get an ERROR when I set the wrong extension's GUCs.\\nHowever, there are likely to be harmful effects, and I may need to think \\nagain.\\n\\n\\nFor now, I'v attached the patch that fixed the compilation error.\\n\\n-- \\nRegards,\\n\\n--\\nShinya Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Fri, 17 Dec 2021 11:25:22 +0900\",\n \"msg_from\": \"Shinya Kato <Shinya11.Kato@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2021/12/17 11:25, Shinya Kato wrote:\\n> For now, I'v attached the patch that fixed the compilation error.\\n\\nThanks for updating the patch! I confirmed that the compilation was completed successfully with this new patch. But the regression test failed. You can see that Patch Tester [1] also reported the failure of regression test for your patch.\\n\\n[1]\\nhttp://commitfest.cputube.org/\\n\\nRegards,\\n\\n-- \\nFujii Masao\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Fri, 17 Dec 2021 14:00:27 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"On 17.12.21 03:25, Shinya Kato wrote:\\n> For now, I'v attached the patch that fixed the compilation error.\\n\\nI think it would be good if you could split the uncontroversial new \\nEmitErrorsOnPlaceholders() calls into a separate patch. And please add \\nan explanation what exactly the rest of the patch changes.\\n\\n\\n\",\n \"msg_date\": \"Fri, 17 Dec 2021 07:42:58 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"On 2021-12-17 15:42, Peter Eisentraut wrote:\\n> On 17.12.21 03:25, Shinya Kato wrote:\\n>> For now, I'v attached the patch that fixed the compilation error.\\n> \\n> I think it would be good if you could split the uncontroversial new\\n> EmitErrorsOnPlaceholders() calls into a separate patch. And please\\n> add an explanation what exactly the rest of the patch changes.\\n\\nThank you for the comment!\\nI splitted the patch.\\n\\n- v6-01-Add-EmitWarningsOnPlaceholders.patch\\nWe should use EmitWarningsOnPlaceholders when we use \\nDefineCustomXXXVariable.\\nI don't think there is any room for debate.\\n\\n- v6-0002-Change-error-level-of-EmitWarningsOnPlaceholders.patch\\nThis is a patch to change the error level of EmitWarningsOnPlaceholders \\nfrom WARNING to ERROR.\\nI think it's OK to emit ERROR as well as when the wrong GUC is set for \\nnon-extensions, or since it does not behave abnormally, it can be left \\nas WARNING.\\nThought?\\n\\n- v6-0003-Emit-error-when-invalid-extensions-GUCs-are-set.patch\\nThis is a patch to emit error when invalid extension's GUCs are set.\\nNo test changes have been made, so the regression test will fail.\\nI have created a patch, but I don't think this is necessary because of \\nthe previous discussion.\\n\\n-- \\nRegards,\\n\\n--\\nShinya Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Mon, 20 Dec 2021 21:05:23 +0900\",\n \"msg_from\": \"Shinya Kato <Shinya11.Kato@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"At Mon, 20 Dec 2021 21:05:23 +0900, Shinya Kato <Shinya11.Kato@oss.nttdata.com> wrote in \\n> - v6-01-Add-EmitWarningsOnPlaceholders.patch\\n> We should use EmitWarningsOnPlaceholders when we use\\n> DefineCustomXXXVariable.\\n> I don't think there is any room for debate.\\n\\nUnfortunately, pltcl.c defines variables both in pltcl and pltclu but\\nthe patch adds the warning only for pltclu.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Tue, 21 Dec 2021 10:08:34 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"At Fri, 17 Dec 2021 11:25:22 +0900, Shinya Kato <Shinya11.Kato@oss.nttdata.com> wrote in \\n> On 2021-12-17 01:55, Fujii Masao wrote:\\n> > I'm still not sure why you were thinking that ERROR is more proper\\n> > here.\\n> \\n> Since I get an ERROR when I set the wrong normal GUCs, I thought I\\n> should also get an ERROR when I set the wrong extension's GUCs.\\n> However, there are likely to be harmful effects, and I may need to\\n> think again.\\n\\nThe GUC placeholder mechanism is evidently designed so that server\\nallows to define variables unknown to the server before loading\\nextension modules. ERRORing out that variables at the timing is\\ncontradicting the objective for the placeholder mechanism.\\n\\nAddition to that EmitWarningsOnPlaceholders()'s objective is to warn\\nfor variables that shouldn't be of a *known* namespace. However, the\\npatch is intending to check out variable definitions of all namespaces\\nthat are seen in configuration file, but it is undeterminable at that\\ntime whether each of the namespaces is either just wrong or unknown\\nyet. And the latter is, as mentioned above, what we are intending to\\nallow.\\n\\nAs the concusion, I think we cannot rule-out \\\"wrong\\\" namespaces unless\\nwe find any means to distinguish whether a namespace is wrong or\\nnot-yet-defined before loading extension modules.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Tue, 21 Dec 2021 10:50:07 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com> writes:\\n> At Mon, 20 Dec 2021 21:05:23 +0900, Shinya Kato <Shinya11.Kato@oss.nttdata.com> wrote in \\n>> We should use EmitWarningsOnPlaceholders when we use\\n>> DefineCustomXXXVariable.\\n>> I don't think there is any room for debate.\\n\\n> Unfortunately, pltcl.c defines variables both in pltcl and pltclu but\\n> the patch adds the warning only for pltclu.\\n\\nRight. The rest of 0001 looks fine, so pushed with that correction.\\n\\nI concur that 0002 is unacceptable. The result of it will be that\\na freshly-loaded extension will fail to hook into the system as it\\nshould, because it will fail to complete its initialization.\\nThat will cause user frustration and bug reports.\\n\\n(BTW, I wonder if EmitWarningsOnPlaceholders should be using LOG\\nrather than WARNING when it's running in the postmaster. Use of\\nWARNING is moderately likely to result in nothing getting printed\\nat all.)\\n\\n0003 looks to me like it was superseded by 75d22069e. I do not\\nparticularly like that patch though; it seems both inefficient\\nand ill-designed. 0003 is adding a check in an equally bizarre\\nplace. Why isn't add_placeholder_variable the place to deal\\nwith this? Also, once we know that a prefix is reserved,\\nwhy don't we throw an error not just a warning for attempts to\\nset some unknown variable under that prefix? We don't allow\\nyou to set unknown non-prefixed variables.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 21 Dec 2021 12:23:25 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> 0003 looks to me like it was superseded by 75d22069e. I do not\\n> particularly like that patch though; it seems both inefficient\\n> and ill-designed. 0003 is adding a check in an equally bizarre\\n> place. Why isn't add_placeholder_variable the place to deal\\n> with this? Also, once we know that a prefix is reserved,\\n> why don't we throw an error not just a warning for attempts to\\n> set some unknown variable under that prefix? We don't allow\\n> you to set unknown non-prefixed variables.\\n\\nConcretely, I think we should do the attached, which removes much of\\n75d22069e and does the check at the point of placeholder creation.\\n(After looking closer, add_placeholder_variable's caller is the\\nfunction that is responsible for rejecting bad names, not\\nadd_placeholder_variable itself.)\\n\\nThis also fixes an indisputable bug in 75d22069e, namely that it\\ndid prefix comparisons incorrectly and would throw warnings for\\ncases it shouldn't. Reserving \\\"plpgsql\\\" shouldn't have the effect\\nof reserving \\\"pl\\\" as well.\\n\\nI'm tempted to propose that we also rename EmitWarningsOnPlaceholders\\nto something like MarkGUCPrefixReserved, to more clearly reflect\\nwhat it does now. (We could provide the old name as a macro alias\\nto avoid breaking extensions needlessly.)\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Tue, 21 Dec 2021 13:33:49 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"On 2021-12-22 02:23, Tom Lane wrote:\\n> Kyotaro Horiguchi <horikyota.ntt@gmail.com> writes:\\n>> At Mon, 20 Dec 2021 21:05:23 +0900, Shinya Kato \\n>> <Shinya11.Kato@oss.nttdata.com> wrote in\\n>>> We should use EmitWarningsOnPlaceholders when we use\\n>>> DefineCustomXXXVariable.\\n>>> I don't think there is any room for debate.\\n> \\n>> Unfortunately, pltcl.c defines variables both in pltcl and pltclu but\\n>> the patch adds the warning only for pltclu.\\n> \\n> Right. The rest of 0001 looks fine, so pushed with that correction.\\n> \\n> I concur that 0002 is unacceptable. The result of it will be that\\n> a freshly-loaded extension will fail to hook into the system as it\\n> should, because it will fail to complete its initialization.\\n> That will cause user frustration and bug reports.\\n> \\n> (BTW, I wonder if EmitWarningsOnPlaceholders should be using LOG\\n> rather than WARNING when it's running in the postmaster. Use of\\n> WARNING is moderately likely to result in nothing getting printed\\n> at all.)\\n> \\n> 0003 looks to me like it was superseded by 75d22069e. I do not\\n> particularly like that patch though; it seems both inefficient\\n> and ill-designed. 0003 is adding a check in an equally bizarre\\n> place. Why isn't add_placeholder_variable the place to deal\\n> with this? Also, once we know that a prefix is reserved,\\n> why don't we throw an error not just a warning for attempts to\\n> set some unknown variable under that prefix? We don't allow\\n> you to set unknown non-prefixed variables.\\n> \\n> \\t\\t\\tregards, tom lane\\n\\nThank you for pushing!\\nThank you all for the reviews, I think I will take down 0002 and 0003.\\n\\n-- \\nRegards,\\n\\n--\\nShinya Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Wed, 22 Dec 2021 16:50:34 +0900\",\n \"msg_from\": \"Shinya Kato <Shinya11.Kato@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"On 2021/12/22 3:33, Tom Lane wrote:\\n> I wrote:\\n>> 0003 looks to me like it was superseded by 75d22069e. I do not\\n>> particularly like that patch though; it seems both inefficient\\n>> and ill-designed. 0003 is adding a check in an equally bizarre\\n>> place. Why isn't add_placeholder_variable the place to deal\\n>> with this? Also, once we know that a prefix is reserved,\\n>> why don't we throw an error not just a warning for attempts to\\n>> set some unknown variable under that prefix? We don't allow\\n>> you to set unknown non-prefixed variables.\\n> Concretely, I think we should do the attached, which removes much of\\n> 75d22069e and does the check at the point of placeholder creation.\\n> (After looking closer, add_placeholder_variable's caller is the\\n> function that is responsible for rejecting bad names, not\\n> add_placeholder_variable itself.)\\n>\\n> This also fixes an indisputable bug in 75d22069e, namely that it\\n> did prefix comparisons incorrectly and would throw warnings for\\n> cases it shouldn't. Reserving \\\"plpgsql\\\" shouldn't have the effect\\n> of reserving \\\"pl\\\" as well.\\n\\nThank you for creating the patch.\\nThis is exactly what I wanted to create. It looks good to me.\\n\\n\\n> I'm tempted to propose that we also rename EmitWarningsOnPlaceholders\\n> to something like MarkGUCPrefixReserved, to more clearly reflect\\n> what it does now. (We could provide the old name as a macro alias\\n> to avoid breaking extensions needlessly.)\\n\\n+1\\n\\n\\n-- \\nRegards,\\n\\n--\\nShinya Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 24 Dec 2021 12:29:47 +0900\",\n \"msg_from\": \"Shinya Kato <Shinya11.Kato@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Concretely, I think we should do the attached, which removes much of\\n> 75d22069e and does the check at the point of placeholder creation.\\n\\nI pushed that, and along the way moved the test case to be beside\\nthe existing tests concerning custom GUC names, rather than appended\\nat the end of guc.sql as it had been. That turns out to make the\\nbuildfarm very unhappy. I had not thought to test that change with\\n\\\"force_parallel_mode = regress\\\"; but with that on, we can see that\\nthe warning (or now error) is reported each time a parallel worker\\nis launched, if the parent process contains a bogus placeholder.\\nSo that accidentally unveiled another deficiency in the design of\\nthe original patch --- we surely don't want that to happen.\\n\\nAs a stopgap to turn the farm green again, I am going to revert\\n75d22069e as well as my followup patches. If we don't want to\\ngive up on that idea altogether, we have to find some way to\\nsuppress the chatter from parallel workers. I wonder whether\\nit would be appropriate to go further than we have, and actively\\ndelete placeholders that turn out to be within an extension's\\nreserved namespace. The core issue here is that workers don't\\nnecessarily set GUCs and load extensions in the same order that\\ntheir parent did, so if we leave any invalid placeholders behind\\nafter reserving an extension's prefix, we're risking issues\\nduring worker start.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 27 Dec 2021 15:48:43 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> As a stopgap to turn the farm green again, I am going to revert\\n> 75d22069e as well as my followup patches. If we don't want to\\n> give up on that idea altogether, we have to find some way to\\n> suppress the chatter from parallel workers. I wonder whether\\n> it would be appropriate to go further than we have, and actively\\n> delete placeholders that turn out to be within an extension's\\n> reserved namespace. The core issue here is that workers don't\\n> necessarily set GUCs and load extensions in the same order that\\n> their parent did, so if we leave any invalid placeholders behind\\n> after reserving an extension's prefix, we're risking issues\\n> during worker start.\\n\\nHere's a delta patch (meant to be applied after reverting cab5b9ab2)\\nthat does things like that. It fixes the parallel-mode problem ...\\nso do we want to tighten things up this much?\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Tue, 28 Dec 2021 12:06:55 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"Il giorno ven 18 feb 2022 alle ore 10:58 Tom Lane <tgl@sss.pgh.pa.us> ha\\nscritto:\\n\\n> I wrote:\\n> > As a stopgap to turn the farm green again, I am going to revert\\n> > 75d22069e as well as my followup patches. If we don't want to\\n> > give up on that idea altogether, we have to find some way to\\n> > suppress the chatter from parallel workers. I wonder whether\\n> > it would be appropriate to go further than we have, and actively\\n> > delete placeholders that turn out to be within an extension's\\n> > reserved namespace. The core issue here is that workers don't\\n> > necessarily set GUCs and load extensions in the same order that\\n> > their parent did, so if we leave any invalid placeholders behind\\n> > after reserving an extension's prefix, we're risking issues\\n> > during worker start.\\n>\\n> Here's a delta patch (meant to be applied after reverting cab5b9ab2)\\n> that does things like that. It fixes the parallel-mode problem ...\\n> so do we want to tighten things up this much?\\n>\\n> regards, tom lan\\n>\\n\\nHello,\\n\\nThank you for taking care of the bug I introduced with 75d22069e,\\nI didn't notice this thread until now.\\n\\nFor what it's worth, I agree with throwing an ERROR if the placeholder is\\nunrecognized. Initially, I didn't want to change too much the liberty of\\nsetting any\\nplaceholder, but mainly to not go unnoticed.\\n\\nIn my humble opinion, I still think that this should go on and disallow\\nbogus placeholders as we do for postgres unrecognized settings.\\n\\nI tested your delta patch with and without parallel mode, and, naturally,\\nit behaves correctly.\\n\\nMy 2 cents.\\n\\n+1\\n\\nCheers,\\nFlorin Irion\\n\\nIl giorno ven 18 feb 2022 alle ore 10:58 Tom Lane <tgl@sss.pgh.pa.us> ha scritto:I wrote:\\n> As a stopgap to turn the farm green again, I am going to revert\\n> 75d22069e as well as my followup patches. If we don't want to\\n> give up on that idea altogether, we have to find some way to\\n> suppress the chatter from parallel workers. I wonder whether\\n> it would be appropriate to go further than we have, and actively\\n> delete placeholders that turn out to be within an extension's\\n> reserved namespace. The core issue here is that workers don't\\n> necessarily set GUCs and load extensions in the same order that\\n> their parent did, so if we leave any invalid placeholders behind\\n> after reserving an extension's prefix, we're risking issues\\n> during worker start.\\n\\nHere's a delta patch (meant to be applied after reverting cab5b9ab2)\\nthat does things like that. It fixes the parallel-mode problem ...\\nso do we want to tighten things up this much?\\n\\n regards, tom lanHello, Thank you for taking care of the bug I introduced with 75d22069e,I didn't notice this thread until now.For what it's worth, I agree with throwing an ERROR if the placeholder isunrecognized. Initially, I didn't want to change too much the liberty of setting anyplaceholder, but mainly to not go unnoticed.In my humble opinion, I still think that this should go on and disallow bogus placeholders as we do for postgres unrecognized settings. I tested your delta patch with and without parallel mode, and, naturally, it behaves correctly.My 2 cents.+1Cheers,Florin Irion\",\n \"msg_date\": \"Fri, 18 Feb 2022 12:47:23 +0100\",\n \"msg_from\": \"Florin Irion <irionr@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"Florin Irion <irionr@gmail.com> writes:\\n> Il giorno ven 18 feb 2022 alle ore 10:58 Tom Lane <tgl@sss.pgh.pa.us> ha\\n> scritto:\\n>> Here's a delta patch (meant to be applied after reverting cab5b9ab2)\\n>> that does things like that. It fixes the parallel-mode problem ...\\n>> so do we want to tighten things up this much?\\n\\n> Thank you for taking care of the bug I introduced with 75d22069e,\\n> I didn't notice this thread until now.\\n\\nYeah, this thread kinda disappeared under the rug during the Christmas\\nholidays, but we need to decide whether we want to push forward or\\nleave things at the status quo.\\n\\n> For what it's worth, I agree with throwing an ERROR if the placeholder is\\n> unrecognized. Initially, I didn't want to change too much the liberty of\\n> setting any placeholder, but mainly to not go unnoticed.\\n\\nI also think that this is probably a good change to make.\\n\\nOne situation where somebody would be unhappy is if they're using GUC\\nvariables as plain custom variables despite them being within the\\nnamespace of an extension they're using. But that seems to me to be\\n(a) far-fetched and (b) mighty dangerous, since some new version of the\\nextension might suddenly start reacting to those variables in ways you\\npresumably didn't expect.\\n\\nPerhaps a more plausible use-case is \\\"I need to work with both versions\\nX and Y of extension E, and Y has setting e.foo while X doesn't --- but\\nI can just set e.foo unconditionally since X will ignore it\\\". With this\\npatch, that could still work, but you'd have to be certain to apply the\\nsetting before loading E.\\n\\nI don't really see any other use-cases favoring the current behavior.\\nOn balance, eliminating possible mistakes seems like enough of a\\nbenefit to justify disallowing these cases.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 18 Feb 2022 17:18:21 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Florin Irion <irionr@gmail.com> writes:\\n>> For what it's worth, I agree with throwing an ERROR if the placeholder is\\n>> unrecognized. Initially, I didn't want to change too much the liberty of\\n>> setting any placeholder, but mainly to not go unnoticed.\\n\\n> I also think that this is probably a good change to make.\\n\\nHearing no objections, done.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 21 Feb 2022 14:12:04 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n },\n {\n \"msg_contents\": \"Il lun 21 feb 2022, 20:12 Tom Lane <tgl@sss.pgh.pa.us> ha scritto:\\n\\n> I wrote:\\n> > Florin Irion <irionr@gmail.com> writes:\\n> >> For what it's worth, I agree with throwing an ERROR if the placeholder\\n> is\\n> >> unrecognized. Initially, I didn't want to change too much the liberty of\\n> >> setting any placeholder, but mainly to not go unnoticed.\\n>\\n> > I also think that this is probably a good change to make.\\n>\\n> Hearing no objections, done.\\n>\\n> regards, tom lane\\n>\\n\\nThank you for taking care of this.\\n\\nCheers,\\nFlorin\\n\\n>\\n\\nIl lun 21 feb 2022, 20:12 Tom Lane <tgl@sss.pgh.pa.us> ha scritto:I wrote:\\n> Florin Irion <irionr@gmail.com> writes:\\n>> For what it's worth, I agree with throwing an ERROR if the placeholder is\\n>> unrecognized. Initially, I didn't want to change too much the liberty of\\n>> setting any placeholder, but mainly to not go unnoticed.\\n\\n> I also think that this is probably a good change to make.\\n\\nHearing no objections, done.\\n\\n regards, tom laneThank you for taking care of this. Cheers, Florin\",\n \"msg_date\": \"Mon, 21 Feb 2022 22:22:55 +0100\",\n \"msg_from\": \"Florin Irion <irionr@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Emit a warning if the extension's GUC is set incorrectly\"\n }\n]","isConversational":false}}},{"rowIdx":53089,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI have been wondering why there is no meta-command for listing\nconstraints in psql. So, I created a POC patch by using my\nexperience developing \\dX command in PG14.\n\nThis feature is helpful for DBAs when they want to check or\nmodify the definition of constraints.\n\nThe current status of the POC patch is as follows:\n\n - Add \"\\dco\" command to list constraints from pg_constraint\n\n - Not implemented yet:\n - NOT NULL constraint, and so on (based on pg_attribute)\n - Tab completion\n - Regression test\n - Document\n\nThe following is test results (See attached test_list_con.sql)\n====================================================================\npostgres=# \\dco\n List of constsraints\n Schema | Name | Definition | Table\n--------+-------------------------+---------------------------------------------------------+----------\n public | t01_chk_price_check | CHECK ((price > (0)::numeric)) | t01_chk\n public | t02_uniq_product_no_key | UNIQUE (product_no) | t02_uniq\n public | t03_pk1_pkey | PRIMARY KEY (product_no) | t03_pk1\n public | t03_pk2_product_no_key | UNIQUE (product_no) | t03_pk2\n public | t04_fk_pkey | PRIMARY KEY (order_id) | t04_fk\n public | t04_fk_product_no_fkey | FOREIGN KEY (product_no) REFERENCES t03_pk1(product_no) | t04_fk\n public | t05_ex_c_excl | EXCLUDE USING gist (c WITH &&) | t05_ex\n(7 rows)\n====================================================================\n\n\nI have the following two questions that need to be discussed.\n\nQuestions:\n(1) What strings should be assigned as meta-command for this feature?\n Currently, \\dc and \\dC are not available, so I tentatively\n assigned \\dco. However, I do not have a strong opinion, so please\n let me know if you have any suggestions.\n\n(2) About domain constraints\n There is the \\dD command to show a list of domain constraints.\n So I think this feature should not include it. Is it Okay?\n\n\nIf I can get \"+1\" for this new feature development, I would like to\nimprove the patch by adding NOT NULL constraints, and so on.\nAny advice or comments would be appreciated.\n\n\nThanks,\nTatsuro Yamada","msg_date":"Mon, 15 Nov 2021 10:38:55 +0900","msg_from":"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>","msg_from_op":true,"msg_subject":"Add psql command to list constraints"},{"msg_contents":"Hi,\n\nOn Mon, Nov 15, 2021 at 10:38:55AM +0900, Tatsuro Yamada wrote:\n> postgres=# \\dco\n> List of constsraints\n> Schema | Name | Definition | Table\n> --------+-------------------------+---------------------------------------------------------+----------\n> public | t01_chk_price_check | CHECK ((price > (0)::numeric)) | t01_chk\n> public | t02_uniq_product_no_key | UNIQUE (product_no) | t02_uniq\n> public | t03_pk1_pkey | PRIMARY KEY (product_no) | t03_pk1\n> public | t03_pk2_product_no_key | UNIQUE (product_no) | t03_pk2\n> public | t04_fk_pkey | PRIMARY KEY (order_id) | t04_fk\n> public | t04_fk_product_no_fkey | FOREIGN KEY (product_no) REFERENCES t03_pk1(product_no) | t04_fk\n> public | t05_ex_c_excl | EXCLUDE USING gist (c WITH &&) | t05_ex\n> (7 rows)\n> ====================================================================\n\nMaybe it ought to be possible to choose the type of constraints to show.\nSimilar to how \\dt shows tables and \\di shows indexes and \\dti shows\ntables+inds, you could run \\dcoc for check constraints and \\dcof for foreign\nkeys. But I think \"\\dco\" is too long of a prefix...\n\n> +\tinitPQExpBuffer(&buf);\n> +\tprintfPQExpBuffer(&buf,\n> +\t\t\t\t\t \"SELECT \\n\"\n> +\t\t\t\t\t \"n.nspname AS \\\"%s\\\", \\n\"\n> +\t\t\t\t\t \"cst.conname AS \\\"%s\\\", \\n\"\n> +\t\t\t\t\t \"pg_catalog.pg_get_constraintdef(cst.oid) AS \\\"%s\\\", \\n\"\n> +\t\t\t\t\t \"c.relname AS \\\"%s\\\" \\n\"\n> +\t\t\t\t\t \"FROM pg_constraint cst \\n\"\n> +\t\t\t\t\t \"JOIN pg_namespace n ON n.oid = cst.connamespace \\n\"\n> +\t\t\t\t\t \"JOIN pg_class c ON c.oid = cst.conrelid \\n\",\n\nYou should write \"pg_catalog.\" prefix for the tables (in addition to the\nfunction).\n\nRather than join to pg_class, you can write conrelid::pg_catalog.regclass,\nsince regclass is supported since at least v7.3 (but ::regnamespace was\nintroduced in v9.5, so the join against pg_namespace is still necessary).\nhttps://www.postgresql.org/docs/9.5/datatype-oid.html\n\n> +\tmyopt.title = _(\"List of constsraints\");\n\nspelling: constraints\n\nI'm not confident that if I would use this, so let's wait to see if someone\nelse wants to give a +1.\n\n-- \nJustin\n\n\n","msg_date":"Sun, 14 Nov 2021 21:56:33 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"Hi Justin,\n\nThanks for your comments and review!\n\n> Maybe it ought to be possible to choose the type of constraints to show.\n> Similar to how \\dt shows tables and \\di shows indexes and \\dti shows\n> tables+inds, you could run \\dcoc for check constraints and \\dcof for foreign\n> keys. But I think \"\\dco\" is too long of a prefix...\n\n\nYeah, agreed.\nI added a filter based on the type of constraints:\n - c for check\n - f for foreign key\n - p for primary key\n - t for trigger\n - u for unique\n - x for exclude c, f, p, u, t, and x.\n\nThe following is examples of \\dcop, \\dcof, and \\dcopf.\n\n========================================================================\npostgres=# \\dcop\n List of constraints\n Schema | Name | Definition | Table\n--------+--------------+--------------------------+---------\n public | t03_pk1_pkey | PRIMARY KEY (product_no) | t03_pk1\n public | t04_fk_pkey | PRIMARY KEY (order_id) | t04_fk\n(2 rows)\n\npostgres=# \\dcof\n List of constraints\n Schema | Name | Definition | Table\n--------+------------------------+---------------------------------------------------------+--------\n public | t04_fk_product_no_fkey | FOREIGN KEY (product_no) REFERENCES t03_pk1(product_no) | t04_fk\n(1 row)\n\npostgres=# \\dcopf\n List of constraints\n Schema | Name | Definition | Table\n--------+------------------------+---------------------------------------------------------+---------\n public | t03_pk1_pkey | PRIMARY KEY (product_no) | t03_pk1\n public | t04_fk_pkey | PRIMARY KEY (order_id) | t04_fk\n public | t04_fk_product_no_fkey | FOREIGN KEY (product_no) REFERENCES t03_pk1(product_no) | t04_fk\n(3 rows)\n========================================================================\n\n\nI too think \\dco is a long name. So, I'd like to get suggestions to be more shortened. :)\n\n\n \n>> +\tinitPQExpBuffer(&buf);\n>> +\tprintfPQExpBuffer(&buf,\n>> +\t\t\t\t\t \"SELECT \\n\"\n>> +\t\t\t\t\t \"n.nspname AS \\\"%s\\\", \\n\"\n>> +\t\t\t\t\t \"cst.conname AS \\\"%s\\\", \\n\"\n>> +\t\t\t\t\t \"pg_catalog.pg_get_constraintdef(cst.oid) AS \\\"%s\\\", \\n\"\n>> +\t\t\t\t\t \"c.relname AS \\\"%s\\\" \\n\"\n>> +\t\t\t\t\t \"FROM pg_constraint cst \\n\"\n>> +\t\t\t\t\t \"JOIN pg_namespace n ON n.oid = cst.connamespace \\n\"\n>> +\t\t\t\t\t \"JOIN pg_class c ON c.oid = cst.conrelid \\n\",\n> \n> You should write \"pg_catalog.\" prefix for the tables (in addition to the\n> function).\n\n\nOops, You are right. Fixed.\n\n \n> Rather than join to pg_class, you can write conrelid::pg_catalog.regclass,\n> since regclass is supported since at least v7.3 (but ::regnamespace was\n> introduced in v9.5, so the join against pg_namespace is still necessary).\n> https://www.postgresql.org/docs/9.5/datatype-oid.html\n>> +\tmyopt.title = _(\"List of constsraints\");\n> \n> spelling: constraints\n\n\nThanks! Fixed.\n\n \n> I'm not confident that if I would use this, so let's wait to see if someone\n> else wants to give a +1.\n\n\nOkay, but you agree that there are DBAs and users who want to see the\nlist of constraints, I think.\n\nCurrently, DBAs need the table name to see the constraint information.\nHowever, with this feature, you can see its definition and table name\nfrom the constraint name.\nFor example, it will be easier to understand how many foreign key\nconstraints are in the DB. The \\d command also displays the constraints\nbut does not list them, so this feature is more beneficial for those who\nwant to check them.\n\n\nAttached new patch includes:\n\n - Add a filter by contype\n - Add pg_catalog prefix\n - Fix typo\n - Fix help message to add \\dco\n\nNot implemented yet:\n - NOT NULL constraint, and so on (based on pg_attribute)\n - Tab completion\n - Regression test\n - Document\n\nAny comments welcome! :-D\n\n\nThanks,\nTatsuro Yamada","msg_date":"Tue, 16 Nov 2021 09:22:20 +0900","msg_from":"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>","msg_from_op":true,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"Thanks for updating the patch :)\n\n> Currently, DBAs need the table name to see the constraint information.\n\nOr, they can query pg_constraint or information_schema: check_constraints,\ntable_constraints.\n\n> -\t\t\t\tsuccess = listConversions(pattern, show_verbose, show_system);\n> +\t\t\t\tif (strncmp(cmd, \"dco\", 3) == 0) /* Constraint */\n> +\t\t\t\t\tswitch (cmd[3])\n> +\t\t\t\t\t{\n> +\t\t\t\t\t\tcase '\\0':\n> +\t\t\t\t\t\tcase '+':\n\nDoes \"+\" do anything ?\n\n> +++ b/src/bin/psql/help.c\n> @@ -231,6 +231,7 @@ slashUsage(unsigned short int pager)\n\n> \tfprintf(output, _(\" \\\\db[+] [PATTERN] list tablespaces\\n\"));\n> \tfprintf(output, _(\" \\\\dc[S+] [PATTERN] list conversions\\n\"));\n> +\tfprintf(output, _(\" \\\\dco[S] [PATTERN] list constraint\\n\"));\n\nShould be plural \"constraints\".\n\nI think \"exclude\" should be called \"exclusion\" (\"exclude\" sounded to me like\nyou're going to provide a way to \"exclude\" types of constraints, like \"xc\"\nwould show everything except check constraints).\n\n-- \nJustin\n\n\n","msg_date":"Mon, 15 Nov 2021 20:02:15 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"On Mon, Nov 15, 2021 at 5:23 PM Tatsuro Yamada <\ntatsuro.yamada.tf@nttcom.co.jp> wrote:\n\n>\n> > I'm not confident that if I would use this, so let's wait to see if\n> someone\n> > else wants to give a +1.\n>\n> Okay, but you agree that there are DBAs and users who want to see the\n> list of constraints, I think.\n>\n\nMy opinion is this doesn't exist because there isn't any demand for it.\n\n\n> For example, it will be easier to understand how many foreign key\n> constraints are in the DB.\n\n\nThat isn't a very compelling metric. Metrics also don't seem to be the\nprimary motivation for the psql \\d commands. I envision them mostly useful\nwhen writing a query and wanting a query refresher as to what is\nvalid/available. In that context looking at constraints in the context of\na single table makes sense. Looking at all constraints is considerably\nless so. Especially since constraints mostly impact insert queries and\nthose only affect a single table.\n\nIf the only motivation for this is \"feature completion\" - since we have so\nmany other \\d commands already implemented - I say we should pass.\n\nDavid J.\n\nOn Mon, Nov 15, 2021 at 5:23 PM Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp> wrote:\n> I'm not confident that if I would use this, so let's wait to see if someone\n> else wants to give a +1.\n\nOkay, but you agree that there are DBAs and users who want to see the\nlist of constraints, I think.My opinion is this doesn't exist because there isn't any demand for it. \nFor example, it will be easier to understand how many foreign key\nconstraints are in the DB.That isn't a very compelling metric. Metrics also don't seem to be the primary motivation for the psql \\d commands. I envision them mostly useful when writing a query and wanting a query refresher as to what is valid/available. In that context looking at constraints in the context of a single table makes sense. Looking at all constraints is considerably less so. Especially since constraints mostly impact insert queries and those only affect a single table.If the only motivation for this is \"feature completion\" - since we have so many other \\d commands already implemented - I say we should pass.David J.","msg_date":"Mon, 15 Nov 2021 19:08:56 -0700","msg_from":"\"David G. Johnston\" <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"On Monday, November 15, 2021, David G. Johnston <david.g.johnston@gmail.com>\nwrote:\n\n>\n> If the only motivation for this is \"feature completion\" - since we have so\n> many other \\d commands already implemented - I say we should pass.\n>\n\nIf anything, doing this for triggers would be a much better feature.\n\nThe other missing listing then would be columns. I would rather add both\ncolumns and constraints, or neither. And add triggers regardless as their\ncombination of rarity and performance impact makes having them listable on\na \\d command would be beneficial.\n\nI also noticed that the “\\dd” command would need to be modified (and maybe\ndeprecated if we actually do simply provide a listing for object type).\n\nDavid J.\n\nOn Monday, November 15, 2021, David G. Johnston <david.g.johnston@gmail.com> wrote:If the only motivation for this is \"feature completion\" - since we have so many other \\d commands already implemented - I say we should pass.If anything, doing this for triggers would be a much better feature.The other missing listing then would be columns. I would rather add both columns and constraints, or neither. And add triggers regardless as their combination of rarity and performance impact makes having them listable on a \\d command would be beneficial.I also noticed that the “\\dd” command would need to be modified (and maybe deprecated if we actually do simply provide a listing for object type).David J.","msg_date":"Mon, 15 Nov 2021 19:49:27 -0700","msg_from":"\"David G. Johnston\" <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"Hi Justin,\n\n\nThanks for your comments!\n\n \n>> Currently, DBAs need the table name to see the constraint information.\n> \n> Or, they can query pg_constraint or information_schema: check_constraints,\n> table_constraints.\n\n\nYeah, right.\nIf they can use the meta-command instead of a long query against pg_constraint\nor information_schema and also pg_attribulte, it would be helpful, I believe. :-D\n\n\n\n>> -\t\t\t\tsuccess = listConversions(pattern, show_verbose, show_system);\n>> +\t\t\t\tif (strncmp(cmd, \"dco\", 3) == 0) /* Constraint */\n>> +\t\t\t\t\tswitch (cmd[3])\n>> +\t\t\t\t\t{\n>> +\t\t\t\t\t\tcase '\\0':\n>> +\t\t\t\t\t\tcase '+':\n> \n> Does \"+\" do anything ?\n\n\nNo, it doesn't. Removed.\n\n\n \n>> +++ b/src/bin/psql/help.c\n>> @@ -231,6 +231,7 @@ slashUsage(unsigned short int pager)\n> \n>> \tfprintf(output, _(\" \\\\db[+] [PATTERN] list tablespaces\\n\"));\n>> \tfprintf(output, _(\" \\\\dc[S+] [PATTERN] list conversions\\n\"));\n>> +\tfprintf(output, _(\" \\\\dco[S] [PATTERN] list constraint\\n\"));\n> \n> Should be plural \"constraints\".\n> \n> I think \"exclude\" should be called \"exclusion\" (\"exclude\" sounded to me like\n> you're going to provide a way to \"exclude\" types of constraints, like \"xc\"\n> would show everything except check constraints).\n\n\nThanks! Fixed the both.\n\n\nAttached file is new patch. It includes:\n\n - Fix help message s/constraint/constraints/\n - s/Exclude/Exclusion/\n - Remove unused modifier \"+\"\n - Add document for \\dco\n\n\n\nThanks,\nTatsuro Yamada","msg_date":"Tue, 16 Nov 2021 12:47:31 +0900","msg_from":"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>","msg_from_op":true,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"Hi David,\n\nThanks for your comments.\n\n> Okay, but you agree that there are DBAs and users who want to see the\n> list of constraints, I think.\n> \n> \n> My opinion is this doesn't exist because there isn't any demand for it.\n\n\nI don't know if this is a good example, but if you look at StackOverflow,\nit seems that people who want to see a list of constraints appear regularly.\n(The other day, I also wanted to see the list, and I arrived at this site)\nTherefore, the only thing that hasn't been implemented so far is that no one\ncould communicate the request to -hackers, and I think there is demand.\n\nhttps://stackoverflow.com/questions/62987794/how-to-list-all-constraints-of-a-table-in-postgresql\n\n\nRegards,\nTatsuro Yamada\n\n\n\n\n\n\n\n\n","msg_date":"Tue, 16 Nov 2021 13:11:20 +0900","msg_from":"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>","msg_from_op":true,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"On Monday, November 15, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\nwrote:\n>\n>\n> I don't know if this is a good example, but if you look at StackOverflow,\n> it seems that people who want to see a list of constraints appear\n> regularly.\n\n\nhttps://stackoverflow.com/questions/62987794/how-to-list-all\n> -constraints-of-a-table-in-postgresql\n>\n>\nGiven the questioner restricted their question to “for a given table” I’d\nsay it supports leaving the status quo, not changing it.\n\nIf, as you suppose, these come up regularly then finding one that asks for\nit “in the entire database”, ideally with some stated goal, should be\ndoable.\n\n David J.\n\nOn Monday, November 15, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp> wrote:\n\nI don't know if this is a good example, but if you look at StackOverflow,\nit seems that people who want to see a list of constraints appear regularly.https://stackoverflow.com/questions/62987794/how-to-list-all-constraints-of-a-table-in-postgresql\nGiven the questioner restricted their question to “for a given table” I’d say it supports leaving the status quo, not changing it.If, as you suppose, these come up regularly then finding one that asks for it “in the entire database”, ideally with some stated goal, should be doable. David J.","msg_date":"Mon, 15 Nov 2021 21:23:04 -0700","msg_from":"\"David G. Johnston\" <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Add psql command to list constraints"},{"msg_contents":"\"David G. Johnston\" <david.g.johnston@gmail.com> writes:\n\n> On Monday, November 15, 2021, Tatsuro Yamada\n> <tatsuro.yamada.tf@nttcom.co.jp> wrote:\n>\n> I don't know if this is a good example, but if you look at\n> StackOverflow,\n> it seems that people who want to see a list of constraints appear\n> regularly.\n>\n> https://stackoverflow.com/questions/62987794/how-to-list-all-constraints-\n> of-a-table-in-postgresql\n> \n>\n> Given the questioner restricted their question to “for a given table”\n> I’d say it supports leaving the status quo, not changing it.\n>\n> If, as you suppose, these come up regularly then finding one that asks\n> for it “in the entire database”, ideally with some stated goal, should\n> be doable.\n>\n> David J.\n>\n>\n\nThis is my review of the patch in\nhttps://commitfest.postgresql.org/37/3468/\n\nThe patch adds the command \"\\dco\" to list constraints in psql. This\nseems useful to me.\n\nThe patch applies cleanly to HEAD, although some hunks have rather large\noffsets.\n\nAs far as I can tell, the \"\\dco\" command works as documented.\n\nI have however found the following issues with the patch:\n\n* A TAB character has been added to doc/src/sgml/ref/psql-ref.sgml -\n this should be replaced with spaces.\n* The call to listConstraints in line src/bin/psql/command.c 794 refers\n to &cmd[2], this should rather be &cmd[3].\n* The patch kills the \"\\dc\" command in src/bin/psql/command.c\n This can be fixed by adding the following at line 800:\n else\n success =\n listConversions(pattern, show_verbose, show_system);\n\nAnother comment is that the \"\\dco\" command outputs quite a lot of\ninformation, which only fits in a wide terminal window. Would it be an\nidea to only display the columns \"Schema\" and \"Name\" by default, and\nuse \"+\" to specify inclusion of the columns \"Definition\" and \"Table\".\n\n\nBest regards\n\nDag Lem\n\n\n","msg_date":"Fri, 04 Feb 2022 12:39:09 +0100","msg_from":"Dag Lem <dag@nimrod.no>","msg_from_op":false,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"Development of this seems to have stalled with the only review of this\npatch expressing some skepticism about whether it's needed at all.\n\nUnless anyone steps forward and says it's interesting I'm going to\nmark it rejected?\n\nI don't actually think it's a terrible idea myself but I think every\nuse case I might have had for it was solved better by directly\nquerying catalog tables.\n\n\n","msg_date":"Fri, 25 Mar 2022 00:27:46 -0400","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"On Fri, Mar 25, 2022 at 12:28 AM Greg Stark <stark@mit.edu> wrote:\n> Development of this seems to have stalled with the only review of this\n> patch expressing some skepticism about whether it's needed at all.\n\nMy opinion on this patch is that we typically handle objects that are\nessentially table properties by showing the output in \\d+ on the\nindividual table. And that already works just fine:\n\nrhaas=# create table duck (quack int unique, check (quack > 0));\nCREATE TABLE\nrhaas=# \\d+ duck\n Table \"public.duck\"\n Column | Type | Collation | Nullable | Default | Storage |\nCompression | Stats target | Description\n--------+---------+-----------+----------+---------+---------+-------------+--------------+-------------\n quack | integer | | | | plain |\n | |\nIndexes:\n \"duck_quack_key\" UNIQUE CONSTRAINT, btree (quack)\nCheck constraints:\n \"duck_quack_check\" CHECK (quack > 0)\nAccess method: heap\n\nNow, there is some precedent for the idea of providing a command that\nlists everything globally. Specifically, we have a \\dp command, also\nknown as \\z, to list privileges across all objects in the database.\nHowever, I have found that command to be relatively useless, because\nif you've got any significant number of grants in the database it just\nproduces too much output. I think this would have the same problem.\nThe other thing that we have that is somewhat similar to this is \\dd,\nwhich lists comments \"for object types which do not have their\ncomments displayed by their own backslash commands.\" However, it says\nthat the object types that it covers are \"constraint, operator class,\noperator family, rule, and trigger,\" and that list is out of date,\nbecause operator classes and families got their own backslash commands\ntwo years ago. We could update that, but honestly I can't see anyone\nbeing too excited about a command that lists comments on every\nconstraint, rule, and trigger in the system: it would be a lot more\nuseful to show those commands in the \\d+ output for the table to which\nthey are bound, and get rid of \\dd (and maybe \\dp and \\z too).\n\nNow that is not to say that what is being proposed here is completely\nuseless. It clearly isn't. It's totally debatable whether we ought to\nstart having commands like this, and maybe we should. It would make\nfor more commands, and that's not entirely great because the command\nnames are increasingly alphabet soup. Who can remember what \\drds or\n\\dAc does? Only real power users. If we add more, it's going to get\neven more difficult, but some people will use it and like it and those\npeople will be happy. It's kind of hard to say whether we'd come out\nahead or not. What I think is fairly certain is that it would\nrepresent a reversal of our current policy.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 25 Mar 2022 15:11:47 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"On Fri, Mar 25, 2022 at 03:11:47PM -0400, Robert Haas wrote:\n> On Fri, Mar 25, 2022 at 12:28 AM Greg Stark <stark@mit.edu> wrote:\n> > Development of this seems to have stalled with the only review of this\n> > patch expressing some skepticism about whether it's needed at all.\n> \n> Now, there is some precedent for the idea of providing a command that\n> lists everything globally. Specifically, we have a \\dp command, also\n> known as \\z, to list privileges across all objects in the database.\n\n> The other thing that we have that is somewhat similar to this is \\dd,\n\n\\dX is similar, and I remember wondering whether it was really useful/needed.\nThe catalog tables are exposed and documented for a reason, and power-users\nwill learn to use them.\n\n+0.5 to mark the patch as RWF or rejected.\n\n-- \nJustin\n\n\n","msg_date":"Fri, 25 Mar 2022 14:20:25 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"On Fri, Mar 25, 2022 at 3:20 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> \\dX is similar, and I remember wondering whether it was really useful/needed.\n> The catalog tables are exposed and documented for a reason, and power-users\n> will learn to use them.\n\nI don't think that \\dX is comparable, because I don't think we should\nregard extended statistics as a table object. Indeed, generalizing\nextended statistics so that they can be generated on a join seems to\nme to be one of the most important things we could do in that area.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 25 Mar 2022 15:24:34 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"In the interests of trying to clean up the CF and keep things moving\nI'll mark the patch rejected.\n\nThat doesn't mean the decision can't change or that nobody's allowed\nto keep discussing it. It's just that that seems to be the decision\nright now and there are too many patches queued up to keep things in a\npending state waiting for a more decisive conclusion. We can always\nchange it if the winds shift...\n\n\n","msg_date":"Mon, 28 Mar 2022 14:59:20 -0400","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"Hi All,\n\n> In the interests of trying to clean up the CF and keep things moving\n> I'll mark the patch rejected.\n\nThank you for managing the commitfest and the comments from many of\nhackers. I apologize for not being able to email you more often due to\nmy busy day job.\n\nFirst of all, I understand to a certain extent your opinion that we\ncan use \\d and look at the constraints on a table-by-table basis as a\nway to check the constraints.\nHowever, suppose We want to reverse lookup a table from a constraint.\nIn that case, there are two ways: (1) use \"\\d\" to lookup all tables,\n(2) execute a select statement against pg_constraint. I think the\nproposed function is more valuable than these.\n\n From another perspective, when looking at the comprehensiveness of\nmetacommands in psql, it seems that only functions that focus on\nconstraints do not exist. Therefore, It would be better to add it in\nterms of comprehensiveness.\n\nI think there is room for other discussions about this patch. Still,\nat least there are people (including myself) who think it is useful.\nI don't think there is anything wrong with this patch that would\ncomplicate the code or cause performance degradation, so I would like to\ncontinue developing it for those who want to use it.\n\nHowever, I understand that it will not be ready in time for PG15, so\nI would like to move forward with PG16. Therefore, the status of the\npatch would be better by changing \"Waiting for Author\" rather than\n\"Rejected\".\n\nP.S.\nI'll send a new patch addressing Dag's comments in the next email.\n\nThanks,\nTatsuro Yamada\n\n\n\n\n","msg_date":"Tue, 29 Mar 2022 12:15:45 +0900","msg_from":"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>","msg_from_op":true,"msg_subject":"Re: Add psql command to list constraints"},{"msg_contents":"Hi Dag,\n\n> The patch adds the command \"\\dco\" to list constraints in psql. This\n> seems useful to me.\n\nThank you!\n\n \n> The patch applies cleanly to HEAD, although some hunks have rather large\n> offsets.\n>\n> As far as I can tell, the \"\\dco\" command works as documented.\n>\n> I have however found the following issues with the patch:\n>\n> * A TAB character has been added to doc/src/sgml/ref/psql-ref.sgml -\n> this should be replaced with spaces.\n\nFixed.\n\n\n> * The call to listConstraints in line src/bin/psql/command.c 794 refers\n> to &cmd[2], this should rather be &cmd[3].\n>\n> * The patch kills the \"\\dc\" command in src/bin/psql/command.c\n> This can be fixed by adding the following at line 800:\n> else\n> success =\n> listConversions(pattern, show_verbose, show_system);\n\n\nOh, you are right! Fixed.\n\n \n> Another comment is that the \"\\dco\" command outputs quite a lot of\n> information, which only fits in a wide terminal window. Would it be an\n> idea to only display the columns \"Schema\" and \"Name\" by default, and\n> use \"+\" to specify inclusion of the columns \"Definition\" and \"Table\".\n \n\nI fixed the output columns as you proposed.\n\nThe current status of this patch is:\n\n - Addressed Dag's comments\n - Not implemented yet:\n - Tab completion\n - Regression test\n - NOT NULL constraint, and so on (based on pg_attribute)\n\nPlease find attached new patch.\n\n\nThanks,\nTatsuro Yamada","msg_date":"Tue, 29 Mar 2022 14:21:07 +0900","msg_from":"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>","msg_from_op":true,"msg_subject":"Re: Add psql command to list constraints"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI have been wondering why there is no meta-command for listing\\nconstraints in psql. So, I created a POC patch by using my\\nexperience developing \\\\dX command in PG14.\\n\\nThis feature is helpful for DBAs when they want to check or\\nmodify the definition of constraints.\\n\\nThe current status of the POC patch is as follows:\\n\\n - Add \\\"\\\\dco\\\" command to list constraints from pg_constraint\\n\\n - Not implemented yet:\\n - NOT NULL constraint, and so on (based on pg_attribute)\\n - Tab completion\\n - Regression test\\n - Document\\n\\nThe following is test results (See attached test_list_con.sql)\\n====================================================================\\npostgres=# \\\\dco\\n List of constsraints\\n Schema | Name | Definition | Table\\n--------+-------------------------+---------------------------------------------------------+----------\\n public | t01_chk_price_check | CHECK ((price > (0)::numeric)) | t01_chk\\n public | t02_uniq_product_no_key | UNIQUE (product_no) | t02_uniq\\n public | t03_pk1_pkey | PRIMARY KEY (product_no) | t03_pk1\\n public | t03_pk2_product_no_key | UNIQUE (product_no) | t03_pk2\\n public | t04_fk_pkey | PRIMARY KEY (order_id) | t04_fk\\n public | t04_fk_product_no_fkey | FOREIGN KEY (product_no) REFERENCES t03_pk1(product_no) | t04_fk\\n public | t05_ex_c_excl | EXCLUDE USING gist (c WITH &&) | t05_ex\\n(7 rows)\\n====================================================================\\n\\n\\nI have the following two questions that need to be discussed.\\n\\nQuestions:\\n(1) What strings should be assigned as meta-command for this feature?\\n Currently, \\\\dc and \\\\dC are not available, so I tentatively\\n assigned \\\\dco. However, I do not have a strong opinion, so please\\n let me know if you have any suggestions.\\n\\n(2) About domain constraints\\n There is the \\\\dD command to show a list of domain constraints.\\n So I think this feature should not include it. Is it Okay?\\n\\n\\nIf I can get \\\"+1\\\" for this new feature development, I would like to\\nimprove the patch by adding NOT NULL constraints, and so on.\\nAny advice or comments would be appreciated.\\n\\n\\nThanks,\\nTatsuro Yamada\",\n \"msg_date\": \"Mon, 15 Nov 2021 10:38:55 +0900\",\n \"msg_from\": \"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Mon, Nov 15, 2021 at 10:38:55AM +0900, Tatsuro Yamada wrote:\\n> postgres=# \\\\dco\\n> List of constsraints\\n> Schema | Name | Definition | Table\\n> --------+-------------------------+---------------------------------------------------------+----------\\n> public | t01_chk_price_check | CHECK ((price > (0)::numeric)) | t01_chk\\n> public | t02_uniq_product_no_key | UNIQUE (product_no) | t02_uniq\\n> public | t03_pk1_pkey | PRIMARY KEY (product_no) | t03_pk1\\n> public | t03_pk2_product_no_key | UNIQUE (product_no) | t03_pk2\\n> public | t04_fk_pkey | PRIMARY KEY (order_id) | t04_fk\\n> public | t04_fk_product_no_fkey | FOREIGN KEY (product_no) REFERENCES t03_pk1(product_no) | t04_fk\\n> public | t05_ex_c_excl | EXCLUDE USING gist (c WITH &&) | t05_ex\\n> (7 rows)\\n> ====================================================================\\n\\nMaybe it ought to be possible to choose the type of constraints to show.\\nSimilar to how \\\\dt shows tables and \\\\di shows indexes and \\\\dti shows\\ntables+inds, you could run \\\\dcoc for check constraints and \\\\dcof for foreign\\nkeys. But I think \\\"\\\\dco\\\" is too long of a prefix...\\n\\n> +\\tinitPQExpBuffer(&buf);\\n> +\\tprintfPQExpBuffer(&buf,\\n> +\\t\\t\\t\\t\\t \\\"SELECT \\\\n\\\"\\n> +\\t\\t\\t\\t\\t \\\"n.nspname AS \\\\\\\"%s\\\\\\\", \\\\n\\\"\\n> +\\t\\t\\t\\t\\t \\\"cst.conname AS \\\\\\\"%s\\\\\\\", \\\\n\\\"\\n> +\\t\\t\\t\\t\\t \\\"pg_catalog.pg_get_constraintdef(cst.oid) AS \\\\\\\"%s\\\\\\\", \\\\n\\\"\\n> +\\t\\t\\t\\t\\t \\\"c.relname AS \\\\\\\"%s\\\\\\\" \\\\n\\\"\\n> +\\t\\t\\t\\t\\t \\\"FROM pg_constraint cst \\\\n\\\"\\n> +\\t\\t\\t\\t\\t \\\"JOIN pg_namespace n ON n.oid = cst.connamespace \\\\n\\\"\\n> +\\t\\t\\t\\t\\t \\\"JOIN pg_class c ON c.oid = cst.conrelid \\\\n\\\",\\n\\nYou should write \\\"pg_catalog.\\\" prefix for the tables (in addition to the\\nfunction).\\n\\nRather than join to pg_class, you can write conrelid::pg_catalog.regclass,\\nsince regclass is supported since at least v7.3 (but ::regnamespace was\\nintroduced in v9.5, so the join against pg_namespace is still necessary).\\nhttps://www.postgresql.org/docs/9.5/datatype-oid.html\\n\\n> +\\tmyopt.title = _(\\\"List of constsraints\\\");\\n\\nspelling: constraints\\n\\nI'm not confident that if I would use this, so let's wait to see if someone\\nelse wants to give a +1.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sun, 14 Nov 2021 21:56:33 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"Hi Justin,\\n\\nThanks for your comments and review!\\n\\n> Maybe it ought to be possible to choose the type of constraints to show.\\n> Similar to how \\\\dt shows tables and \\\\di shows indexes and \\\\dti shows\\n> tables+inds, you could run \\\\dcoc for check constraints and \\\\dcof for foreign\\n> keys. But I think \\\"\\\\dco\\\" is too long of a prefix...\\n\\n\\nYeah, agreed.\\nI added a filter based on the type of constraints:\\n - c for check\\n - f for foreign key\\n - p for primary key\\n - t for trigger\\n - u for unique\\n - x for exclude c, f, p, u, t, and x.\\n\\nThe following is examples of \\\\dcop, \\\\dcof, and \\\\dcopf.\\n\\n========================================================================\\npostgres=# \\\\dcop\\n List of constraints\\n Schema | Name | Definition | Table\\n--------+--------------+--------------------------+---------\\n public | t03_pk1_pkey | PRIMARY KEY (product_no) | t03_pk1\\n public | t04_fk_pkey | PRIMARY KEY (order_id) | t04_fk\\n(2 rows)\\n\\npostgres=# \\\\dcof\\n List of constraints\\n Schema | Name | Definition | Table\\n--------+------------------------+---------------------------------------------------------+--------\\n public | t04_fk_product_no_fkey | FOREIGN KEY (product_no) REFERENCES t03_pk1(product_no) | t04_fk\\n(1 row)\\n\\npostgres=# \\\\dcopf\\n List of constraints\\n Schema | Name | Definition | Table\\n--------+------------------------+---------------------------------------------------------+---------\\n public | t03_pk1_pkey | PRIMARY KEY (product_no) | t03_pk1\\n public | t04_fk_pkey | PRIMARY KEY (order_id) | t04_fk\\n public | t04_fk_product_no_fkey | FOREIGN KEY (product_no) REFERENCES t03_pk1(product_no) | t04_fk\\n(3 rows)\\n========================================================================\\n\\n\\nI too think \\\\dco is a long name. So, I'd like to get suggestions to be more shortened. :)\\n\\n\\n \\n>> +\\tinitPQExpBuffer(&buf);\\n>> +\\tprintfPQExpBuffer(&buf,\\n>> +\\t\\t\\t\\t\\t \\\"SELECT \\\\n\\\"\\n>> +\\t\\t\\t\\t\\t \\\"n.nspname AS \\\\\\\"%s\\\\\\\", \\\\n\\\"\\n>> +\\t\\t\\t\\t\\t \\\"cst.conname AS \\\\\\\"%s\\\\\\\", \\\\n\\\"\\n>> +\\t\\t\\t\\t\\t \\\"pg_catalog.pg_get_constraintdef(cst.oid) AS \\\\\\\"%s\\\\\\\", \\\\n\\\"\\n>> +\\t\\t\\t\\t\\t \\\"c.relname AS \\\\\\\"%s\\\\\\\" \\\\n\\\"\\n>> +\\t\\t\\t\\t\\t \\\"FROM pg_constraint cst \\\\n\\\"\\n>> +\\t\\t\\t\\t\\t \\\"JOIN pg_namespace n ON n.oid = cst.connamespace \\\\n\\\"\\n>> +\\t\\t\\t\\t\\t \\\"JOIN pg_class c ON c.oid = cst.conrelid \\\\n\\\",\\n> \\n> You should write \\\"pg_catalog.\\\" prefix for the tables (in addition to the\\n> function).\\n\\n\\nOops, You are right. Fixed.\\n\\n \\n> Rather than join to pg_class, you can write conrelid::pg_catalog.regclass,\\n> since regclass is supported since at least v7.3 (but ::regnamespace was\\n> introduced in v9.5, so the join against pg_namespace is still necessary).\\n> https://www.postgresql.org/docs/9.5/datatype-oid.html\\n>> +\\tmyopt.title = _(\\\"List of constsraints\\\");\\n> \\n> spelling: constraints\\n\\n\\nThanks! Fixed.\\n\\n \\n> I'm not confident that if I would use this, so let's wait to see if someone\\n> else wants to give a +1.\\n\\n\\nOkay, but you agree that there are DBAs and users who want to see the\\nlist of constraints, I think.\\n\\nCurrently, DBAs need the table name to see the constraint information.\\nHowever, with this feature, you can see its definition and table name\\nfrom the constraint name.\\nFor example, it will be easier to understand how many foreign key\\nconstraints are in the DB. The \\\\d command also displays the constraints\\nbut does not list them, so this feature is more beneficial for those who\\nwant to check them.\\n\\n\\nAttached new patch includes:\\n\\n - Add a filter by contype\\n - Add pg_catalog prefix\\n - Fix typo\\n - Fix help message to add \\\\dco\\n\\nNot implemented yet:\\n - NOT NULL constraint, and so on (based on pg_attribute)\\n - Tab completion\\n - Regression test\\n - Document\\n\\nAny comments welcome! :-D\\n\\n\\nThanks,\\nTatsuro Yamada\",\n \"msg_date\": \"Tue, 16 Nov 2021 09:22:20 +0900\",\n \"msg_from\": \"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"Thanks for updating the patch :)\\n\\n> Currently, DBAs need the table name to see the constraint information.\\n\\nOr, they can query pg_constraint or information_schema: check_constraints,\\ntable_constraints.\\n\\n> -\\t\\t\\t\\tsuccess = listConversions(pattern, show_verbose, show_system);\\n> +\\t\\t\\t\\tif (strncmp(cmd, \\\"dco\\\", 3) == 0) /* Constraint */\\n> +\\t\\t\\t\\t\\tswitch (cmd[3])\\n> +\\t\\t\\t\\t\\t{\\n> +\\t\\t\\t\\t\\t\\tcase '\\\\0':\\n> +\\t\\t\\t\\t\\t\\tcase '+':\\n\\nDoes \\\"+\\\" do anything ?\\n\\n> +++ b/src/bin/psql/help.c\\n> @@ -231,6 +231,7 @@ slashUsage(unsigned short int pager)\\n\\n> \\tfprintf(output, _(\\\" \\\\\\\\db[+] [PATTERN] list tablespaces\\\\n\\\"));\\n> \\tfprintf(output, _(\\\" \\\\\\\\dc[S+] [PATTERN] list conversions\\\\n\\\"));\\n> +\\tfprintf(output, _(\\\" \\\\\\\\dco[S] [PATTERN] list constraint\\\\n\\\"));\\n\\nShould be plural \\\"constraints\\\".\\n\\nI think \\\"exclude\\\" should be called \\\"exclusion\\\" (\\\"exclude\\\" sounded to me like\\nyou're going to provide a way to \\\"exclude\\\" types of constraints, like \\\"xc\\\"\\nwould show everything except check constraints).\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 20:02:15 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 5:23 PM Tatsuro Yamada <\\ntatsuro.yamada.tf@nttcom.co.jp> wrote:\\n\\n>\\n> > I'm not confident that if I would use this, so let's wait to see if\\n> someone\\n> > else wants to give a +1.\\n>\\n> Okay, but you agree that there are DBAs and users who want to see the\\n> list of constraints, I think.\\n>\\n\\nMy opinion is this doesn't exist because there isn't any demand for it.\\n\\n\\n> For example, it will be easier to understand how many foreign key\\n> constraints are in the DB.\\n\\n\\nThat isn't a very compelling metric. Metrics also don't seem to be the\\nprimary motivation for the psql \\\\d commands. I envision them mostly useful\\nwhen writing a query and wanting a query refresher as to what is\\nvalid/available. In that context looking at constraints in the context of\\na single table makes sense. Looking at all constraints is considerably\\nless so. Especially since constraints mostly impact insert queries and\\nthose only affect a single table.\\n\\nIf the only motivation for this is \\\"feature completion\\\" - since we have so\\nmany other \\\\d commands already implemented - I say we should pass.\\n\\nDavid J.\\n\\nOn Mon, Nov 15, 2021 at 5:23 PM Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp> wrote:\\n> I'm not confident that if I would use this, so let's wait to see if someone\\n> else wants to give a +1.\\n\\nOkay, but you agree that there are DBAs and users who want to see the\\nlist of constraints, I think.My opinion is this doesn't exist because there isn't any demand for it. \\nFor example, it will be easier to understand how many foreign key\\nconstraints are in the DB.That isn't a very compelling metric. Metrics also don't seem to be the primary motivation for the psql \\\\d commands. I envision them mostly useful when writing a query and wanting a query refresher as to what is valid/available. In that context looking at constraints in the context of a single table makes sense. Looking at all constraints is considerably less so. Especially since constraints mostly impact insert queries and those only affect a single table.If the only motivation for this is \\\"feature completion\\\" - since we have so many other \\\\d commands already implemented - I say we should pass.David J.\",\n \"msg_date\": \"Mon, 15 Nov 2021 19:08:56 -0700\",\n \"msg_from\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"On Monday, November 15, 2021, David G. Johnston <david.g.johnston@gmail.com>\\nwrote:\\n\\n>\\n> If the only motivation for this is \\\"feature completion\\\" - since we have so\\n> many other \\\\d commands already implemented - I say we should pass.\\n>\\n\\nIf anything, doing this for triggers would be a much better feature.\\n\\nThe other missing listing then would be columns. I would rather add both\\ncolumns and constraints, or neither. And add triggers regardless as their\\ncombination of rarity and performance impact makes having them listable on\\na \\\\d command would be beneficial.\\n\\nI also noticed that the “\\\\dd” command would need to be modified (and maybe\\ndeprecated if we actually do simply provide a listing for object type).\\n\\nDavid J.\\n\\nOn Monday, November 15, 2021, David G. Johnston <david.g.johnston@gmail.com> wrote:If the only motivation for this is \\\"feature completion\\\" - since we have so many other \\\\d commands already implemented - I say we should pass.If anything, doing this for triggers would be a much better feature.The other missing listing then would be columns. I would rather add both columns and constraints, or neither. And add triggers regardless as their combination of rarity and performance impact makes having them listable on a \\\\d command would be beneficial.I also noticed that the “\\\\dd” command would need to be modified (and maybe deprecated if we actually do simply provide a listing for object type).David J.\",\n \"msg_date\": \"Mon, 15 Nov 2021 19:49:27 -0700\",\n \"msg_from\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"Hi Justin,\\n\\n\\nThanks for your comments!\\n\\n \\n>> Currently, DBAs need the table name to see the constraint information.\\n> \\n> Or, they can query pg_constraint or information_schema: check_constraints,\\n> table_constraints.\\n\\n\\nYeah, right.\\nIf they can use the meta-command instead of a long query against pg_constraint\\nor information_schema and also pg_attribulte, it would be helpful, I believe. :-D\\n\\n\\n\\n>> -\\t\\t\\t\\tsuccess = listConversions(pattern, show_verbose, show_system);\\n>> +\\t\\t\\t\\tif (strncmp(cmd, \\\"dco\\\", 3) == 0) /* Constraint */\\n>> +\\t\\t\\t\\t\\tswitch (cmd[3])\\n>> +\\t\\t\\t\\t\\t{\\n>> +\\t\\t\\t\\t\\t\\tcase '\\\\0':\\n>> +\\t\\t\\t\\t\\t\\tcase '+':\\n> \\n> Does \\\"+\\\" do anything ?\\n\\n\\nNo, it doesn't. Removed.\\n\\n\\n \\n>> +++ b/src/bin/psql/help.c\\n>> @@ -231,6 +231,7 @@ slashUsage(unsigned short int pager)\\n> \\n>> \\tfprintf(output, _(\\\" \\\\\\\\db[+] [PATTERN] list tablespaces\\\\n\\\"));\\n>> \\tfprintf(output, _(\\\" \\\\\\\\dc[S+] [PATTERN] list conversions\\\\n\\\"));\\n>> +\\tfprintf(output, _(\\\" \\\\\\\\dco[S] [PATTERN] list constraint\\\\n\\\"));\\n> \\n> Should be plural \\\"constraints\\\".\\n> \\n> I think \\\"exclude\\\" should be called \\\"exclusion\\\" (\\\"exclude\\\" sounded to me like\\n> you're going to provide a way to \\\"exclude\\\" types of constraints, like \\\"xc\\\"\\n> would show everything except check constraints).\\n\\n\\nThanks! Fixed the both.\\n\\n\\nAttached file is new patch. It includes:\\n\\n - Fix help message s/constraint/constraints/\\n - s/Exclude/Exclusion/\\n - Remove unused modifier \\\"+\\\"\\n - Add document for \\\\dco\\n\\n\\n\\nThanks,\\nTatsuro Yamada\",\n \"msg_date\": \"Tue, 16 Nov 2021 12:47:31 +0900\",\n \"msg_from\": \"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"Hi David,\\n\\nThanks for your comments.\\n\\n> Okay, but you agree that there are DBAs and users who want to see the\\n> list of constraints, I think.\\n> \\n> \\n> My opinion is this doesn't exist because there isn't any demand for it.\\n\\n\\nI don't know if this is a good example, but if you look at StackOverflow,\\nit seems that people who want to see a list of constraints appear regularly.\\n(The other day, I also wanted to see the list, and I arrived at this site)\\nTherefore, the only thing that hasn't been implemented so far is that no one\\ncould communicate the request to -hackers, and I think there is demand.\\n\\nhttps://stackoverflow.com/questions/62987794/how-to-list-all-constraints-of-a-table-in-postgresql\\n\\n\\nRegards,\\nTatsuro Yamada\\n\\n\\n\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 13:11:20 +0900\",\n \"msg_from\": \"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"On Monday, November 15, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\\nwrote:\\n>\\n>\\n> I don't know if this is a good example, but if you look at StackOverflow,\\n> it seems that people who want to see a list of constraints appear\\n> regularly.\\n\\n\\nhttps://stackoverflow.com/questions/62987794/how-to-list-all\\n> -constraints-of-a-table-in-postgresql\\n>\\n>\\nGiven the questioner restricted their question to “for a given table” I’d\\nsay it supports leaving the status quo, not changing it.\\n\\nIf, as you suppose, these come up regularly then finding one that asks for\\nit “in the entire database”, ideally with some stated goal, should be\\ndoable.\\n\\n David J.\\n\\nOn Monday, November 15, 2021, Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp> wrote:\\n\\nI don't know if this is a good example, but if you look at StackOverflow,\\nit seems that people who want to see a list of constraints appear regularly.https://stackoverflow.com/questions/62987794/how-to-list-all-constraints-of-a-table-in-postgresql\\nGiven the questioner restricted their question to “for a given table” I’d say it supports leaving the status quo, not changing it.If, as you suppose, these come up regularly then finding one that asks for it “in the entire database”, ideally with some stated goal, should be doable. David J.\",\n \"msg_date\": \"Mon, 15 Nov 2021 21:23:04 -0700\",\n \"msg_from\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com> writes:\\n\\n> On Monday, November 15, 2021, Tatsuro Yamada\\n> <tatsuro.yamada.tf@nttcom.co.jp> wrote:\\n>\\n> I don't know if this is a good example, but if you look at\\n> StackOverflow,\\n> it seems that people who want to see a list of constraints appear\\n> regularly.\\n>\\n> https://stackoverflow.com/questions/62987794/how-to-list-all-constraints-\\n> of-a-table-in-postgresql\\n> \\n>\\n> Given the questioner restricted their question to “for a given table”\\n> I’d say it supports leaving the status quo, not changing it.\\n>\\n> If, as you suppose, these come up regularly then finding one that asks\\n> for it “in the entire database”, ideally with some stated goal, should\\n> be doable.\\n>\\n> David J.\\n>\\n>\\n\\nThis is my review of the patch in\\nhttps://commitfest.postgresql.org/37/3468/\\n\\nThe patch adds the command \\\"\\\\dco\\\" to list constraints in psql. This\\nseems useful to me.\\n\\nThe patch applies cleanly to HEAD, although some hunks have rather large\\noffsets.\\n\\nAs far as I can tell, the \\\"\\\\dco\\\" command works as documented.\\n\\nI have however found the following issues with the patch:\\n\\n* A TAB character has been added to doc/src/sgml/ref/psql-ref.sgml -\\n this should be replaced with spaces.\\n* The call to listConstraints in line src/bin/psql/command.c 794 refers\\n to &cmd[2], this should rather be &cmd[3].\\n* The patch kills the \\\"\\\\dc\\\" command in src/bin/psql/command.c\\n This can be fixed by adding the following at line 800:\\n else\\n success =\\n listConversions(pattern, show_verbose, show_system);\\n\\nAnother comment is that the \\\"\\\\dco\\\" command outputs quite a lot of\\ninformation, which only fits in a wide terminal window. Would it be an\\nidea to only display the columns \\\"Schema\\\" and \\\"Name\\\" by default, and\\nuse \\\"+\\\" to specify inclusion of the columns \\\"Definition\\\" and \\\"Table\\\".\\n\\n\\nBest regards\\n\\nDag Lem\\n\\n\\n\",\n \"msg_date\": \"Fri, 04 Feb 2022 12:39:09 +0100\",\n \"msg_from\": \"Dag Lem <dag@nimrod.no>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"Development of this seems to have stalled with the only review of this\\npatch expressing some skepticism about whether it's needed at all.\\n\\nUnless anyone steps forward and says it's interesting I'm going to\\nmark it rejected?\\n\\nI don't actually think it's a terrible idea myself but I think every\\nuse case I might have had for it was solved better by directly\\nquerying catalog tables.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 00:27:46 -0400\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 25, 2022 at 12:28 AM Greg Stark <stark@mit.edu> wrote:\\n> Development of this seems to have stalled with the only review of this\\n> patch expressing some skepticism about whether it's needed at all.\\n\\nMy opinion on this patch is that we typically handle objects that are\\nessentially table properties by showing the output in \\\\d+ on the\\nindividual table. And that already works just fine:\\n\\nrhaas=# create table duck (quack int unique, check (quack > 0));\\nCREATE TABLE\\nrhaas=# \\\\d+ duck\\n Table \\\"public.duck\\\"\\n Column | Type | Collation | Nullable | Default | Storage |\\nCompression | Stats target | Description\\n--------+---------+-----------+----------+---------+---------+-------------+--------------+-------------\\n quack | integer | | | | plain |\\n | |\\nIndexes:\\n \\\"duck_quack_key\\\" UNIQUE CONSTRAINT, btree (quack)\\nCheck constraints:\\n \\\"duck_quack_check\\\" CHECK (quack > 0)\\nAccess method: heap\\n\\nNow, there is some precedent for the idea of providing a command that\\nlists everything globally. Specifically, we have a \\\\dp command, also\\nknown as \\\\z, to list privileges across all objects in the database.\\nHowever, I have found that command to be relatively useless, because\\nif you've got any significant number of grants in the database it just\\nproduces too much output. I think this would have the same problem.\\nThe other thing that we have that is somewhat similar to this is \\\\dd,\\nwhich lists comments \\\"for object types which do not have their\\ncomments displayed by their own backslash commands.\\\" However, it says\\nthat the object types that it covers are \\\"constraint, operator class,\\noperator family, rule, and trigger,\\\" and that list is out of date,\\nbecause operator classes and families got their own backslash commands\\ntwo years ago. We could update that, but honestly I can't see anyone\\nbeing too excited about a command that lists comments on every\\nconstraint, rule, and trigger in the system: it would be a lot more\\nuseful to show those commands in the \\\\d+ output for the table to which\\nthey are bound, and get rid of \\\\dd (and maybe \\\\dp and \\\\z too).\\n\\nNow that is not to say that what is being proposed here is completely\\nuseless. It clearly isn't. It's totally debatable whether we ought to\\nstart having commands like this, and maybe we should. It would make\\nfor more commands, and that's not entirely great because the command\\nnames are increasingly alphabet soup. Who can remember what \\\\drds or\\n\\\\dAc does? Only real power users. If we add more, it's going to get\\neven more difficult, but some people will use it and like it and those\\npeople will be happy. It's kind of hard to say whether we'd come out\\nahead or not. What I think is fairly certain is that it would\\nrepresent a reversal of our current policy.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 15:11:47 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 25, 2022 at 03:11:47PM -0400, Robert Haas wrote:\\n> On Fri, Mar 25, 2022 at 12:28 AM Greg Stark <stark@mit.edu> wrote:\\n> > Development of this seems to have stalled with the only review of this\\n> > patch expressing some skepticism about whether it's needed at all.\\n> \\n> Now, there is some precedent for the idea of providing a command that\\n> lists everything globally. Specifically, we have a \\\\dp command, also\\n> known as \\\\z, to list privileges across all objects in the database.\\n\\n> The other thing that we have that is somewhat similar to this is \\\\dd,\\n\\n\\\\dX is similar, and I remember wondering whether it was really useful/needed.\\nThe catalog tables are exposed and documented for a reason, and power-users\\nwill learn to use them.\\n\\n+0.5 to mark the patch as RWF or rejected.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 14:20:25 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 25, 2022 at 3:20 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> \\\\dX is similar, and I remember wondering whether it was really useful/needed.\\n> The catalog tables are exposed and documented for a reason, and power-users\\n> will learn to use them.\\n\\nI don't think that \\\\dX is comparable, because I don't think we should\\nregard extended statistics as a table object. Indeed, generalizing\\nextended statistics so that they can be generated on a join seems to\\nme to be one of the most important things we could do in that area.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 15:24:34 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"In the interests of trying to clean up the CF and keep things moving\\nI'll mark the patch rejected.\\n\\nThat doesn't mean the decision can't change or that nobody's allowed\\nto keep discussing it. It's just that that seems to be the decision\\nright now and there are too many patches queued up to keep things in a\\npending state waiting for a more decisive conclusion. We can always\\nchange it if the winds shift...\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 14:59:20 -0400\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"Hi All,\\n\\n> In the interests of trying to clean up the CF and keep things moving\\n> I'll mark the patch rejected.\\n\\nThank you for managing the commitfest and the comments from many of\\nhackers. I apologize for not being able to email you more often due to\\nmy busy day job.\\n\\nFirst of all, I understand to a certain extent your opinion that we\\ncan use \\\\d and look at the constraints on a table-by-table basis as a\\nway to check the constraints.\\nHowever, suppose We want to reverse lookup a table from a constraint.\\nIn that case, there are two ways: (1) use \\\"\\\\d\\\" to lookup all tables,\\n(2) execute a select statement against pg_constraint. I think the\\nproposed function is more valuable than these.\\n\\n From another perspective, when looking at the comprehensiveness of\\nmetacommands in psql, it seems that only functions that focus on\\nconstraints do not exist. Therefore, It would be better to add it in\\nterms of comprehensiveness.\\n\\nI think there is room for other discussions about this patch. Still,\\nat least there are people (including myself) who think it is useful.\\nI don't think there is anything wrong with this patch that would\\ncomplicate the code or cause performance degradation, so I would like to\\ncontinue developing it for those who want to use it.\\n\\nHowever, I understand that it will not be ready in time for PG15, so\\nI would like to move forward with PG16. Therefore, the status of the\\npatch would be better by changing \\\"Waiting for Author\\\" rather than\\n\\\"Rejected\\\".\\n\\nP.S.\\nI'll send a new patch addressing Dag's comments in the next email.\\n\\nThanks,\\nTatsuro Yamada\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 12:15:45 +0900\",\n \"msg_from\": \"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n },\n {\n \"msg_contents\": \"Hi Dag,\\n\\n> The patch adds the command \\\"\\\\dco\\\" to list constraints in psql. This\\n> seems useful to me.\\n\\nThank you!\\n\\n \\n> The patch applies cleanly to HEAD, although some hunks have rather large\\n> offsets.\\n>\\n> As far as I can tell, the \\\"\\\\dco\\\" command works as documented.\\n>\\n> I have however found the following issues with the patch:\\n>\\n> * A TAB character has been added to doc/src/sgml/ref/psql-ref.sgml -\\n> this should be replaced with spaces.\\n\\nFixed.\\n\\n\\n> * The call to listConstraints in line src/bin/psql/command.c 794 refers\\n> to &cmd[2], this should rather be &cmd[3].\\n>\\n> * The patch kills the \\\"\\\\dc\\\" command in src/bin/psql/command.c\\n> This can be fixed by adding the following at line 800:\\n> else\\n> success =\\n> listConversions(pattern, show_verbose, show_system);\\n\\n\\nOh, you are right! Fixed.\\n\\n \\n> Another comment is that the \\\"\\\\dco\\\" command outputs quite a lot of\\n> information, which only fits in a wide terminal window. Would it be an\\n> idea to only display the columns \\\"Schema\\\" and \\\"Name\\\" by default, and\\n> use \\\"+\\\" to specify inclusion of the columns \\\"Definition\\\" and \\\"Table\\\".\\n \\n\\nI fixed the output columns as you proposed.\\n\\nThe current status of this patch is:\\n\\n - Addressed Dag's comments\\n - Not implemented yet:\\n - Tab completion\\n - Regression test\\n - NOT NULL constraint, and so on (based on pg_attribute)\\n\\nPlease find attached new patch.\\n\\n\\nThanks,\\nTatsuro Yamada\",\n \"msg_date\": \"Tue, 29 Mar 2022 14:21:07 +0900\",\n \"msg_from\": \"Tatsuro Yamada <tatsuro.yamada.tf@nttcom.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add psql command to list constraints\"\n }\n]","isConversational":false}}},{"rowIdx":53090,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi hackers,\n\nIn another thread[1], we found the pg_get_publication_tables function will output\nduplicate partition relid when adding both child and parent table to the\npublication(pubviaroot = false).\n\nExample:\ncreate table tbl1 (a int) partition by range (a);\ncreate table tbl1_part1 partition of tbl1 for values from (1) to (10);\ncreate table tbl1_part2 partition of tbl1 for values from (10) to (20);\ncreate publication pub for table\ntbl1, tbl1_part1 with (publish_via_partition_root=false);\n\nselect * from pg_get_publication_tables('pub');\n relid\n-------\n 16387\n 16390\n 16387\n\nThe reason of the duplicate output is that:\n\npg_get_publication_tables invokes function GetPublicationRelations internally.\nIn GetPublicationRelations(), it will add the oid of partition 'tbl1_part1'\ntwice. First time from extracting partitions from the specified parent table\n'tbl1', second time from the explicitly specified partition 'tbl1_part1'.\n\n\nI am not sure is this behavior expected as it seems natural for\npg_get_publication_tables to return duplicate-free relid list. OTOH, there\nseems no harm for the current behavior(duplicate output), it doesn't affect the\ninitial sync and change replication when using logical replication.\n\nPersonally, I think it might be better to make the output of\npg_get_publication_tables duplicate-free, because the change happened on each\noutput relid will only be replicated once. So, it seems more consistent to\noutput each relid only once.\n\nThoughts ?\n\n(Attach a patch which make the output duplicate-free)\n\n[1] https://www.postgresql.org/message-id/CAJcOf-eURu03QNmD%3D37PtsxuNW4nBGN3G_FdRMBx_tpkeyzDUw%40mail.gmail.com\n\nBest regards,\nHou zj","msg_date":"Mon, 15 Nov 2021 08:17:49 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n>\n> Hi hackers,\n>\n> In another thread[1], we found the pg_get_publication_tables function will output\n> duplicate partition relid when adding both child and parent table to the\n> publication(pubviaroot = false).\n>\n> Example:\n> create table tbl1 (a int) partition by range (a);\n> create table tbl1_part1 partition of tbl1 for values from (1) to (10);\n> create table tbl1_part2 partition of tbl1 for values from (10) to (20);\n> create publication pub for table\n> tbl1, tbl1_part1 with (publish_via_partition_root=false);\n>\n> select * from pg_get_publication_tables('pub');\n> relid\n> -------\n> 16387\n> 16390\n> 16387\n>\n> The reason of the duplicate output is that:\n>\n> pg_get_publication_tables invokes function GetPublicationRelations internally.\n> In GetPublicationRelations(), it will add the oid of partition 'tbl1_part1'\n> twice. First time from extracting partitions from the specified parent table\n> 'tbl1', second time from the explicitly specified partition 'tbl1_part1'.\n>\n>\n> I am not sure is this behavior expected as it seems natural for\n> pg_get_publication_tables to return duplicate-free relid list. OTOH, there\n> seems no harm for the current behavior(duplicate output), it doesn't affect the\n> initial sync and change replication when using logical replication.\n>\n> Personally, I think it might be better to make the output of\n> pg_get_publication_tables duplicate-free, because the change happened on each\n> output relid will only be replicated once. So, it seems more consistent to\n> output each relid only once.\n>\n> Thoughts ?\n>\n> (Attach a patch which make the output duplicate-free)\n>\n> [1] https://www.postgresql.org/message-id/CAJcOf-eURu03QNmD%3D37PtsxuNW4nBGN3G_FdRMBx_tpkeyzDUw%40mail.gmail.com\n\nThe users can always specify the distinct clause, see [1]. I don't see\nany problem with the existing way. If required you can specify in the\nview pg_publication_tables documentation that \"This view returns\nmultiple rows for the same relation, if the relation is child to a\nparent partition table and if both the parent and child are specified\nin the publication\" or something similart.\n\nIf at all, the pg_get_publication_tables() is supposed to give unique\noutputs, let's fix it and it is a good idea to specify that the view\npg_publication_tables returns unique rows even though child and parent\npartition tables are specified in the publication.\n\nI have fdw comments on the patch:\n1) Why to do result = list_concat_unique_oid(result, relids); within\nthe while loop which can make the while loop O(n*n*n)\ncomplexity(list_concat_unique_oid is O(n*n)? Can't you use\nlist_append_unique, of course this can also be costlier? Let's avoid\nadding anything to the while loop and use\nlist_sort+list_deduplicate_oid (qsort(O(nlogn)+O(n))\n\n /* Now sort and de-duplicate the result list */\n list_sort(result, list_oid_cmp);\n list_deduplicate_oid(result);\n\n while (HeapTupleIsValid(tup = systable_getnext(scan)))\n {\n Form_pg_publication_rel pubrel;\n+ List *relids = NIL;\n\n pubrel = (Form_pg_publication_rel) GETSTRUCT(tup);\n- result = GetPubPartitionOptionRelations(result, pub_partopt,\n+ relids = GetPubPartitionOptionRelations(relids, pub_partopt,\n pubrel->prrelid);\n+\n+ result = list_concat_unique_oid(result, relids);\n }\n\n2) Are you sure you want GetPublicationRelations to be returning the\nunique relations? I'm just thinking why can't you just do\nlist_sort+list_deduplicate_oid in the caller? I think this makes the\nfunction more restrictive. If required, you can pass a boolean flag\n(bool give_unique and specify in the function comments and if passed\ntrue do list_sort+list_deduplicate_oid at the end of\nGetPublicationRelations ).\n\n[1]\npostgres=# select * from pg_get_publication_tables('pub');\n relid\n-------\n 16387\n 16390\n 16387\n(3 rows)\n\npostgres=# select distinct * from pg_get_publication_tables('pub');\n relid\n-------\n 16387\n 16390\n(2 rows)\n\npostgres=# select * from pg_publication_tables;\n pubname | schemaname | tablename\n---------+------------+------------\n pub | public | tbl1_part1\n pub | public | tbl1_part2\n pub | public | tbl1_part1\n(3 rows)\n\npostgres=# select distinct * from pg_publication_tables;\n pubname | schemaname | tablename\n---------+------------+------------\n pub | public | tbl1_part1\n pub | public | tbl1_part2\n(2 rows)\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Mon, 15 Nov 2021 15:46:48 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"> On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\n> <houzj.fnst@fujitsu.com> wrote:\n\n> > create table tbl1 (a int) partition by range (a);\n> > create table tbl1_part1 partition of tbl1 for values from (1) to (10);\n> > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\n> > create publication pub for table\n> > tbl1, tbl1_part1 with (publish_via_partition_root=false);\n\nIn the name of consistency, I think this situation should be an error --\nI mean, if we detect that the user is trying to add both the partitioned\ntable *and* its partition, then all manner of things are possibly going\nto go wrong in some way, so my inclination is to avoid it altogether.\n\nIs there any reason to allow that?\n\nIf we do that, then we have to be careful with later ALTER PUBLICATION\ntoo: adding a partition is not allowed if its partitioned table is\nthere, and adding a partitioned table should behave in some sensible way\nif one of its partitions is there (either removing the partition at the\nsame time, or outright rejecting the operation.)\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Mon, 15 Nov 2021 10:42:14 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Nov 15, 2021 6:17 PM Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote:\r\n> On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\r\n> <houzj.fnst@fujitsu.com> wrote:\r\n> >\r\n> > Hi hackers,\r\n> >\r\n> > In another thread[1], we found the pg_get_publication_tables function\r\n> > will output duplicate partition relid when adding both child and\r\n> > parent table to the publication(pubviaroot = false).\r\n> >\r\n> > Example:\r\n> > create table tbl1 (a int) partition by range (a); create table\r\n> > tbl1_part1 partition of tbl1 for values from (1) to (10); create table\r\n> > tbl1_part2 partition of tbl1 for values from (10) to (20); create\r\n> > publication pub for table tbl1, tbl1_part1 with\r\n> > (publish_via_partition_root=false);\r\n> >\r\n> > select * from pg_get_publication_tables('pub'); relid\r\n> > -------\r\n> > 16387\r\n> > 16390\r\n> > 16387\r\n> >\r\n> > The reason of the duplicate output is that:\r\n> >\r\n> > pg_get_publication_tables invokes function GetPublicationRelations\r\n> internally.\r\n> > In GetPublicationRelations(), it will add the oid of partition 'tbl1_part1'\r\n> > twice. First time from extracting partitions from the specified parent\r\n> > table 'tbl1', second time from the explicitly specified partition 'tbl1_part1'.\r\n> >\r\n> >\r\n> > I am not sure is this behavior expected as it seems natural for\r\n> > pg_get_publication_tables to return duplicate-free relid list. OTOH,\r\n> > there seems no harm for the current behavior(duplicate output), it\r\n> > doesn't affect the initial sync and change replication when using logical\r\n> replication.\r\n> >\r\n> > Personally, I think it might be better to make the output of\r\n> > pg_get_publication_tables duplicate-free, because the change happened\r\n> > on each output relid will only be replicated once. So, it seems more\r\n> > consistent to output each relid only once.\r\n> >\r\n> > Thoughts ?\r\n> >\r\n> > (Attach a patch which make the output duplicate-free)\r\n> >\r\n> > [1]\r\n> >\r\n> https://www.postgresql.org/message-id/CAJcOf-eURu03QNmD%3D37Ptsxu\r\n> NW4nB\r\n> > GN3G_FdRMBx_tpkeyzDUw%40mail.gmail.com\r\n> \r\n> The users can always specify the distinct clause, see [1]. I don't see any problem\r\n> with the existing way. If required you can specify in the view\r\n> pg_publication_tables documentation that \"This view returns multiple rows for\r\n> the same relation, if the relation is child to a parent partition table and if both\r\n> the parent and child are specified in the publication\" or something similart.\r\n\r\nThanks for the response.\r\nYes, I agreed, as I said there's no harm for the current behavior. If most people don't think\r\nIt's worth changing the behavior, I am fine with that.\r\n\r\n> If at all, the pg_get_publication_tables() is supposed to give unique outputs,\r\n> let's fix it and it is a good idea to specify that the view pg_publication_tables\r\n> returns unique rows even though child and parent partition tables are specified\r\n> in the publication.\r\n> \r\n> I have fdw comments on the patch:\r\n> Let's avoid adding anything to the while\r\n> loop and use list_sort+list_deduplicate_oid (qsort(O(nlogn)+O(n))\r\n> \r\n\r\nThanks for the comments, I will address this if we finally decide to fix it.\r\n\r\n> \r\n> 2) Are you sure you want GetPublicationRelations to be returning the unique\r\n> relations? I'm just thinking why can't you just do list_sort+list_deduplicate_oid\r\n> in the caller? I think this makes the function more restrictive. If required, you\r\n> can pass a boolean flag (bool give_unique and specify in the function\r\n> comments and if passed true do list_sort+list_deduplicate_oid at the end of\r\n> GetPublicationRelations ).\r\n\r\nI change the GetPublicationRelations because I found no caller need the\r\nduplicate oid, and de-duplicate oid in GetPublicationRelations could reduce\r\nsome code change. I think if we need the duplicate oid in the future, we can\r\nadd the flag as you suggested.\r\n\r\nBest regards,\r\nHou zj\r\n","msg_date":"Tue, 16 Nov 2021 01:50:45 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Nov 15, 2021 9:42 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\r\n> > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\r\n> > <houzj.fnst@fujitsu.com> wrote:\r\n> \r\n> > > create table tbl1 (a int) partition by range (a); create table\r\n> > > tbl1_part1 partition of tbl1 for values from (1) to (10); create\r\n> > > table tbl1_part2 partition of tbl1 for values from (10) to (20);\r\n> > > create publication pub for table tbl1, tbl1_part1 with\r\n> > > (publish_via_partition_root=false);\r\n> \r\n> In the name of consistency, I think this situation should be an error -- I mean, if\r\n> we detect that the user is trying to add both the partitioned table *and* its\r\n> partition, then all manner of things are possibly going to go wrong in some way,\r\n> so my inclination is to avoid it altogether.\r\n>\r\n> Is there any reason to allow that?\r\n>\r\n> If we do that, then we have to be careful with later ALTER PUBLICATION\r\n> too: adding a partition is not allowed if its partitioned table is there, and adding\r\n> a partitioned table should behave in some sensible way if one of its partitions is\r\n> there (either removing the partition at the same time, or outright rejecting the\r\n> operation.)\r\n\r\nThanks for the response.\r\n\r\nIf we decide to disallow this case, we seem need to handle some other cases as\r\nwell, for example: We might also need additional check when ATTACH a partition,\r\nbecause the partition's parent table could already be published in the same\r\npublication as the partition. During the check we might also need to lock all\r\nthe parent tables to handle concurrently change which looks complicated to me :(\r\nI am not sure is it worth adding these lock and check. Maybe we can leave the current\r\nbehavior as is ?\r\n\r\nBest regards,\r\nHou zj\r\n","msg_date":"Tue, 16 Nov 2021 01:51:06 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Nov 15, 2021 at 7:12 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>\n> > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\n> > <houzj.fnst@fujitsu.com> wrote:\n>\n> > > create table tbl1 (a int) partition by range (a);\n> > > create table tbl1_part1 partition of tbl1 for values from (1) to (10);\n> > > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\n> > > create publication pub for table\n> > > tbl1, tbl1_part1 with (publish_via_partition_root=false);\n>\n> In the name of consistency, I think this situation should be an error --\n> I mean, if we detect that the user is trying to add both the partitioned\n> table *and* its partition, then all manner of things are possibly going\n> to go wrong in some way, so my inclination is to avoid it altogether.\n>\n> Is there any reason to allow that?\n>\n\nI think it could provide flexibility to users to later change\n\"publish_via_partition_root\" option. Because when that option is\nfalse, we use individual partitions schema to send changes and when it\nis true, we use root table's schema to send changes. Added Amit L. to\nknow if he has any thoughts on this matter as he was the author of\nthis work?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 16 Nov 2021 18:57:05 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n>\n> On Mon, Nov 15, 2021 9:42 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> > > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\n> > > <houzj.fnst@fujitsu.com> wrote:\n> >\n> > > > create table tbl1 (a int) partition by range (a); create table\n> > > > tbl1_part1 partition of tbl1 for values from (1) to (10); create\n> > > > table tbl1_part2 partition of tbl1 for values from (10) to (20);\n> > > > create publication pub for table tbl1, tbl1_part1 with\n> > > > (publish_via_partition_root=false);\n> >\n> > In the name of consistency, I think this situation should be an error -- I mean, if\n> > we detect that the user is trying to add both the partitioned table *and* its\n> > partition, then all manner of things are possibly going to go wrong in some way,\n> > so my inclination is to avoid it altogether.\n> >\n> > Is there any reason to allow that?\n> >\n> > If we do that, then we have to be careful with later ALTER PUBLICATION\n> > too: adding a partition is not allowed if its partitioned table is there, and adding\n> > a partitioned table should behave in some sensible way if one of its partitions is\n> > there (either removing the partition at the same time, or outright rejecting the\n> > operation.)\n>\n> Thanks for the response.\n>\n> If we decide to disallow this case, we seem need to handle some other cases as\n> well, for example: We might also need additional check when ATTACH a partition,\n> because the partition's parent table could already be published in the same\n> publication as the partition.\n>\n\nWhat kind of additional checks you are envisioning and why?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 17 Nov 2021 08:17:06 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Wed, Nov 17, 2021 10:47 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com\r\n> <houzj.fnst@fujitsu.com> wrote:\r\n> >\r\n> > On Mon, Nov 15, 2021 9:42 PM Alvaro Herrera <alvherre@alvh.no-ip.org>\r\n> wrote:\r\n> > > > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\r\n> > > > <houzj.fnst@fujitsu.com> wrote:\r\n> > >\r\n> > > > > create table tbl1 (a int) partition by range (a); create table\r\n> > > > > tbl1_part1 partition of tbl1 for values from (1) to (10); create\r\n> > > > > table tbl1_part2 partition of tbl1 for values from (10) to (20);\r\n> > > > > create publication pub for table tbl1, tbl1_part1 with\r\n> > > > > (publish_via_partition_root=false);\r\n> > >\r\n> > > In the name of consistency, I think this situation should be an\r\n> > > error -- I mean, if we detect that the user is trying to add both\r\n> > > the partitioned table *and* its partition, then all manner of things\r\n> > > are possibly going to go wrong in some way, so my inclination is to avoid it\r\n> > > altogether.\r\n> > >\r\n> > > Is there any reason to allow that?\r\n> > >\r\n> > > If we do that, then we have to be careful with later ALTER\r\n> > > PUBLICATION\r\n> > > too: adding a partition is not allowed if its partitioned table is\r\n> > > there, and adding a partitioned table should behave in some sensible\r\n> > > way if one of its partitions is there (either removing the partition\r\n> > > at the same time, or outright rejecting the\r\n> > > operation.)\r\n> >\r\n> > Thanks for the response.\r\n> >\r\n> > If we decide to disallow this case, we seem need to handle some other\r\n> > cases as well, for example: We might also need additional check when\r\n> > ATTACH a partition, because the partition's parent table could already\r\n> > be published in the same publication as the partition.\r\n> >\r\n> \r\n> What kind of additional checks you are envisioning and why?\r\n\r\nFor example:\r\n\r\ncreate table tbl1 (a int) partition by range (a);\r\ncreate table tbl1_part1 (like tbl1);\r\ncreate table tbl1_part2 partition of tbl1 for values from (10) to (20);\r\ncreate publication pub for table\r\ntbl1, tbl1_part1 with (publish_via_partition_root=false);\r\n\r\n--- We might need addition check here\r\nAlter table tbl1 ATTACH partition tbl1_part1 for values from (1) to (10);\r\n\r\nIn the above cases, 'tbl1_part1' is not a partition of 'tb1' when creating a\r\npublication. After the ATTACH, 'tbl1_part1' would become a partition of 'tbl1'\r\nwhich seems the case we want to disallow(both parent and child table in\r\npublication). So, When ATTACH, I thought we might need to check all the parent\r\ntables to see if any parent table is in the same publication which the table to\r\nbe attached is also belongs to. Does it make sense ?\r\n\r\nBest regards,\r\nHou zj\r\n","msg_date":"Wed, 17 Nov 2021 03:15:37 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Tue, Nov 16, 2021 at 10:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Mon, Nov 15, 2021 at 7:12 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> > > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\n> > > <houzj.fnst@fujitsu.com> wrote:\n> >\n> > > > create table tbl1 (a int) partition by range (a);\n> > > > create table tbl1_part1 partition of tbl1 for values from (1) to (10);\n> > > > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\n> > > > create publication pub for table\n> > > > tbl1, tbl1_part1 with (publish_via_partition_root=false);\n> >\n> > In the name of consistency, I think this situation should be an error --\n> > I mean, if we detect that the user is trying to add both the partitioned\n> > table *and* its partition, then all manner of things are possibly going\n> > to go wrong in some way, so my inclination is to avoid it altogether.\n> >\n> > Is there any reason to allow that?\n>\n> I think it could provide flexibility to users to later change\n> \"publish_via_partition_root\" option. Because when that option is\n> false, we use individual partitions schema to send changes and when it\n> is true, we use root table's schema to send changes. Added Amit L. to\n> know if he has any thoughts on this matter as he was the author of\n> this work?\n\nFWIW, I'm not sure that adding an error in this path, that is, when a\nuser adds both a partitioned table and its partitions to a\npublication, helps much. As for the safety of allowing it, if you\nlook at get_rel_sync_entry(), which handles much of the complexity of\ndetermining whether to publish a relation's changes and the schema to\nuse when doing so, you may be able to see that a partition being added\nduplicatively is harmless, modulo any undiscovered bugs. At least as\nfar as the post-initial-sync replication functionality is concerned.\n\nWhat IS problematic is what a subscriber sees in the\npg_publication_tables view and the problem occurs only in the initial\nsync phase, where the partition is synced duplicatively because of\nbeing found in the view along with the parent in this case, that is,\nwhen publish_via_partiiton_root is true. I was saying on the other\nthread [1] that we should leave it up to the subscriber to decide what\nto do when the view (duplicatively) returns both the parent and the\npartition, citing the use case that a subscriber may want to replicate\nthe parent and the partition as independent tables. Though I now tend\nto agree with Amit K that that may be such a meaningful and all that\ncommon use case, and the implementation on the subscriber side would\nhave to be unnecessarily complex.\n\nSo maybe it makes sense to just do what has been proposed --\nde-duplicate partitions out of the pg_publication_tables view, unless\nwe know of a bigger problem that requires us to hack the subscriber\nside of things too. Actually, I came to know of one such problem\nwhile thinking about this today: when you ATTACH a partition to a\ntable that is present in a publish_via_partition_root=true\npublication, it doesn't get copied via the initial sync even though\nsubsequent replication works just fine. The reason for that is that\nthe subscriber only syncs the partitions that are known at the time\nwhen the parent is first synced, that too via the parent (as SELECT\n<columns..> FROM parent), and then marks the parent as sync-done.\nRefreshing the subscription after ATTACHing doesn't help, because the\nsubscriber can't see any partitions to begin with. Maybe a more\nelaborate solution is needed for this one, though I haven't figured\nwhat it is going to look like yet.\n\nThanks.\n\n--\nAmit Langote\nEDB: http://www.enterprisedb.com\n\n[1] https://www.postgresql.org/message-id/CA%2BHiwqHnDHcT4OOcga9rDFyc7TvDrpN5xFH9J2pyHQo9ptvjmQ%40mail.gmail.com\n\n\n","msg_date":"Wed, 17 Nov 2021 15:09:14 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Wed, Nov 17, 2021 at 12:15 PM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n> On Wed, Nov 17, 2021 10:47 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com wrote:\n> > > If we decide to disallow this case, we seem need to handle some other\n> > > cases as well, for example: We might also need additional check when\n> > > ATTACH a partition, because the partition's parent table could already\n> > > be published in the same publication as the partition.\n> > >\n> >\n> > What kind of additional checks you are envisioning and why?\n>\n> For example:\n>\n> create table tbl1 (a int) partition by range (a);\n> create table tbl1_part1 (like tbl1);\n> create table tbl1_part2 partition of tbl1 for values from (10) to (20);\n> create publication pub for table\n> tbl1, tbl1_part1 with (publish_via_partition_root=false);\n>\n> --- We might need addition check here\n> Alter table tbl1 ATTACH partition tbl1_part1 for values from (1) to (10);\n>\n> In the above cases, 'tbl1_part1' is not a partition of 'tb1' when creating a\n> publication. After the ATTACH, 'tbl1_part1' would become a partition of 'tbl1'\n> which seems the case we want to disallow(both parent and child table in\n> publication). So, When ATTACH, I thought we might need to check all the parent\n> tables to see if any parent table is in the same publication which the table to\n> be attached is also belongs to. Does it make sense ?\n\nI don't think creating or attaching a partition of a table that is\npresent in a publish_via_partition_root=false actually adds the\npartition to pg_publication_rel, the base catalog. A partition's\nmembership in the publication is implicit, unless of course you add it\nto the publication explicitly, like all the examples we have been\ndiscussing. I guess we're only arguing about the problems with the\npg_publication_tables view, which does expand the partitioned table to\nshow the partitions that are not otherwise not present in the base\ncatalog.\n\n-- \nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 17 Nov 2021 15:18:13 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Wed, Nov 17, 2021 at 3:09 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> Though I now tend\n> to agree with Amit K that that may be such a meaningful and all that\n> common use case,\n\nOops I meant: that may NOT be such...\n\n-- \nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 17 Nov 2021 17:49:43 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Wed, Nov 17, 2021 2:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\r\n> On Wed, Nov 17, 2021 at 12:15 PM houzj.fnst@fujitsu.com\r\n> <houzj.fnst@fujitsu.com> wrote:\r\n> > On Wed, Nov 17, 2021 10:47 AM Amit Kapila <amit.kapila16@gmail.com>\r\n> wrote:\r\n> > > On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com wrote:\r\n> > > > If we decide to disallow this case, we seem need to handle some other\r\n> > > > cases as well, for example: We might also need additional check when\r\n> > > > ATTACH a partition, because the partition's parent table could already\r\n> > > > be published in the same publication as the partition.\r\n> > > >\r\n> > >\r\n> > > What kind of additional checks you are envisioning and why?\r\n> >\r\n> > For example:\r\n> >\r\n> > create table tbl1 (a int) partition by range (a);\r\n> > create table tbl1_part1 (like tbl1);\r\n> > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\r\n> > create publication pub for table\r\n> > tbl1, tbl1_part1 with (publish_via_partition_root=false);\r\n> >\r\n> > --- We might need addition check here\r\n> > Alter table tbl1 ATTACH partition tbl1_part1 for values from (1) to (10);\r\n> >\r\n> > In the above cases, 'tbl1_part1' is not a partition of 'tb1' when creating a\r\n> > publication. After the ATTACH, 'tbl1_part1' would become a partition of 'tbl1'\r\n> > which seems the case we want to disallow(both parent and child table in\r\n> > publication). So, When ATTACH, I thought we might need to check all the parent\r\n> > tables to see if any parent table is in the same publication which the table to\r\n> > be attached is also belongs to. Does it make sense ?\r\n> \r\n> I don't think creating or attaching a partition of a table that is\r\n> present in a publish_via_partition_root=false actually adds the\r\n> partition to pg_publication_rel, the base catalog. A partition's\r\n> membership in the publication is implicit, unless of course you add it\r\n> to the publication explicitly, like all the examples we have been\r\n> discussing. I guess we're only arguing about the problems with the\r\n> pg_publication_tables view, which does expand the partitioned table to\r\n> show the partitions that are not otherwise not present in the base\r\n> catalog.\r\n\r\nMaybe I didn't make it clear, I was trying to explain that it would be\r\ncomplicated if we want to completely disallow specifying both child and parent\r\ntable in the publication because of the ATTACH case I gave. In other words, I\r\nthink it's fine to specify both child and parent table in the publication.\r\n\r\nBest regards,\r\nHou zj\r\n","msg_date":"Thu, 18 Nov 2021 00:33:09 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thu, Nov 18, 2021 at 9:33 houzj.fnst@fujitsu.com <houzj.fnst@fujitsu.com>\nwrote:\n\n> On Wed, Nov 17, 2021 2:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > On Wed, Nov 17, 2021 at 12:15 PM houzj.fnst@fujitsu.com\n> > <houzj.fnst@fujitsu.com> wrote:\n> > > On Wed, Nov 17, 2021 10:47 AM Amit Kapila <amit.kapila16@gmail.com>\n> > wrote:\n> > > > On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com wrote:\n> > > > > If we decide to disallow this case, we seem need to handle some\n> other\n> > > > > cases as well, for example: We might also need additional check\n> when\n> > > > > ATTACH a partition, because the partition's parent table could\n> already\n> > > > > be published in the same publication as the partition.\n> > > > >\n> > > >\n> > > > What kind of additional checks you are envisioning and why?\n> > >\n> > > For example:\n> > >\n> > > create table tbl1 (a int) partition by range (a);\n> > > create table tbl1_part1 (like tbl1);\n> > > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\n> > > create publication pub for table\n> > > tbl1, tbl1_part1 with (publish_via_partition_root=false);\n> > >\n> > > --- We might need addition check here\n> > > Alter table tbl1 ATTACH partition tbl1_part1 for values from (1) to\n> (10);\n> > >\n> > > In the above cases, 'tbl1_part1' is not a partition of 'tb1' when\n> creating a\n> > > publication. After the ATTACH, 'tbl1_part1' would become a partition\n> of 'tbl1'\n> > > which seems the case we want to disallow(both parent and child table in\n> > > publication). So, When ATTACH, I thought we might need to check all\n> the parent\n> > > tables to see if any parent table is in the same publication which the\n> table to\n> > > be attached is also belongs to. Does it make sense ?\n> >\n> > I don't think creating or attaching a partition of a table that is\n> > present in a publish_via_partition_root=false actually adds the\n> > partition to pg_publication_rel, the base catalog. A partition's\n> > membership in the publication is implicit, unless of course you add it\n> > to the publication explicitly, like all the examples we have been\n> > discussing. I guess we're only arguing about the problems with the\n> > pg_publication_tables view, which does expand the partitioned table to\n> > show the partitions that are not otherwise not present in the base\n> > catalog.\n>\n> Maybe I didn't make it clear, I was trying to explain that it would be\n> complicated if we want to completely disallow specifying both child and\n> parent\n> table in the publication because of the ATTACH case I gave. In other\n> words, I\n> think it's fine to specify both child and parent table in the publication.\n\n\nAh okay. I see your point. A table that was not once a partition would be\nin the catalog along with the unrelated parent table and if we were to\nenforce the rule that a parent and a partition cannot be in the catalog at\nthe same time, we’d need to delete that table’s catalog record on ATTACHing\nit as a partition.\n\nThanks.\n-- \nAmit Langote\nEDB: http://www.enterprisedb.com\n\nOn Thu, Nov 18, 2021 at 9:33 houzj.fnst@fujitsu.com <houzj.fnst@fujitsu.com> wrote:On Wed, Nov 17, 2021 2:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> On Wed, Nov 17, 2021 at 12:15 PM houzj.fnst@fujitsu.com\n> <houzj.fnst@fujitsu.com> wrote:\n> > On Wed, Nov 17, 2021 10:47 AM Amit Kapila <amit.kapila16@gmail.com>\n> wrote:\n> > > On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com wrote:\n> > > > If we decide to disallow this case, we seem need to handle some other\n> > > > cases as well, for example: We might also need additional check when\n> > > > ATTACH a partition, because the partition's parent table could already\n> > > > be published in the same publication as the partition.\n> > > >\n> > >\n> > > What kind of additional checks you are envisioning and why?\n> >\n> > For example:\n> >\n> > create table tbl1 (a int) partition by range (a);\n> > create table tbl1_part1 (like tbl1);\n> > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\n> > create publication pub for table\n> > tbl1, tbl1_part1 with (publish_via_partition_root=false);\n> >\n> > --- We might need addition check here\n> > Alter table tbl1 ATTACH partition tbl1_part1 for values from (1) to (10);\n> >\n> > In the above cases, 'tbl1_part1' is not a partition of 'tb1' when creating a\n> > publication. After the ATTACH, 'tbl1_part1' would become a partition of 'tbl1'\n> > which seems the case we want to disallow(both parent and child table in\n> > publication). So, When ATTACH, I thought we might need to check all the parent\n> > tables to see if any parent table is in the same publication which the table to\n> > be attached is also belongs to. Does it make sense ?\n> \n> I don't think creating or attaching a partition of a table that is\n> present in a publish_via_partition_root=false actually adds the\n> partition to pg_publication_rel, the base catalog. A partition's\n> membership in the publication is implicit, unless of course you add it\n> to the publication explicitly, like all the examples we have been\n> discussing. I guess we're only arguing about the problems with the\n> pg_publication_tables view, which does expand the partitioned table to\n> show the partitions that are not otherwise not present in the base\n> catalog.\n\nMaybe I didn't make it clear, I was trying to explain that it would be\ncomplicated if we want to completely disallow specifying both child and parent\ntable in the publication because of the ATTACH case I gave. In other words, I\nthink it's fine to specify both child and parent table in the publication.Ah okay. I see your point. A table that was not once a partition would be in the catalog along with the unrelated parent table and if we were to enforce the rule that a parent and a partition cannot be in the catalog at the same time, we’d need to delete that table’s catalog record on ATTACHing it as a partition.Thanks.-- Amit LangoteEDB: http://www.enterprisedb.com","msg_date":"Thu, 18 Nov 2021 09:57:30 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Wed, Nov 17, 2021 at 11:39 AM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Tue, Nov 16, 2021 at 10:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Mon, Nov 15, 2021 at 7:12 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> > > > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\n> > > > <houzj.fnst@fujitsu.com> wrote:\n> > >\n> > > > > create table tbl1 (a int) partition by range (a);\n> > > > > create table tbl1_part1 partition of tbl1 for values from (1) to (10);\n> > > > > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\n> > > > > create publication pub for table\n> > > > > tbl1, tbl1_part1 with (publish_via_partition_root=false);\n> > >\n> > > In the name of consistency, I think this situation should be an error --\n> > > I mean, if we detect that the user is trying to add both the partitioned\n> > > table *and* its partition, then all manner of things are possibly going\n> > > to go wrong in some way, so my inclination is to avoid it altogether.\n> > >\n> > > Is there any reason to allow that?\n> >\n> > I think it could provide flexibility to users to later change\n> > \"publish_via_partition_root\" option. Because when that option is\n> > false, we use individual partitions schema to send changes and when it\n> > is true, we use root table's schema to send changes. Added Amit L. to\n> > know if he has any thoughts on this matter as he was the author of\n> > this work?\n>\n> FWIW, I'm not sure that adding an error in this path, that is, when a\n> user adds both a partitioned table and its partitions to a\n> publication, helps much. As for the safety of allowing it, if you\n> look at get_rel_sync_entry(), which handles much of the complexity of\n> determining whether to publish a relation's changes and the schema to\n> use when doing so, you may be able to see that a partition being added\n> duplicatively is harmless, modulo any undiscovered bugs. At least as\n> far as the post-initial-sync replication functionality is concerned.\n>\n> What IS problematic is what a subscriber sees in the\n> pg_publication_tables view and the problem occurs only in the initial\n> sync phase, where the partition is synced duplicatively because of\n> being found in the view along with the parent in this case, that is,\n> when publish_via_partiiton_root is true. I was saying on the other\n> thread [1] that we should leave it up to the subscriber to decide what\n> to do when the view (duplicatively) returns both the parent and the\n> partition, citing the use case that a subscriber may want to replicate\n> the parent and the partition as independent tables. Though I now tend\n> to agree with Amit K that that may be such a meaningful and all that\n> common use case, and the implementation on the subscriber side would\n> have to be unnecessarily complex.\n>\n> So maybe it makes sense to just do what has been proposed --\n> de-duplicate partitions out of the pg_publication_tables view,\n>\n\nAFAICU, there are actually two problems related to\npg_publication_tables view that are being discussed: (a) when\n'publish_via_partition_root' is true then it returns both parent and\nchild tables (provided both are part of publication), this can lead to\nduplicate data on the subscriber; the fix for this is being discussed\nin thread [1]. (b) when 'publish_via_partition_root' is false, then it\nreturns duplicate entries for child tables (provided both are part of\npublication), this problem is being discussed in this thread.\n\nAs per the proposed patches, it seems both need a separate fix, we can\nfix them together if we want but I am not sure. Is your understanding\nthe same?\n\n> unless\n> we know of a bigger problem that requires us to hack the subscriber\n> side of things too.\n>\n\nThere is yet another issue that might need subscriber side change. See\nthe second issue summarized by Hou-San in the email[2]. I feel it is\nbetter to tackle that separately.\n\n> Actually, I came to know of one such problem\n> while thinking about this today: when you ATTACH a partition to a\n> table that is present in a publish_via_partition_root=true\n> publication, it doesn't get copied via the initial sync even though\n> subsequent replication works just fine. The reason for that is that\n> the subscriber only syncs the partitions that are known at the time\n> when the parent is first synced, that too via the parent (as SELECT\n> <columns..> FROM parent), and then marks the parent as sync-done.\n> Refreshing the subscription after ATTACHing doesn't help, because the\n> subscriber can't see any partitions to begin with.\n>\n\nSorry, it is not clear to me how this can lead to a problem. Even if\nwe just have the root table in the subscriber at the time of sync\nlater shouldn't copy get the newly attached partition's data and\nreplicate it to the required partition for\n\"publish_via_partition_root=true\" case? Anyway, if this is a problem\nwe need to figure the solution for this separately.\n\n[1]: https://www.postgresql.org/message-id/CA%2BHiwqHnDHcT4OOcga9rDFyc7TvDrpN5xFH9J2pyHQo9ptvjmQ%40mail.gmail.com\n[2]: https://www.postgresql.org/message-id/OS0PR01MB5716C756312959F293A822C794869%40OS0PR01MB5716.jpnprd01.prod.outlook.com\n\n\n\n--\nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 18 Nov 2021 10:23:40 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thu, Nov 18, 2021 at 10:23 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> (b) when 'publish_via_partition_root' is false, then it\n> returns duplicate entries for child tables (provided both are part of\n> publication), this problem is being discussed in this thread.\n\n+1 to just focus on fixing the initial problem proposed i.e.\npg_get_publication_tables() outputting duplicate relid. IMO, all other\nissues (if unrelated to this problem and the patch can go to another\nnew thread, if necessary).\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Thu, 18 Nov 2021 10:48:12 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thu, Nov 18, 2021 at 1:53 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Wed, Nov 17, 2021 at 11:39 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> > What IS problematic is what a subscriber sees in the\n> > pg_publication_tables view and the problem occurs only in the initial\n> > sync phase, where the partition is synced duplicatively because of\n> > being found in the view along with the parent in this case, that is,\n> > when publish_via_partiiton_root is true. I was saying on the other\n> > thread [1] that we should leave it up to the subscriber to decide what\n> > to do when the view (duplicatively) returns both the parent and the\n> > partition, citing the use case that a subscriber may want to replicate\n> > the parent and the partition as independent tables. Though I now tend\n> > to agree with Amit K that that may be such a meaningful and all that\n> > common use case, and the implementation on the subscriber side would\n> > have to be unnecessarily complex.\n> >\n> > So maybe it makes sense to just do what has been proposed --\n> > de-duplicate partitions out of the pg_publication_tables view,\n>\n> AFAICU, there are actually two problems related to\n> pg_publication_tables view that are being discussed: (a) when\n> 'publish_via_partition_root' is true then it returns both parent and\n> child tables (provided both are part of publication), this can lead to\n> duplicate data on the subscriber; the fix for this is being discussed\n> in thread [1]. (b) when 'publish_via_partition_root' is false, then it\n> returns duplicate entries for child tables (provided both are part of\n> publication), this problem is being discussed in this thread.\n>\n> As per the proposed patches, it seems both need a separate fix, we can\n> fix them together if we want but I am not sure. Is your understanding\n> the same?\n\nActually, I'm seeing both (a) and (b) as more or less the same thing.\nIf publish_via_partition_root=true, we'd want to remove a partition\nfrom the list if the root parent is also present. If false, we'd want\nto remove a partition from the list because it'd also be added by way\nof expanding the root.\n\nI know that (a) causes the partition-double-sync problem that could be\nfixed by partition OID de-duplication as proposed in the other thread.\nAs a comment on that, it'd be nice to see a test case added to\nsrc/test/subscription suite in that patch, because\npartition-double-sync is the main problem statement I'd think.\n\nLike Bharath said, I don't see (b) as much of a problem, because the\nsubscriber applies DISTINCT when querying pg_publication_tables()\nanyway. Other users, if any out there, may be doing the same by\nfollowing core subscription code's example. That said, I am not\nopposed to applying the patch being proposed here, though I guess we\ndon't have any src/test/subscription test case for this one? As in,\ndo we know of any replication (initial/streaming) misbehavior caused\nby the duplicate partition OIDs in this case or is the only problem\nthat pg_publication_tables output looks odd?\n\n> > unless\n> > we know of a bigger problem that requires us to hack the subscriber\n> > side of things too.\n> >\n>\n> There is yet another issue that might need subscriber side change. See\n> the second issue summarized by Hou-San in the email[2]. I feel it is\n> better to tackle that separately.\n>\n> > Actually, I came to know of one such problem\n> > while thinking about this today: when you ATTACH a partition to a\n> > table that is present in a publish_via_partition_root=true\n> > publication, it doesn't get copied via the initial sync even though\n> > subsequent replication works just fine. The reason for that is that\n> > the subscriber only syncs the partitions that are known at the time\n> > when the parent is first synced, that too via the parent (as SELECT\n> > <columns..> FROM parent), and then marks the parent as sync-done.\n> > Refreshing the subscription after ATTACHing doesn't help, because the\n> > subscriber can't see any partitions to begin with.\n> >\n>\n> Sorry, it is not clear to me how this can lead to a problem. Even if\n> we just have the root table in the subscriber at the time of sync\n> later shouldn't copy get the newly attached partition's data and\n> replicate it to the required partition for\n> \"publish_via_partition_root=true\" case?\n\nNot sure if you thought otherwise but I am talking about attaching a\nnew partition into the partition tree on the *publication side*.\n\nThe problematic case is attaching the partition *after* the subscriber\nhas already marked the root parent as synced and/or ready for\nreplication. Refreshing the subscription doesn't help it discover the\nnewly attached partition, because a publish_via_partition_root only\never tells about the root parent, which would be already synced, so\nthe subscriber would think there's nothing to copy.\n\n> Anyway, if this is a problem\n> we need to figure the solution for this separately.\n\nSure, we might need to do that after all. Though it might be a good\nidea to be sure that we won't need to reconsider the fix we push for\nthe issue(s) being discussed here and elsewhere, because I suspect\nthat the solution to the problem I mentioned is likely to involve\ntweaking pg_publication_tables view output.\n\n--\nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 19 Nov 2021 10:49:32 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Thu, Nov 18, 2021 at 1:53 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > AFAICU, there are actually two problems related to\n> > pg_publication_tables view that are being discussed: (a) when\n> > 'publish_via_partition_root' is true then it returns both parent and\n> > child tables (provided both are part of publication), this can lead to\n> > duplicate data on the subscriber; the fix for this is being discussed\n> > in thread [1]. (b) when 'publish_via_partition_root' is false, then it\n> > returns duplicate entries for child tables (provided both are part of\n> > publication), this problem is being discussed in this thread.\n> >\n> > As per the proposed patches, it seems both need a separate fix, we can\n> > fix them together if we want but I am not sure. Is your understanding\n> > the same?\n>\n> Actually, I'm seeing both (a) and (b) as more or less the same thing.\n> If publish_via_partition_root=true, we'd want to remove a partition\n> from the list if the root parent is also present. If false, we'd want\n> to remove a partition from the list because it'd also be added by way\n> of expanding the root.\n>\n> I know that (a) causes the partition-double-sync problem that could be\n> fixed by partition OID de-duplication as proposed in the other thread.\n> As a comment on that, it'd be nice to see a test case added to\n> src/test/subscription suite in that patch, because\n> partition-double-sync is the main problem statement I'd think.\n>\n> Like Bharath said, I don't see (b) as much of a problem, because the\n> subscriber applies DISTINCT when querying pg_publication_tables()\n> anyway. Other users, if any out there, may be doing the same by\n> following core subscription code's example.\n>\n\nHmm, I don't think we can assume that current or future users of\npg_publication_tables() will use it with DISTINCT unless we specify in\nthe specs/docs. However, I guess we might want to do it only for HEAD\nas there is no direct field report for this and anyway we have some\nworkaround for this.\n\n> That said, I am not\n> opposed to applying the patch being proposed here, though I guess we\n> don't have any src/test/subscription test case for this one?\n>\n\nYeah, we can add tests for this and the other case.\n\n> As in,\n> do we know of any replication (initial/streaming) misbehavior caused\n> by the duplicate partition OIDs in this case or is the only problem\n> that pg_publication_tables output looks odd?\n>\n\nThe latter one but I think either we should document this or change it\nas we can't assume users will follow what subscriber-side code does.\n\n> > > unless\n> > > we know of a bigger problem that requires us to hack the subscriber\n> > > side of things too.\n> > >\n> >\n> > There is yet another issue that might need subscriber side change. See\n> > the second issue summarized by Hou-San in the email[2]. I feel it is\n> > better to tackle that separately.\n> >\n>\n> The problematic case is attaching the partition *after* the subscriber\n> has already marked the root parent as synced and/or ready for\n> replication. Refreshing the subscription doesn't help it discover the\n> newly attached partition, because a publish_via_partition_root only\n> ever tells about the root parent, which would be already synced, so\n> the subscriber would think there's nothing to copy.\n>\n\nOkay, I see this could be a problem but I haven't tried to reproduce it.\n\n> > Anyway, if this is a problem\n> > we need to figure the solution for this separately.\n>\n> Sure, we might need to do that after all. Though it might be a good\n> idea to be sure that we won't need to reconsider the fix we push for\n> the issue(s) being discussed here and elsewhere, because I suspect\n> that the solution to the problem I mentioned is likely to involve\n> tweaking pg_publication_tables view output.\n>\n\nOkay, so we have four known problems in the same area. The first has\nbeen reported at the start of this thread, then the two as summarized\nby Hou-San in his email [1] and the fourth one is the attach of\npartitions after initial sync as you mentioned. It makes sense to have\nsome high-level idea on how to solve each of the problems before\npushing a fix for any one particular problem as one fix could have an\nimpact on other fixes. I'll think over it but please do let me know if\nyou have any ideas.\n\n\n[1] - https://www.postgresql.org/message-id/OS0PR01MB5716C756312959F293A822C794869%40OS0PR01MB5716.jpnprd01.prod.outlook.com\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 19 Nov 2021 10:58:32 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> >\n> > The problematic case is attaching the partition *after* the subscriber\n> > has already marked the root parent as synced and/or ready for\n> > replication. Refreshing the subscription doesn't help it discover the\n> > newly attached partition, because a publish_via_partition_root only\n> > ever tells about the root parent, which would be already synced, so\n> > the subscriber would think there's nothing to copy.\n> >\n>\n> Okay, I see this could be a problem but I haven't tried to reproduce it.\n>\n> > > Anyway, if this is a problem\n> > > we need to figure the solution for this separately.\n> >\n> > Sure, we might need to do that after all. Though it might be a good\n> > idea to be sure that we won't need to reconsider the fix we push for\n> > the issue(s) being discussed here and elsewhere, because I suspect\n> > that the solution to the problem I mentioned is likely to involve\n> > tweaking pg_publication_tables view output.\n> >\n\nI have thought about this problem and I see two possibilities for a\nsolution (a) We could provide a new option say 'truncate' (something\non lines proposed here [1]) which would truncate the table(s) and\nchange its status to 'i' in the pg_subscription_rel, this would allow\nthe newly added partition to be synced after refresh. This could lead\nto a large copy in such a case. (b) We could somehow get and store all\nthe partition info from the publisher-side on the subscriber-side\nwhile initial sync (say in new system table\npg_subscription_rel_members). Now, after the refresh, if this list\nchanges, we can allow to just get the data of that particular\npartition but I guess it would mean that we need to store oids of the\npublisher which might or might not be safe considering oids can\nwraparound before the refresh.\n\nDo you have any other ideas?\n\nOne more thing you mentioned is that the initial sync won't work after\nrefresh but later changes will be replicated but I noticed that later\nchanges also don't get streamed till we restart the subscriber server.\nI am not sure but we might not be invalidating apply workers cache due\nto which it didn't notice the same.\n\n[1] - https://www.postgresql.org/message-id/CF3B6672-2A43-4204-A60A-68F359218A9B%40endpoint.com\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Sat, 20 Nov 2021 17:01:00 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> > As in,\n> > do we know of any replication (initial/streaming) misbehavior caused\n> > by the duplicate partition OIDs in this case or is the only problem\n> > that pg_publication_tables output looks odd?\n>\n> The latter one but I think either we should document this or change it\n> as we can't assume users will follow what subscriber-side code does.\n\nOn second thought, I agree that de-duplicating partitions from this\nview is an improvement.\n\n-- \nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 22 Nov 2021 16:24:45 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Sat, Nov 20, 2021 at 8:31 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > The problematic case is attaching the partition *after* the subscriber\n> > > has already marked the root parent as synced and/or ready for\n> > > replication. Refreshing the subscription doesn't help it discover the\n> > > newly attached partition, because a publish_via_partition_root only\n> > > ever tells about the root parent, which would be already synced, so\n> > > the subscriber would think there's nothing to copy.\n> >\n> > Okay, I see this could be a problem but I haven't tried to reproduce it.\n> >\n> > > > Anyway, if this is a problem\n> > > > we need to figure the solution for this separately.\n> > >\n> > > Sure, we might need to do that after all. Though it might be a good\n> > > idea to be sure that we won't need to reconsider the fix we push for\n> > > the issue(s) being discussed here and elsewhere, because I suspect\n> > > that the solution to the problem I mentioned is likely to involve\n> > > tweaking pg_publication_tables view output.\n>\n> I have thought about this problem and I see two possibilities for a\n> solution (a) We could provide a new option say 'truncate' (something\n> on lines proposed here [1]) which would truncate the table(s) and\n> change its status to 'i' in the pg_subscription_rel, this would allow\n> the newly added partition to be synced after refresh. This could lead\n> to a large copy in such a case.\n\nMaybe I am missing something about the proposal, though I'd think a\nmore automatic solution would be better, something that doesn't need\nto rely on an unrelated feature.\n\n> (b) We could somehow get and store all\n> the partition info from the publisher-side on the subscriber-side\n> while initial sync (say in new system table\n> pg_subscription_rel_members). Now, after the refresh, if this list\n> changes, we can allow to just get the data of that particular\n> partition but I guess it would mean that we need to store oids of the\n> publisher which might or might not be safe considering oids can\n> wraparound before the refresh.\n>\n> Do you have any other ideas?\n\nI thought that the idea I had earlier mentioned at [1] may be useful,\nwhich I can see is similar to your idea (b). I also suspect that it\ncan be implemented without needing a separate catalog and storing\npublication-side relation OIDs in the subscription-side catalog,\nthough maybe I haven't thought hard enough.\n\n> One more thing you mentioned is that the initial sync won't work after\n> refresh but later changes will be replicated but I noticed that later\n> changes also don't get streamed till we restart the subscriber server.\n> I am not sure but we might not be invalidating apply workers cache due\n> to which it didn't notice the same.\n\nOh, that sounds odd and, as you appear to say, a separate problem. I'll check.\n\n-- \nAmit Langote\nEDB: http://www.enterprisedb.com\n\n[1] https://www.postgresql.org/message-id/CA%2BHiwqHnDHcT4OOcga9rDFyc7TvDrpN5xFH9J2pyHQo9ptvjmQ%40mail.gmail.com\n\n\n","msg_date":"Mon, 22 Nov 2021 17:15:37 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Nov 22, 2021 at 1:45 PM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Sat, Nov 20, 2021 at 8:31 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > > The problematic case is attaching the partition *after* the subscriber\n> > > > has already marked the root parent as synced and/or ready for\n> > > > replication. Refreshing the subscription doesn't help it discover the\n> > > > newly attached partition, because a publish_via_partition_root only\n> > > > ever tells about the root parent, which would be already synced, so\n> > > > the subscriber would think there's nothing to copy.\n> > >\n> > > Okay, I see this could be a problem but I haven't tried to reproduce it.\n> > >\n> > > > > Anyway, if this is a problem\n> > > > > we need to figure the solution for this separately.\n> > > >\n> > > > Sure, we might need to do that after all. Though it might be a good\n> > > > idea to be sure that we won't need to reconsider the fix we push for\n> > > > the issue(s) being discussed here and elsewhere, because I suspect\n> > > > that the solution to the problem I mentioned is likely to involve\n> > > > tweaking pg_publication_tables view output.\n> >\n> > I have thought about this problem and I see two possibilities for a\n> > solution (a) We could provide a new option say 'truncate' (something\n> > on lines proposed here [1]) which would truncate the table(s) and\n> > change its status to 'i' in the pg_subscription_rel, this would allow\n> > the newly added partition to be synced after refresh. This could lead\n> > to a large copy in such a case.\n>\n> Maybe I am missing something about the proposal, though I'd think a\n> more automatic solution would be better, something that doesn't need\n> to rely on an unrelated feature.\n>\n\nAgreed, this was more of a workaround for users if we didn't get any\nautomatic solution.\n\n> > (b) We could somehow get and store all\n> > the partition info from the publisher-side on the subscriber-side\n> > while initial sync (say in new system table\n> > pg_subscription_rel_members). Now, after the refresh, if this list\n> > changes, we can allow to just get the data of that particular\n> > partition but I guess it would mean that we need to store oids of the\n> > publisher which might or might not be safe considering oids can\n> > wraparound before the refresh.\n> >\n> > Do you have any other ideas?\n>\n> I thought that the idea I had earlier mentioned at [1] may be useful,\n> which I can see is similar to your idea (b). I also suspect that it\n> can be implemented without needing a separate catalog and storing\n> publication-side relation OIDs in the subscription-side catalog,\n> though maybe I haven't thought hard enough.\n>\n\nThe problem with storing this info in pg_subscription_rel as you were\ndescribing in your proposal is that currently, we ensure that the same\ntable exists in subscriber and then store the subscriber side table id\nin that catalog. I am not sure if we can store publisher-side oids in\nthat catalog and if we store then it would be confusing as now it will\nhave info of both publisher-side oids and subscriber-side oids. Isn't\nthat a problem with this approach?\n\nOne more problem with this overall approach is the risk of OID\nwraparound. Say between Create Subscription and Alter Subscription ..\nRefresh, we detach one partition, the oid wraparounds, and we\ncreate/reattach another partition which gets the same oid as for the\npartition which we detached earlier then we won't be able to identify\nthe new partition even after refresh.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 22 Nov 2021 16:17:31 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Nov 22, 2021 6:48 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> On Mon, Nov 22, 2021 at 1:45 PM Amit Langote <amitlangote09@gmail.com>\r\n> wrote:\r\n> >\r\n> > On Sat, Nov 20, 2021 at 8:31 PM Amit Kapila <amit.kapila16@gmail.com>\r\n> wrote:\r\n> > > On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com>\r\n> wrote:\r\n> > > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote\r\n> <amitlangote09@gmail.com> wrote:\r\n> > > > > The problematic case is attaching the partition *after* the\r\n> > > > > subscriber has already marked the root parent as synced and/or\r\n> > > > > ready for replication. Refreshing the subscription doesn't help\r\n> > > > > it discover the newly attached partition, because a\r\n> > > > > publish_via_partition_root only ever tells about the root\r\n> > > > > parent, which would be already synced, so the subscriber would think\r\n> > > > > there's nothing to copy.\r\n> > > >\r\n> > > > Okay, I see this could be a problem but I haven't tried to reproduce it.\r\n> > > >\r\n> > > > > > Anyway, if this is a problem\r\n> > > > > > we need to figure the solution for this separately.\r\n> > > > >\r\n> > > > > Sure, we might need to do that after all. Though it might be a\r\n> > > > > good idea to be sure that we won't need to reconsider the fix we\r\n> > > > > push for the issue(s) being discussed here and elsewhere,\r\n> > > > > because I suspect that the solution to the problem I mentioned\r\n> > > > > is likely to involve tweaking pg_publication_tables view output.\r\n> > >\r\n> > > I have thought about this problem and I see two possibilities for a\r\n> > > solution (a) We could provide a new option say 'truncate' (something\r\n> > > on lines proposed here [1]) which would truncate the table(s) and\r\n> > > change its status to 'i' in the pg_subscription_rel, this would\r\n> > > allow the newly added partition to be synced after refresh. This\r\n> > > could lead to a large copy in such a case.\r\n> >\r\n> > Maybe I am missing something about the proposal, though I'd think a\r\n> > more automatic solution would be better, something that doesn't need\r\n> > to rely on an unrelated feature.\r\n> >\r\n> \r\n> Agreed, this was more of a workaround for users if we didn't get any automatic\r\n> solution.\r\n> \r\n> > > (b) We could somehow get and store all the partition info from the\r\n> > > publisher-side on the subscriber-side while initial sync (say in new\r\n> > > system table pg_subscription_rel_members). Now, after the refresh,\r\n> > > if this list changes, we can allow to just get the data of that\r\n> > > particular partition but I guess it would mean that we need to store\r\n> > > oids of the publisher which might or might not be safe considering\r\n> > > oids can wraparound before the refresh.\r\n> > >\r\n> > > Do you have any other ideas?\r\n> >\r\n> > I thought that the idea I had earlier mentioned at [1] may be useful,\r\n> > which I can see is similar to your idea (b). I also suspect that it\r\n> > can be implemented without needing a separate catalog and storing\r\n> > publication-side relation OIDs in the subscription-side catalog,\r\n> > though maybe I haven't thought hard enough.\r\n> >\r\n> \r\n> The problem with storing this info in pg_subscription_rel as you were\r\n> describing in your proposal is that currently, we ensure that the same table\r\n> exists in subscriber and then store the subscriber side table id in that catalog. I\r\n> am not sure if we can store publisher-side oids in that catalog and if we store\r\n> then it would be confusing as now it will have info of both publisher-side oids\r\n> and subscriber-side oids. Isn't that a problem with this approach?\r\n> \r\n> One more problem with this overall approach is the risk of OID wraparound.\r\n> Say between Create Subscription and Alter Subscription ..\r\n> Refresh, we detach one partition, the oid wraparounds, and we create/reattach\r\n> another partition which gets the same oid as for the partition which we\r\n> detached earlier then we won't be able to identify the new partition even after\r\n> refresh.\r\n\r\nIf we use this approach, I think maybe we can store the publisher-side's table\r\nname instead of oid. Because for non-partitioned table, it's possible that user\r\ndrop and create a new table with the same name in which case the oid would be\r\nchanged. And in this case, the existing behavior wouldn't sync the table again.\r\nSo I think it might be better to make the partitioned table's behavior\r\nconsistent with the non-partitioned table which only sync the table's data when\r\ndetect new table name.\r\n\r\nBest regards,\r\nHou zj\r\n","msg_date":"Tue, 23 Nov 2021 01:58:05 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Tue, Nov 23, 2021 at 7:28 AM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n>\n> On Mon, Nov 22, 2021 6:48 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Mon, Nov 22, 2021 at 1:45 PM Amit Langote <amitlangote09@gmail.com>\n> > wrote:\n> > >\n> > > On Sat, Nov 20, 2021 at 8:31 PM Amit Kapila <amit.kapila16@gmail.com>\n> > wrote:\n> > > > On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com>\n> > wrote:\n> > > > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote\n> > <amitlangote09@gmail.com> wrote:\n> > > > > > The problematic case is attaching the partition *after* the\n> > > > > > subscriber has already marked the root parent as synced and/or\n> > > > > > ready for replication. Refreshing the subscription doesn't help\n> > > > > > it discover the newly attached partition, because a\n> > > > > > publish_via_partition_root only ever tells about the root\n> > > > > > parent, which would be already synced, so the subscriber would think\n> > > > > > there's nothing to copy.\n> > > > >\n> > > > > Okay, I see this could be a problem but I haven't tried to reproduce it.\n> > > > >\n> > > > > > > Anyway, if this is a problem\n> > > > > > > we need to figure the solution for this separately.\n> > > > > >\n> > > > > > Sure, we might need to do that after all. Though it might be a\n> > > > > > good idea to be sure that we won't need to reconsider the fix we\n> > > > > > push for the issue(s) being discussed here and elsewhere,\n> > > > > > because I suspect that the solution to the problem I mentioned\n> > > > > > is likely to involve tweaking pg_publication_tables view output.\n> > > >\n> > > > I have thought about this problem and I see two possibilities for a\n> > > > solution (a) We could provide a new option say 'truncate' (something\n> > > > on lines proposed here [1]) which would truncate the table(s) and\n> > > > change its status to 'i' in the pg_subscription_rel, this would\n> > > > allow the newly added partition to be synced after refresh. This\n> > > > could lead to a large copy in such a case.\n> > >\n> > > Maybe I am missing something about the proposal, though I'd think a\n> > > more automatic solution would be better, something that doesn't need\n> > > to rely on an unrelated feature.\n> > >\n> >\n> > Agreed, this was more of a workaround for users if we didn't get any automatic\n> > solution.\n> >\n> > > > (b) We could somehow get and store all the partition info from the\n> > > > publisher-side on the subscriber-side while initial sync (say in new\n> > > > system table pg_subscription_rel_members). Now, after the refresh,\n> > > > if this list changes, we can allow to just get the data of that\n> > > > particular partition but I guess it would mean that we need to store\n> > > > oids of the publisher which might or might not be safe considering\n> > > > oids can wraparound before the refresh.\n> > > >\n> > > > Do you have any other ideas?\n> > >\n> > > I thought that the idea I had earlier mentioned at [1] may be useful,\n> > > which I can see is similar to your idea (b). I also suspect that it\n> > > can be implemented without needing a separate catalog and storing\n> > > publication-side relation OIDs in the subscription-side catalog,\n> > > though maybe I haven't thought hard enough.\n> > >\n> >\n> > The problem with storing this info in pg_subscription_rel as you were\n> > describing in your proposal is that currently, we ensure that the same table\n> > exists in subscriber and then store the subscriber side table id in that catalog. I\n> > am not sure if we can store publisher-side oids in that catalog and if we store\n> > then it would be confusing as now it will have info of both publisher-side oids\n> > and subscriber-side oids. Isn't that a problem with this approach?\n> >\n> > One more problem with this overall approach is the risk of OID wraparound.\n> > Say between Create Subscription and Alter Subscription ..\n> > Refresh, we detach one partition, the oid wraparounds, and we create/reattach\n> > another partition which gets the same oid as for the partition which we\n> > detached earlier then we won't be able to identify the new partition even after\n> > refresh.\n>\n> If we use this approach, I think maybe we can store the publisher-side's table\n> name instead of oid. Because for non-partitioned table, it's possible that user\n> drop and create a new table with the same name in which case the oid would be\n> changed. And in this case, the existing behavior wouldn't sync the table again.\n> So I think it might be better to make the partitioned table's behavior\n> consistent with the non-partitioned table which only sync the table's data when\n> detect new table name.\n>\n\nI think firstly you need to then also store schema info. In the case\nof non-partitioned tables, it is derived differently because we remove\nthe table's entry from pg_publication_rel, so the data won't be sent\nto subscriber till the new table is added to the publication. In the\ncase of partitioned table (especially for the case of\npublish_via_partition_root) we won't sync initial data in such a case\n(dropped and re-created the partition with the same name) both with\nand without your solution but the future data will be streamed.\n\nAnother point related to the Attach problem is what if such a\npartition before was also part of publication separately, then if we\ncopy its initial data then that data copy might be considered a\nduplicate. Isn't it better to document this case and explain what\nusers can expect instead of trying to design a solution around this?\nEven if we do so the streaming after attach partition problem as\ndiscussed above should be fixed.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 23 Nov 2021 08:50:57 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Tue, Nov 23, 2021 at 12:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> Isn't it better to document this case and explain what\n> users can expect instead of trying to design a solution around this?\n\nI thought about the problems you've described and it looks like I\nhadn't considered many of the details which complicate implementing a\nsolution for this.\n\nSo yeah, documenting the ATTACH issue as a limitation sounds like the\nbest course for now. I might word it as follows and add it under\nNotes at https://www.postgresql.org/docs/current/sql-createpublication.html:\n\n\"ATTACHing a table into a partition tree whose root is published using\na publication with publish_via_partition_root set to true does not\nresult in the table's existing contents to be replicated.\"\n\nI'm not sure there's a clean enough workaround to this that we can add\nto the paragraph.\n\nDoes that make sense?\n\n> Even if we do so the streaming after attach partition problem as\n> discussed above should be fixed.\n\nI agree. I have reproduced the problem though haven't managed to pin\ndown the cause yet.\n\n--\nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 24 Nov 2021 15:32:04 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Wed, Nov 24, 2021 at 12:02 PM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Tue, Nov 23, 2021 at 12:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > Isn't it better to document this case and explain what\n> > users can expect instead of trying to design a solution around this?\n>\n> I thought about the problems you've described and it looks like I\n> hadn't considered many of the details which complicate implementing a\n> solution for this.\n>\n> So yeah, documenting the ATTACH issue as a limitation sounds like the\n> best course for now. I might word it as follows and add it under\n> Notes at https://www.postgresql.org/docs/current/sql-createpublication.html:\n>\n> \"ATTACHing a table into a partition tree whose root is published using\n> a publication with publish_via_partition_root set to true does not\n> result in the table's existing contents to be replicated.\"\n>\n\nInstead of \"to be\", shall we use \"being\"?\n\n> I'm not sure there's a clean enough workaround to this that we can add\n> to the paragraph.\n>\n> Does that make sense?\n>\n\nYeah, that sounds reasonable. I think the case with\n\"publish_via_partition_root = false\" should work but please test it\nonce. The other thing which we need to consider about all these fixes\nis whether to back patch them or not. I think it is on case to case\nbasis. I feel this limitation should be documented and backpatched,\nwhat do you think? Feel free to submit patches accordingly.\n\n> > Even if we do so the streaming after attach partition problem as\n> > discussed above should be fixed.\n>\n> I agree. I have reproduced the problem though haven't managed to pin\n> down the cause yet.\n>\n\nNo problem, do let us know whatever be your findings at the end. I\nthink now we know how to proceed with this attach issue, maybe we\nshould also try to commit the fixes for other issues.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 24 Nov 2021 14:14:20 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Nov 22, 2021 at 12:55 PM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > As in,\n> > > do we know of any replication (initial/streaming) misbehavior caused\n> > > by the duplicate partition OIDs in this case or is the only problem\n> > > that pg_publication_tables output looks odd?\n> >\n> > The latter one but I think either we should document this or change it\n> > as we can't assume users will follow what subscriber-side code does.\n>\n> On second thought, I agree that de-duplicating partitions from this\n> view is an improvement.\n>\n\nFair enough. Hou-San, Can you please submit the updated patch after\nfixing any pending comments including the test case?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 24 Nov 2021 14:17:54 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Wed, Nov 24, 2021 4:48 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> On Mon, Nov 22, 2021 at 12:55 PM Amit Langote <amitlangote09@gmail.com>\r\n> wrote:\r\n> >\r\n> > On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila <amit.kapila16@gmail.com>\r\n> wrote:\r\n> > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com>\r\n> wrote:\r\n> > > > As in,\r\n> > > > do we know of any replication (initial/streaming) misbehavior\r\n> > > > caused by the duplicate partition OIDs in this case or is the only\r\n> > > > problem that pg_publication_tables output looks odd?\r\n> > >\r\n> > > The latter one but I think either we should document this or change\r\n> > > it as we can't assume users will follow what subscriber-side code does.\r\n> >\r\n> > On second thought, I agree that de-duplicating partitions from this\r\n> > view is an improvement.\r\n> >\r\n> \r\n> Fair enough. Hou-San, Can you please submit the updated patch after fixing\r\n> any pending comments including the test case?\r\n\r\nOK, I will submit the updated patch soon.\r\n\r\nBest regards,\r\nHou zj\r\n","msg_date":"Thu, 25 Nov 2021 01:57:01 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Wed, Nov 24, 2021 at 5:44 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Wed, Nov 24, 2021 at 12:02 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > So yeah, documenting the ATTACH issue as a limitation sounds like the\n> > best course for now. I might word it as follows and add it under\n> > Notes at https://www.postgresql.org/docs/current/sql-createpublication.html:\n> >\n> > \"ATTACHing a table into a partition tree whose root is published using\n> > a publication with publish_via_partition_root set to true does not\n> > result in the table's existing contents to be replicated.\"\n>\n> Instead of \"to be\", shall we use \"being\"?\n\nThat reads better.\n\n> > I'm not sure there's a clean enough workaround to this that we can add\n> > to the paragraph.\n> >\n> > Does that make sense?\n> >\n>\n> Yeah, that sounds reasonable. I think the case with\n> \"publish_via_partition_root = false\" should work but please test it\n> once.\n\nYeah, I did test that case to compare and didn't see a problem with it.\n\n> The other thing which we need to consider about all these fixes\n> is whether to back patch them or not. I think it is on case to case\n> basis. I feel this limitation should be documented and backpatched,\n> what do you think? Feel free to submit patches accordingly.\n\nI agree with backpatching the doc fix. I've attached a diff against\nmaster, though it also appears to apply to 13 and 14 branches.\n\n--\nAmit Langote\nEDB: http://www.enterprisedb.com","msg_date":"Thu, 25 Nov 2021 17:00:36 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thu, Nov 25, 2021 at 1:30 PM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Wed, Nov 24, 2021 at 5:44 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Wed, Nov 24, 2021 at 12:02 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > So yeah, documenting the ATTACH issue as a limitation sounds like the\n> > > best course for now. I might word it as follows and add it under\n> > > Notes at https://www.postgresql.org/docs/current/sql-createpublication.html:\n> > >\n> > > \"ATTACHing a table into a partition tree whose root is published using\n> > > a publication with publish_via_partition_root set to true does not\n> > > result in the table's existing contents to be replicated.\"\n> >\n> > Instead of \"to be\", shall we use \"being\"?\n>\n> That reads better.\n>\n> > > I'm not sure there's a clean enough workaround to this that we can add\n> > > to the paragraph.\n> > >\n>\n> I agree with backpatching the doc fix. I've attached a diff against\n> master, though it also appears to apply to 13 and 14 branches.\n>\n\nI think we can <literal></literal> for publish_via_partition_root,\nother than that the patch looks good to me.\n\nHou-San, others, do you have any opinion about this patch and whether\nto backpatch it or not?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 25 Nov 2021 14:26:58 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thu, Nov 25, 2021 at 5:57 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Thu, Nov 25, 2021 at 1:30 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > I agree with backpatching the doc fix. I've attached a diff against\n> > master, though it also appears to apply to 13 and 14 branches.\n>\n> I think we can <literal></literal> for publish_via_partition_root,\n> other than that the patch looks good to me.\n\nOh right, fixed. I also added the tag for \"true\" in the same sentence.\n\n-- \nAmit Langote\nEDB: http://www.enterprisedb.com","msg_date":"Thu, 25 Nov 2021 21:08:13 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thursday, November 25, 2021 4:57 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> On Thu, Nov 25, 2021 at 1:30 PM Amit Langote <amitlangote09@gmail.com>\r\n> >\r\n> > I agree with backpatching the doc fix. I've attached a diff against\r\n> > master, though it also appears to apply to 13 and 14 branches.\r\n> >\r\n> \r\n> I think we can <literal></literal> for publish_via_partition_root, other than that\r\n> the patch looks good to me.\r\n> \r\n> Hou-San, others, do you have any opinion about this patch and whether to\r\n> backpatch it or not?\r\n\r\nI think it makes sense to backpatch the doc fix, and the patch looks good to me.\r\n\r\nBest regards,\r\nHou zj\r\n","msg_date":"Fri, 26 Nov 2021 01:40:21 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Fri, Nov 26, 2021 at 7:10 AM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n>\n> On Thursday, November 25, 2021 4:57 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Thu, Nov 25, 2021 at 1:30 PM Amit Langote <amitlangote09@gmail.com>\n> > >\n> > > I agree with backpatching the doc fix. I've attached a diff against\n> > > master, though it also appears to apply to 13 and 14 branches.\n> > >\n> >\n> > I think we can <literal></literal> for publish_via_partition_root, other than that\n> > the patch looks good to me.\n> >\n> > Hou-San, others, do you have any opinion about this patch and whether to\n> > backpatch it or not?\n>\n> I think it makes sense to backpatch the doc fix, and the patch looks good to me.\n>\n\nThanks, I'll push this sometime early next week unless there are any\nobjections to it.\n\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 26 Nov 2021 08:38:44 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Wed, Nov 24, 2021 4:48 PM Amit Kapila <amit.kapila16@gmail.com>\r\n> On Mon, Nov 22, 2021 at 12:55 PM Amit Langote <amitlangote09@gmail.com>\r\n> wrote:\r\n> >\r\n> > On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila <amit.kapila16@gmail.com>\r\n> wrote:\r\n> > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com>\r\n> wrote:\r\n> > > > As in,\r\n> > > > do we know of any replication (initial/streaming) misbehavior\r\n> > > > caused by the duplicate partition OIDs in this case or is the only\r\n> > > > problem that pg_publication_tables output looks odd?\r\n> > >\r\n> > > The latter one but I think either we should document this or change\r\n> > > it as we can't assume users will follow what subscriber-side code does.\r\n> >\r\n> > On second thought, I agree that de-duplicating partitions from this\r\n> > view is an improvement.\r\n> >\r\n> \r\n> Fair enough. Hou-San, Can you please submit the updated patch after fixing\r\n> any pending comments including the test case?\r\n\r\nAttach the updated patch which address the comments so far.\r\n\r\nThe patch only adds a testcase in publication.sql because the duplicate\r\noutput doesn't cause unexpected behavior in pub-sub test.\r\n\r\nBest regards,\r\nHou zj","msg_date":"Mon, 29 Nov 2021 09:07:41 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Fri, Nov 26, 2021 at 8:38 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Fri, Nov 26, 2021 at 7:10 AM houzj.fnst@fujitsu.com\n> <houzj.fnst@fujitsu.com> wrote:\n> >\n> > On Thursday, November 25, 2021 4:57 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > On Thu, Nov 25, 2021 at 1:30 PM Amit Langote <amitlangote09@gmail.com>\n> > > >\n> > > > I agree with backpatching the doc fix. I've attached a diff against\n> > > > master, though it also appears to apply to 13 and 14 branches.\n> > > >\n> > >\n> > > I think we can <literal></literal> for publish_via_partition_root, other than that\n> > > the patch looks good to me.\n> > >\n> > > Hou-San, others, do you have any opinion about this patch and whether to\n> > > backpatch it or not?\n> >\n> > I think it makes sense to backpatch the doc fix, and the patch looks good to me.\n> >\n>\n> Thanks, I'll push this sometime early next week unless there are any\n> objections to it.\n>\n\nPushed.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 1 Dec 2021 12:10:39 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Nov 29, 2021 at 2:37 PM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n>\n> On Wed, Nov 24, 2021 4:48 PM Amit Kapila <amit.kapila16@gmail.com>\n> > On Mon, Nov 22, 2021 at 12:55 PM Amit Langote <amitlangote09@gmail.com>\n> > wrote:\n> > >\n> > > On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila <amit.kapila16@gmail.com>\n> > wrote:\n> > > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com>\n> > wrote:\n> > > > > As in,\n> > > > > do we know of any replication (initial/streaming) misbehavior\n> > > > > caused by the duplicate partition OIDs in this case or is the only\n> > > > > problem that pg_publication_tables output looks odd?\n> > > >\n> > > > The latter one but I think either we should document this or change\n> > > > it as we can't assume users will follow what subscriber-side code does.\n> > >\n> > > On second thought, I agree that de-duplicating partitions from this\n> > > view is an improvement.\n> > >\n> >\n> > Fair enough. Hou-San, Can you please submit the updated patch after fixing\n> > any pending comments including the test case?\n>\n> Attach the updated patch which address the comments so far.\n>\n> The patch only adds a testcase in publication.sql because the duplicate\n> output doesn't cause unexpected behavior in pub-sub test.\n>\n\nThanks, the patch looks good to me. I have slightly changed the commit\nmessage in the attached. I would like to commit this only in HEAD as\nthere is no user complaint about this and it might not stop any user's\nservice unless it relies on this view's output for the initial table\nsynchronization.\n\nWhat do you think? Bharath/Amit L., others, do you have any opinion on\nthis matter?\n\n-- \nWith Regards,\nAmit Kapila.","msg_date":"Wed, 1 Dec 2021 12:30:57 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Wed, Dec 1, 2021 3:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> On Mon, Nov 29, 2021 at 2:37 PM houzj.fnst@fujitsu.com\r\n> <houzj.fnst@fujitsu.com> wrote:\r\n> >\r\n> > On Wed, Nov 24, 2021 4:48 PM Amit Kapila <amit.kapila16@gmail.com>\r\n> > > On Mon, Nov 22, 2021 at 12:55 PM Amit Langote\r\n> > > <amitlangote09@gmail.com>\r\n> > > wrote:\r\n> > > >\r\n> > > > On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila\r\n> > > > <amit.kapila16@gmail.com>\r\n> > > wrote:\r\n> > > > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote\r\n> > > > > <amitlangote09@gmail.com>\r\n> > > wrote:\r\n> > > > > > As in,\r\n> > > > > > do we know of any replication (initial/streaming) misbehavior\r\n> > > > > > caused by the duplicate partition OIDs in this case or is the\r\n> > > > > > only problem that pg_publication_tables output looks odd?\r\n> > > > >\r\n> > > > > The latter one but I think either we should document this or\r\n> > > > > change it as we can't assume users will follow what subscriber-side\r\n> > > > > code does.\r\n> > > >\r\n> > > > On second thought, I agree that de-duplicating partitions from\r\n> > > > this view is an improvement.\r\n> > > >\r\n> > >\r\n> > > Fair enough. Hou-San, Can you please submit the updated patch after\r\n> > > fixing any pending comments including the test case?\r\n> >\r\n> > Attach the updated patch which address the comments so far.\r\n> >\r\n> > The patch only adds a testcase in publication.sql because the\r\n> > duplicate output doesn't cause unexpected behavior in pub-sub test.\r\n> >\r\n> \r\n> Thanks, the patch looks good to me. I have slightly changed the commit\r\n> message in the attached. I would like to commit this only in HEAD as there is no\r\n> user complaint about this and it might not stop any user's service unless it relies\r\n> on this view's output for the initial table synchronization.\r\n> \r\n> What do you think?\r\nI agreed that we can commit only in HEAD.\r\n\r\nBest regards,\r\nHou zj\r\n\r\n","msg_date":"Thu, 2 Dec 2021 00:44:42 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thu, Dec 2, 2021 at 9:44 houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n> On Wed, Dec 1, 2021 3:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > > On Mon, Nov 22, 2021 at 12:55 PM Amit Langote\n> > > > <amitlangote09@gmail.com> wrote:\n> > > > > On second thought, I agree that de-duplicating partitions from\n> > > > > this view is an improvement.\n> > > >\n> > > > Fair enough. Hou-San, Can you please submit the updated patch after\n> > > > fixing any pending comments including the test case?\n> > >\n> > > Attach the updated patch which address the comments so far.\n> > >\n> > > The patch only adds a testcase in publication.sql because the\n> > > duplicate output doesn't cause unexpected behavior in pub-sub test.\n> >\n> > Thanks, the patch looks good to me. I have slightly changed the commit\n> > message in the attached. I would like to commit this only in HEAD as there is no\n> > user complaint about this and it might not stop any user's service unless it relies\n> > on this view's output for the initial table synchronization.\n\nThe patch looks good to me too in that it gets the job done.\n\nReading Alvaro's email at the top again gave me a pause to reconsider\nwhat I had said in reply. It might indeed have been nice if the\npublication DDL itself had prevented the cases where a partition and\nits ancestor are added to a publication, though as Hou-san mentioned,\npartition ATTACHes make that a bit tricky to enforce at all times,\nthough of course not impossible. Maybe it's okay to just de-duplicate\npg_publication_tables output as the patch does, even though it may\nappear to be a band-aid solution if we one considers Alvaro's point\nabout consistency. Sorry about too many second thoughts on these\nthreads. :(\n\n> > What do you think?\n> I agreed that we can commit only in HEAD.\n\n+1 to do this only in HEAD.\n\nThank you.\n\n\n--\nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 2 Dec 2021 12:12:25 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Thu, Dec 2, 2021 at 9:44 houzj.fnst@fujitsu.com\n> <houzj.fnst@fujitsu.com> wrote:\n> > On Wed, Dec 1, 2021 3:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > > > On Mon, Nov 22, 2021 at 12:55 PM Amit Langote\n> > > > > <amitlangote09@gmail.com> wrote:\n> > > > > > On second thought, I agree that de-duplicating partitions from\n> > > > > > this view is an improvement.\n> > > > >\n> > > > > Fair enough. Hou-San, Can you please submit the updated patch after\n> > > > > fixing any pending comments including the test case?\n> > > >\n> > > > Attach the updated patch which address the comments so far.\n> > > >\n> > > > The patch only adds a testcase in publication.sql because the\n> > > > duplicate output doesn't cause unexpected behavior in pub-sub test.\n> > >\n> > > Thanks, the patch looks good to me. I have slightly changed the commit\n> > > message in the attached. I would like to commit this only in HEAD as there is no\n> > > user complaint about this and it might not stop any user's service unless it relies\n> > > on this view's output for the initial table synchronization.\n>\n> The patch looks good to me too in that it gets the job done.\n>\n> Reading Alvaro's email at the top again gave me a pause to reconsider\n> what I had said in reply. It might indeed have been nice if the\n> publication DDL itself had prevented the cases where a partition and\n> its ancestor are added to a publication, though as Hou-san mentioned,\n> partition ATTACHes make that a bit tricky to enforce at all times,\n> though of course not impossible. Maybe it's okay to just de-duplicate\n> pg_publication_tables output as the patch does, even though it may\n> appear to be a band-aid solution if we one considers Alvaro's point\n> about consistency.\n>\n\nTrue, I think even if we consider that idea it will be a much bigger\nchange, and also as it will be a behavioral change so we might want to\nkeep it just for HEAD and some of these fixes need to be backpatched.\nHaving said that, if you or someone want to pursue that idea and come\nup with a better solution than what we have currently it is certainly\nwelcome.\n\n\n> > > What do you think?\n> > I agreed that we can commit only in HEAD.\n>\n> +1 to do this only in HEAD.\n>\n\nOkay, If there are no further suggestions or objections, I'll commit\nthe patch early next week (mostly by Monday).\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 2 Dec 2021 10:57:41 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> > Reading Alvaro's email at the top again gave me a pause to reconsider\n> > what I had said in reply. It might indeed have been nice if the\n> > publication DDL itself had prevented the cases where a partition and\n> > its ancestor are added to a publication, though as Hou-san mentioned,\n> > partition ATTACHes make that a bit tricky to enforce at all times,\n> > though of course not impossible. Maybe it's okay to just de-duplicate\n> > pg_publication_tables output as the patch does, even though it may\n> > appear to be a band-aid solution if we one considers Alvaro's point\n> > about consistency.\n>\n> True, I think even if we consider that idea it will be a much bigger\n> change, and also as it will be a behavioral change so we might want to\n> keep it just for HEAD and some of these fixes need to be backpatched.\n> Having said that, if you or someone want to pursue that idea and come\n> up with a better solution than what we have currently it is certainly\n> welcome.\n\nOkay, I did write a PoC patch this morning after sending out my\nearlier email. I polished it a bit, which is attached.\n\n-- \nAmit Langote\nEDB: http://www.enterprisedb.com","msg_date":"Thu, 2 Dec 2021 22:48:08 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > Reading Alvaro's email at the top again gave me a pause to reconsider\n> > > what I had said in reply. It might indeed have been nice if the\n> > > publication DDL itself had prevented the cases where a partition and\n> > > its ancestor are added to a publication, though as Hou-san mentioned,\n> > > partition ATTACHes make that a bit tricky to enforce at all times,\n> > > though of course not impossible. Maybe it's okay to just de-duplicate\n> > > pg_publication_tables output as the patch does, even though it may\n> > > appear to be a band-aid solution if we one considers Alvaro's point\n> > > about consistency.\n> >\n> > True, I think even if we consider that idea it will be a much bigger\n> > change, and also as it will be a behavioral change so we might want to\n> > keep it just for HEAD and some of these fixes need to be backpatched.\n> > Having said that, if you or someone want to pursue that idea and come\n> > up with a better solution than what we have currently it is certainly\n> > welcome.\n>\n> Okay, I did write a PoC patch this morning after sending out my\n> earlier email. I polished it a bit, which is attached.\n>\n\nI see multiple problems with this patch and idea. (a) I think you\nforgot to deal with \"All Tables In Schema\" Publication which will be\nquite tricky to deal with during attach operation. How will you remove\na particular relation from such a publication if there is a need to do\nso? (b) Also, how will we prohibit adding partition and its root in\nthe case of \"All Tables In Schema\" or \"All Tables\" publication? If\nthere is no way to do that, then that would mean we anyway need to\ndeal with parent and child tables as part of the same publication and\nthis new behavior will add an exception to it. (c) I think this can\nmake switching \"publish_via_partition_root\" inconvenient for users.\nSay all or some of the partitions are part of the publication and then\nthe user decides to add root table and change publish option\n\"publish_via_partition_root\" to \"true\" at that moment she needs to\nremove all partitions first.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 3 Dec 2021 09:07:38 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Fri, Dec 3, 2021 at 12:37 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > > Reading Alvaro's email at the top again gave me a pause to reconsider\n> > > > what I had said in reply. It might indeed have been nice if the\n> > > > publication DDL itself had prevented the cases where a partition and\n> > > > its ancestor are added to a publication, though as Hou-san mentioned,\n> > > > partition ATTACHes make that a bit tricky to enforce at all times,\n> > > > though of course not impossible. Maybe it's okay to just de-duplicate\n> > > > pg_publication_tables output as the patch does, even though it may\n> > > > appear to be a band-aid solution if we one considers Alvaro's point\n> > > > about consistency.\n> > >\n> > > True, I think even if we consider that idea it will be a much bigger\n> > > change, and also as it will be a behavioral change so we might want to\n> > > keep it just for HEAD and some of these fixes need to be backpatched.\n> > > Having said that, if you or someone want to pursue that idea and come\n> > > up with a better solution than what we have currently it is certainly\n> > > welcome.\n> >\n> > Okay, I did write a PoC patch this morning after sending out my\n> > earlier email. I polished it a bit, which is attached.\n>\n> I see multiple problems with this patch and idea.\n\nThanks for looking at it. Yeah, I have not looked very closely at ALL\nTABLES [IN SCHEMA], though only because I suspected that those cases\ndeal with partitioning in such a way that the partition duplication\nissue doesn't arise. That is, only the FOR TABLE list_of_tables and\nADD TABLE syntax allow for the duplication issue to occur.\n\n> (a) I think you\n> forgot to deal with \"All Tables In Schema\" Publication which will be\n> quite tricky to deal with during attach operation. How will you remove\n> a particular relation from such a publication if there is a need to do\n> so?\n\nHmm, my understanding of how FOR ALL TABLES... features work is that\none cannot remove a particular relation from such publications?\n\ncreate schema sch;\ncreate table sch.p (a int primary key) partition by list (a);\ncreate table sch.p1 partition of sch.p for values in (1);\ncreate table sch.p2 partition of sch.p for values in (2);\ncreate table p (a int primary key) partition by list (a);\ncreate table p1 partition of p for values in (1);\ncreate table p2 partition of p for values in (2);\ncreate publication puball for all tables;\ncreate publication pubsch for all tables in schema sch;\n\nalter publication puball drop table p;\nERROR: publication \"puball\" is defined as FOR ALL TABLES\nDETAIL: Tables cannot be added to or dropped from FOR ALL TABLES publications.\n\nalter publication pubsch drop table sch.p;\nERROR: relation \"p\" is not part of the publication\n\nWhat am I missing?\n\n> (b) Also, how will we prohibit adding partition and its root in\n> the case of \"All Tables In Schema\" or \"All Tables\" publication? If\n> there is no way to do that, then that would mean we anyway need to\n> deal with parent and child tables as part of the same publication and\n> this new behavior will add an exception to it.\n\nI checked that FOR ALL TABLES publications don't allow adding a table\nexplicitly, but apparently IN SCHEMA one does:\n\nalter publication pubsch add table p2;\n\\dRp+ pubsch\n Publication pubsch\n Owner | All tables | Inserts | Updates | Deletes | Truncates | Via root\n-------+------------+---------+---------+---------+-----------+----------\n amit | f | t | t | t | t | f\nTables:\n \"public.p2\"\nTables from schemas:\n \"sch\"\n\nISTM that the ..IN SCHEMA syntax does not behave like FOR ALL TABLES\nwithout the IN SCHEMA in this regard. Is that correct?\n\n> (c) I think this can\n> make switching \"publish_via_partition_root\" inconvenient for users.\n> Say all or some of the partitions are part of the publication and then\n> the user decides to add root table and change publish option\n> \"publish_via_partition_root\" to \"true\" at that moment she needs to\n> remove all partitions first.\n\nHmm, actually I don't think I have tested the case where some or all\npartitions are present by themselves before adding the root table.\nI'll check that and try to think of sensible behavior for that.\n\n-- \nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 3 Dec 2021 21:34:35 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Fri, Dec 3, 2021 at 21:34 Amit Langote <amitlangote09@gmail.com> wrote:\n\n> On Fri, Dec 3, 2021 at 12:37 PM Amit Kapila <amit.kapila16@gmail.com>\n> wrote:\n> > On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com>\n> wrote:\n> > > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com>\n> wrote:\n> > > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com>\n> wrote:\n> > > > > Reading Alvaro's email at the top again gave me a pause to\n> reconsider\n> > > > > what I had said in reply. It might indeed have been nice if the\n> > > > > publication DDL itself had prevented the cases where a partition\n> and\n> > > > > its ancestor are added to a publication, though as Hou-san\n> mentioned,\n> > > > > partition ATTACHes make that a bit tricky to enforce at all times,\n> > > > > though of course not impossible. Maybe it's okay to just\n> de-duplicate\n> > > > > pg_publication_tables output as the patch does, even though it may\n> > > > > appear to be a band-aid solution if we one considers Alvaro's point\n> > > > > about consistency.\n> > > >\n> > > > True, I think even if we consider that idea it will be a much bigger\n> > > > change, and also as it will be a behavioral change so we might want\n> to\n> > > > keep it just for HEAD and some of these fixes need to be backpatched.\n> > > > Having said that, if you or someone want to pursue that idea and come\n> > > > up with a better solution than what we have currently it is certainly\n> > > > welcome.\n> > >\n> > > Okay, I did write a PoC patch this morning after sending out my\n> > > earlier email. I polished it a bit, which is attached.\n> >\n> > I see multiple problems with this patch and idea.\n>\n> Thanks for looking at it. Yeah, I have not looked very closely at ALL\n> TABLES [IN SCHEMA], though only because I suspected that those cases\n> deal with partitioning in such a way that the partition duplication\n> issue doesn't arise. That is, only the FOR TABLE list_of_tables and\n> ADD TABLE syntax allow for the duplication issue to occur.\n\n\nAnother thing I forgot to mention is that the patch passes check-world.\nPerhaps we don’t have enough tests that would’ve exposed any problems with\nthe patch’s approach.\n-- \nAmit Langote\nEDB: http://www.enterprisedb.com\n\nOn Fri, Dec 3, 2021 at 21:34 Amit Langote <amitlangote09@gmail.com> wrote:On Fri, Dec 3, 2021 at 12:37 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > > Reading Alvaro's email at the top again gave me a pause to reconsider\n> > > > what I had said in reply. It might indeed have been nice if the\n> > > > publication DDL itself had prevented the cases where a partition and\n> > > > its ancestor are added to a publication, though as Hou-san mentioned,\n> > > > partition ATTACHes make that a bit tricky to enforce at all times,\n> > > > though of course not impossible. Maybe it's okay to just de-duplicate\n> > > > pg_publication_tables output as the patch does, even though it may\n> > > > appear to be a band-aid solution if we one considers Alvaro's point\n> > > > about consistency.\n> > >\n> > > True, I think even if we consider that idea it will be a much bigger\n> > > change, and also as it will be a behavioral change so we might want to\n> > > keep it just for HEAD and some of these fixes need to be backpatched.\n> > > Having said that, if you or someone want to pursue that idea and come\n> > > up with a better solution than what we have currently it is certainly\n> > > welcome.\n> >\n> > Okay, I did write a PoC patch this morning after sending out my\n> > earlier email. I polished it a bit, which is attached.\n>\n> I see multiple problems with this patch and idea.\n\nThanks for looking at it. Yeah, I have not looked very closely at ALL\nTABLES [IN SCHEMA], though only because I suspected that those cases\ndeal with partitioning in such a way that the partition duplication\nissue doesn't arise. That is, only the FOR TABLE list_of_tables and\nADD TABLE syntax allow for the duplication issue to occur.Another thing I forgot to mention is that the patch passes check-world. Perhaps we don’t have enough tests that would’ve exposed any problems with the patch’s approach.-- Amit LangoteEDB: http://www.enterprisedb.com","msg_date":"Fri, 3 Dec 2021 22:27:43 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Fri, Dec 3, 2021 at 6:04 PM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Fri, Dec 3, 2021 at 12:37 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > > > Reading Alvaro's email at the top again gave me a pause to reconsider\n> > > > > what I had said in reply. It might indeed have been nice if the\n> > > > > publication DDL itself had prevented the cases where a partition and\n> > > > > its ancestor are added to a publication, though as Hou-san mentioned,\n> > > > > partition ATTACHes make that a bit tricky to enforce at all times,\n> > > > > though of course not impossible. Maybe it's okay to just de-duplicate\n> > > > > pg_publication_tables output as the patch does, even though it may\n> > > > > appear to be a band-aid solution if we one considers Alvaro's point\n> > > > > about consistency.\n> > > >\n> > > > True, I think even if we consider that idea it will be a much bigger\n> > > > change, and also as it will be a behavioral change so we might want to\n> > > > keep it just for HEAD and some of these fixes need to be backpatched.\n> > > > Having said that, if you or someone want to pursue that idea and come\n> > > > up with a better solution than what we have currently it is certainly\n> > > > welcome.\n> > >\n> > > Okay, I did write a PoC patch this morning after sending out my\n> > > earlier email. I polished it a bit, which is attached.\n> >\n> > I see multiple problems with this patch and idea.\n>\n> Thanks for looking at it. Yeah, I have not looked very closely at ALL\n> TABLES [IN SCHEMA], though only because I suspected that those cases\n> deal with partitioning in such a way that the partition duplication\n> issue doesn't arise. That is, only the FOR TABLE list_of_tables and\n> ADD TABLE syntax allow for the duplication issue to occur.\n>\n> > (a) I think you\n> > forgot to deal with \"All Tables In Schema\" Publication which will be\n> > quite tricky to deal with during attach operation. How will you remove\n> > a particular relation from such a publication if there is a need to do\n> > so?\n>\n> Hmm, my understanding of how FOR ALL TABLES... features work is that\n> one cannot remove a particular relation from such publications?\n>\n> create schema sch;\n> create table sch.p (a int primary key) partition by list (a);\n> create table sch.p1 partition of sch.p for values in (1);\n> create table sch.p2 partition of sch.p for values in (2);\n> create table p (a int primary key) partition by list (a);\n> create table p1 partition of p for values in (1);\n> create table p2 partition of p for values in (2);\n> create publication puball for all tables;\n> create publication pubsch for all tables in schema sch;\n>\n> alter publication puball drop table p;\n> ERROR: publication \"puball\" is defined as FOR ALL TABLES\n> DETAIL: Tables cannot be added to or dropped from FOR ALL TABLES publications.\n>\n> alter publication pubsch drop table sch.p;\n> ERROR: relation \"p\" is not part of the publication\n>\n> What am I missing?\n>\n\nCurrently, in your patch, you are trying to remove a particular\nrelation/partition during attach but how will you do that if such a\nrelation is part of All Tables In Schema publication?\n\n> > (b) Also, how will we prohibit adding partition and its root in\n> > the case of \"All Tables In Schema\" or \"All Tables\" publication? If\n> > there is no way to do that, then that would mean we anyway need to\n> > deal with parent and child tables as part of the same publication and\n> > this new behavior will add an exception to it.\n>\n> I checked that FOR ALL TABLES publications don't allow adding a table\n> explicitly, but apparently IN SCHEMA one does:\n>\n> alter publication pubsch add table p2;\n> \\dRp+ pubsch\n> Publication pubsch\n> Owner | All tables | Inserts | Updates | Deletes | Truncates | Via root\n> -------+------------+---------+---------+---------+-----------+----------\n> amit | f | t | t | t | t | f\n> Tables:\n> \"public.p2\"\n> Tables from schemas:\n> \"sch\"\n>\n> ISTM that the ..IN SCHEMA syntax does not behave like FOR ALL TABLES\n> without the IN SCHEMA in this regard. Is that correct?\n>\n\nWe do allow adding additional tables or schema to an exiting All\ntables in schema publication.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 6 Dec 2021 10:29:20 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thursday, December 2, 2021 9:48 PM Amit Langote <amitlangote09@gmail.com> wrote:\r\n> On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com>\r\n> wrote:\r\n> > > Reading Alvaro's email at the top again gave me a pause to\r\n> > > reconsider what I had said in reply. It might indeed have been nice\r\n> > > if the publication DDL itself had prevented the cases where a\r\n> > > partition and its ancestor are added to a publication, though as\r\n> > > Hou-san mentioned, partition ATTACHes make that a bit tricky to\r\n> > > enforce at all times, though of course not impossible. Maybe it's\r\n> > > okay to just de-duplicate pg_publication_tables output as the patch\r\n> > > does, even though it may appear to be a band-aid solution if we one\r\n> > > considers Alvaro's point about consistency.\r\n> >\r\n> > True, I think even if we consider that idea it will be a much bigger\r\n> > change, and also as it will be a behavioral change so we might want to\r\n> > keep it just for HEAD and some of these fixes need to be backpatched.\r\n> > Having said that, if you or someone want to pursue that idea and come\r\n> > up with a better solution than what we have currently it is certainly\r\n> > welcome.\r\n> \r\n> Okay, I did write a PoC patch this morning after sending out my earlier email. I\r\n> polished it a bit, which is attached.\r\n\r\nAfter thinking more on this. I find there might be some usage about adding both\r\nchild and parent to the publication.\r\n\r\nFor the future feature publication row filter(and maybe column filter), It\r\ncould be useful for user to adding child and parent with different filter\r\nexpression. If pubviaroot=true, user can expect the parent's filter take\r\neffect, If pubviaroot=false, they can expect the child's filter take effect.\r\n\r\nIf we disallow adding both child and parent to publication, it could be harder\r\nfor these features to implement.\r\n\r\nSo, personally, I am not sure disallow adding both child and parent to the\r\npublication is a good idea.\r\n\r\nBest regards,\r\nHou zj\r\n","msg_date":"Mon, 6 Dec 2021 06:55:34 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Dec 6, 2021 at 3:55 PM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n> On Thursday, December 2, 2021 9:48 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com>\n> > wrote:\n> > > > Reading Alvaro's email at the top again gave me a pause to\n> > > > reconsider what I had said in reply. It might indeed have been nice\n> > > > if the publication DDL itself had prevented the cases where a\n> > > > partition and its ancestor are added to a publication, though as\n> > > > Hou-san mentioned, partition ATTACHes make that a bit tricky to\n> > > > enforce at all times, though of course not impossible. Maybe it's\n> > > > okay to just de-duplicate pg_publication_tables output as the patch\n> > > > does, even though it may appear to be a band-aid solution if we one\n> > > > considers Alvaro's point about consistency.\n> > >\n> > > True, I think even if we consider that idea it will be a much bigger\n> > > change, and also as it will be a behavioral change so we might want to\n> > > keep it just for HEAD and some of these fixes need to be backpatched.\n> > > Having said that, if you or someone want to pursue that idea and come\n> > > up with a better solution than what we have currently it is certainly\n> > > welcome.\n> >\n> > Okay, I did write a PoC patch this morning after sending out my earlier email. I\n> > polished it a bit, which is attached.\n>\n> After thinking more on this. I find there might be some usage about adding both\n> child and parent to the publication.\n>\n> For the future feature publication row filter(and maybe column filter), It\n> could be useful for user to adding child and parent with different filter\n> expression. If pubviaroot=true, user can expect the parent's filter take\n> effect, If pubviaroot=false, they can expect the child's filter take effect.\n>\n> If we disallow adding both child and parent to publication, it could be harder\n> for these features to implement.\n\nIt's possible that one may want such a feature, and yes, outright\npreventing adding both the parent and a partition to a publication\nwould perhaps make it harder. Mind you though that any such feature\nwould need to fix the current implementation anyway to make the\npublication-behavior-related catalog entries for child tables, which\nwe currently don't, unless children are added explicitly.\n\nThat said, we should also be careful in implementing new features that\noffer a finer per-partition granularity, because the fact that we do\ncurrently offer that for many of the existing old features such as\nconstraints, indexes, etc. limits the scalability of our architecture.\nThat point may be off-topic for the thread, but something to consider,\nor maybe I need to take a look at the other patch to see if my\nconcerns are addressable / have been addressed.\n\n\n--\nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 6 Dec 2021 16:30:00 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Dec 6, 2021 at 1:00 PM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Mon, Dec 6, 2021 at 3:55 PM houzj.fnst@fujitsu.com\n> <houzj.fnst@fujitsu.com> wrote:\n> > On Thursday, December 2, 2021 9:48 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com>\n> > > wrote:\n> > > > > Reading Alvaro's email at the top again gave me a pause to\n> > > > > reconsider what I had said in reply. It might indeed have been nice\n> > > > > if the publication DDL itself had prevented the cases where a\n> > > > > partition and its ancestor are added to a publication, though as\n> > > > > Hou-san mentioned, partition ATTACHes make that a bit tricky to\n> > > > > enforce at all times, though of course not impossible. Maybe it's\n> > > > > okay to just de-duplicate pg_publication_tables output as the patch\n> > > > > does, even though it may appear to be a band-aid solution if we one\n> > > > > considers Alvaro's point about consistency.\n> > > >\n> > > > True, I think even if we consider that idea it will be a much bigger\n> > > > change, and also as it will be a behavioral change so we might want to\n> > > > keep it just for HEAD and some of these fixes need to be backpatched.\n> > > > Having said that, if you or someone want to pursue that idea and come\n> > > > up with a better solution than what we have currently it is certainly\n> > > > welcome.\n> > >\n> > > Okay, I did write a PoC patch this morning after sending out my earlier email. I\n> > > polished it a bit, which is attached.\n> >\n> > After thinking more on this. I find there might be some usage about adding both\n> > child and parent to the publication.\n> >\n> > For the future feature publication row filter(and maybe column filter), It\n> > could be useful for user to adding child and parent with different filter\n> > expression. If pubviaroot=true, user can expect the parent's filter take\n> > effect, If pubviaroot=false, they can expect the child's filter take effect.\n> >\n> > If we disallow adding both child and parent to publication, it could be harder\n> > for these features to implement.\n>\n> It's possible that one may want such a feature, and yes, outright\n> preventing adding both the parent and a partition to a publication\n> would perhaps make it harder. Mind you though that any such feature\n> would need to fix the current implementation anyway to make the\n> publication-behavior-related catalog entries for child tables, which\n> we currently don't, unless children are added explicitly.\n>\n> That said, we should also be careful in implementing new features that\n> offer a finer per-partition granularity, because the fact that we do\n> currently offer that for many of the existing old features such as\n> constraints, indexes, etc. limits the scalability of our architecture.\n> That point may be off-topic for the thread, but something to consider,\n> or maybe I need to take a look at the other patch to see if my\n> concerns are addressable / have been addressed.\n>\n\nYour concern is not very clear to me. Can you be more specific in what\nkind of problem you see with such a design for row filtering? You\nmight want to once look at that thread. You can find a summary of that\npatch in the email [1].\n\n[1] - https://www.postgresql.org/message-id/49ba49f1-8bdb-40b7-ae9e-f17d88b3afcd%40www.fastmail.com\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 6 Dec 2021 14:35:36 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Dec 6, 2021 at 1:59 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Fri, Dec 3, 2021 at 6:04 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > On Fri, Dec 3, 2021 at 12:37 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > > Okay, I did write a PoC patch this morning after sending out my\n> > > > earlier email. I polished it a bit, which is attached.\n> > >\n> > > I see multiple problems with this patch and idea.\n> >\n> > Thanks for looking at it. Yeah, I have not looked very closely at ALL\n> > TABLES [IN SCHEMA], though only because I suspected that those cases\n> > deal with partitioning in such a way that the partition duplication\n> > issue doesn't arise. That is, only the FOR TABLE list_of_tables and\n> > ADD TABLE syntax allow for the duplication issue to occur.\n> >\n> > > (a) I think you\n> > > forgot to deal with \"All Tables In Schema\" Publication which will be\n> > > quite tricky to deal with during attach operation. How will you remove\n> > > a particular relation from such a publication if there is a need to do\n> > > so?\n> >\n> > Hmm, my understanding of how FOR ALL TABLES... features work is that\n> > one cannot remove a particular relation from such publications?\n> >\n> > create schema sch;\n> > create table sch.p (a int primary key) partition by list (a);\n> > create table sch.p1 partition of sch.p for values in (1);\n> > create table sch.p2 partition of sch.p for values in (2);\n> > create table p (a int primary key) partition by list (a);\n> > create table p1 partition of p for values in (1);\n> > create table p2 partition of p for values in (2);\n> > create publication puball for all tables;\n> > create publication pubsch for all tables in schema sch;\n> >\n> > alter publication puball drop table p;\n> > ERROR: publication \"puball\" is defined as FOR ALL TABLES\n> > DETAIL: Tables cannot be added to or dropped from FOR ALL TABLES publications.\n> >\n> > alter publication pubsch drop table sch.p;\n> > ERROR: relation \"p\" is not part of the publication\n> >\n> > What am I missing?\n>\n> Currently, in your patch, you are trying to remove a particular\n> relation/partition during attach but how will you do that if such a\n> relation is part of All Tables In Schema publication?\n\nSo IIUC the scenario of concern is when a table to be attached as a\npartition is in a schema that's present in pg_publication_namespace.\nThe only way to stop it from being published is to move it to another\nschema that is not published using that publication.\n\nI think I misunderstood how the IN SCHEMA feature works.\nSpecifically, I didn't know that one can add a partitioned table to\nthe same publication (or any table other than those in a particular\nschema for that matter). Then the attached partition would still be\npresent in the publication by way of being part of the schema that is\npresent in the publication, along with the partitioned table that is\nadded separately.\n\nYes, my proposal in its current form can't prevent that kind of duplication.\n\n--\nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 6 Dec 2021 23:34:29 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Dec 6, 2021 at 6:05 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Mon, Dec 6, 2021 at 1:00 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > On Mon, Dec 6, 2021 at 3:55 PM houzj.fnst@fujitsu.com\n> > <houzj.fnst@fujitsu.com> wrote:\n> > > After thinking more on this. I find there might be some usage about adding both\n> > > child and parent to the publication.\n> > >\n> > > For the future feature publication row filter(and maybe column filter), It\n> > > could be useful for user to adding child and parent with different filter\n> > > expression. If pubviaroot=true, user can expect the parent's filter take\n> > > effect, If pubviaroot=false, they can expect the child's filter take effect.\n> > >\n> > > If we disallow adding both child and parent to publication, it could be harder\n> > > for these features to implement.\n> >\n> > It's possible that one may want such a feature, and yes, outright\n> > preventing adding both the parent and a partition to a publication\n> > would perhaps make it harder. Mind you though that any such feature\n> > would need to fix the current implementation anyway to make the\n> > publication-behavior-related catalog entries for child tables, which\n> > we currently don't, unless children are added explicitly.\n> >\n> > That said, we should also be careful in implementing new features that\n> > offer a finer per-partition granularity, because the fact that we do\n> > currently offer that for many of the existing old features such as\n> > constraints, indexes, etc. limits the scalability of our architecture.\n> > That point may be off-topic for the thread, but something to consider,\n> > or maybe I need to take a look at the other patch to see if my\n> > concerns are addressable / have been addressed.\n> >\n>\n> Your concern is not very clear to me. Can you be more specific in what\n> kind of problem you see with such a design for row filtering?\n\nI guess my characterization of it is still vague but the problem I see\nis that we'll be adding one more feature that allows manipulating\npartitions directly, which means more system catalogs than there\nalready are that know about partitions as objects. I remember being\ndeliberate about avoiding pg_publication_rel entries for all\npartitions when a partitioned table is added to a publication for that\nvery reason.\n\nIs covering the use case of defining different filters for each\npartition important for developing that feature? I feel that defining\none for the root table and just using that for all partitions (albeit\ntranslating as needed to account for possibly different attribute\nnumbers) would be enough for most users' use cases. Maybe I'm wrong\non that.\n\n> You\n> might want to once look at that thread. You can find a summary of that\n> patch in the email [1].\n\nOkay, I'll try to do that.\n\n\n--\nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 7 Dec 2021 00:29:06 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Dec 6, 2021 at 8:04 PM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> So IIUC the scenario of concern is when a table to be attached as a\n> partition is in a schema that's present in pg_publication_namespace.\n> The only way to stop it from being published is to move it to another\n> schema that is not published using that publication.\n>\n> I think I misunderstood how the IN SCHEMA feature works.\n> Specifically, I didn't know that one can add a partitioned table to\n> the same publication (or any table other than those in a particular\n> schema for that matter). Then the attached partition would still be\n> present in the publication by way of being part of the schema that is\n> present in the publication, along with the partitioned table that is\n> added separately.\n>\n\nRight.\n\n> Yes, my proposal in its current form can't prevent that kind of duplication.\n>\n\nI am not sure how to proceed here. I feel it is better to go ahead\nwith the fix Hou-san proposed here and in another email [1] to fix the\nknow issues, especially because the issue discussed in [1] needs to be\nback-patched. We can evaluate your proposal separately for HEAD. What\ndo you think? I am fine if you think that we should evaluate/analyze\nyour proposal before making a call on whether to fix the existing\nissues with the proposed set of patches.\n\n[1] - https://www.postgresql.org/message-id/OS0PR01MB57165D94CB187A97128F75EA946A9%40OS0PR01MB5716.jpnprd01.prod.outlook.com\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 7 Dec 2021 09:15:46 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Mon, Dec 6, 2021 at 8:59 PM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Mon, Dec 6, 2021 at 6:05 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Mon, Dec 6, 2021 at 1:00 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > On Mon, Dec 6, 2021 at 3:55 PM houzj.fnst@fujitsu.com\n> > > <houzj.fnst@fujitsu.com> wrote:\n> > > > After thinking more on this. I find there might be some usage about adding both\n> > > > child and parent to the publication.\n> > > >\n> > > > For the future feature publication row filter(and maybe column filter), It\n> > > > could be useful for user to adding child and parent with different filter\n> > > > expression. If pubviaroot=true, user can expect the parent's filter take\n> > > > effect, If pubviaroot=false, they can expect the child's filter take effect.\n> > > >\n> > > > If we disallow adding both child and parent to publication, it could be harder\n> > > > for these features to implement.\n> > >\n> > > It's possible that one may want such a feature, and yes, outright\n> > > preventing adding both the parent and a partition to a publication\n> > > would perhaps make it harder. Mind you though that any such feature\n> > > would need to fix the current implementation anyway to make the\n> > > publication-behavior-related catalog entries for child tables, which\n> > > we currently don't, unless children are added explicitly.\n> > >\n> > > That said, we should also be careful in implementing new features that\n> > > offer a finer per-partition granularity, because the fact that we do\n> > > currently offer that for many of the existing old features such as\n> > > constraints, indexes, etc. limits the scalability of our architecture.\n> > > That point may be off-topic for the thread, but something to consider,\n> > > or maybe I need to take a look at the other patch to see if my\n> > > concerns are addressable / have been addressed.\n> > >\n> >\n> > Your concern is not very clear to me. Can you be more specific in what\n> > kind of problem you see with such a design for row filtering?\n>\n> I guess my characterization of it is still vague but the problem I see\n> is that we'll be adding one more feature that allows manipulating\n> partitions directly, which means more system catalogs than there\n> already are that know about partitions as objects. I remember being\n> deliberate about avoiding pg_publication_rel entries for all\n> partitions when a partitioned table is added to a publication for that\n> very reason.\n>\n\nThe same will be true even after this new feature.\n\n> Is covering the use case of defining different filters for each\n> partition important for developing that feature?\n>\n\nUsers can only define the filters for a relation only when they want\nto add a particular partition to the publication. Say, if only root\ntable is added to the publication (and pubviaroot is true) then we\ndon't need row filters to be defined for partitions. Basically, there\nwill be only one additional column in pg_publication_rel for the row\nfilter.\n\n> I feel that defining\n> one for the root table and just using that for all partitions (albeit\n> translating as needed to account for possibly different attribute\n> numbers) would be enough for most users' use cases.\n>\n\nSure, but what if the user just wants to publish only one of the\npartitions? I think in that case along with adding the partition to\npublication, we need to allow specifying row filter.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 7 Dec 2021 09:31:08 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Tue, Dec 7, 2021 at 12:45 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Mon, Dec 6, 2021 at 8:04 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > So IIUC the scenario of concern is when a table to be attached as a\n> > partition is in a schema that's present in pg_publication_namespace.\n> > The only way to stop it from being published is to move it to another\n> > schema that is not published using that publication.\n> >\n> > I think I misunderstood how the IN SCHEMA feature works.\n> > Specifically, I didn't know that one can add a partitioned table to\n> > the same publication (or any table other than those in a particular\n> > schema for that matter). Then the attached partition would still be\n> > present in the publication by way of being part of the schema that is\n> > present in the publication, along with the partitioned table that is\n> > added separately.\n>\n> Right.\n>\n> > Yes, my proposal in its current form can't prevent that kind of duplication.\n>\n> I am not sure how to proceed here. I feel it is better to go ahead\n> with the fix Hou-san proposed here and in another email [1] to fix the\n> know issues, especially because the issue discussed in [1] needs to be\n> back-patched.\n>\n> We can evaluate your proposal separately for HEAD. What\n> do you think?\n\nYeah, maybe. Though given the direction that the row-filters patch\nset is taking in allowing to definw filters on the individual\npartitions, I am not sure if I should be pushing the approach to\ndisallow partitions from being added to publications explicitly\nalongside their parent tables. I'll try to take a look at that thread\nto be sure if that's actually the case.\n\nAlso, for the purposes of the problems that Greg and Hou-san have\ndiscovered, I have no objection with applying Hou-san's patches.\nThose seem better for the back-patching anyway.\n\nThank you.\n\n-- \nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 7 Dec 2021 14:07:56 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Tue, Dec 7, 2021 at 1:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Mon, Dec 6, 2021 at 8:59 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > On Mon, Dec 6, 2021 at 6:05 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > Your concern is not very clear to me. Can you be more specific in what\n> > > kind of problem you see with such a design for row filtering?\n> >\n> > I guess my characterization of it is still vague but the problem I see\n> > is that we'll be adding one more feature that allows manipulating\n> > partitions directly, which means more system catalogs than there\n> > already are that know about partitions as objects. I remember being\n> > deliberate about avoiding pg_publication_rel entries for all\n> > partitions when a partitioned table is added to a publication for that\n> > very reason.\n>\n> The same will be true even after this new feature.\n\nOh, interesting.\n\n> > Is covering the use case of defining different filters for each\n> > partition important for developing that feature?\n> >\n>\n> Users can only define the filters for a relation only when they want\n> to add a particular partition to the publication. Say, if only root\n> table is added to the publication (and pubviaroot is true) then we\n> don't need row filters to be defined for partitions. Basically, there\n> will be only one additional column in pg_publication_rel for the row\n> filter.\n\nOkay, so you are talking about allowing setting filters on a partition\nthat is added *explicitly* to a publication. I guess that's the only\nway to do such a thing given the syntax of FOR TABLE / ADD TABLE.\nFilters specified on a partitioned table that is added to a\npublication apply equally to all partitions, depending on the\npublish_via_partition_root configuration. So if a partition is\nexplicitly present in a publication with a filter defined on it, I\nsuppose following are the possible scenarios:\n\n* publish_via_partition_root=true\n\n1. If the root table is not present in the publication (perhaps\nunlikely), use the partition's filter because there's nothing else to\nconsider.\n\n2. If the root table is present in the publication (very likely), use\nthe root's filter, ignoring/overriding the partition's own if any.\n\n* publish_via_partition_root=false\n\n1. If the root table is not present in the publication, use the\npartition's filter because there's nothing else to consider.\n\n2. If the root table is present in the publication, use the\npartition's own filter if present, else root's.\n\n> > I feel that defining\n> > one for the root table and just using that for all partitions (albeit\n> > translating as needed to account for possibly different attribute\n> > numbers) would be enough for most users' use cases.\n> >\n>\n> Sure, but what if the user just wants to publish only one of the\n> partitions? I think in that case along with adding the partition to\n> publication, we need to allow specifying row filter.\n\nYeah, I understand now that we're talking about partitions that are\nadded explicitly.\n\nThanks for your patience.\n\n-- \nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 7 Dec 2021 14:21:54 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Tue, Dec 7, 2021 at 10:52 AM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Tue, Dec 7, 2021 at 1:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Mon, Dec 6, 2021 at 8:59 PM Amit Langote <amitlangote09@gmail.com> wrote:\n> > > On Mon, Dec 6, 2021 at 6:05 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > > Your concern is not very clear to me. Can you be more specific in what\n> > > > kind of problem you see with such a design for row filtering?\n> > >\n> > > I guess my characterization of it is still vague but the problem I see\n> > > is that we'll be adding one more feature that allows manipulating\n> > > partitions directly, which means more system catalogs than there\n> > > already are that know about partitions as objects. I remember being\n> > > deliberate about avoiding pg_publication_rel entries for all\n> > > partitions when a partitioned table is added to a publication for that\n> > > very reason.\n> >\n> > The same will be true even after this new feature.\n>\n> Oh, interesting.\n>\n> > > Is covering the use case of defining different filters for each\n> > > partition important for developing that feature?\n> > >\n> >\n> > Users can only define the filters for a relation only when they want\n> > to add a particular partition to the publication. Say, if only root\n> > table is added to the publication (and pubviaroot is true) then we\n> > don't need row filters to be defined for partitions. Basically, there\n> > will be only one additional column in pg_publication_rel for the row\n> > filter.\n>\n> Okay, so you are talking about allowing setting filters on a partition\n> that is added *explicitly* to a publication. I guess that's the only\n> way to do such a thing given the syntax of FOR TABLE / ADD TABLE.\n> Filters specified on a partitioned table that is added to a\n> publication apply equally to all partitions, depending on the\n> publish_via_partition_root configuration.\n>\n\nYes.\n\n> So if a partition is\n> explicitly present in a publication with a filter defined on it, I\n> suppose following are the possible scenarios:\n>\n> * publish_via_partition_root=true\n>\n> 1. If the root table is not present in the publication (perhaps\n> unlikely), use the partition's filter because there's nothing else to\n> consider.\n>\n> 2. If the root table is present in the publication (very likely), use\n> the root's filter, ignoring/overriding the partition's own if any.\n>\n> * publish_via_partition_root=false\n>\n> 1. If the root table is not present in the publication, use the\n> partition's filter because there's nothing else to consider.\n>\n> 2. If the root table is present in the publication, use the\n> partition's own filter if present, else root's.\n>\n\nI have not tested/checked each of these scenarios individually but I\nthink it is something like, if publish_via_partition_root is false\nthen we will always try to use partitions row filter and if it is not\nthere then we don't use any filter. Similarly, if\npublish_via_partition_root is true, then we always try to use a\npartitioned table row filter and if it is not there (either because\nthe partitioned table is not part of the publication or because there\nis no filter defined for it) then we don't use any filter.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 7 Dec 2021 11:50:26 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Tue, Dec 7, 2021 at 3:20 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Tue, Dec 7, 2021 at 10:52 AM Amit Langote <amitlangote09@gmail.com> wrote:\n> > So if a partition is\n> > explicitly present in a publication with a filter defined on it, I\n> > suppose following are the possible scenarios:\n> >\n> > * publish_via_partition_root=true\n> >\n> > 1. If the root table is not present in the publication (perhaps\n> > unlikely), use the partition's filter because there's nothing else to\n> > consider.\n> >\n> > 2. If the root table is present in the publication (very likely), use\n> > the root's filter, ignoring/overriding the partition's own if any.\n> >\n> > * publish_via_partition_root=false\n> >\n> > 1. If the root table is not present in the publication, use the\n> > partition's filter because there's nothing else to consider.\n> >\n> > 2. If the root table is present in the publication, use the\n> > partition's own filter if present, else root's.\n>\n> I have not tested/checked each of these scenarios individually but I\n> think it is something like, if publish_via_partition_root is false\n> then we will always try to use partitions row filter and if it is not\n> there then we don't use any filter. Similarly, if\n> publish_via_partition_root is true, then we always try to use a\n> partitioned table row filter and if it is not there (either because\n> the partitioned table is not part of the publication or because there\n> is no filter defined for it) then we don't use any filter.\n\nOkay, thanks for the explanation.\n\n\n--\nAmit Langote\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 7 Dec 2021 21:56:55 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\n>\n> On Thu, Dec 2, 2021 at 9:44 houzj.fnst@fujitsu.com\n> <houzj.fnst@fujitsu.com> wrote:\n> > On Wed, Dec 1, 2021 3:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > > > On Mon, Nov 22, 2021 at 12:55 PM Amit Langote\n> > > > > <amitlangote09@gmail.com> wrote:\n> > > > > > On second thought, I agree that de-duplicating partitions from\n> > > > > > this view is an improvement.\n> > > > >\n> > > > > Fair enough. Hou-San, Can you please submit the updated patch after\n> > > > > fixing any pending comments including the test case?\n> > > >\n> > > > Attach the updated patch which address the comments so far.\n> > > >\n> > > > The patch only adds a testcase in publication.sql because the\n> > > > duplicate output doesn't cause unexpected behavior in pub-sub test.\n> > >\n> > > Thanks, the patch looks good to me. I have slightly changed the commit\n> > > message in the attached. I would like to commit this only in HEAD as there is no\n> > > user complaint about this and it might not stop any user's service unless it relies\n> > > on this view's output for the initial table synchronization.\n>\n> The patch looks good to me too in that it gets the job done.\n>\n\nPushed.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 8 Dec 2021 15:46:50 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Sat, Nov 20, 2021 7:31 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com>\r\n> wrote:\r\n> >\r\n> > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com>\r\n> wrote:\r\n> > >\r\n> > > The problematic case is attaching the partition *after* the\r\n> > > subscriber has already marked the root parent as synced and/or ready\r\n> > > for replication. Refreshing the subscription doesn't help it\r\n> > > discover the newly attached partition, because a\r\n> > > publish_via_partition_root only ever tells about the root parent,\r\n> > > which would be already synced, so the subscriber would think there's\r\n> > > nothing to copy.\r\n> > >\r\n> >\r\n> > Okay, I see this could be a problem but I haven't tried to reproduce it.\r\n> \r\n> One more thing you mentioned is that the initial sync won't work after refresh\r\n> but later changes will be replicated but I noticed that later changes also don't\r\n> get streamed till we restart the subscriber server.\r\n> I am not sure but we might not be invalidating apply workers cache due to\r\n> which it didn't notice the same.\r\n\r\nI investigated this bug recently, and I think the reason is that when receiving\r\nrelcache invalidation message, the callback function[1] in walsender only reset\r\nthe schema sent status while it doesn't reset the replicate_valid flag. So, it\r\nwon’t rebuild the publication actions of the relation.\r\n\r\n[1]\r\nstatic void\r\nrel_sync_cache_relation_cb(Datum arg, Oid relid)\r\n...\r\n\t/*\r\n\t * Reset schema sent status as the relation definition may have changed.\r\n\t * Also free any objects that depended on the earlier definition.\r\n\t */\r\n\tif (entry != NULL)\r\n\t{\r\n\t\tentry->schema_sent = false;\r\n\t\tlist_free(entry->streamed_txns);\r\n...\r\n\r\nAlso, when you DETACH a partition, the publication won’t be rebuilt too because\r\nof the same reason. Which could cause unexpected behavior if we modify the\r\ndetached table's data . And the bug happens regardless of whether pubviaroot is\r\nset or not.\r\n\r\nFor the fix:\r\n\r\nI think if we also reset replicate_valid flag in rel_sync_cache_relation_cb,\r\nthen the bug can be fixed. I have a bit hesitation about this approach, because\r\nit could increase the frequency of invalidating and rebuilding the publication\r\naction. But I haven't produced some other better approaches.\r\n\r\nAnother possibility could be that we add a syscache callback function for\r\npg_inherits table, but we currently don't have syscache for pg_inherits. We\r\nmight need to add the cache pg_inherits first which doesn't seems better\r\nthan the above approach.\r\n\r\nWhat do you think ?\r\n\r\nAttach an initial patch which reset the replicate_valid flag in\r\nrel_sync_cache_relation_cb and add some reproduction tests in 013_partition.pl.\r\n\r\nBest regards,\r\nHou zj","msg_date":"Tue, 14 Dec 2021 07:41:49 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Tuesday, December 14, 2021 3:42 PM houzj.fnst@fujitsu.com <houzj.fnst@fujitsu.com> wrote:\r\n> \r\n> On Sat, Nov 20, 2021 7:31 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> > On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com>\r\n> > wrote:\r\n> > >\r\n> > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote\r\n> > > <amitlangote09@gmail.com>\r\n> > wrote:\r\n> > > >\r\n> > > > The problematic case is attaching the partition *after* the\r\n> > > > subscriber has already marked the root parent as synced and/or\r\n> > > > ready for replication. Refreshing the subscription doesn't help\r\n> > > > it discover the newly attached partition, because a\r\n> > > > publish_via_partition_root only ever tells about the root parent,\r\n> > > > which would be already synced, so the subscriber would think\r\n> > > > there's nothing to copy.\r\n> > > >\r\n> > >\r\n> > > Okay, I see this could be a problem but I haven't tried to reproduce it.\r\n> >\r\n> > One more thing you mentioned is that the initial sync won't work after\r\n> > refresh but later changes will be replicated but I noticed that later\r\n> > changes also don't get streamed till we restart the subscriber server.\r\n> > I am not sure but we might not be invalidating apply workers cache due\r\n> > to which it didn't notice the same.\r\n> \r\n> I investigated this bug recently, and I think the reason is that when receiving\r\n> relcache invalidation message, the callback function[1] in walsender only reset\r\n> the schema sent status while it doesn't reset the replicate_valid flag. So, it\r\n> won’t rebuild the publication actions of the relation.\r\n> \r\n> [1]\r\n> static void\r\n> rel_sync_cache_relation_cb(Datum arg, Oid relid) ...\r\n> \t/*\r\n> \t * Reset schema sent status as the relation definition may have\r\n> changed.\r\n> \t * Also free any objects that depended on the earlier definition.\r\n> \t */\r\n> \tif (entry != NULL)\r\n> \t{\r\n> \t\tentry->schema_sent = false;\r\n> \t\tlist_free(entry->streamed_txns);\r\n> ...\r\n> \r\n> Also, when you DETACH a partition, the publication won’t be rebuilt too\r\n> because of the same reason. Which could cause unexpected behavior if we\r\n> modify the detached table's data . And the bug happens regardless of whether\r\n> pubviaroot is set or not.\r\n> \r\n> For the fix:\r\n> \r\n> I think if we also reset replicate_valid flag in rel_sync_cache_relation_cb, then\r\n> the bug can be fixed. I have a bit hesitation about this approach, because it\r\n> could increase the frequency of invalidating and rebuilding the publication\r\n> action. But I haven't produced some other better approaches.\r\n> \r\n\r\nI have confirmed that the bug of ATTACH PARTITION has been fixed due to recent\r\ncommit 7f481b8. Currently, we always invalidate the RelationSyncCache when\r\nattaching a partition, so the pubactions of the newly attached partition will\r\nbe rebuilt correctly.\r\n\r\nBest regards,\r\nHou zj\r\n\r\n\r\n","msg_date":"Mon, 11 Apr 2022 09:04:28 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On 2022-Apr-11, houzj.fnst@fujitsu.com wrote:\n\n> I have confirmed that the bug of ATTACH PARTITION has been fixed due to recent\n> commit 7f481b8. Currently, we always invalidate the RelationSyncCache when\n> attaching a partition, so the pubactions of the newly attached partition will\n> be rebuilt correctly.\n\nHm, this commit was not backpatched -- it is only in master. I just\nchecked that pg14 behaves as you describe in the opening email of this\nthread; did we decide that we don't want to change behavior of stable\nbranches?\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Thu, 14 Apr 2022 18:45:32 +0200","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: pg_get_publication_tables() output duplicate relid"},{"msg_contents":"On Friday, April 15, 2022 12:46 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\r\n> \r\n> On 2022-Apr-11, houzj.fnst@fujitsu.com wrote:\r\n> \r\n> > I have confirmed that the bug of ATTACH PARTITION has been fixed due to\r\n> recent\r\n> > commit 7f481b8. Currently, we always invalidate the RelationSyncCache when\r\n> > attaching a partition, so the pubactions of the newly attached partition will\r\n> > be rebuilt correctly.\r\n> \r\n> Hm, this commit was not backpatched -- it is only in master. I just\r\n> checked that pg14 behaves as you describe in the opening email of this\r\n> thread; did we decide that we don't want to change behavior of stable\r\n> branches?\r\n\r\nHi,\r\n\r\nDo you mean the original bug: \"pg_get_publication_tables function will output\r\nduplicate partition relid when adding both child and parent table to the\r\npublication(pubviaroot = false). \" ?\r\n\r\nWe decided to commit the fix for it only in HEAD as there is no user complaint\r\nabout this and it might not stop any user's service unless it relies on this\r\nview's output for the initial table synchronization.\r\n\r\nOr do you mean another one mentioned here which is \"the changes on newly\r\nattached partition won't be replicated\" ? For this one, I think it's fine to\r\nbackpatch the fix and I can prepare the patch for back branches if needed. BTW,\r\nif we want to backpatch this fix we might need to keep the order of\r\nRelationSyncEntry struct member unchanged in back branches.\r\n\r\nBest regards,\r\nHou zj\r\n","msg_date":"Mon, 18 Apr 2022 06:25:49 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: pg_get_publication_tables() output duplicate relid"}],"string":"[\n {\n \"msg_contents\": \"Hi hackers,\\n\\nIn another thread[1], we found the pg_get_publication_tables function will output\\nduplicate partition relid when adding both child and parent table to the\\npublication(pubviaroot = false).\\n\\nExample:\\ncreate table tbl1 (a int) partition by range (a);\\ncreate table tbl1_part1 partition of tbl1 for values from (1) to (10);\\ncreate table tbl1_part2 partition of tbl1 for values from (10) to (20);\\ncreate publication pub for table\\ntbl1, tbl1_part1 with (publish_via_partition_root=false);\\n\\nselect * from pg_get_publication_tables('pub');\\n relid\\n-------\\n 16387\\n 16390\\n 16387\\n\\nThe reason of the duplicate output is that:\\n\\npg_get_publication_tables invokes function GetPublicationRelations internally.\\nIn GetPublicationRelations(), it will add the oid of partition 'tbl1_part1'\\ntwice. First time from extracting partitions from the specified parent table\\n'tbl1', second time from the explicitly specified partition 'tbl1_part1'.\\n\\n\\nI am not sure is this behavior expected as it seems natural for\\npg_get_publication_tables to return duplicate-free relid list. OTOH, there\\nseems no harm for the current behavior(duplicate output), it doesn't affect the\\ninitial sync and change replication when using logical replication.\\n\\nPersonally, I think it might be better to make the output of\\npg_get_publication_tables duplicate-free, because the change happened on each\\noutput relid will only be replicated once. So, it seems more consistent to\\noutput each relid only once.\\n\\nThoughts ?\\n\\n(Attach a patch which make the output duplicate-free)\\n\\n[1] https://www.postgresql.org/message-id/CAJcOf-eURu03QNmD%3D37PtsxuNW4nBGN3G_FdRMBx_tpkeyzDUw%40mail.gmail.com\\n\\nBest regards,\\nHou zj\",\n \"msg_date\": \"Mon, 15 Nov 2021 08:17:49 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n>\\n> Hi hackers,\\n>\\n> In another thread[1], we found the pg_get_publication_tables function will output\\n> duplicate partition relid when adding both child and parent table to the\\n> publication(pubviaroot = false).\\n>\\n> Example:\\n> create table tbl1 (a int) partition by range (a);\\n> create table tbl1_part1 partition of tbl1 for values from (1) to (10);\\n> create table tbl1_part2 partition of tbl1 for values from (10) to (20);\\n> create publication pub for table\\n> tbl1, tbl1_part1 with (publish_via_partition_root=false);\\n>\\n> select * from pg_get_publication_tables('pub');\\n> relid\\n> -------\\n> 16387\\n> 16390\\n> 16387\\n>\\n> The reason of the duplicate output is that:\\n>\\n> pg_get_publication_tables invokes function GetPublicationRelations internally.\\n> In GetPublicationRelations(), it will add the oid of partition 'tbl1_part1'\\n> twice. First time from extracting partitions from the specified parent table\\n> 'tbl1', second time from the explicitly specified partition 'tbl1_part1'.\\n>\\n>\\n> I am not sure is this behavior expected as it seems natural for\\n> pg_get_publication_tables to return duplicate-free relid list. OTOH, there\\n> seems no harm for the current behavior(duplicate output), it doesn't affect the\\n> initial sync and change replication when using logical replication.\\n>\\n> Personally, I think it might be better to make the output of\\n> pg_get_publication_tables duplicate-free, because the change happened on each\\n> output relid will only be replicated once. So, it seems more consistent to\\n> output each relid only once.\\n>\\n> Thoughts ?\\n>\\n> (Attach a patch which make the output duplicate-free)\\n>\\n> [1] https://www.postgresql.org/message-id/CAJcOf-eURu03QNmD%3D37PtsxuNW4nBGN3G_FdRMBx_tpkeyzDUw%40mail.gmail.com\\n\\nThe users can always specify the distinct clause, see [1]. I don't see\\nany problem with the existing way. If required you can specify in the\\nview pg_publication_tables documentation that \\\"This view returns\\nmultiple rows for the same relation, if the relation is child to a\\nparent partition table and if both the parent and child are specified\\nin the publication\\\" or something similart.\\n\\nIf at all, the pg_get_publication_tables() is supposed to give unique\\noutputs, let's fix it and it is a good idea to specify that the view\\npg_publication_tables returns unique rows even though child and parent\\npartition tables are specified in the publication.\\n\\nI have fdw comments on the patch:\\n1) Why to do result = list_concat_unique_oid(result, relids); within\\nthe while loop which can make the while loop O(n*n*n)\\ncomplexity(list_concat_unique_oid is O(n*n)? Can't you use\\nlist_append_unique, of course this can also be costlier? Let's avoid\\nadding anything to the while loop and use\\nlist_sort+list_deduplicate_oid (qsort(O(nlogn)+O(n))\\n\\n /* Now sort and de-duplicate the result list */\\n list_sort(result, list_oid_cmp);\\n list_deduplicate_oid(result);\\n\\n while (HeapTupleIsValid(tup = systable_getnext(scan)))\\n {\\n Form_pg_publication_rel pubrel;\\n+ List *relids = NIL;\\n\\n pubrel = (Form_pg_publication_rel) GETSTRUCT(tup);\\n- result = GetPubPartitionOptionRelations(result, pub_partopt,\\n+ relids = GetPubPartitionOptionRelations(relids, pub_partopt,\\n pubrel->prrelid);\\n+\\n+ result = list_concat_unique_oid(result, relids);\\n }\\n\\n2) Are you sure you want GetPublicationRelations to be returning the\\nunique relations? I'm just thinking why can't you just do\\nlist_sort+list_deduplicate_oid in the caller? I think this makes the\\nfunction more restrictive. If required, you can pass a boolean flag\\n(bool give_unique and specify in the function comments and if passed\\ntrue do list_sort+list_deduplicate_oid at the end of\\nGetPublicationRelations ).\\n\\n[1]\\npostgres=# select * from pg_get_publication_tables('pub');\\n relid\\n-------\\n 16387\\n 16390\\n 16387\\n(3 rows)\\n\\npostgres=# select distinct * from pg_get_publication_tables('pub');\\n relid\\n-------\\n 16387\\n 16390\\n(2 rows)\\n\\npostgres=# select * from pg_publication_tables;\\n pubname | schemaname | tablename\\n---------+------------+------------\\n pub | public | tbl1_part1\\n pub | public | tbl1_part2\\n pub | public | tbl1_part1\\n(3 rows)\\n\\npostgres=# select distinct * from pg_publication_tables;\\n pubname | schemaname | tablename\\n---------+------------+------------\\n pub | public | tbl1_part1\\n pub | public | tbl1_part2\\n(2 rows)\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 15:46:48 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"> On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\\n> <houzj.fnst@fujitsu.com> wrote:\\n\\n> > create table tbl1 (a int) partition by range (a);\\n> > create table tbl1_part1 partition of tbl1 for values from (1) to (10);\\n> > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\\n> > create publication pub for table\\n> > tbl1, tbl1_part1 with (publish_via_partition_root=false);\\n\\nIn the name of consistency, I think this situation should be an error --\\nI mean, if we detect that the user is trying to add both the partitioned\\ntable *and* its partition, then all manner of things are possibly going\\nto go wrong in some way, so my inclination is to avoid it altogether.\\n\\nIs there any reason to allow that?\\n\\nIf we do that, then we have to be careful with later ALTER PUBLICATION\\ntoo: adding a partition is not allowed if its partitioned table is\\nthere, and adding a partitioned table should behave in some sensible way\\nif one of its partitions is there (either removing the partition at the\\nsame time, or outright rejecting the operation.)\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Mon, 15 Nov 2021 10:42:14 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 6:17 PM Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote:\\r\\n> On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\\r\\n> <houzj.fnst@fujitsu.com> wrote:\\r\\n> >\\r\\n> > Hi hackers,\\r\\n> >\\r\\n> > In another thread[1], we found the pg_get_publication_tables function\\r\\n> > will output duplicate partition relid when adding both child and\\r\\n> > parent table to the publication(pubviaroot = false).\\r\\n> >\\r\\n> > Example:\\r\\n> > create table tbl1 (a int) partition by range (a); create table\\r\\n> > tbl1_part1 partition of tbl1 for values from (1) to (10); create table\\r\\n> > tbl1_part2 partition of tbl1 for values from (10) to (20); create\\r\\n> > publication pub for table tbl1, tbl1_part1 with\\r\\n> > (publish_via_partition_root=false);\\r\\n> >\\r\\n> > select * from pg_get_publication_tables('pub'); relid\\r\\n> > -------\\r\\n> > 16387\\r\\n> > 16390\\r\\n> > 16387\\r\\n> >\\r\\n> > The reason of the duplicate output is that:\\r\\n> >\\r\\n> > pg_get_publication_tables invokes function GetPublicationRelations\\r\\n> internally.\\r\\n> > In GetPublicationRelations(), it will add the oid of partition 'tbl1_part1'\\r\\n> > twice. First time from extracting partitions from the specified parent\\r\\n> > table 'tbl1', second time from the explicitly specified partition 'tbl1_part1'.\\r\\n> >\\r\\n> >\\r\\n> > I am not sure is this behavior expected as it seems natural for\\r\\n> > pg_get_publication_tables to return duplicate-free relid list. OTOH,\\r\\n> > there seems no harm for the current behavior(duplicate output), it\\r\\n> > doesn't affect the initial sync and change replication when using logical\\r\\n> replication.\\r\\n> >\\r\\n> > Personally, I think it might be better to make the output of\\r\\n> > pg_get_publication_tables duplicate-free, because the change happened\\r\\n> > on each output relid will only be replicated once. So, it seems more\\r\\n> > consistent to output each relid only once.\\r\\n> >\\r\\n> > Thoughts ?\\r\\n> >\\r\\n> > (Attach a patch which make the output duplicate-free)\\r\\n> >\\r\\n> > [1]\\r\\n> >\\r\\n> https://www.postgresql.org/message-id/CAJcOf-eURu03QNmD%3D37Ptsxu\\r\\n> NW4nB\\r\\n> > GN3G_FdRMBx_tpkeyzDUw%40mail.gmail.com\\r\\n> \\r\\n> The users can always specify the distinct clause, see [1]. I don't see any problem\\r\\n> with the existing way. If required you can specify in the view\\r\\n> pg_publication_tables documentation that \\\"This view returns multiple rows for\\r\\n> the same relation, if the relation is child to a parent partition table and if both\\r\\n> the parent and child are specified in the publication\\\" or something similart.\\r\\n\\r\\nThanks for the response.\\r\\nYes, I agreed, as I said there's no harm for the current behavior. If most people don't think\\r\\nIt's worth changing the behavior, I am fine with that.\\r\\n\\r\\n> If at all, the pg_get_publication_tables() is supposed to give unique outputs,\\r\\n> let's fix it and it is a good idea to specify that the view pg_publication_tables\\r\\n> returns unique rows even though child and parent partition tables are specified\\r\\n> in the publication.\\r\\n> \\r\\n> I have fdw comments on the patch:\\r\\n> Let's avoid adding anything to the while\\r\\n> loop and use list_sort+list_deduplicate_oid (qsort(O(nlogn)+O(n))\\r\\n> \\r\\n\\r\\nThanks for the comments, I will address this if we finally decide to fix it.\\r\\n\\r\\n> \\r\\n> 2) Are you sure you want GetPublicationRelations to be returning the unique\\r\\n> relations? I'm just thinking why can't you just do list_sort+list_deduplicate_oid\\r\\n> in the caller? I think this makes the function more restrictive. If required, you\\r\\n> can pass a boolean flag (bool give_unique and specify in the function\\r\\n> comments and if passed true do list_sort+list_deduplicate_oid at the end of\\r\\n> GetPublicationRelations ).\\r\\n\\r\\nI change the GetPublicationRelations because I found no caller need the\\r\\nduplicate oid, and de-duplicate oid in GetPublicationRelations could reduce\\r\\nsome code change. I think if we need the duplicate oid in the future, we can\\r\\nadd the flag as you suggested.\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 01:50:45 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 9:42 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\r\\n> > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\\r\\n> > <houzj.fnst@fujitsu.com> wrote:\\r\\n> \\r\\n> > > create table tbl1 (a int) partition by range (a); create table\\r\\n> > > tbl1_part1 partition of tbl1 for values from (1) to (10); create\\r\\n> > > table tbl1_part2 partition of tbl1 for values from (10) to (20);\\r\\n> > > create publication pub for table tbl1, tbl1_part1 with\\r\\n> > > (publish_via_partition_root=false);\\r\\n> \\r\\n> In the name of consistency, I think this situation should be an error -- I mean, if\\r\\n> we detect that the user is trying to add both the partitioned table *and* its\\r\\n> partition, then all manner of things are possibly going to go wrong in some way,\\r\\n> so my inclination is to avoid it altogether.\\r\\n>\\r\\n> Is there any reason to allow that?\\r\\n>\\r\\n> If we do that, then we have to be careful with later ALTER PUBLICATION\\r\\n> too: adding a partition is not allowed if its partitioned table is there, and adding\\r\\n> a partitioned table should behave in some sensible way if one of its partitions is\\r\\n> there (either removing the partition at the same time, or outright rejecting the\\r\\n> operation.)\\r\\n\\r\\nThanks for the response.\\r\\n\\r\\nIf we decide to disallow this case, we seem need to handle some other cases as\\r\\nwell, for example: We might also need additional check when ATTACH a partition,\\r\\nbecause the partition's parent table could already be published in the same\\r\\npublication as the partition. During the check we might also need to lock all\\r\\nthe parent tables to handle concurrently change which looks complicated to me :(\\r\\nI am not sure is it worth adding these lock and check. Maybe we can leave the current\\r\\nbehavior as is ?\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 01:51:06 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 7:12 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>\\n> > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\\n> > <houzj.fnst@fujitsu.com> wrote:\\n>\\n> > > create table tbl1 (a int) partition by range (a);\\n> > > create table tbl1_part1 partition of tbl1 for values from (1) to (10);\\n> > > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\\n> > > create publication pub for table\\n> > > tbl1, tbl1_part1 with (publish_via_partition_root=false);\\n>\\n> In the name of consistency, I think this situation should be an error --\\n> I mean, if we detect that the user is trying to add both the partitioned\\n> table *and* its partition, then all manner of things are possibly going\\n> to go wrong in some way, so my inclination is to avoid it altogether.\\n>\\n> Is there any reason to allow that?\\n>\\n\\nI think it could provide flexibility to users to later change\\n\\\"publish_via_partition_root\\\" option. Because when that option is\\nfalse, we use individual partitions schema to send changes and when it\\nis true, we use root table's schema to send changes. Added Amit L. to\\nknow if he has any thoughts on this matter as he was the author of\\nthis work?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 18:57:05 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n>\\n> On Mon, Nov 15, 2021 9:42 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> > > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\\n> > > <houzj.fnst@fujitsu.com> wrote:\\n> >\\n> > > > create table tbl1 (a int) partition by range (a); create table\\n> > > > tbl1_part1 partition of tbl1 for values from (1) to (10); create\\n> > > > table tbl1_part2 partition of tbl1 for values from (10) to (20);\\n> > > > create publication pub for table tbl1, tbl1_part1 with\\n> > > > (publish_via_partition_root=false);\\n> >\\n> > In the name of consistency, I think this situation should be an error -- I mean, if\\n> > we detect that the user is trying to add both the partitioned table *and* its\\n> > partition, then all manner of things are possibly going to go wrong in some way,\\n> > so my inclination is to avoid it altogether.\\n> >\\n> > Is there any reason to allow that?\\n> >\\n> > If we do that, then we have to be careful with later ALTER PUBLICATION\\n> > too: adding a partition is not allowed if its partitioned table is there, and adding\\n> > a partitioned table should behave in some sensible way if one of its partitions is\\n> > there (either removing the partition at the same time, or outright rejecting the\\n> > operation.)\\n>\\n> Thanks for the response.\\n>\\n> If we decide to disallow this case, we seem need to handle some other cases as\\n> well, for example: We might also need additional check when ATTACH a partition,\\n> because the partition's parent table could already be published in the same\\n> publication as the partition.\\n>\\n\\nWhat kind of additional checks you are envisioning and why?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 08:17:06 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 10:47 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com\\r\\n> <houzj.fnst@fujitsu.com> wrote:\\r\\n> >\\r\\n> > On Mon, Nov 15, 2021 9:42 PM Alvaro Herrera <alvherre@alvh.no-ip.org>\\r\\n> wrote:\\r\\n> > > > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\\r\\n> > > > <houzj.fnst@fujitsu.com> wrote:\\r\\n> > >\\r\\n> > > > > create table tbl1 (a int) partition by range (a); create table\\r\\n> > > > > tbl1_part1 partition of tbl1 for values from (1) to (10); create\\r\\n> > > > > table tbl1_part2 partition of tbl1 for values from (10) to (20);\\r\\n> > > > > create publication pub for table tbl1, tbl1_part1 with\\r\\n> > > > > (publish_via_partition_root=false);\\r\\n> > >\\r\\n> > > In the name of consistency, I think this situation should be an\\r\\n> > > error -- I mean, if we detect that the user is trying to add both\\r\\n> > > the partitioned table *and* its partition, then all manner of things\\r\\n> > > are possibly going to go wrong in some way, so my inclination is to avoid it\\r\\n> > > altogether.\\r\\n> > >\\r\\n> > > Is there any reason to allow that?\\r\\n> > >\\r\\n> > > If we do that, then we have to be careful with later ALTER\\r\\n> > > PUBLICATION\\r\\n> > > too: adding a partition is not allowed if its partitioned table is\\r\\n> > > there, and adding a partitioned table should behave in some sensible\\r\\n> > > way if one of its partitions is there (either removing the partition\\r\\n> > > at the same time, or outright rejecting the\\r\\n> > > operation.)\\r\\n> >\\r\\n> > Thanks for the response.\\r\\n> >\\r\\n> > If we decide to disallow this case, we seem need to handle some other\\r\\n> > cases as well, for example: We might also need additional check when\\r\\n> > ATTACH a partition, because the partition's parent table could already\\r\\n> > be published in the same publication as the partition.\\r\\n> >\\r\\n> \\r\\n> What kind of additional checks you are envisioning and why?\\r\\n\\r\\nFor example:\\r\\n\\r\\ncreate table tbl1 (a int) partition by range (a);\\r\\ncreate table tbl1_part1 (like tbl1);\\r\\ncreate table tbl1_part2 partition of tbl1 for values from (10) to (20);\\r\\ncreate publication pub for table\\r\\ntbl1, tbl1_part1 with (publish_via_partition_root=false);\\r\\n\\r\\n--- We might need addition check here\\r\\nAlter table tbl1 ATTACH partition tbl1_part1 for values from (1) to (10);\\r\\n\\r\\nIn the above cases, 'tbl1_part1' is not a partition of 'tb1' when creating a\\r\\npublication. After the ATTACH, 'tbl1_part1' would become a partition of 'tbl1'\\r\\nwhich seems the case we want to disallow(both parent and child table in\\r\\npublication). So, When ATTACH, I thought we might need to check all the parent\\r\\ntables to see if any parent table is in the same publication which the table to\\r\\nbe attached is also belongs to. Does it make sense ?\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 03:15:37 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 10:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Mon, Nov 15, 2021 at 7:12 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> > > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\\n> > > <houzj.fnst@fujitsu.com> wrote:\\n> >\\n> > > > create table tbl1 (a int) partition by range (a);\\n> > > > create table tbl1_part1 partition of tbl1 for values from (1) to (10);\\n> > > > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\\n> > > > create publication pub for table\\n> > > > tbl1, tbl1_part1 with (publish_via_partition_root=false);\\n> >\\n> > In the name of consistency, I think this situation should be an error --\\n> > I mean, if we detect that the user is trying to add both the partitioned\\n> > table *and* its partition, then all manner of things are possibly going\\n> > to go wrong in some way, so my inclination is to avoid it altogether.\\n> >\\n> > Is there any reason to allow that?\\n>\\n> I think it could provide flexibility to users to later change\\n> \\\"publish_via_partition_root\\\" option. Because when that option is\\n> false, we use individual partitions schema to send changes and when it\\n> is true, we use root table's schema to send changes. Added Amit L. to\\n> know if he has any thoughts on this matter as he was the author of\\n> this work?\\n\\nFWIW, I'm not sure that adding an error in this path, that is, when a\\nuser adds both a partitioned table and its partitions to a\\npublication, helps much. As for the safety of allowing it, if you\\nlook at get_rel_sync_entry(), which handles much of the complexity of\\ndetermining whether to publish a relation's changes and the schema to\\nuse when doing so, you may be able to see that a partition being added\\nduplicatively is harmless, modulo any undiscovered bugs. At least as\\nfar as the post-initial-sync replication functionality is concerned.\\n\\nWhat IS problematic is what a subscriber sees in the\\npg_publication_tables view and the problem occurs only in the initial\\nsync phase, where the partition is synced duplicatively because of\\nbeing found in the view along with the parent in this case, that is,\\nwhen publish_via_partiiton_root is true. I was saying on the other\\nthread [1] that we should leave it up to the subscriber to decide what\\nto do when the view (duplicatively) returns both the parent and the\\npartition, citing the use case that a subscriber may want to replicate\\nthe parent and the partition as independent tables. Though I now tend\\nto agree with Amit K that that may be such a meaningful and all that\\ncommon use case, and the implementation on the subscriber side would\\nhave to be unnecessarily complex.\\n\\nSo maybe it makes sense to just do what has been proposed --\\nde-duplicate partitions out of the pg_publication_tables view, unless\\nwe know of a bigger problem that requires us to hack the subscriber\\nside of things too. Actually, I came to know of one such problem\\nwhile thinking about this today: when you ATTACH a partition to a\\ntable that is present in a publish_via_partition_root=true\\npublication, it doesn't get copied via the initial sync even though\\nsubsequent replication works just fine. The reason for that is that\\nthe subscriber only syncs the partitions that are known at the time\\nwhen the parent is first synced, that too via the parent (as SELECT\\n<columns..> FROM parent), and then marks the parent as sync-done.\\nRefreshing the subscription after ATTACHing doesn't help, because the\\nsubscriber can't see any partitions to begin with. Maybe a more\\nelaborate solution is needed for this one, though I haven't figured\\nwhat it is going to look like yet.\\n\\nThanks.\\n\\n--\\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n[1] https://www.postgresql.org/message-id/CA%2BHiwqHnDHcT4OOcga9rDFyc7TvDrpN5xFH9J2pyHQo9ptvjmQ%40mail.gmail.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 15:09:14 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 12:15 PM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n> On Wed, Nov 17, 2021 10:47 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com wrote:\\n> > > If we decide to disallow this case, we seem need to handle some other\\n> > > cases as well, for example: We might also need additional check when\\n> > > ATTACH a partition, because the partition's parent table could already\\n> > > be published in the same publication as the partition.\\n> > >\\n> >\\n> > What kind of additional checks you are envisioning and why?\\n>\\n> For example:\\n>\\n> create table tbl1 (a int) partition by range (a);\\n> create table tbl1_part1 (like tbl1);\\n> create table tbl1_part2 partition of tbl1 for values from (10) to (20);\\n> create publication pub for table\\n> tbl1, tbl1_part1 with (publish_via_partition_root=false);\\n>\\n> --- We might need addition check here\\n> Alter table tbl1 ATTACH partition tbl1_part1 for values from (1) to (10);\\n>\\n> In the above cases, 'tbl1_part1' is not a partition of 'tb1' when creating a\\n> publication. After the ATTACH, 'tbl1_part1' would become a partition of 'tbl1'\\n> which seems the case we want to disallow(both parent and child table in\\n> publication). So, When ATTACH, I thought we might need to check all the parent\\n> tables to see if any parent table is in the same publication which the table to\\n> be attached is also belongs to. Does it make sense ?\\n\\nI don't think creating or attaching a partition of a table that is\\npresent in a publish_via_partition_root=false actually adds the\\npartition to pg_publication_rel, the base catalog. A partition's\\nmembership in the publication is implicit, unless of course you add it\\nto the publication explicitly, like all the examples we have been\\ndiscussing. I guess we're only arguing about the problems with the\\npg_publication_tables view, which does expand the partitioned table to\\nshow the partitions that are not otherwise not present in the base\\ncatalog.\\n\\n-- \\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 15:18:13 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 3:09 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> Though I now tend\\n> to agree with Amit K that that may be such a meaningful and all that\\n> common use case,\\n\\nOops I meant: that may NOT be such...\\n\\n-- \\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 17:49:43 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 2:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\\r\\n> On Wed, Nov 17, 2021 at 12:15 PM houzj.fnst@fujitsu.com\\r\\n> <houzj.fnst@fujitsu.com> wrote:\\r\\n> > On Wed, Nov 17, 2021 10:47 AM Amit Kapila <amit.kapila16@gmail.com>\\r\\n> wrote:\\r\\n> > > On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com wrote:\\r\\n> > > > If we decide to disallow this case, we seem need to handle some other\\r\\n> > > > cases as well, for example: We might also need additional check when\\r\\n> > > > ATTACH a partition, because the partition's parent table could already\\r\\n> > > > be published in the same publication as the partition.\\r\\n> > > >\\r\\n> > >\\r\\n> > > What kind of additional checks you are envisioning and why?\\r\\n> >\\r\\n> > For example:\\r\\n> >\\r\\n> > create table tbl1 (a int) partition by range (a);\\r\\n> > create table tbl1_part1 (like tbl1);\\r\\n> > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\\r\\n> > create publication pub for table\\r\\n> > tbl1, tbl1_part1 with (publish_via_partition_root=false);\\r\\n> >\\r\\n> > --- We might need addition check here\\r\\n> > Alter table tbl1 ATTACH partition tbl1_part1 for values from (1) to (10);\\r\\n> >\\r\\n> > In the above cases, 'tbl1_part1' is not a partition of 'tb1' when creating a\\r\\n> > publication. After the ATTACH, 'tbl1_part1' would become a partition of 'tbl1'\\r\\n> > which seems the case we want to disallow(both parent and child table in\\r\\n> > publication). So, When ATTACH, I thought we might need to check all the parent\\r\\n> > tables to see if any parent table is in the same publication which the table to\\r\\n> > be attached is also belongs to. Does it make sense ?\\r\\n> \\r\\n> I don't think creating or attaching a partition of a table that is\\r\\n> present in a publish_via_partition_root=false actually adds the\\r\\n> partition to pg_publication_rel, the base catalog. A partition's\\r\\n> membership in the publication is implicit, unless of course you add it\\r\\n> to the publication explicitly, like all the examples we have been\\r\\n> discussing. I guess we're only arguing about the problems with the\\r\\n> pg_publication_tables view, which does expand the partitioned table to\\r\\n> show the partitions that are not otherwise not present in the base\\r\\n> catalog.\\r\\n\\r\\nMaybe I didn't make it clear, I was trying to explain that it would be\\r\\ncomplicated if we want to completely disallow specifying both child and parent\\r\\ntable in the publication because of the ATTACH case I gave. In other words, I\\r\\nthink it's fine to specify both child and parent table in the publication.\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 00:33:09 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 9:33 houzj.fnst@fujitsu.com <houzj.fnst@fujitsu.com>\\nwrote:\\n\\n> On Wed, Nov 17, 2021 2:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > On Wed, Nov 17, 2021 at 12:15 PM houzj.fnst@fujitsu.com\\n> > <houzj.fnst@fujitsu.com> wrote:\\n> > > On Wed, Nov 17, 2021 10:47 AM Amit Kapila <amit.kapila16@gmail.com>\\n> > wrote:\\n> > > > On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com wrote:\\n> > > > > If we decide to disallow this case, we seem need to handle some\\n> other\\n> > > > > cases as well, for example: We might also need additional check\\n> when\\n> > > > > ATTACH a partition, because the partition's parent table could\\n> already\\n> > > > > be published in the same publication as the partition.\\n> > > > >\\n> > > >\\n> > > > What kind of additional checks you are envisioning and why?\\n> > >\\n> > > For example:\\n> > >\\n> > > create table tbl1 (a int) partition by range (a);\\n> > > create table tbl1_part1 (like tbl1);\\n> > > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\\n> > > create publication pub for table\\n> > > tbl1, tbl1_part1 with (publish_via_partition_root=false);\\n> > >\\n> > > --- We might need addition check here\\n> > > Alter table tbl1 ATTACH partition tbl1_part1 for values from (1) to\\n> (10);\\n> > >\\n> > > In the above cases, 'tbl1_part1' is not a partition of 'tb1' when\\n> creating a\\n> > > publication. After the ATTACH, 'tbl1_part1' would become a partition\\n> of 'tbl1'\\n> > > which seems the case we want to disallow(both parent and child table in\\n> > > publication). So, When ATTACH, I thought we might need to check all\\n> the parent\\n> > > tables to see if any parent table is in the same publication which the\\n> table to\\n> > > be attached is also belongs to. Does it make sense ?\\n> >\\n> > I don't think creating or attaching a partition of a table that is\\n> > present in a publish_via_partition_root=false actually adds the\\n> > partition to pg_publication_rel, the base catalog. A partition's\\n> > membership in the publication is implicit, unless of course you add it\\n> > to the publication explicitly, like all the examples we have been\\n> > discussing. I guess we're only arguing about the problems with the\\n> > pg_publication_tables view, which does expand the partitioned table to\\n> > show the partitions that are not otherwise not present in the base\\n> > catalog.\\n>\\n> Maybe I didn't make it clear, I was trying to explain that it would be\\n> complicated if we want to completely disallow specifying both child and\\n> parent\\n> table in the publication because of the ATTACH case I gave. In other\\n> words, I\\n> think it's fine to specify both child and parent table in the publication.\\n\\n\\nAh okay. I see your point. A table that was not once a partition would be\\nin the catalog along with the unrelated parent table and if we were to\\nenforce the rule that a parent and a partition cannot be in the catalog at\\nthe same time, we’d need to delete that table’s catalog record on ATTACHing\\nit as a partition.\\n\\nThanks.\\n-- \\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\nOn Thu, Nov 18, 2021 at 9:33 houzj.fnst@fujitsu.com <houzj.fnst@fujitsu.com> wrote:On Wed, Nov 17, 2021 2:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> On Wed, Nov 17, 2021 at 12:15 PM houzj.fnst@fujitsu.com\\n> <houzj.fnst@fujitsu.com> wrote:\\n> > On Wed, Nov 17, 2021 10:47 AM Amit Kapila <amit.kapila16@gmail.com>\\n> wrote:\\n> > > On Tue, Nov 16, 2021 at 7:21 AM houzj.fnst@fujitsu.com wrote:\\n> > > > If we decide to disallow this case, we seem need to handle some other\\n> > > > cases as well, for example: We might also need additional check when\\n> > > > ATTACH a partition, because the partition's parent table could already\\n> > > > be published in the same publication as the partition.\\n> > > >\\n> > >\\n> > > What kind of additional checks you are envisioning and why?\\n> >\\n> > For example:\\n> >\\n> > create table tbl1 (a int) partition by range (a);\\n> > create table tbl1_part1 (like tbl1);\\n> > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\\n> > create publication pub for table\\n> > tbl1, tbl1_part1 with (publish_via_partition_root=false);\\n> >\\n> > --- We might need addition check here\\n> > Alter table tbl1 ATTACH partition tbl1_part1 for values from (1) to (10);\\n> >\\n> > In the above cases, 'tbl1_part1' is not a partition of 'tb1' when creating a\\n> > publication. After the ATTACH, 'tbl1_part1' would become a partition of 'tbl1'\\n> > which seems the case we want to disallow(both parent and child table in\\n> > publication). So, When ATTACH, I thought we might need to check all the parent\\n> > tables to see if any parent table is in the same publication which the table to\\n> > be attached is also belongs to. Does it make sense ?\\n> \\n> I don't think creating or attaching a partition of a table that is\\n> present in a publish_via_partition_root=false actually adds the\\n> partition to pg_publication_rel, the base catalog. A partition's\\n> membership in the publication is implicit, unless of course you add it\\n> to the publication explicitly, like all the examples we have been\\n> discussing. I guess we're only arguing about the problems with the\\n> pg_publication_tables view, which does expand the partitioned table to\\n> show the partitions that are not otherwise not present in the base\\n> catalog.\\n\\nMaybe I didn't make it clear, I was trying to explain that it would be\\ncomplicated if we want to completely disallow specifying both child and parent\\ntable in the publication because of the ATTACH case I gave. In other words, I\\nthink it's fine to specify both child and parent table in the publication.Ah okay. I see your point. A table that was not once a partition would be in the catalog along with the unrelated parent table and if we were to enforce the rule that a parent and a partition cannot be in the catalog at the same time, we’d need to delete that table’s catalog record on ATTACHing it as a partition.Thanks.-- Amit LangoteEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Thu, 18 Nov 2021 09:57:30 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 11:39 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Tue, Nov 16, 2021 at 10:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Mon, Nov 15, 2021 at 7:12 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> > > > On Mon, Nov 15, 2021 at 1:48 PM houzj.fnst@fujitsu.com\\n> > > > <houzj.fnst@fujitsu.com> wrote:\\n> > >\\n> > > > > create table tbl1 (a int) partition by range (a);\\n> > > > > create table tbl1_part1 partition of tbl1 for values from (1) to (10);\\n> > > > > create table tbl1_part2 partition of tbl1 for values from (10) to (20);\\n> > > > > create publication pub for table\\n> > > > > tbl1, tbl1_part1 with (publish_via_partition_root=false);\\n> > >\\n> > > In the name of consistency, I think this situation should be an error --\\n> > > I mean, if we detect that the user is trying to add both the partitioned\\n> > > table *and* its partition, then all manner of things are possibly going\\n> > > to go wrong in some way, so my inclination is to avoid it altogether.\\n> > >\\n> > > Is there any reason to allow that?\\n> >\\n> > I think it could provide flexibility to users to later change\\n> > \\\"publish_via_partition_root\\\" option. Because when that option is\\n> > false, we use individual partitions schema to send changes and when it\\n> > is true, we use root table's schema to send changes. Added Amit L. to\\n> > know if he has any thoughts on this matter as he was the author of\\n> > this work?\\n>\\n> FWIW, I'm not sure that adding an error in this path, that is, when a\\n> user adds both a partitioned table and its partitions to a\\n> publication, helps much. As for the safety of allowing it, if you\\n> look at get_rel_sync_entry(), which handles much of the complexity of\\n> determining whether to publish a relation's changes and the schema to\\n> use when doing so, you may be able to see that a partition being added\\n> duplicatively is harmless, modulo any undiscovered bugs. At least as\\n> far as the post-initial-sync replication functionality is concerned.\\n>\\n> What IS problematic is what a subscriber sees in the\\n> pg_publication_tables view and the problem occurs only in the initial\\n> sync phase, where the partition is synced duplicatively because of\\n> being found in the view along with the parent in this case, that is,\\n> when publish_via_partiiton_root is true. I was saying on the other\\n> thread [1] that we should leave it up to the subscriber to decide what\\n> to do when the view (duplicatively) returns both the parent and the\\n> partition, citing the use case that a subscriber may want to replicate\\n> the parent and the partition as independent tables. Though I now tend\\n> to agree with Amit K that that may be such a meaningful and all that\\n> common use case, and the implementation on the subscriber side would\\n> have to be unnecessarily complex.\\n>\\n> So maybe it makes sense to just do what has been proposed --\\n> de-duplicate partitions out of the pg_publication_tables view,\\n>\\n\\nAFAICU, there are actually two problems related to\\npg_publication_tables view that are being discussed: (a) when\\n'publish_via_partition_root' is true then it returns both parent and\\nchild tables (provided both are part of publication), this can lead to\\nduplicate data on the subscriber; the fix for this is being discussed\\nin thread [1]. (b) when 'publish_via_partition_root' is false, then it\\nreturns duplicate entries for child tables (provided both are part of\\npublication), this problem is being discussed in this thread.\\n\\nAs per the proposed patches, it seems both need a separate fix, we can\\nfix them together if we want but I am not sure. Is your understanding\\nthe same?\\n\\n> unless\\n> we know of a bigger problem that requires us to hack the subscriber\\n> side of things too.\\n>\\n\\nThere is yet another issue that might need subscriber side change. See\\nthe second issue summarized by Hou-San in the email[2]. I feel it is\\nbetter to tackle that separately.\\n\\n> Actually, I came to know of one such problem\\n> while thinking about this today: when you ATTACH a partition to a\\n> table that is present in a publish_via_partition_root=true\\n> publication, it doesn't get copied via the initial sync even though\\n> subsequent replication works just fine. The reason for that is that\\n> the subscriber only syncs the partitions that are known at the time\\n> when the parent is first synced, that too via the parent (as SELECT\\n> <columns..> FROM parent), and then marks the parent as sync-done.\\n> Refreshing the subscription after ATTACHing doesn't help, because the\\n> subscriber can't see any partitions to begin with.\\n>\\n\\nSorry, it is not clear to me how this can lead to a problem. Even if\\nwe just have the root table in the subscriber at the time of sync\\nlater shouldn't copy get the newly attached partition's data and\\nreplicate it to the required partition for\\n\\\"publish_via_partition_root=true\\\" case? Anyway, if this is a problem\\nwe need to figure the solution for this separately.\\n\\n[1]: https://www.postgresql.org/message-id/CA%2BHiwqHnDHcT4OOcga9rDFyc7TvDrpN5xFH9J2pyHQo9ptvjmQ%40mail.gmail.com\\n[2]: https://www.postgresql.org/message-id/OS0PR01MB5716C756312959F293A822C794869%40OS0PR01MB5716.jpnprd01.prod.outlook.com\\n\\n\\n\\n--\\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 10:23:40 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 10:23 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> (b) when 'publish_via_partition_root' is false, then it\\n> returns duplicate entries for child tables (provided both are part of\\n> publication), this problem is being discussed in this thread.\\n\\n+1 to just focus on fixing the initial problem proposed i.e.\\npg_get_publication_tables() outputting duplicate relid. IMO, all other\\nissues (if unrelated to this problem and the patch can go to another\\nnew thread, if necessary).\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 10:48:12 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 1:53 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Wed, Nov 17, 2021 at 11:39 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > What IS problematic is what a subscriber sees in the\\n> > pg_publication_tables view and the problem occurs only in the initial\\n> > sync phase, where the partition is synced duplicatively because of\\n> > being found in the view along with the parent in this case, that is,\\n> > when publish_via_partiiton_root is true. I was saying on the other\\n> > thread [1] that we should leave it up to the subscriber to decide what\\n> > to do when the view (duplicatively) returns both the parent and the\\n> > partition, citing the use case that a subscriber may want to replicate\\n> > the parent and the partition as independent tables. Though I now tend\\n> > to agree with Amit K that that may be such a meaningful and all that\\n> > common use case, and the implementation on the subscriber side would\\n> > have to be unnecessarily complex.\\n> >\\n> > So maybe it makes sense to just do what has been proposed --\\n> > de-duplicate partitions out of the pg_publication_tables view,\\n>\\n> AFAICU, there are actually two problems related to\\n> pg_publication_tables view that are being discussed: (a) when\\n> 'publish_via_partition_root' is true then it returns both parent and\\n> child tables (provided both are part of publication), this can lead to\\n> duplicate data on the subscriber; the fix for this is being discussed\\n> in thread [1]. (b) when 'publish_via_partition_root' is false, then it\\n> returns duplicate entries for child tables (provided both are part of\\n> publication), this problem is being discussed in this thread.\\n>\\n> As per the proposed patches, it seems both need a separate fix, we can\\n> fix them together if we want but I am not sure. Is your understanding\\n> the same?\\n\\nActually, I'm seeing both (a) and (b) as more or less the same thing.\\nIf publish_via_partition_root=true, we'd want to remove a partition\\nfrom the list if the root parent is also present. If false, we'd want\\nto remove a partition from the list because it'd also be added by way\\nof expanding the root.\\n\\nI know that (a) causes the partition-double-sync problem that could be\\nfixed by partition OID de-duplication as proposed in the other thread.\\nAs a comment on that, it'd be nice to see a test case added to\\nsrc/test/subscription suite in that patch, because\\npartition-double-sync is the main problem statement I'd think.\\n\\nLike Bharath said, I don't see (b) as much of a problem, because the\\nsubscriber applies DISTINCT when querying pg_publication_tables()\\nanyway. Other users, if any out there, may be doing the same by\\nfollowing core subscription code's example. That said, I am not\\nopposed to applying the patch being proposed here, though I guess we\\ndon't have any src/test/subscription test case for this one? As in,\\ndo we know of any replication (initial/streaming) misbehavior caused\\nby the duplicate partition OIDs in this case or is the only problem\\nthat pg_publication_tables output looks odd?\\n\\n> > unless\\n> > we know of a bigger problem that requires us to hack the subscriber\\n> > side of things too.\\n> >\\n>\\n> There is yet another issue that might need subscriber side change. See\\n> the second issue summarized by Hou-San in the email[2]. I feel it is\\n> better to tackle that separately.\\n>\\n> > Actually, I came to know of one such problem\\n> > while thinking about this today: when you ATTACH a partition to a\\n> > table that is present in a publish_via_partition_root=true\\n> > publication, it doesn't get copied via the initial sync even though\\n> > subsequent replication works just fine. The reason for that is that\\n> > the subscriber only syncs the partitions that are known at the time\\n> > when the parent is first synced, that too via the parent (as SELECT\\n> > <columns..> FROM parent), and then marks the parent as sync-done.\\n> > Refreshing the subscription after ATTACHing doesn't help, because the\\n> > subscriber can't see any partitions to begin with.\\n> >\\n>\\n> Sorry, it is not clear to me how this can lead to a problem. Even if\\n> we just have the root table in the subscriber at the time of sync\\n> later shouldn't copy get the newly attached partition's data and\\n> replicate it to the required partition for\\n> \\\"publish_via_partition_root=true\\\" case?\\n\\nNot sure if you thought otherwise but I am talking about attaching a\\nnew partition into the partition tree on the *publication side*.\\n\\nThe problematic case is attaching the partition *after* the subscriber\\nhas already marked the root parent as synced and/or ready for\\nreplication. Refreshing the subscription doesn't help it discover the\\nnewly attached partition, because a publish_via_partition_root only\\never tells about the root parent, which would be already synced, so\\nthe subscriber would think there's nothing to copy.\\n\\n> Anyway, if this is a problem\\n> we need to figure the solution for this separately.\\n\\nSure, we might need to do that after all. Though it might be a good\\nidea to be sure that we won't need to reconsider the fix we push for\\nthe issue(s) being discussed here and elsewhere, because I suspect\\nthat the solution to the problem I mentioned is likely to involve\\ntweaking pg_publication_tables view output.\\n\\n--\\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 19 Nov 2021 10:49:32 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Thu, Nov 18, 2021 at 1:53 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > AFAICU, there are actually two problems related to\\n> > pg_publication_tables view that are being discussed: (a) when\\n> > 'publish_via_partition_root' is true then it returns both parent and\\n> > child tables (provided both are part of publication), this can lead to\\n> > duplicate data on the subscriber; the fix for this is being discussed\\n> > in thread [1]. (b) when 'publish_via_partition_root' is false, then it\\n> > returns duplicate entries for child tables (provided both are part of\\n> > publication), this problem is being discussed in this thread.\\n> >\\n> > As per the proposed patches, it seems both need a separate fix, we can\\n> > fix them together if we want but I am not sure. Is your understanding\\n> > the same?\\n>\\n> Actually, I'm seeing both (a) and (b) as more or less the same thing.\\n> If publish_via_partition_root=true, we'd want to remove a partition\\n> from the list if the root parent is also present. If false, we'd want\\n> to remove a partition from the list because it'd also be added by way\\n> of expanding the root.\\n>\\n> I know that (a) causes the partition-double-sync problem that could be\\n> fixed by partition OID de-duplication as proposed in the other thread.\\n> As a comment on that, it'd be nice to see a test case added to\\n> src/test/subscription suite in that patch, because\\n> partition-double-sync is the main problem statement I'd think.\\n>\\n> Like Bharath said, I don't see (b) as much of a problem, because the\\n> subscriber applies DISTINCT when querying pg_publication_tables()\\n> anyway. Other users, if any out there, may be doing the same by\\n> following core subscription code's example.\\n>\\n\\nHmm, I don't think we can assume that current or future users of\\npg_publication_tables() will use it with DISTINCT unless we specify in\\nthe specs/docs. However, I guess we might want to do it only for HEAD\\nas there is no direct field report for this and anyway we have some\\nworkaround for this.\\n\\n> That said, I am not\\n> opposed to applying the patch being proposed here, though I guess we\\n> don't have any src/test/subscription test case for this one?\\n>\\n\\nYeah, we can add tests for this and the other case.\\n\\n> As in,\\n> do we know of any replication (initial/streaming) misbehavior caused\\n> by the duplicate partition OIDs in this case or is the only problem\\n> that pg_publication_tables output looks odd?\\n>\\n\\nThe latter one but I think either we should document this or change it\\nas we can't assume users will follow what subscriber-side code does.\\n\\n> > > unless\\n> > > we know of a bigger problem that requires us to hack the subscriber\\n> > > side of things too.\\n> > >\\n> >\\n> > There is yet another issue that might need subscriber side change. See\\n> > the second issue summarized by Hou-San in the email[2]. I feel it is\\n> > better to tackle that separately.\\n> >\\n>\\n> The problematic case is attaching the partition *after* the subscriber\\n> has already marked the root parent as synced and/or ready for\\n> replication. Refreshing the subscription doesn't help it discover the\\n> newly attached partition, because a publish_via_partition_root only\\n> ever tells about the root parent, which would be already synced, so\\n> the subscriber would think there's nothing to copy.\\n>\\n\\nOkay, I see this could be a problem but I haven't tried to reproduce it.\\n\\n> > Anyway, if this is a problem\\n> > we need to figure the solution for this separately.\\n>\\n> Sure, we might need to do that after all. Though it might be a good\\n> idea to be sure that we won't need to reconsider the fix we push for\\n> the issue(s) being discussed here and elsewhere, because I suspect\\n> that the solution to the problem I mentioned is likely to involve\\n> tweaking pg_publication_tables view output.\\n>\\n\\nOkay, so we have four known problems in the same area. The first has\\nbeen reported at the start of this thread, then the two as summarized\\nby Hou-San in his email [1] and the fourth one is the attach of\\npartitions after initial sync as you mentioned. It makes sense to have\\nsome high-level idea on how to solve each of the problems before\\npushing a fix for any one particular problem as one fix could have an\\nimpact on other fixes. I'll think over it but please do let me know if\\nyou have any ideas.\\n\\n\\n[1] - https://www.postgresql.org/message-id/OS0PR01MB5716C756312959F293A822C794869%40OS0PR01MB5716.jpnprd01.prod.outlook.com\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 19 Nov 2021 10:58:32 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> >\\n> > The problematic case is attaching the partition *after* the subscriber\\n> > has already marked the root parent as synced and/or ready for\\n> > replication. Refreshing the subscription doesn't help it discover the\\n> > newly attached partition, because a publish_via_partition_root only\\n> > ever tells about the root parent, which would be already synced, so\\n> > the subscriber would think there's nothing to copy.\\n> >\\n>\\n> Okay, I see this could be a problem but I haven't tried to reproduce it.\\n>\\n> > > Anyway, if this is a problem\\n> > > we need to figure the solution for this separately.\\n> >\\n> > Sure, we might need to do that after all. Though it might be a good\\n> > idea to be sure that we won't need to reconsider the fix we push for\\n> > the issue(s) being discussed here and elsewhere, because I suspect\\n> > that the solution to the problem I mentioned is likely to involve\\n> > tweaking pg_publication_tables view output.\\n> >\\n\\nI have thought about this problem and I see two possibilities for a\\nsolution (a) We could provide a new option say 'truncate' (something\\non lines proposed here [1]) which would truncate the table(s) and\\nchange its status to 'i' in the pg_subscription_rel, this would allow\\nthe newly added partition to be synced after refresh. This could lead\\nto a large copy in such a case. (b) We could somehow get and store all\\nthe partition info from the publisher-side on the subscriber-side\\nwhile initial sync (say in new system table\\npg_subscription_rel_members). Now, after the refresh, if this list\\nchanges, we can allow to just get the data of that particular\\npartition but I guess it would mean that we need to store oids of the\\npublisher which might or might not be safe considering oids can\\nwraparound before the refresh.\\n\\nDo you have any other ideas?\\n\\nOne more thing you mentioned is that the initial sync won't work after\\nrefresh but later changes will be replicated but I noticed that later\\nchanges also don't get streamed till we restart the subscriber server.\\nI am not sure but we might not be invalidating apply workers cache due\\nto which it didn't notice the same.\\n\\n[1] - https://www.postgresql.org/message-id/CF3B6672-2A43-4204-A60A-68F359218A9B%40endpoint.com\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Sat, 20 Nov 2021 17:01:00 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > As in,\\n> > do we know of any replication (initial/streaming) misbehavior caused\\n> > by the duplicate partition OIDs in this case or is the only problem\\n> > that pg_publication_tables output looks odd?\\n>\\n> The latter one but I think either we should document this or change it\\n> as we can't assume users will follow what subscriber-side code does.\\n\\nOn second thought, I agree that de-duplicating partitions from this\\nview is an improvement.\\n\\n-- \\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 22 Nov 2021 16:24:45 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 20, 2021 at 8:31 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > The problematic case is attaching the partition *after* the subscriber\\n> > > has already marked the root parent as synced and/or ready for\\n> > > replication. Refreshing the subscription doesn't help it discover the\\n> > > newly attached partition, because a publish_via_partition_root only\\n> > > ever tells about the root parent, which would be already synced, so\\n> > > the subscriber would think there's nothing to copy.\\n> >\\n> > Okay, I see this could be a problem but I haven't tried to reproduce it.\\n> >\\n> > > > Anyway, if this is a problem\\n> > > > we need to figure the solution for this separately.\\n> > >\\n> > > Sure, we might need to do that after all. Though it might be a good\\n> > > idea to be sure that we won't need to reconsider the fix we push for\\n> > > the issue(s) being discussed here and elsewhere, because I suspect\\n> > > that the solution to the problem I mentioned is likely to involve\\n> > > tweaking pg_publication_tables view output.\\n>\\n> I have thought about this problem and I see two possibilities for a\\n> solution (a) We could provide a new option say 'truncate' (something\\n> on lines proposed here [1]) which would truncate the table(s) and\\n> change its status to 'i' in the pg_subscription_rel, this would allow\\n> the newly added partition to be synced after refresh. This could lead\\n> to a large copy in such a case.\\n\\nMaybe I am missing something about the proposal, though I'd think a\\nmore automatic solution would be better, something that doesn't need\\nto rely on an unrelated feature.\\n\\n> (b) We could somehow get and store all\\n> the partition info from the publisher-side on the subscriber-side\\n> while initial sync (say in new system table\\n> pg_subscription_rel_members). Now, after the refresh, if this list\\n> changes, we can allow to just get the data of that particular\\n> partition but I guess it would mean that we need to store oids of the\\n> publisher which might or might not be safe considering oids can\\n> wraparound before the refresh.\\n>\\n> Do you have any other ideas?\\n\\nI thought that the idea I had earlier mentioned at [1] may be useful,\\nwhich I can see is similar to your idea (b). I also suspect that it\\ncan be implemented without needing a separate catalog and storing\\npublication-side relation OIDs in the subscription-side catalog,\\nthough maybe I haven't thought hard enough.\\n\\n> One more thing you mentioned is that the initial sync won't work after\\n> refresh but later changes will be replicated but I noticed that later\\n> changes also don't get streamed till we restart the subscriber server.\\n> I am not sure but we might not be invalidating apply workers cache due\\n> to which it didn't notice the same.\\n\\nOh, that sounds odd and, as you appear to say, a separate problem. I'll check.\\n\\n-- \\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n[1] https://www.postgresql.org/message-id/CA%2BHiwqHnDHcT4OOcga9rDFyc7TvDrpN5xFH9J2pyHQo9ptvjmQ%40mail.gmail.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 22 Nov 2021 17:15:37 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 22, 2021 at 1:45 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Sat, Nov 20, 2021 at 8:31 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > > The problematic case is attaching the partition *after* the subscriber\\n> > > > has already marked the root parent as synced and/or ready for\\n> > > > replication. Refreshing the subscription doesn't help it discover the\\n> > > > newly attached partition, because a publish_via_partition_root only\\n> > > > ever tells about the root parent, which would be already synced, so\\n> > > > the subscriber would think there's nothing to copy.\\n> > >\\n> > > Okay, I see this could be a problem but I haven't tried to reproduce it.\\n> > >\\n> > > > > Anyway, if this is a problem\\n> > > > > we need to figure the solution for this separately.\\n> > > >\\n> > > > Sure, we might need to do that after all. Though it might be a good\\n> > > > idea to be sure that we won't need to reconsider the fix we push for\\n> > > > the issue(s) being discussed here and elsewhere, because I suspect\\n> > > > that the solution to the problem I mentioned is likely to involve\\n> > > > tweaking pg_publication_tables view output.\\n> >\\n> > I have thought about this problem and I see two possibilities for a\\n> > solution (a) We could provide a new option say 'truncate' (something\\n> > on lines proposed here [1]) which would truncate the table(s) and\\n> > change its status to 'i' in the pg_subscription_rel, this would allow\\n> > the newly added partition to be synced after refresh. This could lead\\n> > to a large copy in such a case.\\n>\\n> Maybe I am missing something about the proposal, though I'd think a\\n> more automatic solution would be better, something that doesn't need\\n> to rely on an unrelated feature.\\n>\\n\\nAgreed, this was more of a workaround for users if we didn't get any\\nautomatic solution.\\n\\n> > (b) We could somehow get and store all\\n> > the partition info from the publisher-side on the subscriber-side\\n> > while initial sync (say in new system table\\n> > pg_subscription_rel_members). Now, after the refresh, if this list\\n> > changes, we can allow to just get the data of that particular\\n> > partition but I guess it would mean that we need to store oids of the\\n> > publisher which might or might not be safe considering oids can\\n> > wraparound before the refresh.\\n> >\\n> > Do you have any other ideas?\\n>\\n> I thought that the idea I had earlier mentioned at [1] may be useful,\\n> which I can see is similar to your idea (b). I also suspect that it\\n> can be implemented without needing a separate catalog and storing\\n> publication-side relation OIDs in the subscription-side catalog,\\n> though maybe I haven't thought hard enough.\\n>\\n\\nThe problem with storing this info in pg_subscription_rel as you were\\ndescribing in your proposal is that currently, we ensure that the same\\ntable exists in subscriber and then store the subscriber side table id\\nin that catalog. I am not sure if we can store publisher-side oids in\\nthat catalog and if we store then it would be confusing as now it will\\nhave info of both publisher-side oids and subscriber-side oids. Isn't\\nthat a problem with this approach?\\n\\nOne more problem with this overall approach is the risk of OID\\nwraparound. Say between Create Subscription and Alter Subscription ..\\nRefresh, we detach one partition, the oid wraparounds, and we\\ncreate/reattach another partition which gets the same oid as for the\\npartition which we detached earlier then we won't be able to identify\\nthe new partition even after refresh.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 22 Nov 2021 16:17:31 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 22, 2021 6:48 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> On Mon, Nov 22, 2021 at 1:45 PM Amit Langote <amitlangote09@gmail.com>\\r\\n> wrote:\\r\\n> >\\r\\n> > On Sat, Nov 20, 2021 at 8:31 PM Amit Kapila <amit.kapila16@gmail.com>\\r\\n> wrote:\\r\\n> > > On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com>\\r\\n> wrote:\\r\\n> > > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote\\r\\n> <amitlangote09@gmail.com> wrote:\\r\\n> > > > > The problematic case is attaching the partition *after* the\\r\\n> > > > > subscriber has already marked the root parent as synced and/or\\r\\n> > > > > ready for replication. Refreshing the subscription doesn't help\\r\\n> > > > > it discover the newly attached partition, because a\\r\\n> > > > > publish_via_partition_root only ever tells about the root\\r\\n> > > > > parent, which would be already synced, so the subscriber would think\\r\\n> > > > > there's nothing to copy.\\r\\n> > > >\\r\\n> > > > Okay, I see this could be a problem but I haven't tried to reproduce it.\\r\\n> > > >\\r\\n> > > > > > Anyway, if this is a problem\\r\\n> > > > > > we need to figure the solution for this separately.\\r\\n> > > > >\\r\\n> > > > > Sure, we might need to do that after all. Though it might be a\\r\\n> > > > > good idea to be sure that we won't need to reconsider the fix we\\r\\n> > > > > push for the issue(s) being discussed here and elsewhere,\\r\\n> > > > > because I suspect that the solution to the problem I mentioned\\r\\n> > > > > is likely to involve tweaking pg_publication_tables view output.\\r\\n> > >\\r\\n> > > I have thought about this problem and I see two possibilities for a\\r\\n> > > solution (a) We could provide a new option say 'truncate' (something\\r\\n> > > on lines proposed here [1]) which would truncate the table(s) and\\r\\n> > > change its status to 'i' in the pg_subscription_rel, this would\\r\\n> > > allow the newly added partition to be synced after refresh. This\\r\\n> > > could lead to a large copy in such a case.\\r\\n> >\\r\\n> > Maybe I am missing something about the proposal, though I'd think a\\r\\n> > more automatic solution would be better, something that doesn't need\\r\\n> > to rely on an unrelated feature.\\r\\n> >\\r\\n> \\r\\n> Agreed, this was more of a workaround for users if we didn't get any automatic\\r\\n> solution.\\r\\n> \\r\\n> > > (b) We could somehow get and store all the partition info from the\\r\\n> > > publisher-side on the subscriber-side while initial sync (say in new\\r\\n> > > system table pg_subscription_rel_members). Now, after the refresh,\\r\\n> > > if this list changes, we can allow to just get the data of that\\r\\n> > > particular partition but I guess it would mean that we need to store\\r\\n> > > oids of the publisher which might or might not be safe considering\\r\\n> > > oids can wraparound before the refresh.\\r\\n> > >\\r\\n> > > Do you have any other ideas?\\r\\n> >\\r\\n> > I thought that the idea I had earlier mentioned at [1] may be useful,\\r\\n> > which I can see is similar to your idea (b). I also suspect that it\\r\\n> > can be implemented without needing a separate catalog and storing\\r\\n> > publication-side relation OIDs in the subscription-side catalog,\\r\\n> > though maybe I haven't thought hard enough.\\r\\n> >\\r\\n> \\r\\n> The problem with storing this info in pg_subscription_rel as you were\\r\\n> describing in your proposal is that currently, we ensure that the same table\\r\\n> exists in subscriber and then store the subscriber side table id in that catalog. I\\r\\n> am not sure if we can store publisher-side oids in that catalog and if we store\\r\\n> then it would be confusing as now it will have info of both publisher-side oids\\r\\n> and subscriber-side oids. Isn't that a problem with this approach?\\r\\n> \\r\\n> One more problem with this overall approach is the risk of OID wraparound.\\r\\n> Say between Create Subscription and Alter Subscription ..\\r\\n> Refresh, we detach one partition, the oid wraparounds, and we create/reattach\\r\\n> another partition which gets the same oid as for the partition which we\\r\\n> detached earlier then we won't be able to identify the new partition even after\\r\\n> refresh.\\r\\n\\r\\nIf we use this approach, I think maybe we can store the publisher-side's table\\r\\nname instead of oid. Because for non-partitioned table, it's possible that user\\r\\ndrop and create a new table with the same name in which case the oid would be\\r\\nchanged. And in this case, the existing behavior wouldn't sync the table again.\\r\\nSo I think it might be better to make the partitioned table's behavior\\r\\nconsistent with the non-partitioned table which only sync the table's data when\\r\\ndetect new table name.\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\",\n \"msg_date\": \"Tue, 23 Nov 2021 01:58:05 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 23, 2021 at 7:28 AM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n>\\n> On Mon, Nov 22, 2021 6:48 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Mon, Nov 22, 2021 at 1:45 PM Amit Langote <amitlangote09@gmail.com>\\n> > wrote:\\n> > >\\n> > > On Sat, Nov 20, 2021 at 8:31 PM Amit Kapila <amit.kapila16@gmail.com>\\n> > wrote:\\n> > > > On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com>\\n> > wrote:\\n> > > > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote\\n> > <amitlangote09@gmail.com> wrote:\\n> > > > > > The problematic case is attaching the partition *after* the\\n> > > > > > subscriber has already marked the root parent as synced and/or\\n> > > > > > ready for replication. Refreshing the subscription doesn't help\\n> > > > > > it discover the newly attached partition, because a\\n> > > > > > publish_via_partition_root only ever tells about the root\\n> > > > > > parent, which would be already synced, so the subscriber would think\\n> > > > > > there's nothing to copy.\\n> > > > >\\n> > > > > Okay, I see this could be a problem but I haven't tried to reproduce it.\\n> > > > >\\n> > > > > > > Anyway, if this is a problem\\n> > > > > > > we need to figure the solution for this separately.\\n> > > > > >\\n> > > > > > Sure, we might need to do that after all. Though it might be a\\n> > > > > > good idea to be sure that we won't need to reconsider the fix we\\n> > > > > > push for the issue(s) being discussed here and elsewhere,\\n> > > > > > because I suspect that the solution to the problem I mentioned\\n> > > > > > is likely to involve tweaking pg_publication_tables view output.\\n> > > >\\n> > > > I have thought about this problem and I see two possibilities for a\\n> > > > solution (a) We could provide a new option say 'truncate' (something\\n> > > > on lines proposed here [1]) which would truncate the table(s) and\\n> > > > change its status to 'i' in the pg_subscription_rel, this would\\n> > > > allow the newly added partition to be synced after refresh. This\\n> > > > could lead to a large copy in such a case.\\n> > >\\n> > > Maybe I am missing something about the proposal, though I'd think a\\n> > > more automatic solution would be better, something that doesn't need\\n> > > to rely on an unrelated feature.\\n> > >\\n> >\\n> > Agreed, this was more of a workaround for users if we didn't get any automatic\\n> > solution.\\n> >\\n> > > > (b) We could somehow get and store all the partition info from the\\n> > > > publisher-side on the subscriber-side while initial sync (say in new\\n> > > > system table pg_subscription_rel_members). Now, after the refresh,\\n> > > > if this list changes, we can allow to just get the data of that\\n> > > > particular partition but I guess it would mean that we need to store\\n> > > > oids of the publisher which might or might not be safe considering\\n> > > > oids can wraparound before the refresh.\\n> > > >\\n> > > > Do you have any other ideas?\\n> > >\\n> > > I thought that the idea I had earlier mentioned at [1] may be useful,\\n> > > which I can see is similar to your idea (b). I also suspect that it\\n> > > can be implemented without needing a separate catalog and storing\\n> > > publication-side relation OIDs in the subscription-side catalog,\\n> > > though maybe I haven't thought hard enough.\\n> > >\\n> >\\n> > The problem with storing this info in pg_subscription_rel as you were\\n> > describing in your proposal is that currently, we ensure that the same table\\n> > exists in subscriber and then store the subscriber side table id in that catalog. I\\n> > am not sure if we can store publisher-side oids in that catalog and if we store\\n> > then it would be confusing as now it will have info of both publisher-side oids\\n> > and subscriber-side oids. Isn't that a problem with this approach?\\n> >\\n> > One more problem with this overall approach is the risk of OID wraparound.\\n> > Say between Create Subscription and Alter Subscription ..\\n> > Refresh, we detach one partition, the oid wraparounds, and we create/reattach\\n> > another partition which gets the same oid as for the partition which we\\n> > detached earlier then we won't be able to identify the new partition even after\\n> > refresh.\\n>\\n> If we use this approach, I think maybe we can store the publisher-side's table\\n> name instead of oid. Because for non-partitioned table, it's possible that user\\n> drop and create a new table with the same name in which case the oid would be\\n> changed. And in this case, the existing behavior wouldn't sync the table again.\\n> So I think it might be better to make the partitioned table's behavior\\n> consistent with the non-partitioned table which only sync the table's data when\\n> detect new table name.\\n>\\n\\nI think firstly you need to then also store schema info. In the case\\nof non-partitioned tables, it is derived differently because we remove\\nthe table's entry from pg_publication_rel, so the data won't be sent\\nto subscriber till the new table is added to the publication. In the\\ncase of partitioned table (especially for the case of\\npublish_via_partition_root) we won't sync initial data in such a case\\n(dropped and re-created the partition with the same name) both with\\nand without your solution but the future data will be streamed.\\n\\nAnother point related to the Attach problem is what if such a\\npartition before was also part of publication separately, then if we\\ncopy its initial data then that data copy might be considered a\\nduplicate. Isn't it better to document this case and explain what\\nusers can expect instead of trying to design a solution around this?\\nEven if we do so the streaming after attach partition problem as\\ndiscussed above should be fixed.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 23 Nov 2021 08:50:57 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 23, 2021 at 12:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> Isn't it better to document this case and explain what\\n> users can expect instead of trying to design a solution around this?\\n\\nI thought about the problems you've described and it looks like I\\nhadn't considered many of the details which complicate implementing a\\nsolution for this.\\n\\nSo yeah, documenting the ATTACH issue as a limitation sounds like the\\nbest course for now. I might word it as follows and add it under\\nNotes at https://www.postgresql.org/docs/current/sql-createpublication.html:\\n\\n\\\"ATTACHing a table into a partition tree whose root is published using\\na publication with publish_via_partition_root set to true does not\\nresult in the table's existing contents to be replicated.\\\"\\n\\nI'm not sure there's a clean enough workaround to this that we can add\\nto the paragraph.\\n\\nDoes that make sense?\\n\\n> Even if we do so the streaming after attach partition problem as\\n> discussed above should be fixed.\\n\\nI agree. I have reproduced the problem though haven't managed to pin\\ndown the cause yet.\\n\\n--\\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 24 Nov 2021 15:32:04 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 24, 2021 at 12:02 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Tue, Nov 23, 2021 at 12:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > Isn't it better to document this case and explain what\\n> > users can expect instead of trying to design a solution around this?\\n>\\n> I thought about the problems you've described and it looks like I\\n> hadn't considered many of the details which complicate implementing a\\n> solution for this.\\n>\\n> So yeah, documenting the ATTACH issue as a limitation sounds like the\\n> best course for now. I might word it as follows and add it under\\n> Notes at https://www.postgresql.org/docs/current/sql-createpublication.html:\\n>\\n> \\\"ATTACHing a table into a partition tree whose root is published using\\n> a publication with publish_via_partition_root set to true does not\\n> result in the table's existing contents to be replicated.\\\"\\n>\\n\\nInstead of \\\"to be\\\", shall we use \\\"being\\\"?\\n\\n> I'm not sure there's a clean enough workaround to this that we can add\\n> to the paragraph.\\n>\\n> Does that make sense?\\n>\\n\\nYeah, that sounds reasonable. I think the case with\\n\\\"publish_via_partition_root = false\\\" should work but please test it\\nonce. The other thing which we need to consider about all these fixes\\nis whether to back patch them or not. I think it is on case to case\\nbasis. I feel this limitation should be documented and backpatched,\\nwhat do you think? Feel free to submit patches accordingly.\\n\\n> > Even if we do so the streaming after attach partition problem as\\n> > discussed above should be fixed.\\n>\\n> I agree. I have reproduced the problem though haven't managed to pin\\n> down the cause yet.\\n>\\n\\nNo problem, do let us know whatever be your findings at the end. I\\nthink now we know how to proceed with this attach issue, maybe we\\nshould also try to commit the fixes for other issues.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 24 Nov 2021 14:14:20 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 22, 2021 at 12:55 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > As in,\\n> > > do we know of any replication (initial/streaming) misbehavior caused\\n> > > by the duplicate partition OIDs in this case or is the only problem\\n> > > that pg_publication_tables output looks odd?\\n> >\\n> > The latter one but I think either we should document this or change it\\n> > as we can't assume users will follow what subscriber-side code does.\\n>\\n> On second thought, I agree that de-duplicating partitions from this\\n> view is an improvement.\\n>\\n\\nFair enough. Hou-San, Can you please submit the updated patch after\\nfixing any pending comments including the test case?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 24 Nov 2021 14:17:54 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 24, 2021 4:48 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> On Mon, Nov 22, 2021 at 12:55 PM Amit Langote <amitlangote09@gmail.com>\\r\\n> wrote:\\r\\n> >\\r\\n> > On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila <amit.kapila16@gmail.com>\\r\\n> wrote:\\r\\n> > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com>\\r\\n> wrote:\\r\\n> > > > As in,\\r\\n> > > > do we know of any replication (initial/streaming) misbehavior\\r\\n> > > > caused by the duplicate partition OIDs in this case or is the only\\r\\n> > > > problem that pg_publication_tables output looks odd?\\r\\n> > >\\r\\n> > > The latter one but I think either we should document this or change\\r\\n> > > it as we can't assume users will follow what subscriber-side code does.\\r\\n> >\\r\\n> > On second thought, I agree that de-duplicating partitions from this\\r\\n> > view is an improvement.\\r\\n> >\\r\\n> \\r\\n> Fair enough. Hou-San, Can you please submit the updated patch after fixing\\r\\n> any pending comments including the test case?\\r\\n\\r\\nOK, I will submit the updated patch soon.\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\",\n \"msg_date\": \"Thu, 25 Nov 2021 01:57:01 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 24, 2021 at 5:44 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Wed, Nov 24, 2021 at 12:02 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > So yeah, documenting the ATTACH issue as a limitation sounds like the\\n> > best course for now. I might word it as follows and add it under\\n> > Notes at https://www.postgresql.org/docs/current/sql-createpublication.html:\\n> >\\n> > \\\"ATTACHing a table into a partition tree whose root is published using\\n> > a publication with publish_via_partition_root set to true does not\\n> > result in the table's existing contents to be replicated.\\\"\\n>\\n> Instead of \\\"to be\\\", shall we use \\\"being\\\"?\\n\\nThat reads better.\\n\\n> > I'm not sure there's a clean enough workaround to this that we can add\\n> > to the paragraph.\\n> >\\n> > Does that make sense?\\n> >\\n>\\n> Yeah, that sounds reasonable. I think the case with\\n> \\\"publish_via_partition_root = false\\\" should work but please test it\\n> once.\\n\\nYeah, I did test that case to compare and didn't see a problem with it.\\n\\n> The other thing which we need to consider about all these fixes\\n> is whether to back patch them or not. I think it is on case to case\\n> basis. I feel this limitation should be documented and backpatched,\\n> what do you think? Feel free to submit patches accordingly.\\n\\nI agree with backpatching the doc fix. I've attached a diff against\\nmaster, though it also appears to apply to 13 and 14 branches.\\n\\n--\\nAmit Langote\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Thu, 25 Nov 2021 17:00:36 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 25, 2021 at 1:30 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Wed, Nov 24, 2021 at 5:44 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Wed, Nov 24, 2021 at 12:02 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > So yeah, documenting the ATTACH issue as a limitation sounds like the\\n> > > best course for now. I might word it as follows and add it under\\n> > > Notes at https://www.postgresql.org/docs/current/sql-createpublication.html:\\n> > >\\n> > > \\\"ATTACHing a table into a partition tree whose root is published using\\n> > > a publication with publish_via_partition_root set to true does not\\n> > > result in the table's existing contents to be replicated.\\\"\\n> >\\n> > Instead of \\\"to be\\\", shall we use \\\"being\\\"?\\n>\\n> That reads better.\\n>\\n> > > I'm not sure there's a clean enough workaround to this that we can add\\n> > > to the paragraph.\\n> > >\\n>\\n> I agree with backpatching the doc fix. I've attached a diff against\\n> master, though it also appears to apply to 13 and 14 branches.\\n>\\n\\nI think we can <literal></literal> for publish_via_partition_root,\\nother than that the patch looks good to me.\\n\\nHou-San, others, do you have any opinion about this patch and whether\\nto backpatch it or not?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 25 Nov 2021 14:26:58 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 25, 2021 at 5:57 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Thu, Nov 25, 2021 at 1:30 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > I agree with backpatching the doc fix. I've attached a diff against\\n> > master, though it also appears to apply to 13 and 14 branches.\\n>\\n> I think we can <literal></literal> for publish_via_partition_root,\\n> other than that the patch looks good to me.\\n\\nOh right, fixed. I also added the tag for \\\"true\\\" in the same sentence.\\n\\n-- \\nAmit Langote\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Thu, 25 Nov 2021 21:08:13 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thursday, November 25, 2021 4:57 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> On Thu, Nov 25, 2021 at 1:30 PM Amit Langote <amitlangote09@gmail.com>\\r\\n> >\\r\\n> > I agree with backpatching the doc fix. I've attached a diff against\\r\\n> > master, though it also appears to apply to 13 and 14 branches.\\r\\n> >\\r\\n> \\r\\n> I think we can <literal></literal> for publish_via_partition_root, other than that\\r\\n> the patch looks good to me.\\r\\n> \\r\\n> Hou-San, others, do you have any opinion about this patch and whether to\\r\\n> backpatch it or not?\\r\\n\\r\\nI think it makes sense to backpatch the doc fix, and the patch looks good to me.\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\",\n \"msg_date\": \"Fri, 26 Nov 2021 01:40:21 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 26, 2021 at 7:10 AM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n>\\n> On Thursday, November 25, 2021 4:57 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Thu, Nov 25, 2021 at 1:30 PM Amit Langote <amitlangote09@gmail.com>\\n> > >\\n> > > I agree with backpatching the doc fix. I've attached a diff against\\n> > > master, though it also appears to apply to 13 and 14 branches.\\n> > >\\n> >\\n> > I think we can <literal></literal> for publish_via_partition_root, other than that\\n> > the patch looks good to me.\\n> >\\n> > Hou-San, others, do you have any opinion about this patch and whether to\\n> > backpatch it or not?\\n>\\n> I think it makes sense to backpatch the doc fix, and the patch looks good to me.\\n>\\n\\nThanks, I'll push this sometime early next week unless there are any\\nobjections to it.\\n\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 26 Nov 2021 08:38:44 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 24, 2021 4:48 PM Amit Kapila <amit.kapila16@gmail.com>\\r\\n> On Mon, Nov 22, 2021 at 12:55 PM Amit Langote <amitlangote09@gmail.com>\\r\\n> wrote:\\r\\n> >\\r\\n> > On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila <amit.kapila16@gmail.com>\\r\\n> wrote:\\r\\n> > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com>\\r\\n> wrote:\\r\\n> > > > As in,\\r\\n> > > > do we know of any replication (initial/streaming) misbehavior\\r\\n> > > > caused by the duplicate partition OIDs in this case or is the only\\r\\n> > > > problem that pg_publication_tables output looks odd?\\r\\n> > >\\r\\n> > > The latter one but I think either we should document this or change\\r\\n> > > it as we can't assume users will follow what subscriber-side code does.\\r\\n> >\\r\\n> > On second thought, I agree that de-duplicating partitions from this\\r\\n> > view is an improvement.\\r\\n> >\\r\\n> \\r\\n> Fair enough. Hou-San, Can you please submit the updated patch after fixing\\r\\n> any pending comments including the test case?\\r\\n\\r\\nAttach the updated patch which address the comments so far.\\r\\n\\r\\nThe patch only adds a testcase in publication.sql because the duplicate\\r\\noutput doesn't cause unexpected behavior in pub-sub test.\\r\\n\\r\\nBest regards,\\r\\nHou zj\",\n \"msg_date\": \"Mon, 29 Nov 2021 09:07:41 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 26, 2021 at 8:38 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Fri, Nov 26, 2021 at 7:10 AM houzj.fnst@fujitsu.com\\n> <houzj.fnst@fujitsu.com> wrote:\\n> >\\n> > On Thursday, November 25, 2021 4:57 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > On Thu, Nov 25, 2021 at 1:30 PM Amit Langote <amitlangote09@gmail.com>\\n> > > >\\n> > > > I agree with backpatching the doc fix. I've attached a diff against\\n> > > > master, though it also appears to apply to 13 and 14 branches.\\n> > > >\\n> > >\\n> > > I think we can <literal></literal> for publish_via_partition_root, other than that\\n> > > the patch looks good to me.\\n> > >\\n> > > Hou-San, others, do you have any opinion about this patch and whether to\\n> > > backpatch it or not?\\n> >\\n> > I think it makes sense to backpatch the doc fix, and the patch looks good to me.\\n> >\\n>\\n> Thanks, I'll push this sometime early next week unless there are any\\n> objections to it.\\n>\\n\\nPushed.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Dec 2021 12:10:39 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 29, 2021 at 2:37 PM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n>\\n> On Wed, Nov 24, 2021 4:48 PM Amit Kapila <amit.kapila16@gmail.com>\\n> > On Mon, Nov 22, 2021 at 12:55 PM Amit Langote <amitlangote09@gmail.com>\\n> > wrote:\\n> > >\\n> > > On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila <amit.kapila16@gmail.com>\\n> > wrote:\\n> > > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com>\\n> > wrote:\\n> > > > > As in,\\n> > > > > do we know of any replication (initial/streaming) misbehavior\\n> > > > > caused by the duplicate partition OIDs in this case or is the only\\n> > > > > problem that pg_publication_tables output looks odd?\\n> > > >\\n> > > > The latter one but I think either we should document this or change\\n> > > > it as we can't assume users will follow what subscriber-side code does.\\n> > >\\n> > > On second thought, I agree that de-duplicating partitions from this\\n> > > view is an improvement.\\n> > >\\n> >\\n> > Fair enough. Hou-San, Can you please submit the updated patch after fixing\\n> > any pending comments including the test case?\\n>\\n> Attach the updated patch which address the comments so far.\\n>\\n> The patch only adds a testcase in publication.sql because the duplicate\\n> output doesn't cause unexpected behavior in pub-sub test.\\n>\\n\\nThanks, the patch looks good to me. I have slightly changed the commit\\nmessage in the attached. I would like to commit this only in HEAD as\\nthere is no user complaint about this and it might not stop any user's\\nservice unless it relies on this view's output for the initial table\\nsynchronization.\\n\\nWhat do you think? Bharath/Amit L., others, do you have any opinion on\\nthis matter?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\",\n \"msg_date\": \"Wed, 1 Dec 2021 12:30:57 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 1, 2021 3:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> On Mon, Nov 29, 2021 at 2:37 PM houzj.fnst@fujitsu.com\\r\\n> <houzj.fnst@fujitsu.com> wrote:\\r\\n> >\\r\\n> > On Wed, Nov 24, 2021 4:48 PM Amit Kapila <amit.kapila16@gmail.com>\\r\\n> > > On Mon, Nov 22, 2021 at 12:55 PM Amit Langote\\r\\n> > > <amitlangote09@gmail.com>\\r\\n> > > wrote:\\r\\n> > > >\\r\\n> > > > On Fri, Nov 19, 2021 at 2:28 PM Amit Kapila\\r\\n> > > > <amit.kapila16@gmail.com>\\r\\n> > > wrote:\\r\\n> > > > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote\\r\\n> > > > > <amitlangote09@gmail.com>\\r\\n> > > wrote:\\r\\n> > > > > > As in,\\r\\n> > > > > > do we know of any replication (initial/streaming) misbehavior\\r\\n> > > > > > caused by the duplicate partition OIDs in this case or is the\\r\\n> > > > > > only problem that pg_publication_tables output looks odd?\\r\\n> > > > >\\r\\n> > > > > The latter one but I think either we should document this or\\r\\n> > > > > change it as we can't assume users will follow what subscriber-side\\r\\n> > > > > code does.\\r\\n> > > >\\r\\n> > > > On second thought, I agree that de-duplicating partitions from\\r\\n> > > > this view is an improvement.\\r\\n> > > >\\r\\n> > >\\r\\n> > > Fair enough. Hou-San, Can you please submit the updated patch after\\r\\n> > > fixing any pending comments including the test case?\\r\\n> >\\r\\n> > Attach the updated patch which address the comments so far.\\r\\n> >\\r\\n> > The patch only adds a testcase in publication.sql because the\\r\\n> > duplicate output doesn't cause unexpected behavior in pub-sub test.\\r\\n> >\\r\\n> \\r\\n> Thanks, the patch looks good to me. I have slightly changed the commit\\r\\n> message in the attached. I would like to commit this only in HEAD as there is no\\r\\n> user complaint about this and it might not stop any user's service unless it relies\\r\\n> on this view's output for the initial table synchronization.\\r\\n> \\r\\n> What do you think?\\r\\nI agreed that we can commit only in HEAD.\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\\r\\n\",\n \"msg_date\": \"Thu, 2 Dec 2021 00:44:42 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 2, 2021 at 9:44 houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n> On Wed, Dec 1, 2021 3:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > > On Mon, Nov 22, 2021 at 12:55 PM Amit Langote\\n> > > > <amitlangote09@gmail.com> wrote:\\n> > > > > On second thought, I agree that de-duplicating partitions from\\n> > > > > this view is an improvement.\\n> > > >\\n> > > > Fair enough. Hou-San, Can you please submit the updated patch after\\n> > > > fixing any pending comments including the test case?\\n> > >\\n> > > Attach the updated patch which address the comments so far.\\n> > >\\n> > > The patch only adds a testcase in publication.sql because the\\n> > > duplicate output doesn't cause unexpected behavior in pub-sub test.\\n> >\\n> > Thanks, the patch looks good to me. I have slightly changed the commit\\n> > message in the attached. I would like to commit this only in HEAD as there is no\\n> > user complaint about this and it might not stop any user's service unless it relies\\n> > on this view's output for the initial table synchronization.\\n\\nThe patch looks good to me too in that it gets the job done.\\n\\nReading Alvaro's email at the top again gave me a pause to reconsider\\nwhat I had said in reply. It might indeed have been nice if the\\npublication DDL itself had prevented the cases where a partition and\\nits ancestor are added to a publication, though as Hou-san mentioned,\\npartition ATTACHes make that a bit tricky to enforce at all times,\\nthough of course not impossible. Maybe it's okay to just de-duplicate\\npg_publication_tables output as the patch does, even though it may\\nappear to be a band-aid solution if we one considers Alvaro's point\\nabout consistency. Sorry about too many second thoughts on these\\nthreads. :(\\n\\n> > What do you think?\\n> I agreed that we can commit only in HEAD.\\n\\n+1 to do this only in HEAD.\\n\\nThank you.\\n\\n\\n--\\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 2 Dec 2021 12:12:25 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Thu, Dec 2, 2021 at 9:44 houzj.fnst@fujitsu.com\\n> <houzj.fnst@fujitsu.com> wrote:\\n> > On Wed, Dec 1, 2021 3:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > > > On Mon, Nov 22, 2021 at 12:55 PM Amit Langote\\n> > > > > <amitlangote09@gmail.com> wrote:\\n> > > > > > On second thought, I agree that de-duplicating partitions from\\n> > > > > > this view is an improvement.\\n> > > > >\\n> > > > > Fair enough. Hou-San, Can you please submit the updated patch after\\n> > > > > fixing any pending comments including the test case?\\n> > > >\\n> > > > Attach the updated patch which address the comments so far.\\n> > > >\\n> > > > The patch only adds a testcase in publication.sql because the\\n> > > > duplicate output doesn't cause unexpected behavior in pub-sub test.\\n> > >\\n> > > Thanks, the patch looks good to me. I have slightly changed the commit\\n> > > message in the attached. I would like to commit this only in HEAD as there is no\\n> > > user complaint about this and it might not stop any user's service unless it relies\\n> > > on this view's output for the initial table synchronization.\\n>\\n> The patch looks good to me too in that it gets the job done.\\n>\\n> Reading Alvaro's email at the top again gave me a pause to reconsider\\n> what I had said in reply. It might indeed have been nice if the\\n> publication DDL itself had prevented the cases where a partition and\\n> its ancestor are added to a publication, though as Hou-san mentioned,\\n> partition ATTACHes make that a bit tricky to enforce at all times,\\n> though of course not impossible. Maybe it's okay to just de-duplicate\\n> pg_publication_tables output as the patch does, even though it may\\n> appear to be a band-aid solution if we one considers Alvaro's point\\n> about consistency.\\n>\\n\\nTrue, I think even if we consider that idea it will be a much bigger\\nchange, and also as it will be a behavioral change so we might want to\\nkeep it just for HEAD and some of these fixes need to be backpatched.\\nHaving said that, if you or someone want to pursue that idea and come\\nup with a better solution than what we have currently it is certainly\\nwelcome.\\n\\n\\n> > > What do you think?\\n> > I agreed that we can commit only in HEAD.\\n>\\n> +1 to do this only in HEAD.\\n>\\n\\nOkay, If there are no further suggestions or objections, I'll commit\\nthe patch early next week (mostly by Monday).\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 2 Dec 2021 10:57:41 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > Reading Alvaro's email at the top again gave me a pause to reconsider\\n> > what I had said in reply. It might indeed have been nice if the\\n> > publication DDL itself had prevented the cases where a partition and\\n> > its ancestor are added to a publication, though as Hou-san mentioned,\\n> > partition ATTACHes make that a bit tricky to enforce at all times,\\n> > though of course not impossible. Maybe it's okay to just de-duplicate\\n> > pg_publication_tables output as the patch does, even though it may\\n> > appear to be a band-aid solution if we one considers Alvaro's point\\n> > about consistency.\\n>\\n> True, I think even if we consider that idea it will be a much bigger\\n> change, and also as it will be a behavioral change so we might want to\\n> keep it just for HEAD and some of these fixes need to be backpatched.\\n> Having said that, if you or someone want to pursue that idea and come\\n> up with a better solution than what we have currently it is certainly\\n> welcome.\\n\\nOkay, I did write a PoC patch this morning after sending out my\\nearlier email. I polished it a bit, which is attached.\\n\\n-- \\nAmit Langote\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Thu, 2 Dec 2021 22:48:08 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > Reading Alvaro's email at the top again gave me a pause to reconsider\\n> > > what I had said in reply. It might indeed have been nice if the\\n> > > publication DDL itself had prevented the cases where a partition and\\n> > > its ancestor are added to a publication, though as Hou-san mentioned,\\n> > > partition ATTACHes make that a bit tricky to enforce at all times,\\n> > > though of course not impossible. Maybe it's okay to just de-duplicate\\n> > > pg_publication_tables output as the patch does, even though it may\\n> > > appear to be a band-aid solution if we one considers Alvaro's point\\n> > > about consistency.\\n> >\\n> > True, I think even if we consider that idea it will be a much bigger\\n> > change, and also as it will be a behavioral change so we might want to\\n> > keep it just for HEAD and some of these fixes need to be backpatched.\\n> > Having said that, if you or someone want to pursue that idea and come\\n> > up with a better solution than what we have currently it is certainly\\n> > welcome.\\n>\\n> Okay, I did write a PoC patch this morning after sending out my\\n> earlier email. I polished it a bit, which is attached.\\n>\\n\\nI see multiple problems with this patch and idea. (a) I think you\\nforgot to deal with \\\"All Tables In Schema\\\" Publication which will be\\nquite tricky to deal with during attach operation. How will you remove\\na particular relation from such a publication if there is a need to do\\nso? (b) Also, how will we prohibit adding partition and its root in\\nthe case of \\\"All Tables In Schema\\\" or \\\"All Tables\\\" publication? If\\nthere is no way to do that, then that would mean we anyway need to\\ndeal with parent and child tables as part of the same publication and\\nthis new behavior will add an exception to it. (c) I think this can\\nmake switching \\\"publish_via_partition_root\\\" inconvenient for users.\\nSay all or some of the partitions are part of the publication and then\\nthe user decides to add root table and change publish option\\n\\\"publish_via_partition_root\\\" to \\\"true\\\" at that moment she needs to\\nremove all partitions first.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 3 Dec 2021 09:07:38 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 3, 2021 at 12:37 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > > Reading Alvaro's email at the top again gave me a pause to reconsider\\n> > > > what I had said in reply. It might indeed have been nice if the\\n> > > > publication DDL itself had prevented the cases where a partition and\\n> > > > its ancestor are added to a publication, though as Hou-san mentioned,\\n> > > > partition ATTACHes make that a bit tricky to enforce at all times,\\n> > > > though of course not impossible. Maybe it's okay to just de-duplicate\\n> > > > pg_publication_tables output as the patch does, even though it may\\n> > > > appear to be a band-aid solution if we one considers Alvaro's point\\n> > > > about consistency.\\n> > >\\n> > > True, I think even if we consider that idea it will be a much bigger\\n> > > change, and also as it will be a behavioral change so we might want to\\n> > > keep it just for HEAD and some of these fixes need to be backpatched.\\n> > > Having said that, if you or someone want to pursue that idea and come\\n> > > up with a better solution than what we have currently it is certainly\\n> > > welcome.\\n> >\\n> > Okay, I did write a PoC patch this morning after sending out my\\n> > earlier email. I polished it a bit, which is attached.\\n>\\n> I see multiple problems with this patch and idea.\\n\\nThanks for looking at it. Yeah, I have not looked very closely at ALL\\nTABLES [IN SCHEMA], though only because I suspected that those cases\\ndeal with partitioning in such a way that the partition duplication\\nissue doesn't arise. That is, only the FOR TABLE list_of_tables and\\nADD TABLE syntax allow for the duplication issue to occur.\\n\\n> (a) I think you\\n> forgot to deal with \\\"All Tables In Schema\\\" Publication which will be\\n> quite tricky to deal with during attach operation. How will you remove\\n> a particular relation from such a publication if there is a need to do\\n> so?\\n\\nHmm, my understanding of how FOR ALL TABLES... features work is that\\none cannot remove a particular relation from such publications?\\n\\ncreate schema sch;\\ncreate table sch.p (a int primary key) partition by list (a);\\ncreate table sch.p1 partition of sch.p for values in (1);\\ncreate table sch.p2 partition of sch.p for values in (2);\\ncreate table p (a int primary key) partition by list (a);\\ncreate table p1 partition of p for values in (1);\\ncreate table p2 partition of p for values in (2);\\ncreate publication puball for all tables;\\ncreate publication pubsch for all tables in schema sch;\\n\\nalter publication puball drop table p;\\nERROR: publication \\\"puball\\\" is defined as FOR ALL TABLES\\nDETAIL: Tables cannot be added to or dropped from FOR ALL TABLES publications.\\n\\nalter publication pubsch drop table sch.p;\\nERROR: relation \\\"p\\\" is not part of the publication\\n\\nWhat am I missing?\\n\\n> (b) Also, how will we prohibit adding partition and its root in\\n> the case of \\\"All Tables In Schema\\\" or \\\"All Tables\\\" publication? If\\n> there is no way to do that, then that would mean we anyway need to\\n> deal with parent and child tables as part of the same publication and\\n> this new behavior will add an exception to it.\\n\\nI checked that FOR ALL TABLES publications don't allow adding a table\\nexplicitly, but apparently IN SCHEMA one does:\\n\\nalter publication pubsch add table p2;\\n\\\\dRp+ pubsch\\n Publication pubsch\\n Owner | All tables | Inserts | Updates | Deletes | Truncates | Via root\\n-------+------------+---------+---------+---------+-----------+----------\\n amit | f | t | t | t | t | f\\nTables:\\n \\\"public.p2\\\"\\nTables from schemas:\\n \\\"sch\\\"\\n\\nISTM that the ..IN SCHEMA syntax does not behave like FOR ALL TABLES\\nwithout the IN SCHEMA in this regard. Is that correct?\\n\\n> (c) I think this can\\n> make switching \\\"publish_via_partition_root\\\" inconvenient for users.\\n> Say all or some of the partitions are part of the publication and then\\n> the user decides to add root table and change publish option\\n> \\\"publish_via_partition_root\\\" to \\\"true\\\" at that moment she needs to\\n> remove all partitions first.\\n\\nHmm, actually I don't think I have tested the case where some or all\\npartitions are present by themselves before adding the root table.\\nI'll check that and try to think of sensible behavior for that.\\n\\n-- \\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 3 Dec 2021 21:34:35 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 3, 2021 at 21:34 Amit Langote <amitlangote09@gmail.com> wrote:\\n\\n> On Fri, Dec 3, 2021 at 12:37 PM Amit Kapila <amit.kapila16@gmail.com>\\n> wrote:\\n> > On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com>\\n> wrote:\\n> > > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com>\\n> wrote:\\n> > > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com>\\n> wrote:\\n> > > > > Reading Alvaro's email at the top again gave me a pause to\\n> reconsider\\n> > > > > what I had said in reply. It might indeed have been nice if the\\n> > > > > publication DDL itself had prevented the cases where a partition\\n> and\\n> > > > > its ancestor are added to a publication, though as Hou-san\\n> mentioned,\\n> > > > > partition ATTACHes make that a bit tricky to enforce at all times,\\n> > > > > though of course not impossible. Maybe it's okay to just\\n> de-duplicate\\n> > > > > pg_publication_tables output as the patch does, even though it may\\n> > > > > appear to be a band-aid solution if we one considers Alvaro's point\\n> > > > > about consistency.\\n> > > >\\n> > > > True, I think even if we consider that idea it will be a much bigger\\n> > > > change, and also as it will be a behavioral change so we might want\\n> to\\n> > > > keep it just for HEAD and some of these fixes need to be backpatched.\\n> > > > Having said that, if you or someone want to pursue that idea and come\\n> > > > up with a better solution than what we have currently it is certainly\\n> > > > welcome.\\n> > >\\n> > > Okay, I did write a PoC patch this morning after sending out my\\n> > > earlier email. I polished it a bit, which is attached.\\n> >\\n> > I see multiple problems with this patch and idea.\\n>\\n> Thanks for looking at it. Yeah, I have not looked very closely at ALL\\n> TABLES [IN SCHEMA], though only because I suspected that those cases\\n> deal with partitioning in such a way that the partition duplication\\n> issue doesn't arise. That is, only the FOR TABLE list_of_tables and\\n> ADD TABLE syntax allow for the duplication issue to occur.\\n\\n\\nAnother thing I forgot to mention is that the patch passes check-world.\\nPerhaps we don’t have enough tests that would’ve exposed any problems with\\nthe patch’s approach.\\n-- \\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\nOn Fri, Dec 3, 2021 at 21:34 Amit Langote <amitlangote09@gmail.com> wrote:On Fri, Dec 3, 2021 at 12:37 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > > Reading Alvaro's email at the top again gave me a pause to reconsider\\n> > > > what I had said in reply. It might indeed have been nice if the\\n> > > > publication DDL itself had prevented the cases where a partition and\\n> > > > its ancestor are added to a publication, though as Hou-san mentioned,\\n> > > > partition ATTACHes make that a bit tricky to enforce at all times,\\n> > > > though of course not impossible. Maybe it's okay to just de-duplicate\\n> > > > pg_publication_tables output as the patch does, even though it may\\n> > > > appear to be a band-aid solution if we one considers Alvaro's point\\n> > > > about consistency.\\n> > >\\n> > > True, I think even if we consider that idea it will be a much bigger\\n> > > change, and also as it will be a behavioral change so we might want to\\n> > > keep it just for HEAD and some of these fixes need to be backpatched.\\n> > > Having said that, if you or someone want to pursue that idea and come\\n> > > up with a better solution than what we have currently it is certainly\\n> > > welcome.\\n> >\\n> > Okay, I did write a PoC patch this morning after sending out my\\n> > earlier email. I polished it a bit, which is attached.\\n>\\n> I see multiple problems with this patch and idea.\\n\\nThanks for looking at it. Yeah, I have not looked very closely at ALL\\nTABLES [IN SCHEMA], though only because I suspected that those cases\\ndeal with partitioning in such a way that the partition duplication\\nissue doesn't arise. That is, only the FOR TABLE list_of_tables and\\nADD TABLE syntax allow for the duplication issue to occur.Another thing I forgot to mention is that the patch passes check-world. Perhaps we don’t have enough tests that would’ve exposed any problems with the patch’s approach.-- Amit LangoteEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Fri, 3 Dec 2021 22:27:43 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 3, 2021 at 6:04 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Fri, Dec 3, 2021 at 12:37 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > > > Reading Alvaro's email at the top again gave me a pause to reconsider\\n> > > > > what I had said in reply. It might indeed have been nice if the\\n> > > > > publication DDL itself had prevented the cases where a partition and\\n> > > > > its ancestor are added to a publication, though as Hou-san mentioned,\\n> > > > > partition ATTACHes make that a bit tricky to enforce at all times,\\n> > > > > though of course not impossible. Maybe it's okay to just de-duplicate\\n> > > > > pg_publication_tables output as the patch does, even though it may\\n> > > > > appear to be a band-aid solution if we one considers Alvaro's point\\n> > > > > about consistency.\\n> > > >\\n> > > > True, I think even if we consider that idea it will be a much bigger\\n> > > > change, and also as it will be a behavioral change so we might want to\\n> > > > keep it just for HEAD and some of these fixes need to be backpatched.\\n> > > > Having said that, if you or someone want to pursue that idea and come\\n> > > > up with a better solution than what we have currently it is certainly\\n> > > > welcome.\\n> > >\\n> > > Okay, I did write a PoC patch this morning after sending out my\\n> > > earlier email. I polished it a bit, which is attached.\\n> >\\n> > I see multiple problems with this patch and idea.\\n>\\n> Thanks for looking at it. Yeah, I have not looked very closely at ALL\\n> TABLES [IN SCHEMA], though only because I suspected that those cases\\n> deal with partitioning in such a way that the partition duplication\\n> issue doesn't arise. That is, only the FOR TABLE list_of_tables and\\n> ADD TABLE syntax allow for the duplication issue to occur.\\n>\\n> > (a) I think you\\n> > forgot to deal with \\\"All Tables In Schema\\\" Publication which will be\\n> > quite tricky to deal with during attach operation. How will you remove\\n> > a particular relation from such a publication if there is a need to do\\n> > so?\\n>\\n> Hmm, my understanding of how FOR ALL TABLES... features work is that\\n> one cannot remove a particular relation from such publications?\\n>\\n> create schema sch;\\n> create table sch.p (a int primary key) partition by list (a);\\n> create table sch.p1 partition of sch.p for values in (1);\\n> create table sch.p2 partition of sch.p for values in (2);\\n> create table p (a int primary key) partition by list (a);\\n> create table p1 partition of p for values in (1);\\n> create table p2 partition of p for values in (2);\\n> create publication puball for all tables;\\n> create publication pubsch for all tables in schema sch;\\n>\\n> alter publication puball drop table p;\\n> ERROR: publication \\\"puball\\\" is defined as FOR ALL TABLES\\n> DETAIL: Tables cannot be added to or dropped from FOR ALL TABLES publications.\\n>\\n> alter publication pubsch drop table sch.p;\\n> ERROR: relation \\\"p\\\" is not part of the publication\\n>\\n> What am I missing?\\n>\\n\\nCurrently, in your patch, you are trying to remove a particular\\nrelation/partition during attach but how will you do that if such a\\nrelation is part of All Tables In Schema publication?\\n\\n> > (b) Also, how will we prohibit adding partition and its root in\\n> > the case of \\\"All Tables In Schema\\\" or \\\"All Tables\\\" publication? If\\n> > there is no way to do that, then that would mean we anyway need to\\n> > deal with parent and child tables as part of the same publication and\\n> > this new behavior will add an exception to it.\\n>\\n> I checked that FOR ALL TABLES publications don't allow adding a table\\n> explicitly, but apparently IN SCHEMA one does:\\n>\\n> alter publication pubsch add table p2;\\n> \\\\dRp+ pubsch\\n> Publication pubsch\\n> Owner | All tables | Inserts | Updates | Deletes | Truncates | Via root\\n> -------+------------+---------+---------+---------+-----------+----------\\n> amit | f | t | t | t | t | f\\n> Tables:\\n> \\\"public.p2\\\"\\n> Tables from schemas:\\n> \\\"sch\\\"\\n>\\n> ISTM that the ..IN SCHEMA syntax does not behave like FOR ALL TABLES\\n> without the IN SCHEMA in this regard. Is that correct?\\n>\\n\\nWe do allow adding additional tables or schema to an exiting All\\ntables in schema publication.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 Dec 2021 10:29:20 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thursday, December 2, 2021 9:48 PM Amit Langote <amitlangote09@gmail.com> wrote:\\r\\n> On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com>\\r\\n> wrote:\\r\\n> > > Reading Alvaro's email at the top again gave me a pause to\\r\\n> > > reconsider what I had said in reply. It might indeed have been nice\\r\\n> > > if the publication DDL itself had prevented the cases where a\\r\\n> > > partition and its ancestor are added to a publication, though as\\r\\n> > > Hou-san mentioned, partition ATTACHes make that a bit tricky to\\r\\n> > > enforce at all times, though of course not impossible. Maybe it's\\r\\n> > > okay to just de-duplicate pg_publication_tables output as the patch\\r\\n> > > does, even though it may appear to be a band-aid solution if we one\\r\\n> > > considers Alvaro's point about consistency.\\r\\n> >\\r\\n> > True, I think even if we consider that idea it will be a much bigger\\r\\n> > change, and also as it will be a behavioral change so we might want to\\r\\n> > keep it just for HEAD and some of these fixes need to be backpatched.\\r\\n> > Having said that, if you or someone want to pursue that idea and come\\r\\n> > up with a better solution than what we have currently it is certainly\\r\\n> > welcome.\\r\\n> \\r\\n> Okay, I did write a PoC patch this morning after sending out my earlier email. I\\r\\n> polished it a bit, which is attached.\\r\\n\\r\\nAfter thinking more on this. I find there might be some usage about adding both\\r\\nchild and parent to the publication.\\r\\n\\r\\nFor the future feature publication row filter(and maybe column filter), It\\r\\ncould be useful for user to adding child and parent with different filter\\r\\nexpression. If pubviaroot=true, user can expect the parent's filter take\\r\\neffect, If pubviaroot=false, they can expect the child's filter take effect.\\r\\n\\r\\nIf we disallow adding both child and parent to publication, it could be harder\\r\\nfor these features to implement.\\r\\n\\r\\nSo, personally, I am not sure disallow adding both child and parent to the\\r\\npublication is a good idea.\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\",\n \"msg_date\": \"Mon, 6 Dec 2021 06:55:34 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 6, 2021 at 3:55 PM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n> On Thursday, December 2, 2021 9:48 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com>\\n> > wrote:\\n> > > > Reading Alvaro's email at the top again gave me a pause to\\n> > > > reconsider what I had said in reply. It might indeed have been nice\\n> > > > if the publication DDL itself had prevented the cases where a\\n> > > > partition and its ancestor are added to a publication, though as\\n> > > > Hou-san mentioned, partition ATTACHes make that a bit tricky to\\n> > > > enforce at all times, though of course not impossible. Maybe it's\\n> > > > okay to just de-duplicate pg_publication_tables output as the patch\\n> > > > does, even though it may appear to be a band-aid solution if we one\\n> > > > considers Alvaro's point about consistency.\\n> > >\\n> > > True, I think even if we consider that idea it will be a much bigger\\n> > > change, and also as it will be a behavioral change so we might want to\\n> > > keep it just for HEAD and some of these fixes need to be backpatched.\\n> > > Having said that, if you or someone want to pursue that idea and come\\n> > > up with a better solution than what we have currently it is certainly\\n> > > welcome.\\n> >\\n> > Okay, I did write a PoC patch this morning after sending out my earlier email. I\\n> > polished it a bit, which is attached.\\n>\\n> After thinking more on this. I find there might be some usage about adding both\\n> child and parent to the publication.\\n>\\n> For the future feature publication row filter(and maybe column filter), It\\n> could be useful for user to adding child and parent with different filter\\n> expression. If pubviaroot=true, user can expect the parent's filter take\\n> effect, If pubviaroot=false, they can expect the child's filter take effect.\\n>\\n> If we disallow adding both child and parent to publication, it could be harder\\n> for these features to implement.\\n\\nIt's possible that one may want such a feature, and yes, outright\\npreventing adding both the parent and a partition to a publication\\nwould perhaps make it harder. Mind you though that any such feature\\nwould need to fix the current implementation anyway to make the\\npublication-behavior-related catalog entries for child tables, which\\nwe currently don't, unless children are added explicitly.\\n\\nThat said, we should also be careful in implementing new features that\\noffer a finer per-partition granularity, because the fact that we do\\ncurrently offer that for many of the existing old features such as\\nconstraints, indexes, etc. limits the scalability of our architecture.\\nThat point may be off-topic for the thread, but something to consider,\\nor maybe I need to take a look at the other patch to see if my\\nconcerns are addressable / have been addressed.\\n\\n\\n--\\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 Dec 2021 16:30:00 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 6, 2021 at 1:00 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Mon, Dec 6, 2021 at 3:55 PM houzj.fnst@fujitsu.com\\n> <houzj.fnst@fujitsu.com> wrote:\\n> > On Thursday, December 2, 2021 9:48 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > On Thu, Dec 2, 2021 at 2:27 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > > On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com>\\n> > > wrote:\\n> > > > > Reading Alvaro's email at the top again gave me a pause to\\n> > > > > reconsider what I had said in reply. It might indeed have been nice\\n> > > > > if the publication DDL itself had prevented the cases where a\\n> > > > > partition and its ancestor are added to a publication, though as\\n> > > > > Hou-san mentioned, partition ATTACHes make that a bit tricky to\\n> > > > > enforce at all times, though of course not impossible. Maybe it's\\n> > > > > okay to just de-duplicate pg_publication_tables output as the patch\\n> > > > > does, even though it may appear to be a band-aid solution if we one\\n> > > > > considers Alvaro's point about consistency.\\n> > > >\\n> > > > True, I think even if we consider that idea it will be a much bigger\\n> > > > change, and also as it will be a behavioral change so we might want to\\n> > > > keep it just for HEAD and some of these fixes need to be backpatched.\\n> > > > Having said that, if you or someone want to pursue that idea and come\\n> > > > up with a better solution than what we have currently it is certainly\\n> > > > welcome.\\n> > >\\n> > > Okay, I did write a PoC patch this morning after sending out my earlier email. I\\n> > > polished it a bit, which is attached.\\n> >\\n> > After thinking more on this. I find there might be some usage about adding both\\n> > child and parent to the publication.\\n> >\\n> > For the future feature publication row filter(and maybe column filter), It\\n> > could be useful for user to adding child and parent with different filter\\n> > expression. If pubviaroot=true, user can expect the parent's filter take\\n> > effect, If pubviaroot=false, they can expect the child's filter take effect.\\n> >\\n> > If we disallow adding both child and parent to publication, it could be harder\\n> > for these features to implement.\\n>\\n> It's possible that one may want such a feature, and yes, outright\\n> preventing adding both the parent and a partition to a publication\\n> would perhaps make it harder. Mind you though that any such feature\\n> would need to fix the current implementation anyway to make the\\n> publication-behavior-related catalog entries for child tables, which\\n> we currently don't, unless children are added explicitly.\\n>\\n> That said, we should also be careful in implementing new features that\\n> offer a finer per-partition granularity, because the fact that we do\\n> currently offer that for many of the existing old features such as\\n> constraints, indexes, etc. limits the scalability of our architecture.\\n> That point may be off-topic for the thread, but something to consider,\\n> or maybe I need to take a look at the other patch to see if my\\n> concerns are addressable / have been addressed.\\n>\\n\\nYour concern is not very clear to me. Can you be more specific in what\\nkind of problem you see with such a design for row filtering? You\\nmight want to once look at that thread. You can find a summary of that\\npatch in the email [1].\\n\\n[1] - https://www.postgresql.org/message-id/49ba49f1-8bdb-40b7-ae9e-f17d88b3afcd%40www.fastmail.com\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 Dec 2021 14:35:36 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 6, 2021 at 1:59 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Fri, Dec 3, 2021 at 6:04 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > On Fri, Dec 3, 2021 at 12:37 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > On Thu, Dec 2, 2021 at 7:18 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > > Okay, I did write a PoC patch this morning after sending out my\\n> > > > earlier email. I polished it a bit, which is attached.\\n> > >\\n> > > I see multiple problems with this patch and idea.\\n> >\\n> > Thanks for looking at it. Yeah, I have not looked very closely at ALL\\n> > TABLES [IN SCHEMA], though only because I suspected that those cases\\n> > deal with partitioning in such a way that the partition duplication\\n> > issue doesn't arise. That is, only the FOR TABLE list_of_tables and\\n> > ADD TABLE syntax allow for the duplication issue to occur.\\n> >\\n> > > (a) I think you\\n> > > forgot to deal with \\\"All Tables In Schema\\\" Publication which will be\\n> > > quite tricky to deal with during attach operation. How will you remove\\n> > > a particular relation from such a publication if there is a need to do\\n> > > so?\\n> >\\n> > Hmm, my understanding of how FOR ALL TABLES... features work is that\\n> > one cannot remove a particular relation from such publications?\\n> >\\n> > create schema sch;\\n> > create table sch.p (a int primary key) partition by list (a);\\n> > create table sch.p1 partition of sch.p for values in (1);\\n> > create table sch.p2 partition of sch.p for values in (2);\\n> > create table p (a int primary key) partition by list (a);\\n> > create table p1 partition of p for values in (1);\\n> > create table p2 partition of p for values in (2);\\n> > create publication puball for all tables;\\n> > create publication pubsch for all tables in schema sch;\\n> >\\n> > alter publication puball drop table p;\\n> > ERROR: publication \\\"puball\\\" is defined as FOR ALL TABLES\\n> > DETAIL: Tables cannot be added to or dropped from FOR ALL TABLES publications.\\n> >\\n> > alter publication pubsch drop table sch.p;\\n> > ERROR: relation \\\"p\\\" is not part of the publication\\n> >\\n> > What am I missing?\\n>\\n> Currently, in your patch, you are trying to remove a particular\\n> relation/partition during attach but how will you do that if such a\\n> relation is part of All Tables In Schema publication?\\n\\nSo IIUC the scenario of concern is when a table to be attached as a\\npartition is in a schema that's present in pg_publication_namespace.\\nThe only way to stop it from being published is to move it to another\\nschema that is not published using that publication.\\n\\nI think I misunderstood how the IN SCHEMA feature works.\\nSpecifically, I didn't know that one can add a partitioned table to\\nthe same publication (or any table other than those in a particular\\nschema for that matter). Then the attached partition would still be\\npresent in the publication by way of being part of the schema that is\\npresent in the publication, along with the partitioned table that is\\nadded separately.\\n\\nYes, my proposal in its current form can't prevent that kind of duplication.\\n\\n--\\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 Dec 2021 23:34:29 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 6, 2021 at 6:05 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Mon, Dec 6, 2021 at 1:00 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > On Mon, Dec 6, 2021 at 3:55 PM houzj.fnst@fujitsu.com\\n> > <houzj.fnst@fujitsu.com> wrote:\\n> > > After thinking more on this. I find there might be some usage about adding both\\n> > > child and parent to the publication.\\n> > >\\n> > > For the future feature publication row filter(and maybe column filter), It\\n> > > could be useful for user to adding child and parent with different filter\\n> > > expression. If pubviaroot=true, user can expect the parent's filter take\\n> > > effect, If pubviaroot=false, they can expect the child's filter take effect.\\n> > >\\n> > > If we disallow adding both child and parent to publication, it could be harder\\n> > > for these features to implement.\\n> >\\n> > It's possible that one may want such a feature, and yes, outright\\n> > preventing adding both the parent and a partition to a publication\\n> > would perhaps make it harder. Mind you though that any such feature\\n> > would need to fix the current implementation anyway to make the\\n> > publication-behavior-related catalog entries for child tables, which\\n> > we currently don't, unless children are added explicitly.\\n> >\\n> > That said, we should also be careful in implementing new features that\\n> > offer a finer per-partition granularity, because the fact that we do\\n> > currently offer that for many of the existing old features such as\\n> > constraints, indexes, etc. limits the scalability of our architecture.\\n> > That point may be off-topic for the thread, but something to consider,\\n> > or maybe I need to take a look at the other patch to see if my\\n> > concerns are addressable / have been addressed.\\n> >\\n>\\n> Your concern is not very clear to me. Can you be more specific in what\\n> kind of problem you see with such a design for row filtering?\\n\\nI guess my characterization of it is still vague but the problem I see\\nis that we'll be adding one more feature that allows manipulating\\npartitions directly, which means more system catalogs than there\\nalready are that know about partitions as objects. I remember being\\ndeliberate about avoiding pg_publication_rel entries for all\\npartitions when a partitioned table is added to a publication for that\\nvery reason.\\n\\nIs covering the use case of defining different filters for each\\npartition important for developing that feature? I feel that defining\\none for the root table and just using that for all partitions (albeit\\ntranslating as needed to account for possibly different attribute\\nnumbers) would be enough for most users' use cases. Maybe I'm wrong\\non that.\\n\\n> You\\n> might want to once look at that thread. You can find a summary of that\\n> patch in the email [1].\\n\\nOkay, I'll try to do that.\\n\\n\\n--\\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 7 Dec 2021 00:29:06 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 6, 2021 at 8:04 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> So IIUC the scenario of concern is when a table to be attached as a\\n> partition is in a schema that's present in pg_publication_namespace.\\n> The only way to stop it from being published is to move it to another\\n> schema that is not published using that publication.\\n>\\n> I think I misunderstood how the IN SCHEMA feature works.\\n> Specifically, I didn't know that one can add a partitioned table to\\n> the same publication (or any table other than those in a particular\\n> schema for that matter). Then the attached partition would still be\\n> present in the publication by way of being part of the schema that is\\n> present in the publication, along with the partitioned table that is\\n> added separately.\\n>\\n\\nRight.\\n\\n> Yes, my proposal in its current form can't prevent that kind of duplication.\\n>\\n\\nI am not sure how to proceed here. I feel it is better to go ahead\\nwith the fix Hou-san proposed here and in another email [1] to fix the\\nknow issues, especially because the issue discussed in [1] needs to be\\nback-patched. We can evaluate your proposal separately for HEAD. What\\ndo you think? I am fine if you think that we should evaluate/analyze\\nyour proposal before making a call on whether to fix the existing\\nissues with the proposed set of patches.\\n\\n[1] - https://www.postgresql.org/message-id/OS0PR01MB57165D94CB187A97128F75EA946A9%40OS0PR01MB5716.jpnprd01.prod.outlook.com\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 7 Dec 2021 09:15:46 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 6, 2021 at 8:59 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Mon, Dec 6, 2021 at 6:05 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Mon, Dec 6, 2021 at 1:00 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > On Mon, Dec 6, 2021 at 3:55 PM houzj.fnst@fujitsu.com\\n> > > <houzj.fnst@fujitsu.com> wrote:\\n> > > > After thinking more on this. I find there might be some usage about adding both\\n> > > > child and parent to the publication.\\n> > > >\\n> > > > For the future feature publication row filter(and maybe column filter), It\\n> > > > could be useful for user to adding child and parent with different filter\\n> > > > expression. If pubviaroot=true, user can expect the parent's filter take\\n> > > > effect, If pubviaroot=false, they can expect the child's filter take effect.\\n> > > >\\n> > > > If we disallow adding both child and parent to publication, it could be harder\\n> > > > for these features to implement.\\n> > >\\n> > > It's possible that one may want such a feature, and yes, outright\\n> > > preventing adding both the parent and a partition to a publication\\n> > > would perhaps make it harder. Mind you though that any such feature\\n> > > would need to fix the current implementation anyway to make the\\n> > > publication-behavior-related catalog entries for child tables, which\\n> > > we currently don't, unless children are added explicitly.\\n> > >\\n> > > That said, we should also be careful in implementing new features that\\n> > > offer a finer per-partition granularity, because the fact that we do\\n> > > currently offer that for many of the existing old features such as\\n> > > constraints, indexes, etc. limits the scalability of our architecture.\\n> > > That point may be off-topic for the thread, but something to consider,\\n> > > or maybe I need to take a look at the other patch to see if my\\n> > > concerns are addressable / have been addressed.\\n> > >\\n> >\\n> > Your concern is not very clear to me. Can you be more specific in what\\n> > kind of problem you see with such a design for row filtering?\\n>\\n> I guess my characterization of it is still vague but the problem I see\\n> is that we'll be adding one more feature that allows manipulating\\n> partitions directly, which means more system catalogs than there\\n> already are that know about partitions as objects. I remember being\\n> deliberate about avoiding pg_publication_rel entries for all\\n> partitions when a partitioned table is added to a publication for that\\n> very reason.\\n>\\n\\nThe same will be true even after this new feature.\\n\\n> Is covering the use case of defining different filters for each\\n> partition important for developing that feature?\\n>\\n\\nUsers can only define the filters for a relation only when they want\\nto add a particular partition to the publication. Say, if only root\\ntable is added to the publication (and pubviaroot is true) then we\\ndon't need row filters to be defined for partitions. Basically, there\\nwill be only one additional column in pg_publication_rel for the row\\nfilter.\\n\\n> I feel that defining\\n> one for the root table and just using that for all partitions (albeit\\n> translating as needed to account for possibly different attribute\\n> numbers) would be enough for most users' use cases.\\n>\\n\\nSure, but what if the user just wants to publish only one of the\\npartitions? I think in that case along with adding the partition to\\npublication, we need to allow specifying row filter.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 7 Dec 2021 09:31:08 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 7, 2021 at 12:45 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Mon, Dec 6, 2021 at 8:04 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > So IIUC the scenario of concern is when a table to be attached as a\\n> > partition is in a schema that's present in pg_publication_namespace.\\n> > The only way to stop it from being published is to move it to another\\n> > schema that is not published using that publication.\\n> >\\n> > I think I misunderstood how the IN SCHEMA feature works.\\n> > Specifically, I didn't know that one can add a partitioned table to\\n> > the same publication (or any table other than those in a particular\\n> > schema for that matter). Then the attached partition would still be\\n> > present in the publication by way of being part of the schema that is\\n> > present in the publication, along with the partitioned table that is\\n> > added separately.\\n>\\n> Right.\\n>\\n> > Yes, my proposal in its current form can't prevent that kind of duplication.\\n>\\n> I am not sure how to proceed here. I feel it is better to go ahead\\n> with the fix Hou-san proposed here and in another email [1] to fix the\\n> know issues, especially because the issue discussed in [1] needs to be\\n> back-patched.\\n>\\n> We can evaluate your proposal separately for HEAD. What\\n> do you think?\\n\\nYeah, maybe. Though given the direction that the row-filters patch\\nset is taking in allowing to definw filters on the individual\\npartitions, I am not sure if I should be pushing the approach to\\ndisallow partitions from being added to publications explicitly\\nalongside their parent tables. I'll try to take a look at that thread\\nto be sure if that's actually the case.\\n\\nAlso, for the purposes of the problems that Greg and Hou-san have\\ndiscovered, I have no objection with applying Hou-san's patches.\\nThose seem better for the back-patching anyway.\\n\\nThank you.\\n\\n-- \\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 7 Dec 2021 14:07:56 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 7, 2021 at 1:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Mon, Dec 6, 2021 at 8:59 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > On Mon, Dec 6, 2021 at 6:05 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > Your concern is not very clear to me. Can you be more specific in what\\n> > > kind of problem you see with such a design for row filtering?\\n> >\\n> > I guess my characterization of it is still vague but the problem I see\\n> > is that we'll be adding one more feature that allows manipulating\\n> > partitions directly, which means more system catalogs than there\\n> > already are that know about partitions as objects. I remember being\\n> > deliberate about avoiding pg_publication_rel entries for all\\n> > partitions when a partitioned table is added to a publication for that\\n> > very reason.\\n>\\n> The same will be true even after this new feature.\\n\\nOh, interesting.\\n\\n> > Is covering the use case of defining different filters for each\\n> > partition important for developing that feature?\\n> >\\n>\\n> Users can only define the filters for a relation only when they want\\n> to add a particular partition to the publication. Say, if only root\\n> table is added to the publication (and pubviaroot is true) then we\\n> don't need row filters to be defined for partitions. Basically, there\\n> will be only one additional column in pg_publication_rel for the row\\n> filter.\\n\\nOkay, so you are talking about allowing setting filters on a partition\\nthat is added *explicitly* to a publication. I guess that's the only\\nway to do such a thing given the syntax of FOR TABLE / ADD TABLE.\\nFilters specified on a partitioned table that is added to a\\npublication apply equally to all partitions, depending on the\\npublish_via_partition_root configuration. So if a partition is\\nexplicitly present in a publication with a filter defined on it, I\\nsuppose following are the possible scenarios:\\n\\n* publish_via_partition_root=true\\n\\n1. If the root table is not present in the publication (perhaps\\nunlikely), use the partition's filter because there's nothing else to\\nconsider.\\n\\n2. If the root table is present in the publication (very likely), use\\nthe root's filter, ignoring/overriding the partition's own if any.\\n\\n* publish_via_partition_root=false\\n\\n1. If the root table is not present in the publication, use the\\npartition's filter because there's nothing else to consider.\\n\\n2. If the root table is present in the publication, use the\\npartition's own filter if present, else root's.\\n\\n> > I feel that defining\\n> > one for the root table and just using that for all partitions (albeit\\n> > translating as needed to account for possibly different attribute\\n> > numbers) would be enough for most users' use cases.\\n> >\\n>\\n> Sure, but what if the user just wants to publish only one of the\\n> partitions? I think in that case along with adding the partition to\\n> publication, we need to allow specifying row filter.\\n\\nYeah, I understand now that we're talking about partitions that are\\nadded explicitly.\\n\\nThanks for your patience.\\n\\n-- \\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 7 Dec 2021 14:21:54 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 7, 2021 at 10:52 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Tue, Dec 7, 2021 at 1:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Mon, Dec 6, 2021 at 8:59 PM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > > On Mon, Dec 6, 2021 at 6:05 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > > Your concern is not very clear to me. Can you be more specific in what\\n> > > > kind of problem you see with such a design for row filtering?\\n> > >\\n> > > I guess my characterization of it is still vague but the problem I see\\n> > > is that we'll be adding one more feature that allows manipulating\\n> > > partitions directly, which means more system catalogs than there\\n> > > already are that know about partitions as objects. I remember being\\n> > > deliberate about avoiding pg_publication_rel entries for all\\n> > > partitions when a partitioned table is added to a publication for that\\n> > > very reason.\\n> >\\n> > The same will be true even after this new feature.\\n>\\n> Oh, interesting.\\n>\\n> > > Is covering the use case of defining different filters for each\\n> > > partition important for developing that feature?\\n> > >\\n> >\\n> > Users can only define the filters for a relation only when they want\\n> > to add a particular partition to the publication. Say, if only root\\n> > table is added to the publication (and pubviaroot is true) then we\\n> > don't need row filters to be defined for partitions. Basically, there\\n> > will be only one additional column in pg_publication_rel for the row\\n> > filter.\\n>\\n> Okay, so you are talking about allowing setting filters on a partition\\n> that is added *explicitly* to a publication. I guess that's the only\\n> way to do such a thing given the syntax of FOR TABLE / ADD TABLE.\\n> Filters specified on a partitioned table that is added to a\\n> publication apply equally to all partitions, depending on the\\n> publish_via_partition_root configuration.\\n>\\n\\nYes.\\n\\n> So if a partition is\\n> explicitly present in a publication with a filter defined on it, I\\n> suppose following are the possible scenarios:\\n>\\n> * publish_via_partition_root=true\\n>\\n> 1. If the root table is not present in the publication (perhaps\\n> unlikely), use the partition's filter because there's nothing else to\\n> consider.\\n>\\n> 2. If the root table is present in the publication (very likely), use\\n> the root's filter, ignoring/overriding the partition's own if any.\\n>\\n> * publish_via_partition_root=false\\n>\\n> 1. If the root table is not present in the publication, use the\\n> partition's filter because there's nothing else to consider.\\n>\\n> 2. If the root table is present in the publication, use the\\n> partition's own filter if present, else root's.\\n>\\n\\nI have not tested/checked each of these scenarios individually but I\\nthink it is something like, if publish_via_partition_root is false\\nthen we will always try to use partitions row filter and if it is not\\nthere then we don't use any filter. Similarly, if\\npublish_via_partition_root is true, then we always try to use a\\npartitioned table row filter and if it is not there (either because\\nthe partitioned table is not part of the publication or because there\\nis no filter defined for it) then we don't use any filter.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 7 Dec 2021 11:50:26 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 7, 2021 at 3:20 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Tue, Dec 7, 2021 at 10:52 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n> > So if a partition is\\n> > explicitly present in a publication with a filter defined on it, I\\n> > suppose following are the possible scenarios:\\n> >\\n> > * publish_via_partition_root=true\\n> >\\n> > 1. If the root table is not present in the publication (perhaps\\n> > unlikely), use the partition's filter because there's nothing else to\\n> > consider.\\n> >\\n> > 2. If the root table is present in the publication (very likely), use\\n> > the root's filter, ignoring/overriding the partition's own if any.\\n> >\\n> > * publish_via_partition_root=false\\n> >\\n> > 1. If the root table is not present in the publication, use the\\n> > partition's filter because there's nothing else to consider.\\n> >\\n> > 2. If the root table is present in the publication, use the\\n> > partition's own filter if present, else root's.\\n>\\n> I have not tested/checked each of these scenarios individually but I\\n> think it is something like, if publish_via_partition_root is false\\n> then we will always try to use partitions row filter and if it is not\\n> there then we don't use any filter. Similarly, if\\n> publish_via_partition_root is true, then we always try to use a\\n> partitioned table row filter and if it is not there (either because\\n> the partitioned table is not part of the publication or because there\\n> is no filter defined for it) then we don't use any filter.\\n\\nOkay, thanks for the explanation.\\n\\n\\n--\\nAmit Langote\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 7 Dec 2021 21:56:55 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 2, 2021 at 8:42 AM Amit Langote <amitlangote09@gmail.com> wrote:\\n>\\n> On Thu, Dec 2, 2021 at 9:44 houzj.fnst@fujitsu.com\\n> <houzj.fnst@fujitsu.com> wrote:\\n> > On Wed, Dec 1, 2021 3:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > > > On Mon, Nov 22, 2021 at 12:55 PM Amit Langote\\n> > > > > <amitlangote09@gmail.com> wrote:\\n> > > > > > On second thought, I agree that de-duplicating partitions from\\n> > > > > > this view is an improvement.\\n> > > > >\\n> > > > > Fair enough. Hou-San, Can you please submit the updated patch after\\n> > > > > fixing any pending comments including the test case?\\n> > > >\\n> > > > Attach the updated patch which address the comments so far.\\n> > > >\\n> > > > The patch only adds a testcase in publication.sql because the\\n> > > > duplicate output doesn't cause unexpected behavior in pub-sub test.\\n> > >\\n> > > Thanks, the patch looks good to me. I have slightly changed the commit\\n> > > message in the attached. I would like to commit this only in HEAD as there is no\\n> > > user complaint about this and it might not stop any user's service unless it relies\\n> > > on this view's output for the initial table synchronization.\\n>\\n> The patch looks good to me too in that it gets the job done.\\n>\\n\\nPushed.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 8 Dec 2021 15:46:50 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 20, 2021 7:31 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com>\\r\\n> wrote:\\r\\n> >\\r\\n> > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <amitlangote09@gmail.com>\\r\\n> wrote:\\r\\n> > >\\r\\n> > > The problematic case is attaching the partition *after* the\\r\\n> > > subscriber has already marked the root parent as synced and/or ready\\r\\n> > > for replication. Refreshing the subscription doesn't help it\\r\\n> > > discover the newly attached partition, because a\\r\\n> > > publish_via_partition_root only ever tells about the root parent,\\r\\n> > > which would be already synced, so the subscriber would think there's\\r\\n> > > nothing to copy.\\r\\n> > >\\r\\n> >\\r\\n> > Okay, I see this could be a problem but I haven't tried to reproduce it.\\r\\n> \\r\\n> One more thing you mentioned is that the initial sync won't work after refresh\\r\\n> but later changes will be replicated but I noticed that later changes also don't\\r\\n> get streamed till we restart the subscriber server.\\r\\n> I am not sure but we might not be invalidating apply workers cache due to\\r\\n> which it didn't notice the same.\\r\\n\\r\\nI investigated this bug recently, and I think the reason is that when receiving\\r\\nrelcache invalidation message, the callback function[1] in walsender only reset\\r\\nthe schema sent status while it doesn't reset the replicate_valid flag. So, it\\r\\nwon’t rebuild the publication actions of the relation.\\r\\n\\r\\n[1]\\r\\nstatic void\\r\\nrel_sync_cache_relation_cb(Datum arg, Oid relid)\\r\\n...\\r\\n\\t/*\\r\\n\\t * Reset schema sent status as the relation definition may have changed.\\r\\n\\t * Also free any objects that depended on the earlier definition.\\r\\n\\t */\\r\\n\\tif (entry != NULL)\\r\\n\\t{\\r\\n\\t\\tentry->schema_sent = false;\\r\\n\\t\\tlist_free(entry->streamed_txns);\\r\\n...\\r\\n\\r\\nAlso, when you DETACH a partition, the publication won’t be rebuilt too because\\r\\nof the same reason. Which could cause unexpected behavior if we modify the\\r\\ndetached table's data . And the bug happens regardless of whether pubviaroot is\\r\\nset or not.\\r\\n\\r\\nFor the fix:\\r\\n\\r\\nI think if we also reset replicate_valid flag in rel_sync_cache_relation_cb,\\r\\nthen the bug can be fixed. I have a bit hesitation about this approach, because\\r\\nit could increase the frequency of invalidating and rebuilding the publication\\r\\naction. But I haven't produced some other better approaches.\\r\\n\\r\\nAnother possibility could be that we add a syscache callback function for\\r\\npg_inherits table, but we currently don't have syscache for pg_inherits. We\\r\\nmight need to add the cache pg_inherits first which doesn't seems better\\r\\nthan the above approach.\\r\\n\\r\\nWhat do you think ?\\r\\n\\r\\nAttach an initial patch which reset the replicate_valid flag in\\r\\nrel_sync_cache_relation_cb and add some reproduction tests in 013_partition.pl.\\r\\n\\r\\nBest regards,\\r\\nHou zj\",\n \"msg_date\": \"Tue, 14 Dec 2021 07:41:49 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Tuesday, December 14, 2021 3:42 PM houzj.fnst@fujitsu.com <houzj.fnst@fujitsu.com> wrote:\\r\\n> \\r\\n> On Sat, Nov 20, 2021 7:31 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> > On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <amit.kapila16@gmail.com>\\r\\n> > wrote:\\r\\n> > >\\r\\n> > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote\\r\\n> > > <amitlangote09@gmail.com>\\r\\n> > wrote:\\r\\n> > > >\\r\\n> > > > The problematic case is attaching the partition *after* the\\r\\n> > > > subscriber has already marked the root parent as synced and/or\\r\\n> > > > ready for replication. Refreshing the subscription doesn't help\\r\\n> > > > it discover the newly attached partition, because a\\r\\n> > > > publish_via_partition_root only ever tells about the root parent,\\r\\n> > > > which would be already synced, so the subscriber would think\\r\\n> > > > there's nothing to copy.\\r\\n> > > >\\r\\n> > >\\r\\n> > > Okay, I see this could be a problem but I haven't tried to reproduce it.\\r\\n> >\\r\\n> > One more thing you mentioned is that the initial sync won't work after\\r\\n> > refresh but later changes will be replicated but I noticed that later\\r\\n> > changes also don't get streamed till we restart the subscriber server.\\r\\n> > I am not sure but we might not be invalidating apply workers cache due\\r\\n> > to which it didn't notice the same.\\r\\n> \\r\\n> I investigated this bug recently, and I think the reason is that when receiving\\r\\n> relcache invalidation message, the callback function[1] in walsender only reset\\r\\n> the schema sent status while it doesn't reset the replicate_valid flag. So, it\\r\\n> won’t rebuild the publication actions of the relation.\\r\\n> \\r\\n> [1]\\r\\n> static void\\r\\n> rel_sync_cache_relation_cb(Datum arg, Oid relid) ...\\r\\n> \\t/*\\r\\n> \\t * Reset schema sent status as the relation definition may have\\r\\n> changed.\\r\\n> \\t * Also free any objects that depended on the earlier definition.\\r\\n> \\t */\\r\\n> \\tif (entry != NULL)\\r\\n> \\t{\\r\\n> \\t\\tentry->schema_sent = false;\\r\\n> \\t\\tlist_free(entry->streamed_txns);\\r\\n> ...\\r\\n> \\r\\n> Also, when you DETACH a partition, the publication won’t be rebuilt too\\r\\n> because of the same reason. Which could cause unexpected behavior if we\\r\\n> modify the detached table's data . And the bug happens regardless of whether\\r\\n> pubviaroot is set or not.\\r\\n> \\r\\n> For the fix:\\r\\n> \\r\\n> I think if we also reset replicate_valid flag in rel_sync_cache_relation_cb, then\\r\\n> the bug can be fixed. I have a bit hesitation about this approach, because it\\r\\n> could increase the frequency of invalidating and rebuilding the publication\\r\\n> action. But I haven't produced some other better approaches.\\r\\n> \\r\\n\\r\\nI have confirmed that the bug of ATTACH PARTITION has been fixed due to recent\\r\\ncommit 7f481b8. Currently, we always invalidate the RelationSyncCache when\\r\\nattaching a partition, so the pubactions of the newly attached partition will\\r\\nbe rebuilt correctly.\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\\r\\n\\r\\n\",\n \"msg_date\": \"Mon, 11 Apr 2022 09:04:28 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On 2022-Apr-11, houzj.fnst@fujitsu.com wrote:\\n\\n> I have confirmed that the bug of ATTACH PARTITION has been fixed due to recent\\n> commit 7f481b8. Currently, we always invalidate the RelationSyncCache when\\n> attaching a partition, so the pubactions of the newly attached partition will\\n> be rebuilt correctly.\\n\\nHm, this commit was not backpatched -- it is only in master. I just\\nchecked that pg14 behaves as you describe in the opening email of this\\nthread; did we decide that we don't want to change behavior of stable\\nbranches?\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Apr 2022 18:45:32 +0200\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_get_publication_tables() output duplicate relid\"\n },\n {\n \"msg_contents\": \"On Friday, April 15, 2022 12:46 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\r\\n> \\r\\n> On 2022-Apr-11, houzj.fnst@fujitsu.com wrote:\\r\\n> \\r\\n> > I have confirmed that the bug of ATTACH PARTITION has been fixed due to\\r\\n> recent\\r\\n> > commit 7f481b8. Currently, we always invalidate the RelationSyncCache when\\r\\n> > attaching a partition, so the pubactions of the newly attached partition will\\r\\n> > be rebuilt correctly.\\r\\n> \\r\\n> Hm, this commit was not backpatched -- it is only in master. I just\\r\\n> checked that pg14 behaves as you describe in the opening email of this\\r\\n> thread; did we decide that we don't want to change behavior of stable\\r\\n> branches?\\r\\n\\r\\nHi,\\r\\n\\r\\nDo you mean the original bug: \\\"pg_get_publication_tables function will output\\r\\nduplicate partition relid when adding both child and parent table to the\\r\\npublication(pubviaroot = false). \\\" ?\\r\\n\\r\\nWe decided to commit the fix for it only in HEAD as there is no user complaint\\r\\nabout this and it might not stop any user's service unless it relies on this\\r\\nview's output for the initial table synchronization.\\r\\n\\r\\nOr do you mean another one mentioned here which is \\\"the changes on newly\\r\\nattached partition won't be replicated\\\" ? For this one, I think it's fine to\\r\\nbackpatch the fix and I can prepare the patch for back branches if needed. BTW,\\r\\nif we want to backpatch this fix we might need to keep the order of\\r\\nRelationSyncEntry struct member unchanged in back branches.\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\",\n \"msg_date\": \"Mon, 18 Apr 2022 06:25:49 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: pg_get_publication_tables() output duplicate relid\"\n }\n]","isConversational":false}}},{"rowIdx":53091,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\npg_waldump options, --follow or -f(to keep polling once per second for\nnew WAL to appear) and --stats or -z don't work well together i.e. the\ncommand stucks [1]. I think the pg_waldump should emit an error. Note\nthat the pg_basebakup does error out on incompatible options.\n\nHere's a small patch for fixing above along with a note in the documentation.\n\nThoughts?\n\n[1] The following commands stuck:\n./pg_waldump -p data/ -s 0/7000060 -f -z\n./pg_waldump -p data/ -s 0/7000060 -f --stats='record'\n./pg_waldump -p data/ -s 0/7000060 -f --stats='rmgr'\n\n[2]\nubuntu3:~/postgres/inst/bin$ ./pg_waldump -p data/ -s 0/7000060 -f -z\npg_waldump: error: --follow and --stats are incompatible options\nTry \"pg_waldump --help\" for more information.\nbharath@bharathubuntu3:~/postgres/inst/bin$\n\nRegards,\nBharath Rupireddy.","msg_date":"Mon, 15 Nov 2021 19:32:38 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Mon, Nov 15, 2021 at 07:32:38PM +0530, Bharath Rupireddy wrote:\n> pg_waldump options, --follow or -f(to keep polling once per second for\n> new WAL to appear) and --stats or -z don't work well together i.e. the\n> command stucks [1]. I think the pg_waldump should emit an error. Note\n> that the pg_basebakup does error out on incompatible options.\n> \n> Here's a small patch for fixing above along with a note in the documentation.\n> \n> Thoughts?\n\nI don't think that we should block this combination of options as you\nare proposing. The existing behavior is useful for users when it\ncomes to an end position specified with -e, to be able to gather some\nstats on a cluster or an archive where we may not have all the\ncontents wanted yet.\n \n> [1] The following commands stuck:\n> ./pg_waldump -p data/ -s 0/7000060 -f -z\n> ./pg_waldump -p data/ -s 0/7000060 -f --stats='record'\n> ./pg_waldump -p data/ -s 0/7000060 -f --stats='rmgr'\n\nSaying that, you are not completely wrong either, as following\nsomething while we won't print any stats at all is not really\nhelpful. Another thing I can think of here is to make pg_waldump\nmore responsive to the dump of the stats when interrupted, via\nXLogDumpDisplayStats().\n--\nMichael","msg_date":"Tue, 16 Nov 2021 11:56:01 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Tue, Nov 16, 2021 at 8:26 AM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Mon, Nov 15, 2021 at 07:32:38PM +0530, Bharath Rupireddy wrote:\n> > pg_waldump options, --follow or -f(to keep polling once per second for\n> > new WAL to appear) and --stats or -z don't work well together i.e. the\n> > command stucks [1]. I think the pg_waldump should emit an error. Note\n> > that the pg_basebakup does error out on incompatible options.\n> >\n> > Here's a small patch for fixing above along with a note in the documentation.\n> >\n> > Thoughts?\n>\n> I don't think that we should block this combination of options as you\n> are proposing. The existing behavior is useful for users when it\n> comes to an end position specified with -e, to be able to gather some\n> stats on a cluster or an archive where we may not have all the\n> contents wanted yet.\n\nYou are right. The \"./pg_waldump -p data/ -s 0/14CC090 -e 0/14CC390 -f\n-z\" would fail with what I proposed which isn't a good thing at all.\nShould we block both the combinations \"-s/-f/-z\", \"-s/-e/-f/-z\"?\n\n> > [1] The following commands stuck:\n> > ./pg_waldump -p data/ -s 0/7000060 -f -z\n> > ./pg_waldump -p data/ -s 0/7000060 -f --stats='record'\n> > ./pg_waldump -p data/ -s 0/7000060 -f --stats='rmgr'\n>\n> Saying that, you are not completely wrong either, as following\n> something while we won't print any stats at all is not really\n> helpful. Another thing I can think of here is to make pg_waldump\n> more responsive to the dump of the stats when interrupted, via\n> XLogDumpDisplayStats().\n\nIt looks okay to be more responsive with the -f option so that the\nabove commands keep emitting stats for every 1 second(the --follow\noption waits for 1 second). Should we really be emitting stats for\nevery 1 second? Isn't there going to be too much information on the\nstdout? Or should we be emitting the stats for every 5 or 10 seconds?\n\nIf we be more responsive and keep emitting stats per 1 or 5 or 10 etc.\nseconds(we can give an option to the user to choose when he/she would\nlike to see the stats with -f option, but this is going to be an\noverkill IMO), can the user make anything out of those loads of stats\ngetting emitted? Another idea is to give some incremental stats but it\nis overkill again IMO. And when the pg_waldump has the capability to\nemit the stats, why to block it?\n\nIn summary, we have the following options:\n1) block the combinations \"-s/-f/-z\", \"-s/-e/-f/-z\"\n2) be more responsive and keep emitting the stats per 1 sec default\nwith -f option\n3) be more responsive and keep emitting the stats per user's choice\nof seconds (a new option that can be used with the -s/-e/-f/-z).\n\nThoughts?\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Tue, 16 Nov 2021 09:34:25 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Tue, Nov 16, 2021 at 09:34:25AM +0530, Bharath Rupireddy wrote:\n> It looks okay to be more responsive with the -f option so that the\n> above commands keep emitting stats for every 1 second(the --follow\n> option waits for 1 second). Should we really be emitting stats for\n> every 1 second? Isn't there going to be too much information on the\n> stdout? Or should we be emitting the stats for every 5 or 10 seconds?\n\nI don't have a perfect answer to this question, but dumping the stats\nwith a fixed frequency is not going to be useful if we don't have in\nthose logs a current timestamp and/or a current LSN. This is\nbasically about how much we want to call XLogDumpDisplayStats() and\nhow useful it is, but note that my argument is a bit different than\nwhat you are describing here: we could try to make\nXLogDumpDisplayStats() responsive on SIGINT or SIGTERM to show\nstatistics reports. This way, a --follow without an --end LSN\nspecified could still provide some information rather than nothing.\nThat could also be useful if defining an --end but interrupting the\ncall.\n\nAt the same time, we could also just let things as they are. --follow\nand --stats being specified together is what the user looked for, so\nthey get what they wanted.\n\n> In summary, we have the following options:\n> 1) block the combinations \"-s/-f/-z\", \"-s/-e/-f/-z\"\n> 2) be more responsive and keep emitting the stats per 1 sec default\n> with -f option\n> 3) be more responsive and keep emitting the stats per user's choice\n> of seconds (a new option that can be used with the -s/-e/-f/-z).\n\nA frequency cannot be measured only in time here, but also in bytes in\nterms of a minimum amount of WAL replayed between two reports. I\ndon't like much the idea of multiple stats reports emitted in a single\npg_waldump call, TBH. This makes things more complicated than they\nshould, and the gain is not obvious, at least to me.\n--\nMichael","msg_date":"Wed, 17 Nov 2021 11:19:52 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Wed, Nov 17, 2021 at 7:49 AM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> At the same time, we could also just let things as they are. --follow\n> and --stats being specified together is what the user looked for, so\n> they get what they wanted.\n\nI think the existing way of pg_waldump getting stuck with the\ncombination of options \"-s/-f/-z\" and \"-s/-e/-f/-z\" doesn't look good\nfor an end user. In the simplest way, the pg_waldump can just emit an\nerror for these combinations. That shouldn't a big problem as we do\nemit errors for a lot of mutually exclusive combination of options\nother SQL statements/commands.\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Wed, 17 Nov 2021 22:24:18 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On 2021-Nov-17, Bharath Rupireddy wrote:\n\n> On Wed, Nov 17, 2021 at 7:49 AM Michael Paquier <michael@paquier.xyz> wrote:\n> >\n> > At the same time, we could also just let things as they are. --follow\n> > and --stats being specified together is what the user looked for, so\n> > they get what they wanted.\n> \n> I think the existing way of pg_waldump getting stuck with the\n> combination of options \"-s/-f/-z\" and \"-s/-e/-f/-z\" doesn't look good\n> for an end user.\n\nI agree that it's not right, but I don't think the solution is to ban\nthe combination. I think what we should do is change the behavior to\nmake the combination potentially useful, so that it waits until program\ntermination and print the summary then. Consider \"strace -c\" as a\nprecedent: it will also \"become stuck\" until you kill it, and then it'll\nprint a nice table, just like the pg_waldump -z gives you.\n\nI think I would even have had occasion to use this as a feature in the\npast ...\n\n-- \nÁlvaro Herrera 39°49'30\"S 73°17'W — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Wed, 17 Nov 2021 15:35:03 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Thu, Nov 18, 2021 at 12:05 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>\n> On 2021-Nov-17, Bharath Rupireddy wrote:\n>\n> > On Wed, Nov 17, 2021 at 7:49 AM Michael Paquier <michael@paquier.xyz> wrote:\n> > >\n> > > At the same time, we could also just let things as they are. --follow\n> > > and --stats being specified together is what the user looked for, so\n> > > they get what they wanted.\n> >\n> > I think the existing way of pg_waldump getting stuck with the\n> > combination of options \"-s/-f/-z\" and \"-s/-e/-f/-z\" doesn't look good\n> > for an end user.\n>\n> I agree that it's not right, but I don't think the solution is to ban\n> the combination. I think what we should do is change the behavior to\n> make the combination potentially useful, so that it waits until program\n> termination and print the summary then. Consider \"strace -c\" as a\n> precedent: it will also \"become stuck\" until you kill it, and then it'll\n> print a nice table, just like the pg_waldump -z gives you.\n>\n> I think I would even have had occasion to use this as a feature in the\n> past ...\n\nHm. So, the pg_waldump can have handlers for SIGINT, SIGTERM, SIGQUIT\nand then it should emit the computed stats in those handlers the\ncomobinations - \"-s/-f/-z\" and \"-s/-e/-f/-z\". I'm okay with this\nbehaviour. Michael Paquier had suggested the same upthread. If okay, I\nwill work on that patch.\n\nThoughts?\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Thu, 18 Nov 2021 13:32:08 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Thu, Nov 18, 2021 at 01:32:08PM +0530, Bharath Rupireddy wrote:\n> Hm. So, the pg_waldump can have handlers for SIGINT, SIGTERM, SIGQUIT\n> and then it should emit the computed stats in those handlers the\n> comobinations - \"-s/-f/-z\" and \"-s/-e/-f/-z\". I'm okay with this\n> behaviour. Michael Paquier had suggested the same upthread. If okay, I\n> will work on that patch.\n\nWhile on it, I am pretty sure that we should add in the stats report\nthe start and end LSNs matching with the reports. If you use --follow\nand the call is interrupted, we would not really know to which part of\nthe WAL stream a report applies to without this information, likely in\nthe shape of an extra header.\n--\nMichael","msg_date":"Thu, 18 Nov 2021 17:21:47 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Thu, Nov 18, 2021 at 1:51 PM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Thu, Nov 18, 2021 at 01:32:08PM +0530, Bharath Rupireddy wrote:\n> > Hm. So, the pg_waldump can have handlers for SIGINT, SIGTERM, SIGQUIT\n> > and then it should emit the computed stats in those handlers the\n> > comobinations - \"-s/-f/-z\" and \"-s/-e/-f/-z\". I'm okay with this\n> > behaviour. Michael Paquier had suggested the same upthread. If okay, I\n> > will work on that patch.\n>\n> While on it, I am pretty sure that we should add in the stats report\n> the start and end LSNs matching with the reports. If you use --follow\n> and the call is interrupted, we would not really know to which part of\n> the WAL stream a report applies to without this information, likely in\n> the shape of an extra header.\n\nCurrently, the SIGINT, SIGTERM, SIGQUIT interruptions will cause the\npg_waldump process with --follow option to get killed. And it is a good\nidea to specify the range of LSNs (start and end) upto which the stats are\ncomputed. Maybe as another printf message at the end of the stats report,\nafter the \"Total\" section or at the beginning of the stats report.\nSomething like \"summary of the statistics computed from << LSN >> to\n<<LSN>> are:\", see [1].\n\n[1] ubuntu3:~/postgres/src/bin/pg_waldump$ ./pg_waldump -p data -s\n0/1480000 --stats\nsummary of the statistics computed from 0/1480000 and 0/158ABCD are:\nType N (%) Record\nsize (%) FPI size (%) Combined size (%)\n---- - ---\n ----------- --- -------- --- -------------\n ---\nXLOG 13 ( 0.13)\n1138 ( 0.18) 8424 ( 2.46) 9562 ( 0.99)\nTransaction 14 ( 0.14)\n1407 ( 0.22) 0 ( 0.00) 1407 ( 0.15)\nStorage 4 ( 0.04)\n 172 ( 0.03) 0 ( 0.00) 172 ( 0.02)\nCLOG 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nDatabase 1 ( 0.01)\n42 ( 0.01) 0 ( 0.00) 42 ( 0.00)\nTablespace 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nMultiXact 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nRelMap 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nStandby 37 ( 0.36)\n2318 ( 0.37) 0 ( 0.00) 2318 ( 0.24)\nHeap2 74 ( 0.73)\n 12841 ( 2.05) 121184 ( 35.46) 134025 ( 13.84)\nHeap 9005 ( 88.23)\n532313 ( 84.91) 84208 ( 24.64) 616521 ( 63.64)\nBtree 1058 ( 10.37)\n 76710 ( 12.24) 127932 ( 37.43) 204642 ( 21.13)\nHash 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nGin 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nGist 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nSequence 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nSPGist 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nBRIN 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nCommitTs 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nReplicationOrigin 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nGeneric 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\nLogicalMessage 0 ( 0.00)\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\n --------\n -------- -------- --------\nTotal 10206\n 626941 [64.72%] 341748 [35.28%] 968689 [100%]\n\nRegards,\nBharath Rupireddy.\n\nOn Thu, Nov 18, 2021 at 1:51 PM Michael Paquier <michael@paquier.xyz> wrote:>> On Thu, Nov 18, 2021 at 01:32:08PM +0530, Bharath Rupireddy wrote:> > Hm. So, the pg_waldump can have handlers for SIGINT, SIGTERM, SIGQUIT> > and then it should emit the computed stats in those handlers the> > comobinations - \"-s/-f/-z\" and \"-s/-e/-f/-z\". I'm okay with this> > behaviour. Michael Paquier had suggested the same upthread. If okay, I> > will work on that patch.>> While on it, I am pretty sure that we should add in the stats report> the start and end LSNs matching with the reports. If you use --follow> and the call is interrupted, we would not really know to which part of> the WAL stream a report applies to without this information, likely in> the shape of an extra header.Currently, the SIGINT, SIGTERM, SIGQUIT interruptions will cause the pg_waldump process with --follow option to get killed. And it is a good idea to specify the range of LSNs (start and end) upto which the stats are computed. Maybe as another printf message at the end of the stats report, after the \"Total\" section or at the beginning of the stats report. Something like \"summary of the statistics computed from << LSN >> to <<LSN>> are:\", see [1].[1] ubuntu3:~/postgres/src/bin/pg_waldump$ ./pg_waldump -p data -s 0/1480000 --statssummary of the statistics computed from 0/1480000 and 0/158ABCD are:Type N (%) Record size (%) FPI size (%) Combined size (%)---- - --- ----------- --- -------- --- ------------- ---XLOG 13 ( 0.13) 1138 ( 0.18) 8424 ( 2.46) 9562 ( 0.99)Transaction 14 ( 0.14) 1407 ( 0.22) 0 ( 0.00) 1407 ( 0.15)Storage 4 ( 0.04) 172 ( 0.03) 0 ( 0.00) 172 ( 0.02)CLOG 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Database 1 ( 0.01) 42 ( 0.01) 0 ( 0.00) 42 ( 0.00)Tablespace 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)MultiXact 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)RelMap 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Standby 37 ( 0.36) 2318 ( 0.37) 0 ( 0.00) 2318 ( 0.24)Heap2 74 ( 0.73) 12841 ( 2.05) 121184 ( 35.46) 134025 ( 13.84)Heap 9005 ( 88.23) 532313 ( 84.91) 84208 ( 24.64) 616521 ( 63.64)Btree 1058 ( 10.37) 76710 ( 12.24) 127932 ( 37.43) 204642 ( 21.13)Hash 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Gin 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Gist 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Sequence 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)SPGist 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)BRIN 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)CommitTs 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)ReplicationOrigin 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Generic 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)LogicalMessage 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) -------- -------- -------- --------Total 10206 626941 [64.72%] 341748 [35.28%] 968689 [100%]Regards,Bharath Rupireddy.","msg_date":"Thu, 18 Nov 2021 14:03:20 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Thu, Nov 18, 2021 at 2:03 PM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n>\n> On Thu, Nov 18, 2021 at 1:51 PM Michael Paquier <michael@paquier.xyz> wrote:\n> >\n> > On Thu, Nov 18, 2021 at 01:32:08PM +0530, Bharath Rupireddy wrote:\n> > > Hm. So, the pg_waldump can have handlers for SIGINT, SIGTERM, SIGQUIT\n> > > and then it should emit the computed stats in those handlers the\n> > > comobinations - \"-s/-f/-z\" and \"-s/-e/-f/-z\". I'm okay with this\n> > > behaviour. Michael Paquier had suggested the same upthread. If okay, I\n> > > will work on that patch.\n> >\n> > While on it, I am pretty sure that we should add in the stats report\n> > the start and end LSNs matching with the reports. If you use --follow\n> > and the call is interrupted, we would not really know to which part of\n> > the WAL stream a report applies to without this information, likely in\n> > the shape of an extra header.\n>\n> Currently, the SIGINT, SIGTERM, SIGQUIT interruptions will cause the pg_waldump process with --follow option to get killed. And it is a good idea to specify the range of LSNs (start and end) upto which the stats are computed. Maybe as another printf message at the end of the stats report, after the \"Total\" section or at the beginning of the stats report. Something like \"summary of the statistics computed from << LSN >> to <<LSN>> are:\", see [1].\n\nHere's the v2 patch with the above changes i.e. pg_waldump now eimits\nthe computed stats at the time of termination. Please review it.\n\nRegards,\nBharath Rupireddy.","msg_date":"Sat, 20 Nov 2021 10:03:27 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Sat, Nov 20, 2021 at 10:03:27AM +0530, Bharath Rupireddy wrote:\n> Here's the v2 patch with the above changes i.e. pg_waldump now eimits\n> the computed stats at the time of termination. Please review it.\n\n+#define MAX_XLINFO_TYPES 16\n+#define fatal_error(...) do { pg_log_fatal(__VA_ARGS__);\nexit(EXIT_FAILURE); } while(0)\n typedef struct XLogDumpPrivate\nMoving around those declarations does not relate to this patch, so I\nwould recommend to leave this out to ease reviews.\n\n+ /*\n+ * Although the pg_free calls (above and below) are unnecessary here on the\n+ * process exit, let's not leak any memory knowingly.\n+ */\n+ pg_free(opts);\n+\n+ exit(EXIT_FAILURE);\nThere is no need to care about the two pg_free() calls here, so let's\nremove all that.\n\n+ pg_free(opts);\n+ pg_free(stats);\n+\n return EXIT_SUCCESS;\n[...]\n bad_argument:\n fprintf(stderr, _(\"Try \\\"%s --help\\\" for more information.\\n\"),\n progname);\n+ pg_free(opts);\n return EXIT_FAILURE;\nSame here.\n\n+static void\n+SignalHandlerForTermination(int signum)\n+{\n+ /* Display the stats only if they are valid */\n+ if (stats &&\n+ !XLogRecPtrIsInvalid(stats_startptr) &&\n+ !XLogRecPtrIsInvalid(stats_endptr))\n+ {\n+ XLogDumpDisplayStats();\n+ pg_free(stats);\n+ }\n+\n+ /*\n+ * Although the pg_free calls (above and below) are unnecessary here on the\n+ * process exit, let's not leak any memory knowingly.\n+ */\n+ pg_free(opts);\n+\n+ exit(EXIT_FAILURE);\n+}\nDoing the dump of the stats followed by the exit() within the signal\nhandler callback is incorrect. The only safe thing that can be set\nand manipulated in a signal handler is roughly a sig_atomic_t flag.\nWhat you should do is to set such a flag in the signal handler, and\nthen check its value in the main loop of pg_waldump, dumping the stats\nif it is detected as turned on because of a signal.\n--\nMichael","msg_date":"Sat, 20 Nov 2021 14:40:13 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Sat, Nov 20, 2021 at 11:10 AM Michael Paquier <michael@paquier.xyz> wrote:\n> What you should do is to set such a flag in the signal handler, and\n> then check its value in the main loop of pg_waldump, dumping the stats\n> if it is detected as turned on because of a signal.\n\nThanks. Here's the v3 patch, a much simpler one. Please review it.\n\nRegards,\nBharath Rupireddy.","msg_date":"Sat, 20 Nov 2021 11:46:35 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Sat, Nov 20, 2021 at 11:46:35AM +0530, Bharath Rupireddy wrote:\n> Thanks. Here's the v3 patch, a much simpler one. Please review it.\n\n+ pqsignal(SIGINT, SignalHandlerForTermination);\n+ pqsignal(SIGTERM, SignalHandlerForTermination);\n+ pqsignal(SIGQUIT, SignalHandlerForTermination);\nFWIW, I think that we should do this stuff only on SIGINT. I would\nimagine that this behavior becomes handy mainly when one wishes to\nCtrl+C the terminal running pg_waldump but still get some\ninformation.\n\n XLogDumpCountRecord(&config, &stats, xlogreader_state);\n+ stats.endptr = xlogreader_state->currRecPtr;\nIsn't what you are looking for here EndRecPtr rather than currRecPtr,\nto track the end of the last record read?\n\n+ When <option>--follow</option> is used with <option>--stats</option> and\n+ the <application>pg_waldump</application> is terminated or interrupted\n+ with signal <systemitem>SIGINT</systemitem> or <systemitem>SIGTERM</systemitem>\n+ or <systemitem>SIGQUIT</systemitem>, the summary statistics computed\n+ as of the termination will be displayed.\nThis description is not completely correct, as the set of stats would\nshow up only by using --stats, in non-quiet mode. Rather than\ndescribing this behavior at the end of the docs, I think that it would\nbe better to add a new paragraph in the description of --stats.\n--\nMichael","msg_date":"Fri, 26 Nov 2021 15:20:08 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Fri, Nov 26, 2021 at 11:50 AM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Sat, Nov 20, 2021 at 11:46:35AM +0530, Bharath Rupireddy wrote:\n> > Thanks. Here's the v3 patch, a much simpler one. Please review it.\n>\n> + pqsignal(SIGINT, SignalHandlerForTermination);\n> + pqsignal(SIGTERM, SignalHandlerForTermination);\n> + pqsignal(SIGQUIT, SignalHandlerForTermination);\n> FWIW, I think that we should do this stuff only on SIGINT. I would\n> imagine that this behavior becomes handy mainly when one wishes to\n> Ctrl+C the terminal running pg_waldump but still get some\n> information.\n\nDone.\n\n> XLogDumpCountRecord(&config, &stats, xlogreader_state);\n> + stats.endptr = xlogreader_state->currRecPtr;\n> Isn't what you are looking for here EndRecPtr rather than currRecPtr,\n> to track the end of the last record read?\n\nYou are right.\n\n> + When <option>--follow</option> is used with <option>--stats</option> and\n> + the <application>pg_waldump</application> is terminated or interrupted\n> + with signal <systemitem>SIGINT</systemitem> or <systemitem>SIGTERM</systemitem>\n> + or <systemitem>SIGQUIT</systemitem>, the summary statistics computed\n> + as of the termination will be displayed.\n> This description is not completely correct, as the set of stats would\n> show up only by using --stats, in non-quiet mode. Rather than\n> describing this behavior at the end of the docs, I think that it would\n> be better to add a new paragraph in the description of --stats.\n\nDone.\n\nPSA v4 patch.\n\nRegards,\nBharath Rupireddy.","msg_date":"Fri, 26 Nov 2021 15:47:30 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Fri, Nov 26, 2021 at 03:47:30PM +0530, Bharath Rupireddy wrote:\n> On Fri, Nov 26, 2021 at 11:50 AM Michael Paquier <michael@paquier.xyz> wrote:\n>> + When <option>--follow</option> is used with <option>--stats</option> and\n>> + the <application>pg_waldump</application> is terminated or interrupted\n>> + with signal <systemitem>SIGINT</systemitem> or <systemitem>SIGTERM</systemitem>\n>> + or <systemitem>SIGQUIT</systemitem>, the summary statistics computed\n>> + as of the termination will be displayed.\n>> This description is not completely correct, as the set of stats would\n>> show up only by using --stats, in non-quiet mode. Rather than\n>> describing this behavior at the end of the docs, I think that it would\n>> be better to add a new paragraph in the description of --stats.\n> \n> Done.\n\nv4 is just moving this paragraph in its correct subsection. My\nwording may have been confusing here, sorry about that. What I meant\nis that there is no need to mention --follow at all, as an\ninterruption done before reaching the end LSN or the end of WAL would\nstill print out the stats accumulated up to the interrupted point.\n--\nMichael","msg_date":"Sun, 28 Nov 2021 13:19:55 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Sun, Nov 28, 2021 at 9:50 AM Michael Paquier <michael@paquier.xyz> wrote:\n> v4 is just moving this paragraph in its correct subsection. My\n> wording may have been confusing here, sorry about that. What I meant\n> is that there is no need to mention --follow at all, as an\n> interruption done before reaching the end LSN or the end of WAL would\n> still print out the stats accumulated up to the interrupted point.\n\nThanks. Here's the v5.\n\nRegards,\nBharath Rupireddy.","msg_date":"Sun, 28 Nov 2021 12:13:08 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Sun, Nov 28, 2021 at 12:13:08PM +0530, Bharath Rupireddy wrote:\n> Thanks. Here's the v5.\n\nBy the way, one thing that I completely forgot here is that SIGINT is\nnot handled on Windows. If we want to make that work for a WIN32\nterminal, we would need to do something similar to\nsrc/fe_utils/cancel.c where we need to use SetConsoleCtrlHandler() and\nhandle the stats print when facing CTRL_C_EVENT or CTRL_BREAK_EVENT as\nevents. Perhaps we should try to think harder and have a more\ncentralized facility for the handler part between a WIN32 terminal and\nSIGINT, as it is not the first time that we need this level of\nhandling. Or we could just discard this issue, document its WIN32\nlimitation and paint some \"#ifdef WIN32\" around all the handler\nportions of the patch.\n\nI would be fine with just doing the latter for now, as this stuff is\nstill useful for most users, but that's worth mentioning. Any\nopinions?\n--\nMichael","msg_date":"Mon, 29 Nov 2021 14:39:51 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Mon, Nov 29, 2021 at 11:09 AM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Sun, Nov 28, 2021 at 12:13:08PM +0530, Bharath Rupireddy wrote:\n> > Thanks. Here's the v5.\n>\n> By the way, one thing that I completely forgot here is that SIGINT is\n> not handled on Windows. If we want to make that work for a WIN32\n> terminal, we would need to do something similar to\n> src/fe_utils/cancel.c where we need to use SetConsoleCtrlHandler() and\n> handle the stats print when facing CTRL_C_EVENT or CTRL_BREAK_EVENT as\n> events. Perhaps we should try to think harder and have a more\n> centralized facility for the handler part between a WIN32 terminal and\n> SIGINT, as it is not the first time that we need this level of\n> handling. Or we could just discard this issue, document its WIN32\n> limitation and paint some \"#ifdef WIN32\" around all the handler\n> portions of the patch.\n>\n> I would be fine with just doing the latter for now, as this stuff is\n> still useful for most users, but that's worth mentioning. Any\n> opinions?\n\nI'm okay to have the same behaviour as pg_receivewal and pg_recvlogical tools.\n\nOn my Windows 64-bit setup (see below), the CTRL+C works and the\nsummary stats are printed as with Linux. I believe on 32-bit platforms\nwe don't have CTRL+C with SIGINT handling right? If yes, I'm not sure\nhow we go about mentioning it in the pg_waldump documentation. I see\nthat the pg_receivewal and pg_recvlogical don't mention anything in\nthe documentation even though their SIGINT handler is defined for\n#ifndef WIN32 platforms.\npostgres=# select version();\n version\n---------------------------------------------------------------\n PostgreSQL 15devel, compiled by Visual C++ build 1916, 64-bit\n(1 row)\npostgres=#\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Mon, 29 Nov 2021 13:16:13 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Mon, Nov 29, 2021 at 1:16 PM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n>\n> On Mon, Nov 29, 2021 at 11:09 AM Michael Paquier <michael@paquier.xyz> wrote:\n> >\n> > On Sun, Nov 28, 2021 at 12:13:08PM +0530, Bharath Rupireddy wrote:\n> > > Thanks. Here's the v5.\n> >\n> > By the way, one thing that I completely forgot here is that SIGINT is\n> > not handled on Windows. If we want to make that work for a WIN32\n> > terminal, we would need to do something similar to\n> > src/fe_utils/cancel.c where we need to use SetConsoleCtrlHandler() and\n> > handle the stats print when facing CTRL_C_EVENT or CTRL_BREAK_EVENT as\n> > events. Perhaps we should try to think harder and have a more\n> > centralized facility for the handler part between a WIN32 terminal and\n> > SIGINT, as it is not the first time that we need this level of\n> > handling. Or we could just discard this issue, document its WIN32\n> > limitation and paint some \"#ifdef WIN32\" around all the handler\n> > portions of the patch.\n> >\n> > I would be fine with just doing the latter for now, as this stuff is\n> > still useful for most users, but that's worth mentioning. Any\n> > opinions?\n>\n> I'm okay to have the same behaviour as pg_receivewal and pg_recvlogical tools.\n\nHere's the v6 patch that has the SIGINT handling enabled for non-WIN32\nplatforms (note that I haven't specified anything in the\ndocumentation) much like pg_receivewal and pg_recvlogical.\n\nRegards,\nBharath Rupireddy.","msg_date":"Wed, 1 Dec 2021 11:44:02 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"},{"msg_contents":"On Wed, Dec 01, 2021 at 11:44:02AM +0530, Bharath Rupireddy wrote:\n> Here's the v6 patch that has the SIGINT handling enabled for non-WIN32\n> platforms (note that I haven't specified anything in the\n> documentation) much like pg_receivewal and pg_recvlogical.\n\nI have added a safeguard in XLogDumpDisplayStats() to not print any\nstats if the end LSN is not set yet, which would be possible if\npg_waldump is signaled at a very early stage, and I have added some\nmore comments. That looked fine after that, so applied.\n--\nMichael","msg_date":"Thu, 2 Dec 2021 13:57:44 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_waldump stucks with options --follow or -f and --stats or -z"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\npg_waldump options, --follow or -f(to keep polling once per second for\\nnew WAL to appear) and --stats or -z don't work well together i.e. the\\ncommand stucks [1]. I think the pg_waldump should emit an error. Note\\nthat the pg_basebakup does error out on incompatible options.\\n\\nHere's a small patch for fixing above along with a note in the documentation.\\n\\nThoughts?\\n\\n[1] The following commands stuck:\\n./pg_waldump -p data/ -s 0/7000060 -f -z\\n./pg_waldump -p data/ -s 0/7000060 -f --stats='record'\\n./pg_waldump -p data/ -s 0/7000060 -f --stats='rmgr'\\n\\n[2]\\nubuntu3:~/postgres/inst/bin$ ./pg_waldump -p data/ -s 0/7000060 -f -z\\npg_waldump: error: --follow and --stats are incompatible options\\nTry \\\"pg_waldump --help\\\" for more information.\\nbharath@bharathubuntu3:~/postgres/inst/bin$\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Mon, 15 Nov 2021 19:32:38 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 07:32:38PM +0530, Bharath Rupireddy wrote:\\n> pg_waldump options, --follow or -f(to keep polling once per second for\\n> new WAL to appear) and --stats or -z don't work well together i.e. the\\n> command stucks [1]. I think the pg_waldump should emit an error. Note\\n> that the pg_basebakup does error out on incompatible options.\\n> \\n> Here's a small patch for fixing above along with a note in the documentation.\\n> \\n> Thoughts?\\n\\nI don't think that we should block this combination of options as you\\nare proposing. The existing behavior is useful for users when it\\ncomes to an end position specified with -e, to be able to gather some\\nstats on a cluster or an archive where we may not have all the\\ncontents wanted yet.\\n \\n> [1] The following commands stuck:\\n> ./pg_waldump -p data/ -s 0/7000060 -f -z\\n> ./pg_waldump -p data/ -s 0/7000060 -f --stats='record'\\n> ./pg_waldump -p data/ -s 0/7000060 -f --stats='rmgr'\\n\\nSaying that, you are not completely wrong either, as following\\nsomething while we won't print any stats at all is not really\\nhelpful. Another thing I can think of here is to make pg_waldump\\nmore responsive to the dump of the stats when interrupted, via\\nXLogDumpDisplayStats().\\n--\\nMichael\",\n \"msg_date\": \"Tue, 16 Nov 2021 11:56:01 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 8:26 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Mon, Nov 15, 2021 at 07:32:38PM +0530, Bharath Rupireddy wrote:\\n> > pg_waldump options, --follow or -f(to keep polling once per second for\\n> > new WAL to appear) and --stats or -z don't work well together i.e. the\\n> > command stucks [1]. I think the pg_waldump should emit an error. Note\\n> > that the pg_basebakup does error out on incompatible options.\\n> >\\n> > Here's a small patch for fixing above along with a note in the documentation.\\n> >\\n> > Thoughts?\\n>\\n> I don't think that we should block this combination of options as you\\n> are proposing. The existing behavior is useful for users when it\\n> comes to an end position specified with -e, to be able to gather some\\n> stats on a cluster or an archive where we may not have all the\\n> contents wanted yet.\\n\\nYou are right. The \\\"./pg_waldump -p data/ -s 0/14CC090 -e 0/14CC390 -f\\n-z\\\" would fail with what I proposed which isn't a good thing at all.\\nShould we block both the combinations \\\"-s/-f/-z\\\", \\\"-s/-e/-f/-z\\\"?\\n\\n> > [1] The following commands stuck:\\n> > ./pg_waldump -p data/ -s 0/7000060 -f -z\\n> > ./pg_waldump -p data/ -s 0/7000060 -f --stats='record'\\n> > ./pg_waldump -p data/ -s 0/7000060 -f --stats='rmgr'\\n>\\n> Saying that, you are not completely wrong either, as following\\n> something while we won't print any stats at all is not really\\n> helpful. Another thing I can think of here is to make pg_waldump\\n> more responsive to the dump of the stats when interrupted, via\\n> XLogDumpDisplayStats().\\n\\nIt looks okay to be more responsive with the -f option so that the\\nabove commands keep emitting stats for every 1 second(the --follow\\noption waits for 1 second). Should we really be emitting stats for\\nevery 1 second? Isn't there going to be too much information on the\\nstdout? Or should we be emitting the stats for every 5 or 10 seconds?\\n\\nIf we be more responsive and keep emitting stats per 1 or 5 or 10 etc.\\nseconds(we can give an option to the user to choose when he/she would\\nlike to see the stats with -f option, but this is going to be an\\noverkill IMO), can the user make anything out of those loads of stats\\ngetting emitted? Another idea is to give some incremental stats but it\\nis overkill again IMO. And when the pg_waldump has the capability to\\nemit the stats, why to block it?\\n\\nIn summary, we have the following options:\\n1) block the combinations \\\"-s/-f/-z\\\", \\\"-s/-e/-f/-z\\\"\\n2) be more responsive and keep emitting the stats per 1 sec default\\nwith -f option\\n3) be more responsive and keep emitting the stats per user's choice\\nof seconds (a new option that can be used with the -s/-e/-f/-z).\\n\\nThoughts?\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 09:34:25 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 09:34:25AM +0530, Bharath Rupireddy wrote:\\n> It looks okay to be more responsive with the -f option so that the\\n> above commands keep emitting stats for every 1 second(the --follow\\n> option waits for 1 second). Should we really be emitting stats for\\n> every 1 second? Isn't there going to be too much information on the\\n> stdout? Or should we be emitting the stats for every 5 or 10 seconds?\\n\\nI don't have a perfect answer to this question, but dumping the stats\\nwith a fixed frequency is not going to be useful if we don't have in\\nthose logs a current timestamp and/or a current LSN. This is\\nbasically about how much we want to call XLogDumpDisplayStats() and\\nhow useful it is, but note that my argument is a bit different than\\nwhat you are describing here: we could try to make\\nXLogDumpDisplayStats() responsive on SIGINT or SIGTERM to show\\nstatistics reports. This way, a --follow without an --end LSN\\nspecified could still provide some information rather than nothing.\\nThat could also be useful if defining an --end but interrupting the\\ncall.\\n\\nAt the same time, we could also just let things as they are. --follow\\nand --stats being specified together is what the user looked for, so\\nthey get what they wanted.\\n\\n> In summary, we have the following options:\\n> 1) block the combinations \\\"-s/-f/-z\\\", \\\"-s/-e/-f/-z\\\"\\n> 2) be more responsive and keep emitting the stats per 1 sec default\\n> with -f option\\n> 3) be more responsive and keep emitting the stats per user's choice\\n> of seconds (a new option that can be used with the -s/-e/-f/-z).\\n\\nA frequency cannot be measured only in time here, but also in bytes in\\nterms of a minimum amount of WAL replayed between two reports. I\\ndon't like much the idea of multiple stats reports emitted in a single\\npg_waldump call, TBH. This makes things more complicated than they\\nshould, and the gain is not obvious, at least to me.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 17 Nov 2021 11:19:52 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 7:49 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> At the same time, we could also just let things as they are. --follow\\n> and --stats being specified together is what the user looked for, so\\n> they get what they wanted.\\n\\nI think the existing way of pg_waldump getting stuck with the\\ncombination of options \\\"-s/-f/-z\\\" and \\\"-s/-e/-f/-z\\\" doesn't look good\\nfor an end user. In the simplest way, the pg_waldump can just emit an\\nerror for these combinations. That shouldn't a big problem as we do\\nemit errors for a lot of mutually exclusive combination of options\\nother SQL statements/commands.\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 22:24:18 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-17, Bharath Rupireddy wrote:\\n\\n> On Wed, Nov 17, 2021 at 7:49 AM Michael Paquier <michael@paquier.xyz> wrote:\\n> >\\n> > At the same time, we could also just let things as they are. --follow\\n> > and --stats being specified together is what the user looked for, so\\n> > they get what they wanted.\\n> \\n> I think the existing way of pg_waldump getting stuck with the\\n> combination of options \\\"-s/-f/-z\\\" and \\\"-s/-e/-f/-z\\\" doesn't look good\\n> for an end user.\\n\\nI agree that it's not right, but I don't think the solution is to ban\\nthe combination. I think what we should do is change the behavior to\\nmake the combination potentially useful, so that it waits until program\\ntermination and print the summary then. Consider \\\"strace -c\\\" as a\\nprecedent: it will also \\\"become stuck\\\" until you kill it, and then it'll\\nprint a nice table, just like the pg_waldump -z gives you.\\n\\nI think I would even have had occasion to use this as a feature in the\\npast ...\\n\\n-- \\nÁlvaro Herrera 39°49'30\\\"S 73°17'W — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 15:35:03 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 12:05 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>\\n> On 2021-Nov-17, Bharath Rupireddy wrote:\\n>\\n> > On Wed, Nov 17, 2021 at 7:49 AM Michael Paquier <michael@paquier.xyz> wrote:\\n> > >\\n> > > At the same time, we could also just let things as they are. --follow\\n> > > and --stats being specified together is what the user looked for, so\\n> > > they get what they wanted.\\n> >\\n> > I think the existing way of pg_waldump getting stuck with the\\n> > combination of options \\\"-s/-f/-z\\\" and \\\"-s/-e/-f/-z\\\" doesn't look good\\n> > for an end user.\\n>\\n> I agree that it's not right, but I don't think the solution is to ban\\n> the combination. I think what we should do is change the behavior to\\n> make the combination potentially useful, so that it waits until program\\n> termination and print the summary then. Consider \\\"strace -c\\\" as a\\n> precedent: it will also \\\"become stuck\\\" until you kill it, and then it'll\\n> print a nice table, just like the pg_waldump -z gives you.\\n>\\n> I think I would even have had occasion to use this as a feature in the\\n> past ...\\n\\nHm. So, the pg_waldump can have handlers for SIGINT, SIGTERM, SIGQUIT\\nand then it should emit the computed stats in those handlers the\\ncomobinations - \\\"-s/-f/-z\\\" and \\\"-s/-e/-f/-z\\\". I'm okay with this\\nbehaviour. Michael Paquier had suggested the same upthread. If okay, I\\nwill work on that patch.\\n\\nThoughts?\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 13:32:08 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 01:32:08PM +0530, Bharath Rupireddy wrote:\\n> Hm. So, the pg_waldump can have handlers for SIGINT, SIGTERM, SIGQUIT\\n> and then it should emit the computed stats in those handlers the\\n> comobinations - \\\"-s/-f/-z\\\" and \\\"-s/-e/-f/-z\\\". I'm okay with this\\n> behaviour. Michael Paquier had suggested the same upthread. If okay, I\\n> will work on that patch.\\n\\nWhile on it, I am pretty sure that we should add in the stats report\\nthe start and end LSNs matching with the reports. If you use --follow\\nand the call is interrupted, we would not really know to which part of\\nthe WAL stream a report applies to without this information, likely in\\nthe shape of an extra header.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 18 Nov 2021 17:21:47 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 1:51 PM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Thu, Nov 18, 2021 at 01:32:08PM +0530, Bharath Rupireddy wrote:\\n> > Hm. So, the pg_waldump can have handlers for SIGINT, SIGTERM, SIGQUIT\\n> > and then it should emit the computed stats in those handlers the\\n> > comobinations - \\\"-s/-f/-z\\\" and \\\"-s/-e/-f/-z\\\". I'm okay with this\\n> > behaviour. Michael Paquier had suggested the same upthread. If okay, I\\n> > will work on that patch.\\n>\\n> While on it, I am pretty sure that we should add in the stats report\\n> the start and end LSNs matching with the reports. If you use --follow\\n> and the call is interrupted, we would not really know to which part of\\n> the WAL stream a report applies to without this information, likely in\\n> the shape of an extra header.\\n\\nCurrently, the SIGINT, SIGTERM, SIGQUIT interruptions will cause the\\npg_waldump process with --follow option to get killed. And it is a good\\nidea to specify the range of LSNs (start and end) upto which the stats are\\ncomputed. Maybe as another printf message at the end of the stats report,\\nafter the \\\"Total\\\" section or at the beginning of the stats report.\\nSomething like \\\"summary of the statistics computed from << LSN >> to\\n<<LSN>> are:\\\", see [1].\\n\\n[1] ubuntu3:~/postgres/src/bin/pg_waldump$ ./pg_waldump -p data -s\\n0/1480000 --stats\\nsummary of the statistics computed from 0/1480000 and 0/158ABCD are:\\nType N (%) Record\\nsize (%) FPI size (%) Combined size (%)\\n---- - ---\\n ----------- --- -------- --- -------------\\n ---\\nXLOG 13 ( 0.13)\\n1138 ( 0.18) 8424 ( 2.46) 9562 ( 0.99)\\nTransaction 14 ( 0.14)\\n1407 ( 0.22) 0 ( 0.00) 1407 ( 0.15)\\nStorage 4 ( 0.04)\\n 172 ( 0.03) 0 ( 0.00) 172 ( 0.02)\\nCLOG 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nDatabase 1 ( 0.01)\\n42 ( 0.01) 0 ( 0.00) 42 ( 0.00)\\nTablespace 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nMultiXact 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nRelMap 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nStandby 37 ( 0.36)\\n2318 ( 0.37) 0 ( 0.00) 2318 ( 0.24)\\nHeap2 74 ( 0.73)\\n 12841 ( 2.05) 121184 ( 35.46) 134025 ( 13.84)\\nHeap 9005 ( 88.23)\\n532313 ( 84.91) 84208 ( 24.64) 616521 ( 63.64)\\nBtree 1058 ( 10.37)\\n 76710 ( 12.24) 127932 ( 37.43) 204642 ( 21.13)\\nHash 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nGin 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nGist 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nSequence 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nSPGist 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nBRIN 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nCommitTs 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nReplicationOrigin 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nGeneric 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\nLogicalMessage 0 ( 0.00)\\n 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)\\n --------\\n -------- -------- --------\\nTotal 10206\\n 626941 [64.72%] 341748 [35.28%] 968689 [100%]\\n\\nRegards,\\nBharath Rupireddy.\\n\\nOn Thu, Nov 18, 2021 at 1:51 PM Michael Paquier <michael@paquier.xyz> wrote:>> On Thu, Nov 18, 2021 at 01:32:08PM +0530, Bharath Rupireddy wrote:> > Hm. So, the pg_waldump can have handlers for SIGINT, SIGTERM, SIGQUIT> > and then it should emit the computed stats in those handlers the> > comobinations - \\\"-s/-f/-z\\\" and \\\"-s/-e/-f/-z\\\". I'm okay with this> > behaviour. Michael Paquier had suggested the same upthread. If okay, I> > will work on that patch.>> While on it, I am pretty sure that we should add in the stats report> the start and end LSNs matching with the reports. If you use --follow> and the call is interrupted, we would not really know to which part of> the WAL stream a report applies to without this information, likely in> the shape of an extra header.Currently, the SIGINT, SIGTERM, SIGQUIT interruptions will cause the pg_waldump process with --follow option to get killed. And it is a good idea to specify the range of LSNs (start and end) upto which the stats are computed. Maybe as another printf message at the end of the stats report, after the \\\"Total\\\" section or at the beginning of the stats report. Something like \\\"summary of the statistics computed from << LSN >> to <<LSN>> are:\\\", see [1].[1] ubuntu3:~/postgres/src/bin/pg_waldump$ ./pg_waldump -p data -s 0/1480000 --statssummary of the statistics computed from 0/1480000 and 0/158ABCD are:Type N (%) Record size (%) FPI size (%) Combined size (%)---- - --- ----------- --- -------- --- ------------- ---XLOG 13 ( 0.13) 1138 ( 0.18) 8424 ( 2.46) 9562 ( 0.99)Transaction 14 ( 0.14) 1407 ( 0.22) 0 ( 0.00) 1407 ( 0.15)Storage 4 ( 0.04) 172 ( 0.03) 0 ( 0.00) 172 ( 0.02)CLOG 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Database 1 ( 0.01) 42 ( 0.01) 0 ( 0.00) 42 ( 0.00)Tablespace 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)MultiXact 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)RelMap 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Standby 37 ( 0.36) 2318 ( 0.37) 0 ( 0.00) 2318 ( 0.24)Heap2 74 ( 0.73) 12841 ( 2.05) 121184 ( 35.46) 134025 ( 13.84)Heap 9005 ( 88.23) 532313 ( 84.91) 84208 ( 24.64) 616521 ( 63.64)Btree 1058 ( 10.37) 76710 ( 12.24) 127932 ( 37.43) 204642 ( 21.13)Hash 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Gin 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Gist 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Sequence 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)SPGist 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)BRIN 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)CommitTs 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)ReplicationOrigin 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)Generic 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00)LogicalMessage 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) 0 ( 0.00) -------- -------- -------- --------Total 10206 626941 [64.72%] 341748 [35.28%] 968689 [100%]Regards,Bharath Rupireddy.\",\n \"msg_date\": \"Thu, 18 Nov 2021 14:03:20 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 2:03 PM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n>\\n> On Thu, Nov 18, 2021 at 1:51 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> >\\n> > On Thu, Nov 18, 2021 at 01:32:08PM +0530, Bharath Rupireddy wrote:\\n> > > Hm. So, the pg_waldump can have handlers for SIGINT, SIGTERM, SIGQUIT\\n> > > and then it should emit the computed stats in those handlers the\\n> > > comobinations - \\\"-s/-f/-z\\\" and \\\"-s/-e/-f/-z\\\". I'm okay with this\\n> > > behaviour. Michael Paquier had suggested the same upthread. If okay, I\\n> > > will work on that patch.\\n> >\\n> > While on it, I am pretty sure that we should add in the stats report\\n> > the start and end LSNs matching with the reports. If you use --follow\\n> > and the call is interrupted, we would not really know to which part of\\n> > the WAL stream a report applies to without this information, likely in\\n> > the shape of an extra header.\\n>\\n> Currently, the SIGINT, SIGTERM, SIGQUIT interruptions will cause the pg_waldump process with --follow option to get killed. And it is a good idea to specify the range of LSNs (start and end) upto which the stats are computed. Maybe as another printf message at the end of the stats report, after the \\\"Total\\\" section or at the beginning of the stats report. Something like \\\"summary of the statistics computed from << LSN >> to <<LSN>> are:\\\", see [1].\\n\\nHere's the v2 patch with the above changes i.e. pg_waldump now eimits\\nthe computed stats at the time of termination. Please review it.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Sat, 20 Nov 2021 10:03:27 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 20, 2021 at 10:03:27AM +0530, Bharath Rupireddy wrote:\\n> Here's the v2 patch with the above changes i.e. pg_waldump now eimits\\n> the computed stats at the time of termination. Please review it.\\n\\n+#define MAX_XLINFO_TYPES 16\\n+#define fatal_error(...) do { pg_log_fatal(__VA_ARGS__);\\nexit(EXIT_FAILURE); } while(0)\\n typedef struct XLogDumpPrivate\\nMoving around those declarations does not relate to this patch, so I\\nwould recommend to leave this out to ease reviews.\\n\\n+ /*\\n+ * Although the pg_free calls (above and below) are unnecessary here on the\\n+ * process exit, let's not leak any memory knowingly.\\n+ */\\n+ pg_free(opts);\\n+\\n+ exit(EXIT_FAILURE);\\nThere is no need to care about the two pg_free() calls here, so let's\\nremove all that.\\n\\n+ pg_free(opts);\\n+ pg_free(stats);\\n+\\n return EXIT_SUCCESS;\\n[...]\\n bad_argument:\\n fprintf(stderr, _(\\\"Try \\\\\\\"%s --help\\\\\\\" for more information.\\\\n\\\"),\\n progname);\\n+ pg_free(opts);\\n return EXIT_FAILURE;\\nSame here.\\n\\n+static void\\n+SignalHandlerForTermination(int signum)\\n+{\\n+ /* Display the stats only if they are valid */\\n+ if (stats &&\\n+ !XLogRecPtrIsInvalid(stats_startptr) &&\\n+ !XLogRecPtrIsInvalid(stats_endptr))\\n+ {\\n+ XLogDumpDisplayStats();\\n+ pg_free(stats);\\n+ }\\n+\\n+ /*\\n+ * Although the pg_free calls (above and below) are unnecessary here on the\\n+ * process exit, let's not leak any memory knowingly.\\n+ */\\n+ pg_free(opts);\\n+\\n+ exit(EXIT_FAILURE);\\n+}\\nDoing the dump of the stats followed by the exit() within the signal\\nhandler callback is incorrect. The only safe thing that can be set\\nand manipulated in a signal handler is roughly a sig_atomic_t flag.\\nWhat you should do is to set such a flag in the signal handler, and\\nthen check its value in the main loop of pg_waldump, dumping the stats\\nif it is detected as turned on because of a signal.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 20 Nov 2021 14:40:13 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 20, 2021 at 11:10 AM Michael Paquier <michael@paquier.xyz> wrote:\\n> What you should do is to set such a flag in the signal handler, and\\n> then check its value in the main loop of pg_waldump, dumping the stats\\n> if it is detected as turned on because of a signal.\\n\\nThanks. Here's the v3 patch, a much simpler one. Please review it.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Sat, 20 Nov 2021 11:46:35 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 20, 2021 at 11:46:35AM +0530, Bharath Rupireddy wrote:\\n> Thanks. Here's the v3 patch, a much simpler one. Please review it.\\n\\n+ pqsignal(SIGINT, SignalHandlerForTermination);\\n+ pqsignal(SIGTERM, SignalHandlerForTermination);\\n+ pqsignal(SIGQUIT, SignalHandlerForTermination);\\nFWIW, I think that we should do this stuff only on SIGINT. I would\\nimagine that this behavior becomes handy mainly when one wishes to\\nCtrl+C the terminal running pg_waldump but still get some\\ninformation.\\n\\n XLogDumpCountRecord(&config, &stats, xlogreader_state);\\n+ stats.endptr = xlogreader_state->currRecPtr;\\nIsn't what you are looking for here EndRecPtr rather than currRecPtr,\\nto track the end of the last record read?\\n\\n+ When <option>--follow</option> is used with <option>--stats</option> and\\n+ the <application>pg_waldump</application> is terminated or interrupted\\n+ with signal <systemitem>SIGINT</systemitem> or <systemitem>SIGTERM</systemitem>\\n+ or <systemitem>SIGQUIT</systemitem>, the summary statistics computed\\n+ as of the termination will be displayed.\\nThis description is not completely correct, as the set of stats would\\nshow up only by using --stats, in non-quiet mode. Rather than\\ndescribing this behavior at the end of the docs, I think that it would\\nbe better to add a new paragraph in the description of --stats.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 26 Nov 2021 15:20:08 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 26, 2021 at 11:50 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Sat, Nov 20, 2021 at 11:46:35AM +0530, Bharath Rupireddy wrote:\\n> > Thanks. Here's the v3 patch, a much simpler one. Please review it.\\n>\\n> + pqsignal(SIGINT, SignalHandlerForTermination);\\n> + pqsignal(SIGTERM, SignalHandlerForTermination);\\n> + pqsignal(SIGQUIT, SignalHandlerForTermination);\\n> FWIW, I think that we should do this stuff only on SIGINT. I would\\n> imagine that this behavior becomes handy mainly when one wishes to\\n> Ctrl+C the terminal running pg_waldump but still get some\\n> information.\\n\\nDone.\\n\\n> XLogDumpCountRecord(&config, &stats, xlogreader_state);\\n> + stats.endptr = xlogreader_state->currRecPtr;\\n> Isn't what you are looking for here EndRecPtr rather than currRecPtr,\\n> to track the end of the last record read?\\n\\nYou are right.\\n\\n> + When <option>--follow</option> is used with <option>--stats</option> and\\n> + the <application>pg_waldump</application> is terminated or interrupted\\n> + with signal <systemitem>SIGINT</systemitem> or <systemitem>SIGTERM</systemitem>\\n> + or <systemitem>SIGQUIT</systemitem>, the summary statistics computed\\n> + as of the termination will be displayed.\\n> This description is not completely correct, as the set of stats would\\n> show up only by using --stats, in non-quiet mode. Rather than\\n> describing this behavior at the end of the docs, I think that it would\\n> be better to add a new paragraph in the description of --stats.\\n\\nDone.\\n\\nPSA v4 patch.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Fri, 26 Nov 2021 15:47:30 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 26, 2021 at 03:47:30PM +0530, Bharath Rupireddy wrote:\\n> On Fri, Nov 26, 2021 at 11:50 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>> + When <option>--follow</option> is used with <option>--stats</option> and\\n>> + the <application>pg_waldump</application> is terminated or interrupted\\n>> + with signal <systemitem>SIGINT</systemitem> or <systemitem>SIGTERM</systemitem>\\n>> + or <systemitem>SIGQUIT</systemitem>, the summary statistics computed\\n>> + as of the termination will be displayed.\\n>> This description is not completely correct, as the set of stats would\\n>> show up only by using --stats, in non-quiet mode. Rather than\\n>> describing this behavior at the end of the docs, I think that it would\\n>> be better to add a new paragraph in the description of --stats.\\n> \\n> Done.\\n\\nv4 is just moving this paragraph in its correct subsection. My\\nwording may have been confusing here, sorry about that. What I meant\\nis that there is no need to mention --follow at all, as an\\ninterruption done before reaching the end LSN or the end of WAL would\\nstill print out the stats accumulated up to the interrupted point.\\n--\\nMichael\",\n \"msg_date\": \"Sun, 28 Nov 2021 13:19:55 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 28, 2021 at 9:50 AM Michael Paquier <michael@paquier.xyz> wrote:\\n> v4 is just moving this paragraph in its correct subsection. My\\n> wording may have been confusing here, sorry about that. What I meant\\n> is that there is no need to mention --follow at all, as an\\n> interruption done before reaching the end LSN or the end of WAL would\\n> still print out the stats accumulated up to the interrupted point.\\n\\nThanks. Here's the v5.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Sun, 28 Nov 2021 12:13:08 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 28, 2021 at 12:13:08PM +0530, Bharath Rupireddy wrote:\\n> Thanks. Here's the v5.\\n\\nBy the way, one thing that I completely forgot here is that SIGINT is\\nnot handled on Windows. If we want to make that work for a WIN32\\nterminal, we would need to do something similar to\\nsrc/fe_utils/cancel.c where we need to use SetConsoleCtrlHandler() and\\nhandle the stats print when facing CTRL_C_EVENT or CTRL_BREAK_EVENT as\\nevents. Perhaps we should try to think harder and have a more\\ncentralized facility for the handler part between a WIN32 terminal and\\nSIGINT, as it is not the first time that we need this level of\\nhandling. Or we could just discard this issue, document its WIN32\\nlimitation and paint some \\\"#ifdef WIN32\\\" around all the handler\\nportions of the patch.\\n\\nI would be fine with just doing the latter for now, as this stuff is\\nstill useful for most users, but that's worth mentioning. Any\\nopinions?\\n--\\nMichael\",\n \"msg_date\": \"Mon, 29 Nov 2021 14:39:51 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 29, 2021 at 11:09 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Sun, Nov 28, 2021 at 12:13:08PM +0530, Bharath Rupireddy wrote:\\n> > Thanks. Here's the v5.\\n>\\n> By the way, one thing that I completely forgot here is that SIGINT is\\n> not handled on Windows. If we want to make that work for a WIN32\\n> terminal, we would need to do something similar to\\n> src/fe_utils/cancel.c where we need to use SetConsoleCtrlHandler() and\\n> handle the stats print when facing CTRL_C_EVENT or CTRL_BREAK_EVENT as\\n> events. Perhaps we should try to think harder and have a more\\n> centralized facility for the handler part between a WIN32 terminal and\\n> SIGINT, as it is not the first time that we need this level of\\n> handling. Or we could just discard this issue, document its WIN32\\n> limitation and paint some \\\"#ifdef WIN32\\\" around all the handler\\n> portions of the patch.\\n>\\n> I would be fine with just doing the latter for now, as this stuff is\\n> still useful for most users, but that's worth mentioning. Any\\n> opinions?\\n\\nI'm okay to have the same behaviour as pg_receivewal and pg_recvlogical tools.\\n\\nOn my Windows 64-bit setup (see below), the CTRL+C works and the\\nsummary stats are printed as with Linux. I believe on 32-bit platforms\\nwe don't have CTRL+C with SIGINT handling right? If yes, I'm not sure\\nhow we go about mentioning it in the pg_waldump documentation. I see\\nthat the pg_receivewal and pg_recvlogical don't mention anything in\\nthe documentation even though their SIGINT handler is defined for\\n#ifndef WIN32 platforms.\\npostgres=# select version();\\n version\\n---------------------------------------------------------------\\n PostgreSQL 15devel, compiled by Visual C++ build 1916, 64-bit\\n(1 row)\\npostgres=#\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Nov 2021 13:16:13 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 29, 2021 at 1:16 PM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n>\\n> On Mon, Nov 29, 2021 at 11:09 AM Michael Paquier <michael@paquier.xyz> wrote:\\n> >\\n> > On Sun, Nov 28, 2021 at 12:13:08PM +0530, Bharath Rupireddy wrote:\\n> > > Thanks. Here's the v5.\\n> >\\n> > By the way, one thing that I completely forgot here is that SIGINT is\\n> > not handled on Windows. If we want to make that work for a WIN32\\n> > terminal, we would need to do something similar to\\n> > src/fe_utils/cancel.c where we need to use SetConsoleCtrlHandler() and\\n> > handle the stats print when facing CTRL_C_EVENT or CTRL_BREAK_EVENT as\\n> > events. Perhaps we should try to think harder and have a more\\n> > centralized facility for the handler part between a WIN32 terminal and\\n> > SIGINT, as it is not the first time that we need this level of\\n> > handling. Or we could just discard this issue, document its WIN32\\n> > limitation and paint some \\\"#ifdef WIN32\\\" around all the handler\\n> > portions of the patch.\\n> >\\n> > I would be fine with just doing the latter for now, as this stuff is\\n> > still useful for most users, but that's worth mentioning. Any\\n> > opinions?\\n>\\n> I'm okay to have the same behaviour as pg_receivewal and pg_recvlogical tools.\\n\\nHere's the v6 patch that has the SIGINT handling enabled for non-WIN32\\nplatforms (note that I haven't specified anything in the\\ndocumentation) much like pg_receivewal and pg_recvlogical.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Wed, 1 Dec 2021 11:44:02 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 01, 2021 at 11:44:02AM +0530, Bharath Rupireddy wrote:\\n> Here's the v6 patch that has the SIGINT handling enabled for non-WIN32\\n> platforms (note that I haven't specified anything in the\\n> documentation) much like pg_receivewal and pg_recvlogical.\\n\\nI have added a safeguard in XLogDumpDisplayStats() to not print any\\nstats if the end LSN is not set yet, which would be possible if\\npg_waldump is signaled at a very early stage, and I have added some\\nmore comments. That looked fine after that, so applied.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 2 Dec 2021 13:57:44 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump stucks with options --follow or -f and --stats or -z\"\n }\n]","isConversational":false}}},{"rowIdx":53092,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nRobert and I wondered whether we would like to rename the startup\n\nprocess. The reason for doing this is that the current name doesn't\n\nmake any sense, as in the stand-by mode replay loop as the main\n\nloop, the startup process stays around.\n\n\nI think a better name for the process may be “recovery” or “WAL_replay”\n\nor maybe someone has a better idea. One can say that startup is a very\n\ngeneric name, so in the future, it gives flexibility in terms of assigning\n\nany sort of 'work’. But in standby mode, the name “startup” doesn’t\n\nmake sense.\n\n\nOpen for suggestions and thoughts.\n\n\n\nRegards,\nRushabh Lathia\nwww.EnterpriseDB.com\n\nHi,Robert and I wondered whether we would like to rename the startupprocess. The reason for doing this is that the current name doesn'tmake any sense, as in the stand-by mode replay loop as the mainloop, the startup process stays around.I think a better name for the process may be “recovery” or “WAL_replay”or maybe someone has a better idea. One can say that startup is a verygeneric name, so in the future, it gives flexibility in terms of assigningany sort of 'work’. But in standby mode, the name “startup” doesn’tmake sense.Open for suggestions and thoughts.Regards,Rushabh Lathiawww.EnterpriseDB.com","msg_date":"Mon, 15 Nov 2021 20:02:59 +0530","msg_from":"Rushabh Lathia <rushabh.lathia@gmail.com>","msg_from_op":true,"msg_subject":"Should rename \"startup process\" to something else?"},{"msg_contents":"On Mon, Nov 15, 2021 at 8:03 PM Rushabh Lathia <rushabh.lathia@gmail.com> wrote:\n> Robert and I wondered whether we would like to rename the startup\n>\n> process. The reason for doing this is that the current name doesn't\n>\n> make any sense, as in the stand-by mode replay loop as the main\n>\n> loop, the startup process stays around.\n\nThe startup process main function looks to be to do some\ninitialization required for recovery and do the recovery, exit if it\nis a crash recovery or stay there if it is a standby recovery. Apart\nfrom these I'm not sure it does any non-recovery tasks. Does the\nstartup process have any relevance or work to do if the server state\nis DB_SHUTDOWNED i.e the server was shutdown properly? I think yes,\nalthough it doesn't enter the recovery/REDO loop, it does a bunch of\nother things like updating the control file, pre-allocating WAL files,\nupdating shared memory state to allow other backends to write WAL and\nso on. By keeping this in mind, just renaming the startup process to\n\"recovery\" or some other doesn't look good IMO. Instead, we can live\nwith it and be happy with the ps display update that we do in\nXLogFileRead.\n\n> I think a better name for the process may be “recovery” or “WAL_replay”\n>\n> or maybe someone has a better idea. One can say that startup is a very\n>\n> generic name, so in the future, it gives flexibility in terms of assigning\n>\n> any sort of 'work’. But in standby mode, the name “startup” doesn’t\n>\n> make sense.\n>\n> Open for suggestions and thoughts.\n\nIf at all, we were to change the startup process to \"recovery\" or some\nother, it's not going to be a tiny change, we need to change the code\nwith some user facing error message, docs, testsm which isn't good\nfrom code maintainability and especially the diff it creates with the\nback branches that we support.\n\nOthers may have better thoughts.\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Thu, 18 Nov 2021 16:04:37 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"\n\n> 15 нояб. 2021 г., в 19:32, Rushabh Lathia <rushabh.lathia@gmail.com> написал(а):\n> \n> Open for suggestions and thoughts.\n\n\nHow about walapplier ?\nSimilar to walsender, walreciver..\n\nBest regards, Andrey Borodin.\n\n","msg_date":"Thu, 18 Nov 2021 17:35:59 +0500","msg_from":"Andrey Borodin <x4mmm@yandex-team.ru>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On Thu, Nov 18, 2021 at 6:06 PM Andrey Borodin <x4mmm@yandex-team.ru> wrote:\n>\n>\n>\n> > 15 нояб. 2021 г., в 19:32, Rushabh Lathia <rushabh.lathia@gmail.com> написал(а):\n> >\n> > Open for suggestions and thoughts.\n>\n>\n> How about walapplier ?\n> Similar to walsender, walreciver..\n>\nOr maybe walreplayer ?\n\nRegards,\nAmul\n\n\n","msg_date":"Thu, 18 Nov 2021 18:09:07 +0530","msg_from":"Amul Sul <sulamul@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On Thu, Nov 18, 2021 at 5:34 AM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n> The startup process main function looks to be to do some\n> initialization required for recovery and do the recovery, exit if it\n> is a crash recovery or stay there if it is a standby recovery. Apart\n> from these I'm not sure it does any non-recovery tasks. Does the\n> startup process have any relevance or work to do if the server state\n> is DB_SHUTDOWNED i.e the server was shutdown properly? I think yes,\n> although it doesn't enter the recovery/REDO loop, it does a bunch of\n> other things like updating the control file, pre-allocating WAL files,\n> updating shared memory state to allow other backends to write WAL and\n> so on.\n\nThat's true, but those tasks are very brief. Nobody's going to get too\nconfused by a \"recovery\" process that shows up for a few milliseconds\nwhen there's no recovery to be done. Having a \"startup\" process that\nsticks around forever on a standy, though, actually is confusing.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 18 Nov 2021 09:17:50 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> That's true, but those tasks are very brief. Nobody's going to get too\n> confused by a \"recovery\" process that shows up for a few milliseconds\n> when there's no recovery to be done. Having a \"startup\" process that\n> sticks around forever on a standy, though, actually is confusing.\n\nYeah, given current usage it would be better to call it the \"recovery\nprocess\". However, I'm feeling dubious that it's worth the cost to\nchange. The \"startup\" name is embedded in a lot of places, I think,\nand people are used to it. I fear changing it would create more\nconfusion than it removes.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 18 Nov 2021 11:05:04 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On Thu, Nov 18, 2021 at 11:05 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Yeah, given current usage it would be better to call it the \"recovery\n> process\". However, I'm feeling dubious that it's worth the cost to\n> change. The \"startup\" name is embedded in a lot of places, I think,\n> and people are used to it. I fear changing it would create more\n> confusion than it removes.\n\nWhat sorts of places are you thinking about?\n\nAs far as being used to it, I think hackers are, but regular users are\nvery much not. And there are a lot more regular users than there are\nhackers, and it's harder to get them to pay attention, too.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 18 Nov 2021 12:08:18 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Thu, Nov 18, 2021 at 11:05 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> Yeah, given current usage it would be better to call it the \"recovery\n>> process\". However, I'm feeling dubious that it's worth the cost to\n>> change. The \"startup\" name is embedded in a lot of places, I think,\n>> and people are used to it. I fear changing it would create more\n>> confusion than it removes.\n\n> What sorts of places are you thinking about?\n\nAside from our own code, I imagine a lot of people have monitoring\nscripts that know about it.\n\n> As far as being used to it, I think hackers are, but regular users are\n> very much not.\n\nBeing hackers ourselves, I'm not sure we're qualified to opine on\nthat. I cannot say that I've noticed any questions about it on\nthe mailing lists, though.\n\nPersonally I think making a glossary entry that explains what the\nprocess does would be a better plan than renaming it.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 18 Nov 2021 12:24:14 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On Tue, Nov 16, 2021 at 3:33 AM Rushabh Lathia <rushabh.lathia@gmail.com> wrote:\n> I think a better name for the process may be “recovery”\n\n+1\n\n\n","msg_date":"Fri, 19 Nov 2021 06:48:13 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On Thu, Nov 18, 2021 at 12:24:14PM -0500, Tom Lane wrote:\n> Robert Haas <robertmhaas@gmail.com> writes:\n> > On Thu, Nov 18, 2021 at 11:05 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >> Yeah, given current usage it would be better to call it the \"recovery\n> >> process\". However, I'm feeling dubious that it's worth the cost to\n> >> change. The \"startup\" name is embedded in a lot of places, I think,\n> >> and people are used to it. I fear changing it would create more\n> >> confusion than it removes.\n> \n> > As far as being used to it, I think hackers are, but regular users are\n> > very much not.\n> \n> Being hackers ourselves, I'm not sure we're qualified to opine on\n> that. I cannot say that I've noticed any questions about it on\n> the mailing lists, though.\n\nA data point: I was recently confused when I observed the \"startup\" process\nrunning for a bit after restarting the instance (because connections were being\nrejected) I concluded that the shutdown was unclean, and started to blame the\nPGDG RPM's initscript [0].\n\nActually, the shutdown was clean, and the \"startup\" process was just slow doing\n$somethingelse (I imagine this will be less confusing in pg15 - 9ce346eabf).\n\n[0] I believe this is configured such that systemd could kill -9 the postmaster\n(but that's not what happened in this case).\nhttps://redmine.postgresql.org/issues/6855\n\nIf you rename \"startup\", I think \"recovery\" would be a bad choice, since it\nseems to imply that recovery/wal replay was necessary.\n\n> Personally I think making a glossary entry that explains what the\n> process does would be a better plan than renaming it.\n\nSince d3014fff4:\nhttps://www.postgresql.org/docs/devel/glossary.html#GLOSSARY-STARTUP-PROCESS\n\n-- \nJustin\n\n\n","msg_date":"Thu, 18 Nov 2021 11:58:48 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"Justin Pryzby <pryzby@telsasoft.com> writes:\n> On Thu, Nov 18, 2021 at 12:24:14PM -0500, Tom Lane wrote:\n>> Personally I think making a glossary entry that explains what the\n>> process does would be a better plan than renaming it.\n\n> Since d3014fff4:\n> https://www.postgresql.org/docs/devel/glossary.html#GLOSSARY-STARTUP-PROCESS\n\nHm ... I might've found that, if it were in alphabetical order ...\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 18 Nov 2021 13:07:09 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On Thu, Nov 18, 2021 at 12:24 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Being hackers ourselves, I'm not sure we're qualified to opine on\n> that. I cannot say that I've noticed any questions about it on\n> the mailing lists, though.\n\nWhat I've noticed when supporting EDB customer is that if, say,\nthere's a problem with recovery, and they're trying to figure out\nwhich process is responsible, they have no intuition that the startup\nprocess is the likely culprit. They could, for example, notice whether\nthe LSN that shows up in the ps status is advancing over time. Or,\nthey could notice whether that process is doing a lot of I/O, or a lot\nof CPU. But since they have no notion that a startup process has\nanything to do with recovery, they don't make that connection. Now you\ncan argue that I ought to be happy about that because, hey, it's job\nsecurity. And you can also argue that even if the process had a better\nname, a lot of people wouldn't figure it out for one reason or\nanother. However, my view is that we do well to make things more\ncomprehensible when we can.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 18 Nov 2021 14:06:38 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On 2021-Nov-18, Tom Lane wrote:\n\n> Justin Pryzby <pryzby@telsasoft.com> writes:\n> > On Thu, Nov 18, 2021 at 12:24:14PM -0500, Tom Lane wrote:\n> >> Personally I think making a glossary entry that explains what the\n> >> process does would be a better plan than renaming it.\n> \n> > Since d3014fff4:\n> > https://www.postgresql.org/docs/devel/glossary.html#GLOSSARY-STARTUP-PROCESS\n> \n> Hm ... I might've found that, if it were in alphabetical order ...\n\nUgh, my mistake. If the consensus is not to rename, I'll move it to the\nright place.\n\nIf we change the name, and I support the idea that we do, I think a\ngood name would be \"wal replay\". I think \"recovery\" is not great\nprecisely because in a standby there is likely no crash that we're\nrecovering from. The word \"replay\" is at odds with the other names,\nwhich stand for the device that carries out the task at hand\n(checkpointer, bgwriter, wal sender/receiver); but the word \"replayer\"\nseems to be extremely uncommon and IMO looks strange. If you see a\nprocess that claims to be \"wal replay\", you know perfectly well what it\nis.\n\nThe glossary entry reads:\n\n\tAn auxiliary process that replays WAL during crash recovery and in a\n\tphysical replica.\n\n\t(The name is historical: the startup process was named before\n\treplication was implemented; the name refers to its task as it relates\n\tto the server startup following a crash.)\n\nIf we rename, we can drop more than half of the entry and replace it\nwith \"This process was previously known as the startup process\".\n\n... oh, I noticed another mistake: \"WAL receiver\" does not say\n\"(process)\", like all other process entries do.\n\n-- \nÁlvaro Herrera 39°49'30\"S 73°17'W — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Thu, 18 Nov 2021 16:11:52 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> If we change the name, and I support the idea that we do, I think a\n> good name would be \"wal replay\". I think \"recovery\" is not great\n> precisely because in a standby there is likely no crash that we're\n> recovering from.\n\nFair point.\n\n> The word \"replay\" is at odds with the other names,\n> which stand for the device that carries out the task at hand\n> (checkpointer, bgwriter, wal sender/receiver); but the word \"replayer\"\n> seems to be extremely uncommon and IMO looks strange. If you see a\n> process that claims to be \"wal replay\", you know perfectly well what it\n> is.\n\nI'm less concerned about the \"er\" than about the fact that the name is\ntwo words. People will immediately shorten it to just \"replay\", eg\nas a part of names in the code, and I feel that that's confusing in\nits own way. Maybe we could run the words together, on the precedent\nof \"walreceiver\", but I never much liked that name either.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 18 Nov 2021 15:22:41 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On 11/18/21, 12:24 PM, \"Tom Lane\" <tgl@sss.pgh.pa.us> wrote:\r\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\r\n>> If we change the name, and I support the idea that we do, I think a\r\n>> good name would be \"wal replay\". I think \"recovery\" is not great\r\n>> precisely because in a standby there is likely no crash that we're\r\n>> recovering from.\r\n>\r\n> Fair point.\r\n>\r\n>> The word \"replay\" is at odds with the other names,\r\n>> which stand for the device that carries out the task at hand\r\n>> (checkpointer, bgwriter, wal sender/receiver); but the word \"replayer\"\r\n>> seems to be extremely uncommon and IMO looks strange. If you see a\r\n>> process that claims to be \"wal replay\", you know perfectly well what it\r\n>> is.\r\n>\r\n> I'm less concerned about the \"er\" than about the fact that the name is\r\n> two words. People will immediately shorten it to just \"replay\", eg\r\n> as a part of names in the code, and I feel that that's confusing in\r\n> its own way. Maybe we could run the words together, on the precedent\r\n> of \"walreceiver\", but I never much liked that name either.\r\n\r\n+1 to something like \"wal replay\" or \"wal apply.\" My view is that\r\nterms like \"startup\" and \"recovery\" indicate the goal of the process\r\nwhile \"replay\" and \"apply\" just explain what it does. This would be\r\nin line with the other processes (e.g., the WAL receiver receives WAL,\r\nthe checkpointer checkpoints, and the archiver archives). I don't\r\nhave any strong opinion about it being two words, but maybe I am just\r\nconditioned from seeing walreceiver, walsender, and walwriter so\r\noften.\r\n\r\nNathan\r\n\r\n","msg_date":"Fri, 19 Nov 2021 23:45:40 +0000","msg_from":"\"Bossart, Nathan\" <bossartn@amazon.com>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On Fri, Nov 19, 2021 at 11:45:40PM +0000, Bossart, Nathan wrote:\n> +1 to something like \"wal replay\" or \"wal apply.\" My view is that\n> terms like \"startup\" and \"recovery\" indicate the goal of the process\n> while \"replay\" and \"apply\" just explain what it does. This would be\n> in line with the other processes (e.g., the WAL receiver receives WAL,\n> the checkpointer checkpoints, and the archiver archives). I don't\n> have any strong opinion about it being two words, but maybe I am just\n> conditioned from seeing walreceiver, walsender, and walwriter so\n> often.\n\nIf we want to play the card of consistency with the existing names,\nthat would be \"walreplayer\", then. Jokes apart, \"WAL replay\" sounds\nfine to me :) \n\nI would be tempted to do more than that, though, bite the bullet and\ngo as far as renaming walsender to \"WAL sender\", \"walreceiver\" to \"WAL\nreceiver\", and \"walwriter\" to \"WAL writer\", even if that means\npotentially breaking scripts monitoring ps outputs.\n--\nMichael","msg_date":"Sat, 20 Nov 2021 14:50:40 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"\nOn 11/18/21 15:22, Tom Lane wrote:\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n>> If we change the name, and I support the idea that we do, I think a\n>> good name would be \"wal replay\". I think \"recovery\" is not great\n>> precisely because in a standby there is likely no crash that we're\n>> recovering from.\n> Fair point.\n>\n>> The word \"replay\" is at odds with the other names,\n>> which stand for the device that carries out the task at hand\n>> (checkpointer, bgwriter, wal sender/receiver); but the word \"replayer\"\n>> seems to be extremely uncommon and IMO looks strange. If you see a\n>> process that claims to be \"wal replay\", you know perfectly well what it\n>> is.\n> I'm less concerned about the \"er\" than about the fact that the name is\n> two words. People will immediately shorten it to just \"replay\", eg\n> as a part of names in the code, and I feel that that's confusing in\n> its own way. Maybe we could run the words together, on the precedent\n> of \"walreceiver\", but I never much liked that name either.\n>\n> \t\t\t\n\n\n\nMaybe something along those lines but using a dash/hyphen would work:\ne.g. wal-replayer\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Sat, 20 Nov 2021 08:56:42 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On 2021-Nov-20, Andrew Dunstan wrote:\n\n> Maybe something along those lines but using a dash/hyphen would work:\n> e.g. wal-replayer\n\nYeah, the idea of a dash occurred to me too.\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\"La grandeza es una experiencia transitoria. Nunca es consistente.\nDepende en gran parte de la imaginación humana creadora de mitos\"\n(Irulan)\n\n\n","msg_date":"Tue, 23 Nov 2021 11:41:36 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> On 2021-Nov-20, Andrew Dunstan wrote:\n>> Maybe something along those lines but using a dash/hyphen would work:\n>> e.g. wal-replayer\n\n> Yeah, the idea of a dash occurred to me too.\n\nBad memories of uuid-ossp float up ... can we use underscore?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 23 Nov 2021 10:04:08 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On 2021-Nov-23, Tom Lane wrote:\n\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> > On 2021-Nov-20, Andrew Dunstan wrote:\n> >> Maybe something along those lines but using a dash/hyphen would work:\n> >> e.g. wal-replayer\n> \n> > Yeah, the idea of a dash occurred to me too.\n> \n> Bad memories of uuid-ossp float up ... can we use underscore?\n\nI don't object to an underscore, but it looks a bit uglier to me.\nAFAIK the main problem with uuid-ossp was that it is used as an identifier, so\nit required quoting, which won't be the case with this process name.\n\n\nAnyway, as a second and separate point, should we rename the other WAL-related\nprocess names to use the same character?\n\nwalreceiver -> wal_receiver / wal-receiver\nwalsender -> wal_sender / wal-sender\nwalwriter -> wal_writer / wal-writer.\n\nI assume we would *not* rename other existing processes that already\nhave spaces in their names, such as \"autovacuum launcher\" to\n\"autovacuum_launcher\", as that would be a net loss IMO.\n\nCurrently, the process list in an idle primary server looks like \n\nUSER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND\nalvherre 1091970 0.0 0.0 199608 22252 pts/0 S+ 12:22 0:00 /pgsql/install/master/bin/postmaster -p 55432\nalvherre 1091983 0.0 0.0 199608 3644 ? Ss 12:22 0:00 \\_ postgres: checkpointer \nalvherre 1091984 0.0 0.0 199608 5432 ? Ss 12:22 0:00 \\_ postgres: background writer \nalvherre 1091986 0.0 0.0 199608 8956 ? Ss 12:22 0:00 \\_ postgres: walwriter \nalvherre 1091987 0.0 0.0 200148 7988 ? Ss 12:22 0:00 \\_ postgres: autovacuum launcher \nalvherre 1091988 0.0 0.0 54432 4156 ? Ss 12:22 0:00 \\_ postgres: stats collector \nalvherre 1091989 0.0 0.0 200036 6416 ? Ss 12:22 0:00 \\_ postgres: logical replication launcher \n\nI think this looks better:\n\nUSER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND\nalvherre 1091970 0.0 0.0 199608 22252 pts/0 S+ 12:22 0:00 /pgsql/install/master/bin/postmaster -p 55432\nalvherre 1091983 0.0 0.0 199608 3644 ? Ss 12:22 0:00 \\_ postgres: checkpointer \nalvherre 1091984 0.0 0.0 199608 5432 ? Ss 12:22 0:00 \\_ postgres: background writer \nalvherre 1091986 0.0 0.0 199608 8956 ? Ss 12:22 0:00 \\_ postgres: wal-writer \nalvherre 1091987 0.0 0.0 200148 7988 ? Ss 12:22 0:00 \\_ postgres: autovacuum launcher \nalvherre 1091988 0.0 0.0 54432 4156 ? Ss 12:22 0:00 \\_ postgres: stats collector \nalvherre 1091989 0.0 0.0 200036 6416 ? Ss 12:22 0:00 \\_ postgres: logical replication launcher \n\n-- \nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Wed, 24 Nov 2021 15:54:37 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> On 2021-Nov-23, Tom Lane wrote:\n>> Bad memories of uuid-ossp float up ... can we use underscore?\n\n> I don't object to an underscore, but it looks a bit uglier to me.\n> AFAIK the main problem with uuid-ossp was that it is used as an identifier, so\n> it required quoting, which won't be the case with this process name.\n\nWell, we can call the process whatever we want, but we will need to\nderive identifiers from that name, so you can't just ignore the issue.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 24 Nov 2021 14:07:34 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"},{"msg_contents":"On Wed, Nov 24, 2021 at 1:54 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> I don't object to an underscore, but it looks a bit uglier to me.\n> AFAIK the main problem with uuid-ossp was that it is used as an identifier, so\n> it required quoting, which won't be the case with this process name.\n\nI agree, and for that reason I would prefer no separator, or a space.\nThat's what we do with other processes, and I think it's fine.\n\nIt's worth thinking too about the fact that we may want to rename\nfunctions, adjust comments, etc. Each of those things has their own\nconventions. For example consider StartupProcessMain(). If we decide\nto call this the WAL replay process, I suppose that is going to become\nWALReplayMain(). For sure it's not going to be come WAL-Replay-Main().\nBut what displays in the 'ps' status should look like what we do in\nother cases.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 24 Nov 2021 14:10:00 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Should rename \"startup process\" to something else?"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nRobert and I wondered whether we would like to rename the startup\\n\\nprocess. The reason for doing this is that the current name doesn't\\n\\nmake any sense, as in the stand-by mode replay loop as the main\\n\\nloop, the startup process stays around.\\n\\n\\nI think a better name for the process may be “recovery” or “WAL_replay”\\n\\nor maybe someone has a better idea. One can say that startup is a very\\n\\ngeneric name, so in the future, it gives flexibility in terms of assigning\\n\\nany sort of 'work’. But in standby mode, the name “startup” doesn’t\\n\\nmake sense.\\n\\n\\nOpen for suggestions and thoughts.\\n\\n\\n\\nRegards,\\nRushabh Lathia\\nwww.EnterpriseDB.com\\n\\nHi,Robert and I wondered whether we would like to rename the startupprocess. The reason for doing this is that the current name doesn'tmake any sense, as in the stand-by mode replay loop as the mainloop, the startup process stays around.I think a better name for the process may be “recovery” or “WAL_replay”or maybe someone has a better idea. One can say that startup is a verygeneric name, so in the future, it gives flexibility in terms of assigningany sort of 'work’. But in standby mode, the name “startup” doesn’tmake sense.Open for suggestions and thoughts.Regards,Rushabh Lathiawww.EnterpriseDB.com\",\n \"msg_date\": \"Mon, 15 Nov 2021 20:02:59 +0530\",\n \"msg_from\": \"Rushabh Lathia <rushabh.lathia@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 8:03 PM Rushabh Lathia <rushabh.lathia@gmail.com> wrote:\\n> Robert and I wondered whether we would like to rename the startup\\n>\\n> process. The reason for doing this is that the current name doesn't\\n>\\n> make any sense, as in the stand-by mode replay loop as the main\\n>\\n> loop, the startup process stays around.\\n\\nThe startup process main function looks to be to do some\\ninitialization required for recovery and do the recovery, exit if it\\nis a crash recovery or stay there if it is a standby recovery. Apart\\nfrom these I'm not sure it does any non-recovery tasks. Does the\\nstartup process have any relevance or work to do if the server state\\nis DB_SHUTDOWNED i.e the server was shutdown properly? I think yes,\\nalthough it doesn't enter the recovery/REDO loop, it does a bunch of\\nother things like updating the control file, pre-allocating WAL files,\\nupdating shared memory state to allow other backends to write WAL and\\nso on. By keeping this in mind, just renaming the startup process to\\n\\\"recovery\\\" or some other doesn't look good IMO. Instead, we can live\\nwith it and be happy with the ps display update that we do in\\nXLogFileRead.\\n\\n> I think a better name for the process may be “recovery” or “WAL_replay”\\n>\\n> or maybe someone has a better idea. One can say that startup is a very\\n>\\n> generic name, so in the future, it gives flexibility in terms of assigning\\n>\\n> any sort of 'work’. But in standby mode, the name “startup” doesn’t\\n>\\n> make sense.\\n>\\n> Open for suggestions and thoughts.\\n\\nIf at all, we were to change the startup process to \\\"recovery\\\" or some\\nother, it's not going to be a tiny change, we need to change the code\\nwith some user facing error message, docs, testsm which isn't good\\nfrom code maintainability and especially the diff it creates with the\\nback branches that we support.\\n\\nOthers may have better thoughts.\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 16:04:37 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"\\n\\n> 15 нояб. 2021 г., в 19:32, Rushabh Lathia <rushabh.lathia@gmail.com> написал(а):\\n> \\n> Open for suggestions and thoughts.\\n\\n\\nHow about walapplier ?\\nSimilar to walsender, walreciver..\\n\\nBest regards, Andrey Borodin.\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 17:35:59 +0500\",\n \"msg_from\": \"Andrey Borodin <x4mmm@yandex-team.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 6:06 PM Andrey Borodin <x4mmm@yandex-team.ru> wrote:\\n>\\n>\\n>\\n> > 15 нояб. 2021 г., в 19:32, Rushabh Lathia <rushabh.lathia@gmail.com> написал(а):\\n> >\\n> > Open for suggestions and thoughts.\\n>\\n>\\n> How about walapplier ?\\n> Similar to walsender, walreciver..\\n>\\nOr maybe walreplayer ?\\n\\nRegards,\\nAmul\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 18:09:07 +0530\",\n \"msg_from\": \"Amul Sul <sulamul@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 5:34 AM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n> The startup process main function looks to be to do some\\n> initialization required for recovery and do the recovery, exit if it\\n> is a crash recovery or stay there if it is a standby recovery. Apart\\n> from these I'm not sure it does any non-recovery tasks. Does the\\n> startup process have any relevance or work to do if the server state\\n> is DB_SHUTDOWNED i.e the server was shutdown properly? I think yes,\\n> although it doesn't enter the recovery/REDO loop, it does a bunch of\\n> other things like updating the control file, pre-allocating WAL files,\\n> updating shared memory state to allow other backends to write WAL and\\n> so on.\\n\\nThat's true, but those tasks are very brief. Nobody's going to get too\\nconfused by a \\\"recovery\\\" process that shows up for a few milliseconds\\nwhen there's no recovery to be done. Having a \\\"startup\\\" process that\\nsticks around forever on a standy, though, actually is confusing.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 09:17:50 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> That's true, but those tasks are very brief. Nobody's going to get too\\n> confused by a \\\"recovery\\\" process that shows up for a few milliseconds\\n> when there's no recovery to be done. Having a \\\"startup\\\" process that\\n> sticks around forever on a standy, though, actually is confusing.\\n\\nYeah, given current usage it would be better to call it the \\\"recovery\\nprocess\\\". However, I'm feeling dubious that it's worth the cost to\\nchange. The \\\"startup\\\" name is embedded in a lot of places, I think,\\nand people are used to it. I fear changing it would create more\\nconfusion than it removes.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 11:05:04 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 11:05 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Yeah, given current usage it would be better to call it the \\\"recovery\\n> process\\\". However, I'm feeling dubious that it's worth the cost to\\n> change. The \\\"startup\\\" name is embedded in a lot of places, I think,\\n> and people are used to it. I fear changing it would create more\\n> confusion than it removes.\\n\\nWhat sorts of places are you thinking about?\\n\\nAs far as being used to it, I think hackers are, but regular users are\\nvery much not. And there are a lot more regular users than there are\\nhackers, and it's harder to get them to pay attention, too.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 12:08:18 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Thu, Nov 18, 2021 at 11:05 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Yeah, given current usage it would be better to call it the \\\"recovery\\n>> process\\\". However, I'm feeling dubious that it's worth the cost to\\n>> change. The \\\"startup\\\" name is embedded in a lot of places, I think,\\n>> and people are used to it. I fear changing it would create more\\n>> confusion than it removes.\\n\\n> What sorts of places are you thinking about?\\n\\nAside from our own code, I imagine a lot of people have monitoring\\nscripts that know about it.\\n\\n> As far as being used to it, I think hackers are, but regular users are\\n> very much not.\\n\\nBeing hackers ourselves, I'm not sure we're qualified to opine on\\nthat. I cannot say that I've noticed any questions about it on\\nthe mailing lists, though.\\n\\nPersonally I think making a glossary entry that explains what the\\nprocess does would be a better plan than renaming it.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 12:24:14 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 3:33 AM Rushabh Lathia <rushabh.lathia@gmail.com> wrote:\\n> I think a better name for the process may be “recovery”\\n\\n+1\\n\\n\\n\",\n \"msg_date\": \"Fri, 19 Nov 2021 06:48:13 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 12:24:14PM -0500, Tom Lane wrote:\\n> Robert Haas <robertmhaas@gmail.com> writes:\\n> > On Thu, Nov 18, 2021 at 11:05 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >> Yeah, given current usage it would be better to call it the \\\"recovery\\n> >> process\\\". However, I'm feeling dubious that it's worth the cost to\\n> >> change. The \\\"startup\\\" name is embedded in a lot of places, I think,\\n> >> and people are used to it. I fear changing it would create more\\n> >> confusion than it removes.\\n> \\n> > As far as being used to it, I think hackers are, but regular users are\\n> > very much not.\\n> \\n> Being hackers ourselves, I'm not sure we're qualified to opine on\\n> that. I cannot say that I've noticed any questions about it on\\n> the mailing lists, though.\\n\\nA data point: I was recently confused when I observed the \\\"startup\\\" process\\nrunning for a bit after restarting the instance (because connections were being\\nrejected) I concluded that the shutdown was unclean, and started to blame the\\nPGDG RPM's initscript [0].\\n\\nActually, the shutdown was clean, and the \\\"startup\\\" process was just slow doing\\n$somethingelse (I imagine this will be less confusing in pg15 - 9ce346eabf).\\n\\n[0] I believe this is configured such that systemd could kill -9 the postmaster\\n(but that's not what happened in this case).\\nhttps://redmine.postgresql.org/issues/6855\\n\\nIf you rename \\\"startup\\\", I think \\\"recovery\\\" would be a bad choice, since it\\nseems to imply that recovery/wal replay was necessary.\\n\\n> Personally I think making a glossary entry that explains what the\\n> process does would be a better plan than renaming it.\\n\\nSince d3014fff4:\\nhttps://www.postgresql.org/docs/devel/glossary.html#GLOSSARY-STARTUP-PROCESS\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 11:58:48 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"Justin Pryzby <pryzby@telsasoft.com> writes:\\n> On Thu, Nov 18, 2021 at 12:24:14PM -0500, Tom Lane wrote:\\n>> Personally I think making a glossary entry that explains what the\\n>> process does would be a better plan than renaming it.\\n\\n> Since d3014fff4:\\n> https://www.postgresql.org/docs/devel/glossary.html#GLOSSARY-STARTUP-PROCESS\\n\\nHm ... I might've found that, if it were in alphabetical order ...\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 13:07:09 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 12:24 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Being hackers ourselves, I'm not sure we're qualified to opine on\\n> that. I cannot say that I've noticed any questions about it on\\n> the mailing lists, though.\\n\\nWhat I've noticed when supporting EDB customer is that if, say,\\nthere's a problem with recovery, and they're trying to figure out\\nwhich process is responsible, they have no intuition that the startup\\nprocess is the likely culprit. They could, for example, notice whether\\nthe LSN that shows up in the ps status is advancing over time. Or,\\nthey could notice whether that process is doing a lot of I/O, or a lot\\nof CPU. But since they have no notion that a startup process has\\nanything to do with recovery, they don't make that connection. Now you\\ncan argue that I ought to be happy about that because, hey, it's job\\nsecurity. And you can also argue that even if the process had a better\\nname, a lot of people wouldn't figure it out for one reason or\\nanother. However, my view is that we do well to make things more\\ncomprehensible when we can.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 14:06:38 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-18, Tom Lane wrote:\\n\\n> Justin Pryzby <pryzby@telsasoft.com> writes:\\n> > On Thu, Nov 18, 2021 at 12:24:14PM -0500, Tom Lane wrote:\\n> >> Personally I think making a glossary entry that explains what the\\n> >> process does would be a better plan than renaming it.\\n> \\n> > Since d3014fff4:\\n> > https://www.postgresql.org/docs/devel/glossary.html#GLOSSARY-STARTUP-PROCESS\\n> \\n> Hm ... I might've found that, if it were in alphabetical order ...\\n\\nUgh, my mistake. If the consensus is not to rename, I'll move it to the\\nright place.\\n\\nIf we change the name, and I support the idea that we do, I think a\\ngood name would be \\\"wal replay\\\". I think \\\"recovery\\\" is not great\\nprecisely because in a standby there is likely no crash that we're\\nrecovering from. The word \\\"replay\\\" is at odds with the other names,\\nwhich stand for the device that carries out the task at hand\\n(checkpointer, bgwriter, wal sender/receiver); but the word \\\"replayer\\\"\\nseems to be extremely uncommon and IMO looks strange. If you see a\\nprocess that claims to be \\\"wal replay\\\", you know perfectly well what it\\nis.\\n\\nThe glossary entry reads:\\n\\n\\tAn auxiliary process that replays WAL during crash recovery and in a\\n\\tphysical replica.\\n\\n\\t(The name is historical: the startup process was named before\\n\\treplication was implemented; the name refers to its task as it relates\\n\\tto the server startup following a crash.)\\n\\nIf we rename, we can drop more than half of the entry and replace it\\nwith \\\"This process was previously known as the startup process\\\".\\n\\n... oh, I noticed another mistake: \\\"WAL receiver\\\" does not say\\n\\\"(process)\\\", like all other process entries do.\\n\\n-- \\nÁlvaro Herrera 39°49'30\\\"S 73°17'W — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 16:11:52 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> If we change the name, and I support the idea that we do, I think a\\n> good name would be \\\"wal replay\\\". I think \\\"recovery\\\" is not great\\n> precisely because in a standby there is likely no crash that we're\\n> recovering from.\\n\\nFair point.\\n\\n> The word \\\"replay\\\" is at odds with the other names,\\n> which stand for the device that carries out the task at hand\\n> (checkpointer, bgwriter, wal sender/receiver); but the word \\\"replayer\\\"\\n> seems to be extremely uncommon and IMO looks strange. If you see a\\n> process that claims to be \\\"wal replay\\\", you know perfectly well what it\\n> is.\\n\\nI'm less concerned about the \\\"er\\\" than about the fact that the name is\\ntwo words. People will immediately shorten it to just \\\"replay\\\", eg\\nas a part of names in the code, and I feel that that's confusing in\\nits own way. Maybe we could run the words together, on the precedent\\nof \\\"walreceiver\\\", but I never much liked that name either.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 15:22:41 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On 11/18/21, 12:24 PM, \\\"Tom Lane\\\" <tgl@sss.pgh.pa.us> wrote:\\r\\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\r\\n>> If we change the name, and I support the idea that we do, I think a\\r\\n>> good name would be \\\"wal replay\\\". I think \\\"recovery\\\" is not great\\r\\n>> precisely because in a standby there is likely no crash that we're\\r\\n>> recovering from.\\r\\n>\\r\\n> Fair point.\\r\\n>\\r\\n>> The word \\\"replay\\\" is at odds with the other names,\\r\\n>> which stand for the device that carries out the task at hand\\r\\n>> (checkpointer, bgwriter, wal sender/receiver); but the word \\\"replayer\\\"\\r\\n>> seems to be extremely uncommon and IMO looks strange. If you see a\\r\\n>> process that claims to be \\\"wal replay\\\", you know perfectly well what it\\r\\n>> is.\\r\\n>\\r\\n> I'm less concerned about the \\\"er\\\" than about the fact that the name is\\r\\n> two words. People will immediately shorten it to just \\\"replay\\\", eg\\r\\n> as a part of names in the code, and I feel that that's confusing in\\r\\n> its own way. Maybe we could run the words together, on the precedent\\r\\n> of \\\"walreceiver\\\", but I never much liked that name either.\\r\\n\\r\\n+1 to something like \\\"wal replay\\\" or \\\"wal apply.\\\" My view is that\\r\\nterms like \\\"startup\\\" and \\\"recovery\\\" indicate the goal of the process\\r\\nwhile \\\"replay\\\" and \\\"apply\\\" just explain what it does. This would be\\r\\nin line with the other processes (e.g., the WAL receiver receives WAL,\\r\\nthe checkpointer checkpoints, and the archiver archives). I don't\\r\\nhave any strong opinion about it being two words, but maybe I am just\\r\\nconditioned from seeing walreceiver, walsender, and walwriter so\\r\\noften.\\r\\n\\r\\nNathan\\r\\n\\r\\n\",\n \"msg_date\": \"Fri, 19 Nov 2021 23:45:40 +0000\",\n \"msg_from\": \"\\\"Bossart, Nathan\\\" <bossartn@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 19, 2021 at 11:45:40PM +0000, Bossart, Nathan wrote:\\n> +1 to something like \\\"wal replay\\\" or \\\"wal apply.\\\" My view is that\\n> terms like \\\"startup\\\" and \\\"recovery\\\" indicate the goal of the process\\n> while \\\"replay\\\" and \\\"apply\\\" just explain what it does. This would be\\n> in line with the other processes (e.g., the WAL receiver receives WAL,\\n> the checkpointer checkpoints, and the archiver archives). I don't\\n> have any strong opinion about it being two words, but maybe I am just\\n> conditioned from seeing walreceiver, walsender, and walwriter so\\n> often.\\n\\nIf we want to play the card of consistency with the existing names,\\nthat would be \\\"walreplayer\\\", then. Jokes apart, \\\"WAL replay\\\" sounds\\nfine to me :) \\n\\nI would be tempted to do more than that, though, bite the bullet and\\ngo as far as renaming walsender to \\\"WAL sender\\\", \\\"walreceiver\\\" to \\\"WAL\\nreceiver\\\", and \\\"walwriter\\\" to \\\"WAL writer\\\", even if that means\\npotentially breaking scripts monitoring ps outputs.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 20 Nov 2021 14:50:40 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"\\nOn 11/18/21 15:22, Tom Lane wrote:\\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n>> If we change the name, and I support the idea that we do, I think a\\n>> good name would be \\\"wal replay\\\". I think \\\"recovery\\\" is not great\\n>> precisely because in a standby there is likely no crash that we're\\n>> recovering from.\\n> Fair point.\\n>\\n>> The word \\\"replay\\\" is at odds with the other names,\\n>> which stand for the device that carries out the task at hand\\n>> (checkpointer, bgwriter, wal sender/receiver); but the word \\\"replayer\\\"\\n>> seems to be extremely uncommon and IMO looks strange. If you see a\\n>> process that claims to be \\\"wal replay\\\", you know perfectly well what it\\n>> is.\\n> I'm less concerned about the \\\"er\\\" than about the fact that the name is\\n> two words. People will immediately shorten it to just \\\"replay\\\", eg\\n> as a part of names in the code, and I feel that that's confusing in\\n> its own way. Maybe we could run the words together, on the precedent\\n> of \\\"walreceiver\\\", but I never much liked that name either.\\n>\\n> \\t\\t\\t\\n\\n\\n\\nMaybe something along those lines but using a dash/hyphen would work:\\ne.g. wal-replayer\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 20 Nov 2021 08:56:42 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-20, Andrew Dunstan wrote:\\n\\n> Maybe something along those lines but using a dash/hyphen would work:\\n> e.g. wal-replayer\\n\\nYeah, the idea of a dash occurred to me too.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\\"La grandeza es una experiencia transitoria. Nunca es consistente.\\nDepende en gran parte de la imaginación humana creadora de mitos\\\"\\n(Irulan)\\n\\n\\n\",\n \"msg_date\": \"Tue, 23 Nov 2021 11:41:36 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> On 2021-Nov-20, Andrew Dunstan wrote:\\n>> Maybe something along those lines but using a dash/hyphen would work:\\n>> e.g. wal-replayer\\n\\n> Yeah, the idea of a dash occurred to me too.\\n\\nBad memories of uuid-ossp float up ... can we use underscore?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 23 Nov 2021 10:04:08 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-23, Tom Lane wrote:\\n\\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> > On 2021-Nov-20, Andrew Dunstan wrote:\\n> >> Maybe something along those lines but using a dash/hyphen would work:\\n> >> e.g. wal-replayer\\n> \\n> > Yeah, the idea of a dash occurred to me too.\\n> \\n> Bad memories of uuid-ossp float up ... can we use underscore?\\n\\nI don't object to an underscore, but it looks a bit uglier to me.\\nAFAIK the main problem with uuid-ossp was that it is used as an identifier, so\\nit required quoting, which won't be the case with this process name.\\n\\n\\nAnyway, as a second and separate point, should we rename the other WAL-related\\nprocess names to use the same character?\\n\\nwalreceiver -> wal_receiver / wal-receiver\\nwalsender -> wal_sender / wal-sender\\nwalwriter -> wal_writer / wal-writer.\\n\\nI assume we would *not* rename other existing processes that already\\nhave spaces in their names, such as \\\"autovacuum launcher\\\" to\\n\\\"autovacuum_launcher\\\", as that would be a net loss IMO.\\n\\nCurrently, the process list in an idle primary server looks like \\n\\nUSER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND\\nalvherre 1091970 0.0 0.0 199608 22252 pts/0 S+ 12:22 0:00 /pgsql/install/master/bin/postmaster -p 55432\\nalvherre 1091983 0.0 0.0 199608 3644 ? Ss 12:22 0:00 \\\\_ postgres: checkpointer \\nalvherre 1091984 0.0 0.0 199608 5432 ? Ss 12:22 0:00 \\\\_ postgres: background writer \\nalvherre 1091986 0.0 0.0 199608 8956 ? Ss 12:22 0:00 \\\\_ postgres: walwriter \\nalvherre 1091987 0.0 0.0 200148 7988 ? Ss 12:22 0:00 \\\\_ postgres: autovacuum launcher \\nalvherre 1091988 0.0 0.0 54432 4156 ? Ss 12:22 0:00 \\\\_ postgres: stats collector \\nalvherre 1091989 0.0 0.0 200036 6416 ? Ss 12:22 0:00 \\\\_ postgres: logical replication launcher \\n\\nI think this looks better:\\n\\nUSER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND\\nalvherre 1091970 0.0 0.0 199608 22252 pts/0 S+ 12:22 0:00 /pgsql/install/master/bin/postmaster -p 55432\\nalvherre 1091983 0.0 0.0 199608 3644 ? Ss 12:22 0:00 \\\\_ postgres: checkpointer \\nalvherre 1091984 0.0 0.0 199608 5432 ? Ss 12:22 0:00 \\\\_ postgres: background writer \\nalvherre 1091986 0.0 0.0 199608 8956 ? Ss 12:22 0:00 \\\\_ postgres: wal-writer \\nalvherre 1091987 0.0 0.0 200148 7988 ? Ss 12:22 0:00 \\\\_ postgres: autovacuum launcher \\nalvherre 1091988 0.0 0.0 54432 4156 ? Ss 12:22 0:00 \\\\_ postgres: stats collector \\nalvherre 1091989 0.0 0.0 200036 6416 ? Ss 12:22 0:00 \\\\_ postgres: logical replication launcher \\n\\n-- \\nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Wed, 24 Nov 2021 15:54:37 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> On 2021-Nov-23, Tom Lane wrote:\\n>> Bad memories of uuid-ossp float up ... can we use underscore?\\n\\n> I don't object to an underscore, but it looks a bit uglier to me.\\n> AFAIK the main problem with uuid-ossp was that it is used as an identifier, so\\n> it required quoting, which won't be the case with this process name.\\n\\nWell, we can call the process whatever we want, but we will need to\\nderive identifiers from that name, so you can't just ignore the issue.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 24 Nov 2021 14:07:34 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 24, 2021 at 1:54 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> I don't object to an underscore, but it looks a bit uglier to me.\\n> AFAIK the main problem with uuid-ossp was that it is used as an identifier, so\\n> it required quoting, which won't be the case with this process name.\\n\\nI agree, and for that reason I would prefer no separator, or a space.\\nThat's what we do with other processes, and I think it's fine.\\n\\nIt's worth thinking too about the fact that we may want to rename\\nfunctions, adjust comments, etc. Each of those things has their own\\nconventions. For example consider StartupProcessMain(). If we decide\\nto call this the WAL replay process, I suppose that is going to become\\nWALReplayMain(). For sure it's not going to be come WAL-Replay-Main().\\nBut what displays in the 'ps' status should look like what we do in\\nother cases.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 24 Nov 2021 14:10:00 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Should rename \\\"startup process\\\" to something else?\"\n }\n]","isConversational":false}}},{"rowIdx":53093,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hackers,\n\nIn the ongoing effort [1] to reduce the number of tasks which require operating under superuser privileges, this patch extends the system to allow, per GUC variable, the ability to SET or ALTER SYSTEM for the variable. A previous patch set was submitted [2] which created hard-coded privileged roles with the authority to manage associated hard-coded sets of GUC variables. This current patch appears superior in several ways:\n\n- It allows much greater flexibility in how roles and GUCs are associated\n- Custom GUC variables defined by extensions can be covered by this approach\n\nand perhaps most importantly,\n\n- It's what Andrew suggested\n\nGranting SET privilege on a USERSET variable makes no practical difference, but for SUSET variables it does, and granting ALTER SYSTEM is meaningful for all variables. The patch does not mandate that non-login roles be created for this, but as a usage suggestion, one could define a non-login role and assign privileges for a set of GUCs, such as:\n\n CREATE ROLE regress_host_resource_admin NOSUPERUSER;\n GRANT SET VALUE, ALTER SYSTEM ON\n autovacuum_work_mem, hash_mem_multiplier, logical_decoding_work_mem,\n maintenance_work_mem, max_stack_depth, min_dynamic_shared_memory,\n shared_buffers, temp_buffers, temp_file_limit, work_mem\n TO regress_host_resource_admin;\n\nand then delegate authority to manage the set of GUCs to a non-superuser by granting membership in non-login role:\n\n CREATE ROLE regress_admin_member IN ROLE regress_host_resource_admin;\n\nOne disadvantage of this approach is that the GUC variables are represented both in the list of C structures in guc.c and in the new system catalog pg_config_param's .dat file. Failure to enter a GUC in the .dat file will result in the inability to grant privileges on the GUC, at least unless/until you run CREATE CONFIGURATION PARAMETER on the GUC. (This is, in fact, how extension scripts deal with the issue.) It would perhaps be better if the list of GUCs were not duplicated, but I wasn't clever enough to find a clean way to do that without greatly expanding the patch (nor did I complete prototyping any such thing.)\n\n\n\n\n\n\n[1] https://www.postgresql.org/message-id/flat/F9408A5A-B20B-42D2-9E7F-49CD3D1547BC%40enterprisedb.com\n[2] https://www.postgresql.org/message-id/flat/0BE45772-31CF-4BE9-9FE8-FB7D8B130240%40enterprisedb.com\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Mon, 15 Nov 2021 12:37:44 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Mon, Nov 15, 2021 at 3:37 PM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n> One disadvantage of this approach is that the GUC variables are represented both in the list of C structures in guc.c and in the new system catalog pg_config_param's .dat file. Failure to enter a GUC in the .dat file will result in the inability to grant privileges on the GUC, at least unless/until you run CREATE CONFIGURATION PARAMETER on the GUC. (This is, in fact, how extension scripts deal with the issue.) It would perhaps be better if the list of GUCs were not duplicated, but I wasn't clever enough to find a clean way to do that without greatly expanding the patch (nor did I complete prototyping any such thing.)\n\nI think this imposes an unnecessary burden on developers. It seems\nlike it would be easy to write some code that lives inside guc.c and\nruns during bootstrap, and it could just iterate over each\nConfigureNamesWhatever array and insert catalog entries for everything\nit finds.\n\nIt's also going to be important to think about what happens with\nextension GUCs. If somebody installs an extension, we can't ask them\nto perform a manual step in order to be able to grant privileges. And\nif somebody then loads up a different .so for that extension, the set\nof GUCs that it provides can change without any DDL being executed.\nNew GUCs could appear, and old GUCs could vanish.\n\nSo I wonder if we should instead not do what I proposed above, and\ninstead just adjust the GRANT command to automatically insert a new\nrow into the relevant catalog if there isn't one already. That seems\nnicer for extensions, and also nicer for core GUCs, since it avoids\nbloating the catalog with a bunch of entries that aren't needed.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 16 Nov 2021 10:03:34 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> It's also going to be important to think about what happens with\n> extension GUCs. If somebody installs an extension, we can't ask them\n> to perform a manual step in order to be able to grant privileges. And\n> if somebody then loads up a different .so for that extension, the set\n> of GUCs that it provides can change without any DDL being executed.\n> New GUCs could appear, and old GUCs could vanish.\n\nRight. I think that any design that involves per-GUC catalog entries\nis going to be an abysmal failure, precisely because the set of GUCs\nis not stable enough. So I'm suspicious of this entire proposal.\nMaybe there's a way to make it work, but that way isn't how.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 16 Nov 2021 10:17:15 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Tue, Nov 16, 2021 at 10:17 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Right. I think that any design that involves per-GUC catalog entries\n> is going to be an abysmal failure, precisely because the set of GUCs\n> is not stable enough.\n\nIn practice it's pretty stable. I think it's just a matter of having a\nplan that covers the cases where it isn't stable reasonably elegantly.\n\nWe already embed GUC names in catalog entries when someone runs ALTER\nUSER SET or ALTER DATABASE SET, so this proposal doesn't seem to be\nmoving the goalposts in any meaningful way.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 16 Nov 2021 10:23:27 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Nov 16, 2021, at 7:03 AM, Robert Haas <robertmhaas@gmail.com> wrote:\n> \n> It's also going to be important to think about what happens with\n> extension GUCs. If somebody installs an extension, we can't ask them\n> to perform a manual step in order to be able to grant privileges.\n\nThe burden isn't on the installer of an extension. As implemented, it's the extension's installation .sql file that sets it up, and the upgrade .sql files that make adjustments, if necessary.\n\n> And\n> if somebody then loads up a different .so for that extension, the set\n> of GUCs that it provides can change without any DDL being executed.\n> New GUCs could appear, and old GUCs could vanish.\n\nWell, the same is true for functions, right? If you add, remove, or redefine functions in the extension, you need an upgrade script that defines the new functions, removes the old ones, changes function signatures, or whatever. The same is true here for GUCs.\n\nI don't think we support using a .so that is mismatched against the version of the extension that is installed.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Tue, 16 Nov 2021 07:24:51 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Tue, Nov 16, 2021 at 10:17 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> Right. I think that any design that involves per-GUC catalog entries\n>> is going to be an abysmal failure, precisely because the set of GUCs\n>> is not stable enough.\n\n> In practice it's pretty stable. I think it's just a matter of having a\n> plan that covers the cases where it isn't stable reasonably elegantly.\n\nUm. Really the point comes down to having sane default behavior\nwhen there's no entry, which ought to eliminate any need to do the\nsort of \"run over all the entries at startup\" processing that you\nseemed to be proposing. So I guess I don't understand why such a\nthing would be needed.\n\n> We already embed GUC names in catalog entries when someone runs ALTER\n> USER SET or ALTER DATABASE SET, so this proposal doesn't seem to be\n> moving the goalposts in any meaningful way.\n\nTrue; as long as the expectation is that entries will exist for only\na tiny subset of GUCs, it's probably fine.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 16 Nov 2021 10:28:57 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> I don't think we support using a .so that is mismatched against the\n> version of the extension that is installed.\n\nYou are entirely mistaken. That's not only \"supported\", it's *required*.\nConsider binary upgrades, where the SQL objects are transferred as-is\nbut the receiving installation may have a different (hopefully newer)\nversion of the .so. That .so has to cope with the older SQL object\ndefinitions; it doesn't get to assume that the upgrade script has been\nrun yet.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 16 Nov 2021 10:32:07 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Nov 16, 2021, at 7:32 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\n>> I don't think we support using a .so that is mismatched against the\n>> version of the extension that is installed.\n> \n> You are entirely mistaken. That's not only \"supported\", it's *required*.\n> Consider binary upgrades, where the SQL objects are transferred as-is\n> but the receiving installation may have a different (hopefully newer)\n> version of the .so. That .so has to cope with the older SQL object\n> definitions; it doesn't get to assume that the upgrade script has been\n> run yet.\n\nYou are talking about mismatches in the other direction, aren't you? I was responding to Robert's point that new gucs could appear, and old gucs disappear. That seems analogous to new functions appearing and old functions disappearing. If you upgrade (not downgrade) the .so, the new gucs and functions will be in the .so, but won't be callable/grantable from sql until the upgrade script runs. The old gucs and functions will be missing from the .so, and attempts to call them/grant them from sql before the upgrade will fail. What am I missing here?\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Tue, 16 Nov 2021 07:45:14 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Nov 16, 2021, at 7:28 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> True; as long as the expectation is that entries will exist for only\n> a tiny subset of GUCs, it's probably fine.\n\nI understand that bloating a frequently used catalog can be pretty harmful to performance. I wasn't aware that the size of an infrequently used catalog was critical. This new catalog would be used during GRANT SET ... and GRANT ALTER SYSTEM commands, which should be rare, and potentially consulted when SET or ALTER SYSTEM commands are issued. Is there a more substantial performance impact to this than I'm aware? It can be a bit challenging to run performance tests on such things, given the way everything interacts with everything else. \n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Tue, 16 Nov 2021 07:52:06 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Tue, Nov 16, 2021 at 10:45 AM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n> You are talking about mismatches in the other direction, aren't you? I was responding to Robert's point that new gucs could appear, and old gucs disappear. That seems analogous to new functions appearing and old functions disappearing. If you upgrade (not downgrade) the .so, the new gucs and functions will be in the .so, but won't be callable/grantable from sql until the upgrade script runs. The old gucs and functions will be missing from the .so, and attempts to call them/grant them from sql before the upgrade will fail. What am I missing here?\n\nIt's true that we could impose such a restriction, but I don't think\nwe should. If I install a different .so, I want the new GUCs to be\ngrantable immediately, without running any separate DDL.\n\nI also don't think we should burden extension authors with putting\nstuff in their upgrade scripts for this. We should solve the problem\nin our code rather than forcing them to do so in theirs.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 16 Nov 2021 11:07:36 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> You are talking about mismatches in the other direction, aren't you? I was responding to Robert's point that new gucs could appear, and old gucs disappear. That seems analogous to new functions appearing and old functions disappearing. If you upgrade (not downgrade) the .so, the new gucs and functions will be in the .so, but won't be callable/grantable from sql until the upgrade script runs.\n\nTrue, but in general users might not care about getting access to new\nfeatures (or at least, they might not care until much later, at which\npoint they'd bother to run ALTER EXTENSION UPDATE).\n\n> The old gucs and functions will be missing from the .so, and attempts to call them/grant them from sql before the upgrade will fail. What am I missing here?\n\nHere, you are describing behavior that's against project policy and would\nbe rejected immediately if anyone submitted a patch that made an extension\ndo that. Newer .so versions are expected to run successfully, not fail,\nwith an older version of their SQL objects. Were this not so, it's almost\ninevitable that a binary upgrade would fail, because the extension is\nlikely to be called into action somewhere before there is any opportunity\nto issue ALTER EXTENSION UPDATE. Even in-place updates of extensions\nwould become problematic: you can't assume that a rollout of new\nexecutables will be instantly followed by ALTER EXTENSION UPDATE.\n\nBasically, I think the proposed new system catalog is both unworkable\nand entirely unnecessary. Maybe it would have been okay if Peter had\ndesigned GUCs like that a couple of decades ago, but at this point\nwe have a ton of infrastructure --- much of it not under the core\nproject's control --- that assumes that GUCs can be created with just\na bit of code. I do not think that this feature is worth the cost of\nchanging that. Or IOW: ALTER SYSTEM SET has gotten along fine without\nsuch a catalog; why can't this feature?\n\nI also notice that the patch seems to intend to introduce tracking\nof which user \"owns\" a GUC, which I think is even more unworkable.\nThey are all going to wind up owned by the bootstrap superuser\n(extension GUCs too), so why bother?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 16 Nov 2021 11:29:06 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> On Nov 16, 2021, at 7:28 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> True; as long as the expectation is that entries will exist for only\n>> a tiny subset of GUCs, it's probably fine.\n\n> I understand that bloating a frequently used catalog can be pretty\n> harmful to performance. I wasn't aware that the size of an infrequently\n> used catalog was critical.\n\nMy concern is not about performance, it's about the difficulty of\nmaintaining a catalog that expects to be a more-or-less exhaustive\nlist of GUCs. I think you need to get rid of that expectation.\nIn the analogy to ALTER DATABASE/USER SET, we don't expect that\npg_db_role_setting catalog entries will exist for all, or even\nvery many, GUCs. Also, the fact that pg_db_role_setting entries\naren't tied very hard to the actual existence of a GUC is a good\nthing from the maintenance and upgrade standpoint.\n\nBTW, if we did create such a catalog, there would need to be\npg_dump and pg_upgrade support for its entries, and you'd have to\nthink about (e.g.) whether pg_upgrade would attempt to preserve\nthe same OIDs. I don't see any indication that the patch has\naddressed that infrastructure ... which is probably just as well,\nsince it's work that I'd be wanting to reject. (Hm, but actually,\ndoesn't pg_dump need work anyway to dump this new type of GRANT?)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 16 Nov 2021 11:44:07 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Nov 16, 2021, at 8:44 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> My concern is not about performance, it's about the difficulty of\n> maintaining a catalog that expects to be a more-or-less exhaustive\n> list of GUCs. I think you need to get rid of that expectation.\n\nI'm preparing a new version of the patch that has the catalog empty to begin with, and only adds values in response to GRANT commands. That also handles the issues of extension upgrades, which I think the old patch handled better than the discussion on this thread suggested, but no matter. The new behavior will allow granting privileges on non-existent gucs, just as ALTER ROLE..SET allows registering settings on non-existent gucs.\n\nThe reason I had resisted allowing grants of privileges on non-existent gucs is that you can get the following sort of behavior (note the last two lines):\n\n DROP USER regress_priv_user7;\n ERROR: role \"regress_priv_user7\" cannot be dropped because some objects depend on it\n DETAIL: privileges for table persons2\n privileges for configuration parameter sort_mem\n privileges for configuration parameter no_such_param\n\nRejecting \"no_such_param\" in the original GRANT statement seemed cleaner to me, but this discussion suggests pretty strongly that I can't do it that way. Changing the historical \"sort_mem\" to the canonical \"work_mem\" name also seems better to me, as otherwise you could have different grants on the same GUC under different names. I'm inclined to keep the canonicalization of names where known, but maybe that runs afoul the rule that these grants should not be tied very hard to the GUC?\n\n> In the analogy to ALTER DATABASE/USER SET, we don't expect that\n> pg_db_role_setting catalog entries will exist for all, or even\n> very many, GUCs. Also, the fact that pg_db_role_setting entries\n> aren't tied very hard to the actual existence of a GUC is a good\n> thing from the maintenance and upgrade standpoint.\n\nDoing it that way....\n\n> BTW, if we did create such a catalog, there would need to be\n> pg_dump and pg_upgrade support for its entries, and you'd have to\n> think about (e.g.) whether pg_upgrade would attempt to preserve\n> the same OIDs. I don't see any indication that the patch has\n> addressed that infrastructure ... which is probably just as well,\n> since it's work that I'd be wanting to reject.\n\nYeah, that's why I didn't write it. I wanted feedback on the basic implementation before doing that work.\n\n> (Hm, but actually,\n> doesn't pg_dump need work anyway to dump this new type of GRANT?)\n\nYes, if the idea of this kind of grant isn't being outright rejected, then I'll need to write that.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Tue, 16 Nov 2021 11:48:16 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 11/16/21 14:48, Mark Dilger wrote:\n>\n>> On Nov 16, 2021, at 8:44 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>\n>> My concern is not about performance, it's about the difficulty of\n>> maintaining a catalog that expects to be a more-or-less exhaustive\n>> list of GUCs. I think you need to get rid of that expectation.\n> I'm preparing a new version of the patch that has the catalog empty to begin with, and only adds values in response to GRANT commands. That also handles the issues of extension upgrades, which I think the old patch handled better than the discussion on this thread suggested, but no matter. The new behavior will allow granting privileges on non-existent gucs, just as ALTER ROLE..SET allows registering settings on non-existent gucs.\n\n\n\nYour original and fairly simple set of patches used hardcoded role names\nand sets of GUCs they could update via ALTER SYSTEM. I suggested to you\nprivately that a more flexible approach would be to drive this from a\ncatalog table. I had in mind a table of more or less <roleid, guc_name>.\nYou could prepopulate it with the roles / GUCs from your original patch\nset. I don't think it needs to be initially empty. But DBAs would be\nable to modify and extend the settings. I agree with Tom that we\nshouldn't try to cover all GUCs in the table - any GUC without an entry\ncan only be updated by a superuser.\n\n\nTo support pg_dump and pg_upgrade, it might be better to have an\nenabled/disabled flag rather than to delete rows.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Tue, 16 Nov 2021 15:38:07 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Nov 16, 2021, at 12:38 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\n> \n> Your original and fairly simple set of patches used hardcoded role names\n> and sets of GUCs they could update via ALTER SYSTEM. I suggested to you\n> privately that a more flexible approach would be to drive this from a\n> catalog table. I had in mind a table of more or less <roleid, guc_name>.\n\nRight. I prefer a table of <guc_name, acl> matching the structure of most of the rest of the grantable objects, so that entries from pg_depend or pg_shdepend can track the dependencies in the usual way and prevent dangling references when DROP ROLE is executed. Is there a reason to prefer an ad hoc tables structure?\n\n> You could prepopulate it with the roles / GUCs from your original patch\n> set. I don't think it needs to be initially empty. But DBAs would be\n> able to modify and extend the settings. I agree with Tom that we\n> shouldn't try to cover all GUCs in the table - any GUC without an entry\n> can only be updated by a superuser.\n\nThe patch already posted on this thread already works that way. Robert and Tom seemed to infer that all gucs need to be present in the catalog, but in fact, not entering them in the catalog simply means that they aren't grantable. I think the confusion arose from the fact that I created a mechanism for extension authors to enter additional gucs into the catalog, which gave the false impression that such entries were required. They are not. I didn't bother explaining that before, because Robert and Tom both seem to expect all gucs to be grantable, an expectation you do not appear to share.\n\n> To support pg_dump and pg_upgrade, it might be better to have an\n> enabled/disabled flag rather than to delete rows.\n\nI'm not really sure what this means. Are you suggesting that the <guc_name, acl> (or in your formulation <roleid, guc_name>) should have a \"bool enabled\" field, and when the guc gets dropped, the \"enabled\" field gets set to false?\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Tue, 16 Nov 2021 13:02:44 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> The patch already posted on this thread already works that way. Robert and Tom seemed to infer that all gucs need to be present in the catalog, but in fact, not entering them in the catalog simply means that they aren't grantable. I think the confusion arose from the fact that I created a mechanism for extension authors to enter additional gucs into the catalog, which gave the false impression that such entries were required. They are not. I didn't bother explaining that before, because Robert and Tom both seem to expect all gucs to be grantable, an expectation you do not appear to share.\n\nThe question is why you need pg_config_param at all, then.\nAFAICS it just adds maintenance complexity we could do without.\nI think we'd be better off with a catalog modeled on the design of\npg_db_role_setting, which would have entries for roles and lists\nof GUC names that those roles could set.\n\nBTW, another objection to pg_config_param as designed here is that\na \"name\" is not an appropriate way to store possibly-qualified\ncustom GUC names. It's not long enough (cf. valid_custom_variable_name).\n\n>> To support pg_dump and pg_upgrade, it might be better to have an\n>> enabled/disabled flag rather than to delete rows.\n\n> I'm not really sure what this means.\n\nI didn't see the point of this either. We really need to KISS here.\nEvery bit of added complexity in the catalog representation is another\nopportunity for bugs-of-omission, not to mention a detail that you\nhave to provide mechanisms to dump and restore.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 16 Nov 2021 17:12:49 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Nov 16, 2021, at 2:12 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> The question is why you need pg_config_param at all, then.\n> AFAICS it just adds maintenance complexity we could do without.\n> I think we'd be better off with a catalog modeled on the design of\n> pg_db_role_setting, which would have entries for roles and lists\n> of GUC names that those roles could set.\n\nOriginally, I was trying to have dependency linkage between two proper types of objects, so that DROP ROLE and DROP CONFIGURATION PARAMETER would behave as expected, complaining about privileges remaining rather than dropping an object and leaving a dangling reference.\n\nI've deleted the whole CREATE CONFIGURATION PARAMETER and DROP CONFIGURATION PARAMETER syntax and implementation, but it still seems odd to me that:\n\n CREATE ROLE somebody;\n GRANT SELECT ON TABLE sometable TO ROLE somebody;\n GRANT ALTER SYSTEM ON someguc TO ROLE somebody;\n DROP ROLE somebody;\n ERROR: role \"somebody\" cannot be dropped because some objects depend on it\n DETAIL: privileges for table sometable\n\nwould not mention privileges for \"someguc\" as well. That's why I want configuration parameters to be proper objects with OIDs and with entries in pg_depend and/or pg_shdepend. Maybe there is some better way to do it, but that's why I've been doing this.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Tue, 16 Nov 2021 14:25:12 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Nov 16, 2021, at 2:12 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> BTW, another objection to pg_config_param as designed here is that\n> a \"name\" is not an appropriate way to store possibly-qualified\n> custom GUC names. It's not long enough (cf. valid_custom_variable_name).\n\nI was aware of that, but figured not all GUCs have to be grantable. If it doesn't fit in a NameData, you can't grant on it.\n\nIf we want to be more accommodating than that, we can store it as text, just like pg_db_role_names does, but then we need more code complexity to look it up and to verify that it is unique. (We wouldn't want multiple records for the same <role,guc> pair.) \n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Tue, 16 Nov 2021 14:45:12 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Tue, Nov 16, 2021 at 2:48 PM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n> I'm preparing a new version of the patch that has the catalog empty to begin with, and only adds values in response to GRANT commands. That also handles the issues of extension upgrades, which I think the old patch handled better than the discussion on this thread suggested, but no matter. The new behavior will allow granting privileges on non-existent gucs, just as ALTER ROLE..SET allows registering settings on non-existent gucs.\n>\n> The reason I had resisted allowing grants of privileges on non-existent gucs is that you can get the following sort of behavior (note the last two lines):\n>\n> DROP USER regress_priv_user7;\n> ERROR: role \"regress_priv_user7\" cannot be dropped because some objects depend on it\n> DETAIL: privileges for table persons2\n> privileges for configuration parameter sort_mem\n> privileges for configuration parameter no_such_param\n>\n> Rejecting \"no_such_param\" in the original GRANT statement seemed cleaner to me, but this discussion suggests pretty strongly that I can't do it that way.\n\nI think it's perfectly fine to refuse a GRANT statement on a GUC that\ndoesn't exist, and I don't believe I ever said otherwise. What I don't\nthink is OK is to require a preparatory statement like CREATE\nCONFIGURATION PARAMETER to be executed before you can grant\npermissions on it. There's no reason for that. If somebody tries to\ngrant privileges on a GUC that does not exist, fail. If the GUC does\nexist but there's no catalog entry, make one. If the GUC exists and\nthe catalog entry also exists, update it.\n\nAt REVOKE time, don't check whether the GUC exists - only check the\ncatalog. That way people can remove privileges on GUCs that don't\nexist any more. If somebody issues a REVOKE and there's no catalog\nentry, do nothing. If somebody issues a REVOKE and there is a catalog\nentry, remove stuff from it; but if that would leave it completely\nempty, instead delete it.\n\nWhenever you create a catalog entry, also add dependency entries\npointing to it. When you delete one, remove those entries.\n\n> Changing the historical \"sort_mem\" to the canonical \"work_mem\" name also seems better to me, as otherwise you could have different grants on the same GUC under different names. I'm inclined to keep the canonicalization of names where known, but maybe that runs afoul the rule that these grants should not be tied very hard to the GUC?\n\nNo. If somebody grants privileges on an old name, record the grant\nunder the canonical name.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 17 Nov 2021 08:26:33 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Tue, Nov 16, 2021 at 3:38 PM Andrew Dunstan <andrew@dunslane.net> wrote:\n> Your original and fairly simple set of patches used hardcoded role names\n> and sets of GUCs they could update via ALTER SYSTEM. I suggested to you\n> privately that a more flexible approach would be to drive this from a\n> catalog table. I had in mind a table of more or less <roleid, guc_name>.\n> You could prepopulate it with the roles / GUCs from your original patch\n> set. I don't think it needs to be initially empty. But DBAs would be\n> able to modify and extend the settings. I agree with Tom that we\n> shouldn't try to cover all GUCs in the table - any GUC without an entry\n> can only be updated by a superuser.\n\nI simply can't understand the point of this. You're basically\nproposing that somebody has to execute one SQL statement to make a GUC\ngrantable, and then a second SQL statement to actually grant access to\nit. What is the value in that? It is the same person doing both\nthings, and the system can work out automatically what needs to be\ndone.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 17 Nov 2021 08:27:59 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Tue, Nov 16, 2021 at 5:45 PM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n> I was aware of that, but figured not all GUCs have to be grantable. If it doesn't fit in a NameData, you can't grant on it.\n\nSuch restrictions are rather counterintuitive for users, and here it\ndoesn't even buy anything. Using 'text' rather than 'name' as the data\ntype isn't going to cost any meaningful amount of performance.\n\n> If we want to be more accommodating than that, we can store it as text, just like pg_db_role_names does, but then we need more code complexity to look it up and to verify that it is unique. (We wouldn't want multiple records for the same <role,guc> pair.)\n\nIf you're verifying that it's unique in any way other than using a\nunique index, I think you're doing it wrong.\n\nAlso, maybe I'm confused here, but why isn't the schema:\n\ngucoid\ngucname\ngucacl\n\nIOW, I don't understand why this table has <role,guc> as the primary\nkey rather than just guc. Everywhere else, we have a single ACL array\nfor the all privileges on an object. Why wouldn't we do the same thing\nhere?\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 17 Nov 2021 08:32:27 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 11/17/21 08:32, Robert Haas wrote:\n> On Tue, Nov 16, 2021 at 5:45 PM Mark Dilger\n> <mark.dilger@enterprisedb.com> wrote:\n>> I was aware of that, but figured not all GUCs have to be grantable. If it doesn't fit in a NameData, you can't grant on it.\n> Such restrictions are rather counterintuitive for users, and here it\n> doesn't even buy anything. Using 'text' rather than 'name' as the data\n> type isn't going to cost any meaningful amount of performance.\n\n\nindeed\n\n\n>> If we want to be more accommodating than that, we can store it as text, just like pg_db_role_names does, but then we need more code complexity to look it up and to verify that it is unique. (We wouldn't want multiple records for the same <role,guc> pair.)\n> If you're verifying that it's unique in any way other than using a\n> unique index, I think you're doing it wrong.\n\n\nyeah\n\n\n>\n> Also, maybe I'm confused here, but why isn't the schema:\n>\n> gucoid\n> gucname\n> gucacl\n>\n> IOW, I don't understand why this table has <role,guc> as the primary\n> key rather than just guc. Everywhere else, we have a single ACL array\n> for the all privileges on an object. Why wouldn't we do the same thing\n> here?\n>\n\n\nYes, that should work, It seems like a better scheme.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 17 Nov 2021 09:24:39 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 11/16/21 17:12, Tom Lane wrote:\n>\n>>> To support pg_dump and pg_upgrade, it might be better to have an\n>>> enabled/disabled flag rather than to delete rows.\n>> I'm not really sure what this means.\n> I didn't see the point of this either. We really need to KISS here.\n> Every bit of added complexity in the catalog representation is another\n> opportunity for bugs-of-omission, not to mention a detail that you\n> have to provide mechanisms to dump and restore.\n>\n> \t\t\t\n\n\nWell, I was trying (perhaps not very well) to imagine how to deal with\nsomeone modifying the permissions of one of the predefined roles. Say\npg_foo has initial permission to set bar and baz, and the DBA removes\npermission to set baz. How is pg_dump going to emit the right commands\nto allow a safe pg_upgrade? Maybe we should say that the permissions for\nthe predefined roles are immutable, so only permissions sets for user\ndefined roles are mutable.\n\n\ncheers\n\n\nandrew\n\n-- \n\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 17 Nov 2021 09:31:40 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Nov 17, 2021, at 5:32 AM, Robert Haas <robertmhaas@gmail.com> wrote:\n> \n>> I was aware of that, but figured not all GUCs have to be grantable. If it doesn't fit in a NameData, you can't grant on it.\n> \n> Such restrictions are rather counterintuitive for users, and here it\n> doesn't even buy anything. Using 'text' rather than 'name' as the data\n> type isn't going to cost any meaningful amount of performance.\n\nThat sounds fine.\n\n>> If we want to be more accommodating than that, we can store it as text, just like pg_db_role_names does, but then we need more code complexity to look it up and to verify that it is unique. (We wouldn't want multiple records for the same <role,guc> pair.)\n> \n> If you're verifying that it's unique in any way other than using a\n> unique index, I think you're doing it wrong.\n\nNo, I'm using a unique index. I was overthinking it, concerned about changing from name_ops to text_ops and needing the toast table, but that's silly, because I need one for the acl anyway.\n\n> Also, maybe I'm confused here, but why isn't the schema:\n> \n> gucoid\n> gucname\n> gucacl\n\nIt is, both in v2 already posted, and in the v3, written but not yet posted, as I haven't finished the pg_dump work, and also I'm waiting to see how this discussion gets resolved before asking for a review of v3.\n\n> IOW, I don't understand why this table has <role,guc> as the primary\n> key rather than just guc.\n\nI was responding to Tom's recommendation that I follow the pattern in pg_db_role_setting, and speculating how that would work. I was not proposing to do it that way.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Wed, 17 Nov 2021 07:57:41 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Nov 17, 2021, at 6:31 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\n> \n> Well, I was trying (perhaps not very well) to imagine how to deal with\n> someone modifying the permissions of one of the predefined roles. Say\n> pg_foo has initial permission to set bar and baz, and the DBA removes\n> permission to set baz. How is pg_dump going to emit the right commands\n> to allow a safe pg_upgrade? Maybe we should say that the permissions for\n> the predefined roles are immutable, so only permissions sets for user\n> defined roles are mutable.\n\nI find this somewhat amusing. When you suggested off-list that I make gucs individually grantable rather than creating predefined roles with privileges, that was a great idea precisely because sites could define their own security policies using their own site-defined roles:\n\n CREATE ROLE admin_type_a NOLOGIN NOSUPERUSER;\n CREATE ROLE admin_type_b NOLOGIN NOSUPERUSER;\n ...\n GRANT ALTER SYSTEM ON guc_a1, guc_a2, guc_a3, ... TO admin_type_a;\n GRANT ALTER SYSTEM ON guc_b1, guc_b2, guc_b3, ... TO admin_type_b;\n ...\n\nThat has all the power of a system based on predefined roles, but with site-specific flexibility, which is better. So it amuses me that we'd now be talking about granting some of these to predefined roles, as that is a regression in flexibility. (How would a site revoke it from one of those predefined roles if they wanted a different policy?)\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Wed, 17 Nov 2021 08:07:04 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Wed, Nov 17, 2021 at 9:31 AM Andrew Dunstan <andrew@dunslane.net> wrote:\n> Well, I was trying (perhaps not very well) to imagine how to deal with\n> someone modifying the permissions of one of the predefined roles. Say\n> pg_foo has initial permission to set bar and baz, and the DBA removes\n> permission to set baz. How is pg_dump going to emit the right commands\n> to allow a safe pg_upgrade? Maybe we should say that the permissions for\n> the predefined roles are immutable, so only permissions sets for user\n> defined roles are mutable.\n\nThat's a great question, but it isn't a new problem. If I create a\nbrand new database and do thIs:\n\nrhaas=# revoke execute on function pg_ls_waldir() from pg_monitor;\nREVOKE\n\nAnd then I do this:\n\n[rhaas pgsql]$ pg_dump\n\nThen the output includes this:\n\nREVOKE ALL ON FUNCTION pg_catalog.pg_ls_waldir(OUT name text, OUT size\nbigint, OUT modification timestamp with time zone) FROM pg_monitor;\n\nI recommend looking at how that works and making this work the same way.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 17 Nov 2021 12:05:53 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 11/17/21 11:07, Mark Dilger wrote:\n>\n>> On Nov 17, 2021, at 6:31 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>\n>> Well, I was trying (perhaps not very well) to imagine how to deal with\n>> someone modifying the permissions of one of the predefined roles. Say\n>> pg_foo has initial permission to set bar and baz, and the DBA removes\n>> permission to set baz. How is pg_dump going to emit the right commands\n>> to allow a safe pg_upgrade? Maybe we should say that the permissions for\n>> the predefined roles are immutable, so only permissions sets for user\n>> defined roles are mutable.\n> I find this somewhat amusing. When you suggested off-list that I make gucs individually grantable rather than creating predefined roles with privileges, that was a great idea precisely because sites could define their own security policies using their own site-defined roles:\n>\n> CREATE ROLE admin_type_a NOLOGIN NOSUPERUSER;\n> CREATE ROLE admin_type_b NOLOGIN NOSUPERUSER;\n> ...\n> GRANT ALTER SYSTEM ON guc_a1, guc_a2, guc_a3, ... TO admin_type_a;\n> GRANT ALTER SYSTEM ON guc_b1, guc_b2, guc_b3, ... TO admin_type_b;\n> ...\n>\n> That has all the power of a system based on predefined roles, but with site-specific flexibility, which is better. So it amuses me that we'd now be talking about granting some of these to predefined roles, as that is a regression in flexibility. (How would a site revoke it from one of those predefined roles if they wanted a different policy?)\n>\n\nI agree it's not ideal. At the time I suggested a more flexible approach\nI hadn't really thought about the problems of upgrading. If you can come\nup with something that works there then I'll be all ears.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 17 Nov 2021 12:06:50 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Nov 17, 2021, at 9:06 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\n> \n> I agree it's not ideal. At the time I suggested a more flexible approach\n> I hadn't really thought about the problems of upgrading. If you can come\n> up with something that works there then I'll be all ears.\n\nAre you talking about upgrades preserving revocations of privileges on gucs from predefined roles, or merely preserving grants and revocation of privileges on gucs to regular roles? I think the former problem is easily handled by not shipping any predefined roles with such privileges. The latter problem would seem to be a mere matter of programming, something I'm working on but don't have finished. (But maybe you see dragons ahead for me that I'm not seeing yet?)\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Wed, 17 Nov 2021 09:12:06 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 11/17/21 12:12, Mark Dilger wrote:\n>\n>> On Nov 17, 2021, at 9:06 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>\n>> I agree it's not ideal. At the time I suggested a more flexible approach\n>> I hadn't really thought about the problems of upgrading. If you can come\n>> up with something that works there then I'll be all ears.\n> Are you talking about upgrades preserving revocations of privileges on gucs from predefined roles, or merely preserving grants and revocation of privileges on gucs to regular roles? I think the former problem is easily handled by not shipping any predefined roles with such privileges. The latter problem would seem to be a mere matter of programming, something I'm working on but don't have finished. (But maybe you see dragons ahead for me that I'm not seeing yet?)\n>\n\nYes, if we don't ship with any preset privileges then the problem I was\nthinking about disappears.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 17 Nov 2021 15:24:09 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Nov 16, 2021, at 7:03 AM, Robert Haas <robertmhaas@gmail.com> wrote:\n> \n> It's also going to be important to think about what happens with\n> extension GUCs. If somebody installs an extension, we can't ask them\n> to perform a manual step in order to be able to grant privileges. And\n> if somebody then loads up a different .so for that extension, the set\n> of GUCs that it provides can change without any DDL being executed.\n> New GUCs could appear, and old GUCs could vanish.\n\nThe v3 patch allows grants on unrecognized guc names. This should allow a grant statement to precede the loading of a new .so which provides the named guc.\n\n> instead just adjust the GRANT command to automatically insert a new\n> row into the relevant catalog if there isn't one already. That seems\n> nicer for extensions, and also nicer for core GUCs, since it avoids\n> bloating the catalog with a bunch of entries that aren't needed.\n\nGrants on GUCs create a new catalog entry if necessary, or update the existing catalog entry if found.\n\nThere is a new information_schema.guc_privileges view, not present in v2.\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Mon, 22 Nov 2021 16:21:42 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Mon, Nov 22, 2021 at 7:21 PM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n> There is a new information_schema.guc_privileges view, not present in v2.\n\nIt's my impression that information_schema is a child of the SQL\nstandard, and that inventions specific to PG go in pg_catalog.\n\nAlso, I think the user-facing name for GUCs is \"settings\".\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 23 Nov 2021 11:07:12 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Nov 23, 2021, at 8:07 AM, Robert Haas <robertmhaas@gmail.com> wrote:\n> \n> It's my impression that information_schema is a child of the SQL\n> standard, and that inventions specific to PG go in pg_catalog.\n> \n> Also, I think the user-facing name for GUCs is \"settings\".\n\nThanks. These issues should be fixed in v4.\n\nAlong the way, I also added has_setting_privilege() functions overlooked in v3 and before.\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Tue, 23 Nov 2021 18:14:54 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 11/23/21 21:14, Mark Dilger wrote:\n>\n>> On Nov 23, 2021, at 8:07 AM, Robert Haas <robertmhaas@gmail.com> wrote:\n>>\n>> It's my impression that information_schema is a child of the SQL\n>> standard, and that inventions specific to PG go in pg_catalog.\n>>\n>> Also, I think the user-facing name for GUCs is \"settings\".\n> Thanks. These issues should be fixed in v4.\n>\n> Along the way, I also added has_setting_privilege() functions overlooked in v3 and before.\n\n\n\nThis patch had bit-rotted slightly, and I was attempting to remedy it.\nHowever, I got a failure running the TAP tests because of this change:\n\n\ndiff --git a/src/test/modules/test_pg_dump/t/001_base.pl\nb/src/test/modules/test_pg_dump/t/001_base.pl\nindex 16f7610883..7fbf2d871b 100644\n--- a/src/test/modules/test_pg_dump/t/001_base.pl\n+++ b/src/test/modules/test_pg_dump/t/001_base.pl\n@@ -9,7 +9,12 @@ use PostgreSQL::Test::Cluster;\n use PostgreSQL::Test::Utils;\n use Test::More;\n \n-my $tempdir = PostgreSQL::Test::Utils::tempdir;\n+# my $tempdir = PostgreSQL::Test::Utils::tempdir;\n+my $tempbase = '/tmp/test_pg_dump';\n+my $subdir = 0;\n+$subdir++ while (-e \"$tempbase/$subdir\");\n+my $tempdir = \"$tempbase/$subdir\";\n+system(\"mkdir $tempdir\");\n \n\n\nWhat's going on here?\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Mon, 13 Dec 2021 15:56:34 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Dec 13, 2021, at 12:56 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\n> \n> This patch had bit-rotted slightly, and I was attempting to remedy it.\n\nI have that already, and getting ready to post. Give me a few minutes and I'll repost.\n\n> However, I got a failure running the TAP tests because of this change:\n> \n> \n> diff --git a/src/test/modules/test_pg_dump/t/001_base.pl\n> b/src/test/modules/test_pg_dump/t/001_base.pl\n> index 16f7610883..7fbf2d871b 100644\n> --- a/src/test/modules/test_pg_dump/t/001_base.pl\n> +++ b/src/test/modules/test_pg_dump/t/001_base.pl\n> @@ -9,7 +9,12 @@ use PostgreSQL::Test::Cluster;\n> use PostgreSQL::Test::Utils;\n> use Test::More;\n> \n> -my $tempdir = PostgreSQL::Test::Utils::tempdir;\n> +# my $tempdir = PostgreSQL::Test::Utils::tempdir;\n> +my $tempbase = '/tmp/test_pg_dump';\n> +my $subdir = 0;\n> +$subdir++ while (-e \"$tempbase/$subdir\");\n> +my $tempdir = \"$tempbase/$subdir\";\n> +system(\"mkdir $tempdir\");\n> \n> \n> \n> What's going on here?\n\nYeah, I hit that, too. That was an accidentally committed bit of local testing. Please ignore it for now.\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Mon, 13 Dec 2021 12:58:25 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Dec 13, 2021, at 12:58 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> \n> Give me a few minutes and I'll repost.\n\nOk, this new version is rebased, cleaned up, and has the ability to revoke SET VALUE on most settings which were previously userset, but are now suset with initdb time granted SET VALUE privilege to public.\n\nI have not yet altered gucs defined by contrib modules, but will repost in a few hours, or tomorrow, with those changes.\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Mon, 13 Dec 2021 13:33:58 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Dec 13, 2021, at 1:33 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> \n> but will repost in a few hours\n\n... and here it is:\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Mon, 13 Dec 2021 14:34:35 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Mon, Dec 13, 2021 at 5:34 PM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n>\n>\n>\n> > On Dec 13, 2021, at 1:33 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> >\n> > but will repost in a few hours\n>\n> ... and here it is:\n\ncurrently there is a failure in check-world (not sure if it's known):\n\ndiff -U3 /src/postgres/src/test/regress/expected/guc_privs.out\n/src/postgres/src/test/regress/results/guc_privs.out\n--- /src/postgres/src/test/regress/expected/guc_privs.out 2021-12-14\n14:11:45.000000000 +0000\n+++ /src/postgres/src/test/regress/results/guc_privs.out 2021-12-14\n15:50:52.219583421 +0000\n@@ -39,8 +39,10 @@\n ALTER SYSTEM RESET autovacuum; -- ok\n -- PGC_SUSET\n SET lc_messages = 'en_US.UTF-8'; -- ok\n+ERROR: invalid value for parameter \"lc_messages\": \"en_US.UTF-8\"\n RESET lc_messages; -- ok\n ALTER SYSTEM SET lc_messages = 'en_US.UTF-8'; -- ok\n+ERROR: invalid value for parameter \"lc_messages\": \"en_US.UTF-8\"\n ALTER SYSTEM RESET lc_messages; -- ok\n -- PGC_SU_BACKEND\n SET jit_debugging_support = OFF; -- fail, cannot be set after connection start\n\nAside from that I've tested this and it seems to function as\nadvertised and in my view is worth adding to PG.\n\nOne thing that seems like an omission to me is the absence of a\nInvokeObjectPostAlterHook in pg_setting_acl_aclcheck or\npg_setting_acl_aclmask so that MAC extensions can also block this,\nInvokeObjectPostCreateHook is already in the create path so a\nPostAlter hook seems appropriate.\n\nThank you.\n\n\n","msg_date":"Tue, 14 Dec 2021 17:26:24 -0500","msg_from":"Joshua Brindle <joshua.brindle@crunchydata.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Dec 14, 2021, at 2:26 PM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\n> \n> currently there is a failure in check-world (not sure if it's known):\n\nThat one is definitely my fault. 'en_US.UTF-8' exists on my platform, so I hadn't noticed. I've changed it to use 'C', which should be portable.\n\n> One thing that seems like an omission to me is the absence of a\n> InvokeObjectPostAlterHook in pg_setting_acl_aclcheck or\n> pg_setting_acl_aclmask so that MAC extensions can also block this,\n> InvokeObjectPostCreateHook is already in the create path so a\n> PostAlter hook seems appropriate.\n\nGood catch, but that seems like a strange place to put a PostAlterHook, so I added it to ExecGrant_Setting for v6, instead. This seems more consistent with the hook in SetDefaultACL.\n\n(If you are really trying to do Managed Access Control (MAC), wouldn't that be a separate patch which adds security hooks into all *_aclcheck functions?)\n\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Wed, 15 Dec 2021 07:56:12 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Wed, Dec 15, 2021 at 10:56 AM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n>\n>\n>\n> > On Dec 14, 2021, at 2:26 PM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\n> >\n> > currently there is a failure in check-world (not sure if it's known):\n>\n> That one is definitely my fault. 'en_US.UTF-8' exists on my platform, so I hadn't noticed. I've changed it to use 'C', which should be portable.\n>\n> > One thing that seems like an omission to me is the absence of a\n> > InvokeObjectPostAlterHook in pg_setting_acl_aclcheck or\n> > pg_setting_acl_aclmask so that MAC extensions can also block this,\n> > InvokeObjectPostCreateHook is already in the create path so a\n> > PostAlter hook seems appropriate.\n>\n> Good catch, but that seems like a strange place to put a PostAlterHook, so I added it to ExecGrant_Setting for v6, instead. This seems more consistent with the hook in SetDefaultACL.\n\nAh, I was actually requesting a hook where the acl check was done for\nsetting a GUC, such that we could deny setting them in a hook,\nsomething that would be useful for the set_user extension\n(github.com/pgaudit/set_user) but having a hook for grant/revoke is\nalso helpful.\n\n> (If you are really trying to do Managed Access Control (MAC), wouldn't that be a separate patch which adds security hooks into all *_aclcheck functions?)\n\nMAC is mandatory access controls, so something that can be layered on\ntop of DAC and can only be enforced even on object owners and\nsuperuser. sepgsql is a MAC extension, for example. It uses the object\naccess hooks to enforce SELinux policy on PG objects.\n\n\n","msg_date":"Wed, 15 Dec 2021 13:02:35 -0500","msg_from":"Joshua Brindle <joshua.brindle@crunchydata.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Dec 15, 2021, at 10:02 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\n> \n> Ah, I was actually requesting a hook where the acl check was done for\n> setting a GUC, such that we could deny setting them in a hook,\n> something that would be useful for the set_user extension\n> (github.com/pgaudit/set_user)\n\nHmm, this seems orthogonal to the patch under discussion. This patch only adds a pg_setting_acl_aclcheck in ExecSetVariableStmt() for settings which have been explicitly granted, otherwise it works the traditional way (checking whether the setting is suset/userset). I don't think you'd get MAC support without finding a way to fire the hook for all settings, regardless of their presence in the new pg_setting_acl table. That is hard, because InvokeObjectPostAlterHook expects the classId (SettingAclRelationId) and the objectId (pg_setting_acl.oid), but you don't have those for many (most?) settings. As discussed upthread, we *do not* want to force an entry into the table for all settings, only for ones that have been explicitly granted.\n\nDo you agree? I'm happy to support MAC in this patch if can explain a simple way of doing so.\n\n> but having a hook for grant/revoke is\n> also helpful.\n\nYes, I see no reason to rip this out.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Wed, 15 Dec 2021 10:18:05 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Wed, Dec 15, 2021 at 1:18 PM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n>\n> > On Dec 15, 2021, at 10:02 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\n> >\n> > Ah, I was actually requesting a hook where the acl check was done for\n> > setting a GUC, such that we could deny setting them in a hook,\n> > something that would be useful for the set_user extension\n> > (github.com/pgaudit/set_user)\n>\n> Hmm, this seems orthogonal to the patch under discussion. This patch only adds a pg_setting_acl_aclcheck in ExecSetVariableStmt() for settings which have been explicitly granted, otherwise it works the traditional way (checking whether the setting is suset/userset). I don't think you'd get MAC support without finding a way to fire the hook for all settings, regardless of their presence in the new pg_setting_acl table. That is hard, because InvokeObjectPostAlterHook expects the classId (SettingAclRelationId) and the objectId (pg_setting_acl.oid), but you don't have those for many (most?) settings. As discussed upthread, we *do not* want to force an entry into the table for all settings, only for ones that have been explicitly granted.\n>\n> Do you agree? I'm happy to support MAC in this patch if can explain a simple way of doing so.\n\nAh, I understand now. Would it be possible to pass the\nSettingAclRelationId if it exists or InvalidOid if not? That way if a\nMAC implementation cares about a particular GUC it'll ensure it's in\npg_setting_acl.\n\nI don't know if others will object to that but it seems like an okay\ncompromise.\n\n> > but having a hook for grant/revoke is\n> > also helpful.\n>\n> Yes, I see no reason to rip this out.\n>\n\n\n","msg_date":"Thu, 16 Dec 2021 10:43:15 -0500","msg_from":"Joshua Brindle <joshua.brindle@crunchydata.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Dec 16, 2021, at 7:43 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\n> \n> Ah, I understand now. Would it be possible to pass the\n> SettingAclRelationId if it exists or InvalidOid if not?\n\nSettingAclRelationId is always defined, so we can always pass that value. But the settingId itself may sometimes be InvalidOid.\n\n> That way if a\n> MAC implementation cares about a particular GUC it'll ensure it's in\n> pg_setting_acl.\n\nA much cleaner solution would be to create new ObjectAccessTypes with a corresponding new Invoke macro and Run function. Those could take setting names, not Oids, and include additional information about whether the operation is SET, RESET or ALTER SYSTEM, what the new value is (if any), what kind of setting it is (bool, int, ...), etc. I don't think such a patch would even be all that hard to write.\n\nWhat do you think?\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Thu, 16 Dec 2021 09:53:45 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Thu, Dec 16, 2021 at 12:53 PM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n>\n>\n>\n> > On Dec 16, 2021, at 7:43 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\n> >\n> > Ah, I understand now. Would it be possible to pass the\n> > SettingAclRelationId if it exists or InvalidOid if not?\n>\n> SettingAclRelationId is always defined, so we can always pass that value. But the settingId itself may sometimes be InvalidOid.\n\nYes, that is what I meant.\n\n> > That way if a\n> > MAC implementation cares about a particular GUC it'll ensure it's in\n> > pg_setting_acl.\n>\n> A much cleaner solution would be to create new ObjectAccessTypes with a corresponding new Invoke macro and Run function. Those could take setting names, not Oids, and include additional information about whether the operation is SET, RESET or ALTER SYSTEM, what the new value is (if any), what kind of setting it is (bool, int, ...), etc. I don't think such a patch would even be all that hard to write.\n>\n> What do you think?\n\nPersonally, I would be happy with that, but since it's a whole new\nhooking method I suspect it'll be an uphill battle. That definitely\nseems like another patchset though, if you do submit this I will test\nand review.\n\nThank you.\n\n\n","msg_date":"Thu, 16 Dec 2021 13:31:58 -0500","msg_from":"Joshua Brindle <joshua.brindle@crunchydata.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Dec 15, 2021, at 7:56 AM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> \n> <v6-0001-Allow-GRANT-of-SET-and-ALTER-SYSTEM-SET-for-gucs.patch>\n\nRebased:\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Wed, 26 Jan 2022 07:24:31 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Jan 26, 2022, at 7:24 AM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> \n> Rebased:\n> \n> <v7-0001-Allow-GRANT-of-SET-and-ALTER-SYSTEM-SET-for-gucs.patch>\n\nHackers,\n\nJoshua and I had some off-list back-and-forth on this patch and I think the attached v8 patch meets his needs. Specifically, v8 adds support for object access hooks in the case of SET and ALTER SYSTEM SET of a guc variable.\n\nThe way InvokeObjectPostAlterHookArg() is written, it expects some things like a Oid that we don't always have. I abused the parameters of that function to pass as much information as possible. Anyone wanting to review this might consider how that is coded.\n\nJoshua, this is byte-for-byte the same patch that you already reviewed off-list (or I attached the wrong patch!) If you are satisfied, could you set the status to ready-for-committer?\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Mon, 31 Jan 2022 11:17:12 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> [ v8-0001-Allow-GRANT-of-SET-and-ALTER-SYSTEM-SET-for-gucs.patch ]\n\nI noticed that this is failing in the cfbot as a side-effect of my\ncc50080a8, so here's a rebase (basically it just removes the no-\nlonger-necessary sanity_check.out hunk).\n\nI haven't reviewed this particularly, or read the thread, but I'm\npretty disturbed by the contrib/ changes that it proposes.\n\n1. If we need to change these two contrib modules, doesn't that imply\na lot of changes forced on external modules as well? What are the\nsecurity implications if somebody doesn't make such a change?\n\n2. It looks to me like if someone installs the updated postgres_fdw.so,\nbut doesn't run ALTER EXTENSION UPDATE, they are going to be left with a\nrather broken extension. This seems not good either, especially if it's\nmultiplied by a boatload of third-party extensions requiring updates.\n\nSo I think some more thought is needed to see if we can't avoid\nthe need to touch extension modules in this patch. Or at least\navoid the need for synchronized C-level and SQL-level updates,\nbecause that is going to create a lot of pain for end users.\n\nI'm also fairly distressed by the number of changes in guc.c,\nmainly because I fear that that means that pending patches that\nadd GUC variables will be subtly incorrect/insecure if they're not\nupdated to account for this. Frankly, I also reject the apparent\nposition that we need to support forbidding users from touching\nmany of these GUCs. Or, if that's our position, why are there\nper-GUC changes at all, rather than just redefining what the\ncontext values mean? (That is, why not redefine USERSET and\nSUSET as simply indicating the default ACL to be applied if there's\nno entry in the catalog.)\n\n\t\t\tregards, tom lane","msg_date":"Sun, 06 Mar 2022 17:13:44 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Mar 6, 2022, at 2:13 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> 1. If we need to change these two contrib modules, doesn't that imply\n> a lot of changes forced on external modules as well? What are the\n> security implications if somebody doesn't make such a change?\n> \n> 2. It looks to me like if someone installs the updated postgres_fdw.so,\n> but doesn't run ALTER EXTENSION UPDATE, they are going to be left with a\n> rather broken extension. This seems not good either, especially if it's\n> multiplied by a boatload of third-party extensions requiring updates.\n> \n> So I think some more thought is needed to see if we can't avoid\n> the need to touch extension modules in this patch. Or at least\n> avoid the need for synchronized C-level and SQL-level updates,\n> because that is going to create a lot of pain for end users.\n> \n> I'm also fairly distressed by the number of changes in guc.c,\n> mainly because I fear that that means that pending patches that\n> add GUC variables will be subtly incorrect/insecure if they're not\n> updated to account for this. Frankly, I also reject the apparent\n> position that we need to support forbidding users from touching\n> many of these GUCs. Or, if that's our position, why are there\n> per-GUC changes at all, rather than just redefining what the\n> context values mean? (That is, why not redefine USERSET and\n> SUSET as simply indicating the default ACL to be applied if there's\n> no entry in the catalog.)\n\nTo my knowledge, there is no mechanism to revoke an implicit privilege. You can revoke a privilege explicitly listed in an aclitem[], but only if the privilege is being tracked that way.\n\nUserset variables are implicitly settable by any user. There was a request, off-list as I recall, to make it possible to revoke the privilege to set variables such as \"work_mem\". To make that possible, but not change the default behavior vis-a-vis prior releases, I upgraded most userset variables to suset with a corresponding grant to public on the variable. Sites which wish to have a more restrictive policy on such variables can revoke that privilege from public and instead issue more restrictive grants. There were a few variables where such treatment didn't seem sensible, such as ones to do with client connections, and I left them alone. I didn't insist on a defense for why any particular setting needed to be revocable in order to apply this treatment. My assumption was that sites should be allowed to determine their own security policies per setting unless there is a technical difficulty for a given setting that would make it overly burdensome to implement.\n\nIt seemed more complete to do the same thing for contrib modules, so I did.\n\nIf a third-party module doesn't do the equivalent operation, they would simply continue with their userset variables working as before. I don't see a specific security concern with that, though I'd be happy to consider arguments to the contrary.\n\nI believe there are also some issues with this patch relating to pg_dump/pg_restore and to pg_upgrade, which EDB is working to address. We intend to repost something soon.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Sun, 6 Mar 2022 14:40:50 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n>> On Mar 6, 2022, at 2:13 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> ... Or, if that's our position, why are there\n>> per-GUC changes at all, rather than just redefining what the\n>> context values mean? (That is, why not redefine USERSET and\n>> SUSET as simply indicating the default ACL to be applied if there's\n>> no entry in the catalog.)\n\n> To my knowledge, there is no mechanism to revoke an implicit privilege. You can revoke a privilege explicitly listed in an aclitem[], but only if the privilege is being tracked that way.\n\nSo? What I'm suggesting is along the lines of\n\n(1) pg_setting_acl starts out empty, or at least mostly empty (maybe\nthere are a few GUCs that need custom values).\n\n(2) If there's a pg_setting_acl entry for a GUC that's to be set,\nwe apply it: either it grants the desired permission or it doesn't.\n\n(3) If there's no entry, then for a USERSET GUC we assume that the\nentry would be like \"=s/postgres\", while for any other context value\nwe assume the ACL grants nothing.\n\nI don't think this is materially different from what we do with\npermissions on (say) functions. If you want to revoke the public\nSET privilege on some USERSET variable, you instantiate the default\nand then revoke. You end up with an empty ACL stored in pg_setting_acl,\nand voila.\n\nIt'd likely be necessary to refuse to record a grant/revoke on\nan unknown GUC, since if we don't know the GUC then we can't know\nwhat the relevant default ACL ought to be. But I bet your existing\npatch has some dubious behavior in that case too.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 06 Mar 2022 17:57:10 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Mar 6, 2022, at 2:57 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> I don't think this is materially different from what we do with\n> permissions on (say) functions. If you want to revoke the public\n> SET privilege on some USERSET variable, you instantiate the default\n> and then revoke. You end up with an empty ACL stored in pg_setting_acl,\n> and voila.\n\nI assume you mean the implementation of REVOKE does this, not that the user needs to do both a grant and a revoke.\n\n> It'd likely be necessary to refuse to record a grant/revoke on\n> an unknown GUC, since if we don't know the GUC then we can't know\n> what the relevant default ACL ought to be. But I bet your existing\n> patch has some dubious behavior in that case too.\n\nThe existing patch allows grants on unknown gucs, because it can't know what guc an upgrade script will introduce, and the grant statement may need to execute before the guc exists. That opens a window for granting privileges on non-existent gucs. That sounds bad, but I don't know of any actual harm that it does, beyond just being ugly.\n\nWith your proposal, it sounds like we could avoid that ugliness, so I'm inclined to at least try doing it as you propose.\n\nThanks for the suggestion!\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Sun, 6 Mar 2022 15:06:43 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> On Mar 6, 2022, at 2:57 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> I don't think this is materially different from what we do with\n>> permissions on (say) functions. If you want to revoke the public\n>> SET privilege on some USERSET variable, you instantiate the default\n>> and then revoke. You end up with an empty ACL stored in pg_setting_acl,\n>> and voila.\n\n> I assume you mean the implementation of REVOKE does this, not that the user needs to do both a grant and a revoke.\n\nRight. Again, look at what happens when you create a function and\nthen revoke its default PUBLIC EXECUTE permission.\n\n>> It'd likely be necessary to refuse to record a grant/revoke on\n>> an unknown GUC, since if we don't know the GUC then we can't know\n>> what the relevant default ACL ought to be. But I bet your existing\n>> patch has some dubious behavior in that case too.\n\n> The existing patch allows grants on unknown gucs, because it can't know what guc an upgrade script will introduce, and the grant statement may need to execute before the guc exists.\n\nYeah, that's the problematic case. It might mostly work to assume that\nan unknown GUC has an empty default ACL. This could fail to retain the\ndefault PUBLIC SET permission if it later turns out the GUC is USERSET\n... but I suspect in most cases anybody who's messing with the permissions\nwould've started out by revoking that anyway. We could make this\ndefinitely work in pg_dump if we teach pg_dump to explicitly grant or\nrevoke the PUBLIC SET permission anytime it's emitting anything for\na GUC, even if it thinks that would be the default state anyway.\nExtension scripts that want to modify permissions for their GUCs\ncould follow that same principle to be sure.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 06 Mar 2022 18:27:43 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Mar 6, 2022, at 3:27 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\n>> The existing patch allows grants on unknown gucs, because it can't know what guc an upgrade script will introduce, and the grant statement may need to execute before the guc exists.\n> \n> Yeah, that's the problematic case. It might mostly work to assume that\n> an unknown GUC has an empty default ACL. This could fail to retain the\n> default PUBLIC SET permission if it later turns out the GUC is USERSET\n\nOn further reflection, I concluded this isn't needed. No current extension, whether in-core or third party, expects to be able to create a new GUC and then grant or revoke permissions on it. They can already specify the guc context (PGC_USERS, etc). Introducing a feature that depends on the dubious assumption that unrecognized GUCs will turn out to be USERSET doesn't seem warranted.\n\nThe patch attributes all grants of setting privileges to the bootstrap superuser. Only superusers can grant or revoke privileges on settings, and all settings are implicitly owned by the bootstrap superuser because there is no explicit owner associated with settings. Consequently, select_best_grantor(some_superuser, ..., BOOTSTRAP_SUPERUSERID, ...) always chooses the bootstrap superuser. I don't see a problem with this, but wouldn't mind a second opinion. Some people might find it surprising when viewing the pg_setting_acl.setacl field.\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Tue, 15 Mar 2022 13:59:08 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 3/15/22 16:59, Mark Dilger wrote:\n>> On Mar 6, 2022, at 3:27 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>\n>> Mark Dilger <mark.dilger@enterprisedb.com> writes:\n>>> The existing patch allows grants on unknown gucs, because it can't know what guc an upgrade script will introduce, and the grant statement may need to execute before the guc exists.\n>> Yeah, that's the problematic case. It might mostly work to assume that\n>> an unknown GUC has an empty default ACL. This could fail to retain the\n>> default PUBLIC SET permission if it later turns out the GUC is USERSET\n> On further reflection, I concluded this isn't needed. No current extension, whether in-core or third party, expects to be able to create a new GUC and then grant or revoke permissions on it. They can already specify the guc context (PGC_USERS, etc). Introducing a feature that depends on the dubious assumption that unrecognized GUCs will turn out to be USERSET doesn't seem warranted.\n\nAgreed.\n\n\n>\n> The patch attributes all grants of setting privileges to the bootstrap superuser. Only superusers can grant or revoke privileges on settings, and all settings are implicitly owned by the bootstrap superuser because there is no explicit owner associated with settings. Consequently, select_best_grantor(some_superuser, ..., BOOTSTRAP_SUPERUSERID, ...) always chooses the bootstrap superuser. I don't see a problem with this, but wouldn't mind a second opinion. Some people might find it surprising when viewing the pg_setting_acl.setacl field.\n\n\nI think it's OK as long as we document it. An alternative might be to\ninvent a pseudo-superuser called, say, 'postgres_system', but that seems\nlike overkill to solve what is in effect a cosmetic problem.\n\nGenerally I think this is now in fairly good shape, I've played with it\nand it seems to do what I expect in every case, and the things I found\nsurprising are gone.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 16 Mar 2022 07:43:28 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Andrew Dunstan <andrew@dunslane.net> writes:\n> Generally I think this is now in fairly good shape, I've played with it\n> and it seems to do what I expect in every case, and the things I found\n> surprising are gone.\n\nStepping back a bit ... do we really want to institutionalize the\nterm \"setting\" for GUC variables? I realize that the view pg_settings\nexists, but the documentation generally prefers the term \"configuration\nparameters\". Where config.sgml uses \"setting\" as a noun, it's usually\ntalking about a specific concrete value for a parameter, and you can\nargue that the view's name comports with that meaning. But you can't\nGRANT a parameter's current value.\n\nI don't have a better name to offer offhand --- I surely am not proposing\nthat we change the syntax to be \"GRANT ... ON CONFIGURATION PARAMETER x\",\nbecause that's way too wordy. But now is the time to bikeshed if we're\ngonna bikeshed, or else admit that we're not interested in precise\nvocabulary.\n\nI'm also fairly allergic to the way that this patch has decided to assign\nmulti-word names to privilege types (ie SET VALUE, ALTER SYSTEM). There\nis no existing precedent for that, and I think it's going to break\nclient-side code that we don't need to break. It's not coincidental that\nthis forces weird changes in rules about whitespace in the has_privilege\nfunctions, for example; and if you think that isn't going to cause\nproblems I think you are wrong. Perhaps we could just use \"SET\" and\n\"ALTER\", or \"SET\" and \"SYSTEM\"?\n\nI also agree with the upthread criticism that this is abusing the\nObjectPostAlterHook API unreasonably much. In particular, it's\ntrying to use pg_setting_acl OIDs as identities for GUCs, which\nthey are not, first because they're unstable and second because\nmost GUCs won't even have an entry there. I therefore judge the\nhook calls added to ExecSetVariableStmt and AlterSystemSetConfigFile\nto be 100% useless, in fact probably counterproductive because they\nintroduce a boatload of worries about whether the right things happen\nif the hook errors out or does something guc.c isn't expecting.\n\nI suggest that what might be saner is to consider that the \"objects\"\nthat the hook calls are concerned with are the pg_setting_acl entries,\nnot the underlying GUCs, and thus that the hooks need be invoked only\nwhen creating, destroying or altering those entries. If we do have\na need for a hook editorializing on GUC value settings, that would\nhave to be an independent API --- but I have heard no calls for\nthe ability to have such a hook, and I don't think that this patch\nis the place to introduce one.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 16 Mar 2022 14:47:24 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Mar 16, 2022, at 11:47 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Stepping back a bit ... do we really want to institutionalize the\n> term \"setting\" for GUC variables? I realize that the view pg_settings\n> exists, but the documentation generally prefers the term \"configuration\n> parameters\". Where config.sgml uses \"setting\" as a noun, it's usually\n> talking about a specific concrete value for a parameter, and you can\n> argue that the view's name comports with that meaning. But you can't\n> GRANT a parameter's current value.\n> \n> I don't have a better name to offer offhand --- I surely am not proposing\n> that we change the syntax to be \"GRANT ... ON CONFIGURATION PARAMETER x\",\n> because that's way too wordy. But now is the time to bikeshed if we're\n> gonna bikeshed, or else admit that we're not interested in precise\n> vocabulary.\n\nInformally, we often use \"GUC\" on this list, but that isn't used formally, leaving \"configuration parameter\" and \"setting\" as the two obvious choices. I preferred \"configuration parameter\" originally and was argued out of it. My take on \"setting\" was also that it more naturally refers to the choice of setting, not the thing being set, such that \"work_mem = 8192\" means the configuration parameter \"work_mem\" has the setting \"8192\". I'm happy to change the patch along these lines, but I vaguely recall it being Robert who liked the \"setting\" language. Robert? (Sorry if I misremember that...)\n\n> I'm also fairly allergic to the way that this patch has decided to assign\n> multi-word names to privilege types (ie SET VALUE, ALTER SYSTEM). There\n> is no existing precedent for that, and I think it's going to break\n> client-side code that we don't need to break. It's not coincidental that\n> this forces weird changes in rules about whitespace in the has_privilege\n> functions, for example; and if you think that isn't going to cause\n> problems I think you are wrong. Perhaps we could just use \"SET\" and\n> \"ALTER\", or \"SET\" and \"SYSTEM\"?\n\nI chose those particular multi-word names to avoid needing to promote any keywords. That's been a while ago, and I don't recall exactly where the shift-reduce or reduce-reduce errors were coming from. I'll play with it to see what I can find.\n\n> I also agree with the upthread criticism that this is abusing the\n> ObjectPostAlterHook API unreasonably much. In particular, it's\n> trying to use pg_setting_acl OIDs as identities for GUCs, which\n> they are not, first because they're unstable and second because\n> most GUCs won't even have an entry there. I therefore judge the\n> hook calls added to ExecSetVariableStmt and AlterSystemSetConfigFile\n> to be 100% useless, in fact probably counterproductive because they\n> introduce a boatload of worries about whether the right things happen\n> if the hook errors out or does something guc.c isn't expecting.\n\nI think Joshua was planning to use these hooks for security purposes. The hooks are supposed to check whether the Oid is valid, and if not, still be able to make choices based on the other information. Joshua, any comment on this?\n\n> I suggest that what might be saner is to consider that the \"objects\"\n> that the hook calls are concerned with are the pg_setting_acl entries,\n> not the underlying GUCs, and thus that the hooks need be invoked only\n> when creating, destroying or altering those entries.\n\nThat's certainly a thing we could do, but I got the impression that Joshua wanted to hook into SET, RESET, and ALTER SYSTEM SET, not merely into GRANT and REVOKE.\n\n> If we do have\n> a need for a hook editorializing on GUC value settings, that would\n> have to be an independent API --- but I have heard no calls for\n> the ability to have such a hook, and I don't think that this patch\n> is the place to introduce one.\n\nWell, the call for it was in this thread, but I'm ok with yanking out that part of the patch if it seems half-baked.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Wed, 16 Mar 2022 11:59:06 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"<snip>\n\n> I suggest that what might be saner is to consider that the \"objects\"\n> that the hook calls are concerned with are the pg_setting_acl entries,\n> not the underlying GUCs, and thus that the hooks need be invoked only\n> when creating, destroying or altering those entries. If we do have\n> a need for a hook editorializing on GUC value settings, that would\n> have to be an independent API --- but I have heard no calls for\n> the ability to have such a hook, and I don't think that this patch\n> is the place to introduce one.\n\nI requested it here:\nhttps://www.postgresql.org/message-id/CAGB%2BVh5pVFAqw8YzeXy4xxmEt_4Hq_8pEUHdCQvv3mCjvC-S-w%40mail.gmail.com\n\nwith compromises here:\nhttps://www.postgresql.org/message-id/CAGB%2BVh6wLJ3FKsno62fi54pfg0FDrZRWcpuuCJBkHcCj-G1ndw%40mail.gmail.com\nhttps://www.postgresql.org/message-id/0A3D3CBA-6548-4C9E-9F46-59D5C51A1F31%40enterprisedb.com\nhttps://www.postgresql.org/message-id/CAGB%2BVh65R5vKC4rEt7r2_pK3kMZd-VY0n99RJwcP8Bic7xvOxQ%40mail.gmail.com\n\nand the new version here:\nhttps://www.postgresql.org/message-id/C8DF19D8-C15D-4C2D-91CA-391390F1E421%40enterprisedb.com\n\nwhich I wrote. a toy module to test and was satisfied with it, despite\nthe limitations.\n\nIf we are adding a DAC Grant for a type of object it seems untenable\nthat it would not come with analogous MAC capable hooks.\n\nThank you.\n\n\n","msg_date":"Wed, 16 Mar 2022 15:00:09 -0400","msg_from":"Joshua Brindle <joshua.brindle@crunchydata.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> On Mar 16, 2022, at 11:47 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> ... I therefore judge the\n>> hook calls added to ExecSetVariableStmt and AlterSystemSetConfigFile\n>> to be 100% useless, in fact probably counterproductive because they\n>> introduce a boatload of worries about whether the right things happen\n>> if the hook errors out or does something guc.c isn't expecting.\n\n> I think Joshua was planning to use these hooks for security purposes. The hooks are supposed to check whether the Oid is valid, and if not, still be able to make choices based on the other information. Joshua, any comment on this?\n\nIt's going to be hard to do anything useful in a hook that (a) does\nnot know which GUC is being assigned to and (b) cannot do catalog\naccesses for fear that we're not inside a transaction. (b), in\nparticular, seems like a rather thorough API break; up to now\nObjectPostAlter hooks could assume that catalog accesses are OK.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 16 Mar 2022 15:06:04 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Wed, Mar 16, 2022 at 3:06 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> > On Mar 16, 2022, at 11:47 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >> ... I therefore judge the\n> >> hook calls added to ExecSetVariableStmt and AlterSystemSetConfigFile\n> >> to be 100% useless, in fact probably counterproductive because they\n> >> introduce a boatload of worries about whether the right things happen\n> >> if the hook errors out or does something guc.c isn't expecting.\n>\n> > I think Joshua was planning to use these hooks for security purposes. The hooks are supposed to check whether the Oid is valid, and if not, still be able to make choices based on the other information. Joshua, any comment on this?\n>\n> It's going to be hard to do anything useful in a hook that (a) does\n> not know which GUC is being assigned to and (b) cannot do catalog\n> accesses for fear that we're not inside a transaction. (b), in\n> particular, seems like a rather thorough API break; up to now\n> ObjectPostAlter hooks could assume that catalog accesses are OK.\n>\n\nCan you elaborate on this point? This is perhaps an area where I don't\nknow the rules of the road, to test this hook I modified the set_user\nextension, which normally uses the parse tree to figure out if someone\nis trying to SET log_statement or ALTER SYSTEM log_statement and\nreplaced it with:\n\n case OAT_POST_ALTER:\n if (classId == SettingAclRelationId)\n {\n ObjectAddress object;\n object.classId = SettingAclRelationId;\n object.objectId = objectId;\n object.objectSubId = 0;\n if (strcmp(getObjectIdentity(&object),\n\"log_statement\") == 0)\n {\n ereport(ERROR,\n (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),\n errmsg(\"Setting %s blocked\",\ngetObjectIdentity(&object))));\n }\n }\n\nIs that inherently unsafe?\n\n\n","msg_date":"Wed, 16 Mar 2022 15:36:21 -0400","msg_from":"Joshua Brindle <joshua.brindle@crunchydata.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Joshua Brindle <joshua.brindle@crunchydata.com> writes:\n> On Wed, Mar 16, 2022 at 3:06 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> It's going to be hard to do anything useful in a hook that (a) does\n>> not know which GUC is being assigned to and (b) cannot do catalog\n>> accesses for fear that we're not inside a transaction. (b), in\n>> particular, seems like a rather thorough API break; up to now\n>> ObjectPostAlter hooks could assume that catalog accesses are OK.\n\n> Can you elaborate on this point?\n\nHmm, I glossed over too many details there perhaps. I was thinking\nabout the restrictions on GUC check_hooks, which can be run outside\na transaction altogether. But that's not quite relevant here.\nExecSetVariableStmt can assume it's inside a transaction, but what\nit *can't* assume is that we've set a transaction snapshot as yet\n(cf. PlannedStmtRequiresSnapshot). If we call an ObjectPostAlter hook\nthere, and it does a catalog access, that's going to break things for\nmodifications of GUCs that are supposed to be modifiable without\nfreezing the transaction snapshot. So the net result is the same:\ncatalog access not okay, at least not in general.\n\nBetween that and the fact that an OID-based API is largely useless here,\nI don't think it's sane to try to use the existing ObjectPostAlter API.\nMaybe there is a case for inventing a separate hook API that could pass\nthe GUC name as a string, instead. I remain of the opinion that this\npatch should not concern itself with that, though.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 16 Mar 2022 15:58:57 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 3/16/22 14:47, Tom Lane wrote:\n> Andrew Dunstan <andrew@dunslane.net> writes:\n>> Generally I think this is now in fairly good shape, I've played with it\n>> and it seems to do what I expect in every case, and the things I found\n>> surprising are gone.\n> Stepping back a bit ... do we really want to institutionalize the\n> term \"setting\" for GUC variables? I realize that the view pg_settings\n> exists, but the documentation generally prefers the term \"configuration\n> parameters\". Where config.sgml uses \"setting\" as a noun, it's usually\n> talking about a specific concrete value for a parameter, and you can\n> argue that the view's name comports with that meaning. But you can't\n> GRANT a parameter's current value.\n>\n> I don't have a better name to offer offhand --- I surely am not proposing\n> that we change the syntax to be \"GRANT ... ON CONFIGURATION PARAMETER x\",\n> because that's way too wordy. But now is the time to bikeshed if we're\n> gonna bikeshed, or else admit that we're not interested in precise\n> vocabulary.\n>\n> I'm also fairly allergic to the way that this patch has decided to assign\n> multi-word names to privilege types (ie SET VALUE, ALTER SYSTEM). There\n> is no existing precedent for that, and I think it's going to break\n> client-side code that we don't need to break. It's not coincidental that\n> this forces weird changes in rules about whitespace in the has_privilege\n> functions, for example; and if you think that isn't going to cause\n> problems I think you are wrong. Perhaps we could just use \"SET\" and\n> \"ALTER\", or \"SET\" and \"SYSTEM\"?\n\n\nThat's going to look weird, ISTM. This is less clear about what it's\ngranting.\n\n GRANT ALTER ON SOMETHING shared_buffers TO myuser;\n\nIf you don't like that maybe ALTER_SYSTEM and SET_VALUE would work,\nalthough mostly we have avoided things like that.\n\nHow about MODIFY instead of SET VALUE and CONFIGURE instead of ALTER SYSTEM?\n\nPersonally I don't have problem with the use of SETTING. I think the\nmeaning is pretty plain in context and unlikely to produce any confusion.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 16 Mar 2022 16:07:43 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Andrew Dunstan <andrew@dunslane.net> writes:\n> On 3/16/22 14:47, Tom Lane wrote:\n>> I'm also fairly allergic to the way that this patch has decided to assign\n>> multi-word names to privilege types (ie SET VALUE, ALTER SYSTEM). There\n>> is no existing precedent for that, and I think it's going to break\n>> client-side code that we don't need to break. It's not coincidental that\n>> this forces weird changes in rules about whitespace in the has_privilege\n>> functions, for example; and if you think that isn't going to cause\n>> problems I think you are wrong. Perhaps we could just use \"SET\" and\n>> \"ALTER\", or \"SET\" and \"SYSTEM\"?\n\n> That's going to look weird, ISTM. This is less clear about what it's\n> granting.\n> GRANT ALTER ON SOMETHING shared_buffers TO myuser;\n\nTrue. I think \"GRANT SET\" is clear enough, and it fits with the custom of\nusing the name of the SQL statement that the privilege allows you to\ninvoke. (I gather from Mark's comments that Bison gave him problems with\nthat, but maybe that can be dealt with.) But I concede that \"ALTER\" by\nitself is pretty vague.\n\n> If you don't like that maybe ALTER_SYSTEM and SET_VALUE would work,\n> although mostly we have avoided things like that.\n> How about MODIFY instead of SET VALUE and CONFIGURE instead of ALTER SYSTEM?\n\nI thought about ALTER_SYSTEM too. It's not great but maybe the best we\ncan do. Not sure that CONFIGURE is better.\n\n> Personally I don't have problem with the use of SETTING. I think the\n> meaning is pretty plain in context and unlikely to produce any confusion.\n\nI'm just unhappy about the disconnect with the documentation. I wonder\nif we could get away with s/configuration parameter/setting/g in the docs.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 16 Mar 2022 16:53:02 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 3/16/22 16:53, Tom Lane wrote:\n\n>> Personally I don't have problem with the use of SETTING. I think the\n>> meaning is pretty plain in context and unlikely to produce any confusion.\n> I'm just unhappy about the disconnect with the documentation. I wonder\n> if we could get away with s/configuration parameter/setting/g in the docs.\n>\n> \t\t\t\n\n\nI don't think we need to fix that here though. If you can live with\nSETTING for now I will undertake to fix the docs post-feature-freeze if\nnecessary.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 16 Mar 2022 17:09:35 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"<snip>\n\n> I remain of the opinion that this\n> patch should not concern itself with that, though.\n\nSo you are saying that people can add new object types to PG with DAC\npermissions and not concern themselves with MAC capable hooks? Is that\nan official PG community stance?\n\n\n","msg_date":"Thu, 17 Mar 2022 09:24:54 -0400","msg_from":"Joshua Brindle <joshua.brindle@crunchydata.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On 16.03.22 19:59, Mark Dilger wrote:\n> Informally, we often use \"GUC\" on this list, but that isn't used formally, leaving \"configuration parameter\" and \"setting\" as the two obvious choices. I preferred \"configuration parameter\" originally and was argued out of it. My take on \"setting\" was also that it more naturally refers to the choice of setting, not the thing being set, such that \"work_mem = 8192\" means the configuration parameter \"work_mem\" has the setting \"8192\".\n\n\"The current setting of the work_mem parameter is 8192.\"\n\nI think something based on \"parameter\" is good. We also use that \nlanguage in the protocol (e.g., ParameterStatus).\n\n\n","msg_date":"Thu, 17 Mar 2022 14:52:34 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On 16.03.22 19:47, Tom Lane wrote:\n> I'm also fairly allergic to the way that this patch has decided to assign\n> multi-word names to privilege types (ie SET VALUE, ALTER SYSTEM). There\n> is no existing precedent for that, and I think it's going to break\n> client-side code that we don't need to break. It's not coincidental that\n> this forces weird changes in rules about whitespace in the has_privilege\n> functions, for example; and if you think that isn't going to cause\n> problems I think you are wrong. Perhaps we could just use \"SET\" and\n> \"ALTER\", or \"SET\" and \"SYSTEM\"?\n\nI think Oracle and MS SQL Server have many multi-word privilege names. \nSo users are quite used to that. And if we want to add more complex \nprivileges, we might run out of sensible single words eventually. So I \nwould not exclude this approach.\n\n\n","msg_date":"Thu, 17 Mar 2022 14:54:57 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> On 16.03.22 19:47, Tom Lane wrote:\n>> ... Perhaps we could just use \"SET\" and\n>> \"ALTER\", or \"SET\" and \"SYSTEM\"?\n\n> I think Oracle and MS SQL Server have many multi-word privilege names. \n> So users are quite used to that. And if we want to add more complex \n> privileges, we might run out of sensible single words eventually. So I \n> would not exclude this approach.\n\nWell, I still say that \"SET\" is sufficient for the one privilege name\n(unless we really can't make Bison handle that, which I doubt). But\nI'm willing to yield on using \"ALTER SYSTEM\" for the other.\n\nIf we go with s/SETTING/PARAMETER/ as per your other message, then\nthat would be adequately consistent with the docs I think. So it'd\nbe\n\nGRANT { SET | ALTER SYSTEM } ON PARAMETER foo TO ...\n\nand the new catalog would be pg_parameter_acl, and so on.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 17 Mar 2022 10:47:03 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 3/17/22 10:47, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>> On 16.03.22 19:47, Tom Lane wrote:\n>>> ... Perhaps we could just use \"SET\" and\n>>> \"ALTER\", or \"SET\" and \"SYSTEM\"?\n>> I think Oracle and MS SQL Server have many multi-word privilege names. \n>> So users are quite used to that. And if we want to add more complex \n>> privileges, we might run out of sensible single words eventually. So I \n>> would not exclude this approach.\n> Well, I still say that \"SET\" is sufficient for the one privilege name\n> (unless we really can't make Bison handle that, which I doubt). But\n> I'm willing to yield on using \"ALTER SYSTEM\" for the other.\n>\n> If we go with s/SETTING/PARAMETER/ as per your other message, then\n> that would be adequately consistent with the docs I think. So it'd\n> be\n>\n> GRANT { SET | ALTER SYSTEM } ON PARAMETER foo TO ...\n>\n> and the new catalog would be pg_parameter_acl, and so on.\n>\n> \t\t\t\n\n\n\nThe upside of this is that it avoids the inelegant\n\n\n GRANT SET ON SETTING ...\n\n\nBut I was just looking again at the grammar, and the only reason we need\nthis keyword at all AFAICS is to disambiguate ALL [PRIVILEGES] cases. \nIf we abandoned that for this form of GRANT/REVOKE I think we could\nprobably get away with\n\n\n GRANT { SET | ALTER SYSTEM } ON setting_name ...\n\n\nI haven't tried it, so I could be all wrong.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Thu, 17 Mar 2022 11:41:38 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Thu, Mar 17, 2022 at 9:25 AM Joshua Brindle\n<joshua.brindle@crunchydata.com> wrote:\n> <snip>\n>\n> > I remain of the opinion that this\n> > patch should not concern itself with that, though.\n>\n> So you are saying that people can add new object types to PG with DAC\n> permissions and not concern themselves with MAC capable hooks? Is that\n> an official PG community stance?\n\nI don't know that the community has an official position on that\ntopic, but I do not think it's reasonable to expect everyone who\ntinkers with MAC permissions to try to make a corresponding equivalent\nfor DAC. The number of people using PostgreSQL with DAC is relatively\nsmall, and the topic is extremely complicated, and a lot of hackers\ndon't really understand it well enough to be sure that whatever they\nmight do is right. I think it's reasonable to expect people who\nunderstand DAC and care about it to put some energy into the topic,\nand not just in terms of telling other people how they have to write\ntheir patches.\n\nI *don't* think it's appropriate for a patch that touches MAC to\ndeliberately sabotage the existing support we have for DAC or to just\nignore it where the right thing to do is obvious. But maintaining a\nmillion lines of code is a lot of work, and I can't think of any\nreason why the burden of maintaining relatively little-used features\nshould fall entirely on people who don't care about them.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 17 Mar 2022 12:04:42 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Wed, Mar 16, 2022 at 2:47 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Stepping back a bit ... do we really want to institutionalize the\n> term \"setting\" for GUC variables? I realize that the view pg_settings\n> exists, but the documentation generally prefers the term \"configuration\n> parameters\". Where config.sgml uses \"setting\" as a noun, it's usually\n> talking about a specific concrete value for a parameter, and you can\n> argue that the view's name comports with that meaning. But you can't\n> GRANT a parameter's current value.\n\nI agree that the lack of a good user-friendly term for GUCs is a real\nproblem. Here at EDB I've observed even relatively non-technical\npeople using that term, which appears nowhere in the documentation and\nis utterly unintelligible to a typical end-user. Somebody gets on the\nphone and tells the customer that they need to set a GUC and the\ncustomer is like \"what's a guck?\" except that they probably don't\nactually ask that question but are just confused and fail to\nunderstand that a postgresql.conf change is being proposed. I hate it.\nIt sucks.\n\nI have sort of been trying to promote the use of the word \"setting\"\nand use it in my own writing, especially to end-users. That is\ndefinitely more intelligible to random users, but it's admittedly also\nawkward. \"Set a setting\" just sounds redundant. But \"set a\nconfiguration variable\" sounds wordy, so I don't know.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 17 Mar 2022 12:18:51 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Mar 17, 2022, at 9:04 AM, Robert Haas <robertmhaas@gmail.com> wrote:\n> \n> not just in terms of telling other people how they have to write\n> their patches.\n\n...\n\n> the burden of maintaining relatively little-used features\n> should fall entirely on people who don't care about them.\n\nJoshua helped test the DAC portion of this patch, and answered a number of my questions on the topic, including in-person back in December. I take your point, Robert, on the general principle, but the archives should reflect that Joshua did contribute to this specific patch.\n\nJoshua, should we drop the DAC portion for postgres 15 and revisit the issue for postgres 16? I think it's getting late in the development cycle to attempt what Tom described upthread, and I'd hate to see the rest of this patch punted.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Thu, 17 Mar 2022 09:19:24 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Thu, Mar 17, 2022 at 12:19 PM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n> Joshua helped test the DAC portion of this patch, and answered a number of my questions on the topic, including in-person back in December. I take your point, Robert, on the general principle, but the archives should reflect that Joshua did contribute to this specific patch.\n\nI wasn't intending to say otherwise, and if the changes needed for DAC\nhere are straightforward and don't make the patch significantly harder\nto finish, then I would say Tom is wrong and we should just go ahead\nand make them. But if they do, then I think it's perfectly fine to say\nthat we're going to leave that alone and let someone with\nsubject-matter expertise sort it out when they have time. We do that\nall the time with other things, most notably MSVC builds, where we\njust can't expect every hacker or even every committer to understand\nexactly what's required to make it all work. I try my best to commit\nthings that don't break it, but sometimes I do, and then Andrew helps\nsort it out, because he understands it and I don't. I think DAC should\nfall into that kind of category as well.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 17 Mar 2022 12:29:11 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Thu, Mar 17, 2022 at 12:04 PM Robert Haas <robertmhaas@gmail.com> wrote:\n>\n> On Thu, Mar 17, 2022 at 9:25 AM Joshua Brindle\n> <joshua.brindle@crunchydata.com> wrote:\n> > <snip>\n> >\n> > > I remain of the opinion that this\n> > > patch should not concern itself with that, though.\n> >\n> > So you are saying that people can add new object types to PG with DAC\n> > permissions and not concern themselves with MAC capable hooks? Is that\n> > an official PG community stance?\n>\n> I don't know that the community has an official position on that\n> topic, but I do not think it's reasonable to expect everyone who\n> tinkers with MAC permissions to try to make a corresponding equivalent\n> for DAC. The number of people using PostgreSQL with DAC is relatively\n> small, and the topic is extremely complicated, and a lot of hackers\n> don't really understand it well enough to be sure that whatever they\n> might do is right. I think it's reasonable to expect people who\n> understand DAC and care about it to put some energy into the topic,\n> and not just in terms of telling other people how they have to write\n> their patches.\n>\n> I *don't* think it's appropriate for a patch that touches MAC to\n> deliberately sabotage the existing support we have for DAC or to just\n> ignore it where the right thing to do is obvious. But maintaining a\n> million lines of code is a lot of work, and I can't think of any\n> reason why the burden of maintaining relatively little-used features\n> should fall entirely on people who don't care about them.\n\nI agree with this view, outside of the mixup between MAC and DAC (DAC\nis in-core, MAC is via hooks)\n\nSo there should be some committers or contributors that keep an eye\nout and make sure new objects have appropriate hooks, and since an Oid\nbased hook for GUCs is inappropriate, I hope this patch can be rolled\ninto the contribution from Mark.\n\nThe attached patch adds a set of hooks for strings, and changes the\nGUC patch from Mark to use it, I tested with this code:\n\nvoid set_user_object_access_str(ObjectAccessType access, Oid classId,\nconst char *objName, int subId, void *arg)\n{\nif (next_object_access_hook_str)\n{\n(*next_object_access_hook_str)(access, classId, objName, subId, arg);\n}\nswitch (access)\n{\ncase OAT_POST_ALTER:\nif (classId == SettingAclRelationId)\n{\nOid userid = GetUserId();\nbool is_superuser = superuser_arg(userid);\nchar *username = GETUSERNAMEFROMID(GetUserId());\nconst char *setting = \"setting value\";\nconst char *altering = \"altering system\";\nconst char *unknown = \"unknown\";\nconst char *action;\nif (subId && ACL_SET_VALUE)\naction = setting;\nelse if (subId && ACL_ALTER_SYSTEM)\naction = altering;\nelse\naction = unknown;\n\nelog(WARNING, \"%s is %s %s (%d)\", username, action, objName, is_superuser);\nif (strcmp(objName, \"log_statement\") == 0)\n{\nereport(ERROR,\n(errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),\nerrmsg(\"Setting %s blocked\", objName)));\n}\n}\ndefault:\nbreak;\n}\n}\n\nstatic void\nset_user_object_access (ObjectAccessType access, Oid classId, Oid\nobjectId, int subId, void *arg)\n{\nif (next_object_access_hook)\n{\n(*next_object_access_hook)(access, classId, objectId, subId, arg);\n}\nswitch (access)\n{\ncase OAT_FUNCTION_EXECUTE:\n{\n/* Update the `set_config` Oid cache if necessary. */\nset_user_cache_proc(InvalidOid);\n\n/* Now see if this function is blocked */\nset_user_block_set_config(objectId);\nbreak;\n}\n\n/* fallthrough */\n\ncase OAT_POST_CREATE:\n{\nif (classId == ProcedureRelationId)\n{\nset_user_cache_proc(objectId);\n}\nbreak;\n}\ndefault:\nbreak;\n}\n}","msg_date":"Thu, 17 Mar 2022 12:29:53 -0400","msg_from":"Joshua Brindle <joshua.brindle@crunchydata.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Thu, Mar 17, 2022 at 12:30 PM Joshua Brindle\n<joshua.brindle@crunchydata.com> wrote:\n> I agree with this view, outside of the mixup between MAC and DAC (DAC\n> is in-core, MAC is via hooks)\n\nAn excellent point! Exactly why we need expert-level help with this stuff! :-)\n\n> So there should be some committers or contributors that keep an eye\n> out and make sure new objects have appropriate hooks, and since an Oid\n> based hook for GUCs is inappropriate, I hope this patch can be rolled\n> into the contribution from Mark.\n\nI'm not going to take a position on this patch right this minute, but\nI appreciate you providing it.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 17 Mar 2022 12:34:25 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Mar 17, 2022, at 9:29 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\n> \n> I hope this patch can be rolled\n> into the contribution from Mark.\n\nWorking on it.... Thanks for the patch!\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Thu, 17 Mar 2022 09:36:49 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Thu, Mar 17, 2022 at 12:36 PM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n> > On Mar 17, 2022, at 9:29 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\n> >\n> > I hope this patch can be rolled\n> > into the contribution from Mark.\n>\n> Working on it.... Thanks for the patch!\n\nGreat, thanks.\n\nI missed one objectId reference (InvokeObjectDropHookStr), fixed\nversion attached.","msg_date":"Thu, 17 Mar 2022 14:19:48 -0400","msg_from":"Joshua Brindle <joshua.brindle@crunchydata.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Mar 17, 2022, at 8:41 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\n> \n> If we abandoned that for this form of GRANT/REVOKE I think we could\n> probably get away with\n> \n> \n> GRANT { SET | ALTER SYSTEM } ON setting_name ...\n> \n> \n> I haven't tried it, so I could be all wrong.\n\nVersion 12 of the patch uses SET and ALTER SYSTEM as the names of the privileges, and PARAMETER as the name of the thing on which the privilege is granted. The catalog table which tracks these grants is now named pg_parameter_acl, and various other parts of the patch have been adjusted to use a \"parameter\" based, rather than a \"setting\" based, naming scheme. One exception to this rule is the \"setacl\" column in pg_parameter_acl, which is much more compact than the \"parameteracl\" name would be, so that remains under the old name.\n\nThe tab completion for grants and revokes of \"set\" and \"alter system\" on parameters has been extended compared to the previously submitted patch. The tab completion for grants and revokes of other privilege types remains fairly thin, as I resisted the urge to fix that generally, given that it is unrelated to this patch. However, in making tab completion for grant/revoke on parameters work, this patch incidentally improves the coverage for other privilege types, in so far as the prefixes for those commands match the prefixes of these new commands. \n\nVersion 12 also introduces \\dcp (pneumonic, \"Describe Configuration Parameter\") for listing parameters, with \\dcp+ also showing the acl, like:\n\nmark.dilger=# \\dcp+ %client%\n List of configuration parameters\n Parameter | Context | Setting | Unit | Access privileges \n----------------------------------+---------+---------+------+--------------------------------\n client_connection_check_interval | user | 0 | ms | \n client_encoding | user | UTF8 | | \"mark.dilger\"=sA/\"mark.dilger\"\n client_min_messages | user | notice | | \"mark.dilger\"=sA/\"mark.dilger\"+\n | | | | =s/\"mark.dilger\" +\n | | | | alice=A*/\"mark.dilger\"\n(3 rows)\n\nThe \"Access privileges\" column is blank for parameters which have neither been granted nor revoked. In the example above, that's true for client_connection_check_interval.\n\n\\dcp[+] only shows \"user\" and \"superuser\" parameters:\n\nmark.dilger=# \\dcp %wal%\n List of configuration parameters\n Parameter | Context \n--------------------------+-----------\n track_wal_io_timing | superuser\n wal_compression | superuser\n wal_consistency_checking | superuser\n wal_init_zero | superuser\n wal_recycle | superuser\n wal_sender_timeout | user\n wal_skip_threshold | user\n(7 rows)\n\nwhereas \\dcpS[+] also shows all parameters (so, also \"sighup\", \"backend\", \"superuser-backend\", \"postmaster\", and \"internal\".)\n\nmark.dilger=# \\dcpS %wal%\n List of configuration parameters\n Parameter | Context \n-------------------------------+------------\n max_slot_wal_keep_size | sighup\n max_wal_senders | postmaster\n max_wal_size | sighup\n min_wal_size | sighup\n track_wal_io_timing | superuser\n wal_block_size | internal\n wal_buffers | postmaster\n wal_compression | superuser\n wal_consistency_checking | superuser\n wal_init_zero | superuser\n wal_keep_size | sighup\n wal_level | postmaster\n wal_log_hints | postmaster\n wal_receiver_create_temp_slot | sighup\n wal_receiver_status_interval | sighup\n wal_receiver_timeout | sighup\n wal_recycle | superuser\n wal_retrieve_retry_interval | sighup\n wal_segment_size | internal\n wal_sender_timeout | user\n wal_skip_threshold | user\n wal_sync_method | sighup\n wal_writer_delay | sighup\n wal_writer_flush_after | sighup\n(24 rows)\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Thu, 24 Mar 2022 09:49:24 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 3/24/22 12:49, Mark Dilger wrote:\n>\n>> On Mar 17, 2022, at 8:41 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>\n>> If we abandoned that for this form of GRANT/REVOKE I think we could\n>> probably get away with\n>>\n>>\n>> GRANT { SET | ALTER SYSTEM } ON setting_name ...\n>>\n>>\n>> I haven't tried it, so I could be all wrong.\n> Version 12 of the patch uses SET and ALTER SYSTEM as the names of the privileges, and PARAMETER as the name of the thing on which the privilege is granted. The catalog table which tracks these grants is now named pg_parameter_acl, and various other parts of the patch have been adjusted to use a \"parameter\" based, rather than a \"setting\" based, naming scheme. One exception to this rule is the \"setacl\" column in pg_parameter_acl, which is much more compact than the \"parameteracl\" name would be, so that remains under the old name.\n\n\nI can live with it I guess, but it seems perverse to me to have\npg_settings but pg_paramater_acl effectively referring to the same set\nof things. If we're going to do this perhaps we should create a\npg_parameters view which is identical to pg_settings and deprecate\npg_settings. I don;t want to hold up this patch, I think this can\nprobably be managed as a follow up item.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Thu, 24 Mar 2022 15:06:21 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Mar 24, 2022, at 12:06 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\n> \n> \n> On 3/24/22 12:49, Mark Dilger wrote:\n>> \n>>> On Mar 17, 2022, at 8:41 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>> \n>>> If we abandoned that for this form of GRANT/REVOKE I think we could\n>>> probably get away with\n>>> \n>>> \n>>> GRANT { SET | ALTER SYSTEM } ON setting_name ...\n>>> \n>>> \n>>> I haven't tried it, so I could be all wrong.\n>> Version 12 of the patch uses SET and ALTER SYSTEM as the names of the privileges, and PARAMETER as the name of the thing on which the privilege is granted. The catalog table which tracks these grants is now named pg_parameter_acl, and various other parts of the patch have been adjusted to use a \"parameter\" based, rather than a \"setting\" based, naming scheme. One exception to this rule is the \"setacl\" column in pg_parameter_acl, which is much more compact than the \"parameteracl\" name would be, so that remains under the old name.\n> \n> \n> I can live with it I guess, but it seems perverse to me to have\n> pg_settings but pg_paramater_acl effectively referring to the same set\n> of things. If we're going to do this perhaps we should create a\n> pg_parameters view which is identical to pg_settings and deprecate\n> pg_settings. I don;t want to hold up this patch, I think this can\n> probably be managed as a follow up item.\n\nRight, the version 12 patch was following Peter's and Tom's comments upthread:\n\n> On Mar 17, 2022, at 7:47 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Well, I still say that \"SET\" is sufficient for the one privilege name\n> (unless we really can't make Bison handle that, which I doubt). But\n> I'm willing to yield on using \"ALTER SYSTEM\" for the other.\n> \n> If we go with s/SETTING/PARAMETER/ as per your other message, then\n> that would be adequately consistent with the docs I think. So it'd\n> be\n> \n> GRANT { SET | ALTER SYSTEM } ON PARAMETER foo TO ...\n> \n> and the new catalog would be pg_parameter_acl, and so on.\n\nWe could debate that again, but it seems awfully late in the development cycle. I'd rather just get this committed, barring any objections?\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Thu, 24 Mar 2022 12:50:42 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"I've started reviewing this patch in earnest, and almost immediately\nhit a serious problem: these regression tests aren't even a little bit\ncommittable. For one thing, they fail if you do \"make installcheck\"\ntwice in a row. This seems to be because the first run leaves some\ncruft behind in pg_parameter_acl that affects the results of\nsubsequent runs. But the bigger problem is that it is ABSOLUTELY NOT\nOKAY to test ALTER SYSTEM during an \"installcheck\" run. That would be\ntrue even if you cleaned up all the settings by the end of the run, as\nthe patch fails to do (and for extra demerit, it leaves a superuser\nrole laying around). Even transient effects on the behavior of\nsessions in other DBs aren't acceptable in installcheck mode, IMO.\n\nI think we probably have to trash the core-regression-tests part of\nthe patch altogether and instead use a TAP test for whatever testing\nwe want to do. It might be all right to test SET privileges without\ntesting ALTER SYSTEM, but I'm not sure there's much point in that.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 28 Mar 2022 14:31:47 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Mar 28, 2022, at 11:31 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> I think we probably have to trash the core-regression-tests part of\n> the patch altogether and instead use a TAP test for whatever testing\n> we want to do. It might be all right to test SET privileges without\n> testing ALTER SYSTEM, but I'm not sure there's much point in that.\n\nHow about putting them under src/test/modules/unsafe_tests ?\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Mon, 28 Mar 2022 11:42:42 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> Version 12 also introduces \\dcp (pneumonic, \"Describe Configuration Parameter\") for listing parameters, with \\dcp+ also showing the acl, like:\n\nThe fact that that code is not dry behind the ears is painfully obvious.\nIt's not documented in psql-ref, not tested anywhere AFAICS, and its\nhandling of the pattern parameter is inconsistent with every other\n\\d command. The wildcard character should be * not %. It only\naccidentally fails to dump core if no pattern is given, too.\n\n> \\dcp[+] only shows \"user\" and \"superuser\" parameters:\n\nWhy make that restriction? Also, I find it astonishing that this doesn't\nshow the GUC's value by default. The non-plus form of the command seems\nuseless as it stands, or at least its intended use-case is so narrow\nI can't see it. If we're to have it at all, it seems like it ought to\nbe a reasonably useful shortcut for interrogating pg_settings. I'd\nexpect the base set of columns to be name, value, and possibly unit;\nthen add ACL with +. I'm not sure that GucContext belongs in this at all,\nbut if it does, it's a + column.\n\nOn the whole perhaps this should be taken out again; it's a bit\nlate in the cycle to be introducing new features, especially ones\nas subject to bikeshedding as a \\d command is. My ideas about what\ncolumns to show probably wouldn't match anyone else's ... and we\nhaven't even gotten to whether \\dcp is an okay choice of name.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 28 Mar 2022 15:11:15 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> On Mar 28, 2022, at 11:31 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> I think we probably have to trash the core-regression-tests part of\n>> the patch altogether and instead use a TAP test for whatever testing\n>> we want to do. It might be all right to test SET privileges without\n>> testing ALTER SYSTEM, but I'm not sure there's much point in that.\n\n> How about putting them under src/test/modules/unsafe_tests ?\n\nAh, that could work; I'd forgotten about that subdirectory.\n\nIt'd still be a good idea if they didn't fail when run twice in a row.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 28 Mar 2022 15:14:04 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Mar 28, 2022, at 12:11 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\n>> Version 12 also introduces \\dcp (pneumonic, \"Describe Configuration Parameter\") for listing parameters, with \\dcp+ also showing the acl, like:\n> \n> The fact that that code is not dry behind the ears is painfully obvious.\n\n<snip>\n\n> On the whole perhaps this should be taken out again\n\nI added the \\dcp command after noticing that doc/src/sgml/ddl.sgml's \"Summary of Access Privileges\" table had an empty column for the \"psql Command\" column for the \"PARAMETER\" row. Maybe that's ok. I thought it would be better if I added a command for that.\n\nI don't feel too strongly about it, so I'm removing \\dcp from this patch with the intention of introducing it as a separate patch for the postgres 16 development cycle. For now, I'm just putting \"none\" in that table cell.\n\n\nThis patch also moves the grant/revoke tests in guc_privs.sql into the unsafe_tests directory, fixes them up to be repeatable, and drops the superuser role that had been negligently left around in prior patches. The tests that had been added in privileges.sql are simply removed.\n\n\n\n \n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Mon, 28 Mar 2022 13:57:50 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"I just came across something odd in v12 that is still there in v13:\nExecGrant_Parameter uses InvokeObjectPostAlterHook not\nInvokeObjectPostAlterHookArgStr. This seems pretty inconsistent.\nIs there a good argument for it?\n\n... or, for that matter, why is there any such call at all?\nNo other GRANT/REVOKE operation calls such a hook.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 28 Mar 2022 17:16:14 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Mar 28, 2022, at 2:16 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> I just came across something odd in v12 that is still there in v13:\n> ExecGrant_Parameter uses InvokeObjectPostAlterHook not\n> InvokeObjectPostAlterHookArgStr. This seems pretty inconsistent.\n> Is there a good argument for it?\n> \n\nFor SET and ALTER SYSTEM, the target of the action may not have an entry in pg_parameter_acl, nor an assigned Oid anywhere, so the only consistent way to pass the argument to the hook is by name. For GRANT/REVOKE, the parameter must have an Oid, at least by the time the hook gets called. Upthread there was some discussion of a hook not being able to assume a snapshot and working transaction, and hence not being able to query the catalogs. I would think that in a GRANT or REVOKE that hasn't already errored, the hook would have a transaction and could look up whatever it likes? There is a CommandCounterIncrement() call issued in objectNamesToOids() for new parameters, so by the time the hook is running it should be able to see the parameter.\n\nAm I reasoning about this the wrong way?\n\n> ... or, for that matter, why is there any such call at all?\n> No other GRANT/REVOKE operation calls such a hook.\n\nI think ALTER DEFAULT PRIVILEGES does, though that's not quite the same thing. I don't have a strong opinion on this. Joshua, what's your take?\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Mon, 28 Mar 2022 14:42:46 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n>> On Mar 28, 2022, at 2:16 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> I just came across something odd in v12 that is still there in v13:\n>> ExecGrant_Parameter uses InvokeObjectPostAlterHook not\n>> InvokeObjectPostAlterHookArgStr. This seems pretty inconsistent.\n>> Is there a good argument for it?\n\n> For SET and ALTER SYSTEM, the target of the action may not have an entry\n> in pg_parameter_acl, nor an assigned Oid anywhere, so the only\n> consistent way to pass the argument to the hook is by name. For\n> GRANT/REVOKE, the parameter must have an Oid, at least by the time the\n> hook gets called.\n\nYeah, I know it's *possible* to make this work. The question is why is\nit good to do it like this rather than to use the string API, now that\nwe have the latter. AFAICS this way just guarantees that the hook must\ndo a catalog lookup in order to figure out what you're talking about.\n\nThe core point here is that the actual identity of a GUC is its name.\nAny OID that may exist in pg_parameter_acl is just a nonce alias that\nmeans nothing to anybody. Anyone who's trying to, say, enforce that\nJoe Blow can't change shared_buffers is going to need to see the GUC\nname. (I am, btw, busy doing a lot of renaming in the patch to try\nto clarify that these OIDs are not identifiers for GUCs; imagining\nthat they are just risks confusion.)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 28 Mar 2022 17:54:28 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Mar 28, 2022, at 2:54 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Yeah, I know it's *possible* to make this work. The question is why is\n> it good to do it like this rather than to use the string API, now that\n> we have the latter. AFAICS this way just guarantees that the hook must\n> do a catalog lookup in order to figure out what you're talking about.\n\nOk, thanks for clarifying. I took the *HookStr versions of the hooks to be an alternative to be used when no Oid was present, something of a last resort. I never thought much about using them under other circumstances.\n\n> The core point here is that the actual identity of a GUC is its name.\n> Any OID that may exist in pg_parameter_acl is just a nonce alias that\n> means nothing to anybody. Anyone who's trying to, say, enforce that\n> Joe Blow can't change shared_buffers is going to need to see the GUC\n> name. (I am, btw, busy doing a lot of renaming in the patch to try\n> to clarify that these OIDs are not identifiers for GUCs; imagining\n> that they are just risks confusion.)\n\nI was about to write another patch using the HookStr form, but if you are already editing, then I'll let you make the change. I don't see a problem with what you are proposing.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Mon, 28 Mar 2022 15:04:12 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"I'm going to be mostly unavailable till Wednesday, but I'll leave\nyou with another thing to chew on:\n\nregression=# create user joe;\nCREATE ROLE\nregression=# grant set on parameter plpgsql.extra_warnings to joe;\nERROR: unrecognized configuration parameter \"plpgsql.extra_warnings\"\n\nThis is problematic, because once plpgsql is loaded it works:\n\nregression=# load 'plpgsql'; \nLOAD\nregression=# grant set on parameter plpgsql.extra_warnings to joe;\nGRANT\n\nIf we then do\n\n$ pg_dumpall -g\n\nit falls over:\npg_dumpall: error: query failed: ERROR: unrecognized configuration parameter \"plpgsql.extra_warnings\"\n\napparently because aclparameterdefault() is allergic to being asked about\nunknown GUCs, and plpgsql is not loaded in pg_dumpall's session. But\nif pg_dumpall hadn't failed, it'd produce a dump script containing that\nsame command, which would fail at load time (because, again, plpgsql\nisn't going to be loaded in the backend reading the restore script).\n\nThis is what I meant by saying that you can't just refuse to GRANT on\nunknown GUCs. It makes custom GUCs into a time bomb for dump/restore.\nAnd that means you need a strategy for dealing with the possibility\nthat you don't know whether the GUC is USERSET or not. I think though\nthat it might work to just assume that it isn't, in which case dumps\non unrecognized GUCs that really are USERSET will end up issuing an\nexplicit GRANT SET TO PUBLIC that we didn't actually need to do, but it\nwon't hurt anything. (Testing that assertion would be a good thing\nto do.)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 28 Mar 2022 18:31:02 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> I was about to write another patch using the HookStr form, but if you are already editing, then I'll let you make the change. I don't see a problem with what you are proposing.\n\nDon't sweat about that, I can easily rebase what I've done so far\nover your updates.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 28 Mar 2022 18:37:07 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Mar 28, 2022, at 3:31 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> This is what I meant by saying that you can't just refuse to GRANT on\n> unknown GUCs. It makes custom GUCs into a time bomb for dump/restore.\n> And that means you need a strategy for dealing with the possibility\n> that you don't know whether the GUC is USERSET or not. I think though\n> that it might work to just assume that it isn't, in which case dumps\n> on unrecognized GUCs that really are USERSET will end up issuing an\n> explicit GRANT SET TO PUBLIC that we didn't actually need to do, but it\n> won't hurt anything. (Testing that assertion would be a good thing\n> to do.)\n\nOk, I returned to the idea upthread for a solution to this problem. A grant or revoke on an unrecognized custom parameter will create a SUSET placeholder, which is not quite right in some cases. However, the installation scripts for modules have been updated to manually grant SET privilege on their custom USERSET parameters, which cleans up the problem, with one exception: if the user executes a \"revoke set on parameter some.such from public\" prior to loading the module which defines parameter some.such, that revoke won't be retained. That doesn't seem entirely wrong to me, since no privilege to set the parameter existed when the revoke was performed, but rather was granted along with the creation of the parameter, but it also doesn't seem entirely right. Maybe revoke commands (but not grant commands) should error on unrecognized custom parameters? I didn't implement that here, but can do so if you think that makes more sense than this new behavior.\n \nI changed add_placeholder_variable() to take a GucContext argument. It previously always used PGC_USERSET, which is what all pre-existing call sites now pass into it, but that seems a bit inappropriate where we're creating a placeholder that we intend to treat as a SUSET variable until such time as a module gets installed saying otherwise. Not changing add_placeholder_variable in this fashion seems to work just fine. I just didn't feel comfortable with doing it that way. But if you feel it generates needless code churn, I could be talked out of doing this.\n\nI also changed the patch to use the ...HookStr functions for parameters. I would really like a comment on this from Joshua, to be sure what I'm doing comports with what he wanted. In particular, I'm uncertain that simply passing the AclMode (in other words, the istmt->privileges field) for the grant/revoke to the hook is sufficient. For one, how does the hook want to distinguish grants from revokes? Do we want a bit for that? And what about distinguishing WITH GRANT OPTION? I think the hooks are usable right now, but they might be made better.\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Tue, 29 Mar 2022 21:00:36 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Tue, Mar 29, 2022 at 9:00 PM Mark Dilger <mark.dilger@enterprisedb.com>\nwrote:\n\n> A grant or revoke on an unrecognized custom parameter will create a SUSET\n> placeholder,\n\n[...]\n\n> which cleans up the problem, with one exception: if the user executes a\n> \"revoke set on parameter some.such from public\" prior to loading the module\n> which defines parameter some.such, that revoke won't be retained. That\n> doesn't seem entirely wrong to me, since no privilege to set the parameter\n> existed when the revoke was performed, but rather was granted along with\n> the creation of the parameter, but it also doesn't seem entirely right.\n> Maybe revoke commands (but not grant commands) should error on unrecognized\n> custom parameters?\n\n\nThe only revoke role target that makes sense here is the default grant\ngiven to public. Aside from that technicality the grant system is purely\nadditive and so only the GRANTS end up retained so far as perpetual state\nis concerned.\n\nFor a revoke that doesn't target public we should remove the corresponding\nunrecognized setting grant if one is present. We don't generally/always\nraise a notice if a revoke doesn't actually cause something to be ungranted\n- though I'm partial to being so informed. I suppose we could distinguish\nthe cases where the not-yet-loaded setting name is unrecognized by the\nsystem from the one where it is recognized but the grant is actually on a\ndifferent role (or a revoke entry from public for the setting name is\npresent).\n\nFor a revoke of an unknown setting from public we should keep an entry\nsomewhere that tells the system that the default grant to public for that\nsetting has been revoked. Maybe there isn't the same timing concern here\nas there is for GRANT, but if only for symmetry it seems like a good thing\nto implement.\n\nI have the impression I'm missing something in what I wrote above but\ncannot quite figure out what. In any case as a first pass at this the\nbehavior described is kinda what I'm expecting.\n\nDavid J.\n\nP.S.\n\nSkimming the patch we are, to my agreement, not touching the ALTER DEFAULT\nPRIVILEGES command to work with this feature. Should the omission be noted\nexplicitly? At least in the commit message I would think. Though the\nsentence in [1] \"Also, these default privilege settings can be overridden\nusing the ALTER DEFAULT PRIVILEGES command.\" is rendered only partially\ncorrect.\n\n[1] https://www.postgresql.org/docs/current/ddl-priv.html\n\nP.P.S.\n\n+ The default privileges for a <literal>user</literal> parameter allow\n+ <literal>PUBLIC</literal> to <command>SET</command> and\n+ <command>RESET</command> the assigned value. By default,\n+ <literal>PUBLIC</literal> has no privileges on\n+ <literal>postmaster</literal>, <literal>superuser-backend</literal>,\n+ <literal>internal</literal>, <literal>backend</literal>,\n+ <literal>sighup</literal>, and <literal>superuser</literal> parameters.\n\nCan we rephrase this to something like:\n\nBy default, PUBLIC has no privileges on parameters in the postmaster, ...,\nand superuser contexts.\n\npg_settings.context exists and those are the values found there. My\ninitial interpretation of the wording was the postmaster, etc..., were\nthemselves parameters, not containers for many parameters.\n\nOn Tue, Mar 29, 2022 at 9:00 PM Mark Dilger <mark.dilger@enterprisedb.com> wrote: A grant or revoke on an unrecognized custom parameter will create a SUSET placeholder,[...] which cleans up the problem, with one exception: if the user executes a \"revoke set on parameter some.such from public\" prior to loading the module which defines parameter some.such, that revoke won't be retained. That doesn't seem entirely wrong to me, since no privilege to set the parameter existed when the revoke was performed, but rather was granted along with the creation of the parameter, but it also doesn't seem entirely right. Maybe revoke commands (but not grant commands) should error on unrecognized custom parameters?The only revoke role target that makes sense here is the default grant given to public. Aside from that technicality the grant system is purely additive and so only the GRANTS end up retained so far as perpetual state is concerned.For a revoke that doesn't target public we should remove the corresponding unrecognized setting grant if one is present. We don't generally/always raise a notice if a revoke doesn't actually cause something to be ungranted - though I'm partial to being so informed. I suppose we could distinguish the cases where the not-yet-loaded setting name is unrecognized by the system from the one where it is recognized but the grant is actually on a different role (or a revoke entry from public for the setting name is present).For a revoke of an unknown setting from public we should keep an entry somewhere that tells the system that the default grant to public for that setting has been revoked. Maybe there isn't the same timing concern here as there is for GRANT, but if only for symmetry it seems like a good thing to implement.I have the impression I'm missing something in what I wrote above but cannot quite figure out what. In any case as a first pass at this the behavior described is kinda what I'm expecting.David J.P.S.Skimming the patch we are, to my agreement, not touching the ALTER DEFAULT PRIVILEGES command to work with this feature. Should the omission be noted explicitly? At least in the commit message I would think. Though the sentence in [1] \"Also, these default privilege settings can be overridden using the ALTER DEFAULT PRIVILEGES command.\" is rendered only partially correct.[1] https://www.postgresql.org/docs/current/ddl-priv.htmlP.P.S.+ The default privileges for a <literal>user</literal> parameter allow+ <literal>PUBLIC</literal> to <command>SET</command> and+ <command>RESET</command> the assigned value. By default,+ <literal>PUBLIC</literal> has no privileges on+ <literal>postmaster</literal>, <literal>superuser-backend</literal>,+ <literal>internal</literal>, <literal>backend</literal>,+ <literal>sighup</literal>, and <literal>superuser</literal> parameters.Can we rephrase this to something like:By default, PUBLIC has no privileges on parameters in the postmaster, ..., and superuser contexts.pg_settings.context exists and those are the values found there. My initial interpretation of the wording was the postmaster, etc..., were themselves parameters, not containers for many parameters.","msg_date":"Tue, 29 Mar 2022 22:27:44 -0700","msg_from":"\"David G. Johnston\" <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Wed, Mar 30, 2022 at 12:00 AM Mark Dilger\n<mark.dilger@enterprisedb.com> wrote:\n>\n> > On Mar 28, 2022, at 3:31 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >\n> > This is what I meant by saying that you can't just refuse to GRANT on\n> > unknown GUCs. It makes custom GUCs into a time bomb for dump/restore.\n> > And that means you need a strategy for dealing with the possibility\n> > that you don't know whether the GUC is USERSET or not. I think though\n> > that it might work to just assume that it isn't, in which case dumps\n> > on unrecognized GUCs that really are USERSET will end up issuing an\n> > explicit GRANT SET TO PUBLIC that we didn't actually need to do, but it\n> > won't hurt anything. (Testing that assertion would be a good thing\n> > to do.)\n>\n> Ok, I returned to the idea upthread for a solution to this problem. A grant or revoke on an unrecognized custom parameter will create a SUSET placeholder, which is not quite right in some cases. However, the installation scripts for modules have been updated to manually grant SET privilege on their custom USERSET parameters, which cleans up the problem, with one exception: if the user executes a \"revoke set on parameter some.such from public\" prior to loading the module which defines parameter some.such, that revoke won't be retained. That doesn't seem entirely wrong to me, since no privilege to set the parameter existed when the revoke was performed, but rather was granted along with the creation of the parameter, but it also doesn't seem entirely right. Maybe revoke commands (but not grant commands) should error on unrecognized custom parameters? I didn't implement that here, but can do so if you think that makes more sense than this new behavior.\n>\n> I changed add_placeholder_variable() to take a GucContext argument. It previously always used PGC_USERSET, which is what all pre-existing call sites now pass into it, but that seems a bit inappropriate where we're creating a placeholder that we intend to treat as a SUSET variable until such time as a module gets installed saying otherwise. Not changing add_placeholder_variable in this fashion seems to work just fine. I just didn't feel comfortable with doing it that way. But if you feel it generates needless code churn, I could be talked out of doing this.\n>\n> I also changed the patch to use the ...HookStr functions for parameters. I would really like a comment on this from Joshua, to be sure what I'm doing comports with what he wanted. In particular, I'm uncertain that simply passing the AclMode (in other words, the istmt->privileges field) for the grant/revoke to the hook is sufficient. For one, how does the hook want to distinguish grants from revokes? Do we want a bit for that? And what about distinguishing WITH GRANT OPTION? I think the hooks are usable right now, but they might be made better.\n\nI had not even been thinking about hooking grant/revoke TBH. In\nSELinux-type MAC systems we don't always try to control DAC permission\nchanges, and when we do it's not granular, all permission changes just\nboil down to setattr permission or something of that nature.\n\nThank you.\n\n\n","msg_date":"Wed, 30 Mar 2022 08:30:42 -0400","msg_from":"Joshua Brindle <joshua.brindle@crunchydata.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"After sleeping on it, I have a modest proposal for simplifying\nthese issues. Consider this design:\n\n1. In the SET code path, we assume (without any catalog lookup)\nthat USERSET GUCs can be set. Only for SUSET GUCs do we perform\na permissions lookup. (ALTER SYSTEM does a lookup in both cases.)\n\n2. Given this, the default ACL for any GUC can be empty, greatly\nsimplifying all these management issues. Superusers could do what\nthey want anyway, so modeling an \"owner's default grant\" becomes\nunnecessary.\n\nWhat this loses is the ability to revoke public SET permissions\non USERSET GUCs. I claim that that is not so valuable as to\njustify all the complication needed to deal with it. (If a GUC\nseems to require some defenses, why is it USERSET?) Avoiding\na permissions lookup in the default SET code path seems like\na pretty important benefit, too. If we force that to happen\nit's going to be a noticeable drag on functions with SET clauses.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 30 Mar 2022 09:26:21 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Mar 30, 2022, at 6:26 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Consider this design:\n\nIsn't that just the design I had implemented in v8 several months ago?\n\nSubject: [PATCH v8] Allow GRANT of SET and ALTER SYSTEM SET for gucs\n\nAllow granting of privilege to set or alter system set variables\nwhich otherwise can only be managed by superusers. Each\n(role,variable,privilege) triple is independently grantable, so a\nuser may be granted privilege to SET but not to ALTER SYSTEM SET on\na variable, or vice versa. The privilege to SET a userset variable\nmay be granted, though doing so has no practical effect, since any\nrole can set userset variables anyway. Worse, there is no way to\nrevoke the privilege to SET a userset variable. To remedy that,\nmost core userset variables have been changed to suset, with\nexplicit grants to set the variable to public.\n\n\nI don't think v9 ever got posted to the list, but v10 has:\n\nSubject: [PATCH v10] Allow grant and revoke of privileges on settings\n\nAllow grant and revoke of privileges to set or alter system set\nconfiguration variables. Each (role,variable,privilege) triple can\nbe independently granted or revoked, so a user may be granted\nprivilege to SET but not to ALTER SYSTEM SET on a variable, or vice\nversa. Privilege to SET a userset variable is implicitly granted to\npublic, but may be revoked.\n\n\nIf we want to backtrack to v8, that's fine. I can rebase that, port some of the other changes from v14 to it, and repost it as v15. We should review the conversation from December and January which included some arguments for allowing revokes of SET on USERSET from PUBLIC. I don't want to keep going around in circles on this.\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Wed, 30 Mar 2022 06:59:48 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Mar 30, 2022, at 6:59 AM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> \n> We should review the conversation from December and January which included some arguments for allowing revokes of SET on USERSET from PUBLIC. I don't want to keep going around in circles on this.\n\nHmm, I guess that conversation was mostly off-list at the PGConn in December. I made a reference to it upthread:\n\n> On Mar 6, 2022, at 2:40 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> \n> Userset variables are implicitly settable by any user. There was a request, off-list as I recall, to make it possible to revoke the privilege to set variables such as \"work_mem\". To make that possible, but not change the default behavior vis-a-vis prior releases, I upgraded most userset variables to suset with a corresponding grant to public on the variable. Sites which wish to have a more restrictive policy on such variables can revoke that privilege from public and instead issue more restrictive grants. There were a few variables where such treatment didn't seem sensible, such as ones to do with client connections, and I left them alone. I didn't insist on a defense for why any particular setting needed to be revocable in order to apply this treatment. My assumption was that sites should be allowed to determine their own security policies per setting unless there is a technical difficulty for a given setting that would make it overly burdensome to implement.\n\nYour proposal to just punt on supporting revocation of set on userset from public seems fine. We could revisit that in the next development cycle if anyone really wants to defend it. In particular, I don't see that committing this feature without that part would create any additional backward compatibility problems when implementing that later.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Wed, 30 Mar 2022 07:45:36 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 3/30/22 09:26, Tom Lane wrote:\n> After sleeping on it, I have a modest proposal for simplifying\n> these issues. Consider this design:\n>\n> 1. In the SET code path, we assume (without any catalog lookup)\n> that USERSET GUCs can be set. Only for SUSET GUCs do we perform\n> a permissions lookup. (ALTER SYSTEM does a lookup in both cases.)\n>\n> 2. Given this, the default ACL for any GUC can be empty, greatly\n> simplifying all these management issues. Superusers could do what\n> they want anyway, so modeling an \"owner's default grant\" becomes\n> unnecessary.\n>\n> What this loses is the ability to revoke public SET permissions\n> on USERSET GUCs. I claim that that is not so valuable as to\n> justify all the complication needed to deal with it. (If a GUC\n> seems to require some defenses, why is it USERSET?) Avoiding\n> a permissions lookup in the default SET code path seems like\n> a pretty important benefit, too. If we force that to happen\n> it's going to be a noticeable drag on functions with SET clauses.\n>\n> \t\t\t\n\n\n\nThe last point is telling, so +1\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 30 Mar 2022 11:12:46 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Wed, Mar 30, 2022 at 8:12 AM Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 3/30/22 09:26, Tom Lane wrote:\n>\n> >\n> > What this loses is the ability to revoke public SET permissions\n> > on USERSET GUCs. I claim that that is not so valuable as to\n> > justify all the complication needed to deal with it.\n\n\nAgreed, and in line with my thinking from last night. These default public\nset grants are indeed the complication and I'm good with the status quo\nwhere they are non-revocable.\n\nI'm finding it curious that we are choosing to document every (all 6)\ncontext that doesn't have this default privilege instead of saying that\nonly the user context variables are granted this default, and now\nirrevocable, default set privilege. This is in addition to making sure we\ndistinguish between parameter and context in my earlier email.\n\n\n> > Avoiding\n> > a permissions lookup in the default SET code path seems like\n> > a pretty important benefit, too. If we force that to happen\n> > it's going to be a noticeable drag on functions with SET clauses.\n> >\n> >\n> The last point is telling, so +1\n>\n>\nIndeed. +1\n\nDavid J.\n\nOn Wed, Mar 30, 2022 at 8:12 AM Andrew Dunstan <andrew@dunslane.net> wrote:\nOn 3/30/22 09:26, Tom Lane wrote:\n>\n> What this loses is the ability to revoke public SET permissions\n> on USERSET GUCs. I claim that that is not so valuable as to\n> justify all the complication needed to deal with it.Agreed, and in line with my thinking from last night. These default public set grants are indeed the complication and I'm good with the status quo where they are non-revocable.I'm finding it curious that we are choosing to document every (all 6) context that doesn't have this default privilege instead of saying that only the user context variables are granted this default, and now irrevocable, default set privilege. This is in addition to making sure we distinguish between parameter and context in my earlier email. > Avoiding\n> a permissions lookup in the default SET code path seems like\n> a pretty important benefit, too. If we force that to happen\n> it's going to be a noticeable drag on functions with SET clauses.\n>\n> \nThe last point is telling, so +1 Indeed. +1 David J.","msg_date":"Wed, 30 Mar 2022 08:44:53 -0700","msg_from":"\"David G. Johnston\" <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> Your proposal to just punt on supporting revocation of set on userset from public seems fine. We could revisit that in the next development cycle if anyone really wants to defend it. In particular, I don't see that committing this feature without that part would create any additional backward compatibility problems when implementing that later.\n\nYeah. Also, as you noted, we could mark some individual built-in\nvariables as SUSET and add a default GRANT. I don't want to do that with\na blunderbuss, but perhaps there's an argument to do it for specific\ncases (search_path comes to mind, though the performance cost could be\nsignificant, since I think setting that in function SET clauses is\ncommon).\n\nFor now, though, saying that you can't restrict SET for USERSET variables\nseems fine --- there's certainly no loss of capability compared to where\nwe stand today. I'd prefer to get the feature committed in that form\nand then look at whether we want to tighten things around the margins.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 30 Mar 2022 11:45:53 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"On Wed, Mar 30, 2022 at 8:46 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> I don't want to do that with\n> a blunderbuss, but perhaps there's an argument to do it for specific\n> cases (search_path comes to mind, though the performance cost could be\n> significant, since I think setting that in function SET clauses is\n> common).\n>\n\n\nI suspect it became considerably moreso when we fixed the search_path CVE\nsince we basically told people that doing so, despite the possible\nperformance hit, was the easiest solution to their immediate dump/restore\nfailures. But ISTM that because that SET has a function invocation context\nit could bypass any such check. Though maybe the DO command exposes a flaw\nin that idea.\nDavid J.\n\nOn Wed, Mar 30, 2022 at 8:46 AM Tom Lane <tgl@sss.pgh.pa.us> wrote: I don't want to do that with\na blunderbuss, but perhaps there's an argument to do it for specific\ncases (search_path comes to mind, though the performance cost could be\nsignificant, since I think setting that in function SET clauses is\ncommon). I suspect it became considerably moreso when we fixed the search_path CVE since we basically told people that doing so, despite the possible performance hit, was the easiest solution to their immediate dump/restore failures. But ISTM that because that SET has a function invocation context it could bypass any such check. Though maybe the DO command exposes a flaw in that idea.David J.","msg_date":"Wed, 30 Mar 2022 08:51:05 -0700","msg_from":"\"David G. Johnston\" <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> If we want to backtrack to v8, that's fine. I can rebase that, port\n> some of the other changes from v14 to it, and repost it as v15.\n\nAre you working on that? I've set aside time this week to hopefully\nget this over the finish line, but I don't want to find out that\nI've been duplicating your effort.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 04 Apr 2022 11:36:18 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Apr 4, 2022, at 8:36 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\n>> If we want to backtrack to v8, that's fine. I can rebase that, port\n>> some of the other changes from v14 to it, and repost it as v15.\n> \n> Are you working on that? I've set aside time this week to hopefully\n> get this over the finish line, but I don't want to find out that\n> I've been duplicating your effort.\n\nYes, I expect to be posting the latest in maybe an hour? I believe the latest patch (just reviewing, adjusting code comments, etc.) that I'm preparing to post has all the changes we've discussed, aside from your parameterId renaming.\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Mon, 4 Apr 2022 08:47:46 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"> On Apr 4, 2022, at 8:47 AM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> \n> posting the latest in maybe an hour\n\nHere it is:\n\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Mon, 4 Apr 2022 09:31:41 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> [ v15 patch ]\n\nThanks. As I'm working through this, I'm kind of inclined to drop\nthe has_parameter_privilege() variants that take an OID as object\nidentifier. This gets back to the fact that I don't think\npg_parameter_acl OIDs have any outside use; we wouldn't even have\nthem except that we need a way to track their role dependencies\nin pg_shdepend. AFAICS users will only ever be interested in\nlooking up a GUC by name. Any objections?\n\nAnother thought here is that I see you're expending some code\nto store the canonical name of a GUC in pg_parameter_acl, but\nI think that's probably going too far. In particular, for the\nlegacy mixed-case names like \"DateStyle\", what ends up in the\ntable is the mixed-case name, which seems problematic. It would\ndefinitely be problematic if an extension used such a name,\nbecause we might or might not be aware of the idiosyncratic\ncasing at the time a GRANT is issued. I'm thinking that we\nreally want to avoid looking up custom GUCs at all during GRANT,\nbecause that can't do anything except create hazards.\n\nSo I think that instead of what you've got here, we should\n(1) apply the map_old_guc_names[] mapping, which is constant\n (for any one PG release anyway)\n(2) smash to lower case\n(3) verify validity per valid_variable_name.\n\nThis also simplifies life on the lookup side, where it's sufficient\nto do steps (1) and (2) before performing a catalog search.\n\nThoughts?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 04 Apr 2022 17:26:35 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\nOn 4/4/22 17:26, Tom Lane wrote:\n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\n>> [ v15 patch ]\n> Thanks. As I'm working through this, I'm kind of inclined to drop\n> the has_parameter_privilege() variants that take an OID as object\n> identifier. This gets back to the fact that I don't think\n> pg_parameter_acl OIDs have any outside use; we wouldn't even have\n> them except that we need a way to track their role dependencies\n> in pg_shdepend. AFAICS users will only ever be interested in\n> looking up a GUC by name. Any objections?\n\n\nNot from me\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Mon, 4 Apr 2022 18:15:19 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Apr 4, 2022, at 2:26 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Thanks. As I'm working through this, I'm kind of inclined to drop\n> the has_parameter_privilege() variants that take an OID as object\n> identifier. This gets back to the fact that I don't think\n> pg_parameter_acl OIDs have any outside use; we wouldn't even have\n> them except that we need a way to track their role dependencies\n> in pg_shdepend. AFAICS users will only ever be interested in\n> looking up a GUC by name. Any objections?\n\nNone.\n\n> Another thought here is that I see you're expending some code\n> to store the canonical name of a GUC in pg_parameter_acl, but\n> I think that's probably going too far. In particular, for the\n> legacy mixed-case names like \"DateStyle\", what ends up in the\n> table is the mixed-case name, which seems problematic. It would\n> definitely be problematic if an extension used such a name,\n> because we might or might not be aware of the idiosyncratic\n> casing at the time a GRANT is issued. I'm thinking that we\n> really want to avoid looking up custom GUCs at all during GRANT,\n> because that can't do anything except create hazards.\n\nYikes. It took a few tries to see what you mean. Yes, if the GRANT happens before the LOAD, that can have bad consequences. So I agree something should be changed.\n\n> So I think that instead of what you've got here, we should\n> (1) apply the map_old_guc_names[] mapping, which is constant\n> (for any one PG release anyway)\n> (2) smash to lower case\n> (3) verify validity per valid_variable_name.\n> \n> This also simplifies life on the lookup side, where it's sufficient\n> to do steps (1) and (2) before performing a catalog search.\n> \n> Thoughts?\n\nThat sounds right. Do you already have something like that coded, or would you like me to post a patch?\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Mon, 4 Apr 2022 17:01:00 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Mark Dilger <mark.dilger@enterprisedb.com> writes:\n> On Apr 4, 2022, at 2:26 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> So I think that instead of what you've got here, we should\n>> (1) apply the map_old_guc_names[] mapping, which is constant\n>> (for any one PG release anyway)\n>> (2) smash to lower case\n>> (3) verify validity per valid_variable_name.\n>> \n>> This also simplifies life on the lookup side, where it's sufficient\n>> to do steps (1) and (2) before performing a catalog search.\n>> \n>> Thoughts?\n\n> That sounds right. Do you already have something like that coded, or would you like me to post a patch?\n\nWrote it already, no need for you to do it.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 04 Apr 2022 20:12:35 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"\n\n> On Apr 4, 2022, at 5:12 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Wrote it already, no need for you to do it.\n\nThanks!\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Mon, 4 Apr 2022 17:16:23 -0700","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"Here's a v16, which I have editorialized on rather heavily.\nI have not yet really looked at the docs, but I'm satisfied\nwith the code and tests now.\n\nThe main non-cosmetic changes I made:\n\n* I got rid of the mechanisms that tried to make has_parameter_privilege()\nact like USERSET GUCs have a forced public SET grant. I think that's\nfundamentally unworkable for custom GUCs, and it's just going to create\nconfusion and inconsistency. As the code stands, it'll report SET\nprivilege if you're superuser or have been explicitly granted that\nprivilege, but having that privilege only matters if the GUC is SUSET.\nNeed to craft some documentation wording to explain that, but I'd rather\nexplain that than say \"it might or might not tell you that SET privileges\nexist\". In any case I think this is perfectly symmetrical with the\nALTER SYSTEM privilege: you can grant that all you want, but it still\nwon't let you set variables whose context value forbids it.\n\n* That meant we have only one default ACL setting, so I was able to\nget rid of aclparameterdefault() altogether.\n\n* I took out the post-alter hook call in ExecGrant_Parameter. I'm\nprepared to listen to arguments why it should be put back in, but\nsaid arguments would have to explain why GRANT ON PARAMETER should\ndo that when no other form of GRANT does. (Or, possibly, it'd be\nsane for all the forms to call that hook; but that would be material\nfor a different patch IMV.)\n\n* I took out the pg_parameter_privileges view, which was undocumented\nanyway, and which I think people would not find useful. The fact that\nit's missing entries for most GUCs and might contain bogus entries\nfor nonexistent GUCs seems like a strong reason why it's not very\nuseful in this form. (But see below.)\n\n* I thought the tab-completion stuff was completely excessive.\nI got rid of the support for tab completion with both privileges\nspelled out, because I think people are just going to write ALL\ninstead. We don't have tab completion support for commands listing\nmultiple privileges of other object types, and I don't see why\nPARAMETER is the place to start. I also merged a bunch of duplicative\nrules by relying on MatchAny where possible.\n\n\nAside from documentation, there is one loose end that perhaps would\nbe best done in a follow-up patch: I'm somewhat of the opinion that\nwe ought to reinvent a \\dcp command for psql. I'd define it as being\na compact version of the pg_settings view, showing the name, value,\nand unit columns and probably not much else by default. With \"+\",\nwe could add on the privileges if any, and perhaps the context since\nyou need that to interpret the privileges. I think this'd be more\nuseful than the pg_parameter_privileges view, and it'd be handier\nthan pg_settings itself since you could do things like\n\t\\dcp autovac*\nwith a lot less typing than it takes to get the same info\nout of pg_settings directly.\n\nAnyway, I'm out of patience with this for today, and I'm throwing\nit up for the cfbot to have a look at. I'll hit the docs tomorrow.\n\n\t\t\tregards, tom lane","msg_date":"Tue, 05 Apr 2022 17:50:02 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"I wrote:\n> Anyway, I'm out of patience with this for today, and I'm throwing\n> it up for the cfbot to have a look at. I'll hit the docs tomorrow.\n\n... and of course, the cfbot gave me the raspberry, because of a\nmerge conflict in gram.y. Here's v17 rebased up to HEAD.\n\n\t\t\tregards, tom lane","msg_date":"Tue, 05 Apr 2022 18:52:18 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"},{"msg_contents":"I wrote:\n> Here's v17 rebased up to HEAD.\n\nPushed after fooling around with the docs.\n\nI have a couple of followup ideas in mind (\\dcp and another one),\nwhich I'll start separate threads about.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 06 Apr 2022 13:29:38 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Granting SET and ALTER SYSTE privileges for GUCs"}],"string":"[\n {\n \"msg_contents\": \"Hackers,\\n\\nIn the ongoing effort [1] to reduce the number of tasks which require operating under superuser privileges, this patch extends the system to allow, per GUC variable, the ability to SET or ALTER SYSTEM for the variable. A previous patch set was submitted [2] which created hard-coded privileged roles with the authority to manage associated hard-coded sets of GUC variables. This current patch appears superior in several ways:\\n\\n- It allows much greater flexibility in how roles and GUCs are associated\\n- Custom GUC variables defined by extensions can be covered by this approach\\n\\nand perhaps most importantly,\\n\\n- It's what Andrew suggested\\n\\nGranting SET privilege on a USERSET variable makes no practical difference, but for SUSET variables it does, and granting ALTER SYSTEM is meaningful for all variables. The patch does not mandate that non-login roles be created for this, but as a usage suggestion, one could define a non-login role and assign privileges for a set of GUCs, such as:\\n\\n CREATE ROLE regress_host_resource_admin NOSUPERUSER;\\n GRANT SET VALUE, ALTER SYSTEM ON\\n autovacuum_work_mem, hash_mem_multiplier, logical_decoding_work_mem,\\n maintenance_work_mem, max_stack_depth, min_dynamic_shared_memory,\\n shared_buffers, temp_buffers, temp_file_limit, work_mem\\n TO regress_host_resource_admin;\\n\\nand then delegate authority to manage the set of GUCs to a non-superuser by granting membership in non-login role:\\n\\n CREATE ROLE regress_admin_member IN ROLE regress_host_resource_admin;\\n\\nOne disadvantage of this approach is that the GUC variables are represented both in the list of C structures in guc.c and in the new system catalog pg_config_param's .dat file. Failure to enter a GUC in the .dat file will result in the inability to grant privileges on the GUC, at least unless/until you run CREATE CONFIGURATION PARAMETER on the GUC. (This is, in fact, how extension scripts deal with the issue.) It would perhaps be better if the list of GUCs were not duplicated, but I wasn't clever enough to find a clean way to do that without greatly expanding the patch (nor did I complete prototyping any such thing.)\\n\\n\\n\\n\\n\\n\\n[1] https://www.postgresql.org/message-id/flat/F9408A5A-B20B-42D2-9E7F-49CD3D1547BC%40enterprisedb.com\\n[2] https://www.postgresql.org/message-id/flat/0BE45772-31CF-4BE9-9FE8-FB7D8B130240%40enterprisedb.com\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 15 Nov 2021 12:37:44 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 15, 2021 at 3:37 PM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n> One disadvantage of this approach is that the GUC variables are represented both in the list of C structures in guc.c and in the new system catalog pg_config_param's .dat file. Failure to enter a GUC in the .dat file will result in the inability to grant privileges on the GUC, at least unless/until you run CREATE CONFIGURATION PARAMETER on the GUC. (This is, in fact, how extension scripts deal with the issue.) It would perhaps be better if the list of GUCs were not duplicated, but I wasn't clever enough to find a clean way to do that without greatly expanding the patch (nor did I complete prototyping any such thing.)\\n\\nI think this imposes an unnecessary burden on developers. It seems\\nlike it would be easy to write some code that lives inside guc.c and\\nruns during bootstrap, and it could just iterate over each\\nConfigureNamesWhatever array and insert catalog entries for everything\\nit finds.\\n\\nIt's also going to be important to think about what happens with\\nextension GUCs. If somebody installs an extension, we can't ask them\\nto perform a manual step in order to be able to grant privileges. And\\nif somebody then loads up a different .so for that extension, the set\\nof GUCs that it provides can change without any DDL being executed.\\nNew GUCs could appear, and old GUCs could vanish.\\n\\nSo I wonder if we should instead not do what I proposed above, and\\ninstead just adjust the GRANT command to automatically insert a new\\nrow into the relevant catalog if there isn't one already. That seems\\nnicer for extensions, and also nicer for core GUCs, since it avoids\\nbloating the catalog with a bunch of entries that aren't needed.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 10:03:34 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> It's also going to be important to think about what happens with\\n> extension GUCs. If somebody installs an extension, we can't ask them\\n> to perform a manual step in order to be able to grant privileges. And\\n> if somebody then loads up a different .so for that extension, the set\\n> of GUCs that it provides can change without any DDL being executed.\\n> New GUCs could appear, and old GUCs could vanish.\\n\\nRight. I think that any design that involves per-GUC catalog entries\\nis going to be an abysmal failure, precisely because the set of GUCs\\nis not stable enough. So I'm suspicious of this entire proposal.\\nMaybe there's a way to make it work, but that way isn't how.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 10:17:15 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 10:17 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Right. I think that any design that involves per-GUC catalog entries\\n> is going to be an abysmal failure, precisely because the set of GUCs\\n> is not stable enough.\\n\\nIn practice it's pretty stable. I think it's just a matter of having a\\nplan that covers the cases where it isn't stable reasonably elegantly.\\n\\nWe already embed GUC names in catalog entries when someone runs ALTER\\nUSER SET or ALTER DATABASE SET, so this proposal doesn't seem to be\\nmoving the goalposts in any meaningful way.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 10:23:27 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 16, 2021, at 7:03 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> \\n> It's also going to be important to think about what happens with\\n> extension GUCs. If somebody installs an extension, we can't ask them\\n> to perform a manual step in order to be able to grant privileges.\\n\\nThe burden isn't on the installer of an extension. As implemented, it's the extension's installation .sql file that sets it up, and the upgrade .sql files that make adjustments, if necessary.\\n\\n> And\\n> if somebody then loads up a different .so for that extension, the set\\n> of GUCs that it provides can change without any DDL being executed.\\n> New GUCs could appear, and old GUCs could vanish.\\n\\nWell, the same is true for functions, right? If you add, remove, or redefine functions in the extension, you need an upgrade script that defines the new functions, removes the old ones, changes function signatures, or whatever. The same is true here for GUCs.\\n\\nI don't think we support using a .so that is mismatched against the version of the extension that is installed.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 07:24:51 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Tue, Nov 16, 2021 at 10:17 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Right. I think that any design that involves per-GUC catalog entries\\n>> is going to be an abysmal failure, precisely because the set of GUCs\\n>> is not stable enough.\\n\\n> In practice it's pretty stable. I think it's just a matter of having a\\n> plan that covers the cases where it isn't stable reasonably elegantly.\\n\\nUm. Really the point comes down to having sane default behavior\\nwhen there's no entry, which ought to eliminate any need to do the\\nsort of \\\"run over all the entries at startup\\\" processing that you\\nseemed to be proposing. So I guess I don't understand why such a\\nthing would be needed.\\n\\n> We already embed GUC names in catalog entries when someone runs ALTER\\n> USER SET or ALTER DATABASE SET, so this proposal doesn't seem to be\\n> moving the goalposts in any meaningful way.\\n\\nTrue; as long as the expectation is that entries will exist for only\\na tiny subset of GUCs, it's probably fine.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 10:28:57 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> I don't think we support using a .so that is mismatched against the\\n> version of the extension that is installed.\\n\\nYou are entirely mistaken. That's not only \\\"supported\\\", it's *required*.\\nConsider binary upgrades, where the SQL objects are transferred as-is\\nbut the receiving installation may have a different (hopefully newer)\\nversion of the .so. That .so has to cope with the older SQL object\\ndefinitions; it doesn't get to assume that the upgrade script has been\\nrun yet.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 10:32:07 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 16, 2021, at 7:32 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n>> I don't think we support using a .so that is mismatched against the\\n>> version of the extension that is installed.\\n> \\n> You are entirely mistaken. That's not only \\\"supported\\\", it's *required*.\\n> Consider binary upgrades, where the SQL objects are transferred as-is\\n> but the receiving installation may have a different (hopefully newer)\\n> version of the .so. That .so has to cope with the older SQL object\\n> definitions; it doesn't get to assume that the upgrade script has been\\n> run yet.\\n\\nYou are talking about mismatches in the other direction, aren't you? I was responding to Robert's point that new gucs could appear, and old gucs disappear. That seems analogous to new functions appearing and old functions disappearing. If you upgrade (not downgrade) the .so, the new gucs and functions will be in the .so, but won't be callable/grantable from sql until the upgrade script runs. The old gucs and functions will be missing from the .so, and attempts to call them/grant them from sql before the upgrade will fail. What am I missing here?\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 07:45:14 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 16, 2021, at 7:28 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> True; as long as the expectation is that entries will exist for only\\n> a tiny subset of GUCs, it's probably fine.\\n\\nI understand that bloating a frequently used catalog can be pretty harmful to performance. I wasn't aware that the size of an infrequently used catalog was critical. This new catalog would be used during GRANT SET ... and GRANT ALTER SYSTEM commands, which should be rare, and potentially consulted when SET or ALTER SYSTEM commands are issued. Is there a more substantial performance impact to this than I'm aware? It can be a bit challenging to run performance tests on such things, given the way everything interacts with everything else. \\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 07:52:06 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 10:45 AM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n> You are talking about mismatches in the other direction, aren't you? I was responding to Robert's point that new gucs could appear, and old gucs disappear. That seems analogous to new functions appearing and old functions disappearing. If you upgrade (not downgrade) the .so, the new gucs and functions will be in the .so, but won't be callable/grantable from sql until the upgrade script runs. The old gucs and functions will be missing from the .so, and attempts to call them/grant them from sql before the upgrade will fail. What am I missing here?\\n\\nIt's true that we could impose such a restriction, but I don't think\\nwe should. If I install a different .so, I want the new GUCs to be\\ngrantable immediately, without running any separate DDL.\\n\\nI also don't think we should burden extension authors with putting\\nstuff in their upgrade scripts for this. We should solve the problem\\nin our code rather than forcing them to do so in theirs.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 11:07:36 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> You are talking about mismatches in the other direction, aren't you? I was responding to Robert's point that new gucs could appear, and old gucs disappear. That seems analogous to new functions appearing and old functions disappearing. If you upgrade (not downgrade) the .so, the new gucs and functions will be in the .so, but won't be callable/grantable from sql until the upgrade script runs.\\n\\nTrue, but in general users might not care about getting access to new\\nfeatures (or at least, they might not care until much later, at which\\npoint they'd bother to run ALTER EXTENSION UPDATE).\\n\\n> The old gucs and functions will be missing from the .so, and attempts to call them/grant them from sql before the upgrade will fail. What am I missing here?\\n\\nHere, you are describing behavior that's against project policy and would\\nbe rejected immediately if anyone submitted a patch that made an extension\\ndo that. Newer .so versions are expected to run successfully, not fail,\\nwith an older version of their SQL objects. Were this not so, it's almost\\ninevitable that a binary upgrade would fail, because the extension is\\nlikely to be called into action somewhere before there is any opportunity\\nto issue ALTER EXTENSION UPDATE. Even in-place updates of extensions\\nwould become problematic: you can't assume that a rollout of new\\nexecutables will be instantly followed by ALTER EXTENSION UPDATE.\\n\\nBasically, I think the proposed new system catalog is both unworkable\\nand entirely unnecessary. Maybe it would have been okay if Peter had\\ndesigned GUCs like that a couple of decades ago, but at this point\\nwe have a ton of infrastructure --- much of it not under the core\\nproject's control --- that assumes that GUCs can be created with just\\na bit of code. I do not think that this feature is worth the cost of\\nchanging that. Or IOW: ALTER SYSTEM SET has gotten along fine without\\nsuch a catalog; why can't this feature?\\n\\nI also notice that the patch seems to intend to introduce tracking\\nof which user \\\"owns\\\" a GUC, which I think is even more unworkable.\\nThey are all going to wind up owned by the bootstrap superuser\\n(extension GUCs too), so why bother?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 11:29:06 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> On Nov 16, 2021, at 7:28 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> True; as long as the expectation is that entries will exist for only\\n>> a tiny subset of GUCs, it's probably fine.\\n\\n> I understand that bloating a frequently used catalog can be pretty\\n> harmful to performance. I wasn't aware that the size of an infrequently\\n> used catalog was critical.\\n\\nMy concern is not about performance, it's about the difficulty of\\nmaintaining a catalog that expects to be a more-or-less exhaustive\\nlist of GUCs. I think you need to get rid of that expectation.\\nIn the analogy to ALTER DATABASE/USER SET, we don't expect that\\npg_db_role_setting catalog entries will exist for all, or even\\nvery many, GUCs. Also, the fact that pg_db_role_setting entries\\naren't tied very hard to the actual existence of a GUC is a good\\nthing from the maintenance and upgrade standpoint.\\n\\nBTW, if we did create such a catalog, there would need to be\\npg_dump and pg_upgrade support for its entries, and you'd have to\\nthink about (e.g.) whether pg_upgrade would attempt to preserve\\nthe same OIDs. I don't see any indication that the patch has\\naddressed that infrastructure ... which is probably just as well,\\nsince it's work that I'd be wanting to reject. (Hm, but actually,\\ndoesn't pg_dump need work anyway to dump this new type of GRANT?)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 11:44:07 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 16, 2021, at 8:44 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> My concern is not about performance, it's about the difficulty of\\n> maintaining a catalog that expects to be a more-or-less exhaustive\\n> list of GUCs. I think you need to get rid of that expectation.\\n\\nI'm preparing a new version of the patch that has the catalog empty to begin with, and only adds values in response to GRANT commands. That also handles the issues of extension upgrades, which I think the old patch handled better than the discussion on this thread suggested, but no matter. The new behavior will allow granting privileges on non-existent gucs, just as ALTER ROLE..SET allows registering settings on non-existent gucs.\\n\\nThe reason I had resisted allowing grants of privileges on non-existent gucs is that you can get the following sort of behavior (note the last two lines):\\n\\n DROP USER regress_priv_user7;\\n ERROR: role \\\"regress_priv_user7\\\" cannot be dropped because some objects depend on it\\n DETAIL: privileges for table persons2\\n privileges for configuration parameter sort_mem\\n privileges for configuration parameter no_such_param\\n\\nRejecting \\\"no_such_param\\\" in the original GRANT statement seemed cleaner to me, but this discussion suggests pretty strongly that I can't do it that way. Changing the historical \\\"sort_mem\\\" to the canonical \\\"work_mem\\\" name also seems better to me, as otherwise you could have different grants on the same GUC under different names. I'm inclined to keep the canonicalization of names where known, but maybe that runs afoul the rule that these grants should not be tied very hard to the GUC?\\n\\n> In the analogy to ALTER DATABASE/USER SET, we don't expect that\\n> pg_db_role_setting catalog entries will exist for all, or even\\n> very many, GUCs. Also, the fact that pg_db_role_setting entries\\n> aren't tied very hard to the actual existence of a GUC is a good\\n> thing from the maintenance and upgrade standpoint.\\n\\nDoing it that way....\\n\\n> BTW, if we did create such a catalog, there would need to be\\n> pg_dump and pg_upgrade support for its entries, and you'd have to\\n> think about (e.g.) whether pg_upgrade would attempt to preserve\\n> the same OIDs. I don't see any indication that the patch has\\n> addressed that infrastructure ... which is probably just as well,\\n> since it's work that I'd be wanting to reject.\\n\\nYeah, that's why I didn't write it. I wanted feedback on the basic implementation before doing that work.\\n\\n> (Hm, but actually,\\n> doesn't pg_dump need work anyway to dump this new type of GRANT?)\\n\\nYes, if the idea of this kind of grant isn't being outright rejected, then I'll need to write that.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 11:48:16 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 11/16/21 14:48, Mark Dilger wrote:\\n>\\n>> On Nov 16, 2021, at 8:44 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>\\n>> My concern is not about performance, it's about the difficulty of\\n>> maintaining a catalog that expects to be a more-or-less exhaustive\\n>> list of GUCs. I think you need to get rid of that expectation.\\n> I'm preparing a new version of the patch that has the catalog empty to begin with, and only adds values in response to GRANT commands. That also handles the issues of extension upgrades, which I think the old patch handled better than the discussion on this thread suggested, but no matter. The new behavior will allow granting privileges on non-existent gucs, just as ALTER ROLE..SET allows registering settings on non-existent gucs.\\n\\n\\n\\nYour original and fairly simple set of patches used hardcoded role names\\nand sets of GUCs they could update via ALTER SYSTEM. I suggested to you\\nprivately that a more flexible approach would be to drive this from a\\ncatalog table. I had in mind a table of more or less <roleid, guc_name>.\\nYou could prepopulate it with the roles / GUCs from your original patch\\nset. I don't think it needs to be initially empty. But DBAs would be\\nable to modify and extend the settings. I agree with Tom that we\\nshouldn't try to cover all GUCs in the table - any GUC without an entry\\ncan only be updated by a superuser.\\n\\n\\nTo support pg_dump and pg_upgrade, it might be better to have an\\nenabled/disabled flag rather than to delete rows.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 15:38:07 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 16, 2021, at 12:38 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> \\n> Your original and fairly simple set of patches used hardcoded role names\\n> and sets of GUCs they could update via ALTER SYSTEM. I suggested to you\\n> privately that a more flexible approach would be to drive this from a\\n> catalog table. I had in mind a table of more or less <roleid, guc_name>.\\n\\nRight. I prefer a table of <guc_name, acl> matching the structure of most of the rest of the grantable objects, so that entries from pg_depend or pg_shdepend can track the dependencies in the usual way and prevent dangling references when DROP ROLE is executed. Is there a reason to prefer an ad hoc tables structure?\\n\\n> You could prepopulate it with the roles / GUCs from your original patch\\n> set. I don't think it needs to be initially empty. But DBAs would be\\n> able to modify and extend the settings. I agree with Tom that we\\n> shouldn't try to cover all GUCs in the table - any GUC without an entry\\n> can only be updated by a superuser.\\n\\nThe patch already posted on this thread already works that way. Robert and Tom seemed to infer that all gucs need to be present in the catalog, but in fact, not entering them in the catalog simply means that they aren't grantable. I think the confusion arose from the fact that I created a mechanism for extension authors to enter additional gucs into the catalog, which gave the false impression that such entries were required. They are not. I didn't bother explaining that before, because Robert and Tom both seem to expect all gucs to be grantable, an expectation you do not appear to share.\\n\\n> To support pg_dump and pg_upgrade, it might be better to have an\\n> enabled/disabled flag rather than to delete rows.\\n\\nI'm not really sure what this means. Are you suggesting that the <guc_name, acl> (or in your formulation <roleid, guc_name>) should have a \\\"bool enabled\\\" field, and when the guc gets dropped, the \\\"enabled\\\" field gets set to false?\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 13:02:44 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> The patch already posted on this thread already works that way. Robert and Tom seemed to infer that all gucs need to be present in the catalog, but in fact, not entering them in the catalog simply means that they aren't grantable. I think the confusion arose from the fact that I created a mechanism for extension authors to enter additional gucs into the catalog, which gave the false impression that such entries were required. They are not. I didn't bother explaining that before, because Robert and Tom both seem to expect all gucs to be grantable, an expectation you do not appear to share.\\n\\nThe question is why you need pg_config_param at all, then.\\nAFAICS it just adds maintenance complexity we could do without.\\nI think we'd be better off with a catalog modeled on the design of\\npg_db_role_setting, which would have entries for roles and lists\\nof GUC names that those roles could set.\\n\\nBTW, another objection to pg_config_param as designed here is that\\na \\\"name\\\" is not an appropriate way to store possibly-qualified\\ncustom GUC names. It's not long enough (cf. valid_custom_variable_name).\\n\\n>> To support pg_dump and pg_upgrade, it might be better to have an\\n>> enabled/disabled flag rather than to delete rows.\\n\\n> I'm not really sure what this means.\\n\\nI didn't see the point of this either. We really need to KISS here.\\nEvery bit of added complexity in the catalog representation is another\\nopportunity for bugs-of-omission, not to mention a detail that you\\nhave to provide mechanisms to dump and restore.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 17:12:49 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 16, 2021, at 2:12 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> The question is why you need pg_config_param at all, then.\\n> AFAICS it just adds maintenance complexity we could do without.\\n> I think we'd be better off with a catalog modeled on the design of\\n> pg_db_role_setting, which would have entries for roles and lists\\n> of GUC names that those roles could set.\\n\\nOriginally, I was trying to have dependency linkage between two proper types of objects, so that DROP ROLE and DROP CONFIGURATION PARAMETER would behave as expected, complaining about privileges remaining rather than dropping an object and leaving a dangling reference.\\n\\nI've deleted the whole CREATE CONFIGURATION PARAMETER and DROP CONFIGURATION PARAMETER syntax and implementation, but it still seems odd to me that:\\n\\n CREATE ROLE somebody;\\n GRANT SELECT ON TABLE sometable TO ROLE somebody;\\n GRANT ALTER SYSTEM ON someguc TO ROLE somebody;\\n DROP ROLE somebody;\\n ERROR: role \\\"somebody\\\" cannot be dropped because some objects depend on it\\n DETAIL: privileges for table sometable\\n\\nwould not mention privileges for \\\"someguc\\\" as well. That's why I want configuration parameters to be proper objects with OIDs and with entries in pg_depend and/or pg_shdepend. Maybe there is some better way to do it, but that's why I've been doing this.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 14:25:12 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 16, 2021, at 2:12 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> BTW, another objection to pg_config_param as designed here is that\\n> a \\\"name\\\" is not an appropriate way to store possibly-qualified\\n> custom GUC names. It's not long enough (cf. valid_custom_variable_name).\\n\\nI was aware of that, but figured not all GUCs have to be grantable. If it doesn't fit in a NameData, you can't grant on it.\\n\\nIf we want to be more accommodating than that, we can store it as text, just like pg_db_role_names does, but then we need more code complexity to look it up and to verify that it is unique. (We wouldn't want multiple records for the same <role,guc> pair.) \\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 14:45:12 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 2:48 PM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n> I'm preparing a new version of the patch that has the catalog empty to begin with, and only adds values in response to GRANT commands. That also handles the issues of extension upgrades, which I think the old patch handled better than the discussion on this thread suggested, but no matter. The new behavior will allow granting privileges on non-existent gucs, just as ALTER ROLE..SET allows registering settings on non-existent gucs.\\n>\\n> The reason I had resisted allowing grants of privileges on non-existent gucs is that you can get the following sort of behavior (note the last two lines):\\n>\\n> DROP USER regress_priv_user7;\\n> ERROR: role \\\"regress_priv_user7\\\" cannot be dropped because some objects depend on it\\n> DETAIL: privileges for table persons2\\n> privileges for configuration parameter sort_mem\\n> privileges for configuration parameter no_such_param\\n>\\n> Rejecting \\\"no_such_param\\\" in the original GRANT statement seemed cleaner to me, but this discussion suggests pretty strongly that I can't do it that way.\\n\\nI think it's perfectly fine to refuse a GRANT statement on a GUC that\\ndoesn't exist, and I don't believe I ever said otherwise. What I don't\\nthink is OK is to require a preparatory statement like CREATE\\nCONFIGURATION PARAMETER to be executed before you can grant\\npermissions on it. There's no reason for that. If somebody tries to\\ngrant privileges on a GUC that does not exist, fail. If the GUC does\\nexist but there's no catalog entry, make one. If the GUC exists and\\nthe catalog entry also exists, update it.\\n\\nAt REVOKE time, don't check whether the GUC exists - only check the\\ncatalog. That way people can remove privileges on GUCs that don't\\nexist any more. If somebody issues a REVOKE and there's no catalog\\nentry, do nothing. If somebody issues a REVOKE and there is a catalog\\nentry, remove stuff from it; but if that would leave it completely\\nempty, instead delete it.\\n\\nWhenever you create a catalog entry, also add dependency entries\\npointing to it. When you delete one, remove those entries.\\n\\n> Changing the historical \\\"sort_mem\\\" to the canonical \\\"work_mem\\\" name also seems better to me, as otherwise you could have different grants on the same GUC under different names. I'm inclined to keep the canonicalization of names where known, but maybe that runs afoul the rule that these grants should not be tied very hard to the GUC?\\n\\nNo. If somebody grants privileges on an old name, record the grant\\nunder the canonical name.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 08:26:33 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 3:38 PM Andrew Dunstan <andrew@dunslane.net> wrote:\\n> Your original and fairly simple set of patches used hardcoded role names\\n> and sets of GUCs they could update via ALTER SYSTEM. I suggested to you\\n> privately that a more flexible approach would be to drive this from a\\n> catalog table. I had in mind a table of more or less <roleid, guc_name>.\\n> You could prepopulate it with the roles / GUCs from your original patch\\n> set. I don't think it needs to be initially empty. But DBAs would be\\n> able to modify and extend the settings. I agree with Tom that we\\n> shouldn't try to cover all GUCs in the table - any GUC without an entry\\n> can only be updated by a superuser.\\n\\nI simply can't understand the point of this. You're basically\\nproposing that somebody has to execute one SQL statement to make a GUC\\ngrantable, and then a second SQL statement to actually grant access to\\nit. What is the value in that? It is the same person doing both\\nthings, and the system can work out automatically what needs to be\\ndone.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 08:27:59 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 5:45 PM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n> I was aware of that, but figured not all GUCs have to be grantable. If it doesn't fit in a NameData, you can't grant on it.\\n\\nSuch restrictions are rather counterintuitive for users, and here it\\ndoesn't even buy anything. Using 'text' rather than 'name' as the data\\ntype isn't going to cost any meaningful amount of performance.\\n\\n> If we want to be more accommodating than that, we can store it as text, just like pg_db_role_names does, but then we need more code complexity to look it up and to verify that it is unique. (We wouldn't want multiple records for the same <role,guc> pair.)\\n\\nIf you're verifying that it's unique in any way other than using a\\nunique index, I think you're doing it wrong.\\n\\nAlso, maybe I'm confused here, but why isn't the schema:\\n\\ngucoid\\ngucname\\ngucacl\\n\\nIOW, I don't understand why this table has <role,guc> as the primary\\nkey rather than just guc. Everywhere else, we have a single ACL array\\nfor the all privileges on an object. Why wouldn't we do the same thing\\nhere?\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 08:32:27 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 11/17/21 08:32, Robert Haas wrote:\\n> On Tue, Nov 16, 2021 at 5:45 PM Mark Dilger\\n> <mark.dilger@enterprisedb.com> wrote:\\n>> I was aware of that, but figured not all GUCs have to be grantable. If it doesn't fit in a NameData, you can't grant on it.\\n> Such restrictions are rather counterintuitive for users, and here it\\n> doesn't even buy anything. Using 'text' rather than 'name' as the data\\n> type isn't going to cost any meaningful amount of performance.\\n\\n\\nindeed\\n\\n\\n>> If we want to be more accommodating than that, we can store it as text, just like pg_db_role_names does, but then we need more code complexity to look it up and to verify that it is unique. (We wouldn't want multiple records for the same <role,guc> pair.)\\n> If you're verifying that it's unique in any way other than using a\\n> unique index, I think you're doing it wrong.\\n\\n\\nyeah\\n\\n\\n>\\n> Also, maybe I'm confused here, but why isn't the schema:\\n>\\n> gucoid\\n> gucname\\n> gucacl\\n>\\n> IOW, I don't understand why this table has <role,guc> as the primary\\n> key rather than just guc. Everywhere else, we have a single ACL array\\n> for the all privileges on an object. Why wouldn't we do the same thing\\n> here?\\n>\\n\\n\\nYes, that should work, It seems like a better scheme.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 09:24:39 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 11/16/21 17:12, Tom Lane wrote:\\n>\\n>>> To support pg_dump and pg_upgrade, it might be better to have an\\n>>> enabled/disabled flag rather than to delete rows.\\n>> I'm not really sure what this means.\\n> I didn't see the point of this either. We really need to KISS here.\\n> Every bit of added complexity in the catalog representation is another\\n> opportunity for bugs-of-omission, not to mention a detail that you\\n> have to provide mechanisms to dump and restore.\\n>\\n> \\t\\t\\t\\n\\n\\nWell, I was trying (perhaps not very well) to imagine how to deal with\\nsomeone modifying the permissions of one of the predefined roles. Say\\npg_foo has initial permission to set bar and baz, and the DBA removes\\npermission to set baz. How is pg_dump going to emit the right commands\\nto allow a safe pg_upgrade? Maybe we should say that the permissions for\\nthe predefined roles are immutable, so only permissions sets for user\\ndefined roles are mutable.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n-- \\n\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 09:31:40 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 17, 2021, at 5:32 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> \\n>> I was aware of that, but figured not all GUCs have to be grantable. If it doesn't fit in a NameData, you can't grant on it.\\n> \\n> Such restrictions are rather counterintuitive for users, and here it\\n> doesn't even buy anything. Using 'text' rather than 'name' as the data\\n> type isn't going to cost any meaningful amount of performance.\\n\\nThat sounds fine.\\n\\n>> If we want to be more accommodating than that, we can store it as text, just like pg_db_role_names does, but then we need more code complexity to look it up and to verify that it is unique. (We wouldn't want multiple records for the same <role,guc> pair.)\\n> \\n> If you're verifying that it's unique in any way other than using a\\n> unique index, I think you're doing it wrong.\\n\\nNo, I'm using a unique index. I was overthinking it, concerned about changing from name_ops to text_ops and needing the toast table, but that's silly, because I need one for the acl anyway.\\n\\n> Also, maybe I'm confused here, but why isn't the schema:\\n> \\n> gucoid\\n> gucname\\n> gucacl\\n\\nIt is, both in v2 already posted, and in the v3, written but not yet posted, as I haven't finished the pg_dump work, and also I'm waiting to see how this discussion gets resolved before asking for a review of v3.\\n\\n> IOW, I don't understand why this table has <role,guc> as the primary\\n> key rather than just guc.\\n\\nI was responding to Tom's recommendation that I follow the pattern in pg_db_role_setting, and speculating how that would work. I was not proposing to do it that way.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 07:57:41 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 17, 2021, at 6:31 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> \\n> Well, I was trying (perhaps not very well) to imagine how to deal with\\n> someone modifying the permissions of one of the predefined roles. Say\\n> pg_foo has initial permission to set bar and baz, and the DBA removes\\n> permission to set baz. How is pg_dump going to emit the right commands\\n> to allow a safe pg_upgrade? Maybe we should say that the permissions for\\n> the predefined roles are immutable, so only permissions sets for user\\n> defined roles are mutable.\\n\\nI find this somewhat amusing. When you suggested off-list that I make gucs individually grantable rather than creating predefined roles with privileges, that was a great idea precisely because sites could define their own security policies using their own site-defined roles:\\n\\n CREATE ROLE admin_type_a NOLOGIN NOSUPERUSER;\\n CREATE ROLE admin_type_b NOLOGIN NOSUPERUSER;\\n ...\\n GRANT ALTER SYSTEM ON guc_a1, guc_a2, guc_a3, ... TO admin_type_a;\\n GRANT ALTER SYSTEM ON guc_b1, guc_b2, guc_b3, ... TO admin_type_b;\\n ...\\n\\nThat has all the power of a system based on predefined roles, but with site-specific flexibility, which is better. So it amuses me that we'd now be talking about granting some of these to predefined roles, as that is a regression in flexibility. (How would a site revoke it from one of those predefined roles if they wanted a different policy?)\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 08:07:04 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 9:31 AM Andrew Dunstan <andrew@dunslane.net> wrote:\\n> Well, I was trying (perhaps not very well) to imagine how to deal with\\n> someone modifying the permissions of one of the predefined roles. Say\\n> pg_foo has initial permission to set bar and baz, and the DBA removes\\n> permission to set baz. How is pg_dump going to emit the right commands\\n> to allow a safe pg_upgrade? Maybe we should say that the permissions for\\n> the predefined roles are immutable, so only permissions sets for user\\n> defined roles are mutable.\\n\\nThat's a great question, but it isn't a new problem. If I create a\\nbrand new database and do thIs:\\n\\nrhaas=# revoke execute on function pg_ls_waldir() from pg_monitor;\\nREVOKE\\n\\nAnd then I do this:\\n\\n[rhaas pgsql]$ pg_dump\\n\\nThen the output includes this:\\n\\nREVOKE ALL ON FUNCTION pg_catalog.pg_ls_waldir(OUT name text, OUT size\\nbigint, OUT modification timestamp with time zone) FROM pg_monitor;\\n\\nI recommend looking at how that works and making this work the same way.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 12:05:53 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 11/17/21 11:07, Mark Dilger wrote:\\n>\\n>> On Nov 17, 2021, at 6:31 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>\\n>> Well, I was trying (perhaps not very well) to imagine how to deal with\\n>> someone modifying the permissions of one of the predefined roles. Say\\n>> pg_foo has initial permission to set bar and baz, and the DBA removes\\n>> permission to set baz. How is pg_dump going to emit the right commands\\n>> to allow a safe pg_upgrade? Maybe we should say that the permissions for\\n>> the predefined roles are immutable, so only permissions sets for user\\n>> defined roles are mutable.\\n> I find this somewhat amusing. When you suggested off-list that I make gucs individually grantable rather than creating predefined roles with privileges, that was a great idea precisely because sites could define their own security policies using their own site-defined roles:\\n>\\n> CREATE ROLE admin_type_a NOLOGIN NOSUPERUSER;\\n> CREATE ROLE admin_type_b NOLOGIN NOSUPERUSER;\\n> ...\\n> GRANT ALTER SYSTEM ON guc_a1, guc_a2, guc_a3, ... TO admin_type_a;\\n> GRANT ALTER SYSTEM ON guc_b1, guc_b2, guc_b3, ... TO admin_type_b;\\n> ...\\n>\\n> That has all the power of a system based on predefined roles, but with site-specific flexibility, which is better. So it amuses me that we'd now be talking about granting some of these to predefined roles, as that is a regression in flexibility. (How would a site revoke it from one of those predefined roles if they wanted a different policy?)\\n>\\n\\nI agree it's not ideal. At the time I suggested a more flexible approach\\nI hadn't really thought about the problems of upgrading. If you can come\\nup with something that works there then I'll be all ears.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 12:06:50 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 17, 2021, at 9:06 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> \\n> I agree it's not ideal. At the time I suggested a more flexible approach\\n> I hadn't really thought about the problems of upgrading. If you can come\\n> up with something that works there then I'll be all ears.\\n\\nAre you talking about upgrades preserving revocations of privileges on gucs from predefined roles, or merely preserving grants and revocation of privileges on gucs to regular roles? I think the former problem is easily handled by not shipping any predefined roles with such privileges. The latter problem would seem to be a mere matter of programming, something I'm working on but don't have finished. (But maybe you see dragons ahead for me that I'm not seeing yet?)\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 09:12:06 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 11/17/21 12:12, Mark Dilger wrote:\\n>\\n>> On Nov 17, 2021, at 9:06 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>\\n>> I agree it's not ideal. At the time I suggested a more flexible approach\\n>> I hadn't really thought about the problems of upgrading. If you can come\\n>> up with something that works there then I'll be all ears.\\n> Are you talking about upgrades preserving revocations of privileges on gucs from predefined roles, or merely preserving grants and revocation of privileges on gucs to regular roles? I think the former problem is easily handled by not shipping any predefined roles with such privileges. The latter problem would seem to be a mere matter of programming, something I'm working on but don't have finished. (But maybe you see dragons ahead for me that I'm not seeing yet?)\\n>\\n\\nYes, if we don't ship with any preset privileges then the problem I was\\nthinking about disappears.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 15:24:09 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Nov 16, 2021, at 7:03 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> \\n> It's also going to be important to think about what happens with\\n> extension GUCs. If somebody installs an extension, we can't ask them\\n> to perform a manual step in order to be able to grant privileges. And\\n> if somebody then loads up a different .so for that extension, the set\\n> of GUCs that it provides can change without any DDL being executed.\\n> New GUCs could appear, and old GUCs could vanish.\\n\\nThe v3 patch allows grants on unrecognized guc names. This should allow a grant statement to precede the loading of a new .so which provides the named guc.\\n\\n> instead just adjust the GRANT command to automatically insert a new\\n> row into the relevant catalog if there isn't one already. That seems\\n> nicer for extensions, and also nicer for core GUCs, since it avoids\\n> bloating the catalog with a bunch of entries that aren't needed.\\n\\nGrants on GUCs create a new catalog entry if necessary, or update the existing catalog entry if found.\\n\\nThere is a new information_schema.guc_privileges view, not present in v2.\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 22 Nov 2021 16:21:42 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 22, 2021 at 7:21 PM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n> There is a new information_schema.guc_privileges view, not present in v2.\\n\\nIt's my impression that information_schema is a child of the SQL\\nstandard, and that inventions specific to PG go in pg_catalog.\\n\\nAlso, I think the user-facing name for GUCs is \\\"settings\\\".\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 23 Nov 2021 11:07:12 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Nov 23, 2021, at 8:07 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> \\n> It's my impression that information_schema is a child of the SQL\\n> standard, and that inventions specific to PG go in pg_catalog.\\n> \\n> Also, I think the user-facing name for GUCs is \\\"settings\\\".\\n\\nThanks. These issues should be fixed in v4.\\n\\nAlong the way, I also added has_setting_privilege() functions overlooked in v3 and before.\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Tue, 23 Nov 2021 18:14:54 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 11/23/21 21:14, Mark Dilger wrote:\\n>\\n>> On Nov 23, 2021, at 8:07 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>\\n>> It's my impression that information_schema is a child of the SQL\\n>> standard, and that inventions specific to PG go in pg_catalog.\\n>>\\n>> Also, I think the user-facing name for GUCs is \\\"settings\\\".\\n> Thanks. These issues should be fixed in v4.\\n>\\n> Along the way, I also added has_setting_privilege() functions overlooked in v3 and before.\\n\\n\\n\\nThis patch had bit-rotted slightly, and I was attempting to remedy it.\\nHowever, I got a failure running the TAP tests because of this change:\\n\\n\\ndiff --git a/src/test/modules/test_pg_dump/t/001_base.pl\\nb/src/test/modules/test_pg_dump/t/001_base.pl\\nindex 16f7610883..7fbf2d871b 100644\\n--- a/src/test/modules/test_pg_dump/t/001_base.pl\\n+++ b/src/test/modules/test_pg_dump/t/001_base.pl\\n@@ -9,7 +9,12 @@ use PostgreSQL::Test::Cluster;\\n use PostgreSQL::Test::Utils;\\n use Test::More;\\n \\n-my $tempdir = PostgreSQL::Test::Utils::tempdir;\\n+# my $tempdir = PostgreSQL::Test::Utils::tempdir;\\n+my $tempbase = '/tmp/test_pg_dump';\\n+my $subdir = 0;\\n+$subdir++ while (-e \\\"$tempbase/$subdir\\\");\\n+my $tempdir = \\\"$tempbase/$subdir\\\";\\n+system(\\\"mkdir $tempdir\\\");\\n \\n\\n\\nWhat's going on here?\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 13 Dec 2021 15:56:34 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Dec 13, 2021, at 12:56 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> \\n> This patch had bit-rotted slightly, and I was attempting to remedy it.\\n\\nI have that already, and getting ready to post. Give me a few minutes and I'll repost.\\n\\n> However, I got a failure running the TAP tests because of this change:\\n> \\n> \\n> diff --git a/src/test/modules/test_pg_dump/t/001_base.pl\\n> b/src/test/modules/test_pg_dump/t/001_base.pl\\n> index 16f7610883..7fbf2d871b 100644\\n> --- a/src/test/modules/test_pg_dump/t/001_base.pl\\n> +++ b/src/test/modules/test_pg_dump/t/001_base.pl\\n> @@ -9,7 +9,12 @@ use PostgreSQL::Test::Cluster;\\n> use PostgreSQL::Test::Utils;\\n> use Test::More;\\n> \\n> -my $tempdir = PostgreSQL::Test::Utils::tempdir;\\n> +# my $tempdir = PostgreSQL::Test::Utils::tempdir;\\n> +my $tempbase = '/tmp/test_pg_dump';\\n> +my $subdir = 0;\\n> +$subdir++ while (-e \\\"$tempbase/$subdir\\\");\\n> +my $tempdir = \\\"$tempbase/$subdir\\\";\\n> +system(\\\"mkdir $tempdir\\\");\\n> \\n> \\n> \\n> What's going on here?\\n\\nYeah, I hit that, too. That was an accidentally committed bit of local testing. Please ignore it for now.\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 13 Dec 2021 12:58:25 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Dec 13, 2021, at 12:58 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> \\n> Give me a few minutes and I'll repost.\\n\\nOk, this new version is rebased, cleaned up, and has the ability to revoke SET VALUE on most settings which were previously userset, but are now suset with initdb time granted SET VALUE privilege to public.\\n\\nI have not yet altered gucs defined by contrib modules, but will repost in a few hours, or tomorrow, with those changes.\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 13 Dec 2021 13:33:58 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Dec 13, 2021, at 1:33 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> \\n> but will repost in a few hours\\n\\n... and here it is:\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 13 Dec 2021 14:34:35 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 13, 2021 at 5:34 PM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n>\\n>\\n>\\n> > On Dec 13, 2021, at 1:33 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> >\\n> > but will repost in a few hours\\n>\\n> ... and here it is:\\n\\ncurrently there is a failure in check-world (not sure if it's known):\\n\\ndiff -U3 /src/postgres/src/test/regress/expected/guc_privs.out\\n/src/postgres/src/test/regress/results/guc_privs.out\\n--- /src/postgres/src/test/regress/expected/guc_privs.out 2021-12-14\\n14:11:45.000000000 +0000\\n+++ /src/postgres/src/test/regress/results/guc_privs.out 2021-12-14\\n15:50:52.219583421 +0000\\n@@ -39,8 +39,10 @@\\n ALTER SYSTEM RESET autovacuum; -- ok\\n -- PGC_SUSET\\n SET lc_messages = 'en_US.UTF-8'; -- ok\\n+ERROR: invalid value for parameter \\\"lc_messages\\\": \\\"en_US.UTF-8\\\"\\n RESET lc_messages; -- ok\\n ALTER SYSTEM SET lc_messages = 'en_US.UTF-8'; -- ok\\n+ERROR: invalid value for parameter \\\"lc_messages\\\": \\\"en_US.UTF-8\\\"\\n ALTER SYSTEM RESET lc_messages; -- ok\\n -- PGC_SU_BACKEND\\n SET jit_debugging_support = OFF; -- fail, cannot be set after connection start\\n\\nAside from that I've tested this and it seems to function as\\nadvertised and in my view is worth adding to PG.\\n\\nOne thing that seems like an omission to me is the absence of a\\nInvokeObjectPostAlterHook in pg_setting_acl_aclcheck or\\npg_setting_acl_aclmask so that MAC extensions can also block this,\\nInvokeObjectPostCreateHook is already in the create path so a\\nPostAlter hook seems appropriate.\\n\\nThank you.\\n\\n\\n\",\n \"msg_date\": \"Tue, 14 Dec 2021 17:26:24 -0500\",\n \"msg_from\": \"Joshua Brindle <joshua.brindle@crunchydata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Dec 14, 2021, at 2:26 PM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\\n> \\n> currently there is a failure in check-world (not sure if it's known):\\n\\nThat one is definitely my fault. 'en_US.UTF-8' exists on my platform, so I hadn't noticed. I've changed it to use 'C', which should be portable.\\n\\n> One thing that seems like an omission to me is the absence of a\\n> InvokeObjectPostAlterHook in pg_setting_acl_aclcheck or\\n> pg_setting_acl_aclmask so that MAC extensions can also block this,\\n> InvokeObjectPostCreateHook is already in the create path so a\\n> PostAlter hook seems appropriate.\\n\\nGood catch, but that seems like a strange place to put a PostAlterHook, so I added it to ExecGrant_Setting for v6, instead. This seems more consistent with the hook in SetDefaultACL.\\n\\n(If you are really trying to do Managed Access Control (MAC), wouldn't that be a separate patch which adds security hooks into all *_aclcheck functions?)\\n\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Wed, 15 Dec 2021 07:56:12 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 15, 2021 at 10:56 AM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n>\\n>\\n>\\n> > On Dec 14, 2021, at 2:26 PM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\\n> >\\n> > currently there is a failure in check-world (not sure if it's known):\\n>\\n> That one is definitely my fault. 'en_US.UTF-8' exists on my platform, so I hadn't noticed. I've changed it to use 'C', which should be portable.\\n>\\n> > One thing that seems like an omission to me is the absence of a\\n> > InvokeObjectPostAlterHook in pg_setting_acl_aclcheck or\\n> > pg_setting_acl_aclmask so that MAC extensions can also block this,\\n> > InvokeObjectPostCreateHook is already in the create path so a\\n> > PostAlter hook seems appropriate.\\n>\\n> Good catch, but that seems like a strange place to put a PostAlterHook, so I added it to ExecGrant_Setting for v6, instead. This seems more consistent with the hook in SetDefaultACL.\\n\\nAh, I was actually requesting a hook where the acl check was done for\\nsetting a GUC, such that we could deny setting them in a hook,\\nsomething that would be useful for the set_user extension\\n(github.com/pgaudit/set_user) but having a hook for grant/revoke is\\nalso helpful.\\n\\n> (If you are really trying to do Managed Access Control (MAC), wouldn't that be a separate patch which adds security hooks into all *_aclcheck functions?)\\n\\nMAC is mandatory access controls, so something that can be layered on\\ntop of DAC and can only be enforced even on object owners and\\nsuperuser. sepgsql is a MAC extension, for example. It uses the object\\naccess hooks to enforce SELinux policy on PG objects.\\n\\n\\n\",\n \"msg_date\": \"Wed, 15 Dec 2021 13:02:35 -0500\",\n \"msg_from\": \"Joshua Brindle <joshua.brindle@crunchydata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Dec 15, 2021, at 10:02 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\\n> \\n> Ah, I was actually requesting a hook where the acl check was done for\\n> setting a GUC, such that we could deny setting them in a hook,\\n> something that would be useful for the set_user extension\\n> (github.com/pgaudit/set_user)\\n\\nHmm, this seems orthogonal to the patch under discussion. This patch only adds a pg_setting_acl_aclcheck in ExecSetVariableStmt() for settings which have been explicitly granted, otherwise it works the traditional way (checking whether the setting is suset/userset). I don't think you'd get MAC support without finding a way to fire the hook for all settings, regardless of their presence in the new pg_setting_acl table. That is hard, because InvokeObjectPostAlterHook expects the classId (SettingAclRelationId) and the objectId (pg_setting_acl.oid), but you don't have those for many (most?) settings. As discussed upthread, we *do not* want to force an entry into the table for all settings, only for ones that have been explicitly granted.\\n\\nDo you agree? I'm happy to support MAC in this patch if can explain a simple way of doing so.\\n\\n> but having a hook for grant/revoke is\\n> also helpful.\\n\\nYes, I see no reason to rip this out.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 15 Dec 2021 10:18:05 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 15, 2021 at 1:18 PM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n>\\n> > On Dec 15, 2021, at 10:02 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\\n> >\\n> > Ah, I was actually requesting a hook where the acl check was done for\\n> > setting a GUC, such that we could deny setting them in a hook,\\n> > something that would be useful for the set_user extension\\n> > (github.com/pgaudit/set_user)\\n>\\n> Hmm, this seems orthogonal to the patch under discussion. This patch only adds a pg_setting_acl_aclcheck in ExecSetVariableStmt() for settings which have been explicitly granted, otherwise it works the traditional way (checking whether the setting is suset/userset). I don't think you'd get MAC support without finding a way to fire the hook for all settings, regardless of their presence in the new pg_setting_acl table. That is hard, because InvokeObjectPostAlterHook expects the classId (SettingAclRelationId) and the objectId (pg_setting_acl.oid), but you don't have those for many (most?) settings. As discussed upthread, we *do not* want to force an entry into the table for all settings, only for ones that have been explicitly granted.\\n>\\n> Do you agree? I'm happy to support MAC in this patch if can explain a simple way of doing so.\\n\\nAh, I understand now. Would it be possible to pass the\\nSettingAclRelationId if it exists or InvalidOid if not? That way if a\\nMAC implementation cares about a particular GUC it'll ensure it's in\\npg_setting_acl.\\n\\nI don't know if others will object to that but it seems like an okay\\ncompromise.\\n\\n> > but having a hook for grant/revoke is\\n> > also helpful.\\n>\\n> Yes, I see no reason to rip this out.\\n>\\n\\n\\n\",\n \"msg_date\": \"Thu, 16 Dec 2021 10:43:15 -0500\",\n \"msg_from\": \"Joshua Brindle <joshua.brindle@crunchydata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Dec 16, 2021, at 7:43 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\\n> \\n> Ah, I understand now. Would it be possible to pass the\\n> SettingAclRelationId if it exists or InvalidOid if not?\\n\\nSettingAclRelationId is always defined, so we can always pass that value. But the settingId itself may sometimes be InvalidOid.\\n\\n> That way if a\\n> MAC implementation cares about a particular GUC it'll ensure it's in\\n> pg_setting_acl.\\n\\nA much cleaner solution would be to create new ObjectAccessTypes with a corresponding new Invoke macro and Run function. Those could take setting names, not Oids, and include additional information about whether the operation is SET, RESET or ALTER SYSTEM, what the new value is (if any), what kind of setting it is (bool, int, ...), etc. I don't think such a patch would even be all that hard to write.\\n\\nWhat do you think?\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 16 Dec 2021 09:53:45 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 16, 2021 at 12:53 PM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n>\\n>\\n>\\n> > On Dec 16, 2021, at 7:43 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\\n> >\\n> > Ah, I understand now. Would it be possible to pass the\\n> > SettingAclRelationId if it exists or InvalidOid if not?\\n>\\n> SettingAclRelationId is always defined, so we can always pass that value. But the settingId itself may sometimes be InvalidOid.\\n\\nYes, that is what I meant.\\n\\n> > That way if a\\n> > MAC implementation cares about a particular GUC it'll ensure it's in\\n> > pg_setting_acl.\\n>\\n> A much cleaner solution would be to create new ObjectAccessTypes with a corresponding new Invoke macro and Run function. Those could take setting names, not Oids, and include additional information about whether the operation is SET, RESET or ALTER SYSTEM, what the new value is (if any), what kind of setting it is (bool, int, ...), etc. I don't think such a patch would even be all that hard to write.\\n>\\n> What do you think?\\n\\nPersonally, I would be happy with that, but since it's a whole new\\nhooking method I suspect it'll be an uphill battle. That definitely\\nseems like another patchset though, if you do submit this I will test\\nand review.\\n\\nThank you.\\n\\n\\n\",\n \"msg_date\": \"Thu, 16 Dec 2021 13:31:58 -0500\",\n \"msg_from\": \"Joshua Brindle <joshua.brindle@crunchydata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Dec 15, 2021, at 7:56 AM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> \\n> <v6-0001-Allow-GRANT-of-SET-and-ALTER-SYSTEM-SET-for-gucs.patch>\\n\\nRebased:\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Wed, 26 Jan 2022 07:24:31 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Jan 26, 2022, at 7:24 AM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> \\n> Rebased:\\n> \\n> <v7-0001-Allow-GRANT-of-SET-and-ALTER-SYSTEM-SET-for-gucs.patch>\\n\\nHackers,\\n\\nJoshua and I had some off-list back-and-forth on this patch and I think the attached v8 patch meets his needs. Specifically, v8 adds support for object access hooks in the case of SET and ALTER SYSTEM SET of a guc variable.\\n\\nThe way InvokeObjectPostAlterHookArg() is written, it expects some things like a Oid that we don't always have. I abused the parameters of that function to pass as much information as possible. Anyone wanting to review this might consider how that is coded.\\n\\nJoshua, this is byte-for-byte the same patch that you already reviewed off-list (or I attached the wrong patch!) If you are satisfied, could you set the status to ready-for-committer?\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 31 Jan 2022 11:17:12 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> [ v8-0001-Allow-GRANT-of-SET-and-ALTER-SYSTEM-SET-for-gucs.patch ]\\n\\nI noticed that this is failing in the cfbot as a side-effect of my\\ncc50080a8, so here's a rebase (basically it just removes the no-\\nlonger-necessary sanity_check.out hunk).\\n\\nI haven't reviewed this particularly, or read the thread, but I'm\\npretty disturbed by the contrib/ changes that it proposes.\\n\\n1. If we need to change these two contrib modules, doesn't that imply\\na lot of changes forced on external modules as well? What are the\\nsecurity implications if somebody doesn't make such a change?\\n\\n2. It looks to me like if someone installs the updated postgres_fdw.so,\\nbut doesn't run ALTER EXTENSION UPDATE, they are going to be left with a\\nrather broken extension. This seems not good either, especially if it's\\nmultiplied by a boatload of third-party extensions requiring updates.\\n\\nSo I think some more thought is needed to see if we can't avoid\\nthe need to touch extension modules in this patch. Or at least\\navoid the need for synchronized C-level and SQL-level updates,\\nbecause that is going to create a lot of pain for end users.\\n\\nI'm also fairly distressed by the number of changes in guc.c,\\nmainly because I fear that that means that pending patches that\\nadd GUC variables will be subtly incorrect/insecure if they're not\\nupdated to account for this. Frankly, I also reject the apparent\\nposition that we need to support forbidding users from touching\\nmany of these GUCs. Or, if that's our position, why are there\\nper-GUC changes at all, rather than just redefining what the\\ncontext values mean? (That is, why not redefine USERSET and\\nSUSET as simply indicating the default ACL to be applied if there's\\nno entry in the catalog.)\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Sun, 06 Mar 2022 17:13:44 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 6, 2022, at 2:13 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> 1. If we need to change these two contrib modules, doesn't that imply\\n> a lot of changes forced on external modules as well? What are the\\n> security implications if somebody doesn't make such a change?\\n> \\n> 2. It looks to me like if someone installs the updated postgres_fdw.so,\\n> but doesn't run ALTER EXTENSION UPDATE, they are going to be left with a\\n> rather broken extension. This seems not good either, especially if it's\\n> multiplied by a boatload of third-party extensions requiring updates.\\n> \\n> So I think some more thought is needed to see if we can't avoid\\n> the need to touch extension modules in this patch. Or at least\\n> avoid the need for synchronized C-level and SQL-level updates,\\n> because that is going to create a lot of pain for end users.\\n> \\n> I'm also fairly distressed by the number of changes in guc.c,\\n> mainly because I fear that that means that pending patches that\\n> add GUC variables will be subtly incorrect/insecure if they're not\\n> updated to account for this. Frankly, I also reject the apparent\\n> position that we need to support forbidding users from touching\\n> many of these GUCs. Or, if that's our position, why are there\\n> per-GUC changes at all, rather than just redefining what the\\n> context values mean? (That is, why not redefine USERSET and\\n> SUSET as simply indicating the default ACL to be applied if there's\\n> no entry in the catalog.)\\n\\nTo my knowledge, there is no mechanism to revoke an implicit privilege. You can revoke a privilege explicitly listed in an aclitem[], but only if the privilege is being tracked that way.\\n\\nUserset variables are implicitly settable by any user. There was a request, off-list as I recall, to make it possible to revoke the privilege to set variables such as \\\"work_mem\\\". To make that possible, but not change the default behavior vis-a-vis prior releases, I upgraded most userset variables to suset with a corresponding grant to public on the variable. Sites which wish to have a more restrictive policy on such variables can revoke that privilege from public and instead issue more restrictive grants. There were a few variables where such treatment didn't seem sensible, such as ones to do with client connections, and I left them alone. I didn't insist on a defense for why any particular setting needed to be revocable in order to apply this treatment. My assumption was that sites should be allowed to determine their own security policies per setting unless there is a technical difficulty for a given setting that would make it overly burdensome to implement.\\n\\nIt seemed more complete to do the same thing for contrib modules, so I did.\\n\\nIf a third-party module doesn't do the equivalent operation, they would simply continue with their userset variables working as before. I don't see a specific security concern with that, though I'd be happy to consider arguments to the contrary.\\n\\nI believe there are also some issues with this patch relating to pg_dump/pg_restore and to pg_upgrade, which EDB is working to address. We intend to repost something soon.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Sun, 6 Mar 2022 14:40:50 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n>> On Mar 6, 2022, at 2:13 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> ... Or, if that's our position, why are there\\n>> per-GUC changes at all, rather than just redefining what the\\n>> context values mean? (That is, why not redefine USERSET and\\n>> SUSET as simply indicating the default ACL to be applied if there's\\n>> no entry in the catalog.)\\n\\n> To my knowledge, there is no mechanism to revoke an implicit privilege. You can revoke a privilege explicitly listed in an aclitem[], but only if the privilege is being tracked that way.\\n\\nSo? What I'm suggesting is along the lines of\\n\\n(1) pg_setting_acl starts out empty, or at least mostly empty (maybe\\nthere are a few GUCs that need custom values).\\n\\n(2) If there's a pg_setting_acl entry for a GUC that's to be set,\\nwe apply it: either it grants the desired permission or it doesn't.\\n\\n(3) If there's no entry, then for a USERSET GUC we assume that the\\nentry would be like \\\"=s/postgres\\\", while for any other context value\\nwe assume the ACL grants nothing.\\n\\nI don't think this is materially different from what we do with\\npermissions on (say) functions. If you want to revoke the public\\nSET privilege on some USERSET variable, you instantiate the default\\nand then revoke. You end up with an empty ACL stored in pg_setting_acl,\\nand voila.\\n\\nIt'd likely be necessary to refuse to record a grant/revoke on\\nan unknown GUC, since if we don't know the GUC then we can't know\\nwhat the relevant default ACL ought to be. But I bet your existing\\npatch has some dubious behavior in that case too.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 06 Mar 2022 17:57:10 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 6, 2022, at 2:57 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> I don't think this is materially different from what we do with\\n> permissions on (say) functions. If you want to revoke the public\\n> SET privilege on some USERSET variable, you instantiate the default\\n> and then revoke. You end up with an empty ACL stored in pg_setting_acl,\\n> and voila.\\n\\nI assume you mean the implementation of REVOKE does this, not that the user needs to do both a grant and a revoke.\\n\\n> It'd likely be necessary to refuse to record a grant/revoke on\\n> an unknown GUC, since if we don't know the GUC then we can't know\\n> what the relevant default ACL ought to be. But I bet your existing\\n> patch has some dubious behavior in that case too.\\n\\nThe existing patch allows grants on unknown gucs, because it can't know what guc an upgrade script will introduce, and the grant statement may need to execute before the guc exists. That opens a window for granting privileges on non-existent gucs. That sounds bad, but I don't know of any actual harm that it does, beyond just being ugly.\\n\\nWith your proposal, it sounds like we could avoid that ugliness, so I'm inclined to at least try doing it as you propose.\\n\\nThanks for the suggestion!\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Sun, 6 Mar 2022 15:06:43 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> On Mar 6, 2022, at 2:57 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> I don't think this is materially different from what we do with\\n>> permissions on (say) functions. If you want to revoke the public\\n>> SET privilege on some USERSET variable, you instantiate the default\\n>> and then revoke. You end up with an empty ACL stored in pg_setting_acl,\\n>> and voila.\\n\\n> I assume you mean the implementation of REVOKE does this, not that the user needs to do both a grant and a revoke.\\n\\nRight. Again, look at what happens when you create a function and\\nthen revoke its default PUBLIC EXECUTE permission.\\n\\n>> It'd likely be necessary to refuse to record a grant/revoke on\\n>> an unknown GUC, since if we don't know the GUC then we can't know\\n>> what the relevant default ACL ought to be. But I bet your existing\\n>> patch has some dubious behavior in that case too.\\n\\n> The existing patch allows grants on unknown gucs, because it can't know what guc an upgrade script will introduce, and the grant statement may need to execute before the guc exists.\\n\\nYeah, that's the problematic case. It might mostly work to assume that\\nan unknown GUC has an empty default ACL. This could fail to retain the\\ndefault PUBLIC SET permission if it later turns out the GUC is USERSET\\n... but I suspect in most cases anybody who's messing with the permissions\\nwould've started out by revoking that anyway. We could make this\\ndefinitely work in pg_dump if we teach pg_dump to explicitly grant or\\nrevoke the PUBLIC SET permission anytime it's emitting anything for\\na GUC, even if it thinks that would be the default state anyway.\\nExtension scripts that want to modify permissions for their GUCs\\ncould follow that same principle to be sure.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 06 Mar 2022 18:27:43 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Mar 6, 2022, at 3:27 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n>> The existing patch allows grants on unknown gucs, because it can't know what guc an upgrade script will introduce, and the grant statement may need to execute before the guc exists.\\n> \\n> Yeah, that's the problematic case. It might mostly work to assume that\\n> an unknown GUC has an empty default ACL. This could fail to retain the\\n> default PUBLIC SET permission if it later turns out the GUC is USERSET\\n\\nOn further reflection, I concluded this isn't needed. No current extension, whether in-core or third party, expects to be able to create a new GUC and then grant or revoke permissions on it. They can already specify the guc context (PGC_USERS, etc). Introducing a feature that depends on the dubious assumption that unrecognized GUCs will turn out to be USERSET doesn't seem warranted.\\n\\nThe patch attributes all grants of setting privileges to the bootstrap superuser. Only superusers can grant or revoke privileges on settings, and all settings are implicitly owned by the bootstrap superuser because there is no explicit owner associated with settings. Consequently, select_best_grantor(some_superuser, ..., BOOTSTRAP_SUPERUSERID, ...) always chooses the bootstrap superuser. I don't see a problem with this, but wouldn't mind a second opinion. Some people might find it surprising when viewing the pg_setting_acl.setacl field.\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Tue, 15 Mar 2022 13:59:08 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 3/15/22 16:59, Mark Dilger wrote:\\n>> On Mar 6, 2022, at 3:27 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>\\n>> Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n>>> The existing patch allows grants on unknown gucs, because it can't know what guc an upgrade script will introduce, and the grant statement may need to execute before the guc exists.\\n>> Yeah, that's the problematic case. It might mostly work to assume that\\n>> an unknown GUC has an empty default ACL. This could fail to retain the\\n>> default PUBLIC SET permission if it later turns out the GUC is USERSET\\n> On further reflection, I concluded this isn't needed. No current extension, whether in-core or third party, expects to be able to create a new GUC and then grant or revoke permissions on it. They can already specify the guc context (PGC_USERS, etc). Introducing a feature that depends on the dubious assumption that unrecognized GUCs will turn out to be USERSET doesn't seem warranted.\\n\\nAgreed.\\n\\n\\n>\\n> The patch attributes all grants of setting privileges to the bootstrap superuser. Only superusers can grant or revoke privileges on settings, and all settings are implicitly owned by the bootstrap superuser because there is no explicit owner associated with settings. Consequently, select_best_grantor(some_superuser, ..., BOOTSTRAP_SUPERUSERID, ...) always chooses the bootstrap superuser. I don't see a problem with this, but wouldn't mind a second opinion. Some people might find it surprising when viewing the pg_setting_acl.setacl field.\\n\\n\\nI think it's OK as long as we document it. An alternative might be to\\ninvent a pseudo-superuser called, say, 'postgres_system', but that seems\\nlike overkill to solve what is in effect a cosmetic problem.\\n\\nGenerally I think this is now in fairly good shape, I've played with it\\nand it seems to do what I expect in every case, and the things I found\\nsurprising are gone.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 07:43:28 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew@dunslane.net> writes:\\n> Generally I think this is now in fairly good shape, I've played with it\\n> and it seems to do what I expect in every case, and the things I found\\n> surprising are gone.\\n\\nStepping back a bit ... do we really want to institutionalize the\\nterm \\\"setting\\\" for GUC variables? I realize that the view pg_settings\\nexists, but the documentation generally prefers the term \\\"configuration\\nparameters\\\". Where config.sgml uses \\\"setting\\\" as a noun, it's usually\\ntalking about a specific concrete value for a parameter, and you can\\nargue that the view's name comports with that meaning. But you can't\\nGRANT a parameter's current value.\\n\\nI don't have a better name to offer offhand --- I surely am not proposing\\nthat we change the syntax to be \\\"GRANT ... ON CONFIGURATION PARAMETER x\\\",\\nbecause that's way too wordy. But now is the time to bikeshed if we're\\ngonna bikeshed, or else admit that we're not interested in precise\\nvocabulary.\\n\\nI'm also fairly allergic to the way that this patch has decided to assign\\nmulti-word names to privilege types (ie SET VALUE, ALTER SYSTEM). There\\nis no existing precedent for that, and I think it's going to break\\nclient-side code that we don't need to break. It's not coincidental that\\nthis forces weird changes in rules about whitespace in the has_privilege\\nfunctions, for example; and if you think that isn't going to cause\\nproblems I think you are wrong. Perhaps we could just use \\\"SET\\\" and\\n\\\"ALTER\\\", or \\\"SET\\\" and \\\"SYSTEM\\\"?\\n\\nI also agree with the upthread criticism that this is abusing the\\nObjectPostAlterHook API unreasonably much. In particular, it's\\ntrying to use pg_setting_acl OIDs as identities for GUCs, which\\nthey are not, first because they're unstable and second because\\nmost GUCs won't even have an entry there. I therefore judge the\\nhook calls added to ExecSetVariableStmt and AlterSystemSetConfigFile\\nto be 100% useless, in fact probably counterproductive because they\\nintroduce a boatload of worries about whether the right things happen\\nif the hook errors out or does something guc.c isn't expecting.\\n\\nI suggest that what might be saner is to consider that the \\\"objects\\\"\\nthat the hook calls are concerned with are the pg_setting_acl entries,\\nnot the underlying GUCs, and thus that the hooks need be invoked only\\nwhen creating, destroying or altering those entries. If we do have\\na need for a hook editorializing on GUC value settings, that would\\nhave to be an independent API --- but I have heard no calls for\\nthe ability to have such a hook, and I don't think that this patch\\nis the place to introduce one.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 14:47:24 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 16, 2022, at 11:47 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Stepping back a bit ... do we really want to institutionalize the\\n> term \\\"setting\\\" for GUC variables? I realize that the view pg_settings\\n> exists, but the documentation generally prefers the term \\\"configuration\\n> parameters\\\". Where config.sgml uses \\\"setting\\\" as a noun, it's usually\\n> talking about a specific concrete value for a parameter, and you can\\n> argue that the view's name comports with that meaning. But you can't\\n> GRANT a parameter's current value.\\n> \\n> I don't have a better name to offer offhand --- I surely am not proposing\\n> that we change the syntax to be \\\"GRANT ... ON CONFIGURATION PARAMETER x\\\",\\n> because that's way too wordy. But now is the time to bikeshed if we're\\n> gonna bikeshed, or else admit that we're not interested in precise\\n> vocabulary.\\n\\nInformally, we often use \\\"GUC\\\" on this list, but that isn't used formally, leaving \\\"configuration parameter\\\" and \\\"setting\\\" as the two obvious choices. I preferred \\\"configuration parameter\\\" originally and was argued out of it. My take on \\\"setting\\\" was also that it more naturally refers to the choice of setting, not the thing being set, such that \\\"work_mem = 8192\\\" means the configuration parameter \\\"work_mem\\\" has the setting \\\"8192\\\". I'm happy to change the patch along these lines, but I vaguely recall it being Robert who liked the \\\"setting\\\" language. Robert? (Sorry if I misremember that...)\\n\\n> I'm also fairly allergic to the way that this patch has decided to assign\\n> multi-word names to privilege types (ie SET VALUE, ALTER SYSTEM). There\\n> is no existing precedent for that, and I think it's going to break\\n> client-side code that we don't need to break. It's not coincidental that\\n> this forces weird changes in rules about whitespace in the has_privilege\\n> functions, for example; and if you think that isn't going to cause\\n> problems I think you are wrong. Perhaps we could just use \\\"SET\\\" and\\n> \\\"ALTER\\\", or \\\"SET\\\" and \\\"SYSTEM\\\"?\\n\\nI chose those particular multi-word names to avoid needing to promote any keywords. That's been a while ago, and I don't recall exactly where the shift-reduce or reduce-reduce errors were coming from. I'll play with it to see what I can find.\\n\\n> I also agree with the upthread criticism that this is abusing the\\n> ObjectPostAlterHook API unreasonably much. In particular, it's\\n> trying to use pg_setting_acl OIDs as identities for GUCs, which\\n> they are not, first because they're unstable and second because\\n> most GUCs won't even have an entry there. I therefore judge the\\n> hook calls added to ExecSetVariableStmt and AlterSystemSetConfigFile\\n> to be 100% useless, in fact probably counterproductive because they\\n> introduce a boatload of worries about whether the right things happen\\n> if the hook errors out or does something guc.c isn't expecting.\\n\\nI think Joshua was planning to use these hooks for security purposes. The hooks are supposed to check whether the Oid is valid, and if not, still be able to make choices based on the other information. Joshua, any comment on this?\\n\\n> I suggest that what might be saner is to consider that the \\\"objects\\\"\\n> that the hook calls are concerned with are the pg_setting_acl entries,\\n> not the underlying GUCs, and thus that the hooks need be invoked only\\n> when creating, destroying or altering those entries.\\n\\nThat's certainly a thing we could do, but I got the impression that Joshua wanted to hook into SET, RESET, and ALTER SYSTEM SET, not merely into GRANT and REVOKE.\\n\\n> If we do have\\n> a need for a hook editorializing on GUC value settings, that would\\n> have to be an independent API --- but I have heard no calls for\\n> the ability to have such a hook, and I don't think that this patch\\n> is the place to introduce one.\\n\\nWell, the call for it was in this thread, but I'm ok with yanking out that part of the patch if it seems half-baked.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 11:59:06 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"<snip>\\n\\n> I suggest that what might be saner is to consider that the \\\"objects\\\"\\n> that the hook calls are concerned with are the pg_setting_acl entries,\\n> not the underlying GUCs, and thus that the hooks need be invoked only\\n> when creating, destroying or altering those entries. If we do have\\n> a need for a hook editorializing on GUC value settings, that would\\n> have to be an independent API --- but I have heard no calls for\\n> the ability to have such a hook, and I don't think that this patch\\n> is the place to introduce one.\\n\\nI requested it here:\\nhttps://www.postgresql.org/message-id/CAGB%2BVh5pVFAqw8YzeXy4xxmEt_4Hq_8pEUHdCQvv3mCjvC-S-w%40mail.gmail.com\\n\\nwith compromises here:\\nhttps://www.postgresql.org/message-id/CAGB%2BVh6wLJ3FKsno62fi54pfg0FDrZRWcpuuCJBkHcCj-G1ndw%40mail.gmail.com\\nhttps://www.postgresql.org/message-id/0A3D3CBA-6548-4C9E-9F46-59D5C51A1F31%40enterprisedb.com\\nhttps://www.postgresql.org/message-id/CAGB%2BVh65R5vKC4rEt7r2_pK3kMZd-VY0n99RJwcP8Bic7xvOxQ%40mail.gmail.com\\n\\nand the new version here:\\nhttps://www.postgresql.org/message-id/C8DF19D8-C15D-4C2D-91CA-391390F1E421%40enterprisedb.com\\n\\nwhich I wrote. a toy module to test and was satisfied with it, despite\\nthe limitations.\\n\\nIf we are adding a DAC Grant for a type of object it seems untenable\\nthat it would not come with analogous MAC capable hooks.\\n\\nThank you.\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 15:00:09 -0400\",\n \"msg_from\": \"Joshua Brindle <joshua.brindle@crunchydata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> On Mar 16, 2022, at 11:47 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> ... I therefore judge the\\n>> hook calls added to ExecSetVariableStmt and AlterSystemSetConfigFile\\n>> to be 100% useless, in fact probably counterproductive because they\\n>> introduce a boatload of worries about whether the right things happen\\n>> if the hook errors out or does something guc.c isn't expecting.\\n\\n> I think Joshua was planning to use these hooks for security purposes. The hooks are supposed to check whether the Oid is valid, and if not, still be able to make choices based on the other information. Joshua, any comment on this?\\n\\nIt's going to be hard to do anything useful in a hook that (a) does\\nnot know which GUC is being assigned to and (b) cannot do catalog\\naccesses for fear that we're not inside a transaction. (b), in\\nparticular, seems like a rather thorough API break; up to now\\nObjectPostAlter hooks could assume that catalog accesses are OK.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 15:06:04 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 16, 2022 at 3:06 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> > On Mar 16, 2022, at 11:47 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >> ... I therefore judge the\\n> >> hook calls added to ExecSetVariableStmt and AlterSystemSetConfigFile\\n> >> to be 100% useless, in fact probably counterproductive because they\\n> >> introduce a boatload of worries about whether the right things happen\\n> >> if the hook errors out or does something guc.c isn't expecting.\\n>\\n> > I think Joshua was planning to use these hooks for security purposes. The hooks are supposed to check whether the Oid is valid, and if not, still be able to make choices based on the other information. Joshua, any comment on this?\\n>\\n> It's going to be hard to do anything useful in a hook that (a) does\\n> not know which GUC is being assigned to and (b) cannot do catalog\\n> accesses for fear that we're not inside a transaction. (b), in\\n> particular, seems like a rather thorough API break; up to now\\n> ObjectPostAlter hooks could assume that catalog accesses are OK.\\n>\\n\\nCan you elaborate on this point? This is perhaps an area where I don't\\nknow the rules of the road, to test this hook I modified the set_user\\nextension, which normally uses the parse tree to figure out if someone\\nis trying to SET log_statement or ALTER SYSTEM log_statement and\\nreplaced it with:\\n\\n case OAT_POST_ALTER:\\n if (classId == SettingAclRelationId)\\n {\\n ObjectAddress object;\\n object.classId = SettingAclRelationId;\\n object.objectId = objectId;\\n object.objectSubId = 0;\\n if (strcmp(getObjectIdentity(&object),\\n\\\"log_statement\\\") == 0)\\n {\\n ereport(ERROR,\\n (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),\\n errmsg(\\\"Setting %s blocked\\\",\\ngetObjectIdentity(&object))));\\n }\\n }\\n\\nIs that inherently unsafe?\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 15:36:21 -0400\",\n \"msg_from\": \"Joshua Brindle <joshua.brindle@crunchydata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Joshua Brindle <joshua.brindle@crunchydata.com> writes:\\n> On Wed, Mar 16, 2022 at 3:06 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> It's going to be hard to do anything useful in a hook that (a) does\\n>> not know which GUC is being assigned to and (b) cannot do catalog\\n>> accesses for fear that we're not inside a transaction. (b), in\\n>> particular, seems like a rather thorough API break; up to now\\n>> ObjectPostAlter hooks could assume that catalog accesses are OK.\\n\\n> Can you elaborate on this point?\\n\\nHmm, I glossed over too many details there perhaps. I was thinking\\nabout the restrictions on GUC check_hooks, which can be run outside\\na transaction altogether. But that's not quite relevant here.\\nExecSetVariableStmt can assume it's inside a transaction, but what\\nit *can't* assume is that we've set a transaction snapshot as yet\\n(cf. PlannedStmtRequiresSnapshot). If we call an ObjectPostAlter hook\\nthere, and it does a catalog access, that's going to break things for\\nmodifications of GUCs that are supposed to be modifiable without\\nfreezing the transaction snapshot. So the net result is the same:\\ncatalog access not okay, at least not in general.\\n\\nBetween that and the fact that an OID-based API is largely useless here,\\nI don't think it's sane to try to use the existing ObjectPostAlter API.\\nMaybe there is a case for inventing a separate hook API that could pass\\nthe GUC name as a string, instead. I remain of the opinion that this\\npatch should not concern itself with that, though.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 15:58:57 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 3/16/22 14:47, Tom Lane wrote:\\n> Andrew Dunstan <andrew@dunslane.net> writes:\\n>> Generally I think this is now in fairly good shape, I've played with it\\n>> and it seems to do what I expect in every case, and the things I found\\n>> surprising are gone.\\n> Stepping back a bit ... do we really want to institutionalize the\\n> term \\\"setting\\\" for GUC variables? I realize that the view pg_settings\\n> exists, but the documentation generally prefers the term \\\"configuration\\n> parameters\\\". Where config.sgml uses \\\"setting\\\" as a noun, it's usually\\n> talking about a specific concrete value for a parameter, and you can\\n> argue that the view's name comports with that meaning. But you can't\\n> GRANT a parameter's current value.\\n>\\n> I don't have a better name to offer offhand --- I surely am not proposing\\n> that we change the syntax to be \\\"GRANT ... ON CONFIGURATION PARAMETER x\\\",\\n> because that's way too wordy. But now is the time to bikeshed if we're\\n> gonna bikeshed, or else admit that we're not interested in precise\\n> vocabulary.\\n>\\n> I'm also fairly allergic to the way that this patch has decided to assign\\n> multi-word names to privilege types (ie SET VALUE, ALTER SYSTEM). There\\n> is no existing precedent for that, and I think it's going to break\\n> client-side code that we don't need to break. It's not coincidental that\\n> this forces weird changes in rules about whitespace in the has_privilege\\n> functions, for example; and if you think that isn't going to cause\\n> problems I think you are wrong. Perhaps we could just use \\\"SET\\\" and\\n> \\\"ALTER\\\", or \\\"SET\\\" and \\\"SYSTEM\\\"?\\n\\n\\nThat's going to look weird, ISTM. This is less clear about what it's\\ngranting.\\n\\n GRANT ALTER ON SOMETHING shared_buffers TO myuser;\\n\\nIf you don't like that maybe ALTER_SYSTEM and SET_VALUE would work,\\nalthough mostly we have avoided things like that.\\n\\nHow about MODIFY instead of SET VALUE and CONFIGURE instead of ALTER SYSTEM?\\n\\nPersonally I don't have problem with the use of SETTING. I think the\\nmeaning is pretty plain in context and unlikely to produce any confusion.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 16:07:43 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew@dunslane.net> writes:\\n> On 3/16/22 14:47, Tom Lane wrote:\\n>> I'm also fairly allergic to the way that this patch has decided to assign\\n>> multi-word names to privilege types (ie SET VALUE, ALTER SYSTEM). There\\n>> is no existing precedent for that, and I think it's going to break\\n>> client-side code that we don't need to break. It's not coincidental that\\n>> this forces weird changes in rules about whitespace in the has_privilege\\n>> functions, for example; and if you think that isn't going to cause\\n>> problems I think you are wrong. Perhaps we could just use \\\"SET\\\" and\\n>> \\\"ALTER\\\", or \\\"SET\\\" and \\\"SYSTEM\\\"?\\n\\n> That's going to look weird, ISTM. This is less clear about what it's\\n> granting.\\n> GRANT ALTER ON SOMETHING shared_buffers TO myuser;\\n\\nTrue. I think \\\"GRANT SET\\\" is clear enough, and it fits with the custom of\\nusing the name of the SQL statement that the privilege allows you to\\ninvoke. (I gather from Mark's comments that Bison gave him problems with\\nthat, but maybe that can be dealt with.) But I concede that \\\"ALTER\\\" by\\nitself is pretty vague.\\n\\n> If you don't like that maybe ALTER_SYSTEM and SET_VALUE would work,\\n> although mostly we have avoided things like that.\\n> How about MODIFY instead of SET VALUE and CONFIGURE instead of ALTER SYSTEM?\\n\\nI thought about ALTER_SYSTEM too. It's not great but maybe the best we\\ncan do. Not sure that CONFIGURE is better.\\n\\n> Personally I don't have problem with the use of SETTING. I think the\\n> meaning is pretty plain in context and unlikely to produce any confusion.\\n\\nI'm just unhappy about the disconnect with the documentation. I wonder\\nif we could get away with s/configuration parameter/setting/g in the docs.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 16:53:02 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 3/16/22 16:53, Tom Lane wrote:\\n\\n>> Personally I don't have problem with the use of SETTING. I think the\\n>> meaning is pretty plain in context and unlikely to produce any confusion.\\n> I'm just unhappy about the disconnect with the documentation. I wonder\\n> if we could get away with s/configuration parameter/setting/g in the docs.\\n>\\n> \\t\\t\\t\\n\\n\\nI don't think we need to fix that here though. If you can live with\\nSETTING for now I will undertake to fix the docs post-feature-freeze if\\nnecessary.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 17:09:35 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"<snip>\\n\\n> I remain of the opinion that this\\n> patch should not concern itself with that, though.\\n\\nSo you are saying that people can add new object types to PG with DAC\\npermissions and not concern themselves with MAC capable hooks? Is that\\nan official PG community stance?\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 09:24:54 -0400\",\n \"msg_from\": \"Joshua Brindle <joshua.brindle@crunchydata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On 16.03.22 19:59, Mark Dilger wrote:\\n> Informally, we often use \\\"GUC\\\" on this list, but that isn't used formally, leaving \\\"configuration parameter\\\" and \\\"setting\\\" as the two obvious choices. I preferred \\\"configuration parameter\\\" originally and was argued out of it. My take on \\\"setting\\\" was also that it more naturally refers to the choice of setting, not the thing being set, such that \\\"work_mem = 8192\\\" means the configuration parameter \\\"work_mem\\\" has the setting \\\"8192\\\".\\n\\n\\\"The current setting of the work_mem parameter is 8192.\\\"\\n\\nI think something based on \\\"parameter\\\" is good. We also use that \\nlanguage in the protocol (e.g., ParameterStatus).\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 14:52:34 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On 16.03.22 19:47, Tom Lane wrote:\\n> I'm also fairly allergic to the way that this patch has decided to assign\\n> multi-word names to privilege types (ie SET VALUE, ALTER SYSTEM). There\\n> is no existing precedent for that, and I think it's going to break\\n> client-side code that we don't need to break. It's not coincidental that\\n> this forces weird changes in rules about whitespace in the has_privilege\\n> functions, for example; and if you think that isn't going to cause\\n> problems I think you are wrong. Perhaps we could just use \\\"SET\\\" and\\n> \\\"ALTER\\\", or \\\"SET\\\" and \\\"SYSTEM\\\"?\\n\\nI think Oracle and MS SQL Server have many multi-word privilege names. \\nSo users are quite used to that. And if we want to add more complex \\nprivileges, we might run out of sensible single words eventually. So I \\nwould not exclude this approach.\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 14:54:57 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> On 16.03.22 19:47, Tom Lane wrote:\\n>> ... Perhaps we could just use \\\"SET\\\" and\\n>> \\\"ALTER\\\", or \\\"SET\\\" and \\\"SYSTEM\\\"?\\n\\n> I think Oracle and MS SQL Server have many multi-word privilege names. \\n> So users are quite used to that. And if we want to add more complex \\n> privileges, we might run out of sensible single words eventually. So I \\n> would not exclude this approach.\\n\\nWell, I still say that \\\"SET\\\" is sufficient for the one privilege name\\n(unless we really can't make Bison handle that, which I doubt). But\\nI'm willing to yield on using \\\"ALTER SYSTEM\\\" for the other.\\n\\nIf we go with s/SETTING/PARAMETER/ as per your other message, then\\nthat would be adequately consistent with the docs I think. So it'd\\nbe\\n\\nGRANT { SET | ALTER SYSTEM } ON PARAMETER foo TO ...\\n\\nand the new catalog would be pg_parameter_acl, and so on.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 10:47:03 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 3/17/22 10:47, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>> On 16.03.22 19:47, Tom Lane wrote:\\n>>> ... Perhaps we could just use \\\"SET\\\" and\\n>>> \\\"ALTER\\\", or \\\"SET\\\" and \\\"SYSTEM\\\"?\\n>> I think Oracle and MS SQL Server have many multi-word privilege names. \\n>> So users are quite used to that. And if we want to add more complex \\n>> privileges, we might run out of sensible single words eventually. So I \\n>> would not exclude this approach.\\n> Well, I still say that \\\"SET\\\" is sufficient for the one privilege name\\n> (unless we really can't make Bison handle that, which I doubt). But\\n> I'm willing to yield on using \\\"ALTER SYSTEM\\\" for the other.\\n>\\n> If we go with s/SETTING/PARAMETER/ as per your other message, then\\n> that would be adequately consistent with the docs I think. So it'd\\n> be\\n>\\n> GRANT { SET | ALTER SYSTEM } ON PARAMETER foo TO ...\\n>\\n> and the new catalog would be pg_parameter_acl, and so on.\\n>\\n> \\t\\t\\t\\n\\n\\n\\nThe upside of this is that it avoids the inelegant\\n\\n\\n GRANT SET ON SETTING ...\\n\\n\\nBut I was just looking again at the grammar, and the only reason we need\\nthis keyword at all AFAICS is to disambiguate ALL [PRIVILEGES] cases. \\nIf we abandoned that for this form of GRANT/REVOKE I think we could\\nprobably get away with\\n\\n\\n GRANT { SET | ALTER SYSTEM } ON setting_name ...\\n\\n\\nI haven't tried it, so I could be all wrong.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 11:41:38 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 17, 2022 at 9:25 AM Joshua Brindle\\n<joshua.brindle@crunchydata.com> wrote:\\n> <snip>\\n>\\n> > I remain of the opinion that this\\n> > patch should not concern itself with that, though.\\n>\\n> So you are saying that people can add new object types to PG with DAC\\n> permissions and not concern themselves with MAC capable hooks? Is that\\n> an official PG community stance?\\n\\nI don't know that the community has an official position on that\\ntopic, but I do not think it's reasonable to expect everyone who\\ntinkers with MAC permissions to try to make a corresponding equivalent\\nfor DAC. The number of people using PostgreSQL with DAC is relatively\\nsmall, and the topic is extremely complicated, and a lot of hackers\\ndon't really understand it well enough to be sure that whatever they\\nmight do is right. I think it's reasonable to expect people who\\nunderstand DAC and care about it to put some energy into the topic,\\nand not just in terms of telling other people how they have to write\\ntheir patches.\\n\\nI *don't* think it's appropriate for a patch that touches MAC to\\ndeliberately sabotage the existing support we have for DAC or to just\\nignore it where the right thing to do is obvious. But maintaining a\\nmillion lines of code is a lot of work, and I can't think of any\\nreason why the burden of maintaining relatively little-used features\\nshould fall entirely on people who don't care about them.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 12:04:42 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 16, 2022 at 2:47 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Stepping back a bit ... do we really want to institutionalize the\\n> term \\\"setting\\\" for GUC variables? I realize that the view pg_settings\\n> exists, but the documentation generally prefers the term \\\"configuration\\n> parameters\\\". Where config.sgml uses \\\"setting\\\" as a noun, it's usually\\n> talking about a specific concrete value for a parameter, and you can\\n> argue that the view's name comports with that meaning. But you can't\\n> GRANT a parameter's current value.\\n\\nI agree that the lack of a good user-friendly term for GUCs is a real\\nproblem. Here at EDB I've observed even relatively non-technical\\npeople using that term, which appears nowhere in the documentation and\\nis utterly unintelligible to a typical end-user. Somebody gets on the\\nphone and tells the customer that they need to set a GUC and the\\ncustomer is like \\\"what's a guck?\\\" except that they probably don't\\nactually ask that question but are just confused and fail to\\nunderstand that a postgresql.conf change is being proposed. I hate it.\\nIt sucks.\\n\\nI have sort of been trying to promote the use of the word \\\"setting\\\"\\nand use it in my own writing, especially to end-users. That is\\ndefinitely more intelligible to random users, but it's admittedly also\\nawkward. \\\"Set a setting\\\" just sounds redundant. But \\\"set a\\nconfiguration variable\\\" sounds wordy, so I don't know.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 12:18:51 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 17, 2022, at 9:04 AM, Robert Haas <robertmhaas@gmail.com> wrote:\\n> \\n> not just in terms of telling other people how they have to write\\n> their patches.\\n\\n...\\n\\n> the burden of maintaining relatively little-used features\\n> should fall entirely on people who don't care about them.\\n\\nJoshua helped test the DAC portion of this patch, and answered a number of my questions on the topic, including in-person back in December. I take your point, Robert, on the general principle, but the archives should reflect that Joshua did contribute to this specific patch.\\n\\nJoshua, should we drop the DAC portion for postgres 15 and revisit the issue for postgres 16? I think it's getting late in the development cycle to attempt what Tom described upthread, and I'd hate to see the rest of this patch punted.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 09:19:24 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 17, 2022 at 12:19 PM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n> Joshua helped test the DAC portion of this patch, and answered a number of my questions on the topic, including in-person back in December. I take your point, Robert, on the general principle, but the archives should reflect that Joshua did contribute to this specific patch.\\n\\nI wasn't intending to say otherwise, and if the changes needed for DAC\\nhere are straightforward and don't make the patch significantly harder\\nto finish, then I would say Tom is wrong and we should just go ahead\\nand make them. But if they do, then I think it's perfectly fine to say\\nthat we're going to leave that alone and let someone with\\nsubject-matter expertise sort it out when they have time. We do that\\nall the time with other things, most notably MSVC builds, where we\\njust can't expect every hacker or even every committer to understand\\nexactly what's required to make it all work. I try my best to commit\\nthings that don't break it, but sometimes I do, and then Andrew helps\\nsort it out, because he understands it and I don't. I think DAC should\\nfall into that kind of category as well.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 12:29:11 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 17, 2022 at 12:04 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> On Thu, Mar 17, 2022 at 9:25 AM Joshua Brindle\\n> <joshua.brindle@crunchydata.com> wrote:\\n> > <snip>\\n> >\\n> > > I remain of the opinion that this\\n> > > patch should not concern itself with that, though.\\n> >\\n> > So you are saying that people can add new object types to PG with DAC\\n> > permissions and not concern themselves with MAC capable hooks? Is that\\n> > an official PG community stance?\\n>\\n> I don't know that the community has an official position on that\\n> topic, but I do not think it's reasonable to expect everyone who\\n> tinkers with MAC permissions to try to make a corresponding equivalent\\n> for DAC. The number of people using PostgreSQL with DAC is relatively\\n> small, and the topic is extremely complicated, and a lot of hackers\\n> don't really understand it well enough to be sure that whatever they\\n> might do is right. I think it's reasonable to expect people who\\n> understand DAC and care about it to put some energy into the topic,\\n> and not just in terms of telling other people how they have to write\\n> their patches.\\n>\\n> I *don't* think it's appropriate for a patch that touches MAC to\\n> deliberately sabotage the existing support we have for DAC or to just\\n> ignore it where the right thing to do is obvious. But maintaining a\\n> million lines of code is a lot of work, and I can't think of any\\n> reason why the burden of maintaining relatively little-used features\\n> should fall entirely on people who don't care about them.\\n\\nI agree with this view, outside of the mixup between MAC and DAC (DAC\\nis in-core, MAC is via hooks)\\n\\nSo there should be some committers or contributors that keep an eye\\nout and make sure new objects have appropriate hooks, and since an Oid\\nbased hook for GUCs is inappropriate, I hope this patch can be rolled\\ninto the contribution from Mark.\\n\\nThe attached patch adds a set of hooks for strings, and changes the\\nGUC patch from Mark to use it, I tested with this code:\\n\\nvoid set_user_object_access_str(ObjectAccessType access, Oid classId,\\nconst char *objName, int subId, void *arg)\\n{\\nif (next_object_access_hook_str)\\n{\\n(*next_object_access_hook_str)(access, classId, objName, subId, arg);\\n}\\nswitch (access)\\n{\\ncase OAT_POST_ALTER:\\nif (classId == SettingAclRelationId)\\n{\\nOid userid = GetUserId();\\nbool is_superuser = superuser_arg(userid);\\nchar *username = GETUSERNAMEFROMID(GetUserId());\\nconst char *setting = \\\"setting value\\\";\\nconst char *altering = \\\"altering system\\\";\\nconst char *unknown = \\\"unknown\\\";\\nconst char *action;\\nif (subId && ACL_SET_VALUE)\\naction = setting;\\nelse if (subId && ACL_ALTER_SYSTEM)\\naction = altering;\\nelse\\naction = unknown;\\n\\nelog(WARNING, \\\"%s is %s %s (%d)\\\", username, action, objName, is_superuser);\\nif (strcmp(objName, \\\"log_statement\\\") == 0)\\n{\\nereport(ERROR,\\n(errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),\\nerrmsg(\\\"Setting %s blocked\\\", objName)));\\n}\\n}\\ndefault:\\nbreak;\\n}\\n}\\n\\nstatic void\\nset_user_object_access (ObjectAccessType access, Oid classId, Oid\\nobjectId, int subId, void *arg)\\n{\\nif (next_object_access_hook)\\n{\\n(*next_object_access_hook)(access, classId, objectId, subId, arg);\\n}\\nswitch (access)\\n{\\ncase OAT_FUNCTION_EXECUTE:\\n{\\n/* Update the `set_config` Oid cache if necessary. */\\nset_user_cache_proc(InvalidOid);\\n\\n/* Now see if this function is blocked */\\nset_user_block_set_config(objectId);\\nbreak;\\n}\\n\\n/* fallthrough */\\n\\ncase OAT_POST_CREATE:\\n{\\nif (classId == ProcedureRelationId)\\n{\\nset_user_cache_proc(objectId);\\n}\\nbreak;\\n}\\ndefault:\\nbreak;\\n}\\n}\",\n \"msg_date\": \"Thu, 17 Mar 2022 12:29:53 -0400\",\n \"msg_from\": \"Joshua Brindle <joshua.brindle@crunchydata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 17, 2022 at 12:30 PM Joshua Brindle\\n<joshua.brindle@crunchydata.com> wrote:\\n> I agree with this view, outside of the mixup between MAC and DAC (DAC\\n> is in-core, MAC is via hooks)\\n\\nAn excellent point! Exactly why we need expert-level help with this stuff! :-)\\n\\n> So there should be some committers or contributors that keep an eye\\n> out and make sure new objects have appropriate hooks, and since an Oid\\n> based hook for GUCs is inappropriate, I hope this patch can be rolled\\n> into the contribution from Mark.\\n\\nI'm not going to take a position on this patch right this minute, but\\nI appreciate you providing it.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 12:34:25 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 17, 2022, at 9:29 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\\n> \\n> I hope this patch can be rolled\\n> into the contribution from Mark.\\n\\nWorking on it.... Thanks for the patch!\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 09:36:49 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 17, 2022 at 12:36 PM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n> > On Mar 17, 2022, at 9:29 AM, Joshua Brindle <joshua.brindle@crunchydata.com> wrote:\\n> >\\n> > I hope this patch can be rolled\\n> > into the contribution from Mark.\\n>\\n> Working on it.... Thanks for the patch!\\n\\nGreat, thanks.\\n\\nI missed one objectId reference (InvokeObjectDropHookStr), fixed\\nversion attached.\",\n \"msg_date\": \"Thu, 17 Mar 2022 14:19:48 -0400\",\n \"msg_from\": \"Joshua Brindle <joshua.brindle@crunchydata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Mar 17, 2022, at 8:41 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> \\n> If we abandoned that for this form of GRANT/REVOKE I think we could\\n> probably get away with\\n> \\n> \\n> GRANT { SET | ALTER SYSTEM } ON setting_name ...\\n> \\n> \\n> I haven't tried it, so I could be all wrong.\\n\\nVersion 12 of the patch uses SET and ALTER SYSTEM as the names of the privileges, and PARAMETER as the name of the thing on which the privilege is granted. The catalog table which tracks these grants is now named pg_parameter_acl, and various other parts of the patch have been adjusted to use a \\\"parameter\\\" based, rather than a \\\"setting\\\" based, naming scheme. One exception to this rule is the \\\"setacl\\\" column in pg_parameter_acl, which is much more compact than the \\\"parameteracl\\\" name would be, so that remains under the old name.\\n\\nThe tab completion for grants and revokes of \\\"set\\\" and \\\"alter system\\\" on parameters has been extended compared to the previously submitted patch. The tab completion for grants and revokes of other privilege types remains fairly thin, as I resisted the urge to fix that generally, given that it is unrelated to this patch. However, in making tab completion for grant/revoke on parameters work, this patch incidentally improves the coverage for other privilege types, in so far as the prefixes for those commands match the prefixes of these new commands. \\n\\nVersion 12 also introduces \\\\dcp (pneumonic, \\\"Describe Configuration Parameter\\\") for listing parameters, with \\\\dcp+ also showing the acl, like:\\n\\nmark.dilger=# \\\\dcp+ %client%\\n List of configuration parameters\\n Parameter | Context | Setting | Unit | Access privileges \\n----------------------------------+---------+---------+------+--------------------------------\\n client_connection_check_interval | user | 0 | ms | \\n client_encoding | user | UTF8 | | \\\"mark.dilger\\\"=sA/\\\"mark.dilger\\\"\\n client_min_messages | user | notice | | \\\"mark.dilger\\\"=sA/\\\"mark.dilger\\\"+\\n | | | | =s/\\\"mark.dilger\\\" +\\n | | | | alice=A*/\\\"mark.dilger\\\"\\n(3 rows)\\n\\nThe \\\"Access privileges\\\" column is blank for parameters which have neither been granted nor revoked. In the example above, that's true for client_connection_check_interval.\\n\\n\\\\dcp[+] only shows \\\"user\\\" and \\\"superuser\\\" parameters:\\n\\nmark.dilger=# \\\\dcp %wal%\\n List of configuration parameters\\n Parameter | Context \\n--------------------------+-----------\\n track_wal_io_timing | superuser\\n wal_compression | superuser\\n wal_consistency_checking | superuser\\n wal_init_zero | superuser\\n wal_recycle | superuser\\n wal_sender_timeout | user\\n wal_skip_threshold | user\\n(7 rows)\\n\\nwhereas \\\\dcpS[+] also shows all parameters (so, also \\\"sighup\\\", \\\"backend\\\", \\\"superuser-backend\\\", \\\"postmaster\\\", and \\\"internal\\\".)\\n\\nmark.dilger=# \\\\dcpS %wal%\\n List of configuration parameters\\n Parameter | Context \\n-------------------------------+------------\\n max_slot_wal_keep_size | sighup\\n max_wal_senders | postmaster\\n max_wal_size | sighup\\n min_wal_size | sighup\\n track_wal_io_timing | superuser\\n wal_block_size | internal\\n wal_buffers | postmaster\\n wal_compression | superuser\\n wal_consistency_checking | superuser\\n wal_init_zero | superuser\\n wal_keep_size | sighup\\n wal_level | postmaster\\n wal_log_hints | postmaster\\n wal_receiver_create_temp_slot | sighup\\n wal_receiver_status_interval | sighup\\n wal_receiver_timeout | sighup\\n wal_recycle | superuser\\n wal_retrieve_retry_interval | sighup\\n wal_segment_size | internal\\n wal_sender_timeout | user\\n wal_skip_threshold | user\\n wal_sync_method | sighup\\n wal_writer_delay | sighup\\n wal_writer_flush_after | sighup\\n(24 rows)\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Thu, 24 Mar 2022 09:49:24 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 3/24/22 12:49, Mark Dilger wrote:\\n>\\n>> On Mar 17, 2022, at 8:41 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>\\n>> If we abandoned that for this form of GRANT/REVOKE I think we could\\n>> probably get away with\\n>>\\n>>\\n>> GRANT { SET | ALTER SYSTEM } ON setting_name ...\\n>>\\n>>\\n>> I haven't tried it, so I could be all wrong.\\n> Version 12 of the patch uses SET and ALTER SYSTEM as the names of the privileges, and PARAMETER as the name of the thing on which the privilege is granted. The catalog table which tracks these grants is now named pg_parameter_acl, and various other parts of the patch have been adjusted to use a \\\"parameter\\\" based, rather than a \\\"setting\\\" based, naming scheme. One exception to this rule is the \\\"setacl\\\" column in pg_parameter_acl, which is much more compact than the \\\"parameteracl\\\" name would be, so that remains under the old name.\\n\\n\\nI can live with it I guess, but it seems perverse to me to have\\npg_settings but pg_paramater_acl effectively referring to the same set\\nof things. If we're going to do this perhaps we should create a\\npg_parameters view which is identical to pg_settings and deprecate\\npg_settings. I don;t want to hold up this patch, I think this can\\nprobably be managed as a follow up item.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Mar 2022 15:06:21 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 24, 2022, at 12:06 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> \\n> \\n> On 3/24/22 12:49, Mark Dilger wrote:\\n>> \\n>>> On Mar 17, 2022, at 8:41 AM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>> \\n>>> If we abandoned that for this form of GRANT/REVOKE I think we could\\n>>> probably get away with\\n>>> \\n>>> \\n>>> GRANT { SET | ALTER SYSTEM } ON setting_name ...\\n>>> \\n>>> \\n>>> I haven't tried it, so I could be all wrong.\\n>> Version 12 of the patch uses SET and ALTER SYSTEM as the names of the privileges, and PARAMETER as the name of the thing on which the privilege is granted. The catalog table which tracks these grants is now named pg_parameter_acl, and various other parts of the patch have been adjusted to use a \\\"parameter\\\" based, rather than a \\\"setting\\\" based, naming scheme. One exception to this rule is the \\\"setacl\\\" column in pg_parameter_acl, which is much more compact than the \\\"parameteracl\\\" name would be, so that remains under the old name.\\n> \\n> \\n> I can live with it I guess, but it seems perverse to me to have\\n> pg_settings but pg_paramater_acl effectively referring to the same set\\n> of things. If we're going to do this perhaps we should create a\\n> pg_parameters view which is identical to pg_settings and deprecate\\n> pg_settings. I don;t want to hold up this patch, I think this can\\n> probably be managed as a follow up item.\\n\\nRight, the version 12 patch was following Peter's and Tom's comments upthread:\\n\\n> On Mar 17, 2022, at 7:47 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Well, I still say that \\\"SET\\\" is sufficient for the one privilege name\\n> (unless we really can't make Bison handle that, which I doubt). But\\n> I'm willing to yield on using \\\"ALTER SYSTEM\\\" for the other.\\n> \\n> If we go with s/SETTING/PARAMETER/ as per your other message, then\\n> that would be adequately consistent with the docs I think. So it'd\\n> be\\n> \\n> GRANT { SET | ALTER SYSTEM } ON PARAMETER foo TO ...\\n> \\n> and the new catalog would be pg_parameter_acl, and so on.\\n\\nWe could debate that again, but it seems awfully late in the development cycle. I'd rather just get this committed, barring any objections?\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Mar 2022 12:50:42 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"I've started reviewing this patch in earnest, and almost immediately\\nhit a serious problem: these regression tests aren't even a little bit\\ncommittable. For one thing, they fail if you do \\\"make installcheck\\\"\\ntwice in a row. This seems to be because the first run leaves some\\ncruft behind in pg_parameter_acl that affects the results of\\nsubsequent runs. But the bigger problem is that it is ABSOLUTELY NOT\\nOKAY to test ALTER SYSTEM during an \\\"installcheck\\\" run. That would be\\ntrue even if you cleaned up all the settings by the end of the run, as\\nthe patch fails to do (and for extra demerit, it leaves a superuser\\nrole laying around). Even transient effects on the behavior of\\nsessions in other DBs aren't acceptable in installcheck mode, IMO.\\n\\nI think we probably have to trash the core-regression-tests part of\\nthe patch altogether and instead use a TAP test for whatever testing\\nwe want to do. It might be all right to test SET privileges without\\ntesting ALTER SYSTEM, but I'm not sure there's much point in that.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 14:31:47 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 28, 2022, at 11:31 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> I think we probably have to trash the core-regression-tests part of\\n> the patch altogether and instead use a TAP test for whatever testing\\n> we want to do. It might be all right to test SET privileges without\\n> testing ALTER SYSTEM, but I'm not sure there's much point in that.\\n\\nHow about putting them under src/test/modules/unsafe_tests ?\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 11:42:42 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> Version 12 also introduces \\\\dcp (pneumonic, \\\"Describe Configuration Parameter\\\") for listing parameters, with \\\\dcp+ also showing the acl, like:\\n\\nThe fact that that code is not dry behind the ears is painfully obvious.\\nIt's not documented in psql-ref, not tested anywhere AFAICS, and its\\nhandling of the pattern parameter is inconsistent with every other\\n\\\\d command. The wildcard character should be * not %. It only\\naccidentally fails to dump core if no pattern is given, too.\\n\\n> \\\\dcp[+] only shows \\\"user\\\" and \\\"superuser\\\" parameters:\\n\\nWhy make that restriction? Also, I find it astonishing that this doesn't\\nshow the GUC's value by default. The non-plus form of the command seems\\nuseless as it stands, or at least its intended use-case is so narrow\\nI can't see it. If we're to have it at all, it seems like it ought to\\nbe a reasonably useful shortcut for interrogating pg_settings. I'd\\nexpect the base set of columns to be name, value, and possibly unit;\\nthen add ACL with +. I'm not sure that GucContext belongs in this at all,\\nbut if it does, it's a + column.\\n\\nOn the whole perhaps this should be taken out again; it's a bit\\nlate in the cycle to be introducing new features, especially ones\\nas subject to bikeshedding as a \\\\d command is. My ideas about what\\ncolumns to show probably wouldn't match anyone else's ... and we\\nhaven't even gotten to whether \\\\dcp is an okay choice of name.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 15:11:15 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> On Mar 28, 2022, at 11:31 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> I think we probably have to trash the core-regression-tests part of\\n>> the patch altogether and instead use a TAP test for whatever testing\\n>> we want to do. It might be all right to test SET privileges without\\n>> testing ALTER SYSTEM, but I'm not sure there's much point in that.\\n\\n> How about putting them under src/test/modules/unsafe_tests ?\\n\\nAh, that could work; I'd forgotten about that subdirectory.\\n\\nIt'd still be a good idea if they didn't fail when run twice in a row.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 15:14:04 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Mar 28, 2022, at 12:11 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n>> Version 12 also introduces \\\\dcp (pneumonic, \\\"Describe Configuration Parameter\\\") for listing parameters, with \\\\dcp+ also showing the acl, like:\\n> \\n> The fact that that code is not dry behind the ears is painfully obvious.\\n\\n<snip>\\n\\n> On the whole perhaps this should be taken out again\\n\\nI added the \\\\dcp command after noticing that doc/src/sgml/ddl.sgml's \\\"Summary of Access Privileges\\\" table had an empty column for the \\\"psql Command\\\" column for the \\\"PARAMETER\\\" row. Maybe that's ok. I thought it would be better if I added a command for that.\\n\\nI don't feel too strongly about it, so I'm removing \\\\dcp from this patch with the intention of introducing it as a separate patch for the postgres 16 development cycle. For now, I'm just putting \\\"none\\\" in that table cell.\\n\\n\\nThis patch also moves the grant/revoke tests in guc_privs.sql into the unsafe_tests directory, fixes them up to be repeatable, and drops the superuser role that had been negligently left around in prior patches. The tests that had been added in privileges.sql are simply removed.\\n\\n\\n\\n \\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 28 Mar 2022 13:57:50 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"I just came across something odd in v12 that is still there in v13:\\nExecGrant_Parameter uses InvokeObjectPostAlterHook not\\nInvokeObjectPostAlterHookArgStr. This seems pretty inconsistent.\\nIs there a good argument for it?\\n\\n... or, for that matter, why is there any such call at all?\\nNo other GRANT/REVOKE operation calls such a hook.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 17:16:14 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 28, 2022, at 2:16 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> I just came across something odd in v12 that is still there in v13:\\n> ExecGrant_Parameter uses InvokeObjectPostAlterHook not\\n> InvokeObjectPostAlterHookArgStr. This seems pretty inconsistent.\\n> Is there a good argument for it?\\n> \\n\\nFor SET and ALTER SYSTEM, the target of the action may not have an entry in pg_parameter_acl, nor an assigned Oid anywhere, so the only consistent way to pass the argument to the hook is by name. For GRANT/REVOKE, the parameter must have an Oid, at least by the time the hook gets called. Upthread there was some discussion of a hook not being able to assume a snapshot and working transaction, and hence not being able to query the catalogs. I would think that in a GRANT or REVOKE that hasn't already errored, the hook would have a transaction and could look up whatever it likes? There is a CommandCounterIncrement() call issued in objectNamesToOids() for new parameters, so by the time the hook is running it should be able to see the parameter.\\n\\nAm I reasoning about this the wrong way?\\n\\n> ... or, for that matter, why is there any such call at all?\\n> No other GRANT/REVOKE operation calls such a hook.\\n\\nI think ALTER DEFAULT PRIVILEGES does, though that's not quite the same thing. I don't have a strong opinion on this. Joshua, what's your take?\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 14:42:46 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n>> On Mar 28, 2022, at 2:16 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> I just came across something odd in v12 that is still there in v13:\\n>> ExecGrant_Parameter uses InvokeObjectPostAlterHook not\\n>> InvokeObjectPostAlterHookArgStr. This seems pretty inconsistent.\\n>> Is there a good argument for it?\\n\\n> For SET and ALTER SYSTEM, the target of the action may not have an entry\\n> in pg_parameter_acl, nor an assigned Oid anywhere, so the only\\n> consistent way to pass the argument to the hook is by name. For\\n> GRANT/REVOKE, the parameter must have an Oid, at least by the time the\\n> hook gets called.\\n\\nYeah, I know it's *possible* to make this work. The question is why is\\nit good to do it like this rather than to use the string API, now that\\nwe have the latter. AFAICS this way just guarantees that the hook must\\ndo a catalog lookup in order to figure out what you're talking about.\\n\\nThe core point here is that the actual identity of a GUC is its name.\\nAny OID that may exist in pg_parameter_acl is just a nonce alias that\\nmeans nothing to anybody. Anyone who's trying to, say, enforce that\\nJoe Blow can't change shared_buffers is going to need to see the GUC\\nname. (I am, btw, busy doing a lot of renaming in the patch to try\\nto clarify that these OIDs are not identifiers for GUCs; imagining\\nthat they are just risks confusion.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 17:54:28 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 28, 2022, at 2:54 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Yeah, I know it's *possible* to make this work. The question is why is\\n> it good to do it like this rather than to use the string API, now that\\n> we have the latter. AFAICS this way just guarantees that the hook must\\n> do a catalog lookup in order to figure out what you're talking about.\\n\\nOk, thanks for clarifying. I took the *HookStr versions of the hooks to be an alternative to be used when no Oid was present, something of a last resort. I never thought much about using them under other circumstances.\\n\\n> The core point here is that the actual identity of a GUC is its name.\\n> Any OID that may exist in pg_parameter_acl is just a nonce alias that\\n> means nothing to anybody. Anyone who's trying to, say, enforce that\\n> Joe Blow can't change shared_buffers is going to need to see the GUC\\n> name. (I am, btw, busy doing a lot of renaming in the patch to try\\n> to clarify that these OIDs are not identifiers for GUCs; imagining\\n> that they are just risks confusion.)\\n\\nI was about to write another patch using the HookStr form, but if you are already editing, then I'll let you make the change. I don't see a problem with what you are proposing.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 15:04:12 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"I'm going to be mostly unavailable till Wednesday, but I'll leave\\nyou with another thing to chew on:\\n\\nregression=# create user joe;\\nCREATE ROLE\\nregression=# grant set on parameter plpgsql.extra_warnings to joe;\\nERROR: unrecognized configuration parameter \\\"plpgsql.extra_warnings\\\"\\n\\nThis is problematic, because once plpgsql is loaded it works:\\n\\nregression=# load 'plpgsql'; \\nLOAD\\nregression=# grant set on parameter plpgsql.extra_warnings to joe;\\nGRANT\\n\\nIf we then do\\n\\n$ pg_dumpall -g\\n\\nit falls over:\\npg_dumpall: error: query failed: ERROR: unrecognized configuration parameter \\\"plpgsql.extra_warnings\\\"\\n\\napparently because aclparameterdefault() is allergic to being asked about\\nunknown GUCs, and plpgsql is not loaded in pg_dumpall's session. But\\nif pg_dumpall hadn't failed, it'd produce a dump script containing that\\nsame command, which would fail at load time (because, again, plpgsql\\nisn't going to be loaded in the backend reading the restore script).\\n\\nThis is what I meant by saying that you can't just refuse to GRANT on\\nunknown GUCs. It makes custom GUCs into a time bomb for dump/restore.\\nAnd that means you need a strategy for dealing with the possibility\\nthat you don't know whether the GUC is USERSET or not. I think though\\nthat it might work to just assume that it isn't, in which case dumps\\non unrecognized GUCs that really are USERSET will end up issuing an\\nexplicit GRANT SET TO PUBLIC that we didn't actually need to do, but it\\nwon't hurt anything. (Testing that assertion would be a good thing\\nto do.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 18:31:02 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> I was about to write another patch using the HookStr form, but if you are already editing, then I'll let you make the change. I don't see a problem with what you are proposing.\\n\\nDon't sweat about that, I can easily rebase what I've done so far\\nover your updates.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 18:37:07 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Mar 28, 2022, at 3:31 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> This is what I meant by saying that you can't just refuse to GRANT on\\n> unknown GUCs. It makes custom GUCs into a time bomb for dump/restore.\\n> And that means you need a strategy for dealing with the possibility\\n> that you don't know whether the GUC is USERSET or not. I think though\\n> that it might work to just assume that it isn't, in which case dumps\\n> on unrecognized GUCs that really are USERSET will end up issuing an\\n> explicit GRANT SET TO PUBLIC that we didn't actually need to do, but it\\n> won't hurt anything. (Testing that assertion would be a good thing\\n> to do.)\\n\\nOk, I returned to the idea upthread for a solution to this problem. A grant or revoke on an unrecognized custom parameter will create a SUSET placeholder, which is not quite right in some cases. However, the installation scripts for modules have been updated to manually grant SET privilege on their custom USERSET parameters, which cleans up the problem, with one exception: if the user executes a \\\"revoke set on parameter some.such from public\\\" prior to loading the module which defines parameter some.such, that revoke won't be retained. That doesn't seem entirely wrong to me, since no privilege to set the parameter existed when the revoke was performed, but rather was granted along with the creation of the parameter, but it also doesn't seem entirely right. Maybe revoke commands (but not grant commands) should error on unrecognized custom parameters? I didn't implement that here, but can do so if you think that makes more sense than this new behavior.\\n \\nI changed add_placeholder_variable() to take a GucContext argument. It previously always used PGC_USERSET, which is what all pre-existing call sites now pass into it, but that seems a bit inappropriate where we're creating a placeholder that we intend to treat as a SUSET variable until such time as a module gets installed saying otherwise. Not changing add_placeholder_variable in this fashion seems to work just fine. I just didn't feel comfortable with doing it that way. But if you feel it generates needless code churn, I could be talked out of doing this.\\n\\nI also changed the patch to use the ...HookStr functions for parameters. I would really like a comment on this from Joshua, to be sure what I'm doing comports with what he wanted. In particular, I'm uncertain that simply passing the AclMode (in other words, the istmt->privileges field) for the grant/revoke to the hook is sufficient. For one, how does the hook want to distinguish grants from revokes? Do we want a bit for that? And what about distinguishing WITH GRANT OPTION? I think the hooks are usable right now, but they might be made better.\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Tue, 29 Mar 2022 21:00:36 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 29, 2022 at 9:00 PM Mark Dilger <mark.dilger@enterprisedb.com>\\nwrote:\\n\\n> A grant or revoke on an unrecognized custom parameter will create a SUSET\\n> placeholder,\\n\\n[...]\\n\\n> which cleans up the problem, with one exception: if the user executes a\\n> \\\"revoke set on parameter some.such from public\\\" prior to loading the module\\n> which defines parameter some.such, that revoke won't be retained. That\\n> doesn't seem entirely wrong to me, since no privilege to set the parameter\\n> existed when the revoke was performed, but rather was granted along with\\n> the creation of the parameter, but it also doesn't seem entirely right.\\n> Maybe revoke commands (but not grant commands) should error on unrecognized\\n> custom parameters?\\n\\n\\nThe only revoke role target that makes sense here is the default grant\\ngiven to public. Aside from that technicality the grant system is purely\\nadditive and so only the GRANTS end up retained so far as perpetual state\\nis concerned.\\n\\nFor a revoke that doesn't target public we should remove the corresponding\\nunrecognized setting grant if one is present. We don't generally/always\\nraise a notice if a revoke doesn't actually cause something to be ungranted\\n- though I'm partial to being so informed. I suppose we could distinguish\\nthe cases where the not-yet-loaded setting name is unrecognized by the\\nsystem from the one where it is recognized but the grant is actually on a\\ndifferent role (or a revoke entry from public for the setting name is\\npresent).\\n\\nFor a revoke of an unknown setting from public we should keep an entry\\nsomewhere that tells the system that the default grant to public for that\\nsetting has been revoked. Maybe there isn't the same timing concern here\\nas there is for GRANT, but if only for symmetry it seems like a good thing\\nto implement.\\n\\nI have the impression I'm missing something in what I wrote above but\\ncannot quite figure out what. In any case as a first pass at this the\\nbehavior described is kinda what I'm expecting.\\n\\nDavid J.\\n\\nP.S.\\n\\nSkimming the patch we are, to my agreement, not touching the ALTER DEFAULT\\nPRIVILEGES command to work with this feature. Should the omission be noted\\nexplicitly? At least in the commit message I would think. Though the\\nsentence in [1] \\\"Also, these default privilege settings can be overridden\\nusing the ALTER DEFAULT PRIVILEGES command.\\\" is rendered only partially\\ncorrect.\\n\\n[1] https://www.postgresql.org/docs/current/ddl-priv.html\\n\\nP.P.S.\\n\\n+ The default privileges for a <literal>user</literal> parameter allow\\n+ <literal>PUBLIC</literal> to <command>SET</command> and\\n+ <command>RESET</command> the assigned value. By default,\\n+ <literal>PUBLIC</literal> has no privileges on\\n+ <literal>postmaster</literal>, <literal>superuser-backend</literal>,\\n+ <literal>internal</literal>, <literal>backend</literal>,\\n+ <literal>sighup</literal>, and <literal>superuser</literal> parameters.\\n\\nCan we rephrase this to something like:\\n\\nBy default, PUBLIC has no privileges on parameters in the postmaster, ...,\\nand superuser contexts.\\n\\npg_settings.context exists and those are the values found there. My\\ninitial interpretation of the wording was the postmaster, etc..., were\\nthemselves parameters, not containers for many parameters.\\n\\nOn Tue, Mar 29, 2022 at 9:00 PM Mark Dilger <mark.dilger@enterprisedb.com> wrote: A grant or revoke on an unrecognized custom parameter will create a SUSET placeholder,[...] which cleans up the problem, with one exception: if the user executes a \\\"revoke set on parameter some.such from public\\\" prior to loading the module which defines parameter some.such, that revoke won't be retained. That doesn't seem entirely wrong to me, since no privilege to set the parameter existed when the revoke was performed, but rather was granted along with the creation of the parameter, but it also doesn't seem entirely right. Maybe revoke commands (but not grant commands) should error on unrecognized custom parameters?The only revoke role target that makes sense here is the default grant given to public. Aside from that technicality the grant system is purely additive and so only the GRANTS end up retained so far as perpetual state is concerned.For a revoke that doesn't target public we should remove the corresponding unrecognized setting grant if one is present. We don't generally/always raise a notice if a revoke doesn't actually cause something to be ungranted - though I'm partial to being so informed. I suppose we could distinguish the cases where the not-yet-loaded setting name is unrecognized by the system from the one where it is recognized but the grant is actually on a different role (or a revoke entry from public for the setting name is present).For a revoke of an unknown setting from public we should keep an entry somewhere that tells the system that the default grant to public for that setting has been revoked. Maybe there isn't the same timing concern here as there is for GRANT, but if only for symmetry it seems like a good thing to implement.I have the impression I'm missing something in what I wrote above but cannot quite figure out what. In any case as a first pass at this the behavior described is kinda what I'm expecting.David J.P.S.Skimming the patch we are, to my agreement, not touching the ALTER DEFAULT PRIVILEGES command to work with this feature. Should the omission be noted explicitly? At least in the commit message I would think. Though the sentence in [1] \\\"Also, these default privilege settings can be overridden using the ALTER DEFAULT PRIVILEGES command.\\\" is rendered only partially correct.[1] https://www.postgresql.org/docs/current/ddl-priv.htmlP.P.S.+ The default privileges for a <literal>user</literal> parameter allow+ <literal>PUBLIC</literal> to <command>SET</command> and+ <command>RESET</command> the assigned value. By default,+ <literal>PUBLIC</literal> has no privileges on+ <literal>postmaster</literal>, <literal>superuser-backend</literal>,+ <literal>internal</literal>, <literal>backend</literal>,+ <literal>sighup</literal>, and <literal>superuser</literal> parameters.Can we rephrase this to something like:By default, PUBLIC has no privileges on parameters in the postmaster, ..., and superuser contexts.pg_settings.context exists and those are the values found there. My initial interpretation of the wording was the postmaster, etc..., were themselves parameters, not containers for many parameters.\",\n \"msg_date\": \"Tue, 29 Mar 2022 22:27:44 -0700\",\n \"msg_from\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 30, 2022 at 12:00 AM Mark Dilger\\n<mark.dilger@enterprisedb.com> wrote:\\n>\\n> > On Mar 28, 2022, at 3:31 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >\\n> > This is what I meant by saying that you can't just refuse to GRANT on\\n> > unknown GUCs. It makes custom GUCs into a time bomb for dump/restore.\\n> > And that means you need a strategy for dealing with the possibility\\n> > that you don't know whether the GUC is USERSET or not. I think though\\n> > that it might work to just assume that it isn't, in which case dumps\\n> > on unrecognized GUCs that really are USERSET will end up issuing an\\n> > explicit GRANT SET TO PUBLIC that we didn't actually need to do, but it\\n> > won't hurt anything. (Testing that assertion would be a good thing\\n> > to do.)\\n>\\n> Ok, I returned to the idea upthread for a solution to this problem. A grant or revoke on an unrecognized custom parameter will create a SUSET placeholder, which is not quite right in some cases. However, the installation scripts for modules have been updated to manually grant SET privilege on their custom USERSET parameters, which cleans up the problem, with one exception: if the user executes a \\\"revoke set on parameter some.such from public\\\" prior to loading the module which defines parameter some.such, that revoke won't be retained. That doesn't seem entirely wrong to me, since no privilege to set the parameter existed when the revoke was performed, but rather was granted along with the creation of the parameter, but it also doesn't seem entirely right. Maybe revoke commands (but not grant commands) should error on unrecognized custom parameters? I didn't implement that here, but can do so if you think that makes more sense than this new behavior.\\n>\\n> I changed add_placeholder_variable() to take a GucContext argument. It previously always used PGC_USERSET, which is what all pre-existing call sites now pass into it, but that seems a bit inappropriate where we're creating a placeholder that we intend to treat as a SUSET variable until such time as a module gets installed saying otherwise. Not changing add_placeholder_variable in this fashion seems to work just fine. I just didn't feel comfortable with doing it that way. But if you feel it generates needless code churn, I could be talked out of doing this.\\n>\\n> I also changed the patch to use the ...HookStr functions for parameters. I would really like a comment on this from Joshua, to be sure what I'm doing comports with what he wanted. In particular, I'm uncertain that simply passing the AclMode (in other words, the istmt->privileges field) for the grant/revoke to the hook is sufficient. For one, how does the hook want to distinguish grants from revokes? Do we want a bit for that? And what about distinguishing WITH GRANT OPTION? I think the hooks are usable right now, but they might be made better.\\n\\nI had not even been thinking about hooking grant/revoke TBH. In\\nSELinux-type MAC systems we don't always try to control DAC permission\\nchanges, and when we do it's not granular, all permission changes just\\nboil down to setattr permission or something of that nature.\\n\\nThank you.\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 08:30:42 -0400\",\n \"msg_from\": \"Joshua Brindle <joshua.brindle@crunchydata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"After sleeping on it, I have a modest proposal for simplifying\\nthese issues. Consider this design:\\n\\n1. In the SET code path, we assume (without any catalog lookup)\\nthat USERSET GUCs can be set. Only for SUSET GUCs do we perform\\na permissions lookup. (ALTER SYSTEM does a lookup in both cases.)\\n\\n2. Given this, the default ACL for any GUC can be empty, greatly\\nsimplifying all these management issues. Superusers could do what\\nthey want anyway, so modeling an \\\"owner's default grant\\\" becomes\\nunnecessary.\\n\\nWhat this loses is the ability to revoke public SET permissions\\non USERSET GUCs. I claim that that is not so valuable as to\\njustify all the complication needed to deal with it. (If a GUC\\nseems to require some defenses, why is it USERSET?) Avoiding\\na permissions lookup in the default SET code path seems like\\na pretty important benefit, too. If we force that to happen\\nit's going to be a noticeable drag on functions with SET clauses.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 09:26:21 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 30, 2022, at 6:26 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Consider this design:\\n\\nIsn't that just the design I had implemented in v8 several months ago?\\n\\nSubject: [PATCH v8] Allow GRANT of SET and ALTER SYSTEM SET for gucs\\n\\nAllow granting of privilege to set or alter system set variables\\nwhich otherwise can only be managed by superusers. Each\\n(role,variable,privilege) triple is independently grantable, so a\\nuser may be granted privilege to SET but not to ALTER SYSTEM SET on\\na variable, or vice versa. The privilege to SET a userset variable\\nmay be granted, though doing so has no practical effect, since any\\nrole can set userset variables anyway. Worse, there is no way to\\nrevoke the privilege to SET a userset variable. To remedy that,\\nmost core userset variables have been changed to suset, with\\nexplicit grants to set the variable to public.\\n\\n\\nI don't think v9 ever got posted to the list, but v10 has:\\n\\nSubject: [PATCH v10] Allow grant and revoke of privileges on settings\\n\\nAllow grant and revoke of privileges to set or alter system set\\nconfiguration variables. Each (role,variable,privilege) triple can\\nbe independently granted or revoked, so a user may be granted\\nprivilege to SET but not to ALTER SYSTEM SET on a variable, or vice\\nversa. Privilege to SET a userset variable is implicitly granted to\\npublic, but may be revoked.\\n\\n\\nIf we want to backtrack to v8, that's fine. I can rebase that, port some of the other changes from v14 to it, and repost it as v15. We should review the conversation from December and January which included some arguments for allowing revokes of SET on USERSET from PUBLIC. I don't want to keep going around in circles on this.\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 06:59:48 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Mar 30, 2022, at 6:59 AM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> \\n> We should review the conversation from December and January which included some arguments for allowing revokes of SET on USERSET from PUBLIC. I don't want to keep going around in circles on this.\\n\\nHmm, I guess that conversation was mostly off-list at the PGConn in December. I made a reference to it upthread:\\n\\n> On Mar 6, 2022, at 2:40 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> \\n> Userset variables are implicitly settable by any user. There was a request, off-list as I recall, to make it possible to revoke the privilege to set variables such as \\\"work_mem\\\". To make that possible, but not change the default behavior vis-a-vis prior releases, I upgraded most userset variables to suset with a corresponding grant to public on the variable. Sites which wish to have a more restrictive policy on such variables can revoke that privilege from public and instead issue more restrictive grants. There were a few variables where such treatment didn't seem sensible, such as ones to do with client connections, and I left them alone. I didn't insist on a defense for why any particular setting needed to be revocable in order to apply this treatment. My assumption was that sites should be allowed to determine their own security policies per setting unless there is a technical difficulty for a given setting that would make it overly burdensome to implement.\\n\\nYour proposal to just punt on supporting revocation of set on userset from public seems fine. We could revisit that in the next development cycle if anyone really wants to defend it. In particular, I don't see that committing this feature without that part would create any additional backward compatibility problems when implementing that later.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 07:45:36 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 3/30/22 09:26, Tom Lane wrote:\\n> After sleeping on it, I have a modest proposal for simplifying\\n> these issues. Consider this design:\\n>\\n> 1. In the SET code path, we assume (without any catalog lookup)\\n> that USERSET GUCs can be set. Only for SUSET GUCs do we perform\\n> a permissions lookup. (ALTER SYSTEM does a lookup in both cases.)\\n>\\n> 2. Given this, the default ACL for any GUC can be empty, greatly\\n> simplifying all these management issues. Superusers could do what\\n> they want anyway, so modeling an \\\"owner's default grant\\\" becomes\\n> unnecessary.\\n>\\n> What this loses is the ability to revoke public SET permissions\\n> on USERSET GUCs. I claim that that is not so valuable as to\\n> justify all the complication needed to deal with it. (If a GUC\\n> seems to require some defenses, why is it USERSET?) Avoiding\\n> a permissions lookup in the default SET code path seems like\\n> a pretty important benefit, too. If we force that to happen\\n> it's going to be a noticeable drag on functions with SET clauses.\\n>\\n> \\t\\t\\t\\n\\n\\n\\nThe last point is telling, so +1\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 11:12:46 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 30, 2022 at 8:12 AM Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 3/30/22 09:26, Tom Lane wrote:\\n>\\n> >\\n> > What this loses is the ability to revoke public SET permissions\\n> > on USERSET GUCs. I claim that that is not so valuable as to\\n> > justify all the complication needed to deal with it.\\n\\n\\nAgreed, and in line with my thinking from last night. These default public\\nset grants are indeed the complication and I'm good with the status quo\\nwhere they are non-revocable.\\n\\nI'm finding it curious that we are choosing to document every (all 6)\\ncontext that doesn't have this default privilege instead of saying that\\nonly the user context variables are granted this default, and now\\nirrevocable, default set privilege. This is in addition to making sure we\\ndistinguish between parameter and context in my earlier email.\\n\\n\\n> > Avoiding\\n> > a permissions lookup in the default SET code path seems like\\n> > a pretty important benefit, too. If we force that to happen\\n> > it's going to be a noticeable drag on functions with SET clauses.\\n> >\\n> >\\n> The last point is telling, so +1\\n>\\n>\\nIndeed. +1\\n\\nDavid J.\\n\\nOn Wed, Mar 30, 2022 at 8:12 AM Andrew Dunstan <andrew@dunslane.net> wrote:\\nOn 3/30/22 09:26, Tom Lane wrote:\\n>\\n> What this loses is the ability to revoke public SET permissions\\n> on USERSET GUCs. I claim that that is not so valuable as to\\n> justify all the complication needed to deal with it.Agreed, and in line with my thinking from last night. These default public set grants are indeed the complication and I'm good with the status quo where they are non-revocable.I'm finding it curious that we are choosing to document every (all 6) context that doesn't have this default privilege instead of saying that only the user context variables are granted this default, and now irrevocable, default set privilege. This is in addition to making sure we distinguish between parameter and context in my earlier email. > Avoiding\\n> a permissions lookup in the default SET code path seems like\\n> a pretty important benefit, too. If we force that to happen\\n> it's going to be a noticeable drag on functions with SET clauses.\\n>\\n> \\nThe last point is telling, so +1 Indeed. +1 David J.\",\n \"msg_date\": \"Wed, 30 Mar 2022 08:44:53 -0700\",\n \"msg_from\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> Your proposal to just punt on supporting revocation of set on userset from public seems fine. We could revisit that in the next development cycle if anyone really wants to defend it. In particular, I don't see that committing this feature without that part would create any additional backward compatibility problems when implementing that later.\\n\\nYeah. Also, as you noted, we could mark some individual built-in\\nvariables as SUSET and add a default GRANT. I don't want to do that with\\na blunderbuss, but perhaps there's an argument to do it for specific\\ncases (search_path comes to mind, though the performance cost could be\\nsignificant, since I think setting that in function SET clauses is\\ncommon).\\n\\nFor now, though, saying that you can't restrict SET for USERSET variables\\nseems fine --- there's certainly no loss of capability compared to where\\nwe stand today. I'd prefer to get the feature committed in that form\\nand then look at whether we want to tighten things around the margins.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 11:45:53 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 30, 2022 at 8:46 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> I don't want to do that with\\n> a blunderbuss, but perhaps there's an argument to do it for specific\\n> cases (search_path comes to mind, though the performance cost could be\\n> significant, since I think setting that in function SET clauses is\\n> common).\\n>\\n\\n\\nI suspect it became considerably moreso when we fixed the search_path CVE\\nsince we basically told people that doing so, despite the possible\\nperformance hit, was the easiest solution to their immediate dump/restore\\nfailures. But ISTM that because that SET has a function invocation context\\nit could bypass any such check. Though maybe the DO command exposes a flaw\\nin that idea.\\nDavid J.\\n\\nOn Wed, Mar 30, 2022 at 8:46 AM Tom Lane <tgl@sss.pgh.pa.us> wrote: I don't want to do that with\\na blunderbuss, but perhaps there's an argument to do it for specific\\ncases (search_path comes to mind, though the performance cost could be\\nsignificant, since I think setting that in function SET clauses is\\ncommon). I suspect it became considerably moreso when we fixed the search_path CVE since we basically told people that doing so, despite the possible performance hit, was the easiest solution to their immediate dump/restore failures. But ISTM that because that SET has a function invocation context it could bypass any such check. Though maybe the DO command exposes a flaw in that idea.David J.\",\n \"msg_date\": \"Wed, 30 Mar 2022 08:51:05 -0700\",\n \"msg_from\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> If we want to backtrack to v8, that's fine. I can rebase that, port\\n> some of the other changes from v14 to it, and repost it as v15.\\n\\nAre you working on that? I've set aside time this week to hopefully\\nget this over the finish line, but I don't want to find out that\\nI've been duplicating your effort.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 04 Apr 2022 11:36:18 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Apr 4, 2022, at 8:36 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n>> If we want to backtrack to v8, that's fine. I can rebase that, port\\n>> some of the other changes from v14 to it, and repost it as v15.\\n> \\n> Are you working on that? I've set aside time this week to hopefully\\n> get this over the finish line, but I don't want to find out that\\n> I've been duplicating your effort.\\n\\nYes, I expect to be posting the latest in maybe an hour? I believe the latest patch (just reviewing, adjusting code comments, etc.) that I'm preparing to post has all the changes we've discussed, aside from your parameterId renaming.\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Apr 2022 08:47:46 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"> On Apr 4, 2022, at 8:47 AM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> \\n> posting the latest in maybe an hour\\n\\nHere it is:\\n\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 4 Apr 2022 09:31:41 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> [ v15 patch ]\\n\\nThanks. As I'm working through this, I'm kind of inclined to drop\\nthe has_parameter_privilege() variants that take an OID as object\\nidentifier. This gets back to the fact that I don't think\\npg_parameter_acl OIDs have any outside use; we wouldn't even have\\nthem except that we need a way to track their role dependencies\\nin pg_shdepend. AFAICS users will only ever be interested in\\nlooking up a GUC by name. Any objections?\\n\\nAnother thought here is that I see you're expending some code\\nto store the canonical name of a GUC in pg_parameter_acl, but\\nI think that's probably going too far. In particular, for the\\nlegacy mixed-case names like \\\"DateStyle\\\", what ends up in the\\ntable is the mixed-case name, which seems problematic. It would\\ndefinitely be problematic if an extension used such a name,\\nbecause we might or might not be aware of the idiosyncratic\\ncasing at the time a GRANT is issued. I'm thinking that we\\nreally want to avoid looking up custom GUCs at all during GRANT,\\nbecause that can't do anything except create hazards.\\n\\nSo I think that instead of what you've got here, we should\\n(1) apply the map_old_guc_names[] mapping, which is constant\\n (for any one PG release anyway)\\n(2) smash to lower case\\n(3) verify validity per valid_variable_name.\\n\\nThis also simplifies life on the lookup side, where it's sufficient\\nto do steps (1) and (2) before performing a catalog search.\\n\\nThoughts?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 04 Apr 2022 17:26:35 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\nOn 4/4/22 17:26, Tom Lane wrote:\\n> Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n>> [ v15 patch ]\\n> Thanks. As I'm working through this, I'm kind of inclined to drop\\n> the has_parameter_privilege() variants that take an OID as object\\n> identifier. This gets back to the fact that I don't think\\n> pg_parameter_acl OIDs have any outside use; we wouldn't even have\\n> them except that we need a way to track their role dependencies\\n> in pg_shdepend. AFAICS users will only ever be interested in\\n> looking up a GUC by name. Any objections?\\n\\n\\nNot from me\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Apr 2022 18:15:19 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Apr 4, 2022, at 2:26 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Thanks. As I'm working through this, I'm kind of inclined to drop\\n> the has_parameter_privilege() variants that take an OID as object\\n> identifier. This gets back to the fact that I don't think\\n> pg_parameter_acl OIDs have any outside use; we wouldn't even have\\n> them except that we need a way to track their role dependencies\\n> in pg_shdepend. AFAICS users will only ever be interested in\\n> looking up a GUC by name. Any objections?\\n\\nNone.\\n\\n> Another thought here is that I see you're expending some code\\n> to store the canonical name of a GUC in pg_parameter_acl, but\\n> I think that's probably going too far. In particular, for the\\n> legacy mixed-case names like \\\"DateStyle\\\", what ends up in the\\n> table is the mixed-case name, which seems problematic. It would\\n> definitely be problematic if an extension used such a name,\\n> because we might or might not be aware of the idiosyncratic\\n> casing at the time a GRANT is issued. I'm thinking that we\\n> really want to avoid looking up custom GUCs at all during GRANT,\\n> because that can't do anything except create hazards.\\n\\nYikes. It took a few tries to see what you mean. Yes, if the GRANT happens before the LOAD, that can have bad consequences. So I agree something should be changed.\\n\\n> So I think that instead of what you've got here, we should\\n> (1) apply the map_old_guc_names[] mapping, which is constant\\n> (for any one PG release anyway)\\n> (2) smash to lower case\\n> (3) verify validity per valid_variable_name.\\n> \\n> This also simplifies life on the lookup side, where it's sufficient\\n> to do steps (1) and (2) before performing a catalog search.\\n> \\n> Thoughts?\\n\\nThat sounds right. Do you already have something like that coded, or would you like me to post a patch?\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Apr 2022 17:01:00 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Mark Dilger <mark.dilger@enterprisedb.com> writes:\\n> On Apr 4, 2022, at 2:26 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> So I think that instead of what you've got here, we should\\n>> (1) apply the map_old_guc_names[] mapping, which is constant\\n>> (for any one PG release anyway)\\n>> (2) smash to lower case\\n>> (3) verify validity per valid_variable_name.\\n>> \\n>> This also simplifies life on the lookup side, where it's sufficient\\n>> to do steps (1) and (2) before performing a catalog search.\\n>> \\n>> Thoughts?\\n\\n> That sounds right. Do you already have something like that coded, or would you like me to post a patch?\\n\\nWrote it already, no need for you to do it.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 04 Apr 2022 20:12:35 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Apr 4, 2022, at 5:12 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Wrote it already, no need for you to do it.\\n\\nThanks!\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Apr 2022 17:16:23 -0700\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"Here's a v16, which I have editorialized on rather heavily.\\nI have not yet really looked at the docs, but I'm satisfied\\nwith the code and tests now.\\n\\nThe main non-cosmetic changes I made:\\n\\n* I got rid of the mechanisms that tried to make has_parameter_privilege()\\nact like USERSET GUCs have a forced public SET grant. I think that's\\nfundamentally unworkable for custom GUCs, and it's just going to create\\nconfusion and inconsistency. As the code stands, it'll report SET\\nprivilege if you're superuser or have been explicitly granted that\\nprivilege, but having that privilege only matters if the GUC is SUSET.\\nNeed to craft some documentation wording to explain that, but I'd rather\\nexplain that than say \\\"it might or might not tell you that SET privileges\\nexist\\\". In any case I think this is perfectly symmetrical with the\\nALTER SYSTEM privilege: you can grant that all you want, but it still\\nwon't let you set variables whose context value forbids it.\\n\\n* That meant we have only one default ACL setting, so I was able to\\nget rid of aclparameterdefault() altogether.\\n\\n* I took out the post-alter hook call in ExecGrant_Parameter. I'm\\nprepared to listen to arguments why it should be put back in, but\\nsaid arguments would have to explain why GRANT ON PARAMETER should\\ndo that when no other form of GRANT does. (Or, possibly, it'd be\\nsane for all the forms to call that hook; but that would be material\\nfor a different patch IMV.)\\n\\n* I took out the pg_parameter_privileges view, which was undocumented\\nanyway, and which I think people would not find useful. The fact that\\nit's missing entries for most GUCs and might contain bogus entries\\nfor nonexistent GUCs seems like a strong reason why it's not very\\nuseful in this form. (But see below.)\\n\\n* I thought the tab-completion stuff was completely excessive.\\nI got rid of the support for tab completion with both privileges\\nspelled out, because I think people are just going to write ALL\\ninstead. We don't have tab completion support for commands listing\\nmultiple privileges of other object types, and I don't see why\\nPARAMETER is the place to start. I also merged a bunch of duplicative\\nrules by relying on MatchAny where possible.\\n\\n\\nAside from documentation, there is one loose end that perhaps would\\nbe best done in a follow-up patch: I'm somewhat of the opinion that\\nwe ought to reinvent a \\\\dcp command for psql. I'd define it as being\\na compact version of the pg_settings view, showing the name, value,\\nand unit columns and probably not much else by default. With \\\"+\\\",\\nwe could add on the privileges if any, and perhaps the context since\\nyou need that to interpret the privileges. I think this'd be more\\nuseful than the pg_parameter_privileges view, and it'd be handier\\nthan pg_settings itself since you could do things like\\n\\t\\\\dcp autovac*\\nwith a lot less typing than it takes to get the same info\\nout of pg_settings directly.\\n\\nAnyway, I'm out of patience with this for today, and I'm throwing\\nit up for the cfbot to have a look at. I'll hit the docs tomorrow.\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Tue, 05 Apr 2022 17:50:02 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Anyway, I'm out of patience with this for today, and I'm throwing\\n> it up for the cfbot to have a look at. I'll hit the docs tomorrow.\\n\\n... and of course, the cfbot gave me the raspberry, because of a\\nmerge conflict in gram.y. Here's v17 rebased up to HEAD.\\n\\n\\t\\t\\tregards, tom lane\",\n \"msg_date\": \"Tue, 05 Apr 2022 18:52:18 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Here's v17 rebased up to HEAD.\\n\\nPushed after fooling around with the docs.\\n\\nI have a couple of followup ideas in mind (\\\\dcp and another one),\\nwhich I'll start separate threads about.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 06 Apr 2022 13:29:38 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Granting SET and ALTER SYSTE privileges for GUCs\"\n }\n]","isConversational":false}}},{"rowIdx":53094,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi.\n\nI noticed recently that the CREATE PUBLICATION docs page [1] does not\nhave any \"See Also\" reference to the CREATE SUBSCRIPTION docs page\n[2].\n\nPerhaps it is deliberate (??) but IMO it just seemed like a\nstrange/annoying omission that there was no easy way to jump both ways\nbetween those two pages.\n\nOTOH, the reverse is OK - the CREATE SUBSCRIPTION docs page *does*\nhave a \"See Also\" reference to CREATE PUBLICATION.\n\nHere is a patch:\n\n--- a/doc/src/sgml/ref/create_publication.sgml\n+++ b/doc/src/sgml/ref/create_publication.sgml\n@@ -333,6 +333,7 @@ CREATE PUBLICATION sales_publication FOR ALL\nTABLES IN SCHEMA marketing, sales;\n <simplelist type=\"inline\">\n <member><xref linkend=\"sql-alterpublication\"/></member>\n <member><xref linkend=\"sql-droppublication\"/></member>\n+ <member><xref linkend=\"sql-createsubscription\"/></member>\n </simplelist>\n </refsect1>\n </refentry>\n\n------\n[1] https://www.postgresql.org/docs/devel/sql-createpublication.html\n[2] https://www.postgresql.org/docs/devel/sql-createsubscription.html\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 16 Nov 2021 09:29:26 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":true,"msg_subject":"CREATE PUBLICATION should \"See Also\" CREATE SUBSCRIPTION"},{"msg_contents":"> On 15 Nov 2021, at 23:29, Peter Smith <smithpb2250@gmail.com> wrote:\n\n> I noticed recently that the CREATE PUBLICATION docs page [1] does not\n> have any \"See Also\" reference to the CREATE SUBSCRIPTION docs page\n> [2].\n\nNice catch.\n\n> <simplelist type=\"inline\">\n> <member><xref linkend=\"sql-alterpublication\"/></member>\n> <member><xref linkend=\"sql-droppublication\"/></member>\n> + <member><xref linkend=\"sql-createsubscription\"/></member>\n\nTo make it analogous with how CREATE/ALTER for SUBSCRIPTION and PUBLICATION\nreference each other, there should IMO be another xref to ALTER SUBSCRIPTION as\nwell.\n\nUnless there are objections, I'll apply that.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Tue, 16 Nov 2021 14:21:39 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: CREATE PUBLICATION should \"See Also\" CREATE SUBSCRIPTION"},{"msg_contents":"On Tue, Nov 16, 2021 at 10:21 PM Daniel Gustafsson <daniel@yesql.se> wrote:\n>\n> > On 15 Nov 2021, at 23:29, Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> > I noticed recently that the CREATE PUBLICATION docs page [1] does not\n> > have any \"See Also\" reference to the CREATE SUBSCRIPTION docs page\n> > [2].\n>\n> Nice catch.\n>\n> > <simplelist type=\"inline\">\n> > <member><xref linkend=\"sql-alterpublication\"/></member>\n> > <member><xref linkend=\"sql-droppublication\"/></member>\n> > + <member><xref linkend=\"sql-createsubscription\"/></member>\n>\n> To make it analogous with how CREATE/ALTER for SUBSCRIPTION and PUBLICATION\n> reference each other, there should IMO be another xref to ALTER SUBSCRIPTION as\n> well.\n\n+1\n\nRegards,\n\n-- \nMasahiko Sawada\nEDB: https://www.enterprisedb.com/\n\n\n","msg_date":"Wed, 17 Nov 2021 17:24:55 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":false,"msg_subject":"Re: CREATE PUBLICATION should \"See Also\" CREATE SUBSCRIPTION"},{"msg_contents":"> On 17 Nov 2021, at 09:24, Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> On Tue, Nov 16, 2021 at 10:21 PM Daniel Gustafsson <daniel@yesql.se> wrote:\n>> \n>>> On 15 Nov 2021, at 23:29, Peter Smith <smithpb2250@gmail.com> wrote:\n\n>>> <simplelist type=\"inline\">\n>>> <member><xref linkend=\"sql-alterpublication\"/></member>\n>>> <member><xref linkend=\"sql-droppublication\"/></member>\n>>> + <member><xref linkend=\"sql-createsubscription\"/></member>\n>> \n>> To make it analogous with how CREATE/ALTER for SUBSCRIPTION and PUBLICATION\n>> reference each other, there should IMO be another xref to ALTER SUBSCRIPTION as\n>> well.\n> \n> +1\n\nPushed, thanks for review!\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Wed, 17 Nov 2021 13:45:03 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: CREATE PUBLICATION should \"See Also\" CREATE SUBSCRIPTION"},{"msg_contents":"On Wed, Nov 17, 2021 at 11:45 PM Daniel Gustafsson <daniel@yesql.se> wrote:\n>\n> Pushed, thanks for review!\n>\n\nThanks!\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Thu, 18 Nov 2021 08:24:54 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":true,"msg_subject":"Re: CREATE PUBLICATION should \"See Also\" CREATE SUBSCRIPTION"}],"string":"[\n {\n \"msg_contents\": \"Hi.\\n\\nI noticed recently that the CREATE PUBLICATION docs page [1] does not\\nhave any \\\"See Also\\\" reference to the CREATE SUBSCRIPTION docs page\\n[2].\\n\\nPerhaps it is deliberate (??) but IMO it just seemed like a\\nstrange/annoying omission that there was no easy way to jump both ways\\nbetween those two pages.\\n\\nOTOH, the reverse is OK - the CREATE SUBSCRIPTION docs page *does*\\nhave a \\\"See Also\\\" reference to CREATE PUBLICATION.\\n\\nHere is a patch:\\n\\n--- a/doc/src/sgml/ref/create_publication.sgml\\n+++ b/doc/src/sgml/ref/create_publication.sgml\\n@@ -333,6 +333,7 @@ CREATE PUBLICATION sales_publication FOR ALL\\nTABLES IN SCHEMA marketing, sales;\\n <simplelist type=\\\"inline\\\">\\n <member><xref linkend=\\\"sql-alterpublication\\\"/></member>\\n <member><xref linkend=\\\"sql-droppublication\\\"/></member>\\n+ <member><xref linkend=\\\"sql-createsubscription\\\"/></member>\\n </simplelist>\\n </refsect1>\\n </refentry>\\n\\n------\\n[1] https://www.postgresql.org/docs/devel/sql-createpublication.html\\n[2] https://www.postgresql.org/docs/devel/sql-createsubscription.html\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 09:29:26 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"CREATE PUBLICATION should \\\"See Also\\\" CREATE SUBSCRIPTION\"\n },\n {\n \"msg_contents\": \"> On 15 Nov 2021, at 23:29, Peter Smith <smithpb2250@gmail.com> wrote:\\n\\n> I noticed recently that the CREATE PUBLICATION docs page [1] does not\\n> have any \\\"See Also\\\" reference to the CREATE SUBSCRIPTION docs page\\n> [2].\\n\\nNice catch.\\n\\n> <simplelist type=\\\"inline\\\">\\n> <member><xref linkend=\\\"sql-alterpublication\\\"/></member>\\n> <member><xref linkend=\\\"sql-droppublication\\\"/></member>\\n> + <member><xref linkend=\\\"sql-createsubscription\\\"/></member>\\n\\nTo make it analogous with how CREATE/ALTER for SUBSCRIPTION and PUBLICATION\\nreference each other, there should IMO be another xref to ALTER SUBSCRIPTION as\\nwell.\\n\\nUnless there are objections, I'll apply that.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 14:21:39 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: CREATE PUBLICATION should \\\"See Also\\\" CREATE SUBSCRIPTION\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 10:21 PM Daniel Gustafsson <daniel@yesql.se> wrote:\\n>\\n> > On 15 Nov 2021, at 23:29, Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> > I noticed recently that the CREATE PUBLICATION docs page [1] does not\\n> > have any \\\"See Also\\\" reference to the CREATE SUBSCRIPTION docs page\\n> > [2].\\n>\\n> Nice catch.\\n>\\n> > <simplelist type=\\\"inline\\\">\\n> > <member><xref linkend=\\\"sql-alterpublication\\\"/></member>\\n> > <member><xref linkend=\\\"sql-droppublication\\\"/></member>\\n> > + <member><xref linkend=\\\"sql-createsubscription\\\"/></member>\\n>\\n> To make it analogous with how CREATE/ALTER for SUBSCRIPTION and PUBLICATION\\n> reference each other, there should IMO be another xref to ALTER SUBSCRIPTION as\\n> well.\\n\\n+1\\n\\nRegards,\\n\\n-- \\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 17:24:55 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: CREATE PUBLICATION should \\\"See Also\\\" CREATE SUBSCRIPTION\"\n },\n {\n \"msg_contents\": \"> On 17 Nov 2021, at 09:24, Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> On Tue, Nov 16, 2021 at 10:21 PM Daniel Gustafsson <daniel@yesql.se> wrote:\\n>> \\n>>> On 15 Nov 2021, at 23:29, Peter Smith <smithpb2250@gmail.com> wrote:\\n\\n>>> <simplelist type=\\\"inline\\\">\\n>>> <member><xref linkend=\\\"sql-alterpublication\\\"/></member>\\n>>> <member><xref linkend=\\\"sql-droppublication\\\"/></member>\\n>>> + <member><xref linkend=\\\"sql-createsubscription\\\"/></member>\\n>> \\n>> To make it analogous with how CREATE/ALTER for SUBSCRIPTION and PUBLICATION\\n>> reference each other, there should IMO be another xref to ALTER SUBSCRIPTION as\\n>> well.\\n> \\n> +1\\n\\nPushed, thanks for review!\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 13:45:03 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: CREATE PUBLICATION should \\\"See Also\\\" CREATE SUBSCRIPTION\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 11:45 PM Daniel Gustafsson <daniel@yesql.se> wrote:\\n>\\n> Pushed, thanks for review!\\n>\\n\\nThanks!\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 08:24:54 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: CREATE PUBLICATION should \\\"See Also\\\" CREATE SUBSCRIPTION\"\n }\n]","isConversational":false}}},{"rowIdx":53095,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hackers,\n\nWhile working on a fix for dangling references to dropped roles in the pg_auth_members.grantor field, I happened to notice we entirely lack regression test coverage of the REVOKE ADMIN OPTION FOR form of the RevokeRoleStmt. I am unaware of any bugs in the current implementation, but future work on roles may benefit if we close the testing gap.\n\nFor consideration:\n\n\n\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Mon, 15 Nov 2021 15:58:34 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Add regression coverage for REVOKE ADMIN OPTION"},{"msg_contents":"> On 16 Nov 2021, at 00:58, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n\n> While working on a fix for dangling references to dropped roles in the pg_auth_members.grantor field, I happened to notice we entirely lack regression test coverage of the REVOKE ADMIN OPTION FOR form of the RevokeRoleStmt. I am unaware of any bugs in the current implementation, but future work on roles may benefit if we close the testing gap.\n\nLGTM. Reading this I realized that the GRANTED BY keyword for RevokeRoleStmt\nisn't working as documented, it's not checking the role at all. I've sent a\ndiff for that with tests on the relevant thread, but I think it would be a good\nto get this in too to boost coverage.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Tue, 16 Nov 2021 15:31:18 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Add regression coverage for REVOKE ADMIN OPTION"},{"msg_contents":"\n\n> On Nov 16, 2021, at 6:31 AM, Daniel Gustafsson <daniel@yesql.se> wrote:\n> \n>> On 16 Nov 2021, at 00:58, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n> \n>> While working on a fix for dangling references to dropped roles in the pg_auth_members.grantor field, I happened to notice we entirely lack regression test coverage of the REVOKE ADMIN OPTION FOR form of the RevokeRoleStmt. I am unaware of any bugs in the current implementation, but future work on roles may benefit if we close the testing gap.\n> \n> LGTM. Reading this I realized that the GRANTED BY keyword for RevokeRoleStmt\n> isn't working as documented, it's not checking the role at all. I've sent a\n> diff for that with tests on the relevant thread, but I think it would be a good\n> to get this in too to boost coverage.\n\nThanks for the review!\n\n—\nMark Dilger\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n\n\n\n","msg_date":"Tue, 16 Nov 2021 06:32:15 -0800","msg_from":"Mark Dilger <mark.dilger@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Add regression coverage for REVOKE ADMIN OPTION"},{"msg_contents":"> On 16 Nov 2021, at 15:32, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\n\n> Thanks for the review!\n\nPushed to master along with the bugfix it helped uncover, thanks!\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Fri, 26 Nov 2021 14:17:54 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Add regression coverage for REVOKE ADMIN OPTION"}],"string":"[\n {\n \"msg_contents\": \"Hackers,\\n\\nWhile working on a fix for dangling references to dropped roles in the pg_auth_members.grantor field, I happened to notice we entirely lack regression test coverage of the REVOKE ADMIN OPTION FOR form of the RevokeRoleStmt. I am unaware of any bugs in the current implementation, but future work on roles may benefit if we close the testing gap.\\n\\nFor consideration:\\n\\n\\n\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 15 Nov 2021 15:58:34 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Add regression coverage for REVOKE ADMIN OPTION\"\n },\n {\n \"msg_contents\": \"> On 16 Nov 2021, at 00:58, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n\\n> While working on a fix for dangling references to dropped roles in the pg_auth_members.grantor field, I happened to notice we entirely lack regression test coverage of the REVOKE ADMIN OPTION FOR form of the RevokeRoleStmt. I am unaware of any bugs in the current implementation, but future work on roles may benefit if we close the testing gap.\\n\\nLGTM. Reading this I realized that the GRANTED BY keyword for RevokeRoleStmt\\nisn't working as documented, it's not checking the role at all. I've sent a\\ndiff for that with tests on the relevant thread, but I think it would be a good\\nto get this in too to boost coverage.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 15:31:18 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add regression coverage for REVOKE ADMIN OPTION\"\n },\n {\n \"msg_contents\": \"\\n\\n> On Nov 16, 2021, at 6:31 AM, Daniel Gustafsson <daniel@yesql.se> wrote:\\n> \\n>> On 16 Nov 2021, at 00:58, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n> \\n>> While working on a fix for dangling references to dropped roles in the pg_auth_members.grantor field, I happened to notice we entirely lack regression test coverage of the REVOKE ADMIN OPTION FOR form of the RevokeRoleStmt. I am unaware of any bugs in the current implementation, but future work on roles may benefit if we close the testing gap.\\n> \\n> LGTM. Reading this I realized that the GRANTED BY keyword for RevokeRoleStmt\\n> isn't working as documented, it's not checking the role at all. I've sent a\\n> diff for that with tests on the relevant thread, but I think it would be a good\\n> to get this in too to boost coverage.\\n\\nThanks for the review!\\n\\n—\\nMark Dilger\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 06:32:15 -0800\",\n \"msg_from\": \"Mark Dilger <mark.dilger@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add regression coverage for REVOKE ADMIN OPTION\"\n },\n {\n \"msg_contents\": \"> On 16 Nov 2021, at 15:32, Mark Dilger <mark.dilger@enterprisedb.com> wrote:\\n\\n> Thanks for the review!\\n\\nPushed to master along with the bugfix it helped uncover, thanks!\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 26 Nov 2021 14:17:54 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add regression coverage for REVOKE ADMIN OPTION\"\n }\n]","isConversational":false}}},{"rowIdx":53096,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nThe attached patch moves the \"LastRec\" variable declaration inside\nStartupXLOG() where it is supposed to be.\n\n-- \nRegards,\nAmul Sul\nEDB: http://www.enterprisedb.com","msg_date":"Tue, 16 Nov 2021 11:17:27 +0530","msg_from":"Amul Sul <sulamul@gmail.com>","msg_from_op":true,"msg_subject":"Unnecessary global variable declared in xlog.c"},{"msg_contents":"On Tue, Nov 16, 2021 at 11:17:27AM +0530, Amul Sul wrote:\n> The attached patch moves the \"LastRec\" variable declaration inside\n> StartupXLOG() where it is supposed to be.\n\nYes, reading through the code, your suggestion makes sense. I'll\ndouble-check and apply if there are no objections.\n--\nMichael","msg_date":"Tue, 16 Nov 2021 15:12:38 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Unnecessary global variable declared in xlog.c"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Tue, Nov 16, 2021 at 11:17:27AM +0530, Amul Sul wrote:\n>> The attached patch moves the \"LastRec\" variable declaration inside\n>> StartupXLOG() where it is supposed to be.\n\n> Yes, reading through the code, your suggestion makes sense. I'll\n> double-check and apply if there are no objections.\n\nI think LastRec was originally referenced by multiple functions\nin xlog.c. But it does look like it could be a local now.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 16 Nov 2021 02:08:54 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Unnecessary global variable declared in xlog.c"},{"msg_contents":"On Tue, Nov 16, 2021 at 02:08:54AM -0500, Tom Lane wrote:\n> I think LastRec was originally referenced by multiple functions\n> in xlog.c. But it does look like it could be a local now.\n\nThanks for double-checking, applied this one as of f975fc3.\n--\nMichael","msg_date":"Wed, 17 Nov 2021 11:06:04 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Unnecessary global variable declared in xlog.c"},{"msg_contents":"On Wed, Nov 17, 2021 at 7:36 AM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Tue, Nov 16, 2021 at 02:08:54AM -0500, Tom Lane wrote:\n> > I think LastRec was originally referenced by multiple functions\n> > in xlog.c. But it does look like it could be a local now.\n>\n> Thanks for double-checking, applied this one as of f975fc3.\n\nThank you, Michael.\n\nRegards,\nAmul\n\n\n","msg_date":"Wed, 17 Nov 2021 09:18:35 +0530","msg_from":"Amul Sul <sulamul@gmail.com>","msg_from_op":true,"msg_subject":"Re: Unnecessary global variable declared in xlog.c"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nThe attached patch moves the \\\"LastRec\\\" variable declaration inside\\nStartupXLOG() where it is supposed to be.\\n\\n-- \\nRegards,\\nAmul Sul\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Tue, 16 Nov 2021 11:17:27 +0530\",\n \"msg_from\": \"Amul Sul <sulamul@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Unnecessary global variable declared in xlog.c\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 11:17:27AM +0530, Amul Sul wrote:\\n> The attached patch moves the \\\"LastRec\\\" variable declaration inside\\n> StartupXLOG() where it is supposed to be.\\n\\nYes, reading through the code, your suggestion makes sense. I'll\\ndouble-check and apply if there are no objections.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 16 Nov 2021 15:12:38 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Unnecessary global variable declared in xlog.c\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Tue, Nov 16, 2021 at 11:17:27AM +0530, Amul Sul wrote:\\n>> The attached patch moves the \\\"LastRec\\\" variable declaration inside\\n>> StartupXLOG() where it is supposed to be.\\n\\n> Yes, reading through the code, your suggestion makes sense. I'll\\n> double-check and apply if there are no objections.\\n\\nI think LastRec was originally referenced by multiple functions\\nin xlog.c. But it does look like it could be a local now.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 02:08:54 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Unnecessary global variable declared in xlog.c\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 16, 2021 at 02:08:54AM -0500, Tom Lane wrote:\\n> I think LastRec was originally referenced by multiple functions\\n> in xlog.c. But it does look like it could be a local now.\\n\\nThanks for double-checking, applied this one as of f975fc3.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 17 Nov 2021 11:06:04 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Unnecessary global variable declared in xlog.c\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 7:36 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Tue, Nov 16, 2021 at 02:08:54AM -0500, Tom Lane wrote:\\n> > I think LastRec was originally referenced by multiple functions\\n> > in xlog.c. But it does look like it could be a local now.\\n>\\n> Thanks for double-checking, applied this one as of f975fc3.\\n\\nThank you, Michael.\\n\\nRegards,\\nAmul\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 09:18:35 +0530\",\n \"msg_from\": \"Amul Sul <sulamul@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Unnecessary global variable declared in xlog.c\"\n }\n]","isConversational":false}}},{"rowIdx":53097,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Fix headerscheck failure in replication/worker_internal.h\n\nBroken by 31c389d8de91\n\nBranch\n------\nmaster\n\nDetails\n-------\nhttps://git.postgresql.org/pg/commitdiff/ad26ee28250c4cd357a7420161a2be321c3dd536\n\nModified Files\n--------------\nsrc/include/replication/worker_internal.h | 1 +\n1 file changed, 1 insertion(+)","msg_date":"Tue, 16 Nov 2021 16:33:02 +0000","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":true,"msg_subject":"pgsql: Fix headerscheck failure in replication/worker_internal.h"},{"msg_contents":"On 2021-Nov-16, Alvaro Herrera wrote:\n\n> Fix headerscheck failure in replication/worker_internal.h\n\nThe other failure is in src/include/libpq/be-gssapi-common.h:\n\nIn file included from /tmp/headerscheck.a6f0um/test.c:2:\n./src/include/libpq/be-gssapi-common.h:20:10: fatal error: gssapi/gssapi.h: No existe el fichero o el directorio\n 20 | #include <gssapi/gssapi.h>\n | ^~~~~~~~~~~~~~~~~\ncompilation terminated.\n\nOne possible solution for this one is to add an exclusion in\nheaderscheck (and cpluspluscheck?); the other possible solution seems to\nbe to wrap the whole contents of the file in \"#ifdef ENABLE_GSS\". I\nthink the latter is roughly the approach used for OpenSSL inclusions.\n\n-- \nÁlvaro Herrera 39°49'30\"S 73°17'W — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Tue, 16 Nov 2021 14:09:36 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":true,"msg_subject":"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h"},{"msg_contents":"Greetings,\n\n* Alvaro Herrera (alvherre@alvh.no-ip.org) wrote:\n> On 2021-Nov-16, Alvaro Herrera wrote:\n> \n> > Fix headerscheck failure in replication/worker_internal.h\n> \n> The other failure is in src/include/libpq/be-gssapi-common.h:\n> \n> In file included from /tmp/headerscheck.a6f0um/test.c:2:\n> ./src/include/libpq/be-gssapi-common.h:20:10: fatal error: gssapi/gssapi.h: No existe el fichero o el directorio\n> 20 | #include <gssapi/gssapi.h>\n> | ^~~~~~~~~~~~~~~~~\n> compilation terminated.\n> \n> One possible solution for this one is to add an exclusion in\n> headerscheck (and cpluspluscheck?); the other possible solution seems to\n> be to wrap the whole contents of the file in \"#ifdef ENABLE_GSS\". I\n> think the latter is roughly the approach used for OpenSSL inclusions.\n\nShort answer is yes, inclusion of be-gssapi-common.h is typically\nwrapped in a #ifdef, see src/backend/libpq/auth.c\n\nThanks,\n\nStephen","msg_date":"Tue, 16 Nov 2021 12:14:31 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h"},{"msg_contents":"On 2021-Nov-16, Stephen Frost wrote:\n\n> Short answer is yes, inclusion of be-gssapi-common.h is typically\n> wrapped in a #ifdef, see src/backend/libpq/auth.c\n\nIt'd be as in the attached, then.\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/","msg_date":"Tue, 16 Nov 2021 14:33:46 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":true,"msg_subject":"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h"},{"msg_contents":"Greetings,\n\nOn Tue, Nov 16, 2021 at 12:33 Alvaro Herrera <alvherre@alvh.no-ip.org>\nwrote:\n\n> On 2021-Nov-16, Stephen Frost wrote:\n>\n> > Short answer is yes, inclusion of be-gssapi-common.h is typically\n> > wrapped in a #ifdef, see src/backend/libpq/auth.c\n>\n> It'd be as in the attached, then.\n\n\nI’ve not looked at this very closely but no, normally it’s the inclusion of\nbe-gssapi-common.h that’s wrapped in the ifdef, not the contents of it,\nagain, see auth.c\n\nNot against possibly changing that but I don’t get the point of including\nbe-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI\nis possible and the reason for including be-gssapi-common.h then there’s\nother things that need to be under a ifdef, again, as in auth.c\n\nIf there isn’t any need to be-gssapi-common.h then maybe just drop that\ninclude instead of adding an ifdef..?\n\nThanks,\n\nStephen\n\n>\n\nGreetings,On Tue, Nov 16, 2021 at 12:33 Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:On 2021-Nov-16, Stephen Frost wrote:\n\n> Short answer is yes, inclusion of be-gssapi-common.h is typically\n> wrapped in a #ifdef, see src/backend/libpq/auth.c\n\nIt'd be as in the attached, then.I’ve not looked at this very closely but no, normally it’s the inclusion of be-gssapi-common.h that’s wrapped in the ifdef, not the contents of it, again, see auth.cNot against possibly changing that but I don’t get the point of including be-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI is possible and the reason for including be-gssapi-common.h then there’s other things that need to be under a ifdef, again, as in auth.cIf there isn’t any need to be-gssapi-common.h then maybe just drop that include instead of adding an ifdef..?Thanks,Stephen","msg_date":"Tue, 16 Nov 2021 12:47:59 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h"},{"msg_contents":"On 2021-Nov-16, Stephen Frost wrote:\n\n> I’ve not looked at this very closely but no, normally it’s the inclusion of\n> be-gssapi-common.h that’s wrapped in the ifdef, not the contents of it,\n> again, see auth.c\n\nI don't think you read my original post very carefully. I'm talking\nabout sanitizing the output of the headerscheck script.\n\n-- \nÁlvaro Herrera 39°49'30\"S 73°17'W — https://www.EnterpriseDB.com/\n\"Java is clearly an example of money oriented programming\" (A. Stepanov)\n\n\n","msg_date":"Tue, 16 Nov 2021 15:01:54 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":true,"msg_subject":"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h"},{"msg_contents":"On 2021-Nov-16, Stephen Frost wrote:\n\n> Not against possibly changing that but I don’t get the point of including\n> be-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI\n> is possible and the reason for including be-gssapi-common.h then there’s\n> other things that need to be under a ifdef, again, as in auth.c\n\nBTW, this is exactly why my first suggestion was to add an exclusion\nrule to headerscheck so that be-gssapi-common.h is not verified by that\nscript. After re-reading your response, that looks like a reasonable\nanswer too.\n\n-- \nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Tue, 16 Nov 2021 15:16:45 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":true,"msg_subject":"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h"},{"msg_contents":"Greetings,\n\nOn Tue, Nov 16, 2021 at 13:16 Alvaro Herrera <alvherre@alvh.no-ip.org>\nwrote:\n\n> On 2021-Nov-16, Stephen Frost wrote:\n>\n> > Not against possibly changing that but I don’t get the point of including\n> > be-gssapi-common.h if it’s not enabled in the build and typically if\n> GSSAPI\n> > is possible and the reason for including be-gssapi-common.h then there’s\n> > other things that need to be under a ifdef, again, as in auth.c\n>\n> BTW, this is exactly why my first suggestion was to add an exclusion\n> rule to headerscheck so that be-gssapi-common.h is not verified by that\n> script. After re-reading your response, that looks like a reasonable\n> answer too.\n\n\nYeah, that seems better to me, though I’ve not yet had time to look deeply\ninto any of this.\n\nThanks,\n\nStephen\n\n>\n\nGreetings,On Tue, Nov 16, 2021 at 13:16 Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:On 2021-Nov-16, Stephen Frost wrote:\n\n> Not against possibly changing that but I don’t get the point of including\n> be-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI\n> is possible and the reason for including be-gssapi-common.h then there’s\n> other things that need to be under a ifdef, again, as in auth.c\n\nBTW, this is exactly why my first suggestion was to add an exclusion\nrule to headerscheck so that be-gssapi-common.h is not verified by that\nscript. After re-reading your response, that looks like a reasonable\nanswer too.Yeah, that seems better to me, though I’ve not yet had time to look deeply into any of this. Thanks,Stephen","msg_date":"Tue, 16 Nov 2021 13:57:15 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h"},{"msg_contents":"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> On 2021-Nov-16, Stephen Frost wrote:\n>> Not against possibly changing that but I don’t get the point of including\n>> be-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI\n>> is possible and the reason for including be-gssapi-common.h then there’s\n>> other things that need to be under a ifdef, again, as in auth.c\n\n> BTW, this is exactly why my first suggestion was to add an exclusion\n> rule to headerscheck so that be-gssapi-common.h is not verified by that\n> script. After re-reading your response, that looks like a reasonable\n> answer too.\n\nI think adding #ifdef ENABLE_GSS as per your prior message is better.\nHeaders have little business making assumptions about the context in\nwhich they're included --- which is exactly why headerscheck exists ---\nso I disagree with Stephen's argument. In any case I am not in favor of\nmaking random exclusions from that script's testing.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 16 Nov 2021 15:12:07 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h"},{"msg_contents":"Greetings,\n\nOn Tue, Nov 16, 2021 at 15:12 Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> > On 2021-Nov-16, Stephen Frost wrote:\n> >> Not against possibly changing that but I don’t get the point of\n> including\n> >> be-gssapi-common.h if it’s not enabled in the build and typically if\n> GSSAPI\n> >> is possible and the reason for including be-gssapi-common.h then there’s\n> >> other things that need to be under a ifdef, again, as in auth.c\n>\n> > BTW, this is exactly why my first suggestion was to add an exclusion\n> > rule to headerscheck so that be-gssapi-common.h is not verified by that\n> > script. After re-reading your response, that looks like a reasonable\n> > answer too.\n>\n> I think adding #ifdef ENABLE_GSS as per your prior message is better.\n> Headers have little business making assumptions about the context in\n> which they're included --- which is exactly why headerscheck exists ---\n> so I disagree with Stephen's argument. In any case I am not in favor of\n> making random exclusions from that script's testing.\n\n\nI don’t feel all that strongly either way, so if you’d rather have it that\nway then that’s fine. Will still need the other ifdefs too anyway though,\nbut I guess it isn’t that big of a deal.\n\nThanks,\n\nStephen\n\nGreetings,On Tue, Nov 16, 2021 at 15:12 Tom Lane <tgl@sss.pgh.pa.us> wrote:Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> On 2021-Nov-16, Stephen Frost wrote:\n>> Not against possibly changing that but I don’t get the point of including\n>> be-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI\n>> is possible and the reason for including be-gssapi-common.h then there’s\n>> other things that need to be under a ifdef, again, as in auth.c\n\n> BTW, this is exactly why my first suggestion was to add an exclusion\n> rule to headerscheck so that be-gssapi-common.h is not verified by that\n> script. After re-reading your response, that looks like a reasonable\n> answer too.\n\nI think adding #ifdef ENABLE_GSS as per your prior message is better.\nHeaders have little business making assumptions about the context in\nwhich they're included --- which is exactly why headerscheck exists ---\nso I disagree with Stephen's argument. In any case I am not in favor of\nmaking random exclusions from that script's testing.I don’t feel all that strongly either way, so if you’d rather have it that way then that’s fine. Will still need the other ifdefs too anyway though, but I guess it isn’t that big of a deal. Thanks,Stephen","msg_date":"Tue, 16 Nov 2021 15:16:13 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h"}],"string":"[\n {\n \"msg_contents\": \"Fix headerscheck failure in replication/worker_internal.h\\n\\nBroken by 31c389d8de91\\n\\nBranch\\n------\\nmaster\\n\\nDetails\\n-------\\nhttps://git.postgresql.org/pg/commitdiff/ad26ee28250c4cd357a7420161a2be321c3dd536\\n\\nModified Files\\n--------------\\nsrc/include/replication/worker_internal.h | 1 +\\n1 file changed, 1 insertion(+)\",\n \"msg_date\": \"Tue, 16 Nov 2021 16:33:02 +0000\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pgsql: Fix headerscheck failure in replication/worker_internal.h\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-16, Alvaro Herrera wrote:\\n\\n> Fix headerscheck failure in replication/worker_internal.h\\n\\nThe other failure is in src/include/libpq/be-gssapi-common.h:\\n\\nIn file included from /tmp/headerscheck.a6f0um/test.c:2:\\n./src/include/libpq/be-gssapi-common.h:20:10: fatal error: gssapi/gssapi.h: No existe el fichero o el directorio\\n 20 | #include <gssapi/gssapi.h>\\n | ^~~~~~~~~~~~~~~~~\\ncompilation terminated.\\n\\nOne possible solution for this one is to add an exclusion in\\nheaderscheck (and cpluspluscheck?); the other possible solution seems to\\nbe to wrap the whole contents of the file in \\\"#ifdef ENABLE_GSS\\\". I\\nthink the latter is roughly the approach used for OpenSSL inclusions.\\n\\n-- \\nÁlvaro Herrera 39°49'30\\\"S 73°17'W — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 14:09:36 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\n* Alvaro Herrera (alvherre@alvh.no-ip.org) wrote:\\n> On 2021-Nov-16, Alvaro Herrera wrote:\\n> \\n> > Fix headerscheck failure in replication/worker_internal.h\\n> \\n> The other failure is in src/include/libpq/be-gssapi-common.h:\\n> \\n> In file included from /tmp/headerscheck.a6f0um/test.c:2:\\n> ./src/include/libpq/be-gssapi-common.h:20:10: fatal error: gssapi/gssapi.h: No existe el fichero o el directorio\\n> 20 | #include <gssapi/gssapi.h>\\n> | ^~~~~~~~~~~~~~~~~\\n> compilation terminated.\\n> \\n> One possible solution for this one is to add an exclusion in\\n> headerscheck (and cpluspluscheck?); the other possible solution seems to\\n> be to wrap the whole contents of the file in \\\"#ifdef ENABLE_GSS\\\". I\\n> think the latter is roughly the approach used for OpenSSL inclusions.\\n\\nShort answer is yes, inclusion of be-gssapi-common.h is typically\\nwrapped in a #ifdef, see src/backend/libpq/auth.c\\n\\nThanks,\\n\\nStephen\",\n \"msg_date\": \"Tue, 16 Nov 2021 12:14:31 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-16, Stephen Frost wrote:\\n\\n> Short answer is yes, inclusion of be-gssapi-common.h is typically\\n> wrapped in a #ifdef, see src/backend/libpq/auth.c\\n\\nIt'd be as in the attached, then.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\",\n \"msg_date\": \"Tue, 16 Nov 2021 14:33:46 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\nOn Tue, Nov 16, 2021 at 12:33 Alvaro Herrera <alvherre@alvh.no-ip.org>\\nwrote:\\n\\n> On 2021-Nov-16, Stephen Frost wrote:\\n>\\n> > Short answer is yes, inclusion of be-gssapi-common.h is typically\\n> > wrapped in a #ifdef, see src/backend/libpq/auth.c\\n>\\n> It'd be as in the attached, then.\\n\\n\\nI’ve not looked at this very closely but no, normally it’s the inclusion of\\nbe-gssapi-common.h that’s wrapped in the ifdef, not the contents of it,\\nagain, see auth.c\\n\\nNot against possibly changing that but I don’t get the point of including\\nbe-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI\\nis possible and the reason for including be-gssapi-common.h then there’s\\nother things that need to be under a ifdef, again, as in auth.c\\n\\nIf there isn’t any need to be-gssapi-common.h then maybe just drop that\\ninclude instead of adding an ifdef..?\\n\\nThanks,\\n\\nStephen\\n\\n>\\n\\nGreetings,On Tue, Nov 16, 2021 at 12:33 Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:On 2021-Nov-16, Stephen Frost wrote:\\n\\n> Short answer is yes, inclusion of be-gssapi-common.h is typically\\n> wrapped in a #ifdef, see src/backend/libpq/auth.c\\n\\nIt'd be as in the attached, then.I’ve not looked at this very closely but no, normally it’s the inclusion of be-gssapi-common.h that’s wrapped in the ifdef, not the contents of it, again, see auth.cNot against possibly changing that but I don’t get the point of including be-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI is possible and the reason for including be-gssapi-common.h then there’s other things that need to be under a ifdef, again, as in auth.cIf there isn’t any need to be-gssapi-common.h then maybe just drop that include instead of adding an ifdef..?Thanks,Stephen\",\n \"msg_date\": \"Tue, 16 Nov 2021 12:47:59 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-16, Stephen Frost wrote:\\n\\n> I’ve not looked at this very closely but no, normally it’s the inclusion of\\n> be-gssapi-common.h that’s wrapped in the ifdef, not the contents of it,\\n> again, see auth.c\\n\\nI don't think you read my original post very carefully. I'm talking\\nabout sanitizing the output of the headerscheck script.\\n\\n-- \\nÁlvaro Herrera 39°49'30\\\"S 73°17'W — https://www.EnterpriseDB.com/\\n\\\"Java is clearly an example of money oriented programming\\\" (A. Stepanov)\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 15:01:54 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h\"\n },\n {\n \"msg_contents\": \"On 2021-Nov-16, Stephen Frost wrote:\\n\\n> Not against possibly changing that but I don’t get the point of including\\n> be-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI\\n> is possible and the reason for including be-gssapi-common.h then there’s\\n> other things that need to be under a ifdef, again, as in auth.c\\n\\nBTW, this is exactly why my first suggestion was to add an exclusion\\nrule to headerscheck so that be-gssapi-common.h is not verified by that\\nscript. After re-reading your response, that looks like a reasonable\\nanswer too.\\n\\n-- \\nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 15:16:45 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\nOn Tue, Nov 16, 2021 at 13:16 Alvaro Herrera <alvherre@alvh.no-ip.org>\\nwrote:\\n\\n> On 2021-Nov-16, Stephen Frost wrote:\\n>\\n> > Not against possibly changing that but I don’t get the point of including\\n> > be-gssapi-common.h if it’s not enabled in the build and typically if\\n> GSSAPI\\n> > is possible and the reason for including be-gssapi-common.h then there’s\\n> > other things that need to be under a ifdef, again, as in auth.c\\n>\\n> BTW, this is exactly why my first suggestion was to add an exclusion\\n> rule to headerscheck so that be-gssapi-common.h is not verified by that\\n> script. After re-reading your response, that looks like a reasonable\\n> answer too.\\n\\n\\nYeah, that seems better to me, though I’ve not yet had time to look deeply\\ninto any of this.\\n\\nThanks,\\n\\nStephen\\n\\n>\\n\\nGreetings,On Tue, Nov 16, 2021 at 13:16 Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:On 2021-Nov-16, Stephen Frost wrote:\\n\\n> Not against possibly changing that but I don’t get the point of including\\n> be-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI\\n> is possible and the reason for including be-gssapi-common.h then there’s\\n> other things that need to be under a ifdef, again, as in auth.c\\n\\nBTW, this is exactly why my first suggestion was to add an exclusion\\nrule to headerscheck so that be-gssapi-common.h is not verified by that\\nscript. After re-reading your response, that looks like a reasonable\\nanswer too.Yeah, that seems better to me, though I’ve not yet had time to look deeply into any of this. Thanks,Stephen\",\n \"msg_date\": \"Tue, 16 Nov 2021 13:57:15 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> On 2021-Nov-16, Stephen Frost wrote:\\n>> Not against possibly changing that but I don’t get the point of including\\n>> be-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI\\n>> is possible and the reason for including be-gssapi-common.h then there’s\\n>> other things that need to be under a ifdef, again, as in auth.c\\n\\n> BTW, this is exactly why my first suggestion was to add an exclusion\\n> rule to headerscheck so that be-gssapi-common.h is not verified by that\\n> script. After re-reading your response, that looks like a reasonable\\n> answer too.\\n\\nI think adding #ifdef ENABLE_GSS as per your prior message is better.\\nHeaders have little business making assumptions about the context in\\nwhich they're included --- which is exactly why headerscheck exists ---\\nso I disagree with Stephen's argument. In any case I am not in favor of\\nmaking random exclusions from that script's testing.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 15:12:07 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\nOn Tue, Nov 16, 2021 at 15:12 Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> > On 2021-Nov-16, Stephen Frost wrote:\\n> >> Not against possibly changing that but I don’t get the point of\\n> including\\n> >> be-gssapi-common.h if it’s not enabled in the build and typically if\\n> GSSAPI\\n> >> is possible and the reason for including be-gssapi-common.h then there’s\\n> >> other things that need to be under a ifdef, again, as in auth.c\\n>\\n> > BTW, this is exactly why my first suggestion was to add an exclusion\\n> > rule to headerscheck so that be-gssapi-common.h is not verified by that\\n> > script. After re-reading your response, that looks like a reasonable\\n> > answer too.\\n>\\n> I think adding #ifdef ENABLE_GSS as per your prior message is better.\\n> Headers have little business making assumptions about the context in\\n> which they're included --- which is exactly why headerscheck exists ---\\n> so I disagree with Stephen's argument. In any case I am not in favor of\\n> making random exclusions from that script's testing.\\n\\n\\nI don’t feel all that strongly either way, so if you’d rather have it that\\nway then that’s fine. Will still need the other ifdefs too anyway though,\\nbut I guess it isn’t that big of a deal.\\n\\nThanks,\\n\\nStephen\\n\\nGreetings,On Tue, Nov 16, 2021 at 15:12 Tom Lane <tgl@sss.pgh.pa.us> wrote:Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> On 2021-Nov-16, Stephen Frost wrote:\\n>> Not against possibly changing that but I don’t get the point of including\\n>> be-gssapi-common.h if it’s not enabled in the build and typically if GSSAPI\\n>> is possible and the reason for including be-gssapi-common.h then there’s\\n>> other things that need to be under a ifdef, again, as in auth.c\\n\\n> BTW, this is exactly why my first suggestion was to add an exclusion\\n> rule to headerscheck so that be-gssapi-common.h is not verified by that\\n> script. After re-reading your response, that looks like a reasonable\\n> answer too.\\n\\nI think adding #ifdef ENABLE_GSS as per your prior message is better.\\nHeaders have little business making assumptions about the context in\\nwhich they're included --- which is exactly why headerscheck exists ---\\nso I disagree with Stephen's argument. In any case I am not in favor of\\nmaking random exclusions from that script's testing.I don’t feel all that strongly either way, so if you’d rather have it that way then that’s fine. Will still need the other ifdefs too anyway though, but I guess it isn’t that big of a deal. Thanks,Stephen\",\n \"msg_date\": \"Tue, 16 Nov 2021 15:16:13 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pgsql: Fix headerscheck failure in replication/worker_internal.h\"\n }\n]","isConversational":false}}},{"rowIdx":53098,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi:\n\nI'm using the isolation test to write some tests, and then I found the\noutput is unclear some time.\nfor example:\n\na b c\n1988-12-14 01:03:031988-12-13 22:03:03-011988-12-14\n\nActually what I want is there are some spaces among the fields. like:\n\na b c\n1988-12-14 01:03:03 1988-12-13 22:03:03-01 1988-12-14\n\nThe simplest reproduce case is:\n\nsetup\n{\n CREATE TABLE c_tm(a timestamp, b timestamptz, c date);\n}\n\nsession \"s1\"\nstep \"s1inserttm\"\n{\n INSERT INTO c_tm VALUES('1988-12-14 01:03:03 +2', '1988-12-14\n01:03:03 +2', '1988-12-14 01:03:03 +2');\n}\n\nstep \"s1read\"\n{\n SELECT * FROM c_tm;\n}\n\nIs there any simple way to improve this?\n\n-- \nBest Regards\nAndy Fan\n\n\n","msg_date":"Wed, 17 Nov 2021 09:44:00 +0800","msg_from":"Andy Fan <zhihui.fan1213@gmail.com>","msg_from_op":true,"msg_subject":"isolation test output format"},{"msg_contents":"Andy Fan <zhihui.fan1213@gmail.com> writes:\n> I'm using the isolation test to write some tests, and then I found the\n> output is unclear some time.\n> for example:\n\n> a b c\n> 1988-12-14 01:03:031988-12-13 22:03:03-011988-12-14\n\nAre you using recent sources? The isolationtester's output hasn't\nlooked like that since June (4a054069a et al).\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 16 Nov 2021 20:56:27 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: isolation test output format"},{"msg_contents":"> Are you using recent sources? The isolationtester's output hasn't\n> looked like that since June (4a054069a et al).\n>\n> regards, tom lane\n\nI am coding with \"Stamp 14beta2\" which are very close to have 4a054069a.\nI have tried the newer version, the issue here is fixed perfectly. Thanks\nfor that!\n\n\n-- \nBest Regards\nAndy Fan\n\n\n","msg_date":"Wed, 17 Nov 2021 14:04:34 +0800","msg_from":"Andy Fan <zhihui.fan1213@gmail.com>","msg_from_op":true,"msg_subject":"Re: isolation test output format"}],"string":"[\n {\n \"msg_contents\": \"Hi:\\n\\nI'm using the isolation test to write some tests, and then I found the\\noutput is unclear some time.\\nfor example:\\n\\na b c\\n1988-12-14 01:03:031988-12-13 22:03:03-011988-12-14\\n\\nActually what I want is there are some spaces among the fields. like:\\n\\na b c\\n1988-12-14 01:03:03 1988-12-13 22:03:03-01 1988-12-14\\n\\nThe simplest reproduce case is:\\n\\nsetup\\n{\\n CREATE TABLE c_tm(a timestamp, b timestamptz, c date);\\n}\\n\\nsession \\\"s1\\\"\\nstep \\\"s1inserttm\\\"\\n{\\n INSERT INTO c_tm VALUES('1988-12-14 01:03:03 +2', '1988-12-14\\n01:03:03 +2', '1988-12-14 01:03:03 +2');\\n}\\n\\nstep \\\"s1read\\\"\\n{\\n SELECT * FROM c_tm;\\n}\\n\\nIs there any simple way to improve this?\\n\\n-- \\nBest Regards\\nAndy Fan\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 09:44:00 +0800\",\n \"msg_from\": \"Andy Fan <zhihui.fan1213@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"isolation test output format\"\n },\n {\n \"msg_contents\": \"Andy Fan <zhihui.fan1213@gmail.com> writes:\\n> I'm using the isolation test to write some tests, and then I found the\\n> output is unclear some time.\\n> for example:\\n\\n> a b c\\n> 1988-12-14 01:03:031988-12-13 22:03:03-011988-12-14\\n\\nAre you using recent sources? The isolationtester's output hasn't\\nlooked like that since June (4a054069a et al).\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Nov 2021 20:56:27 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: isolation test output format\"\n },\n {\n \"msg_contents\": \"> Are you using recent sources? The isolationtester's output hasn't\\n> looked like that since June (4a054069a et al).\\n>\\n> regards, tom lane\\n\\nI am coding with \\\"Stamp 14beta2\\\" which are very close to have 4a054069a.\\nI have tried the newer version, the issue here is fixed perfectly. Thanks\\nfor that!\\n\\n\\n-- \\nBest Regards\\nAndy Fan\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Nov 2021 14:04:34 +0800\",\n \"msg_from\": \"Andy Fan <zhihui.fan1213@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: isolation test output format\"\n }\n]","isConversational":false}}},{"rowIdx":53099,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi hackers,\n\nI noticed that there are some tab completions missing for the following \ncommands:\n-CREATE CONVERSION : missing FOR, TO, FROM\n-CREATE DOMAIN : missing after AS\n-CREATE LANGUAGE : missing after HANDLER\n-CREATE SCHEMA : missing AUTHORIZATION, IF NOT EXISTS\n-CREATE SEQUENCE : missing AS\n-CREATE TRANSFORM : missing after FOR\n\nI made a patch for this, so please have a look.\n\nBest wishes,\n\n-- \nKen Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION","msg_date":"Wed, 17 Nov 2021 10:44:44 +0900","msg_from":"Ken Kato <katouknl@oss.nttdata.com>","msg_from_op":true,"msg_subject":"CREATE tab completion"},{"msg_contents":"On Wed, Nov 17, 2021 at 10:44:44AM +0900, Ken Kato wrote:\n> I made a patch for this, so please have a look.\n\n+ else if (Matches(\"CREATE\", \"TRANSFORM\", \"FOR\", MatchAny)\n+ COMPLETE_WITH(\"LANGUAGE\")\n+ else if (Matches(\"CREATE\", \"TRANSFORM\", \"FOR\", MatchAny, \"LANGUAGE\")\n\nThose three lines are wrong, for two different reasons and three\nmistakes. You may want to compile your code before sending it :)\n\n\"CREATE [TEMP|TEMPORARY] SEQUENCE name AS\" could be completed with the\nsupported types. There are three of them.\n\n+ COMPLETE_WITH_QUERY(Query_for_list_of_schemas\n+ \" UNION SELECT 'AUTORIZATION'\");\nIncorrect completion here, s/AUTORIZATION/AUTHORIZATION/.\n\n+ else if (Matches(\"CREATE\", \"CONVERSION\", MatchAny))\n+ COMPLETE_WITH(\"FOR\");\nWhy didn't you consider DEFAULT for the set of changes with\nconversions?\n\n+ else if (Matches(\"CREATE\", \"SCHEMA\"))\n+ COMPLETE_WITH(\"AUTHORIZATION\", \"IF NOT EXISTS\");\n+ else if (Matches(\"CREATE\", \"SCHEMA\", \"IF\", \"NOT\", \"EXISTS\"))\nWe don't do any completion for INE or IE in the other object types.\n\n+ /* CREATE LANGUAGE */\n+ else if (Matches(\"CREATE\", \"LANGUAGE\", MatchAny))\n+ COMPLETE_WITH(\"HANDLER\");\n+ else if (Matches(\"CREATE\", \"LANGUAGE\", MatchAny, \"HANDLER\", MatchAny))\n+ COMPLETE_WITH(\"INLINE\", \"VALIDATOR\");\nIt looks like you forgot the case of \"OR REPLACE\" here? This is done\nfor triggers, for example.\n--\nMichael","msg_date":"Wed, 17 Nov 2021 16:31:33 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: CREATE tab completion"},{"msg_contents":"> + else if (Matches(\"CREATE\", \"TRANSFORM\", \"FOR\", MatchAny)\n> + COMPLETE_WITH(\"LANGUAGE\")\n> + else if (Matches(\"CREATE\", \"TRANSFORM\", \"FOR\", MatchAny, \n> \"LANGUAGE\")\n> \n> Those three lines are wrong, for two different reasons and three\n> mistakes. You may want to compile your code before sending it :)\n> \n> + COMPLETE_WITH_QUERY(Query_for_list_of_schemas\n> + \" UNION SELECT\n> 'AUTORIZATION'\");\n> Incorrect completion here, s/AUTORIZATION/AUTHORIZATION/.\n\nThank you for the comments.\n\nI am sorry for the compile error and a typo. I will make sure to compile \nbefore sending it and double check typos.\n\n\n> \"CREATE [TEMP|TEMPORARY] SEQUENCE name AS\" could be completed with the\n> supported types. There are three of them.\n\nFor this part, I did the following:\n+\telse if (TailMatches(\"CREATE\", \"SEQUENCE\", MatchAny, \"AS\") ||\n+\t\t\t TailMatches(\"CREATE\", \"TEMP|TEMPORARY\", \"SEQUENCE\", MatchAny, \n\"AS\"))\n+\t\tCOMPLETE_WITH(\"smallint\", \"integer\", \"bigint\");\n\nAm I doing this right? or Are there better ways to do it?\n\n\nBest wishes,\n\n-- \nKen Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION","msg_date":"Thu, 18 Nov 2021 16:25:30 +0900","msg_from":"Ken Kato <katouknl@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Re: CREATE tab completion"},{"msg_contents":"On Thu, Nov 18, 2021 at 04:25:30PM +0900, Ken Kato wrote:\n> For this part, I did the following:\n> +\telse if (TailMatches(\"CREATE\", \"SEQUENCE\", MatchAny, \"AS\") ||\n> +\t\t\t TailMatches(\"CREATE\", \"TEMP|TEMPORARY\", \"SEQUENCE\", MatchAny, \"AS\"))\n> +\t\tCOMPLETE_WITH(\"smallint\", \"integer\", \"bigint\");\n> \n> Am I doing this right? or Are there better ways to do it?\n\nThat looks fine per se.\n\n> +/* CREATE SCHEMA */\n> +\telse if (Matches(\"CREATE\", \"SCHEMA\"))\n> +\t\tCOMPLETE_WITH(\"AUTHORIZATION\");\n> +\telse if (Matches(\"CREATE\", \"SCHEMA\") && TailMatches(\"AUTHORIZATION\"))\n> +\t\tCOMPLETE_WITH_QUERY(Query_for_list_of_roles);\n\nThe part about CREATE SCHEMA was itching me a bit, until I recalled\nthis recent thread which has a more complete logic:\nhttps://www.postgresql.org/message-id/87im0efqhp.fsf@wibble.ilmari.org\n\nThe rest looks good, I'll take care of that in a bit.\n--\nMichael","msg_date":"Thu, 18 Nov 2021 17:17:25 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: CREATE tab completion"},{"msg_contents":"At Thu, 18 Nov 2021 17:17:25 +0900, Michael Paquier <michael@paquier.xyz> wrote in \n> On Thu, Nov 18, 2021 at 04:25:30PM +0900, Ken Kato wrote:\n> > For this part, I did the following:\n> > +\telse if (TailMatches(\"CREATE\", \"SEQUENCE\", MatchAny, \"AS\") ||\n> > +\t\t\t TailMatches(\"CREATE\", \"TEMP|TEMPORARY\", \"SEQUENCE\", MatchAny, \"AS\"))\n> > +\t\tCOMPLETE_WITH(\"smallint\", \"integer\", \"bigint\");\n> > \n> > Am I doing this right? or Are there better ways to do it?\n> \n> That looks fine per se.\n\nFWIW, I would be a bit perplexed to see type names suggested in\nupper-cases, even if it is acceptable by the parser. If you type-in\nthe following phrase:\n\n=# type boo<tab>\n\nit is completed to \"boolean\" but,\n\n=# type BOO<tab>\n\ndoesn't respond.\n\nSo we could use COMPLETE_WITH_CS instead so that CREATE SEQUENCE\nbehavess the same way with the existing behavior of TYPE.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Thu, 18 Nov 2021 17:20:58 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: CREATE tab completion"},{"msg_contents":"On Thu, Nov 18, 2021 at 05:20:58PM +0900, Kyotaro Horiguchi wrote:\n> So we could use COMPLETE_WITH_CS instead so that CREATE SEQUENCE\n> behavess the same way with the existing behavior of TYPE.\n\nMakes sense.\n\nAnother issue I have noticed with the patch is that it forgets to\napply quotes for the FROM/TO clauses of CREATE CONVERSION, making the\nqueries fail. \\encoding does not care about that but we do for this\nset of DDLs. So I have discarded this part.\n\nOne extra issue was in CREATE TRANSFORM, where we should handle \"OR\nREPLACE\" like the others. And a last issue I had was with CREATE\nLANGUAGE where we would miss TRUSTED. This last one cannot be\ncompleted, but we'd better allow its completion if we can. That makes\nthings a bit tense if you add on top of that the handling of \"OR\nREPLACE\".\n\nI have fixed (or discarded) that, and the parts for sequences, domains\nand transforms remained. That looked like good enough on its own, so\napplied those parts of the patch.\n--\nMichael","msg_date":"Fri, 19 Nov 2021 13:21:48 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: CREATE tab completion"},{"msg_contents":"> I have fixed (or discarded) that, and the parts for sequences, domains\n> and transforms remained. That looked like good enough on its own, so\n> applied those parts of the patch.\n> --\n> Michael\nThank you very much!\n\nI assume Michael has committed the modified version of the patch.\nTherefore, I changed the status to\"committed\" in Commitfest 2022-01.\nhttps://commitfest.postgresql.org/36/3418/\n\nBest wishes,\n\n-- \nKen Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION\n\n\n","msg_date":"Fri, 19 Nov 2021 16:19:01 +0900","msg_from":"Ken Kato <katouknl@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Re: CREATE tab completion"},{"msg_contents":"On Fri, Nov 19, 2021 at 04:19:01PM +0900, Ken Kato wrote:\n> I assume Michael has committed the modified version of the patch.\n> Therefore, I changed the status to\"committed\" in Commitfest 2022-01.\n> https://commitfest.postgresql.org/36/3418/\n\nThanks, I did not notice that :)\n--\nMichael","msg_date":"Fri, 19 Nov 2021 16:33:18 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: CREATE tab completion"}],"string":"[\n {\n \"msg_contents\": \"Hi hackers,\\n\\nI noticed that there are some tab completions missing for the following \\ncommands:\\n-CREATE CONVERSION : missing FOR, TO, FROM\\n-CREATE DOMAIN : missing after AS\\n-CREATE LANGUAGE : missing after HANDLER\\n-CREATE SCHEMA : missing AUTHORIZATION, IF NOT EXISTS\\n-CREATE SEQUENCE : missing AS\\n-CREATE TRANSFORM : missing after FOR\\n\\nI made a patch for this, so please have a look.\\n\\nBest wishes,\\n\\n-- \\nKen Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Wed, 17 Nov 2021 10:44:44 +0900\",\n \"msg_from\": \"Ken Kato <katouknl@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"CREATE tab completion\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 17, 2021 at 10:44:44AM +0900, Ken Kato wrote:\\n> I made a patch for this, so please have a look.\\n\\n+ else if (Matches(\\\"CREATE\\\", \\\"TRANSFORM\\\", \\\"FOR\\\", MatchAny)\\n+ COMPLETE_WITH(\\\"LANGUAGE\\\")\\n+ else if (Matches(\\\"CREATE\\\", \\\"TRANSFORM\\\", \\\"FOR\\\", MatchAny, \\\"LANGUAGE\\\")\\n\\nThose three lines are wrong, for two different reasons and three\\nmistakes. You may want to compile your code before sending it :)\\n\\n\\\"CREATE [TEMP|TEMPORARY] SEQUENCE name AS\\\" could be completed with the\\nsupported types. There are three of them.\\n\\n+ COMPLETE_WITH_QUERY(Query_for_list_of_schemas\\n+ \\\" UNION SELECT 'AUTORIZATION'\\\");\\nIncorrect completion here, s/AUTORIZATION/AUTHORIZATION/.\\n\\n+ else if (Matches(\\\"CREATE\\\", \\\"CONVERSION\\\", MatchAny))\\n+ COMPLETE_WITH(\\\"FOR\\\");\\nWhy didn't you consider DEFAULT for the set of changes with\\nconversions?\\n\\n+ else if (Matches(\\\"CREATE\\\", \\\"SCHEMA\\\"))\\n+ COMPLETE_WITH(\\\"AUTHORIZATION\\\", \\\"IF NOT EXISTS\\\");\\n+ else if (Matches(\\\"CREATE\\\", \\\"SCHEMA\\\", \\\"IF\\\", \\\"NOT\\\", \\\"EXISTS\\\"))\\nWe don't do any completion for INE or IE in the other object types.\\n\\n+ /* CREATE LANGUAGE */\\n+ else if (Matches(\\\"CREATE\\\", \\\"LANGUAGE\\\", MatchAny))\\n+ COMPLETE_WITH(\\\"HANDLER\\\");\\n+ else if (Matches(\\\"CREATE\\\", \\\"LANGUAGE\\\", MatchAny, \\\"HANDLER\\\", MatchAny))\\n+ COMPLETE_WITH(\\\"INLINE\\\", \\\"VALIDATOR\\\");\\nIt looks like you forgot the case of \\\"OR REPLACE\\\" here? This is done\\nfor triggers, for example.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 17 Nov 2021 16:31:33 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: CREATE tab completion\"\n },\n {\n \"msg_contents\": \"> + else if (Matches(\\\"CREATE\\\", \\\"TRANSFORM\\\", \\\"FOR\\\", MatchAny)\\n> + COMPLETE_WITH(\\\"LANGUAGE\\\")\\n> + else if (Matches(\\\"CREATE\\\", \\\"TRANSFORM\\\", \\\"FOR\\\", MatchAny, \\n> \\\"LANGUAGE\\\")\\n> \\n> Those three lines are wrong, for two different reasons and three\\n> mistakes. You may want to compile your code before sending it :)\\n> \\n> + COMPLETE_WITH_QUERY(Query_for_list_of_schemas\\n> + \\\" UNION SELECT\\n> 'AUTORIZATION'\\\");\\n> Incorrect completion here, s/AUTORIZATION/AUTHORIZATION/.\\n\\nThank you for the comments.\\n\\nI am sorry for the compile error and a typo. I will make sure to compile \\nbefore sending it and double check typos.\\n\\n\\n> \\\"CREATE [TEMP|TEMPORARY] SEQUENCE name AS\\\" could be completed with the\\n> supported types. There are three of them.\\n\\nFor this part, I did the following:\\n+\\telse if (TailMatches(\\\"CREATE\\\", \\\"SEQUENCE\\\", MatchAny, \\\"AS\\\") ||\\n+\\t\\t\\t TailMatches(\\\"CREATE\\\", \\\"TEMP|TEMPORARY\\\", \\\"SEQUENCE\\\", MatchAny, \\n\\\"AS\\\"))\\n+\\t\\tCOMPLETE_WITH(\\\"smallint\\\", \\\"integer\\\", \\\"bigint\\\");\\n\\nAm I doing this right? or Are there better ways to do it?\\n\\n\\nBest wishes,\\n\\n-- \\nKen Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\",\n \"msg_date\": \"Thu, 18 Nov 2021 16:25:30 +0900\",\n \"msg_from\": \"Ken Kato <katouknl@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: CREATE tab completion\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 04:25:30PM +0900, Ken Kato wrote:\\n> For this part, I did the following:\\n> +\\telse if (TailMatches(\\\"CREATE\\\", \\\"SEQUENCE\\\", MatchAny, \\\"AS\\\") ||\\n> +\\t\\t\\t TailMatches(\\\"CREATE\\\", \\\"TEMP|TEMPORARY\\\", \\\"SEQUENCE\\\", MatchAny, \\\"AS\\\"))\\n> +\\t\\tCOMPLETE_WITH(\\\"smallint\\\", \\\"integer\\\", \\\"bigint\\\");\\n> \\n> Am I doing this right? or Are there better ways to do it?\\n\\nThat looks fine per se.\\n\\n> +/* CREATE SCHEMA */\\n> +\\telse if (Matches(\\\"CREATE\\\", \\\"SCHEMA\\\"))\\n> +\\t\\tCOMPLETE_WITH(\\\"AUTHORIZATION\\\");\\n> +\\telse if (Matches(\\\"CREATE\\\", \\\"SCHEMA\\\") && TailMatches(\\\"AUTHORIZATION\\\"))\\n> +\\t\\tCOMPLETE_WITH_QUERY(Query_for_list_of_roles);\\n\\nThe part about CREATE SCHEMA was itching me a bit, until I recalled\\nthis recent thread which has a more complete logic:\\nhttps://www.postgresql.org/message-id/87im0efqhp.fsf@wibble.ilmari.org\\n\\nThe rest looks good, I'll take care of that in a bit.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 18 Nov 2021 17:17:25 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: CREATE tab completion\"\n },\n {\n \"msg_contents\": \"At Thu, 18 Nov 2021 17:17:25 +0900, Michael Paquier <michael@paquier.xyz> wrote in \\n> On Thu, Nov 18, 2021 at 04:25:30PM +0900, Ken Kato wrote:\\n> > For this part, I did the following:\\n> > +\\telse if (TailMatches(\\\"CREATE\\\", \\\"SEQUENCE\\\", MatchAny, \\\"AS\\\") ||\\n> > +\\t\\t\\t TailMatches(\\\"CREATE\\\", \\\"TEMP|TEMPORARY\\\", \\\"SEQUENCE\\\", MatchAny, \\\"AS\\\"))\\n> > +\\t\\tCOMPLETE_WITH(\\\"smallint\\\", \\\"integer\\\", \\\"bigint\\\");\\n> > \\n> > Am I doing this right? or Are there better ways to do it?\\n> \\n> That looks fine per se.\\n\\nFWIW, I would be a bit perplexed to see type names suggested in\\nupper-cases, even if it is acceptable by the parser. If you type-in\\nthe following phrase:\\n\\n=# type boo<tab>\\n\\nit is completed to \\\"boolean\\\" but,\\n\\n=# type BOO<tab>\\n\\ndoesn't respond.\\n\\nSo we could use COMPLETE_WITH_CS instead so that CREATE SEQUENCE\\nbehavess the same way with the existing behavior of TYPE.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Nov 2021 17:20:58 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: CREATE tab completion\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 18, 2021 at 05:20:58PM +0900, Kyotaro Horiguchi wrote:\\n> So we could use COMPLETE_WITH_CS instead so that CREATE SEQUENCE\\n> behavess the same way with the existing behavior of TYPE.\\n\\nMakes sense.\\n\\nAnother issue I have noticed with the patch is that it forgets to\\napply quotes for the FROM/TO clauses of CREATE CONVERSION, making the\\nqueries fail. \\\\encoding does not care about that but we do for this\\nset of DDLs. So I have discarded this part.\\n\\nOne extra issue was in CREATE TRANSFORM, where we should handle \\\"OR\\nREPLACE\\\" like the others. And a last issue I had was with CREATE\\nLANGUAGE where we would miss TRUSTED. This last one cannot be\\ncompleted, but we'd better allow its completion if we can. That makes\\nthings a bit tense if you add on top of that the handling of \\\"OR\\nREPLACE\\\".\\n\\nI have fixed (or discarded) that, and the parts for sequences, domains\\nand transforms remained. That looked like good enough on its own, so\\napplied those parts of the patch.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 19 Nov 2021 13:21:48 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: CREATE tab completion\"\n },\n {\n \"msg_contents\": \"> I have fixed (or discarded) that, and the parts for sequences, domains\\n> and transforms remained. That looked like good enough on its own, so\\n> applied those parts of the patch.\\n> --\\n> Michael\\nThank you very much!\\n\\nI assume Michael has committed the modified version of the patch.\\nTherefore, I changed the status to\\\"committed\\\" in Commitfest 2022-01.\\nhttps://commitfest.postgresql.org/36/3418/\\n\\nBest wishes,\\n\\n-- \\nKen Kato\\nAdvanced Computing Technology Center\\nResearch and Development Headquarters\\nNTT DATA CORPORATION\\n\\n\\n\",\n \"msg_date\": \"Fri, 19 Nov 2021 16:19:01 +0900\",\n \"msg_from\": \"Ken Kato <katouknl@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: CREATE tab completion\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 19, 2021 at 04:19:01PM +0900, Ken Kato wrote:\\n> I assume Michael has committed the modified version of the patch.\\n> Therefore, I changed the status to\\\"committed\\\" in Commitfest 2022-01.\\n> https://commitfest.postgresql.org/36/3418/\\n\\nThanks, I did not notice that :)\\n--\\nMichael\",\n \"msg_date\": \"Fri, 19 Nov 2021 16:33:18 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: CREATE tab completion\"\n }\n]","isConversational":false}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":530,"numItemsPerPage":100,"numTotalItems":58371,"offset":53000,"length":100}},"jwt":"eyJhbGciOiJFZERTQSJ9.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc4MTAxMTI2NSwic3ViIjoiL2RhdGFzZXRzL3dhbnNoZW5sL3Bnc3FsLWhhY2tlcnMiLCJleHAiOjE3ODEwMTQ4NjUsImlzcyI6Imh0dHBzOi8vaHVnZ2luZ2ZhY2UuY28ifQ.K1zEmWLp17lxJsvxQs5RGoInBFfSwOmU8fXZfTbQyc5uPx6dLSdwj0UQXCFnGd4LkzzFtFCocdlQPuLXH769Cw","displayUrls":true,"splitSizeSummaries":[{"config":"default","split":"train","numRows":58371,"numBytesParquet":368254123}]},"discussionsStats":{"closed":0,"open":0,"total":0},"fullWidth":true,"hasGatedAccess":true,"hasFullAccess":true,"isEmbedded":false,"savedQueries":{"community":[],"user":[]}}">
threads listlengths 1 2.99k |
|---|
[
{
"msg_contents": "As a follow-on to Conflation of member/privs for predefined roles,\nthis removes is_member_of_role from the header to dissuade it's use\nfor privilege checking. Since SET ROLE must use membership rather than\nprivileges a new, explicitly named can_set_role() function is\nexported.\n\nis_membe... |
[
{
"msg_contents": "Postgres weekly came today with this interesting article:\nhttps://hakibenita.com/sql-medium-text-performance\n\nThe problem it explains is when you have no one records being stored on\ntoast, even if they are toastable. That is because the size of that column\nfits on toast_tuple_target size... |
[
{
"msg_contents": "PSA a smalll patch to fix some typos (acccidental 3x repeated\nlettters) in code commments.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia",
"msg_date": "Thu, 28 Oct 2021 07:34:17 +1100",
"msg_from": "Peter Smith <smithpb2250@gmail.com>",
"msg_from_op": true,
"msg_s... |
[
{
"msg_contents": "Attached is an updated version of the patch to replace\nis_member_of_role with has_privs_for_role for predefined roles. It\ndoes not remove is_member_of_role from acl.h but it does add a warning\nnot to use that function for privilege checking.\n\nPlease consider this for the upcoming commitf... |
[
{
"msg_contents": "Add TAP test for archive_cleanup_command and recovery_end_command\n\nThis adds tests checking for the execution of both commands. The\nrecovery test 002_archiving.pl is nicely adapted to that, as promotion\nis triggered already twice there, and even if any of those commands fail\nthey don't ... |
[
{
"msg_contents": "Hi hackers,\r\n\r\nI've seen a few cases now for v13 where the startup process on a\r\nstandby appears to be stuck on StandbyReleaseLockList(). It looks\r\nlike most of the time is spent on list_delete_first(). I believe this\r\nis related to the recent list rewrite (1cff1b9), which has the... |
[
{
"msg_contents": "Hi,\n\nIn xlog_redo, for end-of-recovery case error message describes the\nrecord as a checkpoint record which seems to be incorrect; the\nattached patch corrects that.\n\n--\nRegards,\nAmul Sul\nEDB: http://www.enterprisedb.com",
"msg_date": "Thu, 28 Oct 2021 18:28:41 +0530",
"msg_fr... |
[
{
"msg_contents": "Hi,\n\nAccording to a reported PL/Java issue [0], SubTransactionId in\nPostgres PRO EE 13 has become a typedef uint64 rather than uint32.\n\nWhat are the plans for this type upstream? I notice it is still uint32\nhere, even for 14. Are there plans for it to become uint64 at some point?\nOr to... |
[
{
"msg_contents": "I previously reported[1] that CreateReplicationSlot() was accessing\nthe global variable ThisTimeLineID when it might not be initialized.\nToday I realized that the code in basebackup.c has the same problem.\nperform_base_backup() blithely uses the variable six times without\ndoing anything a... |
[
{
"msg_contents": "Hi,\nI was looking at:\n\ncommit 475dbd0b718de8ac44da144f934651b959e3b705\nAuthor: David Rowley <drowley@postgresql.org>\nDate: Tue Aug 3 11:47:24 2021 +1200\n\n Track a Bitmapset of non-pruned partitions in RelOptInfo\n\nI noticed that the dummy relation is skipped in the loop\nover rel... |
[
{
"msg_contents": "The docs have said this for 8 years:\n| \"As of PostgreSQL 9.3, complete coverage exists only for errors in SQLSTATE class 23 (integrity constraint violation), but this is likely to be expanded in future.\"\n\nIs there any appetite for a patch like this one to improve that ?\n\nOne might argu... |
[
{
"msg_contents": "Hi all,\n\nSince 911e702 (13~), it is possible to define opclass parameters for\nindex attributes as of CREATE INDEX, but we lack an equivalent grammar\nfor ALTER INDEX. I was looking at that, and it seems natural to me to\ndo the same thing as what we do for SET STATISTICS, where we use a\n... |
[
{
"msg_contents": "Hi,\n\nThe scanning tool used by our organization has detected the presence of vulnerable libxml version in the latest Postgres 13.4 release for windows (Zip version).\n\nDetected by Automated Scanning tool:\nlibxml 2.9.10\n\nCan you confirm if this is the same version of libxml used in Pos... |
[
{
"msg_contents": "Hi,\n\nWhen recently debugging a vacuum-related issue, I noticed that there\nis some log-report processing/formatting code at the end of\nlazy_scan_heap() that, unless the vacuum VERBOSE option is specified,\ntypically results in no log output (as the log-level is then DEBUG2).\nI think it is... |
[
{
"msg_contents": "Hi,\n\nIn plpgsql, i'm trying to use a variable initialised in a DECLARE in\na later initialisation expression whithin the same DECLARE:\n\n DECLARE\n x int := 23;\n y int := x + 4\n BEGIN\n ...\n\nExperiments suggest that it works, but the docs are silent about that.\nIs this a goo... |
[
{
"msg_contents": "While preparing commit a2a731d6c, I chanced to notice what seemed\nlike an off-by-one error in the line number reported for an error\noccurring while we initialize a plpgsql variable. For instance\n\nregression=# do $$ declare x int := 1/0;\nregression$# begin end $$;\nERROR: division by ze... |
[
{
"msg_contents": "Hi,\n\nAt times, there can be many snapshot and mapping files under\npg_logical dir that the checkpoint might have to delete/fsync based on\nthe cutoff LSN which can increase the checkpoint time. The\nuser/admin/developer might want to get an answer for \"what's the total\ntime spent in delet... |
[
{
"msg_contents": "The attached patch responds to the discussion at [1]. That thread\npointed out that if you issue \"\\password\" with no argument, psql tries\nto set the password for the role identified by PQuser(conn), that is,\nthe originally requested login name. That's problematic because\nif you've don... |
[
{
"msg_contents": "Hi,\nI've been trying to run the SSL tests against my branch and that was\ntougher than expected because I didn't realize that additional output was\nbeing saved that I couldn't see - it wasn't even getting to the part where\nit could run the tests. This patch adds a note to the README explai... |
[
{
"msg_contents": "Dear Hackers,\n\nI want to propose the feature that checks the health of foreign servers. \nAs a first step I want to add an infrastructure for periodical checking to PostgreSQL core.\nCurrently this is the WIP, it does not contain docs.\n\n## Background \n\nCurrently there is no way to check... |
[
{
"msg_contents": "Hi all,\n\nWhile reviewing the code for opclass parameters with indexes, I have\nnoticed that opclass parameters are lost after a concurrent reindex.\nAs we use a IndexInfo to hold the information of the new index when\ncreating a copy of the old one, it is just a matter of making sure\nthat ... |
[
{
"msg_contents": "Hi,\n\nDue to bug #17245: [1] I spent a considerably amount of time looking at vacuum\nrelated code. And I found a few things that I think could stand improvement:\n\n- There's pretty much no tests. This is way way too complicated feature for\n that. If there had been tests for the obvious e... |
[
{
"msg_contents": "Hi,\n\nAs I said before [1], I’m working on $SUBJECT. Attached is a WIP\npatch for that. The patch is pretty simple: if a server option added\nby the patch “parallel_commit” is enabled, 1) asynchronously send\nCOMMIT TRANSACTION (RELEASE SAVEPOINT) to all remote servers involved\nin a local... |
[
{
"msg_contents": "I want to reactivate $subject. I took Petr Jelinek's patch from [0], \nrebased it, added a bit of testing. It basically works, but as \nmentioned in [0], there are various issues to work out.\n\nThe idea is that the standby runs a background worker to periodically \nfetch replication slot i... |
[
{
"msg_contents": "Hi,\n\nIt is sometimes super important to be able to answer customer\nquestions like: What was the total time taken by the last recovery of\nthe server? What was the time taken by each phase of recovery/redo\nprocessing of the startup process? Why did the recovery take so long?\nWe've encount... |
[
{
"msg_contents": "Hi,\n\nIt seems the checkpoint stats, that are emitted to server logs when\nthe GUC log_checkpoints is enabled, are so important that a postgres\ndatabase provider would ever want to disable the GUC. Right now, the\nGUC is disabled by default and a postgres database provider needs to\nenable ... |
[
{
"msg_contents": "Attached is a WIP patch to add new WAL records to represent a logical\ninsert, update, or delete. These records do not do anything at REDO\ntime, they are only processed during logical decoding/replication.\n\nThese are intended to be used by a custom table AM, like my columnar\ncompression e... |
[
{
"msg_contents": "Hi,\n\nThe last release of python2 was 2.7.18 released 2020-04-20, with support\nalready having ended before that 2020-01-01.\n\nTo me it seems time to drop plpython2 support. Supporting plpython2\nuntil ~7 years after python2 is EOL is already quite long... It'd be one\nthing if it were comp... |
[
{
"msg_contents": "Hi,\nI was going over:\n\ncommit 71f4c8c6f74ba021e55d35b1128d22fb8c6e1629\nAuthor: Alvaro Herrera <alvherre@alvh.no-ip.org>\nDate: Thu Mar 25 18:00:28 2021 -0300\n\n ALTER TABLE ... DETACH PARTITION ... CONCURRENTLY\n\nIt seems there were some missing words in one of the comments.\n\nSee... |
[
{
"msg_contents": "hamerkop has been failing recently [1] with\n\n c:\\\\build-farm-local\\\\buildroot\\\\head\\\\pgsql.build\\\\src\\\\backend\\\\libpq\\\\be-secure-openssl.c(67): error C2143: \\215\\\\\\225\\266\\203G\\203\\211\\201[: ')' \\202\\252 '(' \\202\\314\\221O\\202\\311\\202\\240\\202\\350\\202\\33... |
[
{
"msg_contents": "Hi,\n\nCurrently postgres builds extension shared libraries (i.e. pretty much\neverything but libpq) with global visibilty. I.e. symbols that are not\nstatic will be visible from outside the shared library.\n\nOn the one platform where that behaviour is not available, namely\nwindows, we emul... |
[
{
"msg_contents": "Hi,\n\nI'm developing an extension, after modifying some source file(add some message \nlike gettext('xxxxx')) and execute\n> $ make USE_PGXS=1 update-po\n> make: Nothing to be done for 'update-po'.\n\nI feel strange because I think *.po.new files should be created.\n\nAfter some research, I ... |
[
{
"msg_contents": "Reading 1051867.1635720347@sss.pgh.pa.us I noticed that hamerkop raise a C4819\nwarning on brin_bloom.c, which is defined as:\n\n \"The file contains a character that cannot be represented in the current\n code page (number). Save the file in Unicode format to prevent data loss.\"\n\nTh... |
[
{
"msg_contents": "It is now November 1st AOE and thus the 2021-11 commitfest is now in progress.\nI've switched the status and opened 2022-01 for new patches, but noone has so\nfar raised their hand to run it AFAICT? Do we have a volunteer Commitfest\nManager keen to help the community make progress on closin... |
[
{
"msg_contents": "As I mentioned recently at\n<https://postgr.es/m/5d72f199-dc11-89a8-29d1-f20f9687c86f@dunslane.net>, \nI want to get USE_MODULE_DB working for vcregress.pl. I started out\nwriting code to strip this from the command line or get it from the\nenvironment, but then it struck me that if would be ... |
[
{
"msg_contents": "This thread is a continuation of the thread with the subject\r\n\"parallelizing the archiver\" [0]. That thread had morphed into an\r\neffort to allow creating archive modules, so I've created a new one to\r\nensure that this topic has the proper visibility.\r\n\r\nI've attached the latest p... |
[
{
"msg_contents": "Hi,\n\nThe global variable ThisTimeLineID is rather confusing. During\nrecovery, in the startup process, when we're reading a WAL record, it\nis the timeline of the WAL record we are trying to read or have just\nread, except when we're trying to read the initial checkpoint record,\nwhen it's ... |
[
{
"msg_contents": "Meson depends on ninja, so it behooves us to look into portability\nof ninja if we're considering migrating to meson as the build system.\nI tried it out quickly on some ancient platforms, and here's what\nI found.\n\n1. The documentation is kinda sucky, eg it fails to explain how\nto run nin... |
[
{
"msg_contents": "Hi,\n\nTrying to insert NULL value to the Identity column defined by \"GENERATED BY\nDEFAULT\" is disallowed, but there can be use cases where the user would\nlike to have an identity column where manual NULL insertion is required(and\nit should not error-out by Postgres).\n\nHow about having... |
[
{
"msg_contents": "Hi!\n\nI'm here to answer your questions about contrib/pg_query_state.\n>> I only took a quick look at pg_query_state, I have some questions.\n\n>> pg_query_state seems using shm_mq to expose the plan information, but\n>> there was a discussion that this kind of architecture would be tricky\n... |
[
{
"msg_contents": "Hi,\n\n\nSince we have the regexp_like operator I have found that there is two \nreferences in the documentation about PostgreSQL lacking of LIKE_REGEX \nimplementation. Here is a patch to fix the documentation. I simply \nremove the reference to non exist of LIKE_REGEX in PostgreSQL in chapt... |
[
{
"msg_contents": "Attached is a patch to create a helper function which creates a\ntuplestore to be used by FUNCAPI-callable functions.\n\nIt was suggested in [1] that I start a separate thread for this patch.\n\nA few notes:\n\nThere are a few places with very similar code to the new helper\n(MakeTuplestore()... |
[
{
"msg_contents": "Hi,\n\nAndres mentioned in passing that he'd defined\nPG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY only on Itanium, PPC and x86 but\nnot ARM.\n\nI took a look at https://developer.arm.com/documentation/ddi0487/gb/\nunder \"B2.2.1 Requirements for single-copy atomicity\" and it seemed\nlike we should t... |
[
{
"msg_contents": "Hi list,\n\njust compiled PostgreSQL head on RedHat 9 beta. Generally seems to be fine but when building contrib, a lot of these warnings pop up:\n\n/usr/lib64/perl5/CORE/inline.h:2635:5: warning: '}' and ')' tokens terminating statement expression appear in different macro expansion contexts... |
[
{
"msg_contents": "The code in gistvacuum.c is closely based on similar code in nbtree.c,\nexcept that it only acquires an exclusive lock -- not a\nsuper-exclusive lock. I suspect that that's because it seemed\nunnecessary; nbtree plain index scans have their own special reasons\nfor this, that don't apply to G... |
[
{
"msg_contents": "Hi,\n\nsqlsmith triggers the following assertion when testing REL_14_STABLE:\n\n TRAP: FailedAssertion(\"a1 <= a2\", File: \"brin_minmax_multi.c\", Line: 1879, PID: 631814)\n\nI can reproduce it with the following query on a fresh regression\ndatabase:\n\n insert into public.brintest_mu... |
[
{
"msg_contents": "Hey,\n\nI've noticed that the current documentation doesn't mention IF EXISTS clause for ALTER STATISTICS in the synopsis section, where PG supports it.\nhttps://www.postgresql.org/docs/14/sql-alterstatistics.html\n(Only for the last item, that is ALTER STATISTICS .. SET STATISTICS; for the o... |
[
{
"msg_contents": "Hi,\n\nI see that \"procsignal\" and \"ProcSignal\" are being used in the code\ncomments which look inconsistent. IMO, \"ProcSignal\" is the right word\nto use and let's be consistent across the code comments. Attaching a\ntiny patch for that.\n\nThoughts?\n\nRegards,\nBharath Rupireddy.",
... |
[
{
"msg_contents": "We have amassed quite a lot of patches in the CF app, and while Jaime did a\nvery good job closing patches in the last CF there is still a lot to go\nthrough. I've attempted a brief per-patch triage here to see where we are with\nthese. Reading every email in the threads for these patches i... |
[
{
"msg_contents": "First draft is up at\n\nhttps://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=39387228c8b7043d168bada0c64e9f59e83285f5\n\nAs usual, please send any comments/corrections by Sunday.\n\n\t\t\tregards, tom lane\n\n\n",
"msg_date": "Fri, 05 Nov 2021 20:27:49 -0400",
"msg_from"... |
[
{
"msg_contents": "I recently pushed a commit which formalized what we expect of HOT\nchains during pruning -- see assertions and comments added by commit\n5cd7eb1f, \"Add various assertions to heap pruning code\". I notice that\nverify_heapam() doesn't really do anything with HOT chains, and I'd\nsay that that... |
[
{
"msg_contents": "On Sun, Oct 24, 2021 at 04:35:02PM -0700, Noah Misch wrote:\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kittiwake&dt=2021-10-24%2012%3A01%3A10\n> got an interesting v9.6 failure [...]:\n> \n> 2021-10-24 14:25:29.263 CEST [34569:175] pgbench ERROR: could not read two-phase stat... |
[
{
"msg_contents": "Ok, subject was a bit philosophical but this message I just found is\nquite confusing.\n\n\"\"\"\nregression=# select cast(null as anyrange) &> cast(null as anymultirange);\nERROR: argument declared anymultirange is not a multirange type but type anymultirange\n\"\"\"\n\n-- \nJaime Casanova\... |
[
{
"msg_contents": "Hi,\n\nI have a question that is a specification of permission check\n(visibilityrule) for psql meta-command with schema option.\n\nAccording to the source code [1], there is no check if a schema\noption is added. As a result, a role that is not granted can see\nother roles' object names.\nWe... |
[
{
"msg_contents": "Hi,\n\nConfiguration parameters for pg_stat_statements were not in the index,\nso I added them just like auto_explain configuration parameters.\n\nBest wishes,\n\n-- \nKen Kato\nAdvanced Computing Technology Center\nResearch and Development Headquarters\nNTT DATA CORPORATION",
"msg_date":... |
[
{
"msg_contents": "Hi All,\n\nWhile I was searching for methods to send the result of a query to the other side of the world, because it might be nice to go there once in a while, I noticed\nmy best option, SSL compression, has been disabled as of version 14. Could someone please explain to me, why compression ... |
[
{
"msg_contents": "Hello,\n\nExpanded data types (types that have a different in-memory working\nrepresentation than on disk) are pretty essential to a couple of projects\nthat I work on, and while the prose documentation is great at explaining\nthe rationale and technical overview, there isn't a complete,\nsel... |
[
{
"msg_contents": "Motivation:\n\nI'm working on a columnar compression AM[0]. Currently, it uses generic\nxlog, which works for crash recovery and physical replication, but not\nlogical decoding/replication.\n\nExtensible rmgr would enable the table AM to support its own\nredo/decode hooks and WAL format, so t... |
[
{
"msg_contents": "Hi all,\n\nI have noticed $subject while looking at a patch in this area:\nhttps://www.postgresql.org/message-id/YYiqxtEqhgb5gIWL@paquier.xyz\n\nWhile we don't expect the four callers of XLogReadRecord() related to\nlogical decoding to fail in the code paths changed by the patch\nattached, I ... |
[
{
"msg_contents": "Below is the second pass over the patches in the current commitfest, this time\nthe patches which have been around for 5 or 6 commitfests. The usual\ndisclaimers that I very well might have missed something still apply.\n\n\n2773: libpq compression\n=======================\nThis patch intend... |
[
{
"msg_contents": "Hi,\n\nWhile trying to add some new tests, I found that\nPostgreSQL::Test::Utils is not required. I felt\nPostgreSQL::Test::Utils can be removed from a lot of tap tests which\ndo not require it. I removed it, ran the tests and found the tests to\nbe executing fine. I have made a patch includi... |
[
{
"msg_contents": "Hi,\r\n\r\nAttached please find a draft of the release announcement for 2021-11-11. \r\nPlease provide any feedback no later than Thu, Nov 11, 2021 0:00 AoE[1].\r\n\r\nThanks!\r\n\r\nJonathan\r\n\r\n[1] https://en.wikipedia.org/wiki/Anywhere_on_Earth",
"msg_date": "Tue, 9 Nov 2021 11:25:3... |
[
{
"msg_contents": "Coverity complained about this code in bbstreamer_file.c,\nsaying that the get_gz_error() call could be accessing freed\nmemory:\n\n if (gzclose(mystreamer->gzfile) != 0)\n {\n pg_log_error(\"could not close compressed file \\\"%s\\\": %s\",\n mystreamer->path... |
[
{
"msg_contents": "Hi,\n\nI notice, belatedly, that 141fd1b tweaked the order of some members of\nstruct catctup, moving c_list and my_cache from earlier locations to the\nend of the struct. [0]\n\nThat resulted in this sequence at the end of the struct:\n\n...\n HeapTupleData tuple; /* tuple manageme... |
[
{
"msg_contents": "ISTM that the recently-introduced new frontend logging support\n(common/logging.h et al) could use a second pass now that we have\nsome experience with it. There are two points that are bugging me:\n\n1. The distinction between \"error\" and \"fatal\" levels seems squishy\nto the point of u... |
[
{
"msg_contents": "Hi all,\n\nThis is a follow-up of the discussion that happened here:\nhttps://www.postgresql.org/message-id/YYoJYO475dsqYhta@paquier.xyz\n\nRecently, we are getting warnings in the build of plperl when using at\nleast clang-12 because of their new flag -Wcompound-token-split-by-macro,\nall th... |
[
{
"msg_contents": "I'm currently working on a GiST extension for PostgreSQL and I ran into strange ORDER BY results during my queries.\n\nBecause I can't find the problem in my source code, I want to investigate the issue by looking at the PostgreSQL source,\nmaybe inserting extra log messages.\nWhile trying to... |
[
{
"msg_contents": "Do Aggregate functions/queries reuse the state value they generate for\nsubsequent queries? If not then an explanation would be greatly\nappreciated.\n\nDo Aggregate functions/queries reuse the state value they generate for subsequent queries? If not then an explanation would be greatly appre... |
[
{
"msg_contents": "Hi\n\nI think I found a bug related to logical replication(REPLICA IDENTITY in specific).\nIf I change REPLICA IDENTITY after creating publication, the DELETE/UPDATE operations won't be replicated as expected. \n\nFor example:\n-- publisher\nCREATE TABLE tbl(a int, b int);\nALTER TABLE tbl A... |
[
{
"msg_contents": "Greetings hackers.\n\nIt seems that PostgreSQL 14 allows using the AT TIME ZONE operator within\ngenerated column definitions; according to the docs, that means the\noperator is considered immutable. However, unless I'm mistaken, the result\nof AT TIME ZONE depends on the time zone database, ... |
[
{
"msg_contents": "Hello,\n\nafter plain ./configure without options (including noticeable absence of \n--with-openssl) and make,\nmake -C contrib COPT=-Werror gives error with gcc-11 on REL_9_6_STABLE, \nREL_10_STABLE or REL_11_STABLE.\n\nThe warning/error is about misleading indent in \ncontrib/pgcrypto/imath... |
[
{
"msg_contents": "Hi,\n\nAs discussed in [1], isn't it a better idea to add some of activity\nmessages [2] such as recovery, archive, backup, streaming etc. to\nserver logs at LOG level? They are currently being set into ps display\nwhich is good if the postgres is being run on a standalone box/VM\nwhere users... |
[
{
"msg_contents": "postgres=# SELECT log_time , database, user_name, error_severity sev, left(message,99) FROM postgres_log_2021_11_10_0800 WHERE log_time BETWEEN '2021-11-10 08:57' AND '2021-11-10 08:58' AND database IS NULL;\n log_time | database | user_name | sev | ... |
[
{
"msg_contents": "Hello I have such case: create table test (id int not null, status text);insert into test select i, 'foo' from generate_series(1,1000000) i;update test set status = 'bar' where id <= 10;create index test_id on test (id );create index test_status_partial on test (status) where status = 'bar';a... |
[
{
"msg_contents": "tadarida just reported this [1]:\n\n2021-11-11 14:40:40.040 CET [29620:388] 023_cic_2pc.pl ERROR: could not read two-phase state from WAL at 0/1716C68\n\nThat error message is remarkably unhelpful, and I think unnecessarily so.\nWhy the heck isn't XlogReadTwoPhaseData passing on the error st... |
[
{
"msg_contents": "Hi,\n\nI have noticed that $SUBJECT, and I think it's bad, and I'd like to\npropose some patches to fix it. 0001 is hopefully uncontroversial, but\n0002 might need some discussion as I'm not quite sure what approach to\ntake.\n\nThe reason why RecoveryInProgress() has critical side effects is... |
[
{
"msg_contents": "Hello\n\nRecently I noticed extremely suspicious behavior from queries under JIT.\nA query (details for which sadly I cannot disclose) had an EXPLAIN\nlooking like this:\n\n JIT:\n Functions: 743\n Options: Inlining true, Optimization true, Expressions true, Deforming true\n Timing: Gen... |
[
{
"msg_contents": "\nperldoc Test::More says:\n\n Key feature milestones include:\n\n subtests\n Subtests were released in Test::More 0.94, which came with Perl\n 5.12.0. Subtests did not implicitly call \"done_testing()\" until\n 0.96; the first Perl with that fix... |
[
{
"msg_contents": "Welcome to the third installment of the CF 2021-11 patch triage, looking at the\npatches which have been in four CF's and parts of the ones with three CF's. As\nper usual I might very well have misunderstood something (or many things), but\nthere seem to be a fair few ones which should be cl... |
[
{
"msg_contents": "Hi,\n\nCurrently, the server shuts down with a FATAL error (added by commit\n[1]) when the recovery target isn't reached. This can cause a server\navailability problem, especially in case of disaster recovery (geo\nrestores) where the primary was down and the user is doing a PITR on a\nserver... |
[
{
"msg_contents": "While reading through and working with the hundreds of patches in the CF app a\nsmall feature/process request struck me: it would be really helpful if the\npatch had a brief abstract outlining what it aims to add or fix (or summary,\ndescription or something else; not sure what to call it). ... |
[
{
"msg_contents": "The combination of these two statements in a transaction hits an\nAssert in heapam.c at line 4770 on REL_14_STABLE\n\nBEGIN;\nSELECT * FROM queue LIMIT 1 FOR UPDATE SKIP LOCKED;\n...\nUPDATE queue SET status = 'UPDATED' WHERE id = :id;\nCOMMIT;\n\npgbench reliably finds this, running from ins... |
[
{
"msg_contents": "The following code appears at the end of heap_page_prune():\n\n /*\n * If requested, report the number of tuples reclaimed to pgstats. This is\n * ndeleted minus ndead, because we don't want to count a now-DEAD root\n * item as a deletion for this purpose.\n */\n if (rep... |
[
{
"msg_contents": "I wasn't aware of $SUBJECT ... were you?\n\nHere's a demonstration:\n\ndrop table if exists pk, fk, pk1, pk2;\ncreate table pk (f1 int primary key) partition by list(f1);\ncreate table pk1 partition of pk for values in (1);\ncreate table pk2 partition of pk for values in (2);\ninsert into pk ... |
[
{
"msg_contents": "Re-reading the thread [1] (cannot answer there – don't have those emails in\nmy box anymore), I see that there was strong support for enabling BUFFERS\nin EXPLAIN ANALYZE by default. And there were patches. Commitfest entry [2]\nwas marked Rejected because there were questions to the implemen... |
[
{
"msg_contents": "\nHi, hackers\n\nWhen I try to create table that has a varchar(n) data type, I find an\ninconsistent error message for it.\n\npostgres=# CREATE TABLE tbl (s varchar(2147483647));\nERROR: length for type varchar cannot exceed 10485760\nLINE 1: CREATE TABLE tbl (s varchar(2147483647));\n ... |
[
{
"msg_contents": "\nHi, hackers\n\nWhen I try to insert an Unicode \"\\u0000\", there is an error $subject.\n\npostgres=# CREATE TABLE tbl (s varchar(10));\nCREATE TABLE\npostgres=# INSERT INTO tbl VALUES (E'\\u0000');\nERROR: invalid Unicode escape value at or near \"\\u0000\"\nLINE 1: INSERT INTO tbl VALUES... |
[
{
"msg_contents": "Hi,\n\nMy name is Deng Kaisheng, a graduate student in National University of Sinapore (NUS) now major in computer science. I've read the introduction of your organization and I'm quite interested in what you're doing.\n\nI want to do something and contribute to the community, and I wonder if... |
[
{
"msg_contents": "Hi hackers,\n\nIf wrong GUCs of auth_delay, pg_trgm, postgres_fdw and sepgsql are \ndescribed in postgresql.conf, a warning is not emitted unlike \npg_stat_statements, auto_explain and pg_prewarm.\nSo, I improved it by adding EmitWarningsOnPlaceholders.\nAn example output is shown below.\n---... |
[
{
"msg_contents": "Hi,\n\nI have been wondering why there is no meta-command for listing\nconstraints in psql. So, I created a POC patch by using my\nexperience developing \\dX command in PG14.\n\nThis feature is helpful for DBAs when they want to check or\nmodify the definition of constraints.\n\nThe current s... |
[
{
"msg_contents": "Hi hackers,\n\nIn another thread[1], we found the pg_get_publication_tables function will output\nduplicate partition relid when adding both child and parent table to the\npublication(pubviaroot = false).\n\nExample:\ncreate table tbl1 (a int) partition by range (a);\ncreate table tbl1_part1 ... |
[
{
"msg_contents": "Hi,\n\npg_waldump options, --follow or -f(to keep polling once per second for\nnew WAL to appear) and --stats or -z don't work well together i.e. the\ncommand stucks [1]. I think the pg_waldump should emit an error. Note\nthat the pg_basebakup does error out on incompatible options.\n\nHere's... |
[
{
"msg_contents": "Hi,\n\nRobert and I wondered whether we would like to rename the startup\n\nprocess. The reason for doing this is that the current name doesn't\n\nmake any sense, as in the stand-by mode replay loop as the main\n\nloop, the startup process stays around.\n\n\nI think a better name for the pro... |
[
{
"msg_contents": "Hackers,\n\nIn the ongoing effort [1] to reduce the number of tasks which require operating under superuser privileges, this patch extends the system to allow, per GUC variable, the ability to SET or ALTER SYSTEM for the variable. A previous patch set was submitted [2] which created hard-cod... |
[
{
"msg_contents": "Hi.\n\nI noticed recently that the CREATE PUBLICATION docs page [1] does not\nhave any \"See Also\" reference to the CREATE SUBSCRIPTION docs page\n[2].\n\nPerhaps it is deliberate (??) but IMO it just seemed like a\nstrange/annoying omission that there was no easy way to jump both ways\nbetw... |
[
{
"msg_contents": "Hackers,\n\nWhile working on a fix for dangling references to dropped roles in the pg_auth_members.grantor field, I happened to notice we entirely lack regression test coverage of the REVOKE ADMIN OPTION FOR form of the RevokeRoleStmt. I am unaware of any bugs in the current implementation, ... |
[
{
"msg_contents": "Hi,\n\nThe attached patch moves the \"LastRec\" variable declaration inside\nStartupXLOG() where it is supposed to be.\n\n-- \nRegards,\nAmul Sul\nEDB: http://www.enterprisedb.com",
"msg_date": "Tue, 16 Nov 2021 11:17:27 +0530",
"msg_from": "Amul Sul <sulamul@gmail.com>",
"msg_fro... |
[
{
"msg_contents": "Fix headerscheck failure in replication/worker_internal.h\n\nBroken by 31c389d8de91\n\nBranch\n------\nmaster\n\nDetails\n-------\nhttps://git.postgresql.org/pg/commitdiff/ad26ee28250c4cd357a7420161a2be321c3dd536\n\nModified Files\n--------------\nsrc/include/replication/worker_internal.h | 1... |
[
{
"msg_contents": "Hi:\n\nI'm using the isolation test to write some tests, and then I found the\noutput is unclear some time.\nfor example:\n\na b c\n1988-12-14 01:03:031988-12-13 22:03:03-011988-12-14\n\nActually what I want is there are some spaces among the fields. like:\n\na ... |
[
{
"msg_contents": "Hi hackers,\n\nI noticed that there are some tab completions missing for the following \ncommands:\n-CREATE CONVERSION : missing FOR, TO, FROM\n-CREATE DOMAIN : missing after AS\n-CREATE LANGUAGE : missing after HANDLER\n-CREATE SCHEMA : missing AUTHORIZATION, IF NOT EXISTS\n-CREATE SEQUENCE ... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.