The attached patch teaches psql to redisplay any not-yet-executed\\n> query text after editing with \\\\e. The fact that you don't get to\\n> see what you're about to execute has been complained of before,\\n> most recently at bug #16034 [1]. In that thread I complained that\\n> we needed some probably-not-very-portable readline functionality\\n> to make this work. However, after experimenting with trying to\\n> shove text back into readline's buffer, I realized that there's\\n> not really any need to do that: we just need to print the waiting\\n> text and then collect another line. (As a bonus, it works the\\n> same even if you turned off readline with -n.)\\n\\nThis is a nice improvement.\\n\\nI tried to torture it with a hex editor, but couldn't get it to break.\\n\\nThere were some weird carriage returns in the patch, but after I\\nremoved them, it applied fine.\\n\\nYours,\\nLaurenz Albe\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 22:05:34 +0100\",\n \"msg_from\": \"Laurenz Albe <laurenz.albe@cybertec.at>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting psql to redisplay command after \\\\e\"\n },\n {\n \"msg_contents\": \"\\nHello Tom,\\n\\n>> psql=> select 1...\\n>> psql-> <prompt>\\n>\\n>> I cannot move back with readline to edit further, I'm stuck there, which\\n>> is strange.\\n>\\n> I don't follow. readline doesn't allow you to edit already-entered lines\\n> today, that is, after typing \\\"select 1<return>\\\" you see\\n>\\n> regression=# select 1\\n> regression-#\\n>\\n> and there isn't any way to move back and edit the already-entered line\\n> within readline.\\n\\nYep.\\n\\nMy point is to possibly not implicitely <return> at the end of \\\\e, but to \\nbehave as if we were moving in history, which allows editing the lines, so \\nthat you would get\\n\\n psql=> select 1<cursor>\\n\\nInstead of the above.\\n\\n> I agree it might be nicer if you could do that, but that's *far* beyond \\n> the scope of this patch. It would take entirely fundamental rethinking \\n> of our use of libreadline, if indeed it's possible at all. I also don't \\n> see how we could have syntax-aware per-line prompts if we were allowing \\n> readline to treat the whole query as one line.\\n\\nI was suggesting something much simpler than rethinking readline handling. \\nDoes not mean that it is a good idea, but while testing the patch I would \\nhave liked the unfinished line to be in the current editing buffer, \\nbasically as if I had not typed <nl>.\\n\\nISTM more natural that \\\\e behaves like history when coming back from \\nediting, i.e. the \\\\e-edited line is set as the current buffer for \\nreadline.\\n\\n> In the larger picture, tinkering with how that works would affect\\n> every psql user at the level of \\\"muscle memory\\\" editing habits,\\n> and I suspect that their reactions would not be uniformly positive.\\n> What I propose here doesn't affect anyone who doesn't use \\\\e at all.\\n> Even for \\\\e users it doesn't have any effect on what you need to type.\\n\\n-- \\nFabien.\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 10:09:14 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting psql to redisplay command after \\\\e\"\n },\n {\n \"msg_contents\": \"Fabien COELHO <coelho@cri.ensmp.fr> writes:\\n> My point is to possibly not implicitely <return> at the end of \\\\e, but to \\n> behave as if we were moving in history, which allows editing the lines, so \\n> that you would get\\n> psql=> select 1<cursor>\\n> Instead of the above.\\n\\n>> I agree it might be nicer if you could do that, but that's *far* beyond \\n>> the scope of this patch. It would take entirely fundamental rethinking \\n>> of our use of libreadline, if indeed it's possible at all. I also don't \\n>> see how we could have syntax-aware per-line prompts if we were allowing \\n>> readline to treat the whole query as one line.\\n\\n> I was suggesting something much simpler than rethinking readline handling. \\n> Does not mean that it is a good idea, but while testing the patch I would \\n> have liked the unfinished line to be in the current editing buffer, \\n> basically as if I had not typed <nl>.\\n\\nI did experiment with trying to do that, but I couldn't get it to work,\\neven with the single version of libreadline I had at hand. It appears\\nto me that readline() starts by clearing the internal buffer. Even if\\nwe could persuade it to work in a particular readline version, I think\\nthe odds of making it portable across all the libreadline and libedit\\nversions that are out there aren't very good. And there's definitely\\nno chance of being remotely compatible with that behavior when using the\\nbare tty drivers (psql -n).\\n\\nIn practice, if you decide that you don't like what you're looking at,\\nyou're probably going to go back into the editor to fix it, ie issue\\nanother \\\\e. So I'm not sure that it's worth such pushups to get the\\ndata into readline's buffer.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 02 Nov 2019 12:21:06 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Getting psql to redisplay command after \\\\e\"\n },\n {\n \"msg_contents\": \"Hello Tom,\\n\\n>> I was suggesting something much simpler than rethinking readline handling.\\n>> Does not mean that it is a good idea, but while testing the patch I would\\n>> have liked the unfinished line to be in the current editing buffer,\\n>> basically as if I had not typed <nl>.\\n>\\n> I did experiment with trying to do that, but I couldn't get it to work, \\n> even with the single version of libreadline I had at hand. It appears \\n> to me that readline() starts by clearing the internal buffer. Even if \\n> we could persuade it to work in a particular readline version, I think \\n> the odds of making it portable across all the libreadline and libedit \\n> versions that are out there aren't very good. And there's definitely no \\n> chance of being remotely compatible with that behavior when using the \\n> bare tty drivers (psql -n).\\n\\nArgh, too bad.\\n\\nThis suggests that readline cannot be used to edit simply a known string? \\n:-( \\\"rl_insert_text\\\" looked promising, although probably not portable, and \\nI tried to make it work without much success anyway. Maybe I'll try to \\ninvestigate more deeply later.\\n\\nNote that replacing the current buffer is exactly what history does. So \\nmaybe that could be exploited by appending the edited line into history \\n(add_history) and tell readline to move there (could not find how to do \\nthat automatically, though)? Or some other history handling…\\n\\n> In practice, if you decide that you don't like what you're looking at,\\n> you're probably going to go back into the editor to fix it, ie issue\\n> another \\\\e. So I'm not sure that it's worth such pushups to get the\\n> data into readline's buffer.\\n\\nFor me \\\\e should mean edit, not edit-and-execute, so I should be back to \\nprompt, which is the crux of my unease with how the feature behaves, \\nbecause it combines two functions that IMO shouldn't.\\n\\nAnyway the submitted patch is an improvement to the current status.\\n\\n-- \\nFabien.\",\n \"msg_date\": \"Sun, 3 Nov 2019 20:58:39 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting psql to redisplay command after \\\\e\"\n },\n {\n \"msg_contents\": \"ne 3. 11. 2019 v 20:58 odesílatel Fabien COELHO <coelho@cri.ensmp.fr>\\nnapsal:\\n\\n>\\n> Hello Tom,\\n>\\n> >> I was suggesting something much simpler than rethinking readline\\n> handling.\\n> >> Does not mean that it is a good idea, but while testing the patch I\\n> would\\n> >> have liked the unfinished line to be in the current editing buffer,\\n> >> basically as if I had not typed <nl>.\\n> >\\n> > I did experiment with trying to do that, but I couldn't get it to work,\\n> > even with the single version of libreadline I had at hand. It appears\\n> > to me that readline() starts by clearing the internal buffer. Even if\\n> > we could persuade it to work in a particular readline version, I think\\n> > the odds of making it portable across all the libreadline and libedit\\n> > versions that are out there aren't very good. And there's definitely no\\n> > chance of being remotely compatible with that behavior when using the\\n> > bare tty drivers (psql -n).\\n>\\n> Argh, too bad.\\n>\\n> This suggests that readline cannot be used to edit simply a known string?\\n> :-( \\\"rl_insert_text\\\" looked promising, although probably not portable, and\\n> I tried to make it work without much success anyway. Maybe I'll try to\\n> investigate more deeply later.\\n>\\n\\npspg uses rl_insert_text\\n\\nhttps://github.com/okbob/pspg/blob/59d115cd55926ab1886fc0dedbbc6ce0577b0cb3/src/pspg.c#L2522\\n\\nPavel\\n\\n\\n> Note that replacing the current buffer is exactly what history does. So\\n> maybe that could be exploited by appending the edited line into history\\n> (add_history) and tell readline to move there (could not find how to do\\n> that automatically, though)? Or some other history handling…\\n>\\n> > In practice, if you decide that you don't like what you're looking at,\\n> > you're probably going to go back into the editor to fix it, ie issue\\n> > another \\\\e. So I'm not sure that it's worth such pushups to get the\\n> > data into readline's buffer.\\n>\\n> For me \\\\e should mean edit, not edit-and-execute, so I should be back to\\n> prompt, which is the crux of my unease with how the feature behaves,\\n> because it combines two functions that IMO shouldn't.\\n>\\n> Anyway the submitted patch is an improvement to the current status.\\n>\\n> --\\n> Fabien.\\n\\nne 3. 11. 2019 v 20:58 odesílatel Fabien COELHO <coelho@cri.ensmp.fr> napsal:\\nHello Tom,\\n\\n>> I was suggesting something much simpler than rethinking readline handling.\\n>> Does not mean that it is a good idea, but while testing the patch I would\\n>> have liked the unfinished line to be in the current editing buffer,\\n>> basically as if I had not typed <nl>.\\n>\\n> I did experiment with trying to do that, but I couldn't get it to work, \\n> even with the single version of libreadline I had at hand. It appears \\n> to me that readline() starts by clearing the internal buffer. Even if \\n> we could persuade it to work in a particular readline version, I think \\n> the odds of making it portable across all the libreadline and libedit \\n> versions that are out there aren't very good. And there's definitely no \\n> chance of being remotely compatible with that behavior when using the \\n> bare tty drivers (psql -n).\\n\\nArgh, too bad.\\n\\nThis suggests that readline cannot be used to edit simply a known string? \\n:-( \\\"rl_insert_text\\\" looked promising, although probably not portable, and \\nI tried to make it work without much success anyway. Maybe I'll try to \\ninvestigate more deeply later.pspg uses rl_insert_text https://github.com/okbob/pspg/blob/59d115cd55926ab1886fc0dedbbc6ce0577b0cb3/src/pspg.c#L2522Pavel \\n\\nNote that replacing the current buffer is exactly what history does. So \\nmaybe that could be exploited by appending the edited line into history \\n(add_history) and tell readline to move there (could not find how to do \\nthat automatically, though)? Or some other history handling…\\n\\n> In practice, if you decide that you don't like what you're looking at,\\n> you're probably going to go back into the editor to fix it, ie issue\\n> another \\\\e. So I'm not sure that it's worth such pushups to get the\\n> data into readline's buffer.\\n\\nFor me \\\\e should mean edit, not edit-and-execute, so I should be back to \\nprompt, which is the crux of my unease with how the feature behaves, \\nbecause it combines two functions that IMO shouldn't.\\n\\nAnyway the submitted patch is an improvement to the current status.\\n\\n-- \\nFabien.\",\n \"msg_date\": \"Sun, 3 Nov 2019 21:08:46 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting psql to redisplay command after \\\\e\"\n },\n {\n \"msg_contents\": \"Fabien COELHO <coelho@cri.ensmp.fr> writes:\\n>> I did experiment with trying to do that, but I couldn't get it to work, \\n>> even with the single version of libreadline I had at hand. It appears \\n>> to me that readline() starts by clearing the internal buffer. Even if \\n>> we could persuade it to work in a particular readline version, I think \\n>> the odds of making it portable across all the libreadline and libedit \\n>> versions that are out there aren't very good. And there's definitely no \\n>> chance of being remotely compatible with that behavior when using the \\n>> bare tty drivers (psql -n).\\n\\n> This suggests that readline cannot be used to edit simply a known string? \\n> :-( \\\"rl_insert_text\\\" looked promising, although probably not portable, and \\n> I tried to make it work without much success anyway. Maybe I'll try to \\n> investigate more deeply later.\\n\\nI think that rl_insert_text and friends can probably only be used from\\nreadline callback functions. So in principle maybe you could make it\\nwork by having an rl_startup_hook that injects text if there is any\\nto inject. There would remain the issues of (a) is it portable across\\na wide range of readline and libedit versions, (b) will the prompting\\nbehavior be nice, and (c) do we really want this to work fundamentally\\ndifferently when readline is turned off?\\n\\n(Pavel's code cited nearby seems to me to be a fine example of what\\nwe do *not* want to do. Getting in bed with libreadline to that\\nextent is inevitably going to misbehave in some places.)\\n\\n>> In practice, if you decide that you don't like what you're looking at,\\n>> you're probably going to go back into the editor to fix it, ie issue\\n>> another \\\\e. So I'm not sure that it's worth such pushups to get the\\n>> data into readline's buffer.\\n\\n> For me \\\\e should mean edit, not edit-and-execute, so I should be back to \\n> prompt, which is the crux of my unease with how the feature behaves, \\n> because it combines two functions that IMO shouldn't.\\n\\nI don't understand that complaint at all. My proposed patch does not\\nchange the behavior to force execution, and it does display a prompt ---\\none that reflects whether you've given a complete command.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 03 Nov 2019 19:35:25 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Getting psql to redisplay command after \\\\e\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Fabien COELHO <coelho@cri.ensmp.fr> writes:\\n>> This suggests that readline cannot be used to edit simply a known string? \\n>> :-( \\\"rl_insert_text\\\" looked promising, although probably not portable, and \\n>> I tried to make it work without much success anyway. Maybe I'll try to \\n>> investigate more deeply later.\\n\\n> I think that rl_insert_text and friends can probably only be used from\\n> readline callback functions. So in principle maybe you could make it\\n> work by having an rl_startup_hook that injects text if there is any\\n> to inject. There would remain the issues of (a) is it portable across\\n> a wide range of readline and libedit versions, (b) will the prompting\\n> behavior be nice, and (c) do we really want this to work fundamentally\\n> differently when readline is turned off?\\n\\nI thought maybe you were going to work on this right away, but since\\nyou haven't, I went ahead and pushed what I had. There's certainly\\nplenty of time to reconsider if you find a better answer.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 22 Nov 2019 17:10:24 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Getting psql to redisplay command after \\\\e\"\n },\n {\n \"msg_contents\": \"\\n>> I think that rl_insert_text and friends can probably only be used from\\n>> readline callback functions. So in principle maybe you could make it\\n>> work by having an rl_startup_hook that injects text if there is any\\n>> to inject. There would remain the issues of (a) is it portable across\\n>> a wide range of readline and libedit versions, (b) will the prompting\\n>> behavior be nice, and (c) do we really want this to work fundamentally\\n>> differently when readline is turned off?\\n>\\n> I thought maybe you were going to work on this right away, but since\\n> you haven't, I went ahead and pushed what I had. There's certainly\\n> plenty of time to reconsider if you find a better answer.\\n\\nIndeeed.\\n\\nI started to play with the startup hook, it kind of worked but not exactly \\nas I wished, and I do not have much time available this round.\\n\\n-- \\nFabien.\\n\\n\\n\",\n \"msg_date\": \"Sat, 23 Nov 2019 19:08:11 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting psql to redisplay command after \\\\e\"\n }\n]","isConversational":false}}},{"rowIdx":49356,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello,\nI think this demonstrates a bug, tested in 11.5:\nhttps://gist.github.com/wkalt/a298fe82c564668c803b3537561e67a0\n\nThe same script succeeds if the index on line 11 is either dropped, made to\nbe non-partial on b, or shifted to a different column (the others are used\nin the partitioning; maybe significant).\n\nThis seems somewhat related to\nhttps://www.postgresql.org/message-id/flat/CAFjFpRc0hqO5hc-%3DFNePygo9j8WTtOvvmysesnN8bfkp3vxHPQ%40mail.gmail.com#00ca695e6c71834622a6e42323f5558a\n.\n\nRegards,\nWyatt\n\nHello,I think this demonstrates a bug, tested in 11.5:https://gist.github.com/wkalt/a298fe82c564668c803b3537561e67a0The same script succeeds if the index on line 11 is either dropped, made to be non-partial on b, or shifted to a different column (the others are used in the partitioning; maybe significant).This seems somewhat related to https://www.postgresql.org/message-id/flat/CAFjFpRc0hqO5hc-%3DFNePygo9j8WTtOvvmysesnN8bfkp3vxHPQ%40mail.gmail.com#00ca695e6c71834622a6e42323f5558a.Regards,Wyatt","msg_date":"Mon, 28 Oct 2019 21:00:24 -0700","msg_from":"Wyatt Alt <wyatt.alt@gmail.com>","msg_from_op":true,"msg_subject":"[BUG] Partition creation fails after dropping a column and adding a\n partial index"},{"msg_contents":"Here's a slightly smaller repro:\nhttps://gist.github.com/wkalt/36720f39c97567fa6cb18cf5c05ac60f\n\nHere's a slightly smaller repro: https://gist.github.com/wkalt/36720f39c97567fa6cb18cf5c05ac60f","msg_date":"Mon, 28 Oct 2019 21:16:10 -0700","msg_from":"Wyatt Alt <wyatt.alt@gmail.com>","msg_from_op":true,"msg_subject":"Re: [BUG] Partition creation fails after dropping a column and adding\n a partial index"},{"msg_contents":"On Mon, Oct 28, 2019 at 09:00:24PM -0700, Wyatt Alt wrote:\n> I think this demonstrates a bug, tested in 11.5:\n> https://gist.github.com/wkalt/a298fe82c564668c803b3537561e67a0\n\nIf this source goes away, then we would lose it. It is always better\nto copy directly the example in the message sent to the mailing lists,\nand here it is:\ncreate table demo (\n id int,\n useless int,\n d timestamp,\n b boolean\n) partition by range (id, d);\nalter table demo drop column useless;\n-- only seems to cause failure when it's a partial index on b.\ncreate index on demo(b) where b = 't';\ncreate table demo_1_20191031 partition of demo for values from (1,\n'2019-10-31') to (1, '2019-11-01');\n\n> The same script succeeds if the index on line 11 is either dropped, made to\n> be non-partial on b, or shifted to a different column (the others are used\n> in the partitioning; maybe significant).\n> \n> This seems somewhat related to\n> https://www.postgresql.org/message-id/flat/CAFjFpRc0hqO5hc-%3DFNePygo9j8WTtOvvmysesnN8bfkp3vxHPQ%40mail.gmail.com#00ca695e6c71834622a6e42323f5558a\n\nYes, something looks wrong with that. I have not looked at it in\ndetails yet though. I'll see about that tomorrow.\n--\nMichael","msg_date":"Tue, 29 Oct 2019 13:16:58 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [BUG] Partition creation fails after dropping a column and\n adding a partial index"},{"msg_contents":"On Tue, Oct 29, 2019 at 01:16:58PM +0900, Michael Paquier wrote:\n> Yes, something looks wrong with that. I have not looked at it in\n> details yet though. I'll see about that tomorrow.\n\nSo.. When building the attribute map for a cloned index (with either\nLIKE during the transformation or for partition indexes), we store\neach attribute number with 0 used for dropped columns. Unfortunately,\nif you look at the way the attribute map is built in this case the\ncode correctly generates the mapping with convert_tuples_by_name_map.\nBut, the length of the mapping used is incorrect as this makes use of \nthe number of attributes for the newly-created child relation, and not\nthe parent which includes the dropped column in its count. So the\nanswer is simply to use the parent as reference for the mapping\nlength.\n\nThe patch is rather simple as per the attached, with extended\nregression tests included. I have not checked on back-branches yet,\nbut that's visibly wrong since 8b08f7d down to v11 (will do that when\nback-patching).\n\nThere could be a point in changing convert_tuples_by_name_map & co so\nas they return the length of the map on top of the map to avoid such\nmistakes in the future. That's a more invasive patch not really\nadapted for a back-patch, but we could do that on HEAD once this bug\nis fixed. I have also checked other calls of this API and the\nhandling is done correctly.\n\nWyatt, what do you think?\n--\nMichael","msg_date":"Thu, 31 Oct 2019 13:45:07 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [BUG] Partition creation fails after dropping a column and\n adding a partial index"},{"msg_contents":"On Thu, Oct 31, 2019 at 9:45 AM Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Tue, Oct 29, 2019 at 01:16:58PM +0900, Michael Paquier wrote:\n> > Yes, something looks wrong with that. I have not looked at it in\n> > details yet though. I'll see about that tomorrow.\n>\n> So.. When building the attribute map for a cloned index (with either\n> LIKE during the transformation or for partition indexes), we store\n> each attribute number with 0 used for dropped columns. Unfortunately,\n> if you look at the way the attribute map is built in this case the\n> code correctly generates the mapping with convert_tuples_by_name_map.\n> But, the length of the mapping used is incorrect as this makes use of\n> the number of attributes for the newly-created child relation, and not\n> the parent which includes the dropped column in its count. So the\n> answer is simply to use the parent as reference for the mapping\n> length.\n>\n> The patch is rather simple as per the attached, with extended\n> regression tests included. I have not checked on back-branches yet,\n> but that's visibly wrong since 8b08f7d down to v11 (will do that when\n> back-patching).\n>\n> There could be a point in changing convert_tuples_by_name_map & co so\n> as they return the length of the map on top of the map to avoid such\n> mistakes in the future. That's a more invasive patch not really\n> adapted for a back-patch, but we could do that on HEAD once this bug\n> is fixed. I have also checked other calls of this API and the\n> handling is done correctly.\n>\n> The patch works for me on master and on 12. I have rebased the patch for\nVersion 11.\n\n\n> Wyatt, what do you think?\n> --\n> Michael\n>\n\n\n-- \nIbrar Ahmed","msg_date":"Thu, 31 Oct 2019 19:54:25 +0500","msg_from":"Ibrar Ahmed <ibrar.ahmad@gmail.com>","msg_from_op":false,"msg_subject":"Re: [BUG] Partition creation fails after dropping a column and adding\n a partial index"},{"msg_contents":"On Thu, Oct 31, 2019 at 1:45 PM Michael Paquier <michael@paquier.xyz> wrote:\n> On Tue, Oct 29, 2019 at 01:16:58PM +0900, Michael Paquier wrote:\n> > Yes, something looks wrong with that. I have not looked at it in\n> > details yet though. I'll see about that tomorrow.\n>\n> So.. When building the attribute map for a cloned index (with either\n> LIKE during the transformation or for partition indexes), we store\n> each attribute number with 0 used for dropped columns. Unfortunately,\n> if you look at the way the attribute map is built in this case the\n> code correctly generates the mapping with convert_tuples_by_name_map.\n> But, the length of the mapping used is incorrect as this makes use of\n> the number of attributes for the newly-created child relation, and not\n> the parent which includes the dropped column in its count. So the\n> answer is simply to use the parent as reference for the mapping\n> length.\n>\n> The patch is rather simple as per the attached, with extended\n> regression tests included. I have not checked on back-branches yet,\n> but that's visibly wrong since 8b08f7d down to v11 (will do that when\n> back-patching).\n\nThe patch looks correct and applies to both v12 and v11.\n\n> There could be a point in changing convert_tuples_by_name_map & co so\n> as they return the length of the map on top of the map to avoid such\n> mistakes in the future. That's a more invasive patch not really\n> adapted for a back-patch, but we could do that on HEAD once this bug\n> is fixed. I have also checked other calls of this API and the\n> handling is done correctly.\n\nI've been bitten by this logical error when deciding what length to\nuse for the map, so seems like a good idea.\n\nThanks,\nAmit\n\n\n","msg_date":"Fri, 1 Nov 2019 09:58:26 +0900","msg_from":"Amit Langote <amitlangote09@gmail.com>","msg_from_op":false,"msg_subject":"Re: [BUG] Partition creation fails after dropping a column and adding\n a partial index"},{"msg_contents":"On Fri, Nov 01, 2019 at 09:58:26AM +0900, Amit Langote wrote:\n> On Thu, Oct 31, 2019 at 1:45 PM Michael Paquier <michael@paquier.xyz> wrote:\n>> The patch is rather simple as per the attached, with extended\n>> regression tests included. I have not checked on back-branches yet,\n>> but that's visibly wrong since 8b08f7d down to v11 (will do that when\n>> back-patching).\n> \n> The patch looks correct and applies to both v12 and v11.\n\nThanks for the review, committed down to v11. The version for v11 had\na couple of conflicts actually.\n\n>> There could be a point in changing convert_tuples_by_name_map & co so\n>> as they return the length of the map on top of the map to avoid such\n>> mistakes in the future. That's a more invasive patch not really\n>> adapted for a back-patch, but we could do that on HEAD once this bug\n>> is fixed. I have also checked other calls of this API and the\n>> handling is done correctly.\n> \n> I've been bitten by this logical error when deciding what length to\n> use for the map, so seems like a good idea.\n\nOkay, let's see about that then.\n--\nMichael","msg_date":"Sat, 2 Nov 2019 14:20:01 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [BUG] Partition creation fails after dropping a column and\n adding a partial index"}],"string":"[\n {\n \"msg_contents\": \"Hello,\\nI think this demonstrates a bug, tested in 11.5:\\nhttps://gist.github.com/wkalt/a298fe82c564668c803b3537561e67a0\\n\\nThe same script succeeds if the index on line 11 is either dropped, made to\\nbe non-partial on b, or shifted to a different column (the others are used\\nin the partitioning; maybe significant).\\n\\nThis seems somewhat related to\\nhttps://www.postgresql.org/message-id/flat/CAFjFpRc0hqO5hc-%3DFNePygo9j8WTtOvvmysesnN8bfkp3vxHPQ%40mail.gmail.com#00ca695e6c71834622a6e42323f5558a\\n.\\n\\nRegards,\\nWyatt\\n\\nHello,I think this demonstrates a bug, tested in 11.5:https://gist.github.com/wkalt/a298fe82c564668c803b3537561e67a0The same script succeeds if the index on line 11 is either dropped, made to be non-partial on b, or shifted to a different column (the others are used in the partitioning; maybe significant).This seems somewhat related to https://www.postgresql.org/message-id/flat/CAFjFpRc0hqO5hc-%3DFNePygo9j8WTtOvvmysesnN8bfkp3vxHPQ%40mail.gmail.com#00ca695e6c71834622a6e42323f5558a.Regards,Wyatt\",\n \"msg_date\": \"Mon, 28 Oct 2019 21:00:24 -0700\",\n \"msg_from\": \"Wyatt Alt <wyatt.alt@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[BUG] Partition creation fails after dropping a column and adding a\\n partial index\"\n },\n {\n \"msg_contents\": \"Here's a slightly smaller repro:\\nhttps://gist.github.com/wkalt/36720f39c97567fa6cb18cf5c05ac60f\\n\\nHere's a slightly smaller repro: https://gist.github.com/wkalt/36720f39c97567fa6cb18cf5c05ac60f\",\n \"msg_date\": \"Mon, 28 Oct 2019 21:16:10 -0700\",\n \"msg_from\": \"Wyatt Alt <wyatt.alt@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [BUG] Partition creation fails after dropping a column and adding\\n a partial index\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 28, 2019 at 09:00:24PM -0700, Wyatt Alt wrote:\\n> I think this demonstrates a bug, tested in 11.5:\\n> https://gist.github.com/wkalt/a298fe82c564668c803b3537561e67a0\\n\\nIf this source goes away, then we would lose it. It is always better\\nto copy directly the example in the message sent to the mailing lists,\\nand here it is:\\ncreate table demo (\\n id int,\\n useless int,\\n d timestamp,\\n b boolean\\n) partition by range (id, d);\\nalter table demo drop column useless;\\n-- only seems to cause failure when it's a partial index on b.\\ncreate index on demo(b) where b = 't';\\ncreate table demo_1_20191031 partition of demo for values from (1,\\n'2019-10-31') to (1, '2019-11-01');\\n\\n> The same script succeeds if the index on line 11 is either dropped, made to\\n> be non-partial on b, or shifted to a different column (the others are used\\n> in the partitioning; maybe significant).\\n> \\n> This seems somewhat related to\\n> https://www.postgresql.org/message-id/flat/CAFjFpRc0hqO5hc-%3DFNePygo9j8WTtOvvmysesnN8bfkp3vxHPQ%40mail.gmail.com#00ca695e6c71834622a6e42323f5558a\\n\\nYes, something looks wrong with that. I have not looked at it in\\ndetails yet though. I'll see about that tomorrow.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 29 Oct 2019 13:16:58 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG] Partition creation fails after dropping a column and\\n adding a partial index\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 29, 2019 at 01:16:58PM +0900, Michael Paquier wrote:\\n> Yes, something looks wrong with that. I have not looked at it in\\n> details yet though. I'll see about that tomorrow.\\n\\nSo.. When building the attribute map for a cloned index (with either\\nLIKE during the transformation or for partition indexes), we store\\neach attribute number with 0 used for dropped columns. Unfortunately,\\nif you look at the way the attribute map is built in this case the\\ncode correctly generates the mapping with convert_tuples_by_name_map.\\nBut, the length of the mapping used is incorrect as this makes use of \\nthe number of attributes for the newly-created child relation, and not\\nthe parent which includes the dropped column in its count. So the\\nanswer is simply to use the parent as reference for the mapping\\nlength.\\n\\nThe patch is rather simple as per the attached, with extended\\nregression tests included. I have not checked on back-branches yet,\\nbut that's visibly wrong since 8b08f7d down to v11 (will do that when\\nback-patching).\\n\\nThere could be a point in changing convert_tuples_by_name_map & co so\\nas they return the length of the map on top of the map to avoid such\\nmistakes in the future. That's a more invasive patch not really\\nadapted for a back-patch, but we could do that on HEAD once this bug\\nis fixed. I have also checked other calls of this API and the\\nhandling is done correctly.\\n\\nWyatt, what do you think?\\n--\\nMichael\",\n \"msg_date\": \"Thu, 31 Oct 2019 13:45:07 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG] Partition creation fails after dropping a column and\\n adding a partial index\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 9:45 AM Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Tue, Oct 29, 2019 at 01:16:58PM +0900, Michael Paquier wrote:\\n> > Yes, something looks wrong with that. I have not looked at it in\\n> > details yet though. I'll see about that tomorrow.\\n>\\n> So.. When building the attribute map for a cloned index (with either\\n> LIKE during the transformation or for partition indexes), we store\\n> each attribute number with 0 used for dropped columns. Unfortunately,\\n> if you look at the way the attribute map is built in this case the\\n> code correctly generates the mapping with convert_tuples_by_name_map.\\n> But, the length of the mapping used is incorrect as this makes use of\\n> the number of attributes for the newly-created child relation, and not\\n> the parent which includes the dropped column in its count. So the\\n> answer is simply to use the parent as reference for the mapping\\n> length.\\n>\\n> The patch is rather simple as per the attached, with extended\\n> regression tests included. I have not checked on back-branches yet,\\n> but that's visibly wrong since 8b08f7d down to v11 (will do that when\\n> back-patching).\\n>\\n> There could be a point in changing convert_tuples_by_name_map & co so\\n> as they return the length of the map on top of the map to avoid such\\n> mistakes in the future. That's a more invasive patch not really\\n> adapted for a back-patch, but we could do that on HEAD once this bug\\n> is fixed. I have also checked other calls of this API and the\\n> handling is done correctly.\\n>\\n> The patch works for me on master and on 12. I have rebased the patch for\\nVersion 11.\\n\\n\\n> Wyatt, what do you think?\\n> --\\n> Michael\\n>\\n\\n\\n-- \\nIbrar Ahmed\",\n \"msg_date\": \"Thu, 31 Oct 2019 19:54:25 +0500\",\n \"msg_from\": \"Ibrar Ahmed <ibrar.ahmad@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG] Partition creation fails after dropping a column and adding\\n a partial index\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 1:45 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> On Tue, Oct 29, 2019 at 01:16:58PM +0900, Michael Paquier wrote:\\n> > Yes, something looks wrong with that. I have not looked at it in\\n> > details yet though. I'll see about that tomorrow.\\n>\\n> So.. When building the attribute map for a cloned index (with either\\n> LIKE during the transformation or for partition indexes), we store\\n> each attribute number with 0 used for dropped columns. Unfortunately,\\n> if you look at the way the attribute map is built in this case the\\n> code correctly generates the mapping with convert_tuples_by_name_map.\\n> But, the length of the mapping used is incorrect as this makes use of\\n> the number of attributes for the newly-created child relation, and not\\n> the parent which includes the dropped column in its count. So the\\n> answer is simply to use the parent as reference for the mapping\\n> length.\\n>\\n> The patch is rather simple as per the attached, with extended\\n> regression tests included. I have not checked on back-branches yet,\\n> but that's visibly wrong since 8b08f7d down to v11 (will do that when\\n> back-patching).\\n\\nThe patch looks correct and applies to both v12 and v11.\\n\\n> There could be a point in changing convert_tuples_by_name_map & co so\\n> as they return the length of the map on top of the map to avoid such\\n> mistakes in the future. That's a more invasive patch not really\\n> adapted for a back-patch, but we could do that on HEAD once this bug\\n> is fixed. I have also checked other calls of this API and the\\n> handling is done correctly.\\n\\nI've been bitten by this logical error when deciding what length to\\nuse for the map, so seems like a good idea.\\n\\nThanks,\\nAmit\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 09:58:26 +0900\",\n \"msg_from\": \"Amit Langote <amitlangote09@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG] Partition creation fails after dropping a column and adding\\n a partial index\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 01, 2019 at 09:58:26AM +0900, Amit Langote wrote:\\n> On Thu, Oct 31, 2019 at 1:45 PM Michael Paquier <michael@paquier.xyz> wrote:\\n>> The patch is rather simple as per the attached, with extended\\n>> regression tests included. I have not checked on back-branches yet,\\n>> but that's visibly wrong since 8b08f7d down to v11 (will do that when\\n>> back-patching).\\n> \\n> The patch looks correct and applies to both v12 and v11.\\n\\nThanks for the review, committed down to v11. The version for v11 had\\na couple of conflicts actually.\\n\\n>> There could be a point in changing convert_tuples_by_name_map & co so\\n>> as they return the length of the map on top of the map to avoid such\\n>> mistakes in the future. That's a more invasive patch not really\\n>> adapted for a back-patch, but we could do that on HEAD once this bug\\n>> is fixed. I have also checked other calls of this API and the\\n>> handling is done correctly.\\n> \\n> I've been bitten by this logical error when deciding what length to\\n> use for the map, so seems like a good idea.\\n\\nOkay, let's see about that then.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 2 Nov 2019 14:20:01 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [BUG] Partition creation fails after dropping a column and\\n adding a partial index\"\n }\n]","isConversational":false}}},{"rowIdx":49357,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"The cache_plan argument to ri_PlanCheck has not been used since\ne8c9fd5fdf768323911f7088e8287f63b513c3c6. I propose to remove it.\n\nThat commit said \"I left it alone in case there is any future need for \nit\" but there hasn't been a need in 7 years, and I find it confusing to \nhave an unused function argument without a clear purpose. It would \ntrivial to put it back if needed.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Tue, 29 Oct 2019 10:21:32 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Remove unused function argument"},{"msg_contents":"On Tue, Oct 29, 2019 at 2:51 PM Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n>\n> The cache_plan argument to ri_PlanCheck has not been used since\n> e8c9fd5fdf768323911f7088e8287f63b513c3c6. I propose to remove it.\n>\n> That commit said \"I left it alone in case there is any future need for\n> it\" but there hasn't been a need in 7 years, and I find it confusing to\n> have an unused function argument without a clear purpose. It would\n> trivial to put it back if needed.\n>\nCode changes looks fine to me.\n\nRegards,\nVignesh\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 30 Oct 2019 11:21:59 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":false,"msg_subject":"Re: Remove unused function argument"},{"msg_contents":"On 2019-10-30 06:51, vignesh C wrote:\n> On Tue, Oct 29, 2019 at 2:51 PM Peter Eisentraut\n> <peter.eisentraut@2ndquadrant.com> wrote:\n>>\n>> The cache_plan argument to ri_PlanCheck has not been used since\n>> e8c9fd5fdf768323911f7088e8287f63b513c3c6. I propose to remove it.\n>>\n>> That commit said \"I left it alone in case there is any future need for\n>> it\" but there hasn't been a need in 7 years, and I find it confusing to\n>> have an unused function argument without a clear purpose. It would\n>> trivial to put it back if needed.\n>>\n> Code changes looks fine to me.\n\npushed, thanks\n\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Wed, 6 Nov 2019 08:20:20 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Remove unused function argument"}],"string":"[\n {\n \"msg_contents\": \"The cache_plan argument to ri_PlanCheck has not been used since\\ne8c9fd5fdf768323911f7088e8287f63b513c3c6. I propose to remove it.\\n\\nThat commit said \\\"I left it alone in case there is any future need for \\nit\\\" but there hasn't been a need in 7 years, and I find it confusing to \\nhave an unused function argument without a clear purpose. It would \\ntrivial to put it back if needed.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Tue, 29 Oct 2019 10:21:32 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Remove unused function argument\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 29, 2019 at 2:51 PM Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n>\\n> The cache_plan argument to ri_PlanCheck has not been used since\\n> e8c9fd5fdf768323911f7088e8287f63b513c3c6. I propose to remove it.\\n>\\n> That commit said \\\"I left it alone in case there is any future need for\\n> it\\\" but there hasn't been a need in 7 years, and I find it confusing to\\n> have an unused function argument without a clear purpose. It would\\n> trivial to put it back if needed.\\n>\\nCode changes looks fine to me.\\n\\nRegards,\\nVignesh\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 11:21:59 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove unused function argument\"\n },\n {\n \"msg_contents\": \"On 2019-10-30 06:51, vignesh C wrote:\\n> On Tue, Oct 29, 2019 at 2:51 PM Peter Eisentraut\\n> <peter.eisentraut@2ndquadrant.com> wrote:\\n>>\\n>> The cache_plan argument to ri_PlanCheck has not been used since\\n>> e8c9fd5fdf768323911f7088e8287f63b513c3c6. I propose to remove it.\\n>>\\n>> That commit said \\\"I left it alone in case there is any future need for\\n>> it\\\" but there hasn't been a need in 7 years, and I find it confusing to\\n>> have an unused function argument without a clear purpose. It would\\n>> trivial to put it back if needed.\\n>>\\n> Code changes looks fine to me.\\n\\npushed, thanks\\n\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Nov 2019 08:20:20 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Remove unused function argument\"\n }\n]","isConversational":false}}},{"rowIdx":49358,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Continuing the discussion from [0] and [1], here is a patch that \nautomates the process of updating Unicode derived files. Summary:\n\n- Edit UNICODE_VERSION and/or CLDR_VERSION in src/Makefile.global.in\n- Run make update-unicode\n- Commit\n\nI have added that to the release checklist in RELEASE_NOTES.\n\nThis also includes the script used in [0] that was not committed at that \ntime. Other than that, this just refactors existing build code.\n\nOpen questions that are currently not handled consistently:\n\n- Should the downloaded files be listed in .gitignore?\n- Should the downloaded files be cleaned by make clean (or distclean or \nmaintainer-clean or none)?\n- Should the generated files be excluded from pgindent? Currently, the \ngenerated files will not pass pgindent unchanged, so that could cause \nannoying whitespace battles when these files are updated and re-indented \naround release time.\n\n\n[0]: \nhttps://www.postgresql.org/message-id/flat/bbb19114-af1e-513b-08a9-61272794bd5c%402ndquadrant.com\n[1]: \nhttps://www.postgresql.org/message-id/flat/77f69366-ca31-6437-079f-47fce69bae1b%402ndquadrant.com\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Tue, 29 Oct 2019 11:06:05 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Add support for automatically updating Unicode derived files"},{"msg_contents":"On Tue, Oct 29, 2019 at 6:06 AM Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n>\n> Continuing the discussion from [0] and [1], here is a patch that\n> automates the process of updating Unicode derived files. Summary:\n>\n> - Edit UNICODE_VERSION and/or CLDR_VERSION in src/Makefile.global.in\n> - Run make update-unicode\n> - Commit\n\nHi Peter,\n\nI gave \"make update-unicode\" a try. It's unclear to me what the state\nof the build tree should be when a maintainer runs this, so I'll just\nreport what happens when running naively (on MacOS).\n\nAfter only running configure, \"make update-unicode\" gives this error\nat normalization-check:\n\nld: library not found for -lpgcommon\nclang: error: linker command failed with exit code 1 (use -v to see invocation)\n\nAfter commenting that out, the next command \"$(MAKE) -C\ncontrib/unaccent $@\" failed, seemingly because $(PYTHON) is empty\nunless --with-python was specified at configure time.\n\n> Open questions that are currently not handled consistently:\n>\n> - Should the downloaded files be listed in .gitignore?\n\nThese files are transient byproducts of a build, and we don't want\nthem committed, so they seem like a normal candidate for .gitignore.\n\n> - Should the downloaded files be cleaned by make clean (or distclean or\n> maintainer-clean or none)?\n\nIt seems one would want to make clean without removing these files,\nand maintainer clean is for removing things that are preserved in\ndistribution tarballs. So I would go with distclean.\n\n> - Should the generated files be excluded from pgindent? Currently, the\n> generated files will not pass pgindent unchanged, so that could cause\n> annoying whitespace battles when these files are updated and re-indented\n> around release time.\n\nI see what you mean in the norm table header. I think generated files\nshould not be pgindent'd, since creating them is already a consistent,\nmechanical process, and their presentation is not as important as\nother code.\n\nOther comments:\n\n+print \"/* generated by\nsrc/common/unicode/generate-unicode_combining_table.pl, do not edit\n*/\\n\\n\";\n\nI would print out the full boilerplate like for other generated headers.\n\nLastly, src/common/unicode/README is outdated (and possibly no longer\nuseful at all?).\n\n-- \nJohn Naylor https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 19 Dec 2019 17:48:39 -0500","msg_from":"John Naylor <john.naylor@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Add support for automatically updating Unicode derived files"},{"msg_contents":"On 2019-12-19 23:48, John Naylor wrote:\n> I gave \"make update-unicode\" a try. It's unclear to me what the state\n> of the build tree should be when a maintainer runs this, so I'll just\n> report what happens when running naively (on MacOS).\n\nYeah, that wasn't fully thought through, it appears.\n\n> After only running configure, \"make update-unicode\" gives this error\n> at normalization-check:\n> \n> ld: library not found for -lpgcommon\n> clang: error: linker command failed with exit code 1 (use -v to see invocation)\n\nFixed by adding more make dependencies.\n\n> After commenting that out, the next command \"$(MAKE) -C\n> contrib/unaccent $@\" failed, seemingly because $(PYTHON) is empty\n> unless --with-python was specified at configure time.\n\nI'm not sure whether that's worth addressing.\n\n>> Open questions that are currently not handled consistently:\n>>\n>> - Should the downloaded files be listed in .gitignore?\n> \n> These files are transient byproducts of a build, and we don't want\n> them committed, so they seem like a normal candidate for .gitignore.\n\nOK done\n\n>> - Should the downloaded files be cleaned by make clean (or distclean or\n>> maintainer-clean or none)?\n> \n> It seems one would want to make clean without removing these files,\n> and maintainer clean is for removing things that are preserved in\n> distribution tarballs. So I would go with distclean.\n\nalso done\n\n>> - Should the generated files be excluded from pgindent? Currently, the\n>> generated files will not pass pgindent unchanged, so that could cause\n>> annoying whitespace battles when these files are updated and re-indented\n>> around release time.\n> \n> I see what you mean in the norm table header. I think generated files\n> should not be pgindent'd, since creating them is already a consistent,\n> mechanical process, and their presentation is not as important as\n> other code.\n\nI've left it alone for now because the little indentation problem \ncurrently present might actually go away with my Unicode normalization \nsupport patch.\n\n> Other comments:\n> \n> +print \"/* generated by\n> src/common/unicode/generate-unicode_combining_table.pl, do not edit\n> */\\n\\n\";\n> \n> I would print out the full boilerplate like for other generated headers.\n\nHmm, you are probably comparing with \nsrc/common/unicode/generate-unicode_norm_table.pl, but other file \ngenerating scripts around the tree print out a small header in the style \nthat I have. I'd rather adjust the output of \ngenerate-unicode_norm_table.pl to match those. (It's also not quite \ncorrect to make copyright claims about automatically generated output.)\n\n> Lastly, src/common/unicode/README is outdated (and possibly no longer\n> useful at all?).\n\nupdated\n\nnew patch attached\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 26 Dec 2019 19:38:54 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Add support for automatically updating Unicode derived files"},{"msg_contents":"On Thu, Dec 26, 2019 at 12:39 PM Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n>\n> On 2019-12-19 23:48, John Naylor wrote:\n> > I would print out the full boilerplate like for other generated headers.\n>\n> Hmm, you are probably comparing with\n> src/common/unicode/generate-unicode_norm_table.pl, but other file\n> generating scripts around the tree print out a small header in the style\n> that I have. I'd rather adjust the output of\n> generate-unicode_norm_table.pl to match those. (It's also not quite\n> correct to make copyright claims about automatically generated output.)\n\nHmm, the scripts I'm most familiar with have full headers. Your point\nabout copyright makes sense, and using smaller file headers would aid\nreadability of the scripts, but I also see how others may feel\ndifferently.\n\nv2 looks good to me, marked ready for committer.\n\n--\nJohn Naylor https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 3 Jan 2020 08:13:53 -0600","msg_from":"John Naylor <john.naylor@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Add support for automatically updating Unicode derived files"},{"msg_contents":"On 2020-01-03 15:13, John Naylor wrote:\n> On Thu, Dec 26, 2019 at 12:39 PM Peter Eisentraut\n> <peter.eisentraut@2ndquadrant.com> wrote:\n>>\n>> On 2019-12-19 23:48, John Naylor wrote:\n>>> I would print out the full boilerplate like for other generated headers.\n>>\n>> Hmm, you are probably comparing with\n>> src/common/unicode/generate-unicode_norm_table.pl, but other file\n>> generating scripts around the tree print out a small header in the style\n>> that I have. I'd rather adjust the output of\n>> generate-unicode_norm_table.pl to match those. (It's also not quite\n>> correct to make copyright claims about automatically generated output.)\n> \n> Hmm, the scripts I'm most familiar with have full headers. Your point\n> about copyright makes sense, and using smaller file headers would aid\n> readability of the scripts, but I also see how others may feel\n> differently.\n> \n> v2 looks good to me, marked ready for committer.\n\nCommitted, thanks.\n\nI have added a little tweak so that it works also without --with-python, \nto avoid gratuitous annoyances.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 9 Jan 2020 10:16:19 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Add support for automatically updating Unicode derived files"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> Committed, thanks.\n\nThis patch is making src/tools/pginclude/headerscheck unhappy:\n\n./src/include/common/unicode_combining_table.h:3: error: array type has incomplete element type\n\nI guess that header needs another #include, or else you need to\nmove some declarations around.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 14 Jan 2020 19:37:30 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Add support for automatically updating Unicode derived files"},{"msg_contents":"On 2020-01-15 01:37, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n>> Committed, thanks.\n> \n> This patch is making src/tools/pginclude/headerscheck unhappy:\n> \n> ./src/include/common/unicode_combining_table.h:3: error: array type has incomplete element type\n> \n> I guess that header needs another #include, or else you need to\n> move some declarations around.\n\nHmm, this file is only meant to be included inside one particular \nfunction. Making it standalone includable would seem to be unnecessary. \n What should we do?\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Mon, 20 Jan 2020 09:59:34 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Add support for automatically updating Unicode derived files"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> On 2020-01-15 01:37, Tom Lane wrote:\n>> This patch is making src/tools/pginclude/headerscheck unhappy:\n>> ./src/include/common/unicode_combining_table.h:3: error: array type has incomplete element type\n\n> Hmm, this file is only meant to be included inside one particular \n> function. Making it standalone includable would seem to be unnecessary. \n> What should we do?\n\nWell, we could make it a documented exception in headerscheck and\ncpluspluscheck.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 20 Jan 2020 10:43:32 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Add support for automatically updating Unicode derived files"},{"msg_contents":"On 2020-01-20 16:43, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n>> On 2020-01-15 01:37, Tom Lane wrote:\n>>> This patch is making src/tools/pginclude/headerscheck unhappy:\n>>> ./src/include/common/unicode_combining_table.h:3: error: array type has incomplete element type\n> \n>> Hmm, this file is only meant to be included inside one particular\n>> function. Making it standalone includable would seem to be unnecessary.\n>> What should we do?\n> \n> Well, we could make it a documented exception in headerscheck and\n> cpluspluscheck.\n\nOK, done.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 24 Jan 2020 12:25:06 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Add support for automatically updating Unicode derived files"},{"msg_contents":"I have committed the first Unicode data update using this new \"make \nupdate-unicode\" facility.\n\nCLDR is released regularly every 6 months, so around this time every \nyear would be the appropriate time to pull in the latest updates in \npreparation for our own release.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 24 Apr 2020 10:01:00 +0200","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Add support for automatically updating Unicode derived files"}],"string":"[\n {\n \"msg_contents\": \"Continuing the discussion from [0] and [1], here is a patch that \\nautomates the process of updating Unicode derived files. Summary:\\n\\n- Edit UNICODE_VERSION and/or CLDR_VERSION in src/Makefile.global.in\\n- Run make update-unicode\\n- Commit\\n\\nI have added that to the release checklist in RELEASE_NOTES.\\n\\nThis also includes the script used in [0] that was not committed at that \\ntime. Other than that, this just refactors existing build code.\\n\\nOpen questions that are currently not handled consistently:\\n\\n- Should the downloaded files be listed in .gitignore?\\n- Should the downloaded files be cleaned by make clean (or distclean or \\nmaintainer-clean or none)?\\n- Should the generated files be excluded from pgindent? Currently, the \\ngenerated files will not pass pgindent unchanged, so that could cause \\nannoying whitespace battles when these files are updated and re-indented \\naround release time.\\n\\n\\n[0]: \\nhttps://www.postgresql.org/message-id/flat/bbb19114-af1e-513b-08a9-61272794bd5c%402ndquadrant.com\\n[1]: \\nhttps://www.postgresql.org/message-id/flat/77f69366-ca31-6437-079f-47fce69bae1b%402ndquadrant.com\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Tue, 29 Oct 2019 11:06:05 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Add support for automatically updating Unicode derived files\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 29, 2019 at 6:06 AM Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n>\\n> Continuing the discussion from [0] and [1], here is a patch that\\n> automates the process of updating Unicode derived files. Summary:\\n>\\n> - Edit UNICODE_VERSION and/or CLDR_VERSION in src/Makefile.global.in\\n> - Run make update-unicode\\n> - Commit\\n\\nHi Peter,\\n\\nI gave \\\"make update-unicode\\\" a try. It's unclear to me what the state\\nof the build tree should be when a maintainer runs this, so I'll just\\nreport what happens when running naively (on MacOS).\\n\\nAfter only running configure, \\\"make update-unicode\\\" gives this error\\nat normalization-check:\\n\\nld: library not found for -lpgcommon\\nclang: error: linker command failed with exit code 1 (use -v to see invocation)\\n\\nAfter commenting that out, the next command \\\"$(MAKE) -C\\ncontrib/unaccent $@\\\" failed, seemingly because $(PYTHON) is empty\\nunless --with-python was specified at configure time.\\n\\n> Open questions that are currently not handled consistently:\\n>\\n> - Should the downloaded files be listed in .gitignore?\\n\\nThese files are transient byproducts of a build, and we don't want\\nthem committed, so they seem like a normal candidate for .gitignore.\\n\\n> - Should the downloaded files be cleaned by make clean (or distclean or\\n> maintainer-clean or none)?\\n\\nIt seems one would want to make clean without removing these files,\\nand maintainer clean is for removing things that are preserved in\\ndistribution tarballs. So I would go with distclean.\\n\\n> - Should the generated files be excluded from pgindent? Currently, the\\n> generated files will not pass pgindent unchanged, so that could cause\\n> annoying whitespace battles when these files are updated and re-indented\\n> around release time.\\n\\nI see what you mean in the norm table header. I think generated files\\nshould not be pgindent'd, since creating them is already a consistent,\\nmechanical process, and their presentation is not as important as\\nother code.\\n\\nOther comments:\\n\\n+print \\\"/* generated by\\nsrc/common/unicode/generate-unicode_combining_table.pl, do not edit\\n*/\\\\n\\\\n\\\";\\n\\nI would print out the full boilerplate like for other generated headers.\\n\\nLastly, src/common/unicode/README is outdated (and possibly no longer\\nuseful at all?).\\n\\n-- \\nJohn Naylor https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 19 Dec 2019 17:48:39 -0500\",\n \"msg_from\": \"John Naylor <john.naylor@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add support for automatically updating Unicode derived files\"\n },\n {\n \"msg_contents\": \"On 2019-12-19 23:48, John Naylor wrote:\\n> I gave \\\"make update-unicode\\\" a try. It's unclear to me what the state\\n> of the build tree should be when a maintainer runs this, so I'll just\\n> report what happens when running naively (on MacOS).\\n\\nYeah, that wasn't fully thought through, it appears.\\n\\n> After only running configure, \\\"make update-unicode\\\" gives this error\\n> at normalization-check:\\n> \\n> ld: library not found for -lpgcommon\\n> clang: error: linker command failed with exit code 1 (use -v to see invocation)\\n\\nFixed by adding more make dependencies.\\n\\n> After commenting that out, the next command \\\"$(MAKE) -C\\n> contrib/unaccent $@\\\" failed, seemingly because $(PYTHON) is empty\\n> unless --with-python was specified at configure time.\\n\\nI'm not sure whether that's worth addressing.\\n\\n>> Open questions that are currently not handled consistently:\\n>>\\n>> - Should the downloaded files be listed in .gitignore?\\n> \\n> These files are transient byproducts of a build, and we don't want\\n> them committed, so they seem like a normal candidate for .gitignore.\\n\\nOK done\\n\\n>> - Should the downloaded files be cleaned by make clean (or distclean or\\n>> maintainer-clean or none)?\\n> \\n> It seems one would want to make clean without removing these files,\\n> and maintainer clean is for removing things that are preserved in\\n> distribution tarballs. So I would go with distclean.\\n\\nalso done\\n\\n>> - Should the generated files be excluded from pgindent? Currently, the\\n>> generated files will not pass pgindent unchanged, so that could cause\\n>> annoying whitespace battles when these files are updated and re-indented\\n>> around release time.\\n> \\n> I see what you mean in the norm table header. I think generated files\\n> should not be pgindent'd, since creating them is already a consistent,\\n> mechanical process, and their presentation is not as important as\\n> other code.\\n\\nI've left it alone for now because the little indentation problem \\ncurrently present might actually go away with my Unicode normalization \\nsupport patch.\\n\\n> Other comments:\\n> \\n> +print \\\"/* generated by\\n> src/common/unicode/generate-unicode_combining_table.pl, do not edit\\n> */\\\\n\\\\n\\\";\\n> \\n> I would print out the full boilerplate like for other generated headers.\\n\\nHmm, you are probably comparing with \\nsrc/common/unicode/generate-unicode_norm_table.pl, but other file \\ngenerating scripts around the tree print out a small header in the style \\nthat I have. I'd rather adjust the output of \\ngenerate-unicode_norm_table.pl to match those. (It's also not quite \\ncorrect to make copyright claims about automatically generated output.)\\n\\n> Lastly, src/common/unicode/README is outdated (and possibly no longer\\n> useful at all?).\\n\\nupdated\\n\\nnew patch attached\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 26 Dec 2019 19:38:54 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add support for automatically updating Unicode derived files\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 26, 2019 at 12:39 PM Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n>\\n> On 2019-12-19 23:48, John Naylor wrote:\\n> > I would print out the full boilerplate like for other generated headers.\\n>\\n> Hmm, you are probably comparing with\\n> src/common/unicode/generate-unicode_norm_table.pl, but other file\\n> generating scripts around the tree print out a small header in the style\\n> that I have. I'd rather adjust the output of\\n> generate-unicode_norm_table.pl to match those. (It's also not quite\\n> correct to make copyright claims about automatically generated output.)\\n\\nHmm, the scripts I'm most familiar with have full headers. Your point\\nabout copyright makes sense, and using smaller file headers would aid\\nreadability of the scripts, but I also see how others may feel\\ndifferently.\\n\\nv2 looks good to me, marked ready for committer.\\n\\n--\\nJohn Naylor https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 3 Jan 2020 08:13:53 -0600\",\n \"msg_from\": \"John Naylor <john.naylor@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add support for automatically updating Unicode derived files\"\n },\n {\n \"msg_contents\": \"On 2020-01-03 15:13, John Naylor wrote:\\n> On Thu, Dec 26, 2019 at 12:39 PM Peter Eisentraut\\n> <peter.eisentraut@2ndquadrant.com> wrote:\\n>>\\n>> On 2019-12-19 23:48, John Naylor wrote:\\n>>> I would print out the full boilerplate like for other generated headers.\\n>>\\n>> Hmm, you are probably comparing with\\n>> src/common/unicode/generate-unicode_norm_table.pl, but other file\\n>> generating scripts around the tree print out a small header in the style\\n>> that I have. I'd rather adjust the output of\\n>> generate-unicode_norm_table.pl to match those. (It's also not quite\\n>> correct to make copyright claims about automatically generated output.)\\n> \\n> Hmm, the scripts I'm most familiar with have full headers. Your point\\n> about copyright makes sense, and using smaller file headers would aid\\n> readability of the scripts, but I also see how others may feel\\n> differently.\\n> \\n> v2 looks good to me, marked ready for committer.\\n\\nCommitted, thanks.\\n\\nI have added a little tweak so that it works also without --with-python, \\nto avoid gratuitous annoyances.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Jan 2020 10:16:19 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add support for automatically updating Unicode derived files\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> Committed, thanks.\\n\\nThis patch is making src/tools/pginclude/headerscheck unhappy:\\n\\n./src/include/common/unicode_combining_table.h:3: error: array type has incomplete element type\\n\\nI guess that header needs another #include, or else you need to\\nmove some declarations around.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 14 Jan 2020 19:37:30 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add support for automatically updating Unicode derived files\"\n },\n {\n \"msg_contents\": \"On 2020-01-15 01:37, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n>> Committed, thanks.\\n> \\n> This patch is making src/tools/pginclude/headerscheck unhappy:\\n> \\n> ./src/include/common/unicode_combining_table.h:3: error: array type has incomplete element type\\n> \\n> I guess that header needs another #include, or else you need to\\n> move some declarations around.\\n\\nHmm, this file is only meant to be included inside one particular \\nfunction. Making it standalone includable would seem to be unnecessary. \\n What should we do?\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Mon, 20 Jan 2020 09:59:34 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add support for automatically updating Unicode derived files\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> On 2020-01-15 01:37, Tom Lane wrote:\\n>> This patch is making src/tools/pginclude/headerscheck unhappy:\\n>> ./src/include/common/unicode_combining_table.h:3: error: array type has incomplete element type\\n\\n> Hmm, this file is only meant to be included inside one particular \\n> function. Making it standalone includable would seem to be unnecessary. \\n> What should we do?\\n\\nWell, we could make it a documented exception in headerscheck and\\ncpluspluscheck.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 20 Jan 2020 10:43:32 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add support for automatically updating Unicode derived files\"\n },\n {\n \"msg_contents\": \"On 2020-01-20 16:43, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n>> On 2020-01-15 01:37, Tom Lane wrote:\\n>>> This patch is making src/tools/pginclude/headerscheck unhappy:\\n>>> ./src/include/common/unicode_combining_table.h:3: error: array type has incomplete element type\\n> \\n>> Hmm, this file is only meant to be included inside one particular\\n>> function. Making it standalone includable would seem to be unnecessary.\\n>> What should we do?\\n> \\n> Well, we could make it a documented exception in headerscheck and\\n> cpluspluscheck.\\n\\nOK, done.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 24 Jan 2020 12:25:06 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add support for automatically updating Unicode derived files\"\n },\n {\n \"msg_contents\": \"I have committed the first Unicode data update using this new \\\"make \\nupdate-unicode\\\" facility.\\n\\nCLDR is released regularly every 6 months, so around this time every \\nyear would be the appropriate time to pull in the latest updates in \\npreparation for our own release.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 24 Apr 2020 10:01:00 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add support for automatically updating Unicode derived files\"\n }\n]","isConversational":false}}},{"rowIdx":49359,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"When joining tables with USING, the listed columns are merged and no\nlonger belong to either the left or the right side. That means they can\nno longer be qualified which can often be an inconvenience.\n\n\nSELECT a.x, b.y, z FROM a INNER JOIN b USING (z);\n\n\nThe SQL standard provides a workaround for this by allowing an alias on\nthe join clause. (<join correlation name> in section 7.10)\n\n\nSELECT j.x, j.y, j.z FROM a INNER JOIN b USING (z) AS j;\n\n\nAttached is a patch (based on 517bf2d910) adding this feature.\n\n-- \n\nVik Fearing","msg_date":"Tue, 29 Oct 2019 11:47:48 +0100","msg_from":"Vik Fearing <vik.fearing@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Join Correlation Name"},{"msg_contents":"On 2019-10-29 11:47, Vik Fearing wrote:\n> When joining tables with USING, the listed columns are merged and no\n> longer belong to either the left or the right side. That means they can\n> no longer be qualified which can often be an inconvenience.\n> \n> \n> SELECT a.x, b.y, z FROM a INNER JOIN b USING (z);\n> \n> \n> The SQL standard provides a workaround for this by allowing an alias on\n> the join clause. (<join correlation name> in section 7.10)\n> \n> \n> SELECT j.x, j.y, j.z FROM a INNER JOIN b USING (z) AS j;\n> \n> \n> Attached is a patch (based on 517bf2d910) adding this feature.\n\nIs this the same as https://commitfest.postgresql.org/25/2158/ ?\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Tue, 29 Oct 2019 12:05:29 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Join Correlation Name"},{"msg_contents":"On Tue, 29 Oct 2019 at 07:05, Peter Eisentraut <\npeter.eisentraut@2ndquadrant.com> wrote:\n\n> On 2019-10-29 11:47, Vik Fearing wrote:\n> > When joining tables with USING, the listed columns are merged and no\n> > longer belong to either the left or the right side. That means they can\n> > no longer be qualified which can often be an inconvenience.\n> >\n> >\n> > SELECT a.x, b.y, z FROM a INNER JOIN b USING (z);\n>\n\nI'm confused. As far as I can tell you can qualify the join columns if you\nwant:\n\nodyssey=> select exam_id, sitting_id, room_id, exam_exam_sitting.exam_id\nfrom exam_exam_sitting join exam_exam_sitting_room using (exam_id,\nsitting_id) limit 5;\n exam_id | sitting_id | room_id | exam_id\n---------+------------+---------+---------\n 22235 | 23235 | 22113 | 22235\n 22237 | 23237 | 22113 | 22237\n 23101 | 21101 | 22215 | 23101\n 23101 | 21101 | 22216 | 23101\n 23101 | 21101 | 22224 | 23101\n(5 rows)\n\nodyssey=>\n\nIn the case of a non-inner join it can make a difference whether you use\nthe left side, right side, or non-qualified version. If you need to refer\nspecifically to the non-qualified version in a different part of the query,\nyou can give an alias to the result of the join:\n\n... (a join b using (z)) as t ...\n\n> The SQL standard provides a workaround for this by allowing an alias on\n> > the join clause. (<join correlation name> in section 7.10)\n> >\n> >\n> > SELECT j.x, j.y, j.z FROM a INNER JOIN b USING (z) AS j;\n>\n\nWhat I would like is to be able to use both USING and ON in the same join;\nI more often than I would like find myself saying things like ON ((l.a,\nl.b, lc.) = (r.a, r.b, r.c) AND l.ab = r.bb). Also I would like to be able\nto use and rename differently-named fields in a USING clause, something\nlike USING (a, b, c=d as f).\n\nA bit of thought convinces me that these are both essentially syntactic\nchanges; I think it's already possible to represent these in the existing\ninternal representation, they just aren't supported by the parser.\n\nOn Tue, 29 Oct 2019 at 07:05, Peter Eisentraut <peter.eisentraut@2ndquadrant.com> wrote:On 2019-10-29 11:47, Vik Fearing wrote:\n> When joining tables with USING, the listed columns are merged and no\n> longer belong to either the left or the right side. That means they can\n> no longer be qualified which can often be an inconvenience.\n> \n> \n> SELECT a.x, b.y, z FROM a INNER JOIN b USING (z);I'm confused. As far as I can tell you can qualify the join columns if you want:odyssey=> select exam_id, sitting_id, room_id, exam_exam_sitting.exam_id from exam_exam_sitting join exam_exam_sitting_room using (exam_id, sitting_id) limit 5; exam_id | sitting_id | room_id | exam_id ---------+------------+---------+--------- 22235 | 23235 | 22113 | 22235 22237 | 23237 | 22113 | 22237 23101 | 21101 | 22215 | 23101 23101 | 21101 | 22216 | 23101 23101 | 21101 | 22224 | 23101(5 rows)odyssey=> In the case of a non-inner join it can make a difference whether you use the left side, right side, or non-qualified version. If you need to refer specifically to the non-qualified version in a different part of the query, you can give an alias to the result of the join:... (a join b using (z)) as t ...\n> The SQL standard provides a workaround for this by allowing an alias on\n> the join clause. (<join correlation name> in section 7.10)\n> \n> \n> SELECT j.x, j.y, j.z FROM a INNER JOIN b USING (z) AS j;What I would like is to be able to use both USING and ON in the same join; I more often than I would like find myself saying things like ON ((l.a, l.b, lc.) = (r.a, r.b, r.c) AND l.ab = r.bb). Also I would like to be able to use and rename differently-named fields in a USING clause, something like USING (a, b, c=d as f).A bit of thought convinces me that these are both essentially syntactic changes; I think it's already possible to represent these in the existing internal representation, they just aren't supported by the parser.","msg_date":"Tue, 29 Oct 2019 07:24:28 -0400","msg_from":"Isaac Morland <isaac.morland@gmail.com>","msg_from_op":false,"msg_subject":"Re: Join Correlation Name"},{"msg_contents":"On 29/10/2019 12:05, Peter Eisentraut wrote:\n> On 2019-10-29 11:47, Vik Fearing wrote:\n>> When joining tables with USING, the listed columns are merged and no\n>> longer belong to either the left or the right side. That means they can\n>> no longer be qualified which can often be an inconvenience.\n>>\n>>\n>> SELECT a.x, b.y, z FROM a INNER JOIN b USING (z);\n>>\n>>\n>> The SQL standard provides a workaround for this by allowing an alias on\n>> the join clause. (<join correlation name> in section 7.10)\n>>\n>>\n>> SELECT j.x, j.y, j.z FROM a INNER JOIN b USING (z) AS j;\n>>\n>>\n>> Attached is a patch (based on 517bf2d910) adding this feature.\n>\n> Is this the same as https://commitfest.postgresql.org/25/2158/ ?\n\n\nCrap. Yes, it is.\n\n\n\n","msg_date":"Tue, 29 Oct 2019 12:51:26 +0100","msg_from":"Vik Fearing <vik.fearing@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Join Correlation Name"},{"msg_contents":"On 29/10/2019 12:24, Isaac Morland wrote:\n> If you need to refer specifically to the non-qualified version in a\n> different part of the query, you can give an alias to the result of\n> the join:\n>\n> ... (a join b using (z)) as t ...\n\n\nYes, this is about having standard SQL syntax for that.\n\n\n\n","msg_date":"Tue, 29 Oct 2019 12:55:46 +0100","msg_from":"Vik Fearing <vik.fearing@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Join Correlation Name"},{"msg_contents":"Vik Fearing <vik.fearing@2ndquadrant.com> writes:\n> On 29/10/2019 12:24, Isaac Morland wrote:\n>> If you need to refer specifically to the non-qualified version in a\n>> different part of the query, you can give an alias to the result of\n>> the join:\n>> ... (a join b using (z)) as t ...\n\n> Yes, this is about having standard SQL syntax for that.\n\nPlease present an argument why this proposal is standard SQL syntax.\nI see no support for it in the spec. AFAICS this proposal is just an\ninconsistent wart; it makes it possible to write\n\n\t(a join b using (z) as q) as t\n\nand then what do you do? Moreover, why should you be able to\nattach an alias to a USING join but not other sorts of joins?\n\nAfter digging around in the spec for awhile, it seems like\nthere actually isn't any way to attach an alias to a join\nper spec.\n\nAccording to SQL:2011 7.6 <table reference>, you can attach an\nAS clause to every variant of <table primary> *except* the\n<parenthesized joined table> variant. And there's nothing\nabout AS clauses in 7.7 <joined table>, which is where it would\nhave to be mentioned if this proposal were spec-compliant.\n\nWhat our grammar effectively does is to allow an AS clause to be\nattached to <parenthesized joined table> as well, which seems\nlike the most natural thing to do if the committee ever decide\nto rectify the shortcoming.\n\nAnyway, we already have the functionality covered, and I don't\nthink we need another non-spec, non-orthogonal way to do it.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 29 Oct 2019 10:20:27 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Join Correlation Name"},{"msg_contents":"On 29/10/2019 15:20, Tom Lane wrote:\n> Vik Fearing <vik.fearing@2ndquadrant.com> writes:\n>> On 29/10/2019 12:24, Isaac Morland wrote:\n>>> If you need to refer specifically to the non-qualified version in a\n>>> different part of the query, you can give an alias to the result of\n>>> the join:\n>>> ... (a join b using (z)) as t ...\n>> Yes, this is about having standard SQL syntax for that.\n> Please present an argument why this proposal is standard SQL syntax.\n\n\nIs quoting the spec good enough?\n\nSQL:2016 Part 2 Foundation Section 7.10 <joined table>:\n\n\n<join specification> ::=\n <join condition>\n | <named columns join>\n\n<join condition> ::=\n ON <search condition>\n\n<named columns join> ::=\n USING <left paren> <join column list> <right paren> [ AS <join\ncorrelation name> ]\n\n<join correlation name> ::=\n <correlation name>\n\n\n> I see no support for it in the spec. AFAICS this proposal is just an\n> inconsistent wart; it makes it possible to write\n>\n> \t(a join b using (z) as q) as t\n>\n> and then what do you do? Moreover, why should you be able to\n> attach an alias to a USING join but not other sorts of joins?\n\n\nI think possibly what the spec says (and that neither my patch nor\nPeter's implements) is assigning the alias just to the <join column\nlist>. So my original example query should actually be:\n\n\nSELECT a.x, b.y, j.z FROM a INNER JOIN b USING (z) AS j;\n\n\n> After digging around in the spec for awhile, it seems like\n> there actually isn't any way to attach an alias to a join\n> per spec.\n>\n> According to SQL:2011 7.6 <table reference>, you can attach an\n> AS clause to every variant of <table primary> *except* the\n> <parenthesized joined table> variant. And there's nothing\n> about AS clauses in 7.7 <joined table>, which is where it would\n> have to be mentioned if this proposal were spec-compliant.\n>\n> What our grammar effectively does is to allow an AS clause to be\n> attached to <parenthesized joined table> as well, which seems\n> like the most natural thing to do if the committee ever decide\n> to rectify the shortcoming.\n>\n> Anyway, we already have the functionality covered, and I don't\n> think we need another non-spec, non-orthogonal way to do it.\n\n\nI think the issue here is you're looking at SQL:2011 whereas I am\nlooking at SQL:2016.\n\n-- \n\nVik Fearing\n\n\n\n","msg_date":"Tue, 29 Oct 2019 16:59:40 +0100","msg_from":"Vik Fearing <vik.fearing@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Join Correlation Name"},{"msg_contents":"Bonjour Vik,\n\n> Is quoting the spec good enough?\n> SQL:2016 Part 2 Foundation Section 7.10 <joined table>:\n\nAh, this is the one information I did not have when reviewing Peter's \npatch.\n\n> <named columns join> ::=\n> USING <left paren> <join column list> <right paren> [ AS <join correlation name> ]\n>\n> <join correlation name> ::=\n> <correlation name>\n>\n> I think possibly what the spec says (and that neither my patch nor\n> Peter's implements) is assigning the alias just to the <join column\n> list>. \n\nI think you are right, the alias is only on the identical columns.\n\nIt solves the issue I raised about inaccessible attributes, and explains \nwhy it is only available with USING and no other join variants.\n\n> So my original example query should actually be:\n>\n> SELECT a.x, b.y, j.z FROM a INNER JOIN b USING (z) AS j;\n\nYep, only z should be in j, it is really just about the USING clause.\n\n-- \nFabien.","msg_date":"Wed, 30 Oct 2019 09:04:12 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":false,"msg_subject":"Re: Join Correlation Name"},{"msg_contents":"On 30/10/2019 09:04, Fabien COELHO wrote:\n>\n>> I think possibly what the spec says (and that neither my patch nor\n>> Peter's implements) is assigning the alias just to the <join column\n>> list>. \n>\n> I think you are right, the alias is only on the identical columns.\n>\n> It solves the issue I raised about inaccessible attributes, and\n> explains why it is only available with USING and no other join variants.\n>\n>> So my original example query should actually be:\n>>\n>> SELECT a.x, b.y, j.z FROM a INNER JOIN b USING (z) AS j;\n>\n> Yep, only z should be in j, it is really just about the USING clause.\n\n\nMy reading of SQL:2016-2 7.10 SR 11.a convinces me that this is the case.\n\n\nMy reading of transformFromClauseItem() convinces me that this is way\nover my head and I have to abandon it. :-(\n\n\n\n","msg_date":"Fri, 1 Nov 2019 23:08:35 +0100","msg_from":"Vik Fearing <vik.fearing@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Join Correlation Name"}],"string":"[\n {\n \"msg_contents\": \"When joining tables with USING, the listed columns are merged and no\\nlonger belong to either the left or the right side. That means they can\\nno longer be qualified which can often be an inconvenience.\\n\\n\\nSELECT a.x, b.y, z FROM a INNER JOIN b USING (z);\\n\\n\\nThe SQL standard provides a workaround for this by allowing an alias on\\nthe join clause. (<join correlation name> in section 7.10)\\n\\n\\nSELECT j.x, j.y, j.z FROM a INNER JOIN b USING (z) AS j;\\n\\n\\nAttached is a patch (based on 517bf2d910) adding this feature.\\n\\n-- \\n\\nVik Fearing\",\n \"msg_date\": \"Tue, 29 Oct 2019 11:47:48 +0100\",\n \"msg_from\": \"Vik Fearing <vik.fearing@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Join Correlation Name\"\n },\n {\n \"msg_contents\": \"On 2019-10-29 11:47, Vik Fearing wrote:\\n> When joining tables with USING, the listed columns are merged and no\\n> longer belong to either the left or the right side. That means they can\\n> no longer be qualified which can often be an inconvenience.\\n> \\n> \\n> SELECT a.x, b.y, z FROM a INNER JOIN b USING (z);\\n> \\n> \\n> The SQL standard provides a workaround for this by allowing an alias on\\n> the join clause. (<join correlation name> in section 7.10)\\n> \\n> \\n> SELECT j.x, j.y, j.z FROM a INNER JOIN b USING (z) AS j;\\n> \\n> \\n> Attached is a patch (based on 517bf2d910) adding this feature.\\n\\nIs this the same as https://commitfest.postgresql.org/25/2158/ ?\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 12:05:29 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Join Correlation Name\"\n },\n {\n \"msg_contents\": \"On Tue, 29 Oct 2019 at 07:05, Peter Eisentraut <\\npeter.eisentraut@2ndquadrant.com> wrote:\\n\\n> On 2019-10-29 11:47, Vik Fearing wrote:\\n> > When joining tables with USING, the listed columns are merged and no\\n> > longer belong to either the left or the right side. That means they can\\n> > no longer be qualified which can often be an inconvenience.\\n> >\\n> >\\n> > SELECT a.x, b.y, z FROM a INNER JOIN b USING (z);\\n>\\n\\nI'm confused. As far as I can tell you can qualify the join columns if you\\nwant:\\n\\nodyssey=> select exam_id, sitting_id, room_id, exam_exam_sitting.exam_id\\nfrom exam_exam_sitting join exam_exam_sitting_room using (exam_id,\\nsitting_id) limit 5;\\n exam_id | sitting_id | room_id | exam_id\\n---------+------------+---------+---------\\n 22235 | 23235 | 22113 | 22235\\n 22237 | 23237 | 22113 | 22237\\n 23101 | 21101 | 22215 | 23101\\n 23101 | 21101 | 22216 | 23101\\n 23101 | 21101 | 22224 | 23101\\n(5 rows)\\n\\nodyssey=>\\n\\nIn the case of a non-inner join it can make a difference whether you use\\nthe left side, right side, or non-qualified version. If you need to refer\\nspecifically to the non-qualified version in a different part of the query,\\nyou can give an alias to the result of the join:\\n\\n... (a join b using (z)) as t ...\\n\\n> The SQL standard provides a workaround for this by allowing an alias on\\n> > the join clause. (<join correlation name> in section 7.10)\\n> >\\n> >\\n> > SELECT j.x, j.y, j.z FROM a INNER JOIN b USING (z) AS j;\\n>\\n\\nWhat I would like is to be able to use both USING and ON in the same join;\\nI more often than I would like find myself saying things like ON ((l.a,\\nl.b, lc.) = (r.a, r.b, r.c) AND l.ab = r.bb). Also I would like to be able\\nto use and rename differently-named fields in a USING clause, something\\nlike USING (a, b, c=d as f).\\n\\nA bit of thought convinces me that these are both essentially syntactic\\nchanges; I think it's already possible to represent these in the existing\\ninternal representation, they just aren't supported by the parser.\\n\\nOn Tue, 29 Oct 2019 at 07:05, Peter Eisentraut <peter.eisentraut@2ndquadrant.com> wrote:On 2019-10-29 11:47, Vik Fearing wrote:\\n> When joining tables with USING, the listed columns are merged and no\\n> longer belong to either the left or the right side. That means they can\\n> no longer be qualified which can often be an inconvenience.\\n> \\n> \\n> SELECT a.x, b.y, z FROM a INNER JOIN b USING (z);I'm confused. As far as I can tell you can qualify the join columns if you want:odyssey=> select exam_id, sitting_id, room_id, exam_exam_sitting.exam_id from exam_exam_sitting join exam_exam_sitting_room using (exam_id, sitting_id) limit 5; exam_id | sitting_id | room_id | exam_id ---------+------------+---------+--------- 22235 | 23235 | 22113 | 22235 22237 | 23237 | 22113 | 22237 23101 | 21101 | 22215 | 23101 23101 | 21101 | 22216 | 23101 23101 | 21101 | 22224 | 23101(5 rows)odyssey=> In the case of a non-inner join it can make a difference whether you use the left side, right side, or non-qualified version. If you need to refer specifically to the non-qualified version in a different part of the query, you can give an alias to the result of the join:... (a join b using (z)) as t ...\\n> The SQL standard provides a workaround for this by allowing an alias on\\n> the join clause. (<join correlation name> in section 7.10)\\n> \\n> \\n> SELECT j.x, j.y, j.z FROM a INNER JOIN b USING (z) AS j;What I would like is to be able to use both USING and ON in the same join; I more often than I would like find myself saying things like ON ((l.a, l.b, lc.) = (r.a, r.b, r.c) AND l.ab = r.bb). Also I would like to be able to use and rename differently-named fields in a USING clause, something like USING (a, b, c=d as f).A bit of thought convinces me that these are both essentially syntactic changes; I think it's already possible to represent these in the existing internal representation, they just aren't supported by the parser.\",\n \"msg_date\": \"Tue, 29 Oct 2019 07:24:28 -0400\",\n \"msg_from\": \"Isaac Morland <isaac.morland@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Join Correlation Name\"\n },\n {\n \"msg_contents\": \"On 29/10/2019 12:05, Peter Eisentraut wrote:\\n> On 2019-10-29 11:47, Vik Fearing wrote:\\n>> When joining tables with USING, the listed columns are merged and no\\n>> longer belong to either the left or the right side. That means they can\\n>> no longer be qualified which can often be an inconvenience.\\n>>\\n>>\\n>> SELECT a.x, b.y, z FROM a INNER JOIN b USING (z);\\n>>\\n>>\\n>> The SQL standard provides a workaround for this by allowing an alias on\\n>> the join clause. (<join correlation name> in section 7.10)\\n>>\\n>>\\n>> SELECT j.x, j.y, j.z FROM a INNER JOIN b USING (z) AS j;\\n>>\\n>>\\n>> Attached is a patch (based on 517bf2d910) adding this feature.\\n>\\n> Is this the same as https://commitfest.postgresql.org/25/2158/ ?\\n\\n\\nCrap. Yes, it is.\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 12:51:26 +0100\",\n \"msg_from\": \"Vik Fearing <vik.fearing@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Join Correlation Name\"\n },\n {\n \"msg_contents\": \"On 29/10/2019 12:24, Isaac Morland wrote:\\n> If you need to refer specifically to the non-qualified version in a\\n> different part of the query, you can give an alias to the result of\\n> the join:\\n>\\n> ... (a join b using (z)) as t ...\\n\\n\\nYes, this is about having standard SQL syntax for that.\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 12:55:46 +0100\",\n \"msg_from\": \"Vik Fearing <vik.fearing@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Join Correlation Name\"\n },\n {\n \"msg_contents\": \"Vik Fearing <vik.fearing@2ndquadrant.com> writes:\\n> On 29/10/2019 12:24, Isaac Morland wrote:\\n>> If you need to refer specifically to the non-qualified version in a\\n>> different part of the query, you can give an alias to the result of\\n>> the join:\\n>> ... (a join b using (z)) as t ...\\n\\n> Yes, this is about having standard SQL syntax for that.\\n\\nPlease present an argument why this proposal is standard SQL syntax.\\nI see no support for it in the spec. AFAICS this proposal is just an\\ninconsistent wart; it makes it possible to write\\n\\n\\t(a join b using (z) as q) as t\\n\\nand then what do you do? Moreover, why should you be able to\\nattach an alias to a USING join but not other sorts of joins?\\n\\nAfter digging around in the spec for awhile, it seems like\\nthere actually isn't any way to attach an alias to a join\\nper spec.\\n\\nAccording to SQL:2011 7.6 <table reference>, you can attach an\\nAS clause to every variant of <table primary> *except* the\\n<parenthesized joined table> variant. And there's nothing\\nabout AS clauses in 7.7 <joined table>, which is where it would\\nhave to be mentioned if this proposal were spec-compliant.\\n\\nWhat our grammar effectively does is to allow an AS clause to be\\nattached to <parenthesized joined table> as well, which seems\\nlike the most natural thing to do if the committee ever decide\\nto rectify the shortcoming.\\n\\nAnyway, we already have the functionality covered, and I don't\\nthink we need another non-spec, non-orthogonal way to do it.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 10:20:27 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Join Correlation Name\"\n },\n {\n \"msg_contents\": \"On 29/10/2019 15:20, Tom Lane wrote:\\n> Vik Fearing <vik.fearing@2ndquadrant.com> writes:\\n>> On 29/10/2019 12:24, Isaac Morland wrote:\\n>>> If you need to refer specifically to the non-qualified version in a\\n>>> different part of the query, you can give an alias to the result of\\n>>> the join:\\n>>> ... (a join b using (z)) as t ...\\n>> Yes, this is about having standard SQL syntax for that.\\n> Please present an argument why this proposal is standard SQL syntax.\\n\\n\\nIs quoting the spec good enough?\\n\\nSQL:2016 Part 2 Foundation Section 7.10 <joined table>:\\n\\n\\n<join specification> ::=\\n <join condition>\\n | <named columns join>\\n\\n<join condition> ::=\\n ON <search condition>\\n\\n<named columns join> ::=\\n USING <left paren> <join column list> <right paren> [ AS <join\\ncorrelation name> ]\\n\\n<join correlation name> ::=\\n <correlation name>\\n\\n\\n> I see no support for it in the spec. AFAICS this proposal is just an\\n> inconsistent wart; it makes it possible to write\\n>\\n> \\t(a join b using (z) as q) as t\\n>\\n> and then what do you do? Moreover, why should you be able to\\n> attach an alias to a USING join but not other sorts of joins?\\n\\n\\nI think possibly what the spec says (and that neither my patch nor\\nPeter's implements) is assigning the alias just to the <join column\\nlist>. So my original example query should actually be:\\n\\n\\nSELECT a.x, b.y, j.z FROM a INNER JOIN b USING (z) AS j;\\n\\n\\n> After digging around in the spec for awhile, it seems like\\n> there actually isn't any way to attach an alias to a join\\n> per spec.\\n>\\n> According to SQL:2011 7.6 <table reference>, you can attach an\\n> AS clause to every variant of <table primary> *except* the\\n> <parenthesized joined table> variant. And there's nothing\\n> about AS clauses in 7.7 <joined table>, which is where it would\\n> have to be mentioned if this proposal were spec-compliant.\\n>\\n> What our grammar effectively does is to allow an AS clause to be\\n> attached to <parenthesized joined table> as well, which seems\\n> like the most natural thing to do if the committee ever decide\\n> to rectify the shortcoming.\\n>\\n> Anyway, we already have the functionality covered, and I don't\\n> think we need another non-spec, non-orthogonal way to do it.\\n\\n\\nI think the issue here is you're looking at SQL:2011 whereas I am\\nlooking at SQL:2016.\\n\\n-- \\n\\nVik Fearing\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 16:59:40 +0100\",\n \"msg_from\": \"Vik Fearing <vik.fearing@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Join Correlation Name\"\n },\n {\n \"msg_contents\": \"Bonjour Vik,\\n\\n> Is quoting the spec good enough?\\n> SQL:2016 Part 2 Foundation Section 7.10 <joined table>:\\n\\nAh, this is the one information I did not have when reviewing Peter's \\npatch.\\n\\n> <named columns join> ::=\\n> USING <left paren> <join column list> <right paren> [ AS <join correlation name> ]\\n>\\n> <join correlation name> ::=\\n> <correlation name>\\n>\\n> I think possibly what the spec says (and that neither my patch nor\\n> Peter's implements) is assigning the alias just to the <join column\\n> list>. \\n\\nI think you are right, the alias is only on the identical columns.\\n\\nIt solves the issue I raised about inaccessible attributes, and explains \\nwhy it is only available with USING and no other join variants.\\n\\n> So my original example query should actually be:\\n>\\n> SELECT a.x, b.y, j.z FROM a INNER JOIN b USING (z) AS j;\\n\\nYep, only z should be in j, it is really just about the USING clause.\\n\\n-- \\nFabien.\",\n \"msg_date\": \"Wed, 30 Oct 2019 09:04:12 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Join Correlation Name\"\n },\n {\n \"msg_contents\": \"On 30/10/2019 09:04, Fabien COELHO wrote:\\n>\\n>> I think possibly what the spec says (and that neither my patch nor\\n>> Peter's implements) is assigning the alias just to the <join column\\n>> list>. \\n>\\n> I think you are right, the alias is only on the identical columns.\\n>\\n> It solves the issue I raised about inaccessible attributes, and\\n> explains why it is only available with USING and no other join variants.\\n>\\n>> So my original example query should actually be:\\n>>\\n>> SELECT a.x, b.y, j.z FROM a INNER JOIN b USING (z) AS j;\\n>\\n> Yep, only z should be in j, it is really just about the USING clause.\\n\\n\\nMy reading of SQL:2016-2 7.10 SR 11.a convinces me that this is the case.\\n\\n\\nMy reading of transformFromClauseItem() convinces me that this is way\\nover my head and I have to abandon it. :-(\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 23:08:35 +0100\",\n \"msg_from\": \"Vik Fearing <vik.fearing@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Join Correlation Name\"\n }\n]","isConversational":false}}},{"rowIdx":49360,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi.\n\nThis is not clear from doc, so I have asked on IRC too.\n\nfrom the DOC: https://www.postgresql.org/docs/current/trigger-definition.html\nIn the case of INSTEAD OF triggers, the possibly-modified row returned by each trigger becomes the input to the next trigger\n\nI modify OLD row, thus I expect to get modified version when run next query: \n\n WITH t1 AS( delete from abc returning *)\n select * from t1;\n\nfiddle: https://dbfiddle.uk/?rdbms=postgres_12&fiddle=637730305f66bf531794edb09a462c95\n\n> https://www.postgresql.org/docs/current/trigger-definition.html\nA row-level INSTEAD OF trigger should either return NULL to indicate that it did not modify any data from the view's underlying base tables,\nor it should return the view row that was passed in (the NEW row for INSERT and UPDATE operations, or the OLD row for DELETE operations).\nA nonnull return value is used to signal that the trigger performed the necessary data modifications in the view.\nThis will cause the count of the number of rows affected by the command to be incremented. For INSERT and UPDATE operations, the trigger may\nmodify the NEW row before returning it. This will change the data returned by INSERT RETURNING or UPDATE RETURNING,\nand is useful when the view will not show exactly the same data that was provided.\n\nBut I still does not understand. Doc explicitly do not prohibit modification of OLD and has no examples for DELETE RETURNING case\n\nSo I want to ask clarify doc a bit.\nIf this prohibited, why this is prohibited? have any discussion on this?\nIf not prohibited, does this is not implemented for DELETE RETURNING queries? if so, is it left for later?\n\nI have next use case.\nI am implementing Bi-Temporal tables. The table have columns: id, app_period, value\nfor example I have next data: 7, '[2019-01-01, 2020-01-01)', 130\nYou can imagine this as having value 7 for each day of the year.\nNow I want to delete this value for May month. I setup special variable to period: '[2019-05-01,2019-06-01)' and then delete:\n\n select app_period( '[2019-05-01,2019-06-01)' );\n WITH t1 AS( delete from abc returning *)\n select * from t1;\n\nAlgorithm of deletion is next:\n1. Deactivate target row\n 7, '[2019-01-01, 2020-01-01)', 130\n2. If target row has wider app_period then we insert record that data back:\n NOT '[2019-05-01,2019-06-01)' @> '[2019-01-01, 2020-01-01)'\n INSERT INTO abc ( id, app_period, value ) values \n ( 7, '[2019-01-01,2019-05-01)', 130 ),\n ( 7, '[2019-06-01,2020-01-01)', 130 ),\n3. OLD.app_period = OLD.app_period * app_period(); \n '[2019-01-01, 2020-01-01)' * '[2019-05-01,2019-06-01)' --> '[2019-05-01,2019-06-01)'\n\nBecause only 130 value is deleted from specified period I expect next result for the query above:\n ( 7, '[2019-05-01,2019-06-01)', 130 )\n\nBut despite on OLD was modified, actual result is:\n ( 7, '[2019-01-01,2020-01-01)', 130 )\nYou can see that this is original data.\n\nSo, does INSTEAD OF DELETE support modification of row?\n\n-- \nBest regards,\nEugen Konkov\n\n\n\n","msg_date":"Tue, 29 Oct 2019 17:54:36 +0200","msg_from":"Eugen Konkov <kes-kes@yandex.ru>","msg_from_op":true,"msg_subject":"Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"On Tue, Oct 29, 2019 at 05:54:36PM +0200, Eugen Konkov wrote:\n> Hi.\n> \n> This is not clear from doc, so I have asked on IRC too.\n> \n> from the DOC: https://www.postgresql.org/docs/current/trigger-definition.html\n> In the case of INSTEAD OF triggers, the possibly-modified row returned by each trigger becomes the input to the next trigger\n> \n> I modify OLD row, thus I expect to get modified version when run next query: \n> \n> WITH t1 AS( delete from abc returning *)\n> select * from t1;\n> \n> fiddle: https://dbfiddle.uk/?rdbms=postgres_12&fiddle=637730305f66bf531794edb09a462c95\n\nWow, that is a very nice way to present the queries.\n\n> > https://www.postgresql.org/docs/current/trigger-definition.html\n> A row-level INSTEAD OF trigger should either return NULL to indicate that it did not modify any data from the view's underlying base tables,\n> or it should return the view row that was passed in (the NEW row for INSERT and UPDATE operations, or the OLD row for DELETE operations).\n> A nonnull return value is used to signal that the trigger performed the necessary data modifications in the view.\n> This will cause the count of the number of rows affected by the command to be incremented. For INSERT and UPDATE operations, the trigger may\n> modify the NEW row before returning it. This will change the data returned by INSERT RETURNING or UPDATE RETURNING,\n> and is useful when the view will not show exactly the same data that was provided.\n> \n> But I still does not understand. Doc explicitly do not prohibit modification of OLD and has no examples for DELETE RETURNING case\n\nI looked in the CREATE TRIGGER manual page and found this:\n\n\thttps://www.postgresql.org/docs/12/sql-createtrigger.html\n\tIf the trigger fires before or instead of the event, the trigger\n\tcan skip the operation for the current row, or change the row\n\tbeing inserted (for INSERT and UPDATE operations only).\n\nI don't see the \"(for INSERT and UPDATE operations only)\" language in\nthe main trigger documentation,\nhttps://www.postgresql.org/docs/current/trigger-definition.html. I have\nwritten the attached patch to fix that. Does that help?\n\nAs far as allowing DELETE to modify the trigger row for RETURNING, I am\nnot sure how much work it would take to allow that, but it seems like it\nis a valid requite, and if so, I can add it to the TODO list.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +","msg_date":"Wed, 6 Nov 2019 13:59:35 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"> I looked in the CREATE TRIGGER manual page and found this:\n\n> https://www.postgresql.org/docs/12/sql-createtrigger.html\n> If the trigger fires before or instead of the event, the trigger\n> can skip the operation for the current row, or change the row\n> being inserted (for INSERT and UPDATE operations only).\n\n> I don't see the \"(for INSERT and UPDATE operations only)\" language in\n> the main trigger documentation,\n> https://www.postgresql.org/docs/current/trigger-definition.html. I have\n> written the attached patch to fix that. Does that help?\n\nNo. If we document that PG does not allow to modify OLD at instead\nof trigger, the we can not implement that. Probably we can put note\nthat \"currently modification of the trigger row for RETURNING is not\nimplemented\"\n\n> As far as allowing DELETE to modify the trigger row for RETURNING, I am\n> not sure how much work it would take to allow that, but it seems like it\n> is a valid requite, and if so, I can add it to the TODO list.\n\nYes, Add please into TODO the feature to \"allowing DELETE to modify the trigger row\nfor RETURNING\". Becuase, as I have described at first letter, without\nthis the RETURNING rows **does not correspond actually deleted data**\n\nThank you.\n\n-- \nBest regards,\nEugen Konkov\n\n\n\n","msg_date":"Thu, 7 Nov 2019 11:20:32 +0200","msg_from":"Eugen Konkov <kes-kes@yandex.ru>","msg_from_op":true,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"Hello Eugen,\n\nThursday, November 7, 2019, 11:20:32 AM, you wrote:\n\n>> I looked in the CREATE TRIGGER manual page and found this:\n\n>> https://www.postgresql.org/docs/12/sql-createtrigger.html\n>> If the trigger fires before or instead of the event, the trigger\n>> can skip the operation for the current row, or change the row\n>> being inserted (for INSERT and UPDATE operations only).\n\n>> I don't see the \"(for INSERT and UPDATE operations only)\" language in\n>> the main trigger documentation,\n>> https://www.postgresql.org/docs/current/trigger-definition.html. I have\n>> written the attached patch to fix that. Does that help?\n\n> No. If we document that PG does not allow to modify OLD at instead\n> of trigger, the we can not implement that. Probably we can put note\n> that \"currently modification of the trigger row for RETURNING is not\n> implemented\"\n\nsorry, typo. Please read:\n\"currently modification of the trigger row for DELETE RETURNING is notimplemented\"\n\n\n>> As far as allowing DELETE to modify the trigger row for RETURNING, I am\n>> not sure how much work it would take to allow that, but it seems like it\n>> is a valid requite, and if so, I can add it to the TODO list.\n\n> Yes, Add please into TODO the feature to \"allowing DELETE to modify the trigger row\n> for RETURNING\". Becuase, as I have described at first letter, without\n> this the RETURNING rows **does not correspond actually deleted data**\n\n> Thank you.\n\n\n\n\n-- \nBest regards,\nEugen Konkov\n\n\n\n","msg_date":"Thu, 7 Nov 2019 11:24:29 +0200","msg_from":"Eugen Konkov <kes-kes@yandex.ru>","msg_from_op":true,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"On Thu, Nov 7, 2019 at 11:24:29AM +0200, Eugen Konkov wrote:\n> Hello Eugen,\n> \n> Thursday, November 7, 2019, 11:20:32 AM, you wrote:\n> \n> >> I looked in the CREATE TRIGGER manual page and found this:\n> \n> >> https://www.postgresql.org/docs/12/sql-createtrigger.html\n> >> If the trigger fires before or instead of the event, the trigger\n> >> can skip the operation for the current row, or change the row\n> >> being inserted (for INSERT and UPDATE operations only).\n> \n> >> I don't see the \"(for INSERT and UPDATE operations only)\" language in\n> >> the main trigger documentation,\n> >> https://www.postgresql.org/docs/current/trigger-definition.html. I have\n> >> written the attached patch to fix that. Does that help?\n> \n> > No. If we document that PG does not allow to modify OLD at instead\n> > of trigger, the we can not implement that. Probably we can put note\n> > that \"currently modification of the trigger row for RETURNING is not\n> > implemented\"\n> \n> sorry, typo. Please read:\n> \"currently modification of the trigger row for DELETE RETURNING is notimplemented\"\n\nIn looking at the existing docs, the bullet above the quoted text says:\n\n\tFor row-level INSERT and UPDATE triggers only, the returned row becomes\n\t ----\n\tthe row that will be inserted or will replace the row being updated.\n\tThis allows the trigger function to modify the row being inserted or\n\tupdated.\n\nFirst, notice \"only\", which was missing from the later sentence:\n\n\tFor <command>INSERT</command> and <command>UPDATE</command>\n\toperations [only], the trigger may modify the\n\t<varname>NEW</varname> row before returning it.\n\nwhich I have now added with my applied patch to all supported releases. \n\nThe major use of modifying NEW is to modify the data that goes into the\ndatabase, and its use to modify data seen by later executed triggers, or\nby RETURNING, is only a side-effect of its primary purpose. Therefore,\nit is not surprising that, since DELETE does not modify any data, just\nremoves it, that the modification of OLD to appear in later triggers or\nRETURNING is not supported.\n\n> >> As far as allowing DELETE to modify the trigger row for RETURNING, I am\n> >> not sure how much work it would take to allow that, but it seems like it\n> >> is a valid requite, and if so, I can add it to the TODO list.\n> \n> > Yes, Add please into TODO the feature to \"allowing DELETE to modify the trigger row\n> > for RETURNING\". Becuase, as I have described at first letter, without\n> > this the RETURNING rows **does not correspond actually deleted data**\n> \n> > Thank you.\n\nI have added a TODO item:\n\n\tAllow DELETE triggers to modify rows, for use by RETURNING \n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Thu, 7 Nov 2019 16:26:55 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"On Thu, Nov 7, 2019 at 04:26:55PM -0500, Bruce Momjian wrote:\n> On Thu, Nov 7, 2019 at 11:24:29AM +0200, Eugen Konkov wrote:\n> > >> As far as allowing DELETE to modify the trigger row for RETURNING, I am\n> > >> not sure how much work it would take to allow that, but it seems like it\n> > >> is a valid requite, and if so, I can add it to the TODO list.\n> > \n> > > Yes, Add please into TODO the feature to \"allowing DELETE to modify the trigger row\n> > > for RETURNING\". Becuase, as I have described at first letter, without\n> > > this the RETURNING rows **does not correspond actually deleted data**\n> > \n> > > Thank you.\n> \n> I have added a TODO item:\n> \n> \tAllow DELETE triggers to modify rows, for use by RETURNING \n\nThinking some more on this, I now don't think a TODO makes sense, so I\nhave removed it.\n\nTriggers are designed to check and modify input data, and since DELETE\nhas no input data, it makes no sense. In the attached SQL script, you\ncan see that only the BEFORE INSERT trigger fires, so there is no way\neven with INSERT to change what is passed after the write to RETURNING. \nWhat you can do is to modify the returning expression, which is what I\nhave done for the last query --- hopefully that will help you.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +","msg_date":"Thu, 7 Nov 2019 17:28:18 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"Hello Bruce,\n\nFriday, November 8, 2019, 12:28:18 AM, you wrote:\n\n> On Thu, Nov 7, 2019 at 04:26:55PM -0500, Bruce Momjian wrote:\n>> On Thu, Nov 7, 2019 at 11:24:29AM +0200, Eugen Konkov wrote:\n>> > >> As far as allowing DELETE to modify the trigger row for RETURNING, I am\n>> > >> not sure how much work it would take to allow that, but it seems like it\n>> > >> is a valid requite, and if so, I can add it to the TODO list.\n>> > \n>> > > Yes, Add please into TODO the feature to \"allowing DELETE to modify the trigger row\n>> > > for RETURNING\". Becuase, as I have described at first letter, without\n>> > > this the RETURNING rows **does not correspond actually deleted data**\n>> > \n>> > > Thank you.\n>> \n>> I have added a TODO item:\n>> \n>> Allow DELETE triggers to modify rows, for use by RETURNING \n\n> Thinking some more on this, I now don't think a TODO makes sense, so I\n> have removed it.\n\n> Triggers are designed to check and modify input data, and since DELETE\n> has no input data, it makes no sense. In the attached SQL script, you\n> can see that only the BEFORE INSERT trigger fires, so there is no way\n> even with INSERT to change what is passed after the write to RETURNING.\n> What you can do is to modify the returning expression, which is what I\n> have done for the last query --- hopefully that will help you.\n\nYou lost my idea. First of all I am talking about views and an\nINSTEAD OF triggers.\n\nINSERT/UPDATE operation present which data is added into DB\nDELETE operation present which data is deleted from DB\n(in my case I am not deleted exact that data which matched by where.\nSee example below)\n\nThus INSTEAD OF INSERT/UPDATE triggers are designed to check and modify input data\neg. we can insert/update something different then incoming data (here\nwe are modifying NEW)\n\nThus INSTEAD OF DELETE triggers are designed to check and delete **output** data\neg. we can delete something different then underlaid data (here we are\nmodifying OLD)\n\nfor example, we have next data: 1 2 3 4 5 6 7 8\nit is not presented by eight rows, but instead it is presented as one\nrow with range data type: [1..8]\n\nWhen we insert data we will not get new row, we change current:\ninsert into table values ( 9 ) will result\n[1..9]\ninstead of\n[1..8]\n9\n\nSo lets look into INSTEAD OF DELETE trigger when we deleting\ndata:\ndelete from table where x in ( 5, 6, 7 );\nafter deleting this we should get:\n[1..4]\n[8..9]\n\nthus\nwith t1 as ( delete from table where x in ( 5, 6, 7 ) returning * )\nselect * from t1\nshould return:\n[5..7]\ninstead of\n[1..9]\nbecause we does not delete ALL [1..9], we just delete ONLY [5..7]\n\nThus I need to change matched row OLD.x from [1..9] to [5..7]\n\n\n\nPlease reread my first letter. There I describe more real life example\nwhen I am manipulating bi-temporal data.\n\nwhere some value exist at given period:\nid | app_period | value\n7 [2019-01-01, 2019-04-05) 207\n\nAnd I am deleting third month: [ 2019-03-01, 2019-04-01 )\nwith t1 as ( delete from table where app_period && [ 2019-03-01,\n2019-04-01 ) returning * )\nselect * from t1;\n7 [ 2019-03-01, 2019-04-01 ) 207\n\nselect * from table;\n7 [ 2019-01-01, 2019-03-01 ) 207\n7 [ 2019-04-01, 2019-04-05 ) 207\n\n\n\n\n-- \nBest regards,\nEugen Konkov\n\n\n\n","msg_date":"Sat, 9 Nov 2019 14:05:02 +0200","msg_from":"Eugen Konkov <kes-kes@yandex.ru>","msg_from_op":true,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"Hello Eugen,\n\nSaturday, November 9, 2019, 2:05:02 PM, you wrote:\n\n> Hello Bruce,\n\n> Friday, November 8, 2019, 12:28:18 AM, you wrote:\n\n>> On Thu, Nov 7, 2019 at 04:26:55PM -0500, Bruce Momjian wrote:\n>>> On Thu, Nov 7, 2019 at 11:24:29AM +0200, Eugen Konkov wrote:\n>>> > >> As far as allowing DELETE to modify the trigger row for RETURNING, I am\n>>> > >> not sure how much work it would take to allow that, but it seems like it\n>>> > >> is a valid requite, and if so, I can add it to the TODO list.\n>>> > \n>>> > > Yes, Add please into TODO the feature to \"allowing DELETE to modify the trigger row\n>>> > > for RETURNING\". Becuase, as I have described at first letter, without\n>>> > > this the RETURNING rows **does not correspond actually deleted data**\n>>> > \n>>> > > Thank you.\n>>> \n>>> I have added a TODO item:\n>>> \n>>> Allow DELETE triggers to modify rows, for use by RETURNING \n\n>> Thinking some more on this, I now don't think a TODO makes sense, so I\n>> have removed it.\n\n>> Triggers are designed to check and modify input data, and since DELETE\n>> has no input data, it makes no sense. In the attached SQL script, you\n>> can see that only the BEFORE INSERT trigger fires, so there is no way\n>> even with INSERT to change what is passed after the write to RETURNING.\n>> What you can do is to modify the returning expression, which is what I\n>> have done for the last query --- hopefully that will help you.\n\n> You lost my idea. First of all I am talking about views and an\n> INSTEAD OF triggers.\n\n> INSERT/UPDATE operation present which data is added into DB\n> DELETE operation present which data is deleted from DB\n> (in my case I am not deleted exact that data which matched by where.\n> See example below)\n\n> Thus INSTEAD OF INSERT/UPDATE triggers are designed to check and modify input data\n> eg. we can insert/update something different then incoming data (here\n> we are modifying NEW)\n\n> Thus INSTEAD OF DELETE triggers are designed to check and delete **output** data\n> eg. we can delete something different then underlaid data (here we are\n> modifying OLD)\n\n> for example, we have next data: 1 2 3 4 5 6 7 8\n> it is not presented by eight rows, but instead it is presented as one\n> row with range data type: [1..8]\n\n> When we insert data we will not get new row, we change current:\n> insert into table values ( 9 ) will result\n> [1..9]\n> instead of\n> [1..8]\n> 9\n\n> So lets look into INSTEAD OF DELETE trigger when we deleting\n> data:\n> delete from table where x in ( 5, 6, 7 );\n> after deleting this we should get:\n> [1..4]\n> [8..9]\n\n> thus\n> with t1 as ( delete from table where x in ( 5, 6, 7 ) returning * )\n> select * from t1\n> should return:\n> [5..7]\n> instead of\n> [1..9]\n> because we does not delete ALL [1..9], we just delete ONLY [5..7]\n\n> Thus I need to change matched row OLD.x from [1..9] to [5..7]\n\n\n\n> Please reread my first letter. There I describe more real life example\n> when I am manipulating bi-temporal data.\n\n> where some value exist at given period:\n> id | app_period | value\n> 7 [2019-01-01, 2019-04-05) 207\n\n> And I am deleting third month: [ 2019-03-01, 2019-04-01 )\n> with t1 as ( delete from table where app_period && [ 2019-03-01,\n> 2019-04-01 ) returning * )\n> select * from t1;\n> 7 [ 2019-03-01, 2019-04-01 ) 207\n\n> select * from table;\n> 7 [ 2019-01-01, 2019-03-01 ) 207\n> 7 [ 2019-04-01, 2019-04-05 ) 207\n\nHere when data is deleted the next row is matched:\n 7 [2019-01-01, 2019-04-05) 207\nand assigned to OLD.\nBecause I am deleting data ONLY from [ 2019-03-01, 2019-04-01 ) period\nI am required to change OLD:\n\nOLD.app_period = [ 2019-03-01, 2019-04-01 )\n\nSo I should get:\n> 7 [ 2019-03-01, 2019-04-01 ) 207\ninstead of\n> 7 [2019-01-01, 2019-04-05) 207\n\n\n\n\n\n-- \nBest regards,\nEugen Konkov\n\n\n\n","msg_date":"Sat, 9 Nov 2019 14:10:13 +0200","msg_from":"Eugen Konkov <kes-kes@yandex.ru>","msg_from_op":true,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"\n> 8 нояб. 2019 г., в 0:26, Bruce Momjian <bruce@momjian.us> написал(а):\n> \n> First, notice \"only\", which was missing from the later sentence:\n> \n> For <command>INSERT</command> and <command>UPDATE</command>\n> operations [only], the trigger may modify the\n> <varname>NEW</varname> row before returning it.\n> \n> which I have now added with my applied patch to all supported releases. \n> \n\nHi Bruce, \n\nI happened to browse recent documentation-related commits and I didn’t see this patch in REL_12_STABLE. Judging by the commit message, it should be applied there too.\n\n","msg_date":"Mon, 11 Nov 2019 19:00:22 +0300","msg_from":"Liudmila Mantrova <l.mantrova@gmail.com>","msg_from_op":false,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"On Mon, Nov 11, 2019 at 07:00:22PM +0300, Liudmila Mantrova wrote:\n> \n> > 8 нояб. 2019 г., в 0:26, Bruce Momjian <bruce@momjian.us> написал(а):\n> > \n> > First, notice \"only\", which was missing from the later sentence:\n> > \n> > For <command>INSERT</command> and <command>UPDATE</command>\n> > operations [only], the trigger may modify the\n> > <varname>NEW</varname> row before returning it.\n> > \n> > which I have now added with my applied patch to all supported releases. \n> > \n> \n> Hi Bruce, \n> \n> I happened to browse recent documentation-related commits and I didn’t see this patch in REL_12_STABLE. Judging by the commit message, it should be applied there too.\n\nWow, not sure how that happened, fixed.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Tue, 12 Nov 2019 22:04:59 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"Hi again.\n\n> Thinking some more on this, I now don't think a TODO makes sense, so I\n> have removed it.\n\nPlease look into this example: https://dbfiddle.uk/?rdbms=postgres_12&fiddle=95ed9fab6870d7c4b6266ea4d93def13\nThis is real life code from our production.\n\nYou can see that this is important to get correct info about deleted\ndata\n\n -- EXPECTED app_period: [\"2018-08-20\", \"2018-08-25\")\n -- ACTUAL app_period: [\"2018-08-14\", )\n\n> Triggers are designed to check and modify input data, and since DELETE\n> has no input data, it makes no sense.\n\nPlease put back into TODO list this feature request to allow\ntriggers to modify output data.\n\nINPUT -- receives data OK (behavior is expected)\nUPDATE -- receives and returns data OK (behavior is expected)\nDELETE -- returns data FAIL (behavior is not expected)\n\nThis is inconsistent to allow modify output data for UPDATE and\nrestrict to do this for DELETE\n\n\nThank you\n\n-- \nBest regards,\nEugen Konkov\n\n\n\n","msg_date":"Wed, 4 Dec 2019 13:37:57 +0200","msg_from":"Eugen Konkov <kes-kes@yandex.ru>","msg_from_op":true,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"Hello Eugen,\n\n> https://dbfiddle.uk/?rdbms=postgres_12&fiddle=95ed9fab6870d7c4b6266ea4d93def13\n\nsorry, forget to update link to the latest example:\nhttps://dbfiddle.uk/?rdbms=postgres_12&fiddle=8e114ccc9f15a30ca3115cdc6c70d247\n\n\n-- \nBest regards,\nEugen Konkov\n\n\n\n","msg_date":"Wed, 4 Dec 2019 13:39:43 +0200","msg_from":"Eugen Konkov <kes-kes@yandex.ru>","msg_from_op":true,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"},{"msg_contents":"Hello Bruce,\n\n> Triggers are designed to check and modify input data, and since DELETE\n> has no input data, it makes no sense.\n\nSorry, I am still ambiguous. You say that DELETE has no input data,\nbut doc says that it has:\n\nhttps://www.postgresql.org/docs/current/trigger-definition.html\nFor a row-level trigger, the input data also includes ... the OLD row for ... DELETE triggers\n\n\nAlso restricting DELETE to change the returned data by DELETE\nRETURNING seems as incomplete.\n\nFor example if triggers implement some compression.\n -- insert into field ZZZZZ value\n -- compress and actually store Zx5 into field\n -- Delete this insert row\n -- So user should get back that the value ZZZZZ was deleted and not Zx5.\n Correct?\n\n but currently user will see Zx5, because next code:\n\n OLD.value = uncompress( OLD.value );\n\n does not effect RETURNING =(\n\n-- \nBest regards,\nEugen Konkov\n\n\n\n","msg_date":"Fri, 17 Jan 2020 12:14:03 +0200","msg_from":"Eugen Konkov <kes-kes@yandex.ru>","msg_from_op":true,"msg_subject":"Re: Does 'instead of delete' trigger support modification of OLD"}],"string":"[\n {\n \"msg_contents\": \"Hi.\\n\\nThis is not clear from doc, so I have asked on IRC too.\\n\\nfrom the DOC: https://www.postgresql.org/docs/current/trigger-definition.html\\nIn the case of INSTEAD OF triggers, the possibly-modified row returned by each trigger becomes the input to the next trigger\\n\\nI modify OLD row, thus I expect to get modified version when run next query: \\n\\n WITH t1 AS( delete from abc returning *)\\n select * from t1;\\n\\nfiddle: https://dbfiddle.uk/?rdbms=postgres_12&fiddle=637730305f66bf531794edb09a462c95\\n\\n> https://www.postgresql.org/docs/current/trigger-definition.html\\nA row-level INSTEAD OF trigger should either return NULL to indicate that it did not modify any data from the view's underlying base tables,\\nor it should return the view row that was passed in (the NEW row for INSERT and UPDATE operations, or the OLD row for DELETE operations).\\nA nonnull return value is used to signal that the trigger performed the necessary data modifications in the view.\\nThis will cause the count of the number of rows affected by the command to be incremented. For INSERT and UPDATE operations, the trigger may\\nmodify the NEW row before returning it. This will change the data returned by INSERT RETURNING or UPDATE RETURNING,\\nand is useful when the view will not show exactly the same data that was provided.\\n\\nBut I still does not understand. Doc explicitly do not prohibit modification of OLD and has no examples for DELETE RETURNING case\\n\\nSo I want to ask clarify doc a bit.\\nIf this prohibited, why this is prohibited? have any discussion on this?\\nIf not prohibited, does this is not implemented for DELETE RETURNING queries? if so, is it left for later?\\n\\nI have next use case.\\nI am implementing Bi-Temporal tables. The table have columns: id, app_period, value\\nfor example I have next data: 7, '[2019-01-01, 2020-01-01)', 130\\nYou can imagine this as having value 7 for each day of the year.\\nNow I want to delete this value for May month. I setup special variable to period: '[2019-05-01,2019-06-01)' and then delete:\\n\\n select app_period( '[2019-05-01,2019-06-01)' );\\n WITH t1 AS( delete from abc returning *)\\n select * from t1;\\n\\nAlgorithm of deletion is next:\\n1. Deactivate target row\\n 7, '[2019-01-01, 2020-01-01)', 130\\n2. If target row has wider app_period then we insert record that data back:\\n NOT '[2019-05-01,2019-06-01)' @> '[2019-01-01, 2020-01-01)'\\n INSERT INTO abc ( id, app_period, value ) values \\n ( 7, '[2019-01-01,2019-05-01)', 130 ),\\n ( 7, '[2019-06-01,2020-01-01)', 130 ),\\n3. OLD.app_period = OLD.app_period * app_period(); \\n '[2019-01-01, 2020-01-01)' * '[2019-05-01,2019-06-01)' --> '[2019-05-01,2019-06-01)'\\n\\nBecause only 130 value is deleted from specified period I expect next result for the query above:\\n ( 7, '[2019-05-01,2019-06-01)', 130 )\\n\\nBut despite on OLD was modified, actual result is:\\n ( 7, '[2019-01-01,2020-01-01)', 130 )\\nYou can see that this is original data.\\n\\nSo, does INSTEAD OF DELETE support modification of row?\\n\\n-- \\nBest regards,\\nEugen Konkov\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 17:54:36 +0200\",\n \"msg_from\": \"Eugen Konkov <kes-kes@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 29, 2019 at 05:54:36PM +0200, Eugen Konkov wrote:\\n> Hi.\\n> \\n> This is not clear from doc, so I have asked on IRC too.\\n> \\n> from the DOC: https://www.postgresql.org/docs/current/trigger-definition.html\\n> In the case of INSTEAD OF triggers, the possibly-modified row returned by each trigger becomes the input to the next trigger\\n> \\n> I modify OLD row, thus I expect to get modified version when run next query: \\n> \\n> WITH t1 AS( delete from abc returning *)\\n> select * from t1;\\n> \\n> fiddle: https://dbfiddle.uk/?rdbms=postgres_12&fiddle=637730305f66bf531794edb09a462c95\\n\\nWow, that is a very nice way to present the queries.\\n\\n> > https://www.postgresql.org/docs/current/trigger-definition.html\\n> A row-level INSTEAD OF trigger should either return NULL to indicate that it did not modify any data from the view's underlying base tables,\\n> or it should return the view row that was passed in (the NEW row for INSERT and UPDATE operations, or the OLD row for DELETE operations).\\n> A nonnull return value is used to signal that the trigger performed the necessary data modifications in the view.\\n> This will cause the count of the number of rows affected by the command to be incremented. For INSERT and UPDATE operations, the trigger may\\n> modify the NEW row before returning it. This will change the data returned by INSERT RETURNING or UPDATE RETURNING,\\n> and is useful when the view will not show exactly the same data that was provided.\\n> \\n> But I still does not understand. Doc explicitly do not prohibit modification of OLD and has no examples for DELETE RETURNING case\\n\\nI looked in the CREATE TRIGGER manual page and found this:\\n\\n\\thttps://www.postgresql.org/docs/12/sql-createtrigger.html\\n\\tIf the trigger fires before or instead of the event, the trigger\\n\\tcan skip the operation for the current row, or change the row\\n\\tbeing inserted (for INSERT and UPDATE operations only).\\n\\nI don't see the \\\"(for INSERT and UPDATE operations only)\\\" language in\\nthe main trigger documentation,\\nhttps://www.postgresql.org/docs/current/trigger-definition.html. I have\\nwritten the attached patch to fix that. Does that help?\\n\\nAs far as allowing DELETE to modify the trigger row for RETURNING, I am\\nnot sure how much work it would take to allow that, but it seems like it\\nis a valid requite, and if so, I can add it to the TODO list.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\",\n \"msg_date\": \"Wed, 6 Nov 2019 13:59:35 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"> I looked in the CREATE TRIGGER manual page and found this:\\n\\n> https://www.postgresql.org/docs/12/sql-createtrigger.html\\n> If the trigger fires before or instead of the event, the trigger\\n> can skip the operation for the current row, or change the row\\n> being inserted (for INSERT and UPDATE operations only).\\n\\n> I don't see the \\\"(for INSERT and UPDATE operations only)\\\" language in\\n> the main trigger documentation,\\n> https://www.postgresql.org/docs/current/trigger-definition.html. I have\\n> written the attached patch to fix that. Does that help?\\n\\nNo. If we document that PG does not allow to modify OLD at instead\\nof trigger, the we can not implement that. Probably we can put note\\nthat \\\"currently modification of the trigger row for RETURNING is not\\nimplemented\\\"\\n\\n> As far as allowing DELETE to modify the trigger row for RETURNING, I am\\n> not sure how much work it would take to allow that, but it seems like it\\n> is a valid requite, and if so, I can add it to the TODO list.\\n\\nYes, Add please into TODO the feature to \\\"allowing DELETE to modify the trigger row\\nfor RETURNING\\\". Becuase, as I have described at first letter, without\\nthis the RETURNING rows **does not correspond actually deleted data**\\n\\nThank you.\\n\\n-- \\nBest regards,\\nEugen Konkov\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Nov 2019 11:20:32 +0200\",\n \"msg_from\": \"Eugen Konkov <kes-kes@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"Hello Eugen,\\n\\nThursday, November 7, 2019, 11:20:32 AM, you wrote:\\n\\n>> I looked in the CREATE TRIGGER manual page and found this:\\n\\n>> https://www.postgresql.org/docs/12/sql-createtrigger.html\\n>> If the trigger fires before or instead of the event, the trigger\\n>> can skip the operation for the current row, or change the row\\n>> being inserted (for INSERT and UPDATE operations only).\\n\\n>> I don't see the \\\"(for INSERT and UPDATE operations only)\\\" language in\\n>> the main trigger documentation,\\n>> https://www.postgresql.org/docs/current/trigger-definition.html. I have\\n>> written the attached patch to fix that. Does that help?\\n\\n> No. If we document that PG does not allow to modify OLD at instead\\n> of trigger, the we can not implement that. Probably we can put note\\n> that \\\"currently modification of the trigger row for RETURNING is not\\n> implemented\\\"\\n\\nsorry, typo. Please read:\\n\\\"currently modification of the trigger row for DELETE RETURNING is notimplemented\\\"\\n\\n\\n>> As far as allowing DELETE to modify the trigger row for RETURNING, I am\\n>> not sure how much work it would take to allow that, but it seems like it\\n>> is a valid requite, and if so, I can add it to the TODO list.\\n\\n> Yes, Add please into TODO the feature to \\\"allowing DELETE to modify the trigger row\\n> for RETURNING\\\". Becuase, as I have described at first letter, without\\n> this the RETURNING rows **does not correspond actually deleted data**\\n\\n> Thank you.\\n\\n\\n\\n\\n-- \\nBest regards,\\nEugen Konkov\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Nov 2019 11:24:29 +0200\",\n \"msg_from\": \"Eugen Konkov <kes-kes@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 7, 2019 at 11:24:29AM +0200, Eugen Konkov wrote:\\n> Hello Eugen,\\n> \\n> Thursday, November 7, 2019, 11:20:32 AM, you wrote:\\n> \\n> >> I looked in the CREATE TRIGGER manual page and found this:\\n> \\n> >> https://www.postgresql.org/docs/12/sql-createtrigger.html\\n> >> If the trigger fires before or instead of the event, the trigger\\n> >> can skip the operation for the current row, or change the row\\n> >> being inserted (for INSERT and UPDATE operations only).\\n> \\n> >> I don't see the \\\"(for INSERT and UPDATE operations only)\\\" language in\\n> >> the main trigger documentation,\\n> >> https://www.postgresql.org/docs/current/trigger-definition.html. I have\\n> >> written the attached patch to fix that. Does that help?\\n> \\n> > No. If we document that PG does not allow to modify OLD at instead\\n> > of trigger, the we can not implement that. Probably we can put note\\n> > that \\\"currently modification of the trigger row for RETURNING is not\\n> > implemented\\\"\\n> \\n> sorry, typo. Please read:\\n> \\\"currently modification of the trigger row for DELETE RETURNING is notimplemented\\\"\\n\\nIn looking at the existing docs, the bullet above the quoted text says:\\n\\n\\tFor row-level INSERT and UPDATE triggers only, the returned row becomes\\n\\t ----\\n\\tthe row that will be inserted or will replace the row being updated.\\n\\tThis allows the trigger function to modify the row being inserted or\\n\\tupdated.\\n\\nFirst, notice \\\"only\\\", which was missing from the later sentence:\\n\\n\\tFor <command>INSERT</command> and <command>UPDATE</command>\\n\\toperations [only], the trigger may modify the\\n\\t<varname>NEW</varname> row before returning it.\\n\\nwhich I have now added with my applied patch to all supported releases. \\n\\nThe major use of modifying NEW is to modify the data that goes into the\\ndatabase, and its use to modify data seen by later executed triggers, or\\nby RETURNING, is only a side-effect of its primary purpose. Therefore,\\nit is not surprising that, since DELETE does not modify any data, just\\nremoves it, that the modification of OLD to appear in later triggers or\\nRETURNING is not supported.\\n\\n> >> As far as allowing DELETE to modify the trigger row for RETURNING, I am\\n> >> not sure how much work it would take to allow that, but it seems like it\\n> >> is a valid requite, and if so, I can add it to the TODO list.\\n> \\n> > Yes, Add please into TODO the feature to \\\"allowing DELETE to modify the trigger row\\n> > for RETURNING\\\". Becuase, as I have described at first letter, without\\n> > this the RETURNING rows **does not correspond actually deleted data**\\n> \\n> > Thank you.\\n\\nI have added a TODO item:\\n\\n\\tAllow DELETE triggers to modify rows, for use by RETURNING \\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Nov 2019 16:26:55 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 7, 2019 at 04:26:55PM -0500, Bruce Momjian wrote:\\n> On Thu, Nov 7, 2019 at 11:24:29AM +0200, Eugen Konkov wrote:\\n> > >> As far as allowing DELETE to modify the trigger row for RETURNING, I am\\n> > >> not sure how much work it would take to allow that, but it seems like it\\n> > >> is a valid requite, and if so, I can add it to the TODO list.\\n> > \\n> > > Yes, Add please into TODO the feature to \\\"allowing DELETE to modify the trigger row\\n> > > for RETURNING\\\". Becuase, as I have described at first letter, without\\n> > > this the RETURNING rows **does not correspond actually deleted data**\\n> > \\n> > > Thank you.\\n> \\n> I have added a TODO item:\\n> \\n> \\tAllow DELETE triggers to modify rows, for use by RETURNING \\n\\nThinking some more on this, I now don't think a TODO makes sense, so I\\nhave removed it.\\n\\nTriggers are designed to check and modify input data, and since DELETE\\nhas no input data, it makes no sense. In the attached SQL script, you\\ncan see that only the BEFORE INSERT trigger fires, so there is no way\\neven with INSERT to change what is passed after the write to RETURNING. \\nWhat you can do is to modify the returning expression, which is what I\\nhave done for the last query --- hopefully that will help you.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\",\n \"msg_date\": \"Thu, 7 Nov 2019 17:28:18 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"Hello Bruce,\\n\\nFriday, November 8, 2019, 12:28:18 AM, you wrote:\\n\\n> On Thu, Nov 7, 2019 at 04:26:55PM -0500, Bruce Momjian wrote:\\n>> On Thu, Nov 7, 2019 at 11:24:29AM +0200, Eugen Konkov wrote:\\n>> > >> As far as allowing DELETE to modify the trigger row for RETURNING, I am\\n>> > >> not sure how much work it would take to allow that, but it seems like it\\n>> > >> is a valid requite, and if so, I can add it to the TODO list.\\n>> > \\n>> > > Yes, Add please into TODO the feature to \\\"allowing DELETE to modify the trigger row\\n>> > > for RETURNING\\\". Becuase, as I have described at first letter, without\\n>> > > this the RETURNING rows **does not correspond actually deleted data**\\n>> > \\n>> > > Thank you.\\n>> \\n>> I have added a TODO item:\\n>> \\n>> Allow DELETE triggers to modify rows, for use by RETURNING \\n\\n> Thinking some more on this, I now don't think a TODO makes sense, so I\\n> have removed it.\\n\\n> Triggers are designed to check and modify input data, and since DELETE\\n> has no input data, it makes no sense. In the attached SQL script, you\\n> can see that only the BEFORE INSERT trigger fires, so there is no way\\n> even with INSERT to change what is passed after the write to RETURNING.\\n> What you can do is to modify the returning expression, which is what I\\n> have done for the last query --- hopefully that will help you.\\n\\nYou lost my idea. First of all I am talking about views and an\\nINSTEAD OF triggers.\\n\\nINSERT/UPDATE operation present which data is added into DB\\nDELETE operation present which data is deleted from DB\\n(in my case I am not deleted exact that data which matched by where.\\nSee example below)\\n\\nThus INSTEAD OF INSERT/UPDATE triggers are designed to check and modify input data\\neg. we can insert/update something different then incoming data (here\\nwe are modifying NEW)\\n\\nThus INSTEAD OF DELETE triggers are designed to check and delete **output** data\\neg. we can delete something different then underlaid data (here we are\\nmodifying OLD)\\n\\nfor example, we have next data: 1 2 3 4 5 6 7 8\\nit is not presented by eight rows, but instead it is presented as one\\nrow with range data type: [1..8]\\n\\nWhen we insert data we will not get new row, we change current:\\ninsert into table values ( 9 ) will result\\n[1..9]\\ninstead of\\n[1..8]\\n9\\n\\nSo lets look into INSTEAD OF DELETE trigger when we deleting\\ndata:\\ndelete from table where x in ( 5, 6, 7 );\\nafter deleting this we should get:\\n[1..4]\\n[8..9]\\n\\nthus\\nwith t1 as ( delete from table where x in ( 5, 6, 7 ) returning * )\\nselect * from t1\\nshould return:\\n[5..7]\\ninstead of\\n[1..9]\\nbecause we does not delete ALL [1..9], we just delete ONLY [5..7]\\n\\nThus I need to change matched row OLD.x from [1..9] to [5..7]\\n\\n\\n\\nPlease reread my first letter. There I describe more real life example\\nwhen I am manipulating bi-temporal data.\\n\\nwhere some value exist at given period:\\nid | app_period | value\\n7 [2019-01-01, 2019-04-05) 207\\n\\nAnd I am deleting third month: [ 2019-03-01, 2019-04-01 )\\nwith t1 as ( delete from table where app_period && [ 2019-03-01,\\n2019-04-01 ) returning * )\\nselect * from t1;\\n7 [ 2019-03-01, 2019-04-01 ) 207\\n\\nselect * from table;\\n7 [ 2019-01-01, 2019-03-01 ) 207\\n7 [ 2019-04-01, 2019-04-05 ) 207\\n\\n\\n\\n\\n-- \\nBest regards,\\nEugen Konkov\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 9 Nov 2019 14:05:02 +0200\",\n \"msg_from\": \"Eugen Konkov <kes-kes@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"Hello Eugen,\\n\\nSaturday, November 9, 2019, 2:05:02 PM, you wrote:\\n\\n> Hello Bruce,\\n\\n> Friday, November 8, 2019, 12:28:18 AM, you wrote:\\n\\n>> On Thu, Nov 7, 2019 at 04:26:55PM -0500, Bruce Momjian wrote:\\n>>> On Thu, Nov 7, 2019 at 11:24:29AM +0200, Eugen Konkov wrote:\\n>>> > >> As far as allowing DELETE to modify the trigger row for RETURNING, I am\\n>>> > >> not sure how much work it would take to allow that, but it seems like it\\n>>> > >> is a valid requite, and if so, I can add it to the TODO list.\\n>>> > \\n>>> > > Yes, Add please into TODO the feature to \\\"allowing DELETE to modify the trigger row\\n>>> > > for RETURNING\\\". Becuase, as I have described at first letter, without\\n>>> > > this the RETURNING rows **does not correspond actually deleted data**\\n>>> > \\n>>> > > Thank you.\\n>>> \\n>>> I have added a TODO item:\\n>>> \\n>>> Allow DELETE triggers to modify rows, for use by RETURNING \\n\\n>> Thinking some more on this, I now don't think a TODO makes sense, so I\\n>> have removed it.\\n\\n>> Triggers are designed to check and modify input data, and since DELETE\\n>> has no input data, it makes no sense. In the attached SQL script, you\\n>> can see that only the BEFORE INSERT trigger fires, so there is no way\\n>> even with INSERT to change what is passed after the write to RETURNING.\\n>> What you can do is to modify the returning expression, which is what I\\n>> have done for the last query --- hopefully that will help you.\\n\\n> You lost my idea. First of all I am talking about views and an\\n> INSTEAD OF triggers.\\n\\n> INSERT/UPDATE operation present which data is added into DB\\n> DELETE operation present which data is deleted from DB\\n> (in my case I am not deleted exact that data which matched by where.\\n> See example below)\\n\\n> Thus INSTEAD OF INSERT/UPDATE triggers are designed to check and modify input data\\n> eg. we can insert/update something different then incoming data (here\\n> we are modifying NEW)\\n\\n> Thus INSTEAD OF DELETE triggers are designed to check and delete **output** data\\n> eg. we can delete something different then underlaid data (here we are\\n> modifying OLD)\\n\\n> for example, we have next data: 1 2 3 4 5 6 7 8\\n> it is not presented by eight rows, but instead it is presented as one\\n> row with range data type: [1..8]\\n\\n> When we insert data we will not get new row, we change current:\\n> insert into table values ( 9 ) will result\\n> [1..9]\\n> instead of\\n> [1..8]\\n> 9\\n\\n> So lets look into INSTEAD OF DELETE trigger when we deleting\\n> data:\\n> delete from table where x in ( 5, 6, 7 );\\n> after deleting this we should get:\\n> [1..4]\\n> [8..9]\\n\\n> thus\\n> with t1 as ( delete from table where x in ( 5, 6, 7 ) returning * )\\n> select * from t1\\n> should return:\\n> [5..7]\\n> instead of\\n> [1..9]\\n> because we does not delete ALL [1..9], we just delete ONLY [5..7]\\n\\n> Thus I need to change matched row OLD.x from [1..9] to [5..7]\\n\\n\\n\\n> Please reread my first letter. There I describe more real life example\\n> when I am manipulating bi-temporal data.\\n\\n> where some value exist at given period:\\n> id | app_period | value\\n> 7 [2019-01-01, 2019-04-05) 207\\n\\n> And I am deleting third month: [ 2019-03-01, 2019-04-01 )\\n> with t1 as ( delete from table where app_period && [ 2019-03-01,\\n> 2019-04-01 ) returning * )\\n> select * from t1;\\n> 7 [ 2019-03-01, 2019-04-01 ) 207\\n\\n> select * from table;\\n> 7 [ 2019-01-01, 2019-03-01 ) 207\\n> 7 [ 2019-04-01, 2019-04-05 ) 207\\n\\nHere when data is deleted the next row is matched:\\n 7 [2019-01-01, 2019-04-05) 207\\nand assigned to OLD.\\nBecause I am deleting data ONLY from [ 2019-03-01, 2019-04-01 ) period\\nI am required to change OLD:\\n\\nOLD.app_period = [ 2019-03-01, 2019-04-01 )\\n\\nSo I should get:\\n> 7 [ 2019-03-01, 2019-04-01 ) 207\\ninstead of\\n> 7 [2019-01-01, 2019-04-05) 207\\n\\n\\n\\n\\n\\n-- \\nBest regards,\\nEugen Konkov\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 9 Nov 2019 14:10:13 +0200\",\n \"msg_from\": \"Eugen Konkov <kes-kes@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"\\n> 8 нояб. 2019 г., в 0:26, Bruce Momjian <bruce@momjian.us> написал(а):\\n> \\n> First, notice \\\"only\\\", which was missing from the later sentence:\\n> \\n> For <command>INSERT</command> and <command>UPDATE</command>\\n> operations [only], the trigger may modify the\\n> <varname>NEW</varname> row before returning it.\\n> \\n> which I have now added with my applied patch to all supported releases. \\n> \\n\\nHi Bruce, \\n\\nI happened to browse recent documentation-related commits and I didn’t see this patch in REL_12_STABLE. Judging by the commit message, it should be applied there too.\\n\\n\",\n \"msg_date\": \"Mon, 11 Nov 2019 19:00:22 +0300\",\n \"msg_from\": \"Liudmila Mantrova <l.mantrova@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 11, 2019 at 07:00:22PM +0300, Liudmila Mantrova wrote:\\n> \\n> > 8 нояб. 2019 г., в 0:26, Bruce Momjian <bruce@momjian.us> написал(а):\\n> > \\n> > First, notice \\\"only\\\", which was missing from the later sentence:\\n> > \\n> > For <command>INSERT</command> and <command>UPDATE</command>\\n> > operations [only], the trigger may modify the\\n> > <varname>NEW</varname> row before returning it.\\n> > \\n> > which I have now added with my applied patch to all supported releases. \\n> > \\n> \\n> Hi Bruce, \\n> \\n> I happened to browse recent documentation-related commits and I didn’t see this patch in REL_12_STABLE. Judging by the commit message, it should be applied there too.\\n\\nWow, not sure how that happened, fixed.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Nov 2019 22:04:59 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"Hi again.\\n\\n> Thinking some more on this, I now don't think a TODO makes sense, so I\\n> have removed it.\\n\\nPlease look into this example: https://dbfiddle.uk/?rdbms=postgres_12&fiddle=95ed9fab6870d7c4b6266ea4d93def13\\nThis is real life code from our production.\\n\\nYou can see that this is important to get correct info about deleted\\ndata\\n\\n -- EXPECTED app_period: [\\\"2018-08-20\\\", \\\"2018-08-25\\\")\\n -- ACTUAL app_period: [\\\"2018-08-14\\\", )\\n\\n> Triggers are designed to check and modify input data, and since DELETE\\n> has no input data, it makes no sense.\\n\\nPlease put back into TODO list this feature request to allow\\ntriggers to modify output data.\\n\\nINPUT -- receives data OK (behavior is expected)\\nUPDATE -- receives and returns data OK (behavior is expected)\\nDELETE -- returns data FAIL (behavior is not expected)\\n\\nThis is inconsistent to allow modify output data for UPDATE and\\nrestrict to do this for DELETE\\n\\n\\nThank you\\n\\n-- \\nBest regards,\\nEugen Konkov\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 4 Dec 2019 13:37:57 +0200\",\n \"msg_from\": \"Eugen Konkov <kes-kes@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"Hello Eugen,\\n\\n> https://dbfiddle.uk/?rdbms=postgres_12&fiddle=95ed9fab6870d7c4b6266ea4d93def13\\n\\nsorry, forget to update link to the latest example:\\nhttps://dbfiddle.uk/?rdbms=postgres_12&fiddle=8e114ccc9f15a30ca3115cdc6c70d247\\n\\n\\n-- \\nBest regards,\\nEugen Konkov\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 4 Dec 2019 13:39:43 +0200\",\n \"msg_from\": \"Eugen Konkov <kes-kes@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n },\n {\n \"msg_contents\": \"Hello Bruce,\\n\\n> Triggers are designed to check and modify input data, and since DELETE\\n> has no input data, it makes no sense.\\n\\nSorry, I am still ambiguous. You say that DELETE has no input data,\\nbut doc says that it has:\\n\\nhttps://www.postgresql.org/docs/current/trigger-definition.html\\nFor a row-level trigger, the input data also includes ... the OLD row for ... DELETE triggers\\n\\n\\nAlso restricting DELETE to change the returned data by DELETE\\nRETURNING seems as incomplete.\\n\\nFor example if triggers implement some compression.\\n -- insert into field ZZZZZ value\\n -- compress and actually store Zx5 into field\\n -- Delete this insert row\\n -- So user should get back that the value ZZZZZ was deleted and not Zx5.\\n Correct?\\n\\n but currently user will see Zx5, because next code:\\n\\n OLD.value = uncompress( OLD.value );\\n\\n does not effect RETURNING =(\\n\\n-- \\nBest regards,\\nEugen Konkov\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 17 Jan 2020 12:14:03 +0200\",\n \"msg_from\": \"Eugen Konkov <kes-kes@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Does 'instead of delete' trigger support modification of OLD\"\n }\n]","isConversational":false}}},{"rowIdx":49361,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\none of the most frequent conflicts I see is that two patches add files\nto OBJS (or one of its other spellings), and there are conflicts because\nanother file has been added.\n\nRight now there's two reasons why that's likely to happen:\n1) By listing multiple objects for each line, we get a conflict whenever\n one of the other files on that lines gets modified\n2) Due to our line-length conventions, we have to re-flow long lines,\n which often triggers reflowing subsequent lines too.\n\nNow, obviously these types of conflicts are easy enough to resolve, but\nit's still annoying. It seems that this would be substantially less\noften a problem if we just split such lines to one file per\nline. E.g. instead of\n\nOBJS_COMMON = base64.o config_info.o controldata_utils.o d2s.o exec.o f2s.o \\\n\tfile_perm.o ip.o keywords.o kwlookup.o link-canary.o md5.o \\\n\tpg_lzcompress.o pgfnames.o psprintf.o relpath.o \\\n\trmtree.o saslprep.o scram-common.o string.o stringinfo.o \\\n\tunicode_norm.o username.o wait_error.o\n\nhave\n\nOBJS_COMMON = \\\n\tbase64.o \\\n\tconfig_info.o \\\n\tcontroldata_utils.o \\\n\td2s.o \\\n\texec.o \\\n\tf2s.o \\\n\tfile_perm.o \\\n\tip.o \\\n\tkeywords.o \\\n\tkwlookup.o \\\n\tlink-canary.o \\\n\tmd5.o \\\n\tpg_lzcompress.o \\\n\tpgfnames.o \\\n\tpsprintf.o \\\n\trelpath.o \\\n\trmtree.o \\\n\tsaslprep.o \\\n\tscram-common.o \\\n\tstring.o \\\n\tstringinfo.o \\\n\tunicode_norm.o \\\n\tusername.o \\\n\twait_error.o\n\na one-off conversion of this seems easy enough to script.\n\nComments?\n\n- Andres\n\n\n","msg_date":"Tue, 29 Oct 2019 13:09:01 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"RFC: split OBJS lines to one object per line"},{"msg_contents":"On Tue, Oct 29, 2019 at 1:09 PM Andres Freund <andres@anarazel.de> wrote:\n> Comments?\n\nI think that this is a good idea. I see no downside.\n\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Tue, 29 Oct 2019 13:16:49 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> one of the most frequent conflicts I see is that two patches add files\n> to OBJS (or one of its other spellings), and there are conflicts because\n> another file has been added.\n> ...\n> Now, obviously these types of conflicts are easy enough to resolve, but\n> it's still annoying. It seems that this would be substantially less\n> often a problem if we just split such lines to one file per\n> line.\n\nWe did something similar not too long ago in configure.in (bfa6c5a0c),\nand it seems to have helped. +1\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 29 Oct 2019 16:31:11 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"Hi,\n\nOn 2019-10-29 16:31:11 -0400, Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n> > one of the most frequent conflicts I see is that two patches add files\n> > to OBJS (or one of its other spellings), and there are conflicts because\n> > another file has been added.\n> > ...\n> > Now, obviously these types of conflicts are easy enough to resolve, but\n> > it's still annoying. It seems that this would be substantially less\n> > often a problem if we just split such lines to one file per\n> > line.\n> \n> We did something similar not too long ago in configure.in (bfa6c5a0c),\n> and it seems to have helped. +1\n\nCool. Any opinion on whether to got for\n\nOBJS = \\\n\tdest.o \\\n\tfastpath.o \\\n...\n\nor\n\nOBJS = dest.o \\\n\tfastpath.o \\\n...\n\nI'm mildly inclined to go for the former.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 29 Oct 2019 23:32:09 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2019-10-29 16:31:11 -0400, Tom Lane wrote:\n>> We did something similar not too long ago in configure.in (bfa6c5a0c),\n>> and it seems to have helped. +1\n\n> Cool. Any opinion on whether to got for ...\n\nNot here.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 30 Oct 2019 02:56:14 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"On Tue, Oct 29, 2019 at 11:32:09PM -0700, Andres Freund wrote:\n> Cool. Any opinion on whether to got for\n> \n> OBJS = \\\n> \tdest.o \\\n> \tfastpath.o \\\n> ...\n> \n> or\n> \n> OBJS = dest.o \\\n> \tfastpath.o \\\n> ...\n> \n> I'm mildly inclined to go for the former.\n\nFWIW, I am more used to the latter, but the former is also fine by\nme.\n--\nMichael","msg_date":"Wed, 30 Oct 2019 16:40:45 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"\n\nOn 10/29/19 11:32 PM, Andres Freund wrote:\n> Hi,\n> \n> On 2019-10-29 16:31:11 -0400, Tom Lane wrote:\n>> Andres Freund <andres@anarazel.de> writes:\n>>> one of the most frequent conflicts I see is that two patches add files\n>>> to OBJS (or one of its other spellings), and there are conflicts because\n>>> another file has been added.\n>>> ...\n>>> Now, obviously these types of conflicts are easy enough to resolve, but\n>>> it's still annoying. It seems that this would be substantially less\n>>> often a problem if we just split such lines to one file per\n>>> line.\n>>\n>> We did something similar not too long ago in configure.in (bfa6c5a0c),\n>> and it seems to have helped. +1\n> \n> Cool. Any opinion on whether to got for\n> \n> OBJS = \\\n> \tdest.o \\\n> \tfastpath.o \\\n> ...\n> \n> or\n> \n> OBJS = dest.o \\\n> \tfastpath.o \\\n> ...\n> \n> I'm mildly inclined to go for the former.\n\n+1 for the former.\n\n> \n> Greetings,\n> \n> Andres Freund\n> \n> \n\n\n","msg_date":"Thu, 31 Oct 2019 09:48:46 -0700","msg_from":"Mark Dilger <hornschnorter@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"Hi,\n\nOn 2019-10-29 23:32:09 -0700, Andres Freund wrote:\n> On 2019-10-29 16:31:11 -0400, Tom Lane wrote:\n> > Andres Freund <andres@anarazel.de> writes:\n> > > one of the most frequent conflicts I see is that two patches add files\n> > > to OBJS (or one of its other spellings), and there are conflicts because\n> > > another file has been added.\n> > > ...\n> > > Now, obviously these types of conflicts are easy enough to resolve, but\n> > > it's still annoying. It seems that this would be substantially less\n> > > often a problem if we just split such lines to one file per\n> > > line.\n> > \n> > We did something similar not too long ago in configure.in (bfa6c5a0c),\n> > and it seems to have helped. +1\n> \n> Cool. Any opinion on whether to got for\n> \n> OBJS = \\\n> \tdest.o \\\n> \tfastpath.o \\\n> ...\n> \n> or\n> \n> OBJS = dest.o \\\n> \tfastpath.o \\\n> ...\n> \n> I'm mildly inclined to go for the former.\n\nPushed a patch going with the former. Let's see what the buildfarm\nsays...\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 5 Nov 2019 14:47:55 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"On Tue, Nov 05, 2019 at 02:47:55PM -0800, Andres Freund wrote:\n> Pushed a patch going with the former. Let's see what the buildfarm\n> says...\n\nThanks Andres. On a rather related note, would it make sense to do\nthe same for regression and isolation tests in our in-core modules?\n--\nMichael","msg_date":"Thu, 7 Nov 2019 11:24:37 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Tue, Nov 05, 2019 at 02:47:55PM -0800, Andres Freund wrote:\n>> Pushed a patch going with the former. Let's see what the buildfarm\n>> says...\n\n> Thanks Andres. On a rather related note, would it make sense to do\n> the same for regression and isolation tests in our in-core modules?\n\nI don't think it'd be a great idea to change parallel_schedule like\nthat. Independently adding test scripts to the same parallel batch\nprobably won't end well: you might end up over the concurrency limit,\nor the scripts might conflict through sharing table names or the like.\nSo I'd rather see that there's a conflict to worry about.\n\nAnyway, merge conflicts there aren't so common IME.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 07 Nov 2019 12:02:04 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"Hi,\n\nOn 2019-11-07 11:24:37 +0900, Michael Paquier wrote:\n> On Tue, Nov 05, 2019 at 02:47:55PM -0800, Andres Freund wrote:\n> > Pushed a patch going with the former. Let's see what the buildfarm\n> > says...\n> \n> Thanks Andres. On a rather related note, would it make sense to do\n> the same for regression and isolation tests in our in-core modules?\n\nI don't see them as being frequent sources of conflicts (partially\nbecause we don't change linebreaks due to line length limits, I think),\nso I don't think it's really worthwhile.\n\nOne I could see some benefit in, would be the SUBDIRS makefile lines.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 7 Nov 2019 09:20:02 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"On Thu, Nov 07, 2019 at 12:02:04PM -0500, Tom Lane wrote:\n> I don't think it'd be a great idea to change parallel_schedule like\n> that. Independently adding test scripts to the same parallel batch\n> probably won't end well: you might end up over the concurrency limit,\n> or the scripts might conflict through sharing table names or the like.\n> So I'd rather see that there's a conflict to worry about.\n> \n> Anyway, merge conflicts there aren't so common IME.\n\nFWIW, I was not referring to the schedule files here, just to REGRESS\nand ISOLATION in the modules' Makefiles. If you think that's not\nworth doing it, let's drop my suggestion then.\n--\nMichael","msg_date":"Fri, 8 Nov 2019 18:07:57 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"On Fri, 8 Nov 2019 at 14:38, Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Thu, Nov 07, 2019 at 12:02:04PM -0500, Tom Lane wrote:\n> > I don't think it'd be a great idea to change parallel_schedule like\n> > that. Independently adding test scripts to the same parallel batch\n> > probably won't end well: you might end up over the concurrency limit,\n> > or the scripts might conflict through sharing table names or the like.\n> > So I'd rather see that there's a conflict to worry about.\n> >\n> > Anyway, merge conflicts there aren't so common IME.\n>\n> FWIW, I was not referring to the schedule files here, just to REGRESS\n> and ISOLATION in the modules' Makefiles. If you think that's not\n> worth doing it, let's drop my suggestion then.\n> --\n\nI found some inconsistency in alphabetical order in\nsrc/backend/tsearch/Makefile, src/backend/utils/Makefile and\nsrc/pl/plpython/Makefile files. Attached patch is fixing those order\nrelated inconsistency.\n\nThanks and Regards\nMahendra Thalor\nEnterpriseDB: http://www.enterprisedb.com","msg_date":"Tue, 17 Dec 2019 23:40:17 +0530","msg_from":"Mahendra Singh <mahi6run@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"On Tue, Dec 17, 2019 at 11:40:17PM +0530, Mahendra Singh wrote:\n> I found some inconsistency in alphabetical order in\n> src/backend/tsearch/Makefile, src/backend/utils/Makefile and\n> src/pl/plpython/Makefile files. Attached patch is fixing those order\n> related inconsistency.\n\nThanks, committed. The one-liner style is also used in ifaddrs, but\nfmgrtab.c is generated so I have left that out. Now, have you tried\nto compile plpython before sending this patch? Because as you forgot\nto add one backslash after WIN32RES, compilation was failing there.\nAnd you also forgot to remove two backslashes at the end of the other\ntwo lists modified :)\n--\nMichael","msg_date":"Wed, 18 Dec 2019 10:53:37 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: RFC: split OBJS lines to one object per line"},{"msg_contents":"On Wed, 18 Dec 2019 at 07:23, Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Tue, Dec 17, 2019 at 11:40:17PM +0530, Mahendra Singh wrote:\n> > I found some inconsistency in alphabetical order in\n> > src/backend/tsearch/Makefile, src/backend/utils/Makefile and\n> > src/pl/plpython/Makefile files. Attached patch is fixing those order\n> > related inconsistency.\n>\n> Thanks, committed. The one-liner style is also used in ifaddrs, but\n\nThanks Michael for quick response.\n\n> fmgrtab.c is generated so I have left that out. Now, have you tried\n> to compile plpython before sending this patch? Because as you forgot\n> to add one backslash after WIN32RES, compilation was failing there.\n> And you also forgot to remove two backslashes at the end of the other\n> two lists modified :)\n\nSorry, I forgot to add backslashes. I will take care from next time.\n\nThanks and Regards\nMahendra Thalor\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 18 Dec 2019 08:30:24 +0530","msg_from":"Mahendra Singh <mahi6run@gmail.com>","msg_from_op":false,"msg_subject":"Re: RFC: split OBJS lines to one object per line"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\none of the most frequent conflicts I see is that two patches add files\\nto OBJS (or one of its other spellings), and there are conflicts because\\nanother file has been added.\\n\\nRight now there's two reasons why that's likely to happen:\\n1) By listing multiple objects for each line, we get a conflict whenever\\n one of the other files on that lines gets modified\\n2) Due to our line-length conventions, we have to re-flow long lines,\\n which often triggers reflowing subsequent lines too.\\n\\nNow, obviously these types of conflicts are easy enough to resolve, but\\nit's still annoying. It seems that this would be substantially less\\noften a problem if we just split such lines to one file per\\nline. E.g. instead of\\n\\nOBJS_COMMON = base64.o config_info.o controldata_utils.o d2s.o exec.o f2s.o \\\\\\n\\tfile_perm.o ip.o keywords.o kwlookup.o link-canary.o md5.o \\\\\\n\\tpg_lzcompress.o pgfnames.o psprintf.o relpath.o \\\\\\n\\trmtree.o saslprep.o scram-common.o string.o stringinfo.o \\\\\\n\\tunicode_norm.o username.o wait_error.o\\n\\nhave\\n\\nOBJS_COMMON = \\\\\\n\\tbase64.o \\\\\\n\\tconfig_info.o \\\\\\n\\tcontroldata_utils.o \\\\\\n\\td2s.o \\\\\\n\\texec.o \\\\\\n\\tf2s.o \\\\\\n\\tfile_perm.o \\\\\\n\\tip.o \\\\\\n\\tkeywords.o \\\\\\n\\tkwlookup.o \\\\\\n\\tlink-canary.o \\\\\\n\\tmd5.o \\\\\\n\\tpg_lzcompress.o \\\\\\n\\tpgfnames.o \\\\\\n\\tpsprintf.o \\\\\\n\\trelpath.o \\\\\\n\\trmtree.o \\\\\\n\\tsaslprep.o \\\\\\n\\tscram-common.o \\\\\\n\\tstring.o \\\\\\n\\tstringinfo.o \\\\\\n\\tunicode_norm.o \\\\\\n\\tusername.o \\\\\\n\\twait_error.o\\n\\na one-off conversion of this seems easy enough to script.\\n\\nComments?\\n\\n- Andres\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 13:09:01 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 29, 2019 at 1:09 PM Andres Freund <andres@anarazel.de> wrote:\\n> Comments?\\n\\nI think that this is a good idea. I see no downside.\\n\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 13:16:49 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> one of the most frequent conflicts I see is that two patches add files\\n> to OBJS (or one of its other spellings), and there are conflicts because\\n> another file has been added.\\n> ...\\n> Now, obviously these types of conflicts are easy enough to resolve, but\\n> it's still annoying. It seems that this would be substantially less\\n> often a problem if we just split such lines to one file per\\n> line.\\n\\nWe did something similar not too long ago in configure.in (bfa6c5a0c),\\nand it seems to have helped. +1\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 16:31:11 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-10-29 16:31:11 -0400, Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n> > one of the most frequent conflicts I see is that two patches add files\\n> > to OBJS (or one of its other spellings), and there are conflicts because\\n> > another file has been added.\\n> > ...\\n> > Now, obviously these types of conflicts are easy enough to resolve, but\\n> > it's still annoying. It seems that this would be substantially less\\n> > often a problem if we just split such lines to one file per\\n> > line.\\n> \\n> We did something similar not too long ago in configure.in (bfa6c5a0c),\\n> and it seems to have helped. +1\\n\\nCool. Any opinion on whether to got for\\n\\nOBJS = \\\\\\n\\tdest.o \\\\\\n\\tfastpath.o \\\\\\n...\\n\\nor\\n\\nOBJS = dest.o \\\\\\n\\tfastpath.o \\\\\\n...\\n\\nI'm mildly inclined to go for the former.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 23:32:09 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2019-10-29 16:31:11 -0400, Tom Lane wrote:\\n>> We did something similar not too long ago in configure.in (bfa6c5a0c),\\n>> and it seems to have helped. +1\\n\\n> Cool. Any opinion on whether to got for ...\\n\\nNot here.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 02:56:14 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 29, 2019 at 11:32:09PM -0700, Andres Freund wrote:\\n> Cool. Any opinion on whether to got for\\n> \\n> OBJS = \\\\\\n> \\tdest.o \\\\\\n> \\tfastpath.o \\\\\\n> ...\\n> \\n> or\\n> \\n> OBJS = dest.o \\\\\\n> \\tfastpath.o \\\\\\n> ...\\n> \\n> I'm mildly inclined to go for the former.\\n\\nFWIW, I am more used to the latter, but the former is also fine by\\nme.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 30 Oct 2019 16:40:45 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 10/29/19 11:32 PM, Andres Freund wrote:\\n> Hi,\\n> \\n> On 2019-10-29 16:31:11 -0400, Tom Lane wrote:\\n>> Andres Freund <andres@anarazel.de> writes:\\n>>> one of the most frequent conflicts I see is that two patches add files\\n>>> to OBJS (or one of its other spellings), and there are conflicts because\\n>>> another file has been added.\\n>>> ...\\n>>> Now, obviously these types of conflicts are easy enough to resolve, but\\n>>> it's still annoying. It seems that this would be substantially less\\n>>> often a problem if we just split such lines to one file per\\n>>> line.\\n>>\\n>> We did something similar not too long ago in configure.in (bfa6c5a0c),\\n>> and it seems to have helped. +1\\n> \\n> Cool. Any opinion on whether to got for\\n> \\n> OBJS = \\\\\\n> \\tdest.o \\\\\\n> \\tfastpath.o \\\\\\n> ...\\n> \\n> or\\n> \\n> OBJS = dest.o \\\\\\n> \\tfastpath.o \\\\\\n> ...\\n> \\n> I'm mildly inclined to go for the former.\\n\\n+1 for the former.\\n\\n> \\n> Greetings,\\n> \\n> Andres Freund\\n> \\n> \\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 09:48:46 -0700\",\n \"msg_from\": \"Mark Dilger <hornschnorter@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-10-29 23:32:09 -0700, Andres Freund wrote:\\n> On 2019-10-29 16:31:11 -0400, Tom Lane wrote:\\n> > Andres Freund <andres@anarazel.de> writes:\\n> > > one of the most frequent conflicts I see is that two patches add files\\n> > > to OBJS (or one of its other spellings), and there are conflicts because\\n> > > another file has been added.\\n> > > ...\\n> > > Now, obviously these types of conflicts are easy enough to resolve, but\\n> > > it's still annoying. It seems that this would be substantially less\\n> > > often a problem if we just split such lines to one file per\\n> > > line.\\n> > \\n> > We did something similar not too long ago in configure.in (bfa6c5a0c),\\n> > and it seems to have helped. +1\\n> \\n> Cool. Any opinion on whether to got for\\n> \\n> OBJS = \\\\\\n> \\tdest.o \\\\\\n> \\tfastpath.o \\\\\\n> ...\\n> \\n> or\\n> \\n> OBJS = dest.o \\\\\\n> \\tfastpath.o \\\\\\n> ...\\n> \\n> I'm mildly inclined to go for the former.\\n\\nPushed a patch going with the former. Let's see what the buildfarm\\nsays...\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Nov 2019 14:47:55 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 05, 2019 at 02:47:55PM -0800, Andres Freund wrote:\\n> Pushed a patch going with the former. Let's see what the buildfarm\\n> says...\\n\\nThanks Andres. On a rather related note, would it make sense to do\\nthe same for regression and isolation tests in our in-core modules?\\n--\\nMichael\",\n \"msg_date\": \"Thu, 7 Nov 2019 11:24:37 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Tue, Nov 05, 2019 at 02:47:55PM -0800, Andres Freund wrote:\\n>> Pushed a patch going with the former. Let's see what the buildfarm\\n>> says...\\n\\n> Thanks Andres. On a rather related note, would it make sense to do\\n> the same for regression and isolation tests in our in-core modules?\\n\\nI don't think it'd be a great idea to change parallel_schedule like\\nthat. Independently adding test scripts to the same parallel batch\\nprobably won't end well: you might end up over the concurrency limit,\\nor the scripts might conflict through sharing table names or the like.\\nSo I'd rather see that there's a conflict to worry about.\\n\\nAnyway, merge conflicts there aren't so common IME.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Nov 2019 12:02:04 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-11-07 11:24:37 +0900, Michael Paquier wrote:\\n> On Tue, Nov 05, 2019 at 02:47:55PM -0800, Andres Freund wrote:\\n> > Pushed a patch going with the former. Let's see what the buildfarm\\n> > says...\\n> \\n> Thanks Andres. On a rather related note, would it make sense to do\\n> the same for regression and isolation tests in our in-core modules?\\n\\nI don't see them as being frequent sources of conflicts (partially\\nbecause we don't change linebreaks due to line length limits, I think),\\nso I don't think it's really worthwhile.\\n\\nOne I could see some benefit in, would be the SUBDIRS makefile lines.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Nov 2019 09:20:02 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 07, 2019 at 12:02:04PM -0500, Tom Lane wrote:\\n> I don't think it'd be a great idea to change parallel_schedule like\\n> that. Independently adding test scripts to the same parallel batch\\n> probably won't end well: you might end up over the concurrency limit,\\n> or the scripts might conflict through sharing table names or the like.\\n> So I'd rather see that there's a conflict to worry about.\\n> \\n> Anyway, merge conflicts there aren't so common IME.\\n\\nFWIW, I was not referring to the schedule files here, just to REGRESS\\nand ISOLATION in the modules' Makefiles. If you think that's not\\nworth doing it, let's drop my suggestion then.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 8 Nov 2019 18:07:57 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"On Fri, 8 Nov 2019 at 14:38, Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Thu, Nov 07, 2019 at 12:02:04PM -0500, Tom Lane wrote:\\n> > I don't think it'd be a great idea to change parallel_schedule like\\n> > that. Independently adding test scripts to the same parallel batch\\n> > probably won't end well: you might end up over the concurrency limit,\\n> > or the scripts might conflict through sharing table names or the like.\\n> > So I'd rather see that there's a conflict to worry about.\\n> >\\n> > Anyway, merge conflicts there aren't so common IME.\\n>\\n> FWIW, I was not referring to the schedule files here, just to REGRESS\\n> and ISOLATION in the modules' Makefiles. If you think that's not\\n> worth doing it, let's drop my suggestion then.\\n> --\\n\\nI found some inconsistency in alphabetical order in\\nsrc/backend/tsearch/Makefile, src/backend/utils/Makefile and\\nsrc/pl/plpython/Makefile files. Attached patch is fixing those order\\nrelated inconsistency.\\n\\nThanks and Regards\\nMahendra Thalor\\nEnterpriseDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Tue, 17 Dec 2019 23:40:17 +0530\",\n \"msg_from\": \"Mahendra Singh <mahi6run@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 17, 2019 at 11:40:17PM +0530, Mahendra Singh wrote:\\n> I found some inconsistency in alphabetical order in\\n> src/backend/tsearch/Makefile, src/backend/utils/Makefile and\\n> src/pl/plpython/Makefile files. Attached patch is fixing those order\\n> related inconsistency.\\n\\nThanks, committed. The one-liner style is also used in ifaddrs, but\\nfmgrtab.c is generated so I have left that out. Now, have you tried\\nto compile plpython before sending this patch? Because as you forgot\\nto add one backslash after WIN32RES, compilation was failing there.\\nAnd you also forgot to remove two backslashes at the end of the other\\ntwo lists modified :)\\n--\\nMichael\",\n \"msg_date\": \"Wed, 18 Dec 2019 10:53:37 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n },\n {\n \"msg_contents\": \"On Wed, 18 Dec 2019 at 07:23, Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Tue, Dec 17, 2019 at 11:40:17PM +0530, Mahendra Singh wrote:\\n> > I found some inconsistency in alphabetical order in\\n> > src/backend/tsearch/Makefile, src/backend/utils/Makefile and\\n> > src/pl/plpython/Makefile files. Attached patch is fixing those order\\n> > related inconsistency.\\n>\\n> Thanks, committed. The one-liner style is also used in ifaddrs, but\\n\\nThanks Michael for quick response.\\n\\n> fmgrtab.c is generated so I have left that out. Now, have you tried\\n> to compile plpython before sending this patch? Because as you forgot\\n> to add one backslash after WIN32RES, compilation was failing there.\\n> And you also forgot to remove two backslashes at the end of the other\\n> two lists modified :)\\n\\nSorry, I forgot to add backslashes. I will take care from next time.\\n\\nThanks and Regards\\nMahendra Thalor\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 18 Dec 2019 08:30:24 +0530\",\n \"msg_from\": \"Mahendra Singh <mahi6run@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: RFC: split OBJS lines to one object per line\"\n }\n]","isConversational":false}}},{"rowIdx":49362,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nWhen using -b, --bkp-details pg_waldump outputs an unnecessary newline\nfor blocks that contain an FPW.\n\nIn --bkp-details block references are output on their own lines, like:\n\nrmgr: SPGist len (rec/tot): 4348/ 4348, tx: 980, lsn: 0/01985818, prev 0/01983850, desc: PICKSPLIT ndel 92; nins 93\n blkref #0: rel 1663/16384/16967 fork main blk 3\n blkref #1: rel 1663/16384/16967 fork main blk 6\n blkref #2: rel 1663/16384/16967 fork main blk 5\n blkref #3: rel 1663/16384/16967 fork main blk 1\nrmgr: Heap len (rec/tot): 69/ 69, tx: 980, lsn: 0/01986930, prev 0/01985818, desc: INSERT off 2 flags 0x00\n blkref #0: rel 1663/16384/16961 fork main blk 1\n\nbut unfortunately, when there's actually an FPW present, it looks like:\n\nrmgr: XLOG len (rec/tot): 75/ 11199, tx: 977, lsn: 0/019755E0, prev 0/0194EDD8, desc: FPI\n blkref #0: rel 1663/16384/16960 fork main blk 32 (FPW); hole: offset: 548, length: 4484\n\n blkref #1: rel 1663/16384/16960 fork main blk 33 (FPW); hole: offset: 548, length: 4484\n\n blkref #2: rel 1663/16384/16960 fork main blk 34 (FPW); hole: offset: 548, length: 4484\n\nrmgr: Heap len (rec/tot): 188/ 188, tx: 977, lsn: 0/019781D0, prev 0/019755E0, desc: INPLACE off 23\n\nwhich clearly seems unnecessary. Looking at the code it seems to me that\n\nstatic void\nXLogDumpDisplayRecord(XLogDumpConfig *config, XLogReaderState *record)\n{\n...\n printf(\"\\tblkref #%u: rel %u/%u/%u fork %s blk %u\",\n block_id,\n rnode.spcNode, rnode.dbNode, rnode.relNode,\n forkNames[forknum],\n blk);\n if (XLogRecHasBlockImage(record, block_id))\n {\n if (record->blocks[block_id].bimg_info &\n BKPIMAGE_IS_COMPRESSED)\n {\n printf(\" (FPW%s); hole: offset: %u, length: %u, \"\n \"compression saved: %u\\n\",\n XLogRecBlockImageApply(record, block_id) ?\n \"\" : \" for WAL verification\",\n record->blocks[block_id].hole_offset,\n record->blocks[block_id].hole_length,\n BLCKSZ -\n record->blocks[block_id].hole_length -\n record->blocks[block_id].bimg_len);\n }\n else\n {\n printf(\" (FPW%s); hole: offset: %u, length: %u\\n\",\n XLogRecBlockImageApply(record, block_id) ?\n \"\" : \" for WAL verification\",\n record->blocks[block_id].hole_offset,\n record->blocks[block_id].hole_length);\n }\n }\n putchar('\\n');\n\nwas intended to not actually print a newline in the printfs in the if\npreceding the putchar.\n\nThis is a fairly longstanding bug, introduced in:\n\ncommit 2c03216d831160bedd72d45f712601b6f7d03f1c\nAuthor: Heikki Linnakangas <heikki.linnakangas@iki.fi>\nDate: 2014-11-20 17:56:26 +0200\n\n Revamp the WAL record format.\n\n\nDoes anybody have an opinion about fixing it just in master or also\nbackpatching it? I guess there could be people having written parsers\nfor the waldump output? I'm inclined to backpatch.\n\n\nI also find a second minor bug:\n\nstatic void\nXLogDumpDisplayRecord(XLogDumpConfig *config, XLogReaderState *record)\n{\n...\n const char *id;\n...\n id = desc->rm_identify(info);\n if (id == NULL)\n id = psprintf(\"UNKNOWN (%x)\", info & ~XLR_INFO_MASK);\n...\n printf(\"desc: %s \", id);\n\nafter that \"id\" is not referenced anymore. Which means we would leak\nmemory if there were a lot of UNKNOWN records. This is from\ncommit 604f7956b9460192222dd37bd3baea24cb669a47\nAuthor: Andres Freund <andres@anarazel.de>\nDate: 2014-09-22 16:48:14 +0200\n\n Improve code around the recently added rm_identify rmgr callback.\n\nWhile not a lot of memory, it's not absurd to run pg_waldump against a\nlarge amount of WAL, so backpatching seems mildly advised.\n\nI'm inlined to think that the best fix is to just move the relevant code\nto the callsite, and not psprintf'ing into a temporary buffer. We'd need\nadditional state to free the memory, as rm_identify returns a static\nbuffer.\n\nSo I'll make it\n\n id = desc->rm_identify(info);\n if (id == NULL)\n printf(\"desc: UNKNOWN (%x) \", info & ~XLR_INFO_MASK);\n else\n printf(\"desc: %s \", id);\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 29 Oct 2019 16:33:41 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"pg_waldump erroneously outputs newline for FPWs, and another minor\n bug"},{"msg_contents":"On Tue, Oct 29, 2019 at 4:33 PM Andres Freund <andres@anarazel.de> wrote:\n> Does anybody have an opinion about fixing it just in master or also\n> backpatching it? I guess there could be people having written parsers\n> for the waldump output? I'm inclined to backpatch.\n\nThe same commit from Heikki omitted one field from that record, for no\ngood reason. I backpatched a bugfix to the output format for nbtree\npage splits a few weeks ago, fixing that problem. I agree that we\nshould also backpatch this bugfix.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Tue, 29 Oct 2019 16:42:07 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: pg_waldump erroneously outputs newline for FPWs, and another\n minor bug"},{"msg_contents":"On Tue, Oct 29, 2019 at 04:42:07PM -0700, Peter Geoghegan wrote:\n> The same commit from Heikki omitted one field from that record, for no\n> good reason. I backpatched a bugfix to the output format for nbtree\n> page splits a few weeks ago, fixing that problem. I agree that we\n> should also backpatch this bugfix.\n\nThe output format of pg_waldump may matter for some tools, like\nJehan-Guillaume's PAF [1], but I am ready to bet that any tools like\nthat just skip any noise newlines, so +1 for a backpatch.\n\nI am adding Jehan-Guillaume in CC just in case.\n\n[1]: https://github.com/ClusterLabs/PAF\n--\nMichael","msg_date":"Wed, 30 Oct 2019 09:26:21 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_waldump erroneously outputs newline for FPWs, and another\n minor bug"},{"msg_contents":"Hi,\n\nOn 2019-10-29 16:33:41 -0700, Andres Freund wrote:\n> Hi,\n> \n> When using -b, --bkp-details pg_waldump outputs an unnecessary newline\n> for blocks that contain an FPW.\n> \n> In --bkp-details block references are output on their own lines, like:\n> \n> rmgr: SPGist len (rec/tot): 4348/ 4348, tx: 980, lsn: 0/01985818, prev 0/01983850, desc: PICKSPLIT ndel 92; nins 93\n> blkref #0: rel 1663/16384/16967 fork main blk 3\n> blkref #1: rel 1663/16384/16967 fork main blk 6\n> blkref #2: rel 1663/16384/16967 fork main blk 5\n> blkref #3: rel 1663/16384/16967 fork main blk 1\n> rmgr: Heap len (rec/tot): 69/ 69, tx: 980, lsn: 0/01986930, prev 0/01985818, desc: INSERT off 2 flags 0x00\n> blkref #0: rel 1663/16384/16961 fork main blk 1\n> \n> but unfortunately, when there's actually an FPW present, it looks like:\n> \n> rmgr: XLOG len (rec/tot): 75/ 11199, tx: 977, lsn: 0/019755E0, prev 0/0194EDD8, desc: FPI\n> blkref #0: rel 1663/16384/16960 fork main blk 32 (FPW); hole: offset: 548, length: 4484\n> \n> blkref #1: rel 1663/16384/16960 fork main blk 33 (FPW); hole: offset: 548, length: 4484\n> \n> blkref #2: rel 1663/16384/16960 fork main blk 34 (FPW); hole: offset: 548, length: 4484\n> \n> rmgr: Heap len (rec/tot): 188/ 188, tx: 977, lsn: 0/019781D0, prev 0/019755E0, desc: INPLACE off 23\n> \n> which clearly seems unnecessary. Looking at the code it seems to me that\n> \n> static void\n> XLogDumpDisplayRecord(XLogDumpConfig *config, XLogReaderState *record)\n> {\n> ...\n> printf(\"\\tblkref #%u: rel %u/%u/%u fork %s blk %u\",\n> block_id,\n> rnode.spcNode, rnode.dbNode, rnode.relNode,\n> forkNames[forknum],\n> blk);\n> if (XLogRecHasBlockImage(record, block_id))\n> {\n> if (record->blocks[block_id].bimg_info &\n> BKPIMAGE_IS_COMPRESSED)\n> {\n> printf(\" (FPW%s); hole: offset: %u, length: %u, \"\n> \"compression saved: %u\\n\",\n> XLogRecBlockImageApply(record, block_id) ?\n> \"\" : \" for WAL verification\",\n> record->blocks[block_id].hole_offset,\n> record->blocks[block_id].hole_length,\n> BLCKSZ -\n> record->blocks[block_id].hole_length -\n> record->blocks[block_id].bimg_len);\n> }\n> else\n> {\n> printf(\" (FPW%s); hole: offset: %u, length: %u\\n\",\n> XLogRecBlockImageApply(record, block_id) ?\n> \"\" : \" for WAL verification\",\n> record->blocks[block_id].hole_offset,\n> record->blocks[block_id].hole_length);\n> }\n> }\n> putchar('\\n');\n> \n> was intended to not actually print a newline in the printfs in the if\n> preceding the putchar.\n> \n> This is a fairly longstanding bug, introduced in:\n> \n> commit 2c03216d831160bedd72d45f712601b6f7d03f1c\n> Author: Heikki Linnakangas <heikki.linnakangas@iki.fi>\n> Date: 2014-11-20 17:56:26 +0200\n> \n> Revamp the WAL record format.\n> \n> \n> Does anybody have an opinion about fixing it just in master or also\n> backpatching it? I guess there could be people having written parsers\n> for the waldump output? I'm inclined to backpatch.\n> \n> \n> I also find a second minor bug:\n> \n> static void\n> XLogDumpDisplayRecord(XLogDumpConfig *config, XLogReaderState *record)\n> {\n> ...\n> const char *id;\n> ...\n> id = desc->rm_identify(info);\n> if (id == NULL)\n> id = psprintf(\"UNKNOWN (%x)\", info & ~XLR_INFO_MASK);\n> ...\n> printf(\"desc: %s \", id);\n> \n> after that \"id\" is not referenced anymore. Which means we would leak\n> memory if there were a lot of UNKNOWN records. This is from\n> commit 604f7956b9460192222dd37bd3baea24cb669a47\n> Author: Andres Freund <andres@anarazel.de>\n> Date: 2014-09-22 16:48:14 +0200\n> \n> Improve code around the recently added rm_identify rmgr callback.\n> \n> While not a lot of memory, it's not absurd to run pg_waldump against a\n> large amount of WAL, so backpatching seems mildly advised.\n> \n> I'm inlined to think that the best fix is to just move the relevant code\n> to the callsite, and not psprintf'ing into a temporary buffer. We'd need\n> additional state to free the memory, as rm_identify returns a static\n> buffer.\n> \n> So I'll make it\n> \n> id = desc->rm_identify(info);\n> if (id == NULL)\n> printf(\"desc: UNKNOWN (%x) \", info & ~XLR_INFO_MASK);\n> else\n> printf(\"desc: %s \", id);\n\nPushed fixes for these.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 29 Oct 2019 22:59:42 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: pg_waldump erroneously outputs newline for FPWs, and another\n minor bug"},{"msg_contents":"On Wed, 30 Oct 2019 09:26:21 +0900\nMichael Paquier <michael@paquier.xyz> wrote:\n\n> On Tue, Oct 29, 2019 at 04:42:07PM -0700, Peter Geoghegan wrote:\n> > The same commit from Heikki omitted one field from that record, for no\n> > good reason. I backpatched a bugfix to the output format for nbtree\n> > page splits a few weeks ago, fixing that problem. I agree that we\n> > should also backpatch this bugfix. \n> \n> The output format of pg_waldump may matter for some tools, like\n> Jehan-Guillaume's PAF [1], but I am ready to bet that any tools like\n> that just skip any noise newlines, so +1 for a backpatch.\n> \n> I am adding Jehan-Guillaume in CC just in case.\n\nThank you Michael!\n\n\n","msg_date":"Mon, 4 Nov 2019 12:42:35 +0100","msg_from":"Jehan-Guillaume de Rorthais <jgdr@dalibo.com>","msg_from_op":false,"msg_subject":"Re: pg_waldump erroneously outputs newline for FPWs, and another\n minor bug"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nWhen using -b, --bkp-details pg_waldump outputs an unnecessary newline\\nfor blocks that contain an FPW.\\n\\nIn --bkp-details block references are output on their own lines, like:\\n\\nrmgr: SPGist len (rec/tot): 4348/ 4348, tx: 980, lsn: 0/01985818, prev 0/01983850, desc: PICKSPLIT ndel 92; nins 93\\n blkref #0: rel 1663/16384/16967 fork main blk 3\\n blkref #1: rel 1663/16384/16967 fork main blk 6\\n blkref #2: rel 1663/16384/16967 fork main blk 5\\n blkref #3: rel 1663/16384/16967 fork main blk 1\\nrmgr: Heap len (rec/tot): 69/ 69, tx: 980, lsn: 0/01986930, prev 0/01985818, desc: INSERT off 2 flags 0x00\\n blkref #0: rel 1663/16384/16961 fork main blk 1\\n\\nbut unfortunately, when there's actually an FPW present, it looks like:\\n\\nrmgr: XLOG len (rec/tot): 75/ 11199, tx: 977, lsn: 0/019755E0, prev 0/0194EDD8, desc: FPI\\n blkref #0: rel 1663/16384/16960 fork main blk 32 (FPW); hole: offset: 548, length: 4484\\n\\n blkref #1: rel 1663/16384/16960 fork main blk 33 (FPW); hole: offset: 548, length: 4484\\n\\n blkref #2: rel 1663/16384/16960 fork main blk 34 (FPW); hole: offset: 548, length: 4484\\n\\nrmgr: Heap len (rec/tot): 188/ 188, tx: 977, lsn: 0/019781D0, prev 0/019755E0, desc: INPLACE off 23\\n\\nwhich clearly seems unnecessary. Looking at the code it seems to me that\\n\\nstatic void\\nXLogDumpDisplayRecord(XLogDumpConfig *config, XLogReaderState *record)\\n{\\n...\\n printf(\\\"\\\\tblkref #%u: rel %u/%u/%u fork %s blk %u\\\",\\n block_id,\\n rnode.spcNode, rnode.dbNode, rnode.relNode,\\n forkNames[forknum],\\n blk);\\n if (XLogRecHasBlockImage(record, block_id))\\n {\\n if (record->blocks[block_id].bimg_info &\\n BKPIMAGE_IS_COMPRESSED)\\n {\\n printf(\\\" (FPW%s); hole: offset: %u, length: %u, \\\"\\n \\\"compression saved: %u\\\\n\\\",\\n XLogRecBlockImageApply(record, block_id) ?\\n \\\"\\\" : \\\" for WAL verification\\\",\\n record->blocks[block_id].hole_offset,\\n record->blocks[block_id].hole_length,\\n BLCKSZ -\\n record->blocks[block_id].hole_length -\\n record->blocks[block_id].bimg_len);\\n }\\n else\\n {\\n printf(\\\" (FPW%s); hole: offset: %u, length: %u\\\\n\\\",\\n XLogRecBlockImageApply(record, block_id) ?\\n \\\"\\\" : \\\" for WAL verification\\\",\\n record->blocks[block_id].hole_offset,\\n record->blocks[block_id].hole_length);\\n }\\n }\\n putchar('\\\\n');\\n\\nwas intended to not actually print a newline in the printfs in the if\\npreceding the putchar.\\n\\nThis is a fairly longstanding bug, introduced in:\\n\\ncommit 2c03216d831160bedd72d45f712601b6f7d03f1c\\nAuthor: Heikki Linnakangas <heikki.linnakangas@iki.fi>\\nDate: 2014-11-20 17:56:26 +0200\\n\\n Revamp the WAL record format.\\n\\n\\nDoes anybody have an opinion about fixing it just in master or also\\nbackpatching it? I guess there could be people having written parsers\\nfor the waldump output? I'm inclined to backpatch.\\n\\n\\nI also find a second minor bug:\\n\\nstatic void\\nXLogDumpDisplayRecord(XLogDumpConfig *config, XLogReaderState *record)\\n{\\n...\\n const char *id;\\n...\\n id = desc->rm_identify(info);\\n if (id == NULL)\\n id = psprintf(\\\"UNKNOWN (%x)\\\", info & ~XLR_INFO_MASK);\\n...\\n printf(\\\"desc: %s \\\", id);\\n\\nafter that \\\"id\\\" is not referenced anymore. Which means we would leak\\nmemory if there were a lot of UNKNOWN records. This is from\\ncommit 604f7956b9460192222dd37bd3baea24cb669a47\\nAuthor: Andres Freund <andres@anarazel.de>\\nDate: 2014-09-22 16:48:14 +0200\\n\\n Improve code around the recently added rm_identify rmgr callback.\\n\\nWhile not a lot of memory, it's not absurd to run pg_waldump against a\\nlarge amount of WAL, so backpatching seems mildly advised.\\n\\nI'm inlined to think that the best fix is to just move the relevant code\\nto the callsite, and not psprintf'ing into a temporary buffer. We'd need\\nadditional state to free the memory, as rm_identify returns a static\\nbuffer.\\n\\nSo I'll make it\\n\\n id = desc->rm_identify(info);\\n if (id == NULL)\\n printf(\\\"desc: UNKNOWN (%x) \\\", info & ~XLR_INFO_MASK);\\n else\\n printf(\\\"desc: %s \\\", id);\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 16:33:41 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pg_waldump erroneously outputs newline for FPWs, and another minor\\n bug\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 29, 2019 at 4:33 PM Andres Freund <andres@anarazel.de> wrote:\\n> Does anybody have an opinion about fixing it just in master or also\\n> backpatching it? I guess there could be people having written parsers\\n> for the waldump output? I'm inclined to backpatch.\\n\\nThe same commit from Heikki omitted one field from that record, for no\\ngood reason. I backpatched a bugfix to the output format for nbtree\\npage splits a few weeks ago, fixing that problem. I agree that we\\nshould also backpatch this bugfix.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 16:42:07 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump erroneously outputs newline for FPWs, and another\\n minor bug\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 29, 2019 at 04:42:07PM -0700, Peter Geoghegan wrote:\\n> The same commit from Heikki omitted one field from that record, for no\\n> good reason. I backpatched a bugfix to the output format for nbtree\\n> page splits a few weeks ago, fixing that problem. I agree that we\\n> should also backpatch this bugfix.\\n\\nThe output format of pg_waldump may matter for some tools, like\\nJehan-Guillaume's PAF [1], but I am ready to bet that any tools like\\nthat just skip any noise newlines, so +1 for a backpatch.\\n\\nI am adding Jehan-Guillaume in CC just in case.\\n\\n[1]: https://github.com/ClusterLabs/PAF\\n--\\nMichael\",\n \"msg_date\": \"Wed, 30 Oct 2019 09:26:21 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump erroneously outputs newline for FPWs, and another\\n minor bug\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-10-29 16:33:41 -0700, Andres Freund wrote:\\n> Hi,\\n> \\n> When using -b, --bkp-details pg_waldump outputs an unnecessary newline\\n> for blocks that contain an FPW.\\n> \\n> In --bkp-details block references are output on their own lines, like:\\n> \\n> rmgr: SPGist len (rec/tot): 4348/ 4348, tx: 980, lsn: 0/01985818, prev 0/01983850, desc: PICKSPLIT ndel 92; nins 93\\n> blkref #0: rel 1663/16384/16967 fork main blk 3\\n> blkref #1: rel 1663/16384/16967 fork main blk 6\\n> blkref #2: rel 1663/16384/16967 fork main blk 5\\n> blkref #3: rel 1663/16384/16967 fork main blk 1\\n> rmgr: Heap len (rec/tot): 69/ 69, tx: 980, lsn: 0/01986930, prev 0/01985818, desc: INSERT off 2 flags 0x00\\n> blkref #0: rel 1663/16384/16961 fork main blk 1\\n> \\n> but unfortunately, when there's actually an FPW present, it looks like:\\n> \\n> rmgr: XLOG len (rec/tot): 75/ 11199, tx: 977, lsn: 0/019755E0, prev 0/0194EDD8, desc: FPI\\n> blkref #0: rel 1663/16384/16960 fork main blk 32 (FPW); hole: offset: 548, length: 4484\\n> \\n> blkref #1: rel 1663/16384/16960 fork main blk 33 (FPW); hole: offset: 548, length: 4484\\n> \\n> blkref #2: rel 1663/16384/16960 fork main blk 34 (FPW); hole: offset: 548, length: 4484\\n> \\n> rmgr: Heap len (rec/tot): 188/ 188, tx: 977, lsn: 0/019781D0, prev 0/019755E0, desc: INPLACE off 23\\n> \\n> which clearly seems unnecessary. Looking at the code it seems to me that\\n> \\n> static void\\n> XLogDumpDisplayRecord(XLogDumpConfig *config, XLogReaderState *record)\\n> {\\n> ...\\n> printf(\\\"\\\\tblkref #%u: rel %u/%u/%u fork %s blk %u\\\",\\n> block_id,\\n> rnode.spcNode, rnode.dbNode, rnode.relNode,\\n> forkNames[forknum],\\n> blk);\\n> if (XLogRecHasBlockImage(record, block_id))\\n> {\\n> if (record->blocks[block_id].bimg_info &\\n> BKPIMAGE_IS_COMPRESSED)\\n> {\\n> printf(\\\" (FPW%s); hole: offset: %u, length: %u, \\\"\\n> \\\"compression saved: %u\\\\n\\\",\\n> XLogRecBlockImageApply(record, block_id) ?\\n> \\\"\\\" : \\\" for WAL verification\\\",\\n> record->blocks[block_id].hole_offset,\\n> record->blocks[block_id].hole_length,\\n> BLCKSZ -\\n> record->blocks[block_id].hole_length -\\n> record->blocks[block_id].bimg_len);\\n> }\\n> else\\n> {\\n> printf(\\\" (FPW%s); hole: offset: %u, length: %u\\\\n\\\",\\n> XLogRecBlockImageApply(record, block_id) ?\\n> \\\"\\\" : \\\" for WAL verification\\\",\\n> record->blocks[block_id].hole_offset,\\n> record->blocks[block_id].hole_length);\\n> }\\n> }\\n> putchar('\\\\n');\\n> \\n> was intended to not actually print a newline in the printfs in the if\\n> preceding the putchar.\\n> \\n> This is a fairly longstanding bug, introduced in:\\n> \\n> commit 2c03216d831160bedd72d45f712601b6f7d03f1c\\n> Author: Heikki Linnakangas <heikki.linnakangas@iki.fi>\\n> Date: 2014-11-20 17:56:26 +0200\\n> \\n> Revamp the WAL record format.\\n> \\n> \\n> Does anybody have an opinion about fixing it just in master or also\\n> backpatching it? I guess there could be people having written parsers\\n> for the waldump output? I'm inclined to backpatch.\\n> \\n> \\n> I also find a second minor bug:\\n> \\n> static void\\n> XLogDumpDisplayRecord(XLogDumpConfig *config, XLogReaderState *record)\\n> {\\n> ...\\n> const char *id;\\n> ...\\n> id = desc->rm_identify(info);\\n> if (id == NULL)\\n> id = psprintf(\\\"UNKNOWN (%x)\\\", info & ~XLR_INFO_MASK);\\n> ...\\n> printf(\\\"desc: %s \\\", id);\\n> \\n> after that \\\"id\\\" is not referenced anymore. Which means we would leak\\n> memory if there were a lot of UNKNOWN records. This is from\\n> commit 604f7956b9460192222dd37bd3baea24cb669a47\\n> Author: Andres Freund <andres@anarazel.de>\\n> Date: 2014-09-22 16:48:14 +0200\\n> \\n> Improve code around the recently added rm_identify rmgr callback.\\n> \\n> While not a lot of memory, it's not absurd to run pg_waldump against a\\n> large amount of WAL, so backpatching seems mildly advised.\\n> \\n> I'm inlined to think that the best fix is to just move the relevant code\\n> to the callsite, and not psprintf'ing into a temporary buffer. We'd need\\n> additional state to free the memory, as rm_identify returns a static\\n> buffer.\\n> \\n> So I'll make it\\n> \\n> id = desc->rm_identify(info);\\n> if (id == NULL)\\n> printf(\\\"desc: UNKNOWN (%x) \\\", info & ~XLR_INFO_MASK);\\n> else\\n> printf(\\\"desc: %s \\\", id);\\n\\nPushed fixes for these.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 22:59:42 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_waldump erroneously outputs newline for FPWs, and another\\n minor bug\"\n },\n {\n \"msg_contents\": \"On Wed, 30 Oct 2019 09:26:21 +0900\\nMichael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Tue, Oct 29, 2019 at 04:42:07PM -0700, Peter Geoghegan wrote:\\n> > The same commit from Heikki omitted one field from that record, for no\\n> > good reason. I backpatched a bugfix to the output format for nbtree\\n> > page splits a few weeks ago, fixing that problem. I agree that we\\n> > should also backpatch this bugfix. \\n> \\n> The output format of pg_waldump may matter for some tools, like\\n> Jehan-Guillaume's PAF [1], but I am ready to bet that any tools like\\n> that just skip any noise newlines, so +1 for a backpatch.\\n> \\n> I am adding Jehan-Guillaume in CC just in case.\\n\\nThank you Michael!\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Nov 2019 12:42:35 +0100\",\n \"msg_from\": \"Jehan-Guillaume de Rorthais <jgdr@dalibo.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_waldump erroneously outputs newline for FPWs, and another\\n minor bug\"\n }\n]","isConversational":false}}},{"rowIdx":49363,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nThis patch, in a slightly rougher form, was submitted as part of [1],\nbut it seems worth bringing up separately, rather than just committing\nhearing no objections.\n\n From the commit message:\n\n Make StringInfo available to frontend code.\n\n There's plenty places in frontend code that could benefit from a\n string buffer implementation. Some because it yields simpler and\n faster code, and some others because of the desire to share code\n between backend and frontend.\n\n While there is a string buffer implementation available to frontend\n code, libpq's PQExpBuffer, it is clunkier than stringinfo, it\n introduces a libpq dependency, doesn't allow for sharing between\n frontend and backend code, and has a higher API/ABI stability\n requirement due to being exposed via libpq.\n\n Therefore it seems best to just making StringInfo being usable by\n frontend code. There's not much to do for that, except for rewriting\n two subsequent elog/ereport calls into others types of error\n reporting, and deciding on a maximum string length.\n\n For the maximum string size I decided to privately define MaxAllocSize\n to the same value as used in the backend. It seems likely that we'll\n want to reconsider this for both backend and frontend code in the not\n too far away future.\n\n For now I've left stringinfo.h in lib/, rather than common/, to reduce\n the likelihood of unnecessary breakage. We could alternatively decide\n to provide a redirecting stringinfo.h in lib/, or just not provide\n compatibility.\n\nI'm still using stringinfo in the aforementioned thread, and I also want\nto use it in a few more places. On the more ambitious side I really\nwould like to have a minimal version of elog.h available in the backend,\nand that would really be a lot easier with stringinfo available.\n\nI also would like to submit a few patches expanding stringinfo's\ncapabilities and performance, and it seems to me it'd be better to do so\nafter moving (lest they introduce new FE vs BE compat issues).\n\n\nThis allows us to remove compat.c hackery providing some stringinfo\nfunctionality for pg_waldump (which now actually needs to pass in a\nStringInfo...). I briefly played with converting more code in\npg_waldump.c than just that one call to StringInfo, but it seems that'd\nbe best done separately.\n\nComments?\n\nGreetings,\n\nAndres Freund\n\n[1] https://postgr.es/m/20190920051857.2fhnvhvx4qdddviz@alap3.anarazel.de","msg_date":"Tue, 29 Oct 2019 17:10:01 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Make StringInfo available to frontend code."},{"msg_contents":"At Tue, 29 Oct 2019 17:10:01 -0700, Andres Freund <andres@anarazel.de> wrote in \n> Hi,\n> \n> This patch, in a slightly rougher form, was submitted as part of [1],\n> but it seems worth bringing up separately, rather than just committing\n> hearing no objections.\n..\n> I'm still using stringinfo in the aforementioned thread, and I also want\n> to use it in a few more places. On the more ambitious side I really\n> would like to have a minimal version of elog.h available in the backend,\n> and that would really be a lot easier with stringinfo available.\n> \n> I also would like to submit a few patches expanding stringinfo's\n> capabilities and performance, and it seems to me it'd be better to do so\n> after moving (lest they introduce new FE vs BE compat issues).\n> \n> \n> This allows us to remove compat.c hackery providing some stringinfo\n> functionality for pg_waldump (which now actually needs to pass in a\n> StringInfo...). I briefly played with converting more code in\n> pg_waldump.c than just that one call to StringInfo, but it seems that'd\n> be best done separately.\n> \n> Comments?\n\nIt uses different form for the same message for FE and BE.\n\ncommon/stringinfo.c:289-\n> BE:\tereport(ERROR,\n> \t\t\t(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),\n> \t\t\t errmsg(\"out of memory\"),\n> \t\t\t errdetail(\"Cannot enlarge string buffer containing %d\n> bytes by %d more bytes.\",\n> \n> FE: +\t\t_(\"out of memory\\n\\nCannot enlarge string buffer containing %d\n> bytes by %d more bytes.\\n\"),\n\n.po files will be smaller and more stable if we keep the same\ntranslation unit for the same messages. That being said it doesn't\nmatter if it is tentative and the minimal elog.h for frontend comes\nsoon.\n\n\n> /* It's possible we could use a different value for this in frontend code */\n> #define MaxAllocSize\t((Size) 0x3fffffff) /* 1 gigabyte - 1 */\n\nThe same symbol is defined for frontend in psprint.c. Isn't it better\nmerge them and place it in postgres_fe.h?\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Wed, 30 Oct 2019 10:58:59 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Make StringInfo available to frontend code."},{"msg_contents":"Hi,\n\nOn 2019-10-30 10:58:59 +0900, Kyotaro Horiguchi wrote:\n> At Tue, 29 Oct 2019 17:10:01 -0700, Andres Freund <andres@anarazel.de> wrote in \n> > Hi,\n> > \n> > This patch, in a slightly rougher form, was submitted as part of [1],\n> > but it seems worth bringing up separately, rather than just committing\n> > hearing no objections.\n> ..\n> > I'm still using stringinfo in the aforementioned thread, and I also want\n> > to use it in a few more places. On the more ambitious side I really\n> > would like to have a minimal version of elog.h available in the backend,\n> > and that would really be a lot easier with stringinfo available.\n> > \n> > I also would like to submit a few patches expanding stringinfo's\n> > capabilities and performance, and it seems to me it'd be better to do so\n> > after moving (lest they introduce new FE vs BE compat issues).\n> > \n> > \n> > This allows us to remove compat.c hackery providing some stringinfo\n> > functionality for pg_waldump (which now actually needs to pass in a\n> > StringInfo...). I briefly played with converting more code in\n> > pg_waldump.c than just that one call to StringInfo, but it seems that'd\n> > be best done separately.\n> > \n> > Comments?\n> \n> It uses different form for the same message for FE and BE.\n\n> common/stringinfo.c:289-\n> > BE:\tereport(ERROR,\n> > \t\t\t(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),\n> > \t\t\t errmsg(\"out of memory\"),\n> > \t\t\t errdetail(\"Cannot enlarge string buffer containing %d\n> > bytes by %d more bytes.\",\n> > \n> > FE: +\t\t_(\"out of memory\\n\\nCannot enlarge string buffer containing %d\n> > bytes by %d more bytes.\\n\"),\n> \n> .po files will be smaller and more stable if we keep the same\n> translation unit for the same messages. That being said it doesn't\n> matter if it is tentative and the minimal elog.h for frontend comes\n> soon.\n\nI'm inclined to think that the contortions necessary to allow reusing\nthe translation strings here would be more work than worthwhile. Also,\ndo we even try to share the translations between backend and frontend?\n\n\n> > /* It's possible we could use a different value for this in frontend code */\n> > #define MaxAllocSize\t((Size) 0x3fffffff) /* 1 gigabyte - 1 */\n> \n> The same symbol is defined for frontend in psprint.c. Isn't it better\n> merge them and place it in postgres_fe.h?\n\nNo, I don't think so. I'd rather have less than more code depend on it.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 29 Oct 2019 19:06:38 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: Make StringInfo available to frontend code."},{"msg_contents":"Hello.\n\nAt Tue, 29 Oct 2019 19:06:38 -0700, Andres Freund <andres@anarazel.de> wrote in \n> Hi,\n> \n> On 2019-10-30 10:58:59 +0900, Kyotaro Horiguchi wrote:\n> > At Tue, 29 Oct 2019 17:10:01 -0700, Andres Freund <andres@anarazel.de> wrote in \n> > > Hi,\n> > > \n> > > This patch, in a slightly rougher form, was submitted as part of [1],\n> > > but it seems worth bringing up separately, rather than just committing\n> > > hearing no objections.\n> > ..\n> > > I'm still using stringinfo in the aforementioned thread, and I also want\n> > > to use it in a few more places. On the more ambitious side I really\n> > > would like to have a minimal version of elog.h available in the backend,\n> > > and that would really be a lot easier with stringinfo available.\n> > > \n> > > I also would like to submit a few patches expanding stringinfo's\n> > > capabilities and performance, and it seems to me it'd be better to do so\n> > > after moving (lest they introduce new FE vs BE compat issues).\n> > > \n> > > \n> > > This allows us to remove compat.c hackery providing some stringinfo\n> > > functionality for pg_waldump (which now actually needs to pass in a\n> > > StringInfo...). I briefly played with converting more code in\n> > > pg_waldump.c than just that one call to StringInfo, but it seems that'd\n> > > be best done separately.\n> > > \n> > > Comments?\n> > \n> > It uses different form for the same message for FE and BE.\n> \n> > common/stringinfo.c:289-\n> > > BE:\tereport(ERROR,\n> > > \t\t\t(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),\n> > > \t\t\t errmsg(\"out of memory\"),\n> > > \t\t\t errdetail(\"Cannot enlarge string buffer containing %d\n> > > bytes by %d more bytes.\",\n> > > \n> > > FE: +\t\t_(\"out of memory\\n\\nCannot enlarge string buffer containing %d\n> > > bytes by %d more bytes.\\n\"),\n> > \n> > .po files will be smaller and more stable if we keep the same\n> > translation unit for the same messages. That being said it doesn't\n> > matter if it is tentative and the minimal elog.h for frontend comes\n> > soon.\n> \n> I'm inclined to think that the contortions necessary to allow reusing\n> the translation strings here would be more work than worthwhile. Also,\n\nMaybe so for doing that in this stage. So I expect that elog.h for FE\ncomes.\n\n> the translation strings here would be more work than worthwhile. Also,\n> do we even try to share the translations between backend and frontend?\n\nNo. I don't mean that. FE and BE have their own .po files, anyway.\n\n> > > /* It's possible we could use a different value for this in frontend code */\n> > > #define MaxAllocSize\t((Size) 0x3fffffff) /* 1 gigabyte - 1 */\n> > \n> > The same symbol is defined for frontend in psprint.c. Isn't it better\n> > merge them and place it in postgres_fe.h?\n> \n> No, I don't think so. I'd rather have less than more code depend on it.\n\nOk, understoold. Thanks for the reply.\n\nregards.\n\n\n","msg_date":"Wed, 30 Oct 2019 11:18:56 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Make StringInfo available to frontend code."},{"msg_contents":"> On 30 Oct 2019, at 01:10, Andres Freund <andres@anarazel.de> wrote:\n\n> Make StringInfo available to frontend code.\n\nI’ve certainly wanted just that on multiple occasions, so +1 on this.\n\n> Therefore it seems best to just making StringInfo being usable by\n> frontend code. There's not much to do for that, except for rewriting\n> two subsequent elog/ereport calls into others types of error\n> reporting, and deciding on a maximum string length.\n\nSkimming (but not testing) the patch, it seems a reasonable approach.\n\n+ * StringInfo provides an extensible string data type. It can be used to\n\nIt might be useful to point out the upper bound on the extensibility in the\nrewrite of this sentence, and that it’s not guaranteed to be consistent between\nfrontend and backend.\n\n> I'm still using stringinfo in the aforementioned thread, and I also want\n> to use it in a few more places. On the more ambitious side I really\n> would like to have a minimal version of elog.h available in the backend,\n> and that would really be a lot easier with stringinfo available.\n\ns/backend/frontend/?\n\ncheers ./daniel\n\n","msg_date":"Fri, 1 Nov 2019 23:19:33 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Make StringInfo available to frontend code."},{"msg_contents":"Hi,\n\nOn 2019-11-01 23:19:33 +0100, Daniel Gustafsson wrote:\n> > On 30 Oct 2019, at 01:10, Andres Freund <andres@anarazel.de> wrote:\n> \n> > Make StringInfo available to frontend code.\n> \n> I’ve certainly wanted just that on multiple occasions, so +1 on this.\n\nCool.\n\n\n> + * StringInfo provides an extensible string data type. It can be used to\n> \n> It might be useful to point out the upper bound on the extensibility in the\n> rewrite of this sentence, and that it’s not guaranteed to be consistent between\n> frontend and backend.\n\nHm. Something like 'Currently the maximum length of a StringInfo is\n1GB.'? I don't really think it's worth pointing out that they may not be\nconsistent, when they currently are...\n\nAnd I suspect we should just fix the length limit to be higher for both,\nperhaps somehow allowing to limit the length for the backend cases where\nwe want to error out if a string gets too long (possibly adding a\nseparate initialization API that allows to specify the memory allocation\nflags or such).\n\n\n> > I'm still using stringinfo in the aforementioned thread, and I also want\n> > to use it in a few more places. On the more ambitious side I really\n> > would like to have a minimal version of elog.h available in the backend,\n> > and that would really be a lot easier with stringinfo available.\n> \n> s/backend/frontend/?\n\nIndeed.\n\nThanks for looking,\n\nAndres Freund\n\n\n","msg_date":"Fri, 1 Nov 2019 19:21:18 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: Make StringInfo available to frontend code."},{"msg_contents":"> On 2 Nov 2019, at 03:21, Andres Freund <andres@anarazel.de> wrote:\n> On 2019-11-01 23:19:33 +0100, Daniel Gustafsson wrote:\n\n>> + * StringInfo provides an extensible string data type. It can be used to\n>> \n>> It might be useful to point out the upper bound on the extensibility in the\n>> rewrite of this sentence, and that it’s not guaranteed to be consistent between\n>> frontend and backend.\n> \n> Hm. Something like 'Currently the maximum length of a StringInfo is\n> 1GB.’?\n\nSomething along those lines (or the define/mechanism with which the upper\nboundary controlled, simply stating the limit seems more straightforward.)\n\n> I don't really think it's worth pointing out that they may not be\n> consistent, when they currently are…\n\nGood point.\n\n> And I suspect we should just fix the length limit to be higher for both,\n> perhaps somehow allowing to limit the length for the backend cases where\n> we want to error out if a string gets too long (possibly adding a\n> separate initialization API that allows to specify the memory allocation\n> flags or such).\n\nSounds reasonable, maybe even where one can set errhint/errdetail on the “out\nof memory” error to get better reporting on failures.\n\ncheers ./daniel\n\n","msg_date":"Sat, 2 Nov 2019 23:57:06 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Make StringInfo available to frontend code."},{"msg_contents":"Hi,\n\nOn 2019-11-02 23:57:06 +0100, Daniel Gustafsson wrote:\n> > On 2 Nov 2019, at 03:21, Andres Freund <andres@anarazel.de> wrote:\n> > On 2019-11-01 23:19:33 +0100, Daniel Gustafsson wrote:\n> \n> >> + * StringInfo provides an extensible string data type. It can be used to\n> >> \n> >> It might be useful to point out the upper bound on the extensibility in the\n> >> rewrite of this sentence, and that it’s not guaranteed to be consistent between\n> >> frontend and backend.\n> > \n> > Hm. Something like 'Currently the maximum length of a StringInfo is\n> > 1GB.’?\n> \n> Something along those lines (or the define/mechanism with which the upper\n> boundary controlled, simply stating the limit seems more straightforward.)\n\nPushed now, with a variation of my suggestion above.\n\nAs nobody commented on that, I've not adjusted the location of\nstringinfo.h. If somebody has feelings about it being in the wrong\nplace, and whether to put a redirecting header in place, or whether to\njust break extensions using stringinfo, ...\n\n\n> > And I suspect we should just fix the length limit to be higher for both,\n> > perhaps somehow allowing to limit the length for the backend cases where\n> > we want to error out if a string gets too long (possibly adding a\n> > separate initialization API that allows to specify the memory allocation\n> > flags or such).\n> \n> Sounds reasonable, maybe even where one can set errhint/errdetail on the “out\n> of memory” error to get better reporting on failures.\n\nThat seems too much customization for the caller / too much added\ncomplexity, for the little gain it'd provide. Normally if a caller wants\nsomething like that, they can push something onto the error context\nstack, and get context information out that way.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 5 Nov 2019 16:43:00 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: Make StringInfo available to frontend code."}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nThis patch, in a slightly rougher form, was submitted as part of [1],\\nbut it seems worth bringing up separately, rather than just committing\\nhearing no objections.\\n\\n From the commit message:\\n\\n Make StringInfo available to frontend code.\\n\\n There's plenty places in frontend code that could benefit from a\\n string buffer implementation. Some because it yields simpler and\\n faster code, and some others because of the desire to share code\\n between backend and frontend.\\n\\n While there is a string buffer implementation available to frontend\\n code, libpq's PQExpBuffer, it is clunkier than stringinfo, it\\n introduces a libpq dependency, doesn't allow for sharing between\\n frontend and backend code, and has a higher API/ABI stability\\n requirement due to being exposed via libpq.\\n\\n Therefore it seems best to just making StringInfo being usable by\\n frontend code. There's not much to do for that, except for rewriting\\n two subsequent elog/ereport calls into others types of error\\n reporting, and deciding on a maximum string length.\\n\\n For the maximum string size I decided to privately define MaxAllocSize\\n to the same value as used in the backend. It seems likely that we'll\\n want to reconsider this for both backend and frontend code in the not\\n too far away future.\\n\\n For now I've left stringinfo.h in lib/, rather than common/, to reduce\\n the likelihood of unnecessary breakage. We could alternatively decide\\n to provide a redirecting stringinfo.h in lib/, or just not provide\\n compatibility.\\n\\nI'm still using stringinfo in the aforementioned thread, and I also want\\nto use it in a few more places. On the more ambitious side I really\\nwould like to have a minimal version of elog.h available in the backend,\\nand that would really be a lot easier with stringinfo available.\\n\\nI also would like to submit a few patches expanding stringinfo's\\ncapabilities and performance, and it seems to me it'd be better to do so\\nafter moving (lest they introduce new FE vs BE compat issues).\\n\\n\\nThis allows us to remove compat.c hackery providing some stringinfo\\nfunctionality for pg_waldump (which now actually needs to pass in a\\nStringInfo...). I briefly played with converting more code in\\npg_waldump.c than just that one call to StringInfo, but it seems that'd\\nbe best done separately.\\n\\nComments?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n[1] https://postgr.es/m/20190920051857.2fhnvhvx4qdddviz@alap3.anarazel.de\",\n \"msg_date\": \"Tue, 29 Oct 2019 17:10:01 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Make StringInfo available to frontend code.\"\n },\n {\n \"msg_contents\": \"At Tue, 29 Oct 2019 17:10:01 -0700, Andres Freund <andres@anarazel.de> wrote in \\n> Hi,\\n> \\n> This patch, in a slightly rougher form, was submitted as part of [1],\\n> but it seems worth bringing up separately, rather than just committing\\n> hearing no objections.\\n..\\n> I'm still using stringinfo in the aforementioned thread, and I also want\\n> to use it in a few more places. On the more ambitious side I really\\n> would like to have a minimal version of elog.h available in the backend,\\n> and that would really be a lot easier with stringinfo available.\\n> \\n> I also would like to submit a few patches expanding stringinfo's\\n> capabilities and performance, and it seems to me it'd be better to do so\\n> after moving (lest they introduce new FE vs BE compat issues).\\n> \\n> \\n> This allows us to remove compat.c hackery providing some stringinfo\\n> functionality for pg_waldump (which now actually needs to pass in a\\n> StringInfo...). I briefly played with converting more code in\\n> pg_waldump.c than just that one call to StringInfo, but it seems that'd\\n> be best done separately.\\n> \\n> Comments?\\n\\nIt uses different form for the same message for FE and BE.\\n\\ncommon/stringinfo.c:289-\\n> BE:\\tereport(ERROR,\\n> \\t\\t\\t(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),\\n> \\t\\t\\t errmsg(\\\"out of memory\\\"),\\n> \\t\\t\\t errdetail(\\\"Cannot enlarge string buffer containing %d\\n> bytes by %d more bytes.\\\",\\n> \\n> FE: +\\t\\t_(\\\"out of memory\\\\n\\\\nCannot enlarge string buffer containing %d\\n> bytes by %d more bytes.\\\\n\\\"),\\n\\n.po files will be smaller and more stable if we keep the same\\ntranslation unit for the same messages. That being said it doesn't\\nmatter if it is tentative and the minimal elog.h for frontend comes\\nsoon.\\n\\n\\n> /* It's possible we could use a different value for this in frontend code */\\n> #define MaxAllocSize\\t((Size) 0x3fffffff) /* 1 gigabyte - 1 */\\n\\nThe same symbol is defined for frontend in psprint.c. Isn't it better\\nmerge them and place it in postgres_fe.h?\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 10:58:59 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Make StringInfo available to frontend code.\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-10-30 10:58:59 +0900, Kyotaro Horiguchi wrote:\\n> At Tue, 29 Oct 2019 17:10:01 -0700, Andres Freund <andres@anarazel.de> wrote in \\n> > Hi,\\n> > \\n> > This patch, in a slightly rougher form, was submitted as part of [1],\\n> > but it seems worth bringing up separately, rather than just committing\\n> > hearing no objections.\\n> ..\\n> > I'm still using stringinfo in the aforementioned thread, and I also want\\n> > to use it in a few more places. On the more ambitious side I really\\n> > would like to have a minimal version of elog.h available in the backend,\\n> > and that would really be a lot easier with stringinfo available.\\n> > \\n> > I also would like to submit a few patches expanding stringinfo's\\n> > capabilities and performance, and it seems to me it'd be better to do so\\n> > after moving (lest they introduce new FE vs BE compat issues).\\n> > \\n> > \\n> > This allows us to remove compat.c hackery providing some stringinfo\\n> > functionality for pg_waldump (which now actually needs to pass in a\\n> > StringInfo...). I briefly played with converting more code in\\n> > pg_waldump.c than just that one call to StringInfo, but it seems that'd\\n> > be best done separately.\\n> > \\n> > Comments?\\n> \\n> It uses different form for the same message for FE and BE.\\n\\n> common/stringinfo.c:289-\\n> > BE:\\tereport(ERROR,\\n> > \\t\\t\\t(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),\\n> > \\t\\t\\t errmsg(\\\"out of memory\\\"),\\n> > \\t\\t\\t errdetail(\\\"Cannot enlarge string buffer containing %d\\n> > bytes by %d more bytes.\\\",\\n> > \\n> > FE: +\\t\\t_(\\\"out of memory\\\\n\\\\nCannot enlarge string buffer containing %d\\n> > bytes by %d more bytes.\\\\n\\\"),\\n> \\n> .po files will be smaller and more stable if we keep the same\\n> translation unit for the same messages. That being said it doesn't\\n> matter if it is tentative and the minimal elog.h for frontend comes\\n> soon.\\n\\nI'm inclined to think that the contortions necessary to allow reusing\\nthe translation strings here would be more work than worthwhile. Also,\\ndo we even try to share the translations between backend and frontend?\\n\\n\\n> > /* It's possible we could use a different value for this in frontend code */\\n> > #define MaxAllocSize\\t((Size) 0x3fffffff) /* 1 gigabyte - 1 */\\n> \\n> The same symbol is defined for frontend in psprint.c. Isn't it better\\n> merge them and place it in postgres_fe.h?\\n\\nNo, I don't think so. I'd rather have less than more code depend on it.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Oct 2019 19:06:38 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Make StringInfo available to frontend code.\"\n },\n {\n \"msg_contents\": \"Hello.\\n\\nAt Tue, 29 Oct 2019 19:06:38 -0700, Andres Freund <andres@anarazel.de> wrote in \\n> Hi,\\n> \\n> On 2019-10-30 10:58:59 +0900, Kyotaro Horiguchi wrote:\\n> > At Tue, 29 Oct 2019 17:10:01 -0700, Andres Freund <andres@anarazel.de> wrote in \\n> > > Hi,\\n> > > \\n> > > This patch, in a slightly rougher form, was submitted as part of [1],\\n> > > but it seems worth bringing up separately, rather than just committing\\n> > > hearing no objections.\\n> > ..\\n> > > I'm still using stringinfo in the aforementioned thread, and I also want\\n> > > to use it in a few more places. On the more ambitious side I really\\n> > > would like to have a minimal version of elog.h available in the backend,\\n> > > and that would really be a lot easier with stringinfo available.\\n> > > \\n> > > I also would like to submit a few patches expanding stringinfo's\\n> > > capabilities and performance, and it seems to me it'd be better to do so\\n> > > after moving (lest they introduce new FE vs BE compat issues).\\n> > > \\n> > > \\n> > > This allows us to remove compat.c hackery providing some stringinfo\\n> > > functionality for pg_waldump (which now actually needs to pass in a\\n> > > StringInfo...). I briefly played with converting more code in\\n> > > pg_waldump.c than just that one call to StringInfo, but it seems that'd\\n> > > be best done separately.\\n> > > \\n> > > Comments?\\n> > \\n> > It uses different form for the same message for FE and BE.\\n> \\n> > common/stringinfo.c:289-\\n> > > BE:\\tereport(ERROR,\\n> > > \\t\\t\\t(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),\\n> > > \\t\\t\\t errmsg(\\\"out of memory\\\"),\\n> > > \\t\\t\\t errdetail(\\\"Cannot enlarge string buffer containing %d\\n> > > bytes by %d more bytes.\\\",\\n> > > \\n> > > FE: +\\t\\t_(\\\"out of memory\\\\n\\\\nCannot enlarge string buffer containing %d\\n> > > bytes by %d more bytes.\\\\n\\\"),\\n> > \\n> > .po files will be smaller and more stable if we keep the same\\n> > translation unit for the same messages. That being said it doesn't\\n> > matter if it is tentative and the minimal elog.h for frontend comes\\n> > soon.\\n> \\n> I'm inclined to think that the contortions necessary to allow reusing\\n> the translation strings here would be more work than worthwhile. Also,\\n\\nMaybe so for doing that in this stage. So I expect that elog.h for FE\\ncomes.\\n\\n> the translation strings here would be more work than worthwhile. Also,\\n> do we even try to share the translations between backend and frontend?\\n\\nNo. I don't mean that. FE and BE have their own .po files, anyway.\\n\\n> > > /* It's possible we could use a different value for this in frontend code */\\n> > > #define MaxAllocSize\\t((Size) 0x3fffffff) /* 1 gigabyte - 1 */\\n> > \\n> > The same symbol is defined for frontend in psprint.c. Isn't it better\\n> > merge them and place it in postgres_fe.h?\\n> \\n> No, I don't think so. I'd rather have less than more code depend on it.\\n\\nOk, understoold. Thanks for the reply.\\n\\nregards.\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 11:18:56 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Make StringInfo available to frontend code.\"\n },\n {\n \"msg_contents\": \"> On 30 Oct 2019, at 01:10, Andres Freund <andres@anarazel.de> wrote:\\n\\n> Make StringInfo available to frontend code.\\n\\nI’ve certainly wanted just that on multiple occasions, so +1 on this.\\n\\n> Therefore it seems best to just making StringInfo being usable by\\n> frontend code. There's not much to do for that, except for rewriting\\n> two subsequent elog/ereport calls into others types of error\\n> reporting, and deciding on a maximum string length.\\n\\nSkimming (but not testing) the patch, it seems a reasonable approach.\\n\\n+ * StringInfo provides an extensible string data type. It can be used to\\n\\nIt might be useful to point out the upper bound on the extensibility in the\\nrewrite of this sentence, and that it’s not guaranteed to be consistent between\\nfrontend and backend.\\n\\n> I'm still using stringinfo in the aforementioned thread, and I also want\\n> to use it in a few more places. On the more ambitious side I really\\n> would like to have a minimal version of elog.h available in the backend,\\n> and that would really be a lot easier with stringinfo available.\\n\\ns/backend/frontend/?\\n\\ncheers ./daniel\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 23:19:33 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Make StringInfo available to frontend code.\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-11-01 23:19:33 +0100, Daniel Gustafsson wrote:\\n> > On 30 Oct 2019, at 01:10, Andres Freund <andres@anarazel.de> wrote:\\n> \\n> > Make StringInfo available to frontend code.\\n> \\n> I’ve certainly wanted just that on multiple occasions, so +1 on this.\\n\\nCool.\\n\\n\\n> + * StringInfo provides an extensible string data type. It can be used to\\n> \\n> It might be useful to point out the upper bound on the extensibility in the\\n> rewrite of this sentence, and that it’s not guaranteed to be consistent between\\n> frontend and backend.\\n\\nHm. Something like 'Currently the maximum length of a StringInfo is\\n1GB.'? I don't really think it's worth pointing out that they may not be\\nconsistent, when they currently are...\\n\\nAnd I suspect we should just fix the length limit to be higher for both,\\nperhaps somehow allowing to limit the length for the backend cases where\\nwe want to error out if a string gets too long (possibly adding a\\nseparate initialization API that allows to specify the memory allocation\\nflags or such).\\n\\n\\n> > I'm still using stringinfo in the aforementioned thread, and I also want\\n> > to use it in a few more places. On the more ambitious side I really\\n> > would like to have a minimal version of elog.h available in the backend,\\n> > and that would really be a lot easier with stringinfo available.\\n> \\n> s/backend/frontend/?\\n\\nIndeed.\\n\\nThanks for looking,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 19:21:18 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Make StringInfo available to frontend code.\"\n },\n {\n \"msg_contents\": \"> On 2 Nov 2019, at 03:21, Andres Freund <andres@anarazel.de> wrote:\\n> On 2019-11-01 23:19:33 +0100, Daniel Gustafsson wrote:\\n\\n>> + * StringInfo provides an extensible string data type. It can be used to\\n>> \\n>> It might be useful to point out the upper bound on the extensibility in the\\n>> rewrite of this sentence, and that it’s not guaranteed to be consistent between\\n>> frontend and backend.\\n> \\n> Hm. Something like 'Currently the maximum length of a StringInfo is\\n> 1GB.’?\\n\\nSomething along those lines (or the define/mechanism with which the upper\\nboundary controlled, simply stating the limit seems more straightforward.)\\n\\n> I don't really think it's worth pointing out that they may not be\\n> consistent, when they currently are…\\n\\nGood point.\\n\\n> And I suspect we should just fix the length limit to be higher for both,\\n> perhaps somehow allowing to limit the length for the backend cases where\\n> we want to error out if a string gets too long (possibly adding a\\n> separate initialization API that allows to specify the memory allocation\\n> flags or such).\\n\\nSounds reasonable, maybe even where one can set errhint/errdetail on the “out\\nof memory” error to get better reporting on failures.\\n\\ncheers ./daniel\\n\\n\",\n \"msg_date\": \"Sat, 2 Nov 2019 23:57:06 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Make StringInfo available to frontend code.\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-11-02 23:57:06 +0100, Daniel Gustafsson wrote:\\n> > On 2 Nov 2019, at 03:21, Andres Freund <andres@anarazel.de> wrote:\\n> > On 2019-11-01 23:19:33 +0100, Daniel Gustafsson wrote:\\n> \\n> >> + * StringInfo provides an extensible string data type. It can be used to\\n> >> \\n> >> It might be useful to point out the upper bound on the extensibility in the\\n> >> rewrite of this sentence, and that it’s not guaranteed to be consistent between\\n> >> frontend and backend.\\n> > \\n> > Hm. Something like 'Currently the maximum length of a StringInfo is\\n> > 1GB.’?\\n> \\n> Something along those lines (or the define/mechanism with which the upper\\n> boundary controlled, simply stating the limit seems more straightforward.)\\n\\nPushed now, with a variation of my suggestion above.\\n\\nAs nobody commented on that, I've not adjusted the location of\\nstringinfo.h. If somebody has feelings about it being in the wrong\\nplace, and whether to put a redirecting header in place, or whether to\\njust break extensions using stringinfo, ...\\n\\n\\n> > And I suspect we should just fix the length limit to be higher for both,\\n> > perhaps somehow allowing to limit the length for the backend cases where\\n> > we want to error out if a string gets too long (possibly adding a\\n> > separate initialization API that allows to specify the memory allocation\\n> > flags or such).\\n> \\n> Sounds reasonable, maybe even where one can set errhint/errdetail on the “out\\n> of memory” error to get better reporting on failures.\\n\\nThat seems too much customization for the caller / too much added\\ncomplexity, for the little gain it'd provide. Normally if a caller wants\\nsomething like that, they can push something onto the error context\\nstack, and get context information out that way.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Nov 2019 16:43:00 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Make StringInfo available to frontend code.\"\n }\n]","isConversational":false}}},{"rowIdx":49364,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello,\n\n\nI propose new simple sql query, which shows total block numbers in the \nrelation.\n\nI now reviewing this patch (https://commitfest.postgresql.org/25/2211/) \nand I think,\nit is usefull for knowing how many blocks there are in the relation to \ndetermine whether we use VACUUM RESUME or not.\n\nOf cource, we can know this value such as\n\nselect (pg_relation_size('t') / \ncurrent_setting('block_size')::bigint)::int;\n\n\nbut I think it is a litte bit complex.\n\n\n\nComment and feedback are very welcome.\n\nRegards ,\n\n\nYu Kimura","msg_date":"Wed, 30 Oct 2019 16:48:00 +0900","msg_from":"btkimurayuzk <btkimurayuzk@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Add SQL function to show total block numbers in the relation"},{"msg_contents":"btkimurayuzk <btkimurayuzk@oss.nttdata.com> writes:\n> I propose new simple sql query, which shows total block numbers in the \n> relation.\n> ...\n> Of cource, we can know this value such as\n> select (pg_relation_size('t') / \n> current_setting('block_size')::bigint)::int;\n\nI don't really see why the existing solution isn't sufficient.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 30 Oct 2019 10:09:47 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Add SQL function to show total block numbers in the relation"},{"msg_contents":"On Wed, Oct 30, 2019 at 10:09:47AM -0400, Tom Lane wrote:\n> btkimurayuzk <btkimurayuzk@oss.nttdata.com> writes:\n>> I propose new simple sql query, which shows total block numbers in the \n>> relation.\n>> ...\n>> Of cource, we can know this value such as\n>> select (pg_relation_size('t') / \n>> current_setting('block_size')::bigint)::int;\n> \n> I don't really see why the existing solution isn't sufficient.\n\n+1.\n--\nMichael","msg_date":"Thu, 31 Oct 2019 12:29:45 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add SQL function to show total block numbers in the relation"},{"msg_contents":"> btkimurayuzk <btkimurayuzk@oss.nttdata.com> writes:\n>> I propose new simple sql query, which shows total block numbers in the\n>> relation.\n>> ...\n>> Of cource, we can know this value such as\n>> select (pg_relation_size('t') /\n>> current_setting('block_size')::bigint)::int;\n> \n> I don't really see why the existing solution isn't sufficient.\n\nI think it's a little difficult to introduce the block size using two \nvalues `current block size` and `reference size`\nfor beginners who are not familiar with the internal structure of \nPostgres,\n\nThis is the reason why the existing solution was insufficient.\n\nWhat do you think?\n\nRegards,\nYu Kimura\n\n\n","msg_date":"Thu, 07 Nov 2019 17:04:51 +0900","msg_from":"btkimurayuzk <btkimurayuzk@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Re: Add SQL function to show total block numbers in the relation"},{"msg_contents":"Hello, Kimura-san.\n\nAt Thu, 07 Nov 2019 17:04:51 +0900, btkimurayuzk <btkimurayuzk@oss.nttdata.com> wrote in \n> > btkimurayuzk <btkimurayuzk@oss.nttdata.com> writes:\n> >> I propose new simple sql query, which shows total block numbers in the\n> >> relation.\n> >> ...\n> >> Of cource, we can know this value such as\n> >> select (pg_relation_size('t') /\n> >> current_setting('block_size')::bigint)::int;\n> > I don't really see why the existing solution isn't sufficient.\n> \n> I think it's a little difficult to introduce the block size using two\n> values `current block size` and `reference size`\n> for beginners who are not familiar with the internal structure of\n> Postgres,\n> \n> This is the reason why the existing solution was insufficient.\n> \n> What do you think?\n\nSorry, but I also vote -1 for the new function.\n\nSize in block number is useless for those who doesn't understand the\nnotion of block, or block size. Those who understands the notion\nshould come up with the simple formula (except the annoying\ncasts). Anyone can find the clue to the base values by searching the\ndocument in the Web with the keywords \"block size\" and \"relation size\"\nor even with \"table size\". (FWIW, I would even do the same for the new\nfunction if any...) If they need it so frequently, a user-defined\nfunction is easily made up.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Thu, 07 Nov 2019 18:01:34 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add SQL function to show total block numbers in the relation"},{"msg_contents":"On Thu, Nov 07, 2019 at 06:01:34PM +0900, Kyotaro Horiguchi wrote:\n> Sorry, but I also vote -1 for the new function.\n\nSo do I. If there are no objections, I will mark the patch as\nrejected in the CF app.\n\n> If they need it so frequently, a user-defined function is easily\n> made up.\n\nYep.\n--\nMichael","msg_date":"Fri, 8 Nov 2019 09:30:56 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add SQL function to show total block numbers in the relation"},{"msg_contents":"On Fri, Nov 08, 2019 at 09:30:56AM +0900, Michael Paquier wrote:\n> On Thu, Nov 07, 2019 at 06:01:34PM +0900, Kyotaro Horiguchi wrote:\n>> Sorry, but I also vote -1 for the new function.\n> \n> So do I. If there are no objections, I will mark the patch as\n> rejected in the CF app.\n\nAnd done.\n--\nMichael","msg_date":"Tue, 12 Nov 2019 13:54:32 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add SQL function to show total block numbers in the relation"},{"msg_contents":"> Size in block number is useless for those who doesn't understand the\n> notion of block, or block size. Those who understands the notion\n> should come up with the simple formula (except the annoying\n> casts). Anyone can find the clue to the base values by searching the\n> document in the Web with the keywords \"block size\" and \"relation size\"\n> or even with \"table size\". (FWIW, I would even do the same for the new\n> function if any...) If they need it so frequently, a user-defined\n> function is easily made up.\n> \n> regards.\n\n\nI didn't know about the existence of the user-defined function .\nI fully understood , Thanks .\n\nRegards,\n\nYu Kimura\n\n\n\n","msg_date":"Wed, 13 Nov 2019 10:55:36 +0900","msg_from":"btkimurayuzk <btkimurayuzk@oss.nttdata.com>","msg_from_op":true,"msg_subject":"Re: Add SQL function to show total block numbers in the relation"}],"string":"[\n {\n \"msg_contents\": \"Hello,\\n\\n\\nI propose new simple sql query, which shows total block numbers in the \\nrelation.\\n\\nI now reviewing this patch (https://commitfest.postgresql.org/25/2211/) \\nand I think,\\nit is usefull for knowing how many blocks there are in the relation to \\ndetermine whether we use VACUUM RESUME or not.\\n\\nOf cource, we can know this value such as\\n\\nselect (pg_relation_size('t') / \\ncurrent_setting('block_size')::bigint)::int;\\n\\n\\nbut I think it is a litte bit complex.\\n\\n\\n\\nComment and feedback are very welcome.\\n\\nRegards ,\\n\\n\\nYu Kimura\",\n \"msg_date\": \"Wed, 30 Oct 2019 16:48:00 +0900\",\n \"msg_from\": \"btkimurayuzk <btkimurayuzk@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Add SQL function to show total block numbers in the relation\"\n },\n {\n \"msg_contents\": \"btkimurayuzk <btkimurayuzk@oss.nttdata.com> writes:\\n> I propose new simple sql query, which shows total block numbers in the \\n> relation.\\n> ...\\n> Of cource, we can know this value such as\\n> select (pg_relation_size('t') / \\n> current_setting('block_size')::bigint)::int;\\n\\nI don't really see why the existing solution isn't sufficient.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 10:09:47 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add SQL function to show total block numbers in the relation\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 30, 2019 at 10:09:47AM -0400, Tom Lane wrote:\\n> btkimurayuzk <btkimurayuzk@oss.nttdata.com> writes:\\n>> I propose new simple sql query, which shows total block numbers in the \\n>> relation.\\n>> ...\\n>> Of cource, we can know this value such as\\n>> select (pg_relation_size('t') / \\n>> current_setting('block_size')::bigint)::int;\\n> \\n> I don't really see why the existing solution isn't sufficient.\\n\\n+1.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 31 Oct 2019 12:29:45 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add SQL function to show total block numbers in the relation\"\n },\n {\n \"msg_contents\": \"> btkimurayuzk <btkimurayuzk@oss.nttdata.com> writes:\\n>> I propose new simple sql query, which shows total block numbers in the\\n>> relation.\\n>> ...\\n>> Of cource, we can know this value such as\\n>> select (pg_relation_size('t') /\\n>> current_setting('block_size')::bigint)::int;\\n> \\n> I don't really see why the existing solution isn't sufficient.\\n\\nI think it's a little difficult to introduce the block size using two \\nvalues `current block size` and `reference size`\\nfor beginners who are not familiar with the internal structure of \\nPostgres,\\n\\nThis is the reason why the existing solution was insufficient.\\n\\nWhat do you think?\\n\\nRegards,\\nYu Kimura\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Nov 2019 17:04:51 +0900\",\n \"msg_from\": \"btkimurayuzk <btkimurayuzk@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add SQL function to show total block numbers in the relation\"\n },\n {\n \"msg_contents\": \"Hello, Kimura-san.\\n\\nAt Thu, 07 Nov 2019 17:04:51 +0900, btkimurayuzk <btkimurayuzk@oss.nttdata.com> wrote in \\n> > btkimurayuzk <btkimurayuzk@oss.nttdata.com> writes:\\n> >> I propose new simple sql query, which shows total block numbers in the\\n> >> relation.\\n> >> ...\\n> >> Of cource, we can know this value such as\\n> >> select (pg_relation_size('t') /\\n> >> current_setting('block_size')::bigint)::int;\\n> > I don't really see why the existing solution isn't sufficient.\\n> \\n> I think it's a little difficult to introduce the block size using two\\n> values `current block size` and `reference size`\\n> for beginners who are not familiar with the internal structure of\\n> Postgres,\\n> \\n> This is the reason why the existing solution was insufficient.\\n> \\n> What do you think?\\n\\nSorry, but I also vote -1 for the new function.\\n\\nSize in block number is useless for those who doesn't understand the\\nnotion of block, or block size. Those who understands the notion\\nshould come up with the simple formula (except the annoying\\ncasts). Anyone can find the clue to the base values by searching the\\ndocument in the Web with the keywords \\\"block size\\\" and \\\"relation size\\\"\\nor even with \\\"table size\\\". (FWIW, I would even do the same for the new\\nfunction if any...) If they need it so frequently, a user-defined\\nfunction is easily made up.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Nov 2019 18:01:34 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add SQL function to show total block numbers in the relation\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 07, 2019 at 06:01:34PM +0900, Kyotaro Horiguchi wrote:\\n> Sorry, but I also vote -1 for the new function.\\n\\nSo do I. If there are no objections, I will mark the patch as\\nrejected in the CF app.\\n\\n> If they need it so frequently, a user-defined function is easily\\n> made up.\\n\\nYep.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 8 Nov 2019 09:30:56 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add SQL function to show total block numbers in the relation\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 08, 2019 at 09:30:56AM +0900, Michael Paquier wrote:\\n> On Thu, Nov 07, 2019 at 06:01:34PM +0900, Kyotaro Horiguchi wrote:\\n>> Sorry, but I also vote -1 for the new function.\\n> \\n> So do I. If there are no objections, I will mark the patch as\\n> rejected in the CF app.\\n\\nAnd done.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 12 Nov 2019 13:54:32 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add SQL function to show total block numbers in the relation\"\n },\n {\n \"msg_contents\": \"> Size in block number is useless for those who doesn't understand the\\n> notion of block, or block size. Those who understands the notion\\n> should come up with the simple formula (except the annoying\\n> casts). Anyone can find the clue to the base values by searching the\\n> document in the Web with the keywords \\\"block size\\\" and \\\"relation size\\\"\\n> or even with \\\"table size\\\". (FWIW, I would even do the same for the new\\n> function if any...) If they need it so frequently, a user-defined\\n> function is easily made up.\\n> \\n> regards.\\n\\n\\nI didn't know about the existence of the user-defined function .\\nI fully understood , Thanks .\\n\\nRegards,\\n\\nYu Kimura\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Nov 2019 10:55:36 +0900\",\n \"msg_from\": \"btkimurayuzk <btkimurayuzk@oss.nttdata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add SQL function to show total block numbers in the relation\"\n }\n]","isConversational":false}}},{"rowIdx":49365,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"A global index by very definition is a single index on the parent table\nthat maps to many\nunderlying table partitions. The parent table itself does not have any\nunderlying storage,\nso it must, therefore, retrieve the data satisfying index constraints from\nthe underlying tables.\nIn very crude terms, it is an accumulation of data from table partitions so\nthat data spanning\nacross multiple partitions are accessed in one go as opposed to\nindividually querying each\npartition.\n\nFor the initial version of this work, we are only considering to build\nb-tree global indexes.\n\n- Partitioned Index (Index Partitioning)\nWhen global indexes become too large, then those are partitioned to keep\nthe performance\nand maintenance overhead manageable. These are not within the scope of this\nwork.\n\n\n- Local Index\nA local index is an index that is local to a specific table partition; i.e.\nit doesn’t span across\nmultiple partitions. So, when we create an index on a parent table, it will\ncreate a separate\nindex for all its partitions. Unfortunately, PostgreSQL uses the\nterminology of “partitioned index”\nwhen it refers to local indexes. This work with fix this terminology for\nPostgreSQL so that the\nnomenclature remains consistent with other DBMS.\n\n\n- Why We Need Global Index?\nA global index is expected to give two very important upgrades to the\npartitioning feature set in\nPostgreSQL. It is expected to give a significant improvement in\nread-performance for queries\ntargeting multiple local indexes of partitions. It also adds a unique\nconstraint across partitions.\n\n\n- Unique Constraint\nData uniqueness is a critical requirement for building an index. For global\nindexes that span across\nmultiple partitions, uniqueness will have to be enforced on index\ncolumn(s). This effectively translates\ninto a unique constraint.\n\n\n- Performance\nCurrently, the pseudo index created on the parent table of partitions does\nnot contain any\ndata. Rather, it dereferences to the local indexes when an index search is\nrequired. This\nmeans that multiple indexes will have to be evaluated and data to be\ncombined thereafter.\nHowever, with the global indexes, data will reside with global index\ndeclared on the parent\ntable. This avoids the need for multi-level index lookups. So read\nperformance is expected\nto be significantly higher in cases. There will however be a negative\nperformance impact\nduring write (insert/update) of data. This is discussed in more detail\nlater on.\n\n\n- Creating a GLOBAL Index - Syntax\nA global index may be created with the addition of a “GLOBAL” keyword to\nthe index statement.\nAlternatively, one could specify the “LOCAL” keyword to create local\nindexes on partitions.\nWe are suggesting to call this set of keywords: “partition_index_type”. By\ndefault,\npartition_index_type will be set as LOCAL. Here is a sample of the create\nindex syntax.\n\nCREATE Index idx parent (columns) [GLOBAL | LOCAL];\n\nNote: There is no shift/reduced by adding these options.\n\n\n- Pointing Index to Tuple\nCurrently, CTID carries a page and offset information for a known heap\n(table name). However,\nin the context of global indexes, this information within an index is\ninsufficient. Since the index is\nexpected to carry tuples from multiple partitions (heaps), CTID alone will\nnot be able to link an index\nnode to a tuple. This requires carrying additional data for the heap name\nto be stored with each\nindex node.\n\nHow this should be implemented is a point to be discussed. A few\npossibilities are listed below:\n\n-- Expand CTID to include a relfilenode id. In PostgreSQL-Conf Asia, Bruce\nsuggested having the OID\ninstead of relfilenode as relfilenode can be duplicated across tablespaces.\n -- Using OID introduces another complication where we would need to\nquery catalog for OID to\n heap mapping.\n\n-- The second option is to have a variable-length CTID. We can reserve some\ntop-level bit for segregation of\nGlobal CTID or Standard CTID. Robert Haas suggested in PostgreSQL-EU to\ndiscuss this with Peter Geoghegan.\n -- I discussed it with Peter and he believes that it is a very\ninvasive approach that requires a whole lot of\n the effort to get committed.\n\n-- Heikki pointed me to include heap specific information using the INCLUDE\nkeyword so that heap information\nis stored with each index node as data.\n -- We (Peter and I) also discussed that option and this looks a more\neasy and non-invasive option.\n\n\n- Optimizer\nThe challenge with optimizer is a selection between local and global\nindexes when both are present.\nHow do we:\n\n-- Evaluate the cost of scanning a global index?\n-- When should the LOCAL index be preferred over the GLOBAL index and vice\nversa?\n -- Should we hit a GLOBAL index when the query is targeting a\ncouple of partitions only?\n -- We need to consider the sizes of those partitions being hit and\nthe sizes of partitions not being hit.\n-- Bruce suggested that we prioritize a GLOBAL index in the first version\nso that in every case,\nthe GLOBAL index is utilized.\n\n\n- Write Performance and Vacuum\nThere will be some write performance degradation because every change in\npartition tables must\npropagate upwards to the GLOBAL index on the parent table. This can be\nthought of as another index\non a table, however, the [slight] performance degradation will be due to\nthe fact that the GLOBAL\nindex may carry a much bigger dataset with data from multiple partitions\nresulting in a higher tree\ntraversal and update time. This applies to both write and vacuum processes.\n\nIt is still an open question though on how this will be handled within the\ncode and how we can better\noptimize this process.\n\nI have a POC patch and working on finalizing the patch, Hamid Akhtar is\nalso working with me on this\nwork.\n\n\n\n--\nIbrar Ahmed\n\nA global index by very definition is a single index on the parent table that maps to manyunderlying table partitions. The parent table itself does not have any underlying storage,so it must, therefore, retrieve the data satisfying index constraints from the underlying tables.In very crude terms, it is an accumulation of data from table partitions so that data spanningacross multiple partitions are accessed in one go as opposed to individually querying eachpartition.For the initial version of this work, we are only considering to build b-tree global indexes.- Partitioned Index (Index Partitioning)When global indexes become too large, then those are partitioned to keep the performanceand maintenance overhead manageable. These are not within the scope of this work.- Local IndexA local index is an index that is local to a specific table partition; i.e. it doesn’t span acrossmultiple partitions. So, when we create an index on a parent table, it will create a separateindex for all its partitions. Unfortunately, PostgreSQL uses the terminology of “partitioned index”when it refers to local indexes. This work with fix this terminology for PostgreSQL so that thenomenclature remains consistent with other DBMS. - Why We Need Global Index?A global index is expected to give two very important upgrades to the partitioning feature set inPostgreSQL. It is expected to give a significant improvement in read-performance for queriestargeting multiple local indexes of partitions. It also adds a unique constraint across partitions.- Unique ConstraintData uniqueness is a critical requirement for building an index. For global indexes that span acrossmultiple partitions, uniqueness will have to be enforced on index column(s). This effectively translatesinto a unique constraint.- PerformanceCurrently, the pseudo index created on the parent table of partitions does not contain anydata. Rather, it dereferences to the local indexes when an index search is required. Thismeans that multiple indexes will have to be evaluated and data to be combined thereafter.However, with the global indexes, data will reside with global index declared on the parenttable. This avoids the need for multi-level index lookups. So read performance is expectedto be significantly higher in cases. There will however be a negative performance impactduring write (insert/update) of data. This is discussed in more detail later on.- Creating a GLOBAL Index - SyntaxA global index may be created with the addition of a “GLOBAL” keyword to the index statement.Alternatively, one could specify the “LOCAL” keyword to create local indexes on partitions.We are suggesting to call this set of keywords: “partition_index_type”. By default,partition_index_type will be set as LOCAL. Here is a sample of the create index syntax.CREATE Index idx parent (columns) [GLOBAL | LOCAL];Note: There is no shift/reduced by adding these options.- Pointing Index to TupleCurrently, CTID carries a page and offset information for a known heap (table name). However,in the context of global indexes, this information within an index is insufficient. Since the index isexpected to carry tuples from multiple partitions (heaps), CTID alone will not be able to link an indexnode to a tuple. This requires carrying additional data for the heap name to be stored with eachindex node.How this should be implemented is a point to be discussed. A few possibilities are listed below:-- Expand CTID to include a relfilenode id. In PostgreSQL-Conf Asia, Bruce suggested having the OIDinstead of relfilenode as relfilenode can be duplicated across tablespaces. -- Using OID introduces another complication where we would need to query catalog for OID to heap mapping.-- The second option is to have a variable-length CTID. We can reserve some top-level bit for segregation ofGlobal CTID or Standard CTID. Robert Haas suggested in PostgreSQL-EU to discuss this with Peter Geoghegan. -- I discussed it with Peter and he believes that it is a very invasive approach that requires a whole lot of the effort to get committed. -- Heikki pointed me to include heap specific information using the INCLUDE keyword so that heap informationis stored with each index node as data. -- We (Peter and I) also discussed that option and this looks a more easy and non-invasive option.- OptimizerThe challenge with optimizer is a selection between local and global indexes when both are present.How do we: -- Evaluate the cost of scanning a global index?-- When should the LOCAL index be preferred over the GLOBAL index and vice versa? -- Should we hit a GLOBAL index when the query is targeting a couple of partitions only? -- We need to consider the sizes of those partitions being hit and the sizes of partitions not being hit.-- Bruce suggested that we prioritize a GLOBAL index in the first version so that in every case,the GLOBAL index is utilized. - Write Performance and VacuumThere will be some write performance degradation because every change in partition tables mustpropagate upwards to the GLOBAL index on the parent table. This can be thought of as another indexon a table, however, the [slight] performance degradation will be due to the fact that the GLOBALindex may carry a much bigger dataset with data from multiple partitions resulting in a higher treetraversal and update time. This applies to both write and vacuum processes.It is still an open question though on how this will be handled within the code and how we can betteroptimize this process.I have a POC patch and working on finalizing the patch, Hamid Akhtar is also working with me on thiswork.--Ibrar Ahmed","msg_date":"Wed, 30 Oct 2019 13:38:29 +0500","msg_from":"Ibrar Ahmed <ibrar.ahmad@gmail.com>","msg_from_op":true,"msg_subject":"Proposal: Global Index"},{"msg_contents":"Ibrar Ahmed <ibrar.ahmad@gmail.com> writes:\n> A global index by very definition is a single index on the parent table\n> that maps to many\n> underlying table partitions.\n\nI believe that the current design of partitioning is explicitly intended\nto avoid the need for such a construct. It'd be absolutely disastrous\nto have such a thing from many standpoints, including the breadth of\nlocking needed to work with the global index, the difficulty of vacuuming,\nand the impossibility of cheaply attaching or detaching partitions.\n\nIn other words, this is a \"feature\" we do not want.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 30 Oct 2019 10:13:26 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Wed, Oct 30, 2019 at 10:13 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> I believe that the current design of partitioning is explicitly intended\n> to avoid the need for such a construct. It'd be absolutely disastrous\n> to have such a thing from many standpoints, including the breadth of\n> locking needed to work with the global index, the difficulty of vacuuming,\n> and the impossibility of cheaply attaching or detaching partitions.\n>\n> In other words, this is a \"feature\" we do not want.\n\nI don't think that's true. Certainly, a lot of EnterpriseDB customers\nwant this feature - it comes up regularly in discussions here. But\nthat is not to say that the technical challenges are not formidable,\nand I don't think this proposal really grapples with any of them.\nHowever, that doesn't mean that the feature isn't desirable.\n\nOne of the biggest reasons why people want it is to enforce uniqueness\nfor secondary keys - e.g. the employees table is partitioned by\nemployee ID, but SSN should also be unique, at least among employees\nfor whom it's not NULL.\n\nBut people also want it for faster data retrieval: if you're looking\nfor a commonly-occurring value, an index per partition is fine. But if\nyou're looking for values that occur only once or a few times across\nthe whole hierarchy, an index scan per partition is very costly.\nConsider, e.g.:\n\nNested Loop\n-> Seq Scan\n-> Append\n -> Index Scan on each_partition\n\nYou don't have to have very many partitions for that to suck, and it's\na thing that people want to do. Runtime partition pruning helps with\nthis case a lot, but, once again, only for the primary key. Secondary\nkeys are a big problem for partitioning today, in many ways.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 30 Oct 2019 12:12:57 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Wed, Oct 30, 2019 at 10:13 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> I believe that the current design of partitioning is explicitly intended\n>> to avoid the need for such a construct. It'd be absolutely disastrous\n>> to have such a thing from many standpoints, including the breadth of\n>> locking needed to work with the global index, the difficulty of vacuuming,\n>> and the impossibility of cheaply attaching or detaching partitions.\n>> In other words, this is a \"feature\" we do not want.\n\n> I don't think that's true. Certainly, a lot of EnterpriseDB customers\n> want this feature - it comes up regularly in discussions here. But\n> that is not to say that the technical challenges are not formidable,\n> and I don't think this proposal really grapples with any of them.\n> However, that doesn't mean that the feature isn't desirable.\n\nWell, the *effects* of the feature seem desirable, but that doesn't\nmean that we want an implementation that actually has a shared index.\nAs soon as you do that, you've thrown away most of the benefits of\nhaving a partitioned data structure in the first place.\n\nNo, I don't have an idea how we might support, eg, uniqueness of\nnon-partition-key columns without that. But we need to spend our\neffort on figuring that out, not on building a complicated mechanism\nwhose performance is never going to not suck.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 30 Oct 2019 12:23:23 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Wed, Oct 30, 2019 at 9:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Well, the *effects* of the feature seem desirable, but that doesn't\n> mean that we want an implementation that actually has a shared index.\n> As soon as you do that, you've thrown away most of the benefits of\n> having a partitioned data structure in the first place.\n\nRight, but that's only the case for the global index. Global indexes\nare useful when used judiciously. They enable the use of partitioning\nfor use cases where not being able to enforce uniqueness across all\npartitions happens to be a deal breaker. I bet that this is fairly\ncommon.\n\n> No, I don't have an idea how we might support, eg, uniqueness of\n> non-partition-key columns without that. But we need to spend our\n> effort on figuring that out, not on building a complicated mechanism\n> whose performance is never going to not suck.\n\nI don't think that there is a way to solve the problem that doesn't\nlook very much like a global index. Also, being able to push down a\npartition number when scanning a global index seems like it would be\nvery compelling in some scenarios.\n\nI'm a bit worried about the complexity that will need to be added to\nnbtree to make global indexes work, but it's probably possible to come\nup with something that isn't too bad. GIN already uses an\nimplementation level attribute number column for multi-column GIN\nindexes, which is a little like what Ibrar has in mind. The really\ncomplicated new code required for global indexes will be in places\nlike vacuumlazy.c.\n\n\n--\nPeter Geoghegan\n\n\n","msg_date":"Wed, 30 Oct 2019 09:48:37 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"Peter Geoghegan <pg@bowt.ie> writes:\n> On Wed, Oct 30, 2019 at 9:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> Well, the *effects* of the feature seem desirable, but that doesn't\n>> mean that we want an implementation that actually has a shared index.\n>> As soon as you do that, you've thrown away most of the benefits of\n>> having a partitioned data structure in the first place.\n\n> Right, but that's only the case for the global index. Global indexes\n> are useful when used judiciously.\n\nBut ... why bother with partitioning then? To me, the main reasons\nwhy you might want a partitioned table are\n\n* ability to cheaply add and remove partitions, primarily so that\nyou can cheaply do things like \"delete the oldest month's data\".\n\n* ability to scale past our limits on the physical size of one table\n--- both the hard BlockNumber-based limit, and the performance\nconstraints of e.g. vacuuming a very large table.\n\nBoth of those go out the window with a global index. So you might\nas well just have one table and forget all the overhead.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 30 Oct 2019 13:05:57 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"Hi,\n\nOn 2019-10-30 13:05:57 -0400, Tom Lane wrote:\n> Peter Geoghegan <pg@bowt.ie> writes:\n> > On Wed, Oct 30, 2019 at 9:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >> Well, the *effects* of the feature seem desirable, but that doesn't\n> >> mean that we want an implementation that actually has a shared index.\n> >> As soon as you do that, you've thrown away most of the benefits of\n> >> having a partitioned data structure in the first place.\n> \n> > Right, but that's only the case for the global index. Global indexes\n> > are useful when used judiciously.\n> \n> But ... why bother with partitioning then? To me, the main reasons\n> why you might want a partitioned table are\n\nQuite commonly there's a lot of *other* indexes, often on a lot wider\ndata than the primary key, that don't need to be global. And whereas in\na lot of cases the primary key in a partitioned table has pretty good\nlocality (and thus will be mostly buffered IO), other indexes will often\nnot have that property (i.e. not have much correlation with table\nposition).\n\n\n> * ability to cheaply add and remove partitions, primarily so that\n> you can cheaply do things like \"delete the oldest month's data\".\n\nYou can still do that to some degree with a global index. Imagine\ne.g. keeping a 'partition id' as a sort-of column in the global\nindex. That allows you to drop the partition, without having to\nimmediately rebuild the index, by checking the partition id against the\nlive partitions during lookup. So sure, your'e wasting space for a bit\nin the global index, but it'll also be space that is likely to be fairly\nefficiently reclaimed the next time vacuum touches the index. And if\nnot the global index can be rebuilt concurrently without blocking\nwrites.\n\n\n> * ability to scale past our limits on the physical size of one table\n> --- both the hard BlockNumber-based limit, and the performance\n> constraints of e.g. vacuuming a very large table.\n\nFor that to be a problem for a global index the global index (which will\noften be something like two int4 or int8 columns) itself would need to\nbe above the block number based limit - which doesn't seem that close.\n\nWRT vacuuming - based on my observations the table itself isn't a\nperformance problem for vacuuming all that commonly anymore, it's the\nassociated index scans. So yea, that's a problem.\n\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 30 Oct 2019 10:27:02 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"Greetings,\n\n* Peter Geoghegan (pg@bowt.ie) wrote:\n> On Wed, Oct 30, 2019 at 9:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > Well, the *effects* of the feature seem desirable, but that doesn't\n> > mean that we want an implementation that actually has a shared index.\n> > As soon as you do that, you've thrown away most of the benefits of\n> > having a partitioned data structure in the first place.\n> \n> Right, but that's only the case for the global index. Global indexes\n> are useful when used judiciously. They enable the use of partitioning\n> for use cases where not being able to enforce uniqueness across all\n> partitions happens to be a deal breaker. I bet that this is fairly\n> common.\n\nAbsolutely- our lack of such is a common point of issue when folks are\nconsidering using or migrating to PostgreSQL.\n\nThanks,\n\nStephen","msg_date":"Thu, 31 Oct 2019 14:50:35 -0400","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Thu, 31 Oct 2019 at 14:50, Stephen Frost <sfrost@snowman.net> wrote:\n\n> Greetings,\n>\n> * Peter Geoghegan (pg@bowt.ie) wrote:\n>\n[....]\n\n>\n> Absolutely- our lack of such is a common point of issue when folks are\n> considering using or migrating to PostgreSQL.\n>\n\nNot sure how similar my situation really is, but I find myself wanting to\nhave indices that cross non-partition members of an inheritance hierarchy:\n\ncreate table t (\n id int,\n primary key (id)\n);\n\ncreate table t1 (\n a text\n) inherits (t);\n\ncreate table t2 (\n b int,\n c int\n) inherits (t);\n\nSo \"t\"s are identified by an integer; and one kind of \"t\" has a single text\nattribute while a different kind of \"t\" has 2 int attributes. The idea is\nthat there is a single primary key constraint on the whole hierarchy that\nensures only one record with a particular id can exist in all the tables\ntogether. I can imagine wanting to do this with other unique constraints\nalso.\n\nAt present I don't actually use inheritance; instead I put triggers on the\nchild tables that do an insert on the parent table, which has the effect of\nenforcing the uniqueness I want.\n\nOn Thu, 31 Oct 2019 at 14:50, Stephen Frost <sfrost@snowman.net> wrote:Greetings,\n\n* Peter Geoghegan (pg@bowt.ie) wrote:\n[....] \n\nAbsolutely- our lack of such is a common point of issue when folks are\nconsidering using or migrating to PostgreSQL.Not sure how similar my situation really is, but I find myself wanting to have indices that cross non-partition members of an inheritance hierarchy:create table t ( id int, primary key (id));create table t1 ( a text) inherits (t);create table t2 ( b int, c int) inherits (t);So \"t\"s are identified by an integer; and one kind of \"t\" has a single text attribute while a different kind of \"t\" has 2 int attributes. The idea is that there is a single primary key constraint on the whole hierarchy that ensures only one record with a particular id can exist in all the tables together. I can imagine wanting to do this with other unique constraints also.At present I don't actually use inheritance; instead I put triggers on the child tables that do an insert on the parent table, which has the effect of enforcing the uniqueness I want.","msg_date":"Thu, 31 Oct 2019 15:02:40 -0400","msg_from":"Isaac Morland <isaac.morland@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Thu, Oct 31, 2019 at 03:02:40PM -0400, Isaac Morland wrote:\n>On Thu, 31 Oct 2019 at 14:50, Stephen Frost <sfrost@snowman.net> wrote:\n>\n>> Greetings,\n>>\n>> * Peter Geoghegan (pg@bowt.ie) wrote:\n>>\n>[....]\n>\n>>\n>> Absolutely- our lack of such is a common point of issue when folks are\n>> considering using or migrating to PostgreSQL.\n>>\n>\n>Not sure how similar my situation really is, but I find myself wanting to\n>have indices that cross non-partition members of an inheritance hierarchy:\n>\n>create table t (\n> id int,\n> primary key (id)\n>);\n>\n>create table t1 (\n> a text\n>) inherits (t);\n>\n>create table t2 (\n> b int,\n> c int\n>) inherits (t);\n>\n>So \"t\"s are identified by an integer; and one kind of \"t\" has a single text\n>attribute while a different kind of \"t\" has 2 int attributes. The idea is\n>that there is a single primary key constraint on the whole hierarchy that\n>ensures only one record with a particular id can exist in all the tables\n>together. I can imagine wanting to do this with other unique constraints\n>also.\n>\n\nIMO the chances of us supporting global indexes with generic inheritance\nhierarchies are about zero. We don't even support creating \"partition\"\nindexes on those hierarchies ...\n\n>At present I don't actually use inheritance; instead I put triggers on the\n>child tables that do an insert on the parent table, which has the effect of\n>enforcing the uniqueness I want.\n\nDoes it? Are you sure it actually works in READ COMMITTED? What exactly\ndoes the trigger do?\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 31 Oct 2019 20:21:55 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On 10/30/19 10:27, Andres Freund wrote:\n> On 2019-10-30 13:05:57 -0400, Tom Lane wrote:\n>> Peter Geoghegan <pg@bowt.ie> writes:\n>>> On Wed, Oct 30, 2019 at 9:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>>> Well, the *effects* of the feature seem desirable, but that doesn't\n>>>> mean that we want an implementation that actually has a shared index.\n>>>> As soon as you do that, you've thrown away most of the benefits of\n>>>> having a partitioned data structure in the first place.\n>>\n>>> Right, but that's only the case for the global index. Global indexes\n>>> are useful when used judiciously.\n>>\n>> But ... why bother with partitioning then? To me, the main reasons\n>> why you might want a partitioned table are\n> \n> Quite commonly there's a lot of *other* indexes, often on a lot wider\n> data than the primary key, that don't need to be global. And whereas in\n> a lot of cases the primary key in a partitioned table has pretty good\n> locality (and thus will be mostly buffered IO), other indexes will often\n> not have that property (i.e. not have much correlation with table\n> position).\n\nI asked around a little bit and got some interesting responses. Thought\nI'd pass two of them along.\n\nOne person worked on a payments network (150,000+ installed readers),\nthe transaction table was date partitioned (1 per day) based on insert\ntimestamp, but lookups and updates were typically by the unique\ntransaction id. Oracle DB, they kept 180 daily partitions, several\nmillion rows per day. Transactions did not arrive in order, and could be\ndelayed if some part of the network was slow (they opted to allow the $2\ncharge rather than reject sales) and when the cash transaction records\nwere uploaded. Step one for their PG conversion created a read replica\nin PG 9.6, and the cost of doing the individual index lookups across 180\npartitions (and 180 indexes) was very high, so they stored max and min\ntxn id per partition and would generate a query with all the dates that\na txn id could have been in so that only a small number of partition\nindexes would be accessed. They wanted a global index on txn id for\nperformance, not for uniqueness – id generated on reader with guid-like\nsemantics.\n\nA second person worked on several large-scale systems and he relayed\nthat in some cases where they used Oracle global indexes on partitioned\ntables, they ended up deciding to reverse that decision as things scaled\nbecause of restrictive locking during partition maintenance (this is the\nexact issue Tom points out). So even on a database _with_ the option of\nusing a global index, they've sometimes opted for \"workaround\" design\npatterns instead:\n* To solve uniqueness, manage serialization at the appliation level.\nIsolate operations (e.g. using a queue) and use that to make sure that\ntwo sessions don’t try to insert the same record at the same time. From\nan RDBMS, this looks like a separate, smaller table that is being used\nto manage work activity.\n* To solve the additional IO for a global table scan ... We often don’t\nneed to do this because the load in this pattern is not typically highly\nconcurrent. If we are looking for higher concurrency, we can usually\nadd a hack/workaround that filters on a partition key to provide “pretty\ngood” pruning. The net result is that you get 2-3x the IO due to the\nlack of global index (same workaround as first story above).\n\nQuote: \"So ... I don’t actually like the idea of introducing this.\nUnless, someone can solve the ugly challenges we have had [around\npartition maintenance operations].\"\n\nI actually don't think those challenges are so un-solvable. I think that\nglobal indexes will be irrelevant to most workloads. I'm not entirely\nconvinced that they won't be useful for a few people with specific\nworkloads and large amounts of data in PostgreSQL where the benefits\noutweigh the costs. I definitely agree that care needs to be taken\naround index maintenance operations if there's an effort here.\n\n\n>> * ability to cheaply add and remove partitions, primarily so that\n>> you can cheaply do things like \"delete the oldest month's data\".\n> \n> You can still do that to some degree with a global index. Imagine\n> e.g. keeping a 'partition id' as a sort-of column in the global\n> index. That allows you to drop the partition, without having to\n> immediately rebuild the index, by checking the partition id against the\n> live partitions during lookup. So sure, your'e wasting space for a bit\n> in the global index, but it'll also be space that is likely to be fairly\n> efficiently reclaimed the next time vacuum touches the index. And if\n> not the global index can be rebuilt concurrently without blocking\n> writes.\n\nAnother idea might be to leverage PostgreSQL's partial indexes. If the\nindex is created \"where date>2020\" and you're dropping an index from\n2019 then you can entirely ignore the index. Not a panacea for every\nindex maintenance operation, but for the super-common case of dropping\nthe oldest partition you can now:\n\n1) create new index concurrently \"where dt>2020\"\n2) drop the old index\n3) drop the 2019 partition\n\ndoesn't solve world hunger but there's lots of benefit for such a simple\nhack.\n\n\n>> * ability to scale past our limits on the physical size of one table\n>> --- both the hard BlockNumber-based limit, and the performance\n>> constraints of e.g. vacuuming a very large table.\n> \n> For that to be a problem for a global index the global index (which will\n> often be something like two int4 or int8 columns) itself would need to\n> be above the block number based limit - which doesn't seem that close.\n> \n> WRT vacuuming - based on my observations the table itself isn't a\n> performance problem for vacuuming all that commonly anymore, it's the\n> associated index scans. So yea, that's a problem.\n\nI'm sure zheap will make all our dreams come true, right? :D\n\n-Jeremy\n\n\n-- \nJeremy Schneider\nDatabase Engineer\nAmazon Web Services\n\n\n","msg_date":"Mon, 25 Nov 2019 15:05:03 -0800","msg_from":"Jeremy Schneider <schnjere@amazon.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On 11/25/19 15:05, Jeremy Schneider wrote:\n> ... the cost of doing the individual index lookups across 180\n> partitions (and 180 indexes) was very high, so they stored max and min\n> txn id per partition and would generate a query with all the dates that\n> a txn id could have been in so that only a small number of partition\n> indexes would be accessed. \n> \n> .. If we are looking for higher concurrency, we can usually\n> add a hack/workaround that filters on a partition key to provide “pretty\n> good” pruning. The net result is that you get 2-3x the IO due to the\n> lack of global index (same workaround as first story above).\n\nIs that basically like a global BRIN index with granularity at the\npartition level?\n\n-J\n\n-- \nJeremy Schneider\nDatabase Engineer\nAmazon Web Services\n\n\n","msg_date":"Mon, 25 Nov 2019 15:44:39 -0800","msg_from":"Jeremy Schneider <schnjere@amazon.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Mon, Nov 25, 2019 at 03:44:39PM -0800, Jeremy Schneider wrote:\n> On 11/25/19 15:05, Jeremy Schneider wrote:\n> > ... the cost of doing the individual index lookups across 180\n> > partitions (and 180 indexes) was very high, so they stored max and min\n> > txn id per partition and would generate a query with all the dates that\n> > a txn id could have been in so that only a small number of partition\n> > indexes would be accessed. \n> > \n> > .. If we are looking for higher concurrency, we can usually\n> > add a hack/workaround that filters on a partition key to provide “pretty\n> > good” pruning. The net result is that you get 2-3x the IO due to the\n> > lack of global index (same workaround as first story above).\n> \n> Is that basically like a global BRIN index with granularity at the\n> partition level?\n\nExactly! :-)\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Wed, 18 Dec 2019 22:03:28 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On 19/12/19 4:03, Bruce Momjian wrote:\n> On Mon, Nov 25, 2019 at 03:44:39PM -0800, Jeremy Schneider wrote:\n>> On 11/25/19 15:05, Jeremy Schneider wrote:\n>>> ... the cost of doing the individual index lookups across 180\n>>> partitions (and 180 indexes) was very high, so they stored max and min\n>>> txn id per partition and would generate a query with all the dates that\n>>> a txn id could have been in so that only a small number of partition\n>>> indexes would be accessed.\n>>>\n>>> .. If we are looking for higher concurrency, we can usually\n>>> add a hack/workaround that filters on a partition key to provide “pretty\n>>> good” pruning. The net result is that you get 2-3x the IO due to the\n>>> lack of global index (same workaround as first story above).\n>> Is that basically like a global BRIN index with granularity at the\n>> partition level?\n> Exactly! :-)\n\nActually, one \"kind of\" BRIN index *per partitioned table* mapping (key \nrange) -> (partition oid)... and so concurrency doesn't need to be very \naffected.\n\n(we don't need to do things just like other RDBMS do, ya know... ;)\n\n\nIIRC, this precise approach was suggested around 2016 when initially \ndiscussing the \"declarative partitioning\" which originated Postgres' \ncurrent partitioning scheme, in order to optimize partition pruning.\n\n\nJust my .02€\n\n / J.L.\n\n\n\n\n","msg_date":"Thu, 19 Dec 2019 09:48:40 +0100","msg_from":"Jose Luis Tallon <jltallon@adv-solutions.net>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Thu, Dec 19, 2019 at 09:48:40AM +0100, Jose Luis Tallon wrote:\n> On 19/12/19 4:03, Bruce Momjian wrote:\n> > On Mon, Nov 25, 2019 at 03:44:39PM -0800, Jeremy Schneider wrote:\n> > > On 11/25/19 15:05, Jeremy Schneider wrote:\n> > > > ... the cost of doing the individual index lookups across 180\n> > > > partitions (and 180 indexes) was very high, so they stored max and min\n> > > > txn id per partition and would generate a query with all the dates that\n> > > > a txn id could have been in so that only a small number of partition\n> > > > indexes would be accessed.\n> > > > \n> > > > .. If we are looking for higher concurrency, we can usually\n> > > > add a hack/workaround that filters on a partition key to provide “pretty\n> > > > good” pruning. The net result is that you get 2-3x the IO due to the\n> > > > lack of global index (same workaround as first story above).\n> > > Is that basically like a global BRIN index with granularity at the\n> > > partition level?\n> > Exactly! :-)\n> \n> Actually, one \"kind of\" BRIN index *per partitioned table* mapping (key\n> range) -> (partition oid)... and so concurrency doesn't need to be very\n> affected.\n> \n> (we don't need to do things just like other RDBMS do, ya know... ;)\n> \n> \n> IIRC, this precise approach was suggested around 2016 when initially\n> discussing the \"declarative partitioning\" which originated Postgres' current\n> partitioning scheme, in order to optimize partition pruning.\n\nRobert Haas identified two needs for global indexes:\n\n\thttps://www.postgresql.org/message-id/CA+Tgmob_J2M2+QKWrhg2NjQEkMEwZNTfd7a6Ubg34fJuZPkN2g@mail.gmail.com\n\t\n\tOne of the biggest reasons why people want it is to enforce uniqueness\n\tfor secondary keys - e.g. the employees table is partitioned by\n\temployee ID, but SSN should also be unique, at least among employees\n\tfor whom it's not NULL.\n\t\n\tBut people also want it for faster data retrieval: if you're looking\n\tfor a commonly-occurring value, an index per partition is fine. But if\n\tyou're looking for values that occur only once or a few times across\n\tthe whole hierarchy, an index scan per partition is very costly.\n\nI don't see lossy BRIN indexes helping with the uniqueness use-case, and\nI am not sure they would help with the rare case either. They would\nhelp for range-based partitions, but I thought our existing facilities\nworked in that case.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Thu, 19 Dec 2019 11:12:07 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On 12/19/19 08:12, Bruce Momjian wrote:\n> I don't see lossy BRIN indexes helping with the uniqueness use-case, and\n> I am not sure they would help with the rare case either. They would\n> help for range-based partitions, but I thought our existing facilities\n> worked in that case.\n\nCorrelated data. The existing facilities work if the filtering column\nis exactly the same as the partition column. But it's not at all\nuncommon to have other columns with correlated data, perhaps the most\nobvious of which is timeseries tables with many date columns of various\ndefinitions (row first update, row latest update, invoice date, payment\ndate, process date, ship date, etc).\n\nWhat if you could use *two* indexes in a single execution plan? Use the\nglobal BRIN to narrow down to 2 or 3 out of a hundred or more\npartitions, then use local indexes to find specific rows in the\npartitions of interest? That might work, without being too overly\ncomplicated.\n\n-J\n\n\n-- \nJeremy Schneider\nDatabase Engineer\nAmazon Web Services\n\n\n","msg_date":"Thu, 19 Dec 2019 11:28:55 -0800","msg_from":"Jeremy Schneider <schnjere@amazon.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Thu, Dec 19, 2019 at 11:28:55AM -0800, Jeremy Schneider wrote:\n> On 12/19/19 08:12, Bruce Momjian wrote:\n> > I don't see lossy BRIN indexes helping with the uniqueness use-case, and\n> > I am not sure they would help with the rare case either. They would\n> > help for range-based partitions, but I thought our existing facilities\n> > worked in that case.\n> \n> Correlated data. The existing facilities work if the filtering column\n> is exactly the same as the partition column. But it's not at all\n> uncommon to have other columns with correlated data, perhaps the most\n> obvious of which is timeseries tables with many date columns of various\n> definitions (row first update, row latest update, invoice date, payment\n> date, process date, ship date, etc).\n> \n> What if you could use *two* indexes in a single execution plan? Use the\n> global BRIN to narrow down to 2 or 3 out of a hundred or more\n> partitions, then use local indexes to find specific rows in the\n> partitions of interest? That might work, without being too overly\n> complicated.\n\nNo, that is very interesting --- having secondary indexes for\npartitioned tables that trim most partitions. Would index lookups on\neach partition index be very slow? BRIN indexes? I am assuming global\nindexes would only avoid such lookups.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Thu, 19 Dec 2019 21:34:33 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"\n> On Oct 30, 2019, at 12:05 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> But ... why bother with partitioning then? To me, the main reasons\n> why you might want a partitioned table are\n> \n> * ability to cheaply add and remove partitions, primarily so that\n> you can cheaply do things like \"delete the oldest month's data\".\n> \n> * ability to scale past our limits on the physical size of one table\n> --- both the hard BlockNumber-based limit, and the performance\n> constraints of e.g. vacuuming a very large table.\n\nA third case is data locality. In that case global indexes would be useful for queries that do not correlate will with hot data.\n\n> Both of those go out the window with a global index. So you might\n> as well just have one table and forget all the overhead.\n\nPartition pruning could still be valuable even with global indexes, provided that we teach vacuum how to clean up tuples in an index that point at a partition that has been deleted.\n\n","msg_date":"Tue, 7 Jan 2020 22:55:13 +0000","msg_from":"\"Nasby, Jim\" <nasbyj@amazon.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"I've been following this topic for a long time. It's been a year since the last response.\nIt was clear that our customers wanted this feature as well, and a large number of them mentioned it.\n\nSo, I wish the whole feature to mature as soon as possible.\nI summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\n\nNext, I encountered some difficulties when implementing the DDL of the partition table with global index, and I hope to get some help from the community\n\nHere are some details what has been implemented\n1 Definition of global index\nUsing the INCLUDE keyword to include the tableoid of the partitioned table.\n\n2. Maintenance of global index by partition table DML.\nBoth INSERT and UPDATE of a partitioned table maintain global index\n\n3. Global index scan\nPlanner: Processes predicate conditions on the primary partition, generating paths and plans for the global index.\nExecuter: index scan get indextup, get the tableoid from indextup, and verify the visibility of the data in the partition.\n\n4. Vacuum partition table maintains global index\nEach partitioned table VACUUM cleans its own garbage data in the global index.\n\nAfter the above function point is completed, the global index can be used without partition table DDL.\n\nDemo:\n--Use pgbench to create the test partition table\npgbench -i -s 1000 --partitions=6 --partition-method=range\n\n—- create global index on bid, bid is not partition key\nCREATE INDEX idx_pgbench_accounts_bid on pgbench_accounts(bid) global;\n\n— check global index status\nselect * , sum(alivetup) over()as sum_alivetup, sum(deadtup) over() as sum_deadtup from bt_get_global_index_status('idx_pgbench_accounts_bid');\n relname | alivetup | deadtup | sum_alivetup | sum_deadtup \n--------------------+----------+---------+--------------+-------------\n pgbench_accounts_1 | 16666667 | 0 | 100000000 | 0\n pgbench_accounts_2 | 16666667 | 0 | 100000000 | 0\n pgbench_accounts_3 | 16666667 | 0 | 100000000 | 0\n pgbench_accounts_4 | 16666667 | 0 | 100000000 | 0\n pgbench_accounts_5 | 16666667 | 0 | 100000000 | 0\n pgbench_accounts_6 | 16666665 | 0 | 100000000 | 0\n(6 rows)\n\n— run pgbench for for a while\npgbench -M prepared -j 32 -c 32 -T 60 -P1\n\n\n—- check global index, The index has bloated\npostgres=# select * , sum(alivetup) over()as sum_alivetup, sum(deadtup) over() as sum_deadtup from bt_get_global_index_status('idx_pgbench_accounts_bid');\n relname | alivetup | deadtup | sum_alivetup | sum_deadtup \n--------------------+----------+---------+--------------+-------------\n pgbench_accounts_1 | 16717733 | 0 | 100306102 | 0\n pgbench_accounts_2 | 16717409 | 0 | 100306102 | 0\n pgbench_accounts_3 | 16717540 | 0 | 100306102 | 0\n pgbench_accounts_4 | 16717972 | 0 | 100306102 | 0\n pgbench_accounts_5 | 16717578 | 0 | 100306102 | 0\n pgbench_accounts_6 | 16717870 | 0 | 100306102 | 0\n(6 rows)\n\n—- vacuum partition table\nvacuum pgbench_accounts;\n\n—- Garbage is collected, global index still looks correct and valid.\npostgres=# select * , sum(alivetup) over()as sum_alivetup, sum(deadtup) over() as sum_deadtup from bt_get_global_index_status('idx_pgbench_accounts_bid');\n relname | alivetup | deadtup | sum_alivetup | sum_deadtup \n--------------------+----------+---------+--------------+-------------\n pgbench_accounts_1 | 16666667 | 0 | 100000000 | 0\n pgbench_accounts_2 | 16666667 | 0 | 100000000 | 0\n pgbench_accounts_3 | 16666667 | 0 | 100000000 | 0\n pgbench_accounts_4 | 16666667 | 0 | 100000000 | 0\n pgbench_accounts_5 | 16666667 | 0 | 100000000 | 0\n pgbench_accounts_6 | 16666665 | 0 | 100000000 | 0\n(6 rows)\n\n—-\n\n—- global index scan works well\npostgres=# select tableoid ,count(*) from pgbench_accounts where bid = 834 group by tableoid;\n tableoid | count \n----------+-------\n 16455 | 33335\n 16458 | 66665\n(2 rows)\n\npostgres=# explain select tableoid ,count(*) from pgbench_accounts where bid = 834 group by tableoid;\n QUERY PLAN \n--------------------------------------------------------------------------------------------------------------------\n HashAggregate (cost=2945.23..2945.24 rows=1 width=12)\n Group Key: pgbench_accounts.tableoid\n -> Global Index Scan using idx_pgbench_accounts_bid on pgbench_accounts (cost=0.50..10.18 rows=587011 width=4)\n Index Cond: (bid = 834)\n(4 rows)\n\n\nThe following is how to implement DDL of global index. How to maintain global index of DDL of partitioned table.\nThis seems to be more difficult than the previous work.\n\nI understand there are four main parts\n\n1 Build global index or reindex, especially in concurrent mode\n\n2 Detach partition\nWould it be a good idea to make a flag to global index and let VACUUM handle the index data of the Detach partition?\n\n3 Attach partition\nIt is easy to Attach a new empty partition, but adding a new one with data is not.\nIf there is a unique key conflict, do we slowly clean up the garbage or invalidate the entire index?\n\n4 Truncate partition with global index\nDo we need to process the heap and index data separately in multiple transactions?\nThis will lose the ability to roll back for Truncate operation.\nIs it worth it?\n\n\nLooking forward to your feedback.\n\nThanks!\n\nWenjing","msg_date":"Thu, 7 Jan 2021 17:44:01 +0800","msg_from":"=?utf-8?B?5pu+5paH5peM?= <wenjing.zwj@alibaba-inc.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Thu, Jan 7, 2021 at 05:44:01PM +0800, 曾文旌 wrote:\n> I've been following this topic for a long time. It's been a year since the last response.\n> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\n> \n> So, I wish the whole feature to mature as soon as possible.\n> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\n\nI think you need to address the items mentioned in this blog, and the\nemail link it mentions:\n\n\thttps://momjian.us/main/blogs/pgblog/2020.html#July_1_2020\n\nI am not clear this is a feature we will want. Yes, people ask for it,\nbut if the experience will be bad for them and they will regret using\nit, I am not sure we want it. Of course, if you code it up and we get\na good user experience, we would want it --- I am just saying it is not\nclear right now.\n\n-- \n Bruce Momjian <bruce@momjian.us> https://momjian.us\n EnterpriseDB https://enterprisedb.com\n\n The usefulness of a cup is in its emptiness, Bruce Lee\n\n\n\n","msg_date":"Thu, 7 Jan 2021 09:16:20 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Thu, Jan 7, 2021 at 4:44 AM 曾文旌 <wenjing.zwj@alibaba-inc.com> wrote:\n> I've been following this topic for a long time. It's been a year since the last response.\n> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\n>\n> So, I wish the whole feature to mature as soon as possible.\n> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\n\nYou need to summarize the basic design choices you've made here. Like,\nwhat did you do about the fact that TIDs have to be 6 bytes?\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 7 Jan 2021 10:04:54 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"> 2021年1月7日 22:16，Bruce Momjian <bruce@momjian.us> 写道：\n> \n> On Thu, Jan 7, 2021 at 05:44:01PM +0800, 曾文旌 wrote:\n>> I've been following this topic for a long time. It's been a year since the last response.\n>> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\n>> \n>> So, I wish the whole feature to mature as soon as possible.\n>> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\n> \n> I think you need to address the items mentioned in this blog, and the\n> email link it mentions:\n> \n> \thttps://momjian.us/main/blogs/pgblog/2020.html#July_1_2020\n\nThank you for your reply.\nI read your blog and it helped me a lot.\n\nThe blog mentions a specific problem: \"A large global index might also reintroduce problems that prompted the creation of partitioning in the first place. \"\nI don't quite understand, could you give some specific information?\n\nIn addition you mentioned: \"It is still unclear if these use-cases justify the architectural changes needed to enable global indexes.\"\nPlease also describe the problems you see, I will confirm each specific issue one by one.\n\n\nThanks\n\nWenjing\n\n\n> \n> I am not clear this is a feature we will want. Yes, people ask for it,\n> but if the experience will be bad for them and they will regret using\n> it, I am not sure we want it. Of course, if you code it up and we get\n> a good user experience, we would want it --- I am just saying it is not\n> clear right now.\n> \n> -- \n> Bruce Momjian <bruce@momjian.us> https://momjian.us\n> EnterpriseDB https://enterprisedb.com\n> \n> The usefulness of a cup is in its emptiness, Bruce Lee","msg_date":"Fri, 8 Jan 2021 11:26:48 +0800","msg_from":"=?utf-8?B?5pu+5paH5peM?= <wenjing.zwj@alibaba-inc.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Fri, Jan 8, 2021 at 4:02 PM 曾文旌 <wenjing.zwj@alibaba-inc.com> wrote:\n>\n> > 2021年1月7日 22:16，Bruce Momjian <bruce@momjian.us> 写道：\n> >\n> > On Thu, Jan 7, 2021 at 05:44:01PM +0800, 曾文旌 wrote:\n> >> I've been following this topic for a long time. It's been a year since the last response.\n> >> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\n> >>\n> >> So, I wish the whole feature to mature as soon as possible.\n> >> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\n> >\n> > I think you need to address the items mentioned in this blog, and the\n> > email link it mentions:\n> >\n> > https://momjian.us/main/blogs/pgblog/2020.html#July_1_2020\n>\n> Thank you for your reply.\n> I read your blog and it helped me a lot.\n>\n> The blog mentions a specific problem: \"A large global index might also reintroduce problems that prompted the creation of partitioning in the first place. \"\n> I don't quite understand, could you give some specific information?\n>\n> In addition you mentioned: \"It is still unclear if these use-cases justify the architectural changes needed to enable global indexes.\"\n> Please also describe the problems you see, I will confirm each specific issue one by one.\n\nOne example is date partitioning. People frequently need to store\nonly the most recent data. For instance doing a monthly partitioning\nand dropping the oldest partition every month once you hit the wanted\nretention is very efficient for that use case, as it should be almost\ninstant (provided that you can acquire the necessary locks\nimmediately). But if you have a global index, you basically lose the\nadvantage of partitioning as it'll require heavy changes on that\nindex.\n\n\n","msg_date":"Fri, 8 Jan 2021 16:26:56 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Fri, Jan 8, 2021 at 11:26:48AM +0800, 曾文旌 wrote:\n> > On Thu, Jan 7, 2021 at 05:44:01PM +0800, 曾文旌 wrote:\n> >> I've been following this topic for a long time. It's been a year since the last response.\n> >> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\n> >> \n> >> So, I wish the whole feature to mature as soon as possible.\n> >> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\n> > \n> > I think you need to address the items mentioned in this blog, and the\n> > email link it mentions:\n> > \n> > \thttps://momjian.us/main/blogs/pgblog/2020.html#July_1_2020\n> \n> Thank you for your reply.\n> I read your blog and it helped me a lot.\n> \n> The blog mentions a specific problem: \"A large global index might also reintroduce problems that prompted the creation of partitioning in the first place. \"\n> I don't quite understand, could you give some specific information?\n\nWell, if you created partitions, you probably did so because:\n\n1. heap files are smaller, allowing for more targeted sequential scans\n2. smaller indexes\n3. the ability to easily drop tables/indexes that are too old\n\nIf you have global indexes, #1 is the same, but #2 is not longer true,\nand for #3, you can drop the heap but the index entries still exist in\nthe global index and must be removed.\n\nSo, if you created partitions for one of the three reasons, once you\nhave global indexes, some of those advantage of partitioning are no\nlonger true. I am sure there are some workloads where the advantages of\npartitioning, minus the advantages lost when using global indexes, are\nuseful, but are there enough of them to make the feature useful? I\ndon't know.\n\n> In addition you mentioned: \"It is still unclear if these use-cases justify the architectural changes needed to enable global indexes.\"\n> Please also describe the problems you see, I will confirm each specific issue one by one.\n\nWell, the email thread I linked to has a lot of them, but the\nfundamental issue is that you have to break the logic that a single\nindex serves a single heap file. Considering what I said above, is\nthere enough usefulness to warrant such an architectural change?\n\n-- \n Bruce Momjian <bruce@momjian.us> https://momjian.us\n EnterpriseDB https://enterprisedb.com\n\n The usefulness of a cup is in its emptiness, Bruce Lee\n\n\n\n","msg_date":"Fri, 8 Jan 2021 10:50:40 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"> 2021年1月8日 16:26，Julien Rouhaud <rjuju123@gmail.com> 写道：\n> \n> On Fri, Jan 8, 2021 at 4:02 PM 曾文旌 <wenjing.zwj@alibaba-inc.com> wrote:\n>> \n>>> 2021年1月7日 22:16，Bruce Momjian <bruce@momjian.us> 写道：\n>>> \n>>> On Thu, Jan 7, 2021 at 05:44:01PM +0800, 曾文旌 wrote:\n>>>> I've been following this topic for a long time. It's been a year since the last response.\n>>>> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\n>>>> \n>>>> So, I wish the whole feature to mature as soon as possible.\n>>>> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\n>>> \n>>> I think you need to address the items mentioned in this blog, and the\n>>> email link it mentions:\n>>> \n>>> https://momjian.us/main/blogs/pgblog/2020.html#July_1_2020\n>> \n>> Thank you for your reply.\n>> I read your blog and it helped me a lot.\n>> \n>> The blog mentions a specific problem: \"A large global index might also reintroduce problems that prompted the creation of partitioning in the first place. \"\n>> I don't quite understand, could you give some specific information?\n>> \n>> In addition you mentioned: \"It is still unclear if these use-cases justify the architectural changes needed to enable global indexes.\"\n>> Please also describe the problems you see, I will confirm each specific issue one by one.\n> \n> One example is date partitioning. People frequently need to store\n> only the most recent data. For instance doing a monthly partitioning\n> and dropping the oldest partition every month once you hit the wanted\n> retention is very efficient for that use case, as it should be almost\n> instant (provided that you can acquire the necessary locks\n> immediately). But if you have a global index, you basically lose the\n> advantage of partitioning as it'll require heavy changes on that\n> index.\nIf the global index removes all the major benefits of partitioned tables, then it is not worth having it.\n\nThis is indeed a typical scenario for a partitioned table.\nthere are two basic operations\n1) Monthly DETACH old child table\n2) Monthly ATTACH new child table\n\nFor 1) The DETACH old child table can be finished immediately, global index can be kept valid after DETACH is completed, and the cleanup of garbage data in global index can be deferred to VACUUM.\nThis is similar to the global index optimization done by Oracle12c.\nFor 2) ATTACH new empty child table can also be completed immediately.\nIf this is the case, many of the advantages of partitioned tables will be retained, while the advantages of global indexes will be gained.","msg_date":"Mon, 11 Jan 2021 19:40:18 +0800","msg_from":"=?utf-8?B?5pu+5paH5peM?= <wenjing.zwj@alibaba-inc.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"> 2021年1月7日 23:04，Robert Haas <robertmhaas@gmail.com> 写道：\n> \n> On Thu, Jan 7, 2021 at 4:44 AM 曾文旌 <wenjing.zwj@alibaba-inc.com> wrote:\n>> I've been following this topic for a long time. It's been a year since the last response.\n>> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\n>> \n>> So, I wish the whole feature to mature as soon as possible.\n>> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\n> \n> You need to summarize the basic design choices you've made here. Like,\n> what did you do about the fact that TIDs have to be 6 bytes?\n\nThese are the basic choices, and most of them come from discussions in previous emails.\n\n1 Definition of global index\nObviously, we need to expand Index address info(CTID) to include child table info in GlobalIndexTuple.\n\n1.1 As mentioned in the previous email, Bruce suggested having the OID\ninstead of relfilenode as relfilenode can be duplicated across tablespaces. \nI agree with that.\n\n1.2 And Heikki pointed me to include heap specific information using the INCLUDE keyword so that heap information\nis stored with each index node as data.\n\nSo ,In POC stage, I choose use INCLUDE keyword to INCLUDE the tableoid of global index. This will add 4 bytes to each IndexTuple.\n\nConsidering that if a single partitioned table does not exceed 65535 child tables, perhaps two bytes for tracking which child table the data belongs to is sufficient.\n\n2. Maintenance of global index by partition table DML.\nThe DML of each child table of the partitioned table needs to maintain the global index on the partitioned table.\n\n3. Global index scan\nPlanner: \nProcesses predicate on the primary partition, generating paths and plans for the global index.\nThe cost model of the global index may need to be considered. We need to make the global index or the local index selected in their respective advantageous scenarios.\n\nExecuter: \nThe index scan get indextup, get the tableoid from indextup, and verify the visibility of the data in the child table.\nIf a child table is DETACH, then the index item of this table is ignored during the index scan until VACUUM finishes cleaning up the global index.\n\n4. Vacuum partition table maintains global index.\nOld data in the global index also needs to be cleaned up, and vaccum is suitable for it.\nEach child table in VACUUM, while vacuuming its own index, also vacuums the global index on the partitioned table.\n\n5. Other\nThe global index indexes all of the child tables, which makes the global index large and has many levels. \nFollow the technical route, The partitioned indexes are a further target.\n\nThis is my basic idea for implementing global index.\nLooking forward to your feedback.\n\nThanks!\n\nWenjing\n\n> \n> -- \n> Robert Haas\n> EDB: http://www.enterprisedb.com","msg_date":"Mon, 11 Jan 2021 21:00:45 +0800","msg_from":"=?utf-8?B?5pu+5paH5peM?= <wenjing.zwj@alibaba-inc.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Mon, Jan 11, 2021 at 07:40:18PM +0800, 曾文旌 wrote:\n> >> In addition you mentioned: \"It is still unclear if these use-cases justify the architectural changes needed to enable global indexes.\"\n> >> Please also describe the problems you see, I will confirm each specific issue one by one.\n> > \n> > One example is date partitioning. People frequently need to store\n> > only the most recent data. For instance doing a monthly partitioning\n> > and dropping the oldest partition every month once you hit the wanted\n> > retention is very efficient for that use case, as it should be almost\n> > instant (provided that you can acquire the necessary locks\n> > immediately). But if you have a global index, you basically lose the\n> > advantage of partitioning as it'll require heavy changes on that\n> > index.\n> If the global index removes all the major benefits of partitioned tables, then it is not worth having it.\n> \n> This is indeed a typical scenario for a partitioned table.\n> there are two basic operations\n> 1) Monthly DETACH old child table\n> 2) Monthly ATTACH new child table\n> \n> For 1) The DETACH old child table can be finished immediately, global index can be kept valid after DETACH is completed, and the cleanup of garbage data in global index can be deferred to VACUUM.\n> This is similar to the global index optimization done by Oracle12c.\n> For 2) ATTACH new empty child table can also be completed immediately.\n> If this is the case, many of the advantages of partitioned tables will be retained, while the advantages of global indexes will be gained.\n\nYes, we can keep the index rows for the deleted partition and clean them\nup later, but what is the advantage of partitioning then? Just heap\ndeletion quickly? Is that enough of a value?\n\n-- \n Bruce Momjian <bruce@momjian.us> https://momjian.us\n EnterpriseDB https://enterprisedb.com\n\n The usefulness of a cup is in its emptiness, Bruce Lee\n\n\n\n","msg_date":"Mon, 11 Jan 2021 12:46:49 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"Bruce Momjian <bruce@momjian.us> writes:\n> On Mon, Jan 11, 2021 at 07:40:18PM +0800, 曾文旌 wrote:\n>> This is indeed a typical scenario for a partitioned table.\n>> there are two basic operations\n>> 1) Monthly DETACH old child table\n>> 2) Monthly ATTACH new child table\n>> \n>> For 1) The DETACH old child table can be finished immediately, global index can be kept valid after DETACH is completed, and the cleanup of garbage data in global index can be deferred to VACUUM.\n\n> Yes, we can keep the index rows for the deleted partition and clean them\n> up later, but what is the advantage of partitioning then? Just heap\n> deletion quickly? Is that enough of a value?\n\nMore to the point, you still have a massive index cleanup operation to do.\nDeferred or not, that's going to take a lot of cycles, and it will leave\nyou with a bloated global index. I find it hard to believe that this\napproach will seem like an improvement over doing partitioning the way\nwe do it now.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 11 Jan 2021 13:34:32 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Mon, Jan 11, 2021 at 12:46 PM Bruce Momjian <bruce@momjian.us> wrote:\n> > For 1) The DETACH old child table can be finished immediately, global index can be kept valid after DETACH is completed, and the cleanup of garbage data in global index can be deferred to VACUUM.\n> > This is similar to the global index optimization done by Oracle12c.\n> > For 2) ATTACH new empty child table can also be completed immediately.\n> > If this is the case, many of the advantages of partitioned tables will be retained, while the advantages of global indexes will be gained.\n>\n> Yes, we can keep the index rows for the deleted partition and clean them\n> up later, but what is the advantage of partitioning then? Just heap\n> deletion quickly? Is that enough of a value?\n\nI actually think the idea of lazily deleting the index entries is\npretty good, but it won't work if the way the global index is\nimplemented is by adding a tableoid column. Because then, I might\ndetach a partition and later reattach it and the old index entries are\nstill there but the table contents might have changed. Worse yet, the\ntable might be dropped and the table OID reused for a completely\nunrelated table with completely unrelated contents, which could then\nbe attached as a new partition.\n\nOne of the big selling points of global indexes is that they allow you\nto enforce uniqueness on a column unrelated to the partitioning\ncolumn. Another is that you can look up a value by doing a single\nindex scan on the global index rather than an index scan per\npartition. Those things are no less valuable for performing index\ndeletion lazily.\n\nHowever, there is a VACUUM amplification effect to worry about here\nwhich Wenjing seems not to be considering. Suppose I have a table\nwhich is not partitioned and it is 1TB in size with an index that is\n128GB in size. To vacuum the table, I need to do 1TB + 128GB of I/O.\nNow, suppose I now partition the table into 1024 partitions each with\nits own local index. Each partition is 1GB in size and the index on\neach partition is 128MB in size. To vacuum an individual partition\nrequires 1GB + 128MB of I/O, so to vacuum all the partitions requires\nthe same amount of total I/O as before. But, now suppose that I have a\nsingle global index instead of a local index per partition. First, how\nbig will that index be? It will not be 128GB, but somewhat bigger,\nbecause it needs extra space for every indexed tuple. Let's say 140GB.\nFurthermore, it will need to be vacuumed whenever any child is\nvacuumed, because it contains some index entries from every child. So\nthe total I/O to vacuum all partitions is now 1GB * 1024 + 140GB *\n1024 = 141TB, which is a heck of a lot worse than the 1.125TB I\nrequired with the unpartitioned table or the locally partitioned\ntable.\n\nThat's not necessarily a death sentence for every use case, but it's\ngoing to be pretty bad for tables that are big and heavily updated.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 11 Jan 2021 13:37:02 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Mon, Jan 11, 2021 at 10:37 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> I actually think the idea of lazily deleting the index entries is\n> pretty good, but it won't work if the way the global index is\n> implemented is by adding a tableoid column.\n\nPerhaps there is an opportunity to apply some of the infrastructure\nthat Masahiko Sawada has been working on, that makes VACUUM more\nincremental in certain specific scenarios:\n\nhttps://postgr.es/m/CAD21AoD0SkE11fMw4jD4RENAwBMcw1wasVnwpJVw3tVqPOQgAw@mail.gmail.com\n\nI think that VACUUM can be taught to skip the ambulkdelete() step for\nindexes in many common scenarios. Global indexes might be one place in\nwhich that's almost essential.\n\n> However, there is a VACUUM amplification effect to worry about here\n> which Wenjing seems not to be considering.\n\n> That's not necessarily a death sentence for every use case, but it's\n> going to be pretty bad for tables that are big and heavily updated.\n\nThe main way in which index vacuuming is currently a death sentence\nfor this design (as you put it) is that it's an all-or-nothing thing.\nPresumably you'll need to VACUUM the entire global index for each\npartition that receives even one UPDATE. That seems pretty extreme,\nand probably not acceptable. In a way it's not really a new problem,\nbut the fact remains: it makes global indexes much less valuable.\n\nHowever, it probably would be okay if a global index feature performed\npoorly in scenarios where partitions get lots of UPDATEs that produce\nlots of index bloat and cause lots of LP_DEAD line pointers to\naccumulate in heap pages. It is probably reasonable to just expect\nusers to not do that if they want to get acceptable performance while\nusing a global index. Especially since it probably is not so bad if\nthe index bloat situation gets out of hand for just one of the\npartitions (say the most recent one) every once in a while. You at\nleast don't have the same crazy I/O multiplier effect that you\ndescribed.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Mon, 11 Jan 2021 11:01:20 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Mon, Jan 11, 2021 at 01:37:02PM -0500, Robert Haas wrote:\n> However, there is a VACUUM amplification effect to worry about here\n...\n> That's not necessarily a death sentence for every use case, but it's\n> going to be pretty bad for tables that are big and heavily updated.\n\nYeah, I had not really gotten that far in my thinking, but someone is\ngoing to need to create a POC and then we need to test it to see if it\noffers a reasonably valuable feature.\n\n-- \n Bruce Momjian <bruce@momjian.us> https://momjian.us\n EnterpriseDB https://enterprisedb.com\n\n The usefulness of a cup is in its emptiness, Bruce Lee\n\n\n\n","msg_date":"Mon, 11 Jan 2021 14:23:13 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Mon, Jan 11, 2021 at 11:01:20AM -0800, Peter Geoghegan wrote:\n> However, it probably would be okay if a global index feature performed\n> poorly in scenarios where partitions get lots of UPDATEs that produce\n> lots of index bloat and cause lots of LP_DEAD line pointers to\n> accumulate in heap pages. It is probably reasonable to just expect\n> users to not do that if they want to get acceptable performance while\n> using a global index. Especially since it probably is not so bad if\n> the index bloat situation gets out of hand for just one of the\n> partitions (say the most recent one) every once in a while. You at\n> least don't have the same crazy I/O multiplier effect that you\n> described.\n\nOnce you layer on all the places a global index will be worse than just\ncreating a single large table, or a partitioned table with an index per\nchild, there might not be much usefulness left. A POC patch might tell\nus that, and might allow us to mark it as \"not wanted\".\n\n-- \n Bruce Momjian <bruce@momjian.us> https://momjian.us\n EnterpriseDB https://enterprisedb.com\n\n The usefulness of a cup is in its emptiness, Bruce Lee\n\n\n\n","msg_date":"Mon, 11 Jan 2021 14:25:55 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Mon, Jan 11, 2021 at 11:25 AM Bruce Momjian <bruce@momjian.us> wrote:\n> Once you layer on all the places a global index will be worse than just\n> creating a single large table, or a partitioned table with an index per\n> child, there might not be much usefulness left. A POC patch might tell\n> us that, and might allow us to mark it as \"not wanted\".\n\nI'm confused. Of course it's true to some degree that having a global\nindex \"defeats the purpose\" of having a partitioned table. But only to\na degree. And for some users it will make the difference between using\npartitioning and not using partitioning -- they simply won't be able\nto tolerate not having it available (e.g. because of a requirement for\na unique constraint that does not cover the partitioning key).\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Mon, 11 Jan 2021 12:05:43 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"On Mon, Jan 11, 2021 at 12:05:43PM -0800, Peter Geoghegan wrote:\n> On Mon, Jan 11, 2021 at 11:25 AM Bruce Momjian <bruce@momjian.us> wrote:\n> > Once you layer on all the places a global index will be worse than just\n> > creating a single large table, or a partitioned table with an index per\n> > child, there might not be much usefulness left. A POC patch might tell\n> > us that, and might allow us to mark it as \"not wanted\".\n> \n> I'm confused. Of course it's true to some degree that having a global\n> index \"defeats the purpose\" of having a partitioned table. But only to\n> a degree. And for some users it will make the difference between using\n> partitioning and not using partitioning -- they simply won't be able\n> to tolerate not having it available (e.g. because of a requirement for\n> a unique constraint that does not cover the partitioning key).\n\nYes, that is a good point. For those cases, I think we need to look at\nthe code complexity/overhead of supporting that feature. There are\ngoing to be a few cases it is a win, but will the code complexity be\nworth it?\n\n-- \n Bruce Momjian <bruce@momjian.us> https://momjian.us\n EnterpriseDB https://enterprisedb.com\n\n The usefulness of a cup is in its emptiness, Bruce Lee\n\n\n\n","msg_date":"Mon, 11 Jan 2021 15:24:06 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"},{"msg_contents":"> 2021年1月12日 02:37，Robert Haas <robertmhaas@gmail.com> 写道：\n> \n> On Mon, Jan 11, 2021 at 12:46 PM Bruce Momjian <bruce@momjian.us> wrote:\n>>> For 1) The DETACH old child table can be finished immediately, global index can be kept valid after DETACH is completed, and the cleanup of garbage data in global index can be deferred to VACUUM.\n>>> This is similar to the global index optimization done by Oracle12c.\n>>> For 2) ATTACH new empty child table can also be completed immediately.\n>>> If this is the case, many of the advantages of partitioned tables will be retained, while the advantages of global indexes will be gained.\n>> \n>> Yes, we can keep the index rows for the deleted partition and clean them\n>> up later, but what is the advantage of partitioning then? Just heap\n>> deletion quickly? Is that enough of a value?\n> \n> I actually think the idea of lazily deleting the index entries is\n> pretty good, but it won't work if the way the global index is\n> implemented is by adding a tableoid column. Because then, I might\n> detach a partition and later reattach it and the old index entries are\n> still there but the table contents might have changed. Worse yet, the\n> table might be dropped and the table OID reused for a completely\n> unrelated table with completely unrelated contents, which could then\n> be attached as a new partition.\n> \n> One of the big selling points of global indexes is that they allow you\n> to enforce uniqueness on a column unrelated to the partitioning\n> column. Another is that you can look up a value by doing a single\n> index scan on the global index rather than an index scan per\n> partition. Those things are no less valuable for performing index\n> deletion lazily.\n> \n> However, there is a VACUUM amplification effect to worry about here\n> which Wenjing seems not to be considering. Suppose I have a table\n> which is not partitioned and it is 1TB in size with an index that is\n> 128GB in size. To vacuum the table, I need to do 1TB + 128GB of I/O.\n> Now, suppose I now partition the table into 1024 partitions each with\n> its own local index. Each partition is 1GB in size and the index on\n> each partition is 128MB in size. To vacuum an individual partition\n> requires 1GB + 128MB of I/O, so to vacuum all the partitions requires\n> the same amount of total I/O as before. But, now suppose that I have a\n> single global index instead of a local index per partition. First, how\n> big will that index be? It will not be 128GB, but somewhat bigger,\n> because it needs extra space for every indexed tuple. Let's say 140GB.\n> Furthermore, it will need to be vacuumed whenever any child is\n> vacuumed, because it contains some index entries from every child. So\n> the total I/O to vacuum all partitions is now 1GB * 1024 + 140GB *\n> 1024 = 141TB, which is a heck of a lot worse than the 1.125TB I\n> required with the unpartitioned table or the locally partitioned\n> table.\nThank you for pointing this out.\nIt seems that some optimization can be done, but there is no good way\nto completely eliminate the vacuum amplification effect of the global index.\nMaybe we can only count on Zheap, which doesn't need to do Vaccum.\n\n\n\n> \n> That's not necessarily a death sentence for every use case, but it's\n> going to be pretty bad for tables that are big and heavily updated.\n> \n> -- \n> Robert Haas\n> EDB: http://www.enterprisedb.com","msg_date":"Mon, 18 Jan 2021 18:00:40 +0800","msg_from":"=?utf-8?B?5pu+5paH5peM?= <wenjing.zwj@alibaba-inc.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Global Index"}],"string":"[\n {\n \"msg_contents\": \"A global index by very definition is a single index on the parent table\\nthat maps to many\\nunderlying table partitions. The parent table itself does not have any\\nunderlying storage,\\nso it must, therefore, retrieve the data satisfying index constraints from\\nthe underlying tables.\\nIn very crude terms, it is an accumulation of data from table partitions so\\nthat data spanning\\nacross multiple partitions are accessed in one go as opposed to\\nindividually querying each\\npartition.\\n\\nFor the initial version of this work, we are only considering to build\\nb-tree global indexes.\\n\\n- Partitioned Index (Index Partitioning)\\nWhen global indexes become too large, then those are partitioned to keep\\nthe performance\\nand maintenance overhead manageable. These are not within the scope of this\\nwork.\\n\\n\\n- Local Index\\nA local index is an index that is local to a specific table partition; i.e.\\nit doesn’t span across\\nmultiple partitions. So, when we create an index on a parent table, it will\\ncreate a separate\\nindex for all its partitions. Unfortunately, PostgreSQL uses the\\nterminology of “partitioned index”\\nwhen it refers to local indexes. This work with fix this terminology for\\nPostgreSQL so that the\\nnomenclature remains consistent with other DBMS.\\n\\n\\n- Why We Need Global Index?\\nA global index is expected to give two very important upgrades to the\\npartitioning feature set in\\nPostgreSQL. It is expected to give a significant improvement in\\nread-performance for queries\\ntargeting multiple local indexes of partitions. It also adds a unique\\nconstraint across partitions.\\n\\n\\n- Unique Constraint\\nData uniqueness is a critical requirement for building an index. For global\\nindexes that span across\\nmultiple partitions, uniqueness will have to be enforced on index\\ncolumn(s). This effectively translates\\ninto a unique constraint.\\n\\n\\n- Performance\\nCurrently, the pseudo index created on the parent table of partitions does\\nnot contain any\\ndata. Rather, it dereferences to the local indexes when an index search is\\nrequired. This\\nmeans that multiple indexes will have to be evaluated and data to be\\ncombined thereafter.\\nHowever, with the global indexes, data will reside with global index\\ndeclared on the parent\\ntable. This avoids the need for multi-level index lookups. So read\\nperformance is expected\\nto be significantly higher in cases. There will however be a negative\\nperformance impact\\nduring write (insert/update) of data. This is discussed in more detail\\nlater on.\\n\\n\\n- Creating a GLOBAL Index - Syntax\\nA global index may be created with the addition of a “GLOBAL” keyword to\\nthe index statement.\\nAlternatively, one could specify the “LOCAL” keyword to create local\\nindexes on partitions.\\nWe are suggesting to call this set of keywords: “partition_index_type”. By\\ndefault,\\npartition_index_type will be set as LOCAL. Here is a sample of the create\\nindex syntax.\\n\\nCREATE Index idx parent (columns) [GLOBAL | LOCAL];\\n\\nNote: There is no shift/reduced by adding these options.\\n\\n\\n- Pointing Index to Tuple\\nCurrently, CTID carries a page and offset information for a known heap\\n(table name). However,\\nin the context of global indexes, this information within an index is\\ninsufficient. Since the index is\\nexpected to carry tuples from multiple partitions (heaps), CTID alone will\\nnot be able to link an index\\nnode to a tuple. This requires carrying additional data for the heap name\\nto be stored with each\\nindex node.\\n\\nHow this should be implemented is a point to be discussed. A few\\npossibilities are listed below:\\n\\n-- Expand CTID to include a relfilenode id. In PostgreSQL-Conf Asia, Bruce\\nsuggested having the OID\\ninstead of relfilenode as relfilenode can be duplicated across tablespaces.\\n -- Using OID introduces another complication where we would need to\\nquery catalog for OID to\\n heap mapping.\\n\\n-- The second option is to have a variable-length CTID. We can reserve some\\ntop-level bit for segregation of\\nGlobal CTID or Standard CTID. Robert Haas suggested in PostgreSQL-EU to\\ndiscuss this with Peter Geoghegan.\\n -- I discussed it with Peter and he believes that it is a very\\ninvasive approach that requires a whole lot of\\n the effort to get committed.\\n\\n-- Heikki pointed me to include heap specific information using the INCLUDE\\nkeyword so that heap information\\nis stored with each index node as data.\\n -- We (Peter and I) also discussed that option and this looks a more\\neasy and non-invasive option.\\n\\n\\n- Optimizer\\nThe challenge with optimizer is a selection between local and global\\nindexes when both are present.\\nHow do we:\\n\\n-- Evaluate the cost of scanning a global index?\\n-- When should the LOCAL index be preferred over the GLOBAL index and vice\\nversa?\\n -- Should we hit a GLOBAL index when the query is targeting a\\ncouple of partitions only?\\n -- We need to consider the sizes of those partitions being hit and\\nthe sizes of partitions not being hit.\\n-- Bruce suggested that we prioritize a GLOBAL index in the first version\\nso that in every case,\\nthe GLOBAL index is utilized.\\n\\n\\n- Write Performance and Vacuum\\nThere will be some write performance degradation because every change in\\npartition tables must\\npropagate upwards to the GLOBAL index on the parent table. This can be\\nthought of as another index\\non a table, however, the [slight] performance degradation will be due to\\nthe fact that the GLOBAL\\nindex may carry a much bigger dataset with data from multiple partitions\\nresulting in a higher tree\\ntraversal and update time. This applies to both write and vacuum processes.\\n\\nIt is still an open question though on how this will be handled within the\\ncode and how we can better\\noptimize this process.\\n\\nI have a POC patch and working on finalizing the patch, Hamid Akhtar is\\nalso working with me on this\\nwork.\\n\\n\\n\\n--\\nIbrar Ahmed\\n\\nA global index by very definition is a single index on the parent table that maps to manyunderlying table partitions. The parent table itself does not have any underlying storage,so it must, therefore, retrieve the data satisfying index constraints from the underlying tables.In very crude terms, it is an accumulation of data from table partitions so that data spanningacross multiple partitions are accessed in one go as opposed to individually querying eachpartition.For the initial version of this work, we are only considering to build b-tree global indexes.- Partitioned Index (Index Partitioning)When global indexes become too large, then those are partitioned to keep the performanceand maintenance overhead manageable. These are not within the scope of this work.- Local IndexA local index is an index that is local to a specific table partition; i.e. it doesn’t span acrossmultiple partitions. So, when we create an index on a parent table, it will create a separateindex for all its partitions. Unfortunately, PostgreSQL uses the terminology of “partitioned index”when it refers to local indexes. This work with fix this terminology for PostgreSQL so that thenomenclature remains consistent with other DBMS. - Why We Need Global Index?A global index is expected to give two very important upgrades to the partitioning feature set inPostgreSQL. It is expected to give a significant improvement in read-performance for queriestargeting multiple local indexes of partitions. It also adds a unique constraint across partitions.- Unique ConstraintData uniqueness is a critical requirement for building an index. For global indexes that span acrossmultiple partitions, uniqueness will have to be enforced on index column(s). This effectively translatesinto a unique constraint.- PerformanceCurrently, the pseudo index created on the parent table of partitions does not contain anydata. Rather, it dereferences to the local indexes when an index search is required. Thismeans that multiple indexes will have to be evaluated and data to be combined thereafter.However, with the global indexes, data will reside with global index declared on the parenttable. This avoids the need for multi-level index lookups. So read performance is expectedto be significantly higher in cases. There will however be a negative performance impactduring write (insert/update) of data. This is discussed in more detail later on.- Creating a GLOBAL Index - SyntaxA global index may be created with the addition of a “GLOBAL” keyword to the index statement.Alternatively, one could specify the “LOCAL” keyword to create local indexes on partitions.We are suggesting to call this set of keywords: “partition_index_type”. By default,partition_index_type will be set as LOCAL. Here is a sample of the create index syntax.CREATE Index idx parent (columns) [GLOBAL | LOCAL];Note: There is no shift/reduced by adding these options.- Pointing Index to TupleCurrently, CTID carries a page and offset information for a known heap (table name). However,in the context of global indexes, this information within an index is insufficient. Since the index isexpected to carry tuples from multiple partitions (heaps), CTID alone will not be able to link an indexnode to a tuple. This requires carrying additional data for the heap name to be stored with eachindex node.How this should be implemented is a point to be discussed. A few possibilities are listed below:-- Expand CTID to include a relfilenode id. In PostgreSQL-Conf Asia, Bruce suggested having the OIDinstead of relfilenode as relfilenode can be duplicated across tablespaces. -- Using OID introduces another complication where we would need to query catalog for OID to heap mapping.-- The second option is to have a variable-length CTID. We can reserve some top-level bit for segregation ofGlobal CTID or Standard CTID. Robert Haas suggested in PostgreSQL-EU to discuss this with Peter Geoghegan. -- I discussed it with Peter and he believes that it is a very invasive approach that requires a whole lot of the effort to get committed. -- Heikki pointed me to include heap specific information using the INCLUDE keyword so that heap informationis stored with each index node as data. -- We (Peter and I) also discussed that option and this looks a more easy and non-invasive option.- OptimizerThe challenge with optimizer is a selection between local and global indexes when both are present.How do we: -- Evaluate the cost of scanning a global index?-- When should the LOCAL index be preferred over the GLOBAL index and vice versa? -- Should we hit a GLOBAL index when the query is targeting a couple of partitions only? -- We need to consider the sizes of those partitions being hit and the sizes of partitions not being hit.-- Bruce suggested that we prioritize a GLOBAL index in the first version so that in every case,the GLOBAL index is utilized. - Write Performance and VacuumThere will be some write performance degradation because every change in partition tables mustpropagate upwards to the GLOBAL index on the parent table. This can be thought of as another indexon a table, however, the [slight] performance degradation will be due to the fact that the GLOBALindex may carry a much bigger dataset with data from multiple partitions resulting in a higher treetraversal and update time. This applies to both write and vacuum processes.It is still an open question though on how this will be handled within the code and how we can betteroptimize this process.I have a POC patch and working on finalizing the patch, Hamid Akhtar is also working with me on thiswork.--Ibrar Ahmed\",\n \"msg_date\": \"Wed, 30 Oct 2019 13:38:29 +0500\",\n \"msg_from\": \"Ibrar Ahmed <ibrar.ahmad@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"Ibrar Ahmed <ibrar.ahmad@gmail.com> writes:\\n> A global index by very definition is a single index on the parent table\\n> that maps to many\\n> underlying table partitions.\\n\\nI believe that the current design of partitioning is explicitly intended\\nto avoid the need for such a construct. It'd be absolutely disastrous\\nto have such a thing from many standpoints, including the breadth of\\nlocking needed to work with the global index, the difficulty of vacuuming,\\nand the impossibility of cheaply attaching or detaching partitions.\\n\\nIn other words, this is a \\\"feature\\\" we do not want.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 10:13:26 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 30, 2019 at 10:13 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> I believe that the current design of partitioning is explicitly intended\\n> to avoid the need for such a construct. It'd be absolutely disastrous\\n> to have such a thing from many standpoints, including the breadth of\\n> locking needed to work with the global index, the difficulty of vacuuming,\\n> and the impossibility of cheaply attaching or detaching partitions.\\n>\\n> In other words, this is a \\\"feature\\\" we do not want.\\n\\nI don't think that's true. Certainly, a lot of EnterpriseDB customers\\nwant this feature - it comes up regularly in discussions here. But\\nthat is not to say that the technical challenges are not formidable,\\nand I don't think this proposal really grapples with any of them.\\nHowever, that doesn't mean that the feature isn't desirable.\\n\\nOne of the biggest reasons why people want it is to enforce uniqueness\\nfor secondary keys - e.g. the employees table is partitioned by\\nemployee ID, but SSN should also be unique, at least among employees\\nfor whom it's not NULL.\\n\\nBut people also want it for faster data retrieval: if you're looking\\nfor a commonly-occurring value, an index per partition is fine. But if\\nyou're looking for values that occur only once or a few times across\\nthe whole hierarchy, an index scan per partition is very costly.\\nConsider, e.g.:\\n\\nNested Loop\\n-> Seq Scan\\n-> Append\\n -> Index Scan on each_partition\\n\\nYou don't have to have very many partitions for that to suck, and it's\\na thing that people want to do. Runtime partition pruning helps with\\nthis case a lot, but, once again, only for the primary key. Secondary\\nkeys are a big problem for partitioning today, in many ways.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 12:12:57 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Wed, Oct 30, 2019 at 10:13 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> I believe that the current design of partitioning is explicitly intended\\n>> to avoid the need for such a construct. It'd be absolutely disastrous\\n>> to have such a thing from many standpoints, including the breadth of\\n>> locking needed to work with the global index, the difficulty of vacuuming,\\n>> and the impossibility of cheaply attaching or detaching partitions.\\n>> In other words, this is a \\\"feature\\\" we do not want.\\n\\n> I don't think that's true. Certainly, a lot of EnterpriseDB customers\\n> want this feature - it comes up regularly in discussions here. But\\n> that is not to say that the technical challenges are not formidable,\\n> and I don't think this proposal really grapples with any of them.\\n> However, that doesn't mean that the feature isn't desirable.\\n\\nWell, the *effects* of the feature seem desirable, but that doesn't\\nmean that we want an implementation that actually has a shared index.\\nAs soon as you do that, you've thrown away most of the benefits of\\nhaving a partitioned data structure in the first place.\\n\\nNo, I don't have an idea how we might support, eg, uniqueness of\\nnon-partition-key columns without that. But we need to spend our\\neffort on figuring that out, not on building a complicated mechanism\\nwhose performance is never going to not suck.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 12:23:23 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 30, 2019 at 9:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Well, the *effects* of the feature seem desirable, but that doesn't\\n> mean that we want an implementation that actually has a shared index.\\n> As soon as you do that, you've thrown away most of the benefits of\\n> having a partitioned data structure in the first place.\\n\\nRight, but that's only the case for the global index. Global indexes\\nare useful when used judiciously. They enable the use of partitioning\\nfor use cases where not being able to enforce uniqueness across all\\npartitions happens to be a deal breaker. I bet that this is fairly\\ncommon.\\n\\n> No, I don't have an idea how we might support, eg, uniqueness of\\n> non-partition-key columns without that. But we need to spend our\\n> effort on figuring that out, not on building a complicated mechanism\\n> whose performance is never going to not suck.\\n\\nI don't think that there is a way to solve the problem that doesn't\\nlook very much like a global index. Also, being able to push down a\\npartition number when scanning a global index seems like it would be\\nvery compelling in some scenarios.\\n\\nI'm a bit worried about the complexity that will need to be added to\\nnbtree to make global indexes work, but it's probably possible to come\\nup with something that isn't too bad. GIN already uses an\\nimplementation level attribute number column for multi-column GIN\\nindexes, which is a little like what Ibrar has in mind. The really\\ncomplicated new code required for global indexes will be in places\\nlike vacuumlazy.c.\\n\\n\\n--\\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 09:48:37 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"Peter Geoghegan <pg@bowt.ie> writes:\\n> On Wed, Oct 30, 2019 at 9:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Well, the *effects* of the feature seem desirable, but that doesn't\\n>> mean that we want an implementation that actually has a shared index.\\n>> As soon as you do that, you've thrown away most of the benefits of\\n>> having a partitioned data structure in the first place.\\n\\n> Right, but that's only the case for the global index. Global indexes\\n> are useful when used judiciously.\\n\\nBut ... why bother with partitioning then? To me, the main reasons\\nwhy you might want a partitioned table are\\n\\n* ability to cheaply add and remove partitions, primarily so that\\nyou can cheaply do things like \\\"delete the oldest month's data\\\".\\n\\n* ability to scale past our limits on the physical size of one table\\n--- both the hard BlockNumber-based limit, and the performance\\nconstraints of e.g. vacuuming a very large table.\\n\\nBoth of those go out the window with a global index. So you might\\nas well just have one table and forget all the overhead.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 13:05:57 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-10-30 13:05:57 -0400, Tom Lane wrote:\\n> Peter Geoghegan <pg@bowt.ie> writes:\\n> > On Wed, Oct 30, 2019 at 9:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >> Well, the *effects* of the feature seem desirable, but that doesn't\\n> >> mean that we want an implementation that actually has a shared index.\\n> >> As soon as you do that, you've thrown away most of the benefits of\\n> >> having a partitioned data structure in the first place.\\n> \\n> > Right, but that's only the case for the global index. Global indexes\\n> > are useful when used judiciously.\\n> \\n> But ... why bother with partitioning then? To me, the main reasons\\n> why you might want a partitioned table are\\n\\nQuite commonly there's a lot of *other* indexes, often on a lot wider\\ndata than the primary key, that don't need to be global. And whereas in\\na lot of cases the primary key in a partitioned table has pretty good\\nlocality (and thus will be mostly buffered IO), other indexes will often\\nnot have that property (i.e. not have much correlation with table\\nposition).\\n\\n\\n> * ability to cheaply add and remove partitions, primarily so that\\n> you can cheaply do things like \\\"delete the oldest month's data\\\".\\n\\nYou can still do that to some degree with a global index. Imagine\\ne.g. keeping a 'partition id' as a sort-of column in the global\\nindex. That allows you to drop the partition, without having to\\nimmediately rebuild the index, by checking the partition id against the\\nlive partitions during lookup. So sure, your'e wasting space for a bit\\nin the global index, but it'll also be space that is likely to be fairly\\nefficiently reclaimed the next time vacuum touches the index. And if\\nnot the global index can be rebuilt concurrently without blocking\\nwrites.\\n\\n\\n> * ability to scale past our limits on the physical size of one table\\n> --- both the hard BlockNumber-based limit, and the performance\\n> constraints of e.g. vacuuming a very large table.\\n\\nFor that to be a problem for a global index the global index (which will\\noften be something like two int4 or int8 columns) itself would need to\\nbe above the block number based limit - which doesn't seem that close.\\n\\nWRT vacuuming - based on my observations the table itself isn't a\\nperformance problem for vacuuming all that commonly anymore, it's the\\nassociated index scans. So yea, that's a problem.\\n\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 10:27:02 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\n* Peter Geoghegan (pg@bowt.ie) wrote:\\n> On Wed, Oct 30, 2019 at 9:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > Well, the *effects* of the feature seem desirable, but that doesn't\\n> > mean that we want an implementation that actually has a shared index.\\n> > As soon as you do that, you've thrown away most of the benefits of\\n> > having a partitioned data structure in the first place.\\n> \\n> Right, but that's only the case for the global index. Global indexes\\n> are useful when used judiciously. They enable the use of partitioning\\n> for use cases where not being able to enforce uniqueness across all\\n> partitions happens to be a deal breaker. I bet that this is fairly\\n> common.\\n\\nAbsolutely- our lack of such is a common point of issue when folks are\\nconsidering using or migrating to PostgreSQL.\\n\\nThanks,\\n\\nStephen\",\n \"msg_date\": \"Thu, 31 Oct 2019 14:50:35 -0400\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Thu, 31 Oct 2019 at 14:50, Stephen Frost <sfrost@snowman.net> wrote:\\n\\n> Greetings,\\n>\\n> * Peter Geoghegan (pg@bowt.ie) wrote:\\n>\\n[....]\\n\\n>\\n> Absolutely- our lack of such is a common point of issue when folks are\\n> considering using or migrating to PostgreSQL.\\n>\\n\\nNot sure how similar my situation really is, but I find myself wanting to\\nhave indices that cross non-partition members of an inheritance hierarchy:\\n\\ncreate table t (\\n id int,\\n primary key (id)\\n);\\n\\ncreate table t1 (\\n a text\\n) inherits (t);\\n\\ncreate table t2 (\\n b int,\\n c int\\n) inherits (t);\\n\\nSo \\\"t\\\"s are identified by an integer; and one kind of \\\"t\\\" has a single text\\nattribute while a different kind of \\\"t\\\" has 2 int attributes. The idea is\\nthat there is a single primary key constraint on the whole hierarchy that\\nensures only one record with a particular id can exist in all the tables\\ntogether. I can imagine wanting to do this with other unique constraints\\nalso.\\n\\nAt present I don't actually use inheritance; instead I put triggers on the\\nchild tables that do an insert on the parent table, which has the effect of\\nenforcing the uniqueness I want.\\n\\nOn Thu, 31 Oct 2019 at 14:50, Stephen Frost <sfrost@snowman.net> wrote:Greetings,\\n\\n* Peter Geoghegan (pg@bowt.ie) wrote:\\n[....] \\n\\nAbsolutely- our lack of such is a common point of issue when folks are\\nconsidering using or migrating to PostgreSQL.Not sure how similar my situation really is, but I find myself wanting to have indices that cross non-partition members of an inheritance hierarchy:create table t ( id int, primary key (id));create table t1 ( a text) inherits (t);create table t2 ( b int, c int) inherits (t);So \\\"t\\\"s are identified by an integer; and one kind of \\\"t\\\" has a single text attribute while a different kind of \\\"t\\\" has 2 int attributes. The idea is that there is a single primary key constraint on the whole hierarchy that ensures only one record with a particular id can exist in all the tables together. I can imagine wanting to do this with other unique constraints also.At present I don't actually use inheritance; instead I put triggers on the child tables that do an insert on the parent table, which has the effect of enforcing the uniqueness I want.\",\n \"msg_date\": \"Thu, 31 Oct 2019 15:02:40 -0400\",\n \"msg_from\": \"Isaac Morland <isaac.morland@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 03:02:40PM -0400, Isaac Morland wrote:\\n>On Thu, 31 Oct 2019 at 14:50, Stephen Frost <sfrost@snowman.net> wrote:\\n>\\n>> Greetings,\\n>>\\n>> * Peter Geoghegan (pg@bowt.ie) wrote:\\n>>\\n>[....]\\n>\\n>>\\n>> Absolutely- our lack of such is a common point of issue when folks are\\n>> considering using or migrating to PostgreSQL.\\n>>\\n>\\n>Not sure how similar my situation really is, but I find myself wanting to\\n>have indices that cross non-partition members of an inheritance hierarchy:\\n>\\n>create table t (\\n> id int,\\n> primary key (id)\\n>);\\n>\\n>create table t1 (\\n> a text\\n>) inherits (t);\\n>\\n>create table t2 (\\n> b int,\\n> c int\\n>) inherits (t);\\n>\\n>So \\\"t\\\"s are identified by an integer; and one kind of \\\"t\\\" has a single text\\n>attribute while a different kind of \\\"t\\\" has 2 int attributes. The idea is\\n>that there is a single primary key constraint on the whole hierarchy that\\n>ensures only one record with a particular id can exist in all the tables\\n>together. I can imagine wanting to do this with other unique constraints\\n>also.\\n>\\n\\nIMO the chances of us supporting global indexes with generic inheritance\\nhierarchies are about zero. We don't even support creating \\\"partition\\\"\\nindexes on those hierarchies ...\\n\\n>At present I don't actually use inheritance; instead I put triggers on the\\n>child tables that do an insert on the parent table, which has the effect of\\n>enforcing the uniqueness I want.\\n\\nDoes it? Are you sure it actually works in READ COMMITTED? What exactly\\ndoes the trigger do?\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 20:21:55 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On 10/30/19 10:27, Andres Freund wrote:\\n> On 2019-10-30 13:05:57 -0400, Tom Lane wrote:\\n>> Peter Geoghegan <pg@bowt.ie> writes:\\n>>> On Wed, Oct 30, 2019 at 9:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>>> Well, the *effects* of the feature seem desirable, but that doesn't\\n>>>> mean that we want an implementation that actually has a shared index.\\n>>>> As soon as you do that, you've thrown away most of the benefits of\\n>>>> having a partitioned data structure in the first place.\\n>>\\n>>> Right, but that's only the case for the global index. Global indexes\\n>>> are useful when used judiciously.\\n>>\\n>> But ... why bother with partitioning then? To me, the main reasons\\n>> why you might want a partitioned table are\\n> \\n> Quite commonly there's a lot of *other* indexes, often on a lot wider\\n> data than the primary key, that don't need to be global. And whereas in\\n> a lot of cases the primary key in a partitioned table has pretty good\\n> locality (and thus will be mostly buffered IO), other indexes will often\\n> not have that property (i.e. not have much correlation with table\\n> position).\\n\\nI asked around a little bit and got some interesting responses. Thought\\nI'd pass two of them along.\\n\\nOne person worked on a payments network (150,000+ installed readers),\\nthe transaction table was date partitioned (1 per day) based on insert\\ntimestamp, but lookups and updates were typically by the unique\\ntransaction id. Oracle DB, they kept 180 daily partitions, several\\nmillion rows per day. Transactions did not arrive in order, and could be\\ndelayed if some part of the network was slow (they opted to allow the $2\\ncharge rather than reject sales) and when the cash transaction records\\nwere uploaded. Step one for their PG conversion created a read replica\\nin PG 9.6, and the cost of doing the individual index lookups across 180\\npartitions (and 180 indexes) was very high, so they stored max and min\\ntxn id per partition and would generate a query with all the dates that\\na txn id could have been in so that only a small number of partition\\nindexes would be accessed. They wanted a global index on txn id for\\nperformance, not for uniqueness – id generated on reader with guid-like\\nsemantics.\\n\\nA second person worked on several large-scale systems and he relayed\\nthat in some cases where they used Oracle global indexes on partitioned\\ntables, they ended up deciding to reverse that decision as things scaled\\nbecause of restrictive locking during partition maintenance (this is the\\nexact issue Tom points out). So even on a database _with_ the option of\\nusing a global index, they've sometimes opted for \\\"workaround\\\" design\\npatterns instead:\\n* To solve uniqueness, manage serialization at the appliation level.\\nIsolate operations (e.g. using a queue) and use that to make sure that\\ntwo sessions don’t try to insert the same record at the same time. From\\nan RDBMS, this looks like a separate, smaller table that is being used\\nto manage work activity.\\n* To solve the additional IO for a global table scan ... We often don’t\\nneed to do this because the load in this pattern is not typically highly\\nconcurrent. If we are looking for higher concurrency, we can usually\\nadd a hack/workaround that filters on a partition key to provide “pretty\\ngood” pruning. The net result is that you get 2-3x the IO due to the\\nlack of global index (same workaround as first story above).\\n\\nQuote: \\\"So ... I don’t actually like the idea of introducing this.\\nUnless, someone can solve the ugly challenges we have had [around\\npartition maintenance operations].\\\"\\n\\nI actually don't think those challenges are so un-solvable. I think that\\nglobal indexes will be irrelevant to most workloads. I'm not entirely\\nconvinced that they won't be useful for a few people with specific\\nworkloads and large amounts of data in PostgreSQL where the benefits\\noutweigh the costs. I definitely agree that care needs to be taken\\naround index maintenance operations if there's an effort here.\\n\\n\\n>> * ability to cheaply add and remove partitions, primarily so that\\n>> you can cheaply do things like \\\"delete the oldest month's data\\\".\\n> \\n> You can still do that to some degree with a global index. Imagine\\n> e.g. keeping a 'partition id' as a sort-of column in the global\\n> index. That allows you to drop the partition, without having to\\n> immediately rebuild the index, by checking the partition id against the\\n> live partitions during lookup. So sure, your'e wasting space for a bit\\n> in the global index, but it'll also be space that is likely to be fairly\\n> efficiently reclaimed the next time vacuum touches the index. And if\\n> not the global index can be rebuilt concurrently without blocking\\n> writes.\\n\\nAnother idea might be to leverage PostgreSQL's partial indexes. If the\\nindex is created \\\"where date>2020\\\" and you're dropping an index from\\n2019 then you can entirely ignore the index. Not a panacea for every\\nindex maintenance operation, but for the super-common case of dropping\\nthe oldest partition you can now:\\n\\n1) create new index concurrently \\\"where dt>2020\\\"\\n2) drop the old index\\n3) drop the 2019 partition\\n\\ndoesn't solve world hunger but there's lots of benefit for such a simple\\nhack.\\n\\n\\n>> * ability to scale past our limits on the physical size of one table\\n>> --- both the hard BlockNumber-based limit, and the performance\\n>> constraints of e.g. vacuuming a very large table.\\n> \\n> For that to be a problem for a global index the global index (which will\\n> often be something like two int4 or int8 columns) itself would need to\\n> be above the block number based limit - which doesn't seem that close.\\n> \\n> WRT vacuuming - based on my observations the table itself isn't a\\n> performance problem for vacuuming all that commonly anymore, it's the\\n> associated index scans. So yea, that's a problem.\\n\\nI'm sure zheap will make all our dreams come true, right? :D\\n\\n-Jeremy\\n\\n\\n-- \\nJeremy Schneider\\nDatabase Engineer\\nAmazon Web Services\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Nov 2019 15:05:03 -0800\",\n \"msg_from\": \"Jeremy Schneider <schnjere@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On 11/25/19 15:05, Jeremy Schneider wrote:\\n> ... the cost of doing the individual index lookups across 180\\n> partitions (and 180 indexes) was very high, so they stored max and min\\n> txn id per partition and would generate a query with all the dates that\\n> a txn id could have been in so that only a small number of partition\\n> indexes would be accessed. \\n> \\n> .. If we are looking for higher concurrency, we can usually\\n> add a hack/workaround that filters on a partition key to provide “pretty\\n> good” pruning. The net result is that you get 2-3x the IO due to the\\n> lack of global index (same workaround as first story above).\\n\\nIs that basically like a global BRIN index with granularity at the\\npartition level?\\n\\n-J\\n\\n-- \\nJeremy Schneider\\nDatabase Engineer\\nAmazon Web Services\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Nov 2019 15:44:39 -0800\",\n \"msg_from\": \"Jeremy Schneider <schnjere@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 25, 2019 at 03:44:39PM -0800, Jeremy Schneider wrote:\\n> On 11/25/19 15:05, Jeremy Schneider wrote:\\n> > ... the cost of doing the individual index lookups across 180\\n> > partitions (and 180 indexes) was very high, so they stored max and min\\n> > txn id per partition and would generate a query with all the dates that\\n> > a txn id could have been in so that only a small number of partition\\n> > indexes would be accessed. \\n> > \\n> > .. If we are looking for higher concurrency, we can usually\\n> > add a hack/workaround that filters on a partition key to provide “pretty\\n> > good” pruning. The net result is that you get 2-3x the IO due to the\\n> > lack of global index (same workaround as first story above).\\n> \\n> Is that basically like a global BRIN index with granularity at the\\n> partition level?\\n\\nExactly! :-)\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Wed, 18 Dec 2019 22:03:28 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On 19/12/19 4:03, Bruce Momjian wrote:\\n> On Mon, Nov 25, 2019 at 03:44:39PM -0800, Jeremy Schneider wrote:\\n>> On 11/25/19 15:05, Jeremy Schneider wrote:\\n>>> ... the cost of doing the individual index lookups across 180\\n>>> partitions (and 180 indexes) was very high, so they stored max and min\\n>>> txn id per partition and would generate a query with all the dates that\\n>>> a txn id could have been in so that only a small number of partition\\n>>> indexes would be accessed.\\n>>>\\n>>> .. If we are looking for higher concurrency, we can usually\\n>>> add a hack/workaround that filters on a partition key to provide “pretty\\n>>> good” pruning. The net result is that you get 2-3x the IO due to the\\n>>> lack of global index (same workaround as first story above).\\n>> Is that basically like a global BRIN index with granularity at the\\n>> partition level?\\n> Exactly! :-)\\n\\nActually, one \\\"kind of\\\" BRIN index *per partitioned table* mapping (key \\nrange) -> (partition oid)... and so concurrency doesn't need to be very \\naffected.\\n\\n(we don't need to do things just like other RDBMS do, ya know... ;)\\n\\n\\nIIRC, this precise approach was suggested around 2016 when initially \\ndiscussing the \\\"declarative partitioning\\\" which originated Postgres' \\ncurrent partitioning scheme, in order to optimize partition pruning.\\n\\n\\nJust my .02€\\n\\n / J.L.\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 19 Dec 2019 09:48:40 +0100\",\n \"msg_from\": \"Jose Luis Tallon <jltallon@adv-solutions.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 19, 2019 at 09:48:40AM +0100, Jose Luis Tallon wrote:\\n> On 19/12/19 4:03, Bruce Momjian wrote:\\n> > On Mon, Nov 25, 2019 at 03:44:39PM -0800, Jeremy Schneider wrote:\\n> > > On 11/25/19 15:05, Jeremy Schneider wrote:\\n> > > > ... the cost of doing the individual index lookups across 180\\n> > > > partitions (and 180 indexes) was very high, so they stored max and min\\n> > > > txn id per partition and would generate a query with all the dates that\\n> > > > a txn id could have been in so that only a small number of partition\\n> > > > indexes would be accessed.\\n> > > > \\n> > > > .. If we are looking for higher concurrency, we can usually\\n> > > > add a hack/workaround that filters on a partition key to provide “pretty\\n> > > > good” pruning. The net result is that you get 2-3x the IO due to the\\n> > > > lack of global index (same workaround as first story above).\\n> > > Is that basically like a global BRIN index with granularity at the\\n> > > partition level?\\n> > Exactly! :-)\\n> \\n> Actually, one \\\"kind of\\\" BRIN index *per partitioned table* mapping (key\\n> range) -> (partition oid)... and so concurrency doesn't need to be very\\n> affected.\\n> \\n> (we don't need to do things just like other RDBMS do, ya know... ;)\\n> \\n> \\n> IIRC, this precise approach was suggested around 2016 when initially\\n> discussing the \\\"declarative partitioning\\\" which originated Postgres' current\\n> partitioning scheme, in order to optimize partition pruning.\\n\\nRobert Haas identified two needs for global indexes:\\n\\n\\thttps://www.postgresql.org/message-id/CA+Tgmob_J2M2+QKWrhg2NjQEkMEwZNTfd7a6Ubg34fJuZPkN2g@mail.gmail.com\\n\\t\\n\\tOne of the biggest reasons why people want it is to enforce uniqueness\\n\\tfor secondary keys - e.g. the employees table is partitioned by\\n\\temployee ID, but SSN should also be unique, at least among employees\\n\\tfor whom it's not NULL.\\n\\t\\n\\tBut people also want it for faster data retrieval: if you're looking\\n\\tfor a commonly-occurring value, an index per partition is fine. But if\\n\\tyou're looking for values that occur only once or a few times across\\n\\tthe whole hierarchy, an index scan per partition is very costly.\\n\\nI don't see lossy BRIN indexes helping with the uniqueness use-case, and\\nI am not sure they would help with the rare case either. They would\\nhelp for range-based partitions, but I thought our existing facilities\\nworked in that case.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Thu, 19 Dec 2019 11:12:07 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On 12/19/19 08:12, Bruce Momjian wrote:\\n> I don't see lossy BRIN indexes helping with the uniqueness use-case, and\\n> I am not sure they would help with the rare case either. They would\\n> help for range-based partitions, but I thought our existing facilities\\n> worked in that case.\\n\\nCorrelated data. The existing facilities work if the filtering column\\nis exactly the same as the partition column. But it's not at all\\nuncommon to have other columns with correlated data, perhaps the most\\nobvious of which is timeseries tables with many date columns of various\\ndefinitions (row first update, row latest update, invoice date, payment\\ndate, process date, ship date, etc).\\n\\nWhat if you could use *two* indexes in a single execution plan? Use the\\nglobal BRIN to narrow down to 2 or 3 out of a hundred or more\\npartitions, then use local indexes to find specific rows in the\\npartitions of interest? That might work, without being too overly\\ncomplicated.\\n\\n-J\\n\\n\\n-- \\nJeremy Schneider\\nDatabase Engineer\\nAmazon Web Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 19 Dec 2019 11:28:55 -0800\",\n \"msg_from\": \"Jeremy Schneider <schnjere@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 19, 2019 at 11:28:55AM -0800, Jeremy Schneider wrote:\\n> On 12/19/19 08:12, Bruce Momjian wrote:\\n> > I don't see lossy BRIN indexes helping with the uniqueness use-case, and\\n> > I am not sure they would help with the rare case either. They would\\n> > help for range-based partitions, but I thought our existing facilities\\n> > worked in that case.\\n> \\n> Correlated data. The existing facilities work if the filtering column\\n> is exactly the same as the partition column. But it's not at all\\n> uncommon to have other columns with correlated data, perhaps the most\\n> obvious of which is timeseries tables with many date columns of various\\n> definitions (row first update, row latest update, invoice date, payment\\n> date, process date, ship date, etc).\\n> \\n> What if you could use *two* indexes in a single execution plan? Use the\\n> global BRIN to narrow down to 2 or 3 out of a hundred or more\\n> partitions, then use local indexes to find specific rows in the\\n> partitions of interest? That might work, without being too overly\\n> complicated.\\n\\nNo, that is very interesting --- having secondary indexes for\\npartitioned tables that trim most partitions. Would index lookups on\\neach partition index be very slow? BRIN indexes? I am assuming global\\nindexes would only avoid such lookups.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Thu, 19 Dec 2019 21:34:33 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"\\n> On Oct 30, 2019, at 12:05 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> But ... why bother with partitioning then? To me, the main reasons\\n> why you might want a partitioned table are\\n> \\n> * ability to cheaply add and remove partitions, primarily so that\\n> you can cheaply do things like \\\"delete the oldest month's data\\\".\\n> \\n> * ability to scale past our limits on the physical size of one table\\n> --- both the hard BlockNumber-based limit, and the performance\\n> constraints of e.g. vacuuming a very large table.\\n\\nA third case is data locality. In that case global indexes would be useful for queries that do not correlate will with hot data.\\n\\n> Both of those go out the window with a global index. So you might\\n> as well just have one table and forget all the overhead.\\n\\nPartition pruning could still be valuable even with global indexes, provided that we teach vacuum how to clean up tuples in an index that point at a partition that has been deleted.\\n\\n\",\n \"msg_date\": \"Tue, 7 Jan 2020 22:55:13 +0000\",\n \"msg_from\": \"\\\"Nasby, Jim\\\" <nasbyj@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"I've been following this topic for a long time. It's been a year since the last response.\\nIt was clear that our customers wanted this feature as well, and a large number of them mentioned it.\\n\\nSo, I wish the whole feature to mature as soon as possible.\\nI summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\\n\\nNext, I encountered some difficulties when implementing the DDL of the partition table with global index, and I hope to get some help from the community\\n\\nHere are some details what has been implemented\\n1 Definition of global index\\nUsing the INCLUDE keyword to include the tableoid of the partitioned table.\\n\\n2. Maintenance of global index by partition table DML.\\nBoth INSERT and UPDATE of a partitioned table maintain global index\\n\\n3. Global index scan\\nPlanner: Processes predicate conditions on the primary partition, generating paths and plans for the global index.\\nExecuter: index scan get indextup, get the tableoid from indextup, and verify the visibility of the data in the partition.\\n\\n4. Vacuum partition table maintains global index\\nEach partitioned table VACUUM cleans its own garbage data in the global index.\\n\\nAfter the above function point is completed, the global index can be used without partition table DDL.\\n\\nDemo:\\n--Use pgbench to create the test partition table\\npgbench -i -s 1000 --partitions=6 --partition-method=range\\n\\n—- create global index on bid, bid is not partition key\\nCREATE INDEX idx_pgbench_accounts_bid on pgbench_accounts(bid) global;\\n\\n— check global index status\\nselect * , sum(alivetup) over()as sum_alivetup, sum(deadtup) over() as sum_deadtup from bt_get_global_index_status('idx_pgbench_accounts_bid');\\n relname | alivetup | deadtup | sum_alivetup | sum_deadtup \\n--------------------+----------+---------+--------------+-------------\\n pgbench_accounts_1 | 16666667 | 0 | 100000000 | 0\\n pgbench_accounts_2 | 16666667 | 0 | 100000000 | 0\\n pgbench_accounts_3 | 16666667 | 0 | 100000000 | 0\\n pgbench_accounts_4 | 16666667 | 0 | 100000000 | 0\\n pgbench_accounts_5 | 16666667 | 0 | 100000000 | 0\\n pgbench_accounts_6 | 16666665 | 0 | 100000000 | 0\\n(6 rows)\\n\\n— run pgbench for for a while\\npgbench -M prepared -j 32 -c 32 -T 60 -P1\\n\\n\\n—- check global index, The index has bloated\\npostgres=# select * , sum(alivetup) over()as sum_alivetup, sum(deadtup) over() as sum_deadtup from bt_get_global_index_status('idx_pgbench_accounts_bid');\\n relname | alivetup | deadtup | sum_alivetup | sum_deadtup \\n--------------------+----------+---------+--------------+-------------\\n pgbench_accounts_1 | 16717733 | 0 | 100306102 | 0\\n pgbench_accounts_2 | 16717409 | 0 | 100306102 | 0\\n pgbench_accounts_3 | 16717540 | 0 | 100306102 | 0\\n pgbench_accounts_4 | 16717972 | 0 | 100306102 | 0\\n pgbench_accounts_5 | 16717578 | 0 | 100306102 | 0\\n pgbench_accounts_6 | 16717870 | 0 | 100306102 | 0\\n(6 rows)\\n\\n—- vacuum partition table\\nvacuum pgbench_accounts;\\n\\n—- Garbage is collected, global index still looks correct and valid.\\npostgres=# select * , sum(alivetup) over()as sum_alivetup, sum(deadtup) over() as sum_deadtup from bt_get_global_index_status('idx_pgbench_accounts_bid');\\n relname | alivetup | deadtup | sum_alivetup | sum_deadtup \\n--------------------+----------+---------+--------------+-------------\\n pgbench_accounts_1 | 16666667 | 0 | 100000000 | 0\\n pgbench_accounts_2 | 16666667 | 0 | 100000000 | 0\\n pgbench_accounts_3 | 16666667 | 0 | 100000000 | 0\\n pgbench_accounts_4 | 16666667 | 0 | 100000000 | 0\\n pgbench_accounts_5 | 16666667 | 0 | 100000000 | 0\\n pgbench_accounts_6 | 16666665 | 0 | 100000000 | 0\\n(6 rows)\\n\\n—-\\n\\n—- global index scan works well\\npostgres=# select tableoid ,count(*) from pgbench_accounts where bid = 834 group by tableoid;\\n tableoid | count \\n----------+-------\\n 16455 | 33335\\n 16458 | 66665\\n(2 rows)\\n\\npostgres=# explain select tableoid ,count(*) from pgbench_accounts where bid = 834 group by tableoid;\\n QUERY PLAN \\n--------------------------------------------------------------------------------------------------------------------\\n HashAggregate (cost=2945.23..2945.24 rows=1 width=12)\\n Group Key: pgbench_accounts.tableoid\\n -> Global Index Scan using idx_pgbench_accounts_bid on pgbench_accounts (cost=0.50..10.18 rows=587011 width=4)\\n Index Cond: (bid = 834)\\n(4 rows)\\n\\n\\nThe following is how to implement DDL of global index. How to maintain global index of DDL of partitioned table.\\nThis seems to be more difficult than the previous work.\\n\\nI understand there are four main parts\\n\\n1 Build global index or reindex, especially in concurrent mode\\n\\n2 Detach partition\\nWould it be a good idea to make a flag to global index and let VACUUM handle the index data of the Detach partition?\\n\\n3 Attach partition\\nIt is easy to Attach a new empty partition, but adding a new one with data is not.\\nIf there is a unique key conflict, do we slowly clean up the garbage or invalidate the entire index?\\n\\n4 Truncate partition with global index\\nDo we need to process the heap and index data separately in multiple transactions?\\nThis will lose the ability to roll back for Truncate operation.\\nIs it worth it?\\n\\n\\nLooking forward to your feedback.\\n\\nThanks!\\n\\nWenjing\",\n \"msg_date\": \"Thu, 7 Jan 2021 17:44:01 +0800\",\n \"msg_from\": \"=?utf-8?B?5pu+5paH5peM?= <wenjing.zwj@alibaba-inc.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 7, 2021 at 05:44:01PM +0800, 曾文旌 wrote:\\n> I've been following this topic for a long time. It's been a year since the last response.\\n> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\\n> \\n> So, I wish the whole feature to mature as soon as possible.\\n> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\\n\\nI think you need to address the items mentioned in this blog, and the\\nemail link it mentions:\\n\\n\\thttps://momjian.us/main/blogs/pgblog/2020.html#July_1_2020\\n\\nI am not clear this is a feature we will want. Yes, people ask for it,\\nbut if the experience will be bad for them and they will regret using\\nit, I am not sure we want it. Of course, if you code it up and we get\\na good user experience, we would want it --- I am just saying it is not\\nclear right now.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> https://momjian.us\\n EnterpriseDB https://enterprisedb.com\\n\\n The usefulness of a cup is in its emptiness, Bruce Lee\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Jan 2021 09:16:20 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 7, 2021 at 4:44 AM 曾文旌 <wenjing.zwj@alibaba-inc.com> wrote:\\n> I've been following this topic for a long time. It's been a year since the last response.\\n> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\\n>\\n> So, I wish the whole feature to mature as soon as possible.\\n> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\\n\\nYou need to summarize the basic design choices you've made here. Like,\\nwhat did you do about the fact that TIDs have to be 6 bytes?\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Jan 2021 10:04:54 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"> 2021年1月7日 22:16，Bruce Momjian <bruce@momjian.us> 写道：\\n> \\n> On Thu, Jan 7, 2021 at 05:44:01PM +0800, 曾文旌 wrote:\\n>> I've been following this topic for a long time. It's been a year since the last response.\\n>> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\\n>> \\n>> So, I wish the whole feature to mature as soon as possible.\\n>> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\\n> \\n> I think you need to address the items mentioned in this blog, and the\\n> email link it mentions:\\n> \\n> \\thttps://momjian.us/main/blogs/pgblog/2020.html#July_1_2020\\n\\nThank you for your reply.\\nI read your blog and it helped me a lot.\\n\\nThe blog mentions a specific problem: \\\"A large global index might also reintroduce problems that prompted the creation of partitioning in the first place. \\\"\\nI don't quite understand, could you give some specific information?\\n\\nIn addition you mentioned: \\\"It is still unclear if these use-cases justify the architectural changes needed to enable global indexes.\\\"\\nPlease also describe the problems you see, I will confirm each specific issue one by one.\\n\\n\\nThanks\\n\\nWenjing\\n\\n\\n> \\n> I am not clear this is a feature we will want. Yes, people ask for it,\\n> but if the experience will be bad for them and they will regret using\\n> it, I am not sure we want it. Of course, if you code it up and we get\\n> a good user experience, we would want it --- I am just saying it is not\\n> clear right now.\\n> \\n> -- \\n> Bruce Momjian <bruce@momjian.us> https://momjian.us\\n> EnterpriseDB https://enterprisedb.com\\n> \\n> The usefulness of a cup is in its emptiness, Bruce Lee\",\n \"msg_date\": \"Fri, 8 Jan 2021 11:26:48 +0800\",\n \"msg_from\": \"=?utf-8?B?5pu+5paH5peM?= <wenjing.zwj@alibaba-inc.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 8, 2021 at 4:02 PM 曾文旌 <wenjing.zwj@alibaba-inc.com> wrote:\\n>\\n> > 2021年1月7日 22:16，Bruce Momjian <bruce@momjian.us> 写道：\\n> >\\n> > On Thu, Jan 7, 2021 at 05:44:01PM +0800, 曾文旌 wrote:\\n> >> I've been following this topic for a long time. It's been a year since the last response.\\n> >> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\\n> >>\\n> >> So, I wish the whole feature to mature as soon as possible.\\n> >> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\\n> >\\n> > I think you need to address the items mentioned in this blog, and the\\n> > email link it mentions:\\n> >\\n> > https://momjian.us/main/blogs/pgblog/2020.html#July_1_2020\\n>\\n> Thank you for your reply.\\n> I read your blog and it helped me a lot.\\n>\\n> The blog mentions a specific problem: \\\"A large global index might also reintroduce problems that prompted the creation of partitioning in the first place. \\\"\\n> I don't quite understand, could you give some specific information?\\n>\\n> In addition you mentioned: \\\"It is still unclear if these use-cases justify the architectural changes needed to enable global indexes.\\\"\\n> Please also describe the problems you see, I will confirm each specific issue one by one.\\n\\nOne example is date partitioning. People frequently need to store\\nonly the most recent data. For instance doing a monthly partitioning\\nand dropping the oldest partition every month once you hit the wanted\\nretention is very efficient for that use case, as it should be almost\\ninstant (provided that you can acquire the necessary locks\\nimmediately). But if you have a global index, you basically lose the\\nadvantage of partitioning as it'll require heavy changes on that\\nindex.\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Jan 2021 16:26:56 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 8, 2021 at 11:26:48AM +0800, 曾文旌 wrote:\\n> > On Thu, Jan 7, 2021 at 05:44:01PM +0800, 曾文旌 wrote:\\n> >> I've been following this topic for a long time. It's been a year since the last response.\\n> >> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\\n> >> \\n> >> So, I wish the whole feature to mature as soon as possible.\\n> >> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\\n> > \\n> > I think you need to address the items mentioned in this blog, and the\\n> > email link it mentions:\\n> > \\n> > \\thttps://momjian.us/main/blogs/pgblog/2020.html#July_1_2020\\n> \\n> Thank you for your reply.\\n> I read your blog and it helped me a lot.\\n> \\n> The blog mentions a specific problem: \\\"A large global index might also reintroduce problems that prompted the creation of partitioning in the first place. \\\"\\n> I don't quite understand, could you give some specific information?\\n\\nWell, if you created partitions, you probably did so because:\\n\\n1. heap files are smaller, allowing for more targeted sequential scans\\n2. smaller indexes\\n3. the ability to easily drop tables/indexes that are too old\\n\\nIf you have global indexes, #1 is the same, but #2 is not longer true,\\nand for #3, you can drop the heap but the index entries still exist in\\nthe global index and must be removed.\\n\\nSo, if you created partitions for one of the three reasons, once you\\nhave global indexes, some of those advantage of partitioning are no\\nlonger true. I am sure there are some workloads where the advantages of\\npartitioning, minus the advantages lost when using global indexes, are\\nuseful, but are there enough of them to make the feature useful? I\\ndon't know.\\n\\n> In addition you mentioned: \\\"It is still unclear if these use-cases justify the architectural changes needed to enable global indexes.\\\"\\n> Please also describe the problems you see, I will confirm each specific issue one by one.\\n\\nWell, the email thread I linked to has a lot of them, but the\\nfundamental issue is that you have to break the logic that a single\\nindex serves a single heap file. Considering what I said above, is\\nthere enough usefulness to warrant such an architectural change?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> https://momjian.us\\n EnterpriseDB https://enterprisedb.com\\n\\n The usefulness of a cup is in its emptiness, Bruce Lee\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Jan 2021 10:50:40 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"> 2021年1月8日 16:26，Julien Rouhaud <rjuju123@gmail.com> 写道：\\n> \\n> On Fri, Jan 8, 2021 at 4:02 PM 曾文旌 <wenjing.zwj@alibaba-inc.com> wrote:\\n>> \\n>>> 2021年1月7日 22:16，Bruce Momjian <bruce@momjian.us> 写道：\\n>>> \\n>>> On Thu, Jan 7, 2021 at 05:44:01PM +0800, 曾文旌 wrote:\\n>>>> I've been following this topic for a long time. It's been a year since the last response.\\n>>>> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\\n>>>> \\n>>>> So, I wish the whole feature to mature as soon as possible.\\n>>>> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\\n>>> \\n>>> I think you need to address the items mentioned in this blog, and the\\n>>> email link it mentions:\\n>>> \\n>>> https://momjian.us/main/blogs/pgblog/2020.html#July_1_2020\\n>> \\n>> Thank you for your reply.\\n>> I read your blog and it helped me a lot.\\n>> \\n>> The blog mentions a specific problem: \\\"A large global index might also reintroduce problems that prompted the creation of partitioning in the first place. \\\"\\n>> I don't quite understand, could you give some specific information?\\n>> \\n>> In addition you mentioned: \\\"It is still unclear if these use-cases justify the architectural changes needed to enable global indexes.\\\"\\n>> Please also describe the problems you see, I will confirm each specific issue one by one.\\n> \\n> One example is date partitioning. People frequently need to store\\n> only the most recent data. For instance doing a monthly partitioning\\n> and dropping the oldest partition every month once you hit the wanted\\n> retention is very efficient for that use case, as it should be almost\\n> instant (provided that you can acquire the necessary locks\\n> immediately). But if you have a global index, you basically lose the\\n> advantage of partitioning as it'll require heavy changes on that\\n> index.\\nIf the global index removes all the major benefits of partitioned tables, then it is not worth having it.\\n\\nThis is indeed a typical scenario for a partitioned table.\\nthere are two basic operations\\n1) Monthly DETACH old child table\\n2) Monthly ATTACH new child table\\n\\nFor 1) The DETACH old child table can be finished immediately, global index can be kept valid after DETACH is completed, and the cleanup of garbage data in global index can be deferred to VACUUM.\\nThis is similar to the global index optimization done by Oracle12c.\\nFor 2) ATTACH new empty child table can also be completed immediately.\\nIf this is the case, many of the advantages of partitioned tables will be retained, while the advantages of global indexes will be gained.\",\n \"msg_date\": \"Mon, 11 Jan 2021 19:40:18 +0800\",\n \"msg_from\": \"=?utf-8?B?5pu+5paH5peM?= <wenjing.zwj@alibaba-inc.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"> 2021年1月7日 23:04，Robert Haas <robertmhaas@gmail.com> 写道：\\n> \\n> On Thu, Jan 7, 2021 at 4:44 AM 曾文旌 <wenjing.zwj@alibaba-inc.com> wrote:\\n>> I've been following this topic for a long time. It's been a year since the last response.\\n>> It was clear that our customers wanted this feature as well, and a large number of them mentioned it.\\n>> \\n>> So, I wish the whole feature to mature as soon as possible.\\n>> I summarized the scheme mentioned in the email and completed the POC patch(base on PG_13).\\n> \\n> You need to summarize the basic design choices you've made here. Like,\\n> what did you do about the fact that TIDs have to be 6 bytes?\\n\\nThese are the basic choices, and most of them come from discussions in previous emails.\\n\\n1 Definition of global index\\nObviously, we need to expand Index address info(CTID) to include child table info in GlobalIndexTuple.\\n\\n1.1 As mentioned in the previous email, Bruce suggested having the OID\\ninstead of relfilenode as relfilenode can be duplicated across tablespaces. \\nI agree with that.\\n\\n1.2 And Heikki pointed me to include heap specific information using the INCLUDE keyword so that heap information\\nis stored with each index node as data.\\n\\nSo ,In POC stage, I choose use INCLUDE keyword to INCLUDE the tableoid of global index. This will add 4 bytes to each IndexTuple.\\n\\nConsidering that if a single partitioned table does not exceed 65535 child tables, perhaps two bytes for tracking which child table the data belongs to is sufficient.\\n\\n2. Maintenance of global index by partition table DML.\\nThe DML of each child table of the partitioned table needs to maintain the global index on the partitioned table.\\n\\n3. Global index scan\\nPlanner: \\nProcesses predicate on the primary partition, generating paths and plans for the global index.\\nThe cost model of the global index may need to be considered. We need to make the global index or the local index selected in their respective advantageous scenarios.\\n\\nExecuter: \\nThe index scan get indextup, get the tableoid from indextup, and verify the visibility of the data in the child table.\\nIf a child table is DETACH, then the index item of this table is ignored during the index scan until VACUUM finishes cleaning up the global index.\\n\\n4. Vacuum partition table maintains global index.\\nOld data in the global index also needs to be cleaned up, and vaccum is suitable for it.\\nEach child table in VACUUM, while vacuuming its own index, also vacuums the global index on the partitioned table.\\n\\n5. Other\\nThe global index indexes all of the child tables, which makes the global index large and has many levels. \\nFollow the technical route, The partitioned indexes are a further target.\\n\\nThis is my basic idea for implementing global index.\\nLooking forward to your feedback.\\n\\nThanks!\\n\\nWenjing\\n\\n> \\n> -- \\n> Robert Haas\\n> EDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Mon, 11 Jan 2021 21:00:45 +0800\",\n \"msg_from\": \"=?utf-8?B?5pu+5paH5peM?= <wenjing.zwj@alibaba-inc.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 11, 2021 at 07:40:18PM +0800, 曾文旌 wrote:\\n> >> In addition you mentioned: \\\"It is still unclear if these use-cases justify the architectural changes needed to enable global indexes.\\\"\\n> >> Please also describe the problems you see, I will confirm each specific issue one by one.\\n> > \\n> > One example is date partitioning. People frequently need to store\\n> > only the most recent data. For instance doing a monthly partitioning\\n> > and dropping the oldest partition every month once you hit the wanted\\n> > retention is very efficient for that use case, as it should be almost\\n> > instant (provided that you can acquire the necessary locks\\n> > immediately). But if you have a global index, you basically lose the\\n> > advantage of partitioning as it'll require heavy changes on that\\n> > index.\\n> If the global index removes all the major benefits of partitioned tables, then it is not worth having it.\\n> \\n> This is indeed a typical scenario for a partitioned table.\\n> there are two basic operations\\n> 1) Monthly DETACH old child table\\n> 2) Monthly ATTACH new child table\\n> \\n> For 1) The DETACH old child table can be finished immediately, global index can be kept valid after DETACH is completed, and the cleanup of garbage data in global index can be deferred to VACUUM.\\n> This is similar to the global index optimization done by Oracle12c.\\n> For 2) ATTACH new empty child table can also be completed immediately.\\n> If this is the case, many of the advantages of partitioned tables will be retained, while the advantages of global indexes will be gained.\\n\\nYes, we can keep the index rows for the deleted partition and clean them\\nup later, but what is the advantage of partitioning then? Just heap\\ndeletion quickly? Is that enough of a value?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> https://momjian.us\\n EnterpriseDB https://enterprisedb.com\\n\\n The usefulness of a cup is in its emptiness, Bruce Lee\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Jan 2021 12:46:49 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"Bruce Momjian <bruce@momjian.us> writes:\\n> On Mon, Jan 11, 2021 at 07:40:18PM +0800, 曾文旌 wrote:\\n>> This is indeed a typical scenario for a partitioned table.\\n>> there are two basic operations\\n>> 1) Monthly DETACH old child table\\n>> 2) Monthly ATTACH new child table\\n>> \\n>> For 1) The DETACH old child table can be finished immediately, global index can be kept valid after DETACH is completed, and the cleanup of garbage data in global index can be deferred to VACUUM.\\n\\n> Yes, we can keep the index rows for the deleted partition and clean them\\n> up later, but what is the advantage of partitioning then? Just heap\\n> deletion quickly? Is that enough of a value?\\n\\nMore to the point, you still have a massive index cleanup operation to do.\\nDeferred or not, that's going to take a lot of cycles, and it will leave\\nyou with a bloated global index. I find it hard to believe that this\\napproach will seem like an improvement over doing partitioning the way\\nwe do it now.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Jan 2021 13:34:32 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 11, 2021 at 12:46 PM Bruce Momjian <bruce@momjian.us> wrote:\\n> > For 1) The DETACH old child table can be finished immediately, global index can be kept valid after DETACH is completed, and the cleanup of garbage data in global index can be deferred to VACUUM.\\n> > This is similar to the global index optimization done by Oracle12c.\\n> > For 2) ATTACH new empty child table can also be completed immediately.\\n> > If this is the case, many of the advantages of partitioned tables will be retained, while the advantages of global indexes will be gained.\\n>\\n> Yes, we can keep the index rows for the deleted partition and clean them\\n> up later, but what is the advantage of partitioning then? Just heap\\n> deletion quickly? Is that enough of a value?\\n\\nI actually think the idea of lazily deleting the index entries is\\npretty good, but it won't work if the way the global index is\\nimplemented is by adding a tableoid column. Because then, I might\\ndetach a partition and later reattach it and the old index entries are\\nstill there but the table contents might have changed. Worse yet, the\\ntable might be dropped and the table OID reused for a completely\\nunrelated table with completely unrelated contents, which could then\\nbe attached as a new partition.\\n\\nOne of the big selling points of global indexes is that they allow you\\nto enforce uniqueness on a column unrelated to the partitioning\\ncolumn. Another is that you can look up a value by doing a single\\nindex scan on the global index rather than an index scan per\\npartition. Those things are no less valuable for performing index\\ndeletion lazily.\\n\\nHowever, there is a VACUUM amplification effect to worry about here\\nwhich Wenjing seems not to be considering. Suppose I have a table\\nwhich is not partitioned and it is 1TB in size with an index that is\\n128GB in size. To vacuum the table, I need to do 1TB + 128GB of I/O.\\nNow, suppose I now partition the table into 1024 partitions each with\\nits own local index. Each partition is 1GB in size and the index on\\neach partition is 128MB in size. To vacuum an individual partition\\nrequires 1GB + 128MB of I/O, so to vacuum all the partitions requires\\nthe same amount of total I/O as before. But, now suppose that I have a\\nsingle global index instead of a local index per partition. First, how\\nbig will that index be? It will not be 128GB, but somewhat bigger,\\nbecause it needs extra space for every indexed tuple. Let's say 140GB.\\nFurthermore, it will need to be vacuumed whenever any child is\\nvacuumed, because it contains some index entries from every child. So\\nthe total I/O to vacuum all partitions is now 1GB * 1024 + 140GB *\\n1024 = 141TB, which is a heck of a lot worse than the 1.125TB I\\nrequired with the unpartitioned table or the locally partitioned\\ntable.\\n\\nThat's not necessarily a death sentence for every use case, but it's\\ngoing to be pretty bad for tables that are big and heavily updated.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Jan 2021 13:37:02 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 11, 2021 at 10:37 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> I actually think the idea of lazily deleting the index entries is\\n> pretty good, but it won't work if the way the global index is\\n> implemented is by adding a tableoid column.\\n\\nPerhaps there is an opportunity to apply some of the infrastructure\\nthat Masahiko Sawada has been working on, that makes VACUUM more\\nincremental in certain specific scenarios:\\n\\nhttps://postgr.es/m/CAD21AoD0SkE11fMw4jD4RENAwBMcw1wasVnwpJVw3tVqPOQgAw@mail.gmail.com\\n\\nI think that VACUUM can be taught to skip the ambulkdelete() step for\\nindexes in many common scenarios. Global indexes might be one place in\\nwhich that's almost essential.\\n\\n> However, there is a VACUUM amplification effect to worry about here\\n> which Wenjing seems not to be considering.\\n\\n> That's not necessarily a death sentence for every use case, but it's\\n> going to be pretty bad for tables that are big and heavily updated.\\n\\nThe main way in which index vacuuming is currently a death sentence\\nfor this design (as you put it) is that it's an all-or-nothing thing.\\nPresumably you'll need to VACUUM the entire global index for each\\npartition that receives even one UPDATE. That seems pretty extreme,\\nand probably not acceptable. In a way it's not really a new problem,\\nbut the fact remains: it makes global indexes much less valuable.\\n\\nHowever, it probably would be okay if a global index feature performed\\npoorly in scenarios where partitions get lots of UPDATEs that produce\\nlots of index bloat and cause lots of LP_DEAD line pointers to\\naccumulate in heap pages. It is probably reasonable to just expect\\nusers to not do that if they want to get acceptable performance while\\nusing a global index. Especially since it probably is not so bad if\\nthe index bloat situation gets out of hand for just one of the\\npartitions (say the most recent one) every once in a while. You at\\nleast don't have the same crazy I/O multiplier effect that you\\ndescribed.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Jan 2021 11:01:20 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 11, 2021 at 01:37:02PM -0500, Robert Haas wrote:\\n> However, there is a VACUUM amplification effect to worry about here\\n...\\n> That's not necessarily a death sentence for every use case, but it's\\n> going to be pretty bad for tables that are big and heavily updated.\\n\\nYeah, I had not really gotten that far in my thinking, but someone is\\ngoing to need to create a POC and then we need to test it to see if it\\noffers a reasonably valuable feature.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> https://momjian.us\\n EnterpriseDB https://enterprisedb.com\\n\\n The usefulness of a cup is in its emptiness, Bruce Lee\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Jan 2021 14:23:13 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 11, 2021 at 11:01:20AM -0800, Peter Geoghegan wrote:\\n> However, it probably would be okay if a global index feature performed\\n> poorly in scenarios where partitions get lots of UPDATEs that produce\\n> lots of index bloat and cause lots of LP_DEAD line pointers to\\n> accumulate in heap pages. It is probably reasonable to just expect\\n> users to not do that if they want to get acceptable performance while\\n> using a global index. Especially since it probably is not so bad if\\n> the index bloat situation gets out of hand for just one of the\\n> partitions (say the most recent one) every once in a while. You at\\n> least don't have the same crazy I/O multiplier effect that you\\n> described.\\n\\nOnce you layer on all the places a global index will be worse than just\\ncreating a single large table, or a partitioned table with an index per\\nchild, there might not be much usefulness left. A POC patch might tell\\nus that, and might allow us to mark it as \\\"not wanted\\\".\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> https://momjian.us\\n EnterpriseDB https://enterprisedb.com\\n\\n The usefulness of a cup is in its emptiness, Bruce Lee\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Jan 2021 14:25:55 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 11, 2021 at 11:25 AM Bruce Momjian <bruce@momjian.us> wrote:\\n> Once you layer on all the places a global index will be worse than just\\n> creating a single large table, or a partitioned table with an index per\\n> child, there might not be much usefulness left. A POC patch might tell\\n> us that, and might allow us to mark it as \\\"not wanted\\\".\\n\\nI'm confused. Of course it's true to some degree that having a global\\nindex \\\"defeats the purpose\\\" of having a partitioned table. But only to\\na degree. And for some users it will make the difference between using\\npartitioning and not using partitioning -- they simply won't be able\\nto tolerate not having it available (e.g. because of a requirement for\\na unique constraint that does not cover the partitioning key).\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Jan 2021 12:05:43 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 11, 2021 at 12:05:43PM -0800, Peter Geoghegan wrote:\\n> On Mon, Jan 11, 2021 at 11:25 AM Bruce Momjian <bruce@momjian.us> wrote:\\n> > Once you layer on all the places a global index will be worse than just\\n> > creating a single large table, or a partitioned table with an index per\\n> > child, there might not be much usefulness left. A POC patch might tell\\n> > us that, and might allow us to mark it as \\\"not wanted\\\".\\n> \\n> I'm confused. Of course it's true to some degree that having a global\\n> index \\\"defeats the purpose\\\" of having a partitioned table. But only to\\n> a degree. And for some users it will make the difference between using\\n> partitioning and not using partitioning -- they simply won't be able\\n> to tolerate not having it available (e.g. because of a requirement for\\n> a unique constraint that does not cover the partitioning key).\\n\\nYes, that is a good point. For those cases, I think we need to look at\\nthe code complexity/overhead of supporting that feature. There are\\ngoing to be a few cases it is a win, but will the code complexity be\\nworth it?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> https://momjian.us\\n EnterpriseDB https://enterprisedb.com\\n\\n The usefulness of a cup is in its emptiness, Bruce Lee\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Jan 2021 15:24:06 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n },\n {\n \"msg_contents\": \"> 2021年1月12日 02:37，Robert Haas <robertmhaas@gmail.com> 写道：\\n> \\n> On Mon, Jan 11, 2021 at 12:46 PM Bruce Momjian <bruce@momjian.us> wrote:\\n>>> For 1) The DETACH old child table can be finished immediately, global index can be kept valid after DETACH is completed, and the cleanup of garbage data in global index can be deferred to VACUUM.\\n>>> This is similar to the global index optimization done by Oracle12c.\\n>>> For 2) ATTACH new empty child table can also be completed immediately.\\n>>> If this is the case, many of the advantages of partitioned tables will be retained, while the advantages of global indexes will be gained.\\n>> \\n>> Yes, we can keep the index rows for the deleted partition and clean them\\n>> up later, but what is the advantage of partitioning then? Just heap\\n>> deletion quickly? Is that enough of a value?\\n> \\n> I actually think the idea of lazily deleting the index entries is\\n> pretty good, but it won't work if the way the global index is\\n> implemented is by adding a tableoid column. Because then, I might\\n> detach a partition and later reattach it and the old index entries are\\n> still there but the table contents might have changed. Worse yet, the\\n> table might be dropped and the table OID reused for a completely\\n> unrelated table with completely unrelated contents, which could then\\n> be attached as a new partition.\\n> \\n> One of the big selling points of global indexes is that they allow you\\n> to enforce uniqueness on a column unrelated to the partitioning\\n> column. Another is that you can look up a value by doing a single\\n> index scan on the global index rather than an index scan per\\n> partition. Those things are no less valuable for performing index\\n> deletion lazily.\\n> \\n> However, there is a VACUUM amplification effect to worry about here\\n> which Wenjing seems not to be considering. Suppose I have a table\\n> which is not partitioned and it is 1TB in size with an index that is\\n> 128GB in size. To vacuum the table, I need to do 1TB + 128GB of I/O.\\n> Now, suppose I now partition the table into 1024 partitions each with\\n> its own local index. Each partition is 1GB in size and the index on\\n> each partition is 128MB in size. To vacuum an individual partition\\n> requires 1GB + 128MB of I/O, so to vacuum all the partitions requires\\n> the same amount of total I/O as before. But, now suppose that I have a\\n> single global index instead of a local index per partition. First, how\\n> big will that index be? It will not be 128GB, but somewhat bigger,\\n> because it needs extra space for every indexed tuple. Let's say 140GB.\\n> Furthermore, it will need to be vacuumed whenever any child is\\n> vacuumed, because it contains some index entries from every child. So\\n> the total I/O to vacuum all partitions is now 1GB * 1024 + 140GB *\\n> 1024 = 141TB, which is a heck of a lot worse than the 1.125TB I\\n> required with the unpartitioned table or the locally partitioned\\n> table.\\nThank you for pointing this out.\\nIt seems that some optimization can be done, but there is no good way\\nto completely eliminate the vacuum amplification effect of the global index.\\nMaybe we can only count on Zheap, which doesn't need to do Vaccum.\\n\\n\\n\\n> \\n> That's not necessarily a death sentence for every use case, but it's\\n> going to be pretty bad for tables that are big and heavily updated.\\n> \\n> -- \\n> Robert Haas\\n> EDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Mon, 18 Jan 2021 18:00:40 +0800\",\n \"msg_from\": \"=?utf-8?B?5pu+5paH5peM?= <wenjing.zwj@alibaba-inc.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Global Index\"\n }\n]","isConversational":false}}},{"rowIdx":49366,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Please consider this scenario (race conditions):\n\n1. FlushBuffer() has written the buffer but hasn't yet managed to clear the\nBM_DIRTY flag (however BM_JUST_DIRTIED could be cleared by now).\n\n2. Another backend modified a hint bit and called MarkBufferDirtyHint().\n\n3. In MarkBufferDirtyHint(), if XLogHintBitIsNeeded() evaluates to true\n(e.g. due to checksums enabled), new LSN is computed, however it's not\nassigned to the page because the buffer is still dirty:\n\n\tif (!(buf_state & BM_DIRTY))\n\t{\n\t\t...\n\n\t\tif (!XLogRecPtrIsInvalid(lsn))\n\t\t\tPageSetLSN(page, lsn);\n\t}\n\n4. MarkBufferDirtyHint() completes.\n\n5. In the first session, FlushBuffer()->TerminateBufferIO() will not clear\nBM_DIRTY because MarkBufferDirtyHint() has eventually set\nBM_JUST_DIRTIED. Thus the hint bit change itself will be written by the next\ncall of FlushBuffer(). However page LSN is hasn't been updated so the\nrequirement that WAL must be flushed first is not met.\n\nI think that PageSetLSN() should be called regardless BM_DIRTY. Do I miss any\nsubtle detail?\n\n-- \nAntonin Houska\nWeb: https://www.cybertec-postgresql.com\n\n\n","msg_date":"Wed, 30 Oct 2019 14:44:18 +0100","msg_from":"Antonin Houska <ah@cybertec.at>","msg_from_op":true,"msg_subject":"MarkBufferDirtyHint() and LSN update"},{"msg_contents":"On Wed, Oct 30, 2019 at 02:44:18PM +0100, Antonin Houska wrote:\n>Please consider this scenario (race conditions):\n>\n>1. FlushBuffer() has written the buffer but hasn't yet managed to clear the\n>BM_DIRTY flag (however BM_JUST_DIRTIED could be cleared by now).\n>\n>2. Another backend modified a hint bit and called MarkBufferDirtyHint().\n>\n>3. In MarkBufferDirtyHint(), if XLogHintBitIsNeeded() evaluates to true\n>(e.g. due to checksums enabled), new LSN is computed, however it's not\n>assigned to the page because the buffer is still dirty:\n>\n>\tif (!(buf_state & BM_DIRTY))\n>\t{\n>\t\t...\n>\n>\t\tif (!XLogRecPtrIsInvalid(lsn))\n>\t\t\tPageSetLSN(page, lsn);\n>\t}\n>\n>4. MarkBufferDirtyHint() completes.\n>\n>5. In the first session, FlushBuffer()->TerminateBufferIO() will not clear\n>BM_DIRTY because MarkBufferDirtyHint() has eventually set\n>BM_JUST_DIRTIED. Thus the hint bit change itself will be written by the next\n>call of FlushBuffer(). However page LSN is hasn't been updated so the\n>requirement that WAL must be flushed first is not met.\n>\n>I think that PageSetLSN() should be called regardless BM_DIRTY. Do I miss any\n>subtle detail?\n>\n\nIsn't this prevented by locking of the buffer header? Both FlushBuffer\nand MarkBufferDirtyHint do obtain that lock. I see MarkBufferDirtyHint\ndoes a bit of work before, but that's related to BM_PERMANENT.\n\nIf there really is a race condition, it shouldn't be that difficult to\ntrigger it by adding a sleep or a breakpoint, I guess.\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Wed, 30 Oct 2019 20:42:37 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:\n\n> On Wed, Oct 30, 2019 at 02:44:18PM +0100, Antonin Houska wrote:\n> >Please consider this scenario (race conditions):\n> >\n> >1. FlushBuffer() has written the buffer but hasn't yet managed to clear the\n> >BM_DIRTY flag (however BM_JUST_DIRTIED could be cleared by now).\n> >\n> >2. Another backend modified a hint bit and called MarkBufferDirtyHint().\n> >\n> >3. In MarkBufferDirtyHint(), if XLogHintBitIsNeeded() evaluates to true\n> >(e.g. due to checksums enabled), new LSN is computed, however it's not\n> >assigned to the page because the buffer is still dirty:\n> >\n> >\tif (!(buf_state & BM_DIRTY))\n> >\t{\n> >\t\t...\n> >\n> >\t\tif (!XLogRecPtrIsInvalid(lsn))\n> >\t\t\tPageSetLSN(page, lsn);\n> >\t}\n> >\n> >4. MarkBufferDirtyHint() completes.\n> >\n> >5. In the first session, FlushBuffer()->TerminateBufferIO() will not clear\n> >BM_DIRTY because MarkBufferDirtyHint() has eventually set\n> >BM_JUST_DIRTIED. Thus the hint bit change itself will be written by the next\n> >call of FlushBuffer(). However page LSN is hasn't been updated so the\n> >requirement that WAL must be flushed first is not met.\n> >\n> >I think that PageSetLSN() should be called regardless BM_DIRTY. Do I miss any\n> >subtle detail?\n> >\n> \n> Isn't this prevented by locking of the buffer header? Both FlushBuffer\n> and MarkBufferDirtyHint do obtain that lock. I see MarkBufferDirtyHint\n> does a bit of work before, but that's related to BM_PERMANENT.\n> \n> If there really is a race condition, it shouldn't be that difficult to\n> trigger it by adding a sleep or a breakpoint, I guess.\n\nYes, I had verified it using gdb: inserted a row into a table, then attached\ngdb to checkpointer and stopped it when FlushBuffer() was about to call\nTerminateBufferIO(). Then, when scanning the table, I saw that\nMarkBufferDirtyHint() skipped the call of PageSetLSN(). Finally, before\ncheckpointer unlocked the buffer header in TerminateBufferIO(), buf_state was\n3553624066 ~ 0b11010011110100000000000000000010.\n\nWith BM_DIRTY defined as\n\n\t#define BM_DIRTY\t\t\t\t(1U << 23)\n\nthe flag appeared to be set.\n\n-- \nAntonin Houska\nWeb: https://www.cybertec-postgresql.com\n\n\n","msg_date":"Thu, 31 Oct 2019 09:43:47 +0100","msg_from":"Antonin Houska <ah@cybertec.at>","msg_from_op":true,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"On Wed, Oct 30, 2019 at 9:43 AM Antonin Houska <ah@cybertec.at> wrote:\n> 5. In the first session, FlushBuffer()->TerminateBufferIO() will not clear\n> BM_DIRTY because MarkBufferDirtyHint() has eventually set\n> BM_JUST_DIRTIED. Thus the hint bit change itself will be written by the next\n> call of FlushBuffer(). However page LSN is hasn't been updated so the\n> requirement that WAL must be flushed first is not met.\n\nThis part confuses me. Are you saying that MarkBufferDirtyHint() can\nset BM_JUST_DIRTIED without setting BM_DIRTY?\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 1 Nov 2019 11:23:20 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> wrote:\n\n> On Wed, Oct 30, 2019 at 9:43 AM Antonin Houska <ah@cybertec.at> wrote:\n> > 5. In the first session, FlushBuffer()->TerminateBufferIO() will not clear\n> > BM_DIRTY because MarkBufferDirtyHint() has eventually set\n> > BM_JUST_DIRTIED. Thus the hint bit change itself will be written by the next\n> > call of FlushBuffer(). However page LSN is hasn't been updated so the\n> > requirement that WAL must be flushed first is not met.\n> \n> This part confuses me. Are you saying that MarkBufferDirtyHint() can\n> set BM_JUST_DIRTIED without setting BM_DIRTY?\n\nNo, I'm saying that MarkBufferDirtyHint() leaves BM_DIRTY set, as\nexpected. However, if things happen in the order I described, then LSN\nreturned by XLogSaveBufferForHint() is not assigned to the page.\n\n-- \nAntonin Houska\nWeb: https://www.cybertec-postgresql.com\n\n\n","msg_date":"Fri, 01 Nov 2019 18:51:26 +0100","msg_from":"Antonin Houska <ah@cybertec.at>","msg_from_op":true,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"On Thu, Oct 31, 2019 at 09:43:47AM +0100, Antonin Houska wrote:\n> Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:\n>> Isn't this prevented by locking of the buffer header? Both FlushBuffer\n>> and MarkBufferDirtyHint do obtain that lock. I see MarkBufferDirtyHint\n>> does a bit of work before, but that's related to BM_PERMANENT.\n\nIn FlushBuffer() you have a window after fetching the buffer state and\nthe header is once unlocked until TerminateBufferIO() gets called\n(this would take again a lock on the buffer header), so it is\nlogically possible for the buffer to be marked as dirty once again\ncausing the update of the LSN on the page to be missed even if a\nbackup block has been written in WAL.\n\n> Yes, I had verified it using gdb: inserted a row into a table, then attached\n> gdb to checkpointer and stopped it when FlushBuffer() was about to call\n> TerminateBufferIO(). Then, when scanning the table, I saw that\n> MarkBufferDirtyHint() skipped the call of PageSetLSN(). Finally, before\n> checkpointer unlocked the buffer header in TerminateBufferIO(), buf_state was\n> 3553624066 ~ 0b11010011110100000000000000000010.\n\nSmall typo here: 11010011110100000000000000000010...\n\n> With BM_DIRTY defined as\n> \n> \t#define BM_DIRTY\t\t\t\t(1U << 23)\n> \n> the flag appeared to be set.\n\n... Still, the bit is set here.\n\nDoes something like the attached patch make sense? Reviews are\nwelcome.\n--\nMichael","msg_date":"Mon, 11 Nov 2019 14:47:37 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> wrote:\n> Does something like the attached patch make sense? Reviews are\n> welcome.\n\nThis looks good to me.\n\n-- \nAntonin Houska\nWeb: https://www.cybertec-postgresql.com\n\n\n","msg_date":"Mon, 11 Nov 2019 10:03:14 +0100","msg_from":"Antonin Houska <ah@cybertec.at>","msg_from_op":true,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"At Mon, 11 Nov 2019 10:03:14 +0100, Antonin Houska <ah@cybertec.at> wrote in \n> Michael Paquier <michael@paquier.xyz> wrote:\n> > Does something like the attached patch make sense? Reviews are\n> > welcome.\n> \n> This looks good to me.\n\nI have a qustion.\n\nThe current code assumes that !BM_DIRTY means that the function is\ndirtying the page. But if !BM_JUST_DIRTIED, the function actually is\ngoing to re-dirty the page even if BM_DIRTY.\n\nIf this is correct, the trigger for stats update is not !BM_DIRTY but\n!BM_JUST_DIRTIED, or the fact that we passed the line of\nXLogSaveBufferForHint() could be the trigger, regardless whether the\nLSN is valid or not.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Tue, 12 Nov 2019 21:31:34 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote:\n\n> At Mon, 11 Nov 2019 10:03:14 +0100, Antonin Houska <ah@cybertec.at> wrote in \n> > Michael Paquier <michael@paquier.xyz> wrote:\n> > > Does something like the attached patch make sense? Reviews are\n> > > welcome.\n> > \n> > This looks good to me.\n> \n> I have a qustion.\n> \n> The current code assumes that !BM_DIRTY means that the function is\n> dirtying the page. But if !BM_JUST_DIRTIED, the function actually is\n> going to re-dirty the page even if BM_DIRTY.\n\nIt makes sense to me. I can imagine the following:\n\n1. FlushBuffer() cleared BM_JUST_DIRTIED, wrote the page to disk but hasn't\nyet cleared BM_DIRTY.\n\n2. Another backend changed a hint bit in shared memory and called\nMarkBufferDirtyHint().\n\nThus FlushBuffer() missed the current hint bit change, so we need to dirty the\npage again.\n\n> If this is correct, the trigger for stats update is not !BM_DIRTY but\n> !BM_JUST_DIRTIED, or the fact that we passed the line of\n> XLogSaveBufferForHint() could be the trigger, regardless whether the\n> LSN is valid or not.\n\nI agree.\n\n-- \nAntonin Houska\nWeb: https://www.cybertec-postgresql.com\n\n\n","msg_date":"Tue, 12 Nov 2019 14:27:16 +0100","msg_from":"Antonin Houska <ah@cybertec.at>","msg_from_op":true,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"On Mon, Nov 11, 2019 at 10:03:14AM +0100, Antonin Houska wrote:\n> This looks good to me.\n\nActually, no, this is not good. I have been studying more the patch,\nand after stressing more this code path with a cluster having\nchecksums enabled and shared_buffers at 1MB, I have been able to make\na couple of page's LSNs go backwards with pgbench -s 100. The cause\nwas simply that the page got flushed with a newer LSN than what was\nreturned by XLogSaveBufferForHint() before taking the buffer header\nlock, so updating only the LSN for a non-dirty page was simply\nguarding against that.\n--\nMichael","msg_date":"Wed, 13 Nov 2019 21:17:03 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"On Wed, Nov 13, 2019 at 09:17:03PM +0900, Michael Paquier wrote:\n> Actually, no, this is not good. I have been studying more the patch,\n> and after stressing more this code path with a cluster having\n> checksums enabled and shared_buffers at 1MB, I have been able to make\n> a couple of page's LSNs go backwards with pgbench -s 100. The cause\n> was simply that the page got flushed with a newer LSN than what was\n> returned by XLogSaveBufferForHint() before taking the buffer header\n> lock, so updating only the LSN for a non-dirty page was simply\n> guarding against that.\n\nfor the reference attached is the trick I have used, adding an extra\nassertion check in PageSetLSN(). \n--\nMichael","msg_date":"Thu, 14 Nov 2019 12:01:29 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"At Thu, 14 Nov 2019 12:01:29 +0900, Michael Paquier <michael@paquier.xyz> wrote in \n> On Wed, Nov 13, 2019 at 09:17:03PM +0900, Michael Paquier wrote:\n> > Actually, no, this is not good. I have been studying more the patch,\n> > and after stressing more this code path with a cluster having\n> > checksums enabled and shared_buffers at 1MB, I have been able to make\n> > a couple of page's LSNs go backwards with pgbench -s 100. The cause\n> > was simply that the page got flushed with a newer LSN than what was\n> > returned by XLogSaveBufferForHint() before taking the buffer header\n> > lock, so updating only the LSN for a non-dirty page was simply\n> > guarding against that.\n\nI thought of something like that but forgot to mention. But after\nthinking of that, couldn't the current code can do the same think even\nthough with a far small probability? Still a session with smaller hint\nLSN but didn't entered the header lock section yet can be cut-in by\nanother session with larger hint LSN.\n\n> for the reference attached is the trick I have used, adding an extra\n> assertion check in PageSetLSN(). \n\nI believe that all locations where Page-LSN is set is in the same\nbuffer-ex-lock section with XLogInsert.. but not sure.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Thu, 14 Nov 2019 16:59:49 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Mon, Nov 11, 2019 at 10:03:14AM +0100, Antonin Houska wrote:\n> > This looks good to me.\n> \n> Actually, no, this is not good. I have been studying more the patch,\n> and after stressing more this code path with a cluster having\n> checksums enabled and shared_buffers at 1MB, I have been able to make\n> a couple of page's LSNs go backwards with pgbench -s 100. The cause\n> was simply that the page got flushed with a newer LSN than what was\n> returned by XLogSaveBufferForHint() before taking the buffer header\n> lock, so updating only the LSN for a non-dirty page was simply\n> guarding against that.\n\nInteresting. Now that I know about the problem, I could have reproduced it\nusing gdb: MarkBufferDirtyHint() was called by 2 backends concurrently in such\na way that the first backend generates the LSN, but before it manages to\nassign it to the page, another backend generates another LSN and sets it.\n\nCan't we just apply the attached diff on the top of your patch?\n\nAlso I wonder how checksums helped you to discover the problem? Although I\ncould simulate the setting of lower LSN, I could not see any broken\nchecksum. And I wouldn't even expect that since FlushBuffer() copies the\nbuffer into backend local memory before it calculates the checksum.\n\n-- \nAntonin Houska\nWeb: https://www.cybertec-postgresql.com","msg_date":"Thu, 14 Nov 2019 15:48:31 +0100","msg_from":"Antonin Houska <ah@cybertec.at>","msg_from_op":true,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"Antonin Houska <ah@cybertec.at> wrote:\n\n> Michael Paquier <michael@paquier.xyz> wrote:\n> \n> > On Mon, Nov 11, 2019 at 10:03:14AM +0100, Antonin Houska wrote:\n> > > This looks good to me.\n> > \n> > Actually, no, this is not good. I have been studying more the patch,\n> > and after stressing more this code path with a cluster having\n> > checksums enabled and shared_buffers at 1MB, I have been able to make\n> > a couple of page's LSNs go backwards with pgbench -s 100. The cause\n> > was simply that the page got flushed with a newer LSN than what was\n> > returned by XLogSaveBufferForHint() before taking the buffer header\n> > lock, so updating only the LSN for a non-dirty page was simply\n> > guarding against that.\n> \n> Interesting. Now that I know about the problem, I could have reproduced it\n> using gdb: MarkBufferDirtyHint() was called by 2 backends concurrently in such\n> a way that the first backend generates the LSN, but before it manages to\n> assign it to the page, another backend generates another LSN and sets it.\n> \n> Can't we just apply the attached diff on the top of your patch?\n\nI wanted to register the patch for the next CF so it's not forgotten, but see\nit's already there. Why have you set the status to \"withdrawn\"?\n\n-- \nAntonin Houska\nWeb: https://www.cybertec-postgresql.com\n\n\n","msg_date":"Fri, 20 Dec 2019 16:30:38 +0100","msg_from":"Antonin Houska <ah@cybertec.at>","msg_from_op":true,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"},{"msg_contents":"On Fri, Dec 20, 2019 at 04:30:38PM +0100, Antonin Houska wrote:\n> I wanted to register the patch for the next CF so it's not forgotten, but see\n> it's already there. Why have you set the status to \"withdrawn\"?\n\nBecause my patch was incorrect, and I did not make enough bandwidth to\nthink more on the matter. I am actually not sure that what you are\nproposing is better.. If you wish to get that reviewed, could you add\na new CF entry?\n--\nMichael","msg_date":"Sat, 21 Dec 2019 11:44:25 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: MarkBufferDirtyHint() and LSN update"}],"string":"[\n {\n \"msg_contents\": \"Please consider this scenario (race conditions):\\n\\n1. FlushBuffer() has written the buffer but hasn't yet managed to clear the\\nBM_DIRTY flag (however BM_JUST_DIRTIED could be cleared by now).\\n\\n2. Another backend modified a hint bit and called MarkBufferDirtyHint().\\n\\n3. In MarkBufferDirtyHint(), if XLogHintBitIsNeeded() evaluates to true\\n(e.g. due to checksums enabled), new LSN is computed, however it's not\\nassigned to the page because the buffer is still dirty:\\n\\n\\tif (!(buf_state & BM_DIRTY))\\n\\t{\\n\\t\\t...\\n\\n\\t\\tif (!XLogRecPtrIsInvalid(lsn))\\n\\t\\t\\tPageSetLSN(page, lsn);\\n\\t}\\n\\n4. MarkBufferDirtyHint() completes.\\n\\n5. In the first session, FlushBuffer()->TerminateBufferIO() will not clear\\nBM_DIRTY because MarkBufferDirtyHint() has eventually set\\nBM_JUST_DIRTIED. Thus the hint bit change itself will be written by the next\\ncall of FlushBuffer(). However page LSN is hasn't been updated so the\\nrequirement that WAL must be flushed first is not met.\\n\\nI think that PageSetLSN() should be called regardless BM_DIRTY. Do I miss any\\nsubtle detail?\\n\\n-- \\nAntonin Houska\\nWeb: https://www.cybertec-postgresql.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 14:44:18 +0100\",\n \"msg_from\": \"Antonin Houska <ah@cybertec.at>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 30, 2019 at 02:44:18PM +0100, Antonin Houska wrote:\\n>Please consider this scenario (race conditions):\\n>\\n>1. FlushBuffer() has written the buffer but hasn't yet managed to clear the\\n>BM_DIRTY flag (however BM_JUST_DIRTIED could be cleared by now).\\n>\\n>2. Another backend modified a hint bit and called MarkBufferDirtyHint().\\n>\\n>3. In MarkBufferDirtyHint(), if XLogHintBitIsNeeded() evaluates to true\\n>(e.g. due to checksums enabled), new LSN is computed, however it's not\\n>assigned to the page because the buffer is still dirty:\\n>\\n>\\tif (!(buf_state & BM_DIRTY))\\n>\\t{\\n>\\t\\t...\\n>\\n>\\t\\tif (!XLogRecPtrIsInvalid(lsn))\\n>\\t\\t\\tPageSetLSN(page, lsn);\\n>\\t}\\n>\\n>4. MarkBufferDirtyHint() completes.\\n>\\n>5. In the first session, FlushBuffer()->TerminateBufferIO() will not clear\\n>BM_DIRTY because MarkBufferDirtyHint() has eventually set\\n>BM_JUST_DIRTIED. Thus the hint bit change itself will be written by the next\\n>call of FlushBuffer(). However page LSN is hasn't been updated so the\\n>requirement that WAL must be flushed first is not met.\\n>\\n>I think that PageSetLSN() should be called regardless BM_DIRTY. Do I miss any\\n>subtle detail?\\n>\\n\\nIsn't this prevented by locking of the buffer header? Both FlushBuffer\\nand MarkBufferDirtyHint do obtain that lock. I see MarkBufferDirtyHint\\ndoes a bit of work before, but that's related to BM_PERMANENT.\\n\\nIf there really is a race condition, it shouldn't be that difficult to\\ntrigger it by adding a sleep or a breakpoint, I guess.\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 20:42:37 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:\\n\\n> On Wed, Oct 30, 2019 at 02:44:18PM +0100, Antonin Houska wrote:\\n> >Please consider this scenario (race conditions):\\n> >\\n> >1. FlushBuffer() has written the buffer but hasn't yet managed to clear the\\n> >BM_DIRTY flag (however BM_JUST_DIRTIED could be cleared by now).\\n> >\\n> >2. Another backend modified a hint bit and called MarkBufferDirtyHint().\\n> >\\n> >3. In MarkBufferDirtyHint(), if XLogHintBitIsNeeded() evaluates to true\\n> >(e.g. due to checksums enabled), new LSN is computed, however it's not\\n> >assigned to the page because the buffer is still dirty:\\n> >\\n> >\\tif (!(buf_state & BM_DIRTY))\\n> >\\t{\\n> >\\t\\t...\\n> >\\n> >\\t\\tif (!XLogRecPtrIsInvalid(lsn))\\n> >\\t\\t\\tPageSetLSN(page, lsn);\\n> >\\t}\\n> >\\n> >4. MarkBufferDirtyHint() completes.\\n> >\\n> >5. In the first session, FlushBuffer()->TerminateBufferIO() will not clear\\n> >BM_DIRTY because MarkBufferDirtyHint() has eventually set\\n> >BM_JUST_DIRTIED. Thus the hint bit change itself will be written by the next\\n> >call of FlushBuffer(). However page LSN is hasn't been updated so the\\n> >requirement that WAL must be flushed first is not met.\\n> >\\n> >I think that PageSetLSN() should be called regardless BM_DIRTY. Do I miss any\\n> >subtle detail?\\n> >\\n> \\n> Isn't this prevented by locking of the buffer header? Both FlushBuffer\\n> and MarkBufferDirtyHint do obtain that lock. I see MarkBufferDirtyHint\\n> does a bit of work before, but that's related to BM_PERMANENT.\\n> \\n> If there really is a race condition, it shouldn't be that difficult to\\n> trigger it by adding a sleep or a breakpoint, I guess.\\n\\nYes, I had verified it using gdb: inserted a row into a table, then attached\\ngdb to checkpointer and stopped it when FlushBuffer() was about to call\\nTerminateBufferIO(). Then, when scanning the table, I saw that\\nMarkBufferDirtyHint() skipped the call of PageSetLSN(). Finally, before\\ncheckpointer unlocked the buffer header in TerminateBufferIO(), buf_state was\\n3553624066 ~ 0b11010011110100000000000000000010.\\n\\nWith BM_DIRTY defined as\\n\\n\\t#define BM_DIRTY\\t\\t\\t\\t(1U << 23)\\n\\nthe flag appeared to be set.\\n\\n-- \\nAntonin Houska\\nWeb: https://www.cybertec-postgresql.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 09:43:47 +0100\",\n \"msg_from\": \"Antonin Houska <ah@cybertec.at>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 30, 2019 at 9:43 AM Antonin Houska <ah@cybertec.at> wrote:\\n> 5. In the first session, FlushBuffer()->TerminateBufferIO() will not clear\\n> BM_DIRTY because MarkBufferDirtyHint() has eventually set\\n> BM_JUST_DIRTIED. Thus the hint bit change itself will be written by the next\\n> call of FlushBuffer(). However page LSN is hasn't been updated so the\\n> requirement that WAL must be flushed first is not met.\\n\\nThis part confuses me. Are you saying that MarkBufferDirtyHint() can\\nset BM_JUST_DIRTIED without setting BM_DIRTY?\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 11:23:20 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> On Wed, Oct 30, 2019 at 9:43 AM Antonin Houska <ah@cybertec.at> wrote:\\n> > 5. In the first session, FlushBuffer()->TerminateBufferIO() will not clear\\n> > BM_DIRTY because MarkBufferDirtyHint() has eventually set\\n> > BM_JUST_DIRTIED. Thus the hint bit change itself will be written by the next\\n> > call of FlushBuffer(). However page LSN is hasn't been updated so the\\n> > requirement that WAL must be flushed first is not met.\\n> \\n> This part confuses me. Are you saying that MarkBufferDirtyHint() can\\n> set BM_JUST_DIRTIED without setting BM_DIRTY?\\n\\nNo, I'm saying that MarkBufferDirtyHint() leaves BM_DIRTY set, as\\nexpected. However, if things happen in the order I described, then LSN\\nreturned by XLogSaveBufferForHint() is not assigned to the page.\\n\\n-- \\nAntonin Houska\\nWeb: https://www.cybertec-postgresql.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 01 Nov 2019 18:51:26 +0100\",\n \"msg_from\": \"Antonin Houska <ah@cybertec.at>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 09:43:47AM +0100, Antonin Houska wrote:\\n> Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:\\n>> Isn't this prevented by locking of the buffer header? Both FlushBuffer\\n>> and MarkBufferDirtyHint do obtain that lock. I see MarkBufferDirtyHint\\n>> does a bit of work before, but that's related to BM_PERMANENT.\\n\\nIn FlushBuffer() you have a window after fetching the buffer state and\\nthe header is once unlocked until TerminateBufferIO() gets called\\n(this would take again a lock on the buffer header), so it is\\nlogically possible for the buffer to be marked as dirty once again\\ncausing the update of the LSN on the page to be missed even if a\\nbackup block has been written in WAL.\\n\\n> Yes, I had verified it using gdb: inserted a row into a table, then attached\\n> gdb to checkpointer and stopped it when FlushBuffer() was about to call\\n> TerminateBufferIO(). Then, when scanning the table, I saw that\\n> MarkBufferDirtyHint() skipped the call of PageSetLSN(). Finally, before\\n> checkpointer unlocked the buffer header in TerminateBufferIO(), buf_state was\\n> 3553624066 ~ 0b11010011110100000000000000000010.\\n\\nSmall typo here: 11010011110100000000000000000010...\\n\\n> With BM_DIRTY defined as\\n> \\n> \\t#define BM_DIRTY\\t\\t\\t\\t(1U << 23)\\n> \\n> the flag appeared to be set.\\n\\n... Still, the bit is set here.\\n\\nDoes something like the attached patch make sense? Reviews are\\nwelcome.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 11 Nov 2019 14:47:37 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> wrote:\\n> Does something like the attached patch make sense? Reviews are\\n> welcome.\\n\\nThis looks good to me.\\n\\n-- \\nAntonin Houska\\nWeb: https://www.cybertec-postgresql.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Nov 2019 10:03:14 +0100\",\n \"msg_from\": \"Antonin Houska <ah@cybertec.at>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"At Mon, 11 Nov 2019 10:03:14 +0100, Antonin Houska <ah@cybertec.at> wrote in \\n> Michael Paquier <michael@paquier.xyz> wrote:\\n> > Does something like the attached patch make sense? Reviews are\\n> > welcome.\\n> \\n> This looks good to me.\\n\\nI have a qustion.\\n\\nThe current code assumes that !BM_DIRTY means that the function is\\ndirtying the page. But if !BM_JUST_DIRTIED, the function actually is\\ngoing to re-dirty the page even if BM_DIRTY.\\n\\nIf this is correct, the trigger for stats update is not !BM_DIRTY but\\n!BM_JUST_DIRTIED, or the fact that we passed the line of\\nXLogSaveBufferForHint() could be the trigger, regardless whether the\\nLSN is valid or not.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Nov 2019 21:31:34 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote:\\n\\n> At Mon, 11 Nov 2019 10:03:14 +0100, Antonin Houska <ah@cybertec.at> wrote in \\n> > Michael Paquier <michael@paquier.xyz> wrote:\\n> > > Does something like the attached patch make sense? Reviews are\\n> > > welcome.\\n> > \\n> > This looks good to me.\\n> \\n> I have a qustion.\\n> \\n> The current code assumes that !BM_DIRTY means that the function is\\n> dirtying the page. But if !BM_JUST_DIRTIED, the function actually is\\n> going to re-dirty the page even if BM_DIRTY.\\n\\nIt makes sense to me. I can imagine the following:\\n\\n1. FlushBuffer() cleared BM_JUST_DIRTIED, wrote the page to disk but hasn't\\nyet cleared BM_DIRTY.\\n\\n2. Another backend changed a hint bit in shared memory and called\\nMarkBufferDirtyHint().\\n\\nThus FlushBuffer() missed the current hint bit change, so we need to dirty the\\npage again.\\n\\n> If this is correct, the trigger for stats update is not !BM_DIRTY but\\n> !BM_JUST_DIRTIED, or the fact that we passed the line of\\n> XLogSaveBufferForHint() could be the trigger, regardless whether the\\n> LSN is valid or not.\\n\\nI agree.\\n\\n-- \\nAntonin Houska\\nWeb: https://www.cybertec-postgresql.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Nov 2019 14:27:16 +0100\",\n \"msg_from\": \"Antonin Houska <ah@cybertec.at>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 11, 2019 at 10:03:14AM +0100, Antonin Houska wrote:\\n> This looks good to me.\\n\\nActually, no, this is not good. I have been studying more the patch,\\nand after stressing more this code path with a cluster having\\nchecksums enabled and shared_buffers at 1MB, I have been able to make\\na couple of page's LSNs go backwards with pgbench -s 100. The cause\\nwas simply that the page got flushed with a newer LSN than what was\\nreturned by XLogSaveBufferForHint() before taking the buffer header\\nlock, so updating only the LSN for a non-dirty page was simply\\nguarding against that.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 13 Nov 2019 21:17:03 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 13, 2019 at 09:17:03PM +0900, Michael Paquier wrote:\\n> Actually, no, this is not good. I have been studying more the patch,\\n> and after stressing more this code path with a cluster having\\n> checksums enabled and shared_buffers at 1MB, I have been able to make\\n> a couple of page's LSNs go backwards with pgbench -s 100. The cause\\n> was simply that the page got flushed with a newer LSN than what was\\n> returned by XLogSaveBufferForHint() before taking the buffer header\\n> lock, so updating only the LSN for a non-dirty page was simply\\n> guarding against that.\\n\\nfor the reference attached is the trick I have used, adding an extra\\nassertion check in PageSetLSN(). \\n--\\nMichael\",\n \"msg_date\": \"Thu, 14 Nov 2019 12:01:29 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"At Thu, 14 Nov 2019 12:01:29 +0900, Michael Paquier <michael@paquier.xyz> wrote in \\n> On Wed, Nov 13, 2019 at 09:17:03PM +0900, Michael Paquier wrote:\\n> > Actually, no, this is not good. I have been studying more the patch,\\n> > and after stressing more this code path with a cluster having\\n> > checksums enabled and shared_buffers at 1MB, I have been able to make\\n> > a couple of page's LSNs go backwards with pgbench -s 100. The cause\\n> > was simply that the page got flushed with a newer LSN than what was\\n> > returned by XLogSaveBufferForHint() before taking the buffer header\\n> > lock, so updating only the LSN for a non-dirty page was simply\\n> > guarding against that.\\n\\nI thought of something like that but forgot to mention. But after\\nthinking of that, couldn't the current code can do the same think even\\nthough with a far small probability? Still a session with smaller hint\\nLSN but didn't entered the header lock section yet can be cut-in by\\nanother session with larger hint LSN.\\n\\n> for the reference attached is the trick I have used, adding an extra\\n> assertion check in PageSetLSN(). \\n\\nI believe that all locations where Page-LSN is set is in the same\\nbuffer-ex-lock section with XLogInsert.. but not sure.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Nov 2019 16:59:49 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Mon, Nov 11, 2019 at 10:03:14AM +0100, Antonin Houska wrote:\\n> > This looks good to me.\\n> \\n> Actually, no, this is not good. I have been studying more the patch,\\n> and after stressing more this code path with a cluster having\\n> checksums enabled and shared_buffers at 1MB, I have been able to make\\n> a couple of page's LSNs go backwards with pgbench -s 100. The cause\\n> was simply that the page got flushed with a newer LSN than what was\\n> returned by XLogSaveBufferForHint() before taking the buffer header\\n> lock, so updating only the LSN for a non-dirty page was simply\\n> guarding against that.\\n\\nInteresting. Now that I know about the problem, I could have reproduced it\\nusing gdb: MarkBufferDirtyHint() was called by 2 backends concurrently in such\\na way that the first backend generates the LSN, but before it manages to\\nassign it to the page, another backend generates another LSN and sets it.\\n\\nCan't we just apply the attached diff on the top of your patch?\\n\\nAlso I wonder how checksums helped you to discover the problem? Although I\\ncould simulate the setting of lower LSN, I could not see any broken\\nchecksum. And I wouldn't even expect that since FlushBuffer() copies the\\nbuffer into backend local memory before it calculates the checksum.\\n\\n-- \\nAntonin Houska\\nWeb: https://www.cybertec-postgresql.com\",\n \"msg_date\": \"Thu, 14 Nov 2019 15:48:31 +0100\",\n \"msg_from\": \"Antonin Houska <ah@cybertec.at>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"Antonin Houska <ah@cybertec.at> wrote:\\n\\n> Michael Paquier <michael@paquier.xyz> wrote:\\n> \\n> > On Mon, Nov 11, 2019 at 10:03:14AM +0100, Antonin Houska wrote:\\n> > > This looks good to me.\\n> > \\n> > Actually, no, this is not good. I have been studying more the patch,\\n> > and after stressing more this code path with a cluster having\\n> > checksums enabled and shared_buffers at 1MB, I have been able to make\\n> > a couple of page's LSNs go backwards with pgbench -s 100. The cause\\n> > was simply that the page got flushed with a newer LSN than what was\\n> > returned by XLogSaveBufferForHint() before taking the buffer header\\n> > lock, so updating only the LSN for a non-dirty page was simply\\n> > guarding against that.\\n> \\n> Interesting. Now that I know about the problem, I could have reproduced it\\n> using gdb: MarkBufferDirtyHint() was called by 2 backends concurrently in such\\n> a way that the first backend generates the LSN, but before it manages to\\n> assign it to the page, another backend generates another LSN and sets it.\\n> \\n> Can't we just apply the attached diff on the top of your patch?\\n\\nI wanted to register the patch for the next CF so it's not forgotten, but see\\nit's already there. Why have you set the status to \\\"withdrawn\\\"?\\n\\n-- \\nAntonin Houska\\nWeb: https://www.cybertec-postgresql.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 20 Dec 2019 16:30:38 +0100\",\n \"msg_from\": \"Antonin Houska <ah@cybertec.at>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 20, 2019 at 04:30:38PM +0100, Antonin Houska wrote:\\n> I wanted to register the patch for the next CF so it's not forgotten, but see\\n> it's already there. Why have you set the status to \\\"withdrawn\\\"?\\n\\nBecause my patch was incorrect, and I did not make enough bandwidth to\\nthink more on the matter. I am actually not sure that what you are\\nproposing is better.. If you wish to get that reviewed, could you add\\na new CF entry?\\n--\\nMichael\",\n \"msg_date\": \"Sat, 21 Dec 2019 11:44:25 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: MarkBufferDirtyHint() and LSN update\"\n }\n]","isConversational":false}}},{"rowIdx":49367,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"HAVE_LONG_LONG_INT is now implied by the requirement for C99, so the \nseparate Autoconf check can be removed. The uses are almost all in ecpg \ncode, and AFAICT the check was originally added specifically for ecpg.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Wed, 30 Oct 2019 14:49:38 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Remove HAVE_LONG_LONG_INT"},{"msg_contents":"On 2019-10-30 14:49, Peter Eisentraut wrote:\n> HAVE_LONG_LONG_INT is now implied by the requirement for C99, so the\n> separate Autoconf check can be removed. The uses are almost all in ecpg\n> code, and AFAICT the check was originally added specifically for ecpg.\n\ncommitted\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 7 Nov 2019 13:31:31 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Remove HAVE_LONG_LONG_INT"}],"string":"[\n {\n \"msg_contents\": \"HAVE_LONG_LONG_INT is now implied by the requirement for C99, so the \\nseparate Autoconf check can be removed. The uses are almost all in ecpg \\ncode, and AFAICT the check was originally added specifically for ecpg.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 30 Oct 2019 14:49:38 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Remove HAVE_LONG_LONG_INT\"\n },\n {\n \"msg_contents\": \"On 2019-10-30 14:49, Peter Eisentraut wrote:\\n> HAVE_LONG_LONG_INT is now implied by the requirement for C99, so the\\n> separate Autoconf check can be removed. The uses are almost all in ecpg\\n> code, and AFAICT the check was originally added specifically for ecpg.\\n\\ncommitted\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Nov 2019 13:31:31 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Remove HAVE_LONG_LONG_INT\"\n }\n]","isConversational":false}}},{"rowIdx":49368,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"While fooling with the NetBSD-vs-libpython issue noted in a nearby\nthread, I observed that the core regression tests sometimes hang up\nin the \"stats\" test on this platform (NetBSD 8.1/amd64). Investigation\nfound that the stats collector process was sometimes exiting like\nthis:\n\n2019-10-29 19:38:14.563 EDT [7018] FATAL: could not read statistics message: No buffer space available\n2019-10-29 19:38:14.563 EDT [7932] LOG: statistics collector process (PID 7018) exited with exit code 1\n\nThe postmaster then restarts the collector, but possibly with a time\ndelay (see the PGSTAT_RESTART_INTERVAL respawn throttling logic).\nThis seems to interact badly with the wait-for-stats-collector logic\nin backend_read_statsfile, so that each cycle of the wait_for_stats()\ntest function takes a long time ... and we will do 300 of those\nunconditionally. (Possibly wait_for_stats ought to be modified so\nthat it pays attention to elapsed wall-clock time rather than\niterating for a fixed number of times?)\n\nNetBSD's recv() man page glosses ENOBUFS as \"A message was not\ndelivered because it would have overflowed the buffer\", but I don't\nbelieve that's actually what's happening. (Just to be sure,\nI added an Assert on the sending side that no message exceeds\nsizeof(PgStat_Msg). I wonder why we didn't have one already.)\nTrawling the NetBSD kernel code, it seems like ENOBUFS could get\nreturned as a result of transient shortages of kernel working\nmemory --- most of the uses of that error code seem to be on the\nsending side, but I found some that seem to be in receiving code.\n\nIn short: it's evidently possible to get ENOBUFS as a transient\nfailure condition on this platform, and having the stats collector\ndie seems like an overreaction. I'm inclined to have it log the\nerror and press on, instead. Looking at the POSIX spec for\nrecv() suggests that ENOMEM is another plausible transient failure,\nso maybe we should do the same with that. Roughly:\n\n if (len < 0)\n {\n+ /* silently ignore these cases (no data available) */\n if (errno == EAGAIN || errno == EWOULDBLOCK || errno == EINTR)\n break; /* out of inner loop */\n+ /* noisily ignore these cases (soft errors) */\n+ if (errno == ENOBUFS || errno == ENOMEM)\n+ {\n+ ereport(LOG,\n+ (errcode_for_socket_access(),\n+ errmsg(\"could not read statistics message: %m\")));\n+ break; /* out of inner loop */\n+ }\n+ /* hard failure, but maybe hara-kiri will fix it */\n ereport(ERROR,\n (errcode_for_socket_access(),\n errmsg(\"could not read statistics message: %m\")));\n }\n\nA variant idea is to treat all unexpected errnos as LOG-and-push-on,\nbut maybe we ought to have a limit on how many times we'll do that\nbefore erroring out.\n\nThoughts?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 30 Oct 2019 15:58:34 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"pgstat.c has brittle response to transient problems"}],"string":"[\n {\n \"msg_contents\": \"While fooling with the NetBSD-vs-libpython issue noted in a nearby\\nthread, I observed that the core regression tests sometimes hang up\\nin the \\\"stats\\\" test on this platform (NetBSD 8.1/amd64). Investigation\\nfound that the stats collector process was sometimes exiting like\\nthis:\\n\\n2019-10-29 19:38:14.563 EDT [7018] FATAL: could not read statistics message: No buffer space available\\n2019-10-29 19:38:14.563 EDT [7932] LOG: statistics collector process (PID 7018) exited with exit code 1\\n\\nThe postmaster then restarts the collector, but possibly with a time\\ndelay (see the PGSTAT_RESTART_INTERVAL respawn throttling logic).\\nThis seems to interact badly with the wait-for-stats-collector logic\\nin backend_read_statsfile, so that each cycle of the wait_for_stats()\\ntest function takes a long time ... and we will do 300 of those\\nunconditionally. (Possibly wait_for_stats ought to be modified so\\nthat it pays attention to elapsed wall-clock time rather than\\niterating for a fixed number of times?)\\n\\nNetBSD's recv() man page glosses ENOBUFS as \\\"A message was not\\ndelivered because it would have overflowed the buffer\\\", but I don't\\nbelieve that's actually what's happening. (Just to be sure,\\nI added an Assert on the sending side that no message exceeds\\nsizeof(PgStat_Msg). I wonder why we didn't have one already.)\\nTrawling the NetBSD kernel code, it seems like ENOBUFS could get\\nreturned as a result of transient shortages of kernel working\\nmemory --- most of the uses of that error code seem to be on the\\nsending side, but I found some that seem to be in receiving code.\\n\\nIn short: it's evidently possible to get ENOBUFS as a transient\\nfailure condition on this platform, and having the stats collector\\ndie seems like an overreaction. I'm inclined to have it log the\\nerror and press on, instead. Looking at the POSIX spec for\\nrecv() suggests that ENOMEM is another plausible transient failure,\\nso maybe we should do the same with that. Roughly:\\n\\n if (len < 0)\\n {\\n+ /* silently ignore these cases (no data available) */\\n if (errno == EAGAIN || errno == EWOULDBLOCK || errno == EINTR)\\n break; /* out of inner loop */\\n+ /* noisily ignore these cases (soft errors) */\\n+ if (errno == ENOBUFS || errno == ENOMEM)\\n+ {\\n+ ereport(LOG,\\n+ (errcode_for_socket_access(),\\n+ errmsg(\\\"could not read statistics message: %m\\\")));\\n+ break; /* out of inner loop */\\n+ }\\n+ /* hard failure, but maybe hara-kiri will fix it */\\n ereport(ERROR,\\n (errcode_for_socket_access(),\\n errmsg(\\\"could not read statistics message: %m\\\")));\\n }\\n\\nA variant idea is to treat all unexpected errnos as LOG-and-push-on,\\nbut maybe we ought to have a limit on how many times we'll do that\\nbefore erroring out.\\n\\nThoughts?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Oct 2019 15:58:34 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pgstat.c has brittle response to transient problems\"\n }\n]","isConversational":false}}},{"rowIdx":49369,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nCurrently CREATE OR REPLACE VIEW command fails if the column names\nare changed. For example,\n\n =# CREATE VIEW test AS SELECT 0 AS a;\n =# CREATE OR REPLACE VIEW test AS SELECT 0 AS x;\n ERROR: cannot change name of view column \"a\" to \"x\"\n\nI'd like to propose the attached patch that allows CREATE OR REPLACE VIEW\nto rename the view columns. Thought?\n\nRegards,\n\n-- \nFujii Masao","msg_date":"Thu, 31 Oct 2019 11:27:12 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":true,"msg_subject":"Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"Fujii Masao <masao.fujii@gmail.com> writes:\n> Currently CREATE OR REPLACE VIEW command fails if the column names\n> are changed.\n\nThat is, I believe, intentional. It's an effective aid to catching\nmistakes in view redefinitions, such as misaligning the new set of\ncolumns relative to the old. That's particularly important given\nthat we allow you to add columns during CREATE OR REPLACE VIEW.\nConsider the oversimplified case where you start with\n\nCREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n\nand you want to add a column z, and you get sloppy and write\n\nCREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n\nIf we did not throw an error on this, references that formerly\npointed to column y would now point to z (as that's still attnum 2),\nwhich is highly unlikely to be what you wanted.\n\nThe right way to handle a column rename in a view is to do a separate\nALTER VIEW RENAME COLUMN, making it totally clear what you intend to\nhappen. (Right now, we make you spell that \"ALTER TABLE RENAME COLUMN\",\nbut it'd be reasonable to add that syntax to ALTER VIEW too.) I don't\nthink this functionality should be folded into redefinition of the content\nof the view. It'd add more opportunity for mistakes than anything else.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 31 Oct 2019 00:42:49 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Fujii Masao <masao.fujii@gmail.com> writes:\n> > Currently CREATE OR REPLACE VIEW command fails if the column names\n> > are changed.\n>\n> That is, I believe, intentional. It's an effective aid to catching\n> mistakes in view redefinitions, such as misaligning the new set of\n> columns relative to the old. That's particularly important given\n> that we allow you to add columns during CREATE OR REPLACE VIEW.\n> Consider the oversimplified case where you start with\n>\n> CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n>\n> and you want to add a column z, and you get sloppy and write\n>\n> CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n>\n> If we did not throw an error on this, references that formerly\n> pointed to column y would now point to z (as that's still attnum 2),\n> which is highly unlikely to be what you wanted.\n\nThis example makes me wonder if the addtion of column by\nCREATE OR REPLACE VIEW also has the same (or even worse) issue.\nThat is, it may increase the oppotunity for users' mistake.\nI'm thinking the case where users mistakenly added new column\ninto the view when replacing the view definition. This mistake can\nhappen because CREATE OR REPLACE VIEW allows new column to\nbe added. But what's the worse is that, currently there is no way to\ndrop the column from the view, except recreation of the view.\nNeither CREATE OR REPLACE VIEW nor ALTER TABLE support\nthe drop of the column from the view. So, to fix the mistake,\nusers would need to drop the view itself and recreate it. If there are\nsome objects depending the view, they also might need to be recreated.\nThis looks not good. Since the feature has been supported,\nit's too late to say that, though...\n\nAt least, the support for ALTER VIEW DROP COLUMN might be\nnecessary to alleviate that situation.\n\nRegards,\n\n-- \nFujii Masao\n\n\n","msg_date":"Thu, 31 Oct 2019 16:31:52 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":true,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> wrote:\n\n> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >\n> > Fujii Masao <masao.fujii@gmail.com> writes:\n> > > Currently CREATE OR REPLACE VIEW command fails if the column names\n> > > are changed.\n> >\n> > That is, I believe, intentional. It's an effective aid to catching\n> > mistakes in view redefinitions, such as misaligning the new set of\n> > columns relative to the old. That's particularly important given\n> > that we allow you to add columns during CREATE OR REPLACE VIEW.\n> > Consider the oversimplified case where you start with\n> >\n> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n> >\n> > and you want to add a column z, and you get sloppy and write\n> >\n> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n> >\n> > If we did not throw an error on this, references that formerly\n> > pointed to column y would now point to z (as that's still attnum 2),\n> > which is highly unlikely to be what you wanted.\n>\n> This example makes me wonder if the addtion of column by\n> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n> That is, it may increase the oppotunity for users' mistake.\n> I'm thinking the case where users mistakenly added new column\n> into the view when replacing the view definition. This mistake can\n> happen because CREATE OR REPLACE VIEW allows new column to\n> be added. But what's the worse is that, currently there is no way to\n> drop the column from the view, except recreation of the view.\n> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\n> the drop of the column from the view. So, to fix the mistake,\n> users would need to drop the view itself and recreate it. If there are\n> some objects depending the view, they also might need to be recreated.\n> This looks not good. Since the feature has been supported,\n> it's too late to say that, though...\n>\n> At least, the support for ALTER VIEW DROP COLUMN might be\n> necessary to alleviate that situation.\n>\n>\n- Is this intentional not implemented the \"RENAME COLUMN\" statement for\nVIEW because it is implemented for Materialized View? I have made just a\nsimilar\nchange to view and it works.\n\nALTER VIEW v RENAME COLUMN d to e;\n\n- For \"DROP COLUMN\" for VIEW is throwing error.\n\npostgres=# alter view v drop column e;\nERROR: \"v\" is not a table, composite type, or foreign table\n\n\n\nRegards,\n>\n> --\n> Fujii Masao\n>\n>\n>\n\n-- \nIbrar Ahmed","msg_date":"Thu, 31 Oct 2019 15:58:49 +0500","msg_from":"Ibrar Ahmed <ibrar.ahmad@gmail.com>","msg_from_op":false,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n>\n>\n>\n> On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> wrote:\n>>\n>> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> >\n>> > Fujii Masao <masao.fujii@gmail.com> writes:\n>> > > Currently CREATE OR REPLACE VIEW command fails if the column names\n>> > > are changed.\n>> >\n>> > That is, I believe, intentional. It's an effective aid to catching\n>> > mistakes in view redefinitions, such as misaligning the new set of\n>> > columns relative to the old. That's particularly important given\n>> > that we allow you to add columns during CREATE OR REPLACE VIEW.\n>> > Consider the oversimplified case where you start with\n>> >\n>> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n>> >\n>> > and you want to add a column z, and you get sloppy and write\n>> >\n>> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n>> >\n>> > If we did not throw an error on this, references that formerly\n>> > pointed to column y would now point to z (as that's still attnum 2),\n>> > which is highly unlikely to be what you wanted.\n>>\n>> This example makes me wonder if the addtion of column by\n>> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n>> That is, it may increase the oppotunity for users' mistake.\n>> I'm thinking the case where users mistakenly added new column\n>> into the view when replacing the view definition. This mistake can\n>> happen because CREATE OR REPLACE VIEW allows new column to\n>> be added. But what's the worse is that, currently there is no way to\n>> drop the column from the view, except recreation of the view.\n>> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\n>> the drop of the column from the view. So, to fix the mistake,\n>> users would need to drop the view itself and recreate it. If there are\n>> some objects depending the view, they also might need to be recreated.\n>> This looks not good. Since the feature has been supported,\n>> it's too late to say that, though...\n>>\n>> At least, the support for ALTER VIEW DROP COLUMN might be\n>> necessary to alleviate that situation.\n>>\n>\n> - Is this intentional not implemented the \"RENAME COLUMN\" statement for\n> VIEW because it is implemented for Materialized View?\n\nNot sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\nsounds reasonable whether we support the rename of columns when replacing\nthe view definition, or not. Attached is the patch that adds support for\nALTER VIEW RENAME COLUMN command.\n\n> I have made just a similar\n> change to view and it works.\n\nYeah, ISTM that we made the same patch at the same time. You changed gram.y,\nbut I added the following changes additionally.\n\n- Update the doc\n- Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the columns\n- Update tab-complete.c\n- Add regression test\n\nOne issue I've not addressed yet is about the command tag of\n\"ALTER VIEW RENAME COLUMN\". Currently \"ALTER TABLE\" is returned as the tag\nlike ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \"ALTER VIEW\"\nis better. I started the discussion about the command tag of\n\"ALTER MATERIALIZED VIEW\" at another thread. I will update the patch according\nto the result of that discussion.\nhttps://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\n\nRegards,\n\n-- \nFujii Masao","msg_date":"Thu, 31 Oct 2019 21:01:02 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":true,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com> wrote:\n\n> On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n> >\n> >\n> >\n> > On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com>\n> wrote:\n> >>\n> >> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >> >\n> >> > Fujii Masao <masao.fujii@gmail.com> writes:\n> >> > > Currently CREATE OR REPLACE VIEW command fails if the column names\n> >> > > are changed.\n> >> >\n> >> > That is, I believe, intentional. It's an effective aid to catching\n> >> > mistakes in view redefinitions, such as misaligning the new set of\n> >> > columns relative to the old. That's particularly important given\n> >> > that we allow you to add columns during CREATE OR REPLACE VIEW.\n> >> > Consider the oversimplified case where you start with\n> >> >\n> >> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n> >> >\n> >> > and you want to add a column z, and you get sloppy and write\n> >> >\n> >> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n> >> >\n> >> > If we did not throw an error on this, references that formerly\n> >> > pointed to column y would now point to z (as that's still attnum 2),\n> >> > which is highly unlikely to be what you wanted.\n> >>\n> >> This example makes me wonder if the addtion of column by\n> >> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n> >> That is, it may increase the oppotunity for users' mistake.\n> >> I'm thinking the case where users mistakenly added new column\n> >> into the view when replacing the view definition. This mistake can\n> >> happen because CREATE OR REPLACE VIEW allows new column to\n> >> be added. But what's the worse is that, currently there is no way to\n> >> drop the column from the view, except recreation of the view.\n> >> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\n> >> the drop of the column from the view. So, to fix the mistake,\n> >> users would need to drop the view itself and recreate it. If there are\n> >> some objects depending the view, they also might need to be recreated.\n> >> This looks not good. Since the feature has been supported,\n> >> it's too late to say that, though...\n> >>\n> >> At least, the support for ALTER VIEW DROP COLUMN might be\n> >> necessary to alleviate that situation.\n> >>\n> >\n> > - Is this intentional not implemented the \"RENAME COLUMN\" statement for\n> > VIEW because it is implemented for Materialized View?\n>\n> Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\n> sounds reasonable whether we support the rename of columns when replacing\n> the view definition, or not. Attached is the patch that adds support for\n> ALTER VIEW RENAME COLUMN command.\n>\n> > I have made just a similar\n> > change to view and it works.\n>\n> Yeah, ISTM that we made the same patch at the same time. You changed\n> gram.y,\n> but I added the following changes additionally.\n>\n> - Update the doc\n> - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the\n> columns\n> - Update tab-complete.c\n> - Add regression test\n>\n>\nOh, I just sent the patch to ask, good you do that in all the places.\n\nOne issue I've not addressed yet is about the command tag of\n> \"ALTER VIEW RENAME COLUMN\". Currently \"ALTER TABLE\" is returned as the tag\n> like ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \"ALTER VIEW\"\n> is better. I started the discussion about the command tag of\n> \"ALTER MATERIALIZED VIEW\" at another thread. I will update the patch\n> according\n> to the result of that discussion.\n>\n> https://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\n>\n> Attached patch contain small change for ALTER MATERIALIZED VIEW.\n\n\n\n> Regards,\n>\n> --\n> Fujii Masao\n>\n\n\n-- \nIbrar Ahmed","msg_date":"Thu, 31 Oct 2019 17:11:57 +0500","msg_from":"Ibrar Ahmed <ibrar.ahmad@gmail.com>","msg_from_op":false,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"On Thu, Oct 31, 2019 at 5:11 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n\n>\n>\n> On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com> wrote:\n>\n>> On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com>\n>> wrote:\n>> >\n>> >\n>> >\n>> > On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com>\n>> wrote:\n>> >>\n>> >> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> >> >\n>> >> > Fujii Masao <masao.fujii@gmail.com> writes:\n>> >> > > Currently CREATE OR REPLACE VIEW command fails if the column names\n>> >> > > are changed.\n>> >> >\n>> >> > That is, I believe, intentional. It's an effective aid to catching\n>> >> > mistakes in view redefinitions, such as misaligning the new set of\n>> >> > columns relative to the old. That's particularly important given\n>> >> > that we allow you to add columns during CREATE OR REPLACE VIEW.\n>> >> > Consider the oversimplified case where you start with\n>> >> >\n>> >> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n>> >> >\n>> >> > and you want to add a column z, and you get sloppy and write\n>> >> >\n>> >> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n>> >> >\n>> >> > If we did not throw an error on this, references that formerly\n>> >> > pointed to column y would now point to z (as that's still attnum 2),\n>> >> > which is highly unlikely to be what you wanted.\n>> >>\n>> >> This example makes me wonder if the addtion of column by\n>> >> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n>> >> That is, it may increase the oppotunity for users' mistake.\n>> >> I'm thinking the case where users mistakenly added new column\n>> >> into the view when replacing the view definition. This mistake can\n>> >> happen because CREATE OR REPLACE VIEW allows new column to\n>> >> be added. But what's the worse is that, currently there is no way to\n>> >> drop the column from the view, except recreation of the view.\n>> >> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\n>> >> the drop of the column from the view. So, to fix the mistake,\n>> >> users would need to drop the view itself and recreate it. If there are\n>> >> some objects depending the view, they also might need to be recreated.\n>> >> This looks not good. Since the feature has been supported,\n>> >> it's too late to say that, though...\n>> >>\n>> >> At least, the support for ALTER VIEW DROP COLUMN might be\n>> >> necessary to alleviate that situation.\n>> >>\n>> >\n>> > - Is this intentional not implemented the \"RENAME COLUMN\" statement for\n>> > VIEW because it is implemented for Materialized View?\n>>\n>> Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\n>> sounds reasonable whether we support the rename of columns when replacing\n>> the view definition, or not. Attached is the patch that adds support for\n>> ALTER VIEW RENAME COLUMN command.\n>>\n>> > I have made just a similar\n>> > change to view and it works.\n>>\n>> Yeah, ISTM that we made the same patch at the same time. You changed\n>> gram.y,\n>> but I added the following changes additionally.\n>>\n>> - Update the doc\n>> - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the\n>> columns\n>> - Update tab-complete.c\n>> - Add regression test\n>>\n>>\n> Oh, I just sent the patch to ask, good you do that in all the places.\n>\n> One issue I've not addressed yet is about the command tag of\n>> \"ALTER VIEW RENAME COLUMN\". Currently \"ALTER TABLE\" is returned as the tag\n>> like ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \"ALTER VIEW\"\n>> is better. I started the discussion about the command tag of\n>> \"ALTER MATERIALIZED VIEW\" at another thread. I will update the patch\n>> according\n>> to the result of that discussion.\n>>\n>> https://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\n>>\n>> Attached patch contain small change for ALTER MATERIALIZED VIEW.\n>\n>\nHmm, my small change of \"ALTER MATERIALIZED VIEW\" does not work in some\ncases need more work on that.\n\n\n>\n>\n>> Regards,\n>>\n>> --\n>> Fujii Masao\n>>\n>\n>\n> --\n> Ibrar Ahmed\n>\n\n\n-- \nIbrar Ahmed\n\nOn Thu, Oct 31, 2019 at 5:11 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com> wrote:On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n>\n>\n>\n> On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> wrote:\n>>\n>> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> >\n>> > Fujii Masao <masao.fujii@gmail.com> writes:\n>> > > Currently CREATE OR REPLACE VIEW command fails if the column names\n>> > > are changed.\n>> >\n>> > That is, I believe, intentional. It's an effective aid to catching\n>> > mistakes in view redefinitions, such as misaligning the new set of\n>> > columns relative to the old. That's particularly important given\n>> > that we allow you to add columns during CREATE OR REPLACE VIEW.\n>> > Consider the oversimplified case where you start with\n>> >\n>> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n>> >\n>> > and you want to add a column z, and you get sloppy and write\n>> >\n>> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n>> >\n>> > If we did not throw an error on this, references that formerly\n>> > pointed to column y would now point to z (as that's still attnum 2),\n>> > which is highly unlikely to be what you wanted.\n>>\n>> This example makes me wonder if the addtion of column by\n>> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n>> That is, it may increase the oppotunity for users' mistake.\n>> I'm thinking the case where users mistakenly added new column\n>> into the view when replacing the view definition. This mistake can\n>> happen because CREATE OR REPLACE VIEW allows new column to\n>> be added. But what's the worse is that, currently there is no way to\n>> drop the column from the view, except recreation of the view.\n>> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\n>> the drop of the column from the view. So, to fix the mistake,\n>> users would need to drop the view itself and recreate it. If there are\n>> some objects depending the view, they also might need to be recreated.\n>> This looks not good. Since the feature has been supported,\n>> it's too late to say that, though...\n>>\n>> At least, the support for ALTER VIEW DROP COLUMN might be\n>> necessary to alleviate that situation.\n>>\n>\n> - Is this intentional not implemented the \"RENAME COLUMN\" statement for\n> VIEW because it is implemented for Materialized View?\n\nNot sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\nsounds reasonable whether we support the rename of columns when replacing\nthe view definition, or not. Attached is the patch that adds support for\nALTER VIEW RENAME COLUMN command.\n\n> I have made just a similar\n> change to view and it works.\n\nYeah, ISTM that we made the same patch at the same time. You changed gram.y,\nbut I added the following changes additionally.\n\n- Update the doc\n- Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the columns\n- Update tab-complete.c\n- Add regression test\nOh, I just sent the patch to ask, good you do that in all the places. \nOne issue I've not addressed yet is about the command tag of\n\"ALTER VIEW RENAME COLUMN\". Currently \"ALTER TABLE\" is returned as the tag\nlike ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \"ALTER VIEW\"\nis better. I started the discussion about the command tag of\n\"ALTER MATERIALIZED VIEW\" at another thread. I will update the patch according\nto the result of that discussion.\nhttps://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\nAttached patch contain small change for ALTER MATERIALIZED VIEW.Hmm, my small change of \"ALTER MATERIALIZED VIEW\" does not work in some cases need more work on that. \nRegards,\n\n-- \nFujii Masao\n-- Ibrar Ahmed\n-- Ibrar Ahmed","msg_date":"Thu, 31 Oct 2019 17:28:00 +0500","msg_from":"Ibrar Ahmed <ibrar.ahmad@gmail.com>","msg_from_op":false,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"On Thu, Oct 31, 2019 at 5:28 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n\n>\n>\n> On Thu, Oct 31, 2019 at 5:11 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n>\n>>\n>>\n>> On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com>\n>> wrote:\n>>\n>>> On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com>\n>>> wrote:\n>>> >\n>>> >\n>>> >\n>>> > On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com>\n>>> wrote:\n>>> >>\n>>> >> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>> >> >\n>>> >> > Fujii Masao <masao.fujii@gmail.com> writes:\n>>> >> > > Currently CREATE OR REPLACE VIEW command fails if the column names\n>>> >> > > are changed.\n>>> >> >\n>>> >> > That is, I believe, intentional. It's an effective aid to catching\n>>> >> > mistakes in view redefinitions, such as misaligning the new set of\n>>> >> > columns relative to the old. That's particularly important given\n>>> >> > that we allow you to add columns during CREATE OR REPLACE VIEW.\n>>> >> > Consider the oversimplified case where you start with\n>>> >> >\n>>> >> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n>>> >> >\n>>> >> > and you want to add a column z, and you get sloppy and write\n>>> >> >\n>>> >> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n>>> >> >\n>>> >> > If we did not throw an error on this, references that formerly\n>>> >> > pointed to column y would now point to z (as that's still attnum 2),\n>>> >> > which is highly unlikely to be what you wanted.\n>>> >>\n>>> >> This example makes me wonder if the addtion of column by\n>>> >> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n>>> >> That is, it may increase the oppotunity for users' mistake.\n>>> >> I'm thinking the case where users mistakenly added new column\n>>> >> into the view when replacing the view definition. This mistake can\n>>> >> happen because CREATE OR REPLACE VIEW allows new column to\n>>> >> be added. But what's the worse is that, currently there is no way to\n>>> >> drop the column from the view, except recreation of the view.\n>>> >> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\n>>> >> the drop of the column from the view. So, to fix the mistake,\n>>> >> users would need to drop the view itself and recreate it. If there are\n>>> >> some objects depending the view, they also might need to be recreated.\n>>> >> This looks not good. Since the feature has been supported,\n>>> >> it's too late to say that, though...\n>>> >>\n>>> >> At least, the support for ALTER VIEW DROP COLUMN might be\n>>> >> necessary to alleviate that situation.\n>>> >>\n>>> >\n>>> > - Is this intentional not implemented the \"RENAME COLUMN\" statement for\n>>> > VIEW because it is implemented for Materialized View?\n>>>\n>>> Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\n>>> sounds reasonable whether we support the rename of columns when replacing\n>>> the view definition, or not. Attached is the patch that adds support for\n>>> ALTER VIEW RENAME COLUMN command.\n>>>\n>>> > I have made just a similar\n>>> > change to view and it works.\n>>>\n>>> Yeah, ISTM that we made the same patch at the same time. You changed\n>>> gram.y,\n>>> but I added the following changes additionally.\n>>>\n>>> - Update the doc\n>>> - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the\n>>> columns\n>>> - Update tab-complete.c\n>>> - Add regression test\n>>>\n>>>\n>> Oh, I just sent the patch to ask, good you do that in all the places.\n>>\n>> One issue I've not addressed yet is about the command tag of\n>>> \"ALTER VIEW RENAME COLUMN\". Currently \"ALTER TABLE\" is returned as the\n>>> tag\n>>> like ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \"ALTER VIEW\"\n>>> is better. I started the discussion about the command tag of\n>>> \"ALTER MATERIALIZED VIEW\" at another thread. I will update the patch\n>>> according\n>>> to the result of that discussion.\n>>>\n>>> https://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\n>>>\n>>> Attached patch contain small change for ALTER MATERIALIZED VIEW.\n>>\n>>\n> Hmm, my small change of \"ALTER MATERIALIZED VIEW\" does not work in some\n> cases need more work on that.\n>\n>\n\nThe AlterObjectTypeCommandTag function just take one parameter, but to\nshow \"ALTER MATERIALIZED VIEW\" instead of ALTER TABLE we need to\npass \"relationType = OBJECT_MATVIEW\" along with \"renameType = OBJECT_COLUMN\"\nand handle that in the function. The \"AlterObjectTypeCommandTag\" function\nhas many\ncalls. Do you think just for the command tag, we should do all this change?\n\n\n>\n>>\n>>> Regards,\n>>>\n>>> --\n>>> Fujii Masao\n>>>\n>>\n>>\n>> --\n>> Ibrar Ahmed\n>>\n>\n>\n> --\n> Ibrar Ahmed\n>\n\n\n-- \nIbrar Ahmed\n\nOn Thu, Oct 31, 2019 at 5:28 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:On Thu, Oct 31, 2019 at 5:11 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com> wrote:On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n>\n>\n>\n> On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> wrote:\n>>\n>> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> >\n>> > Fujii Masao <masao.fujii@gmail.com> writes:\n>> > > Currently CREATE OR REPLACE VIEW command fails if the column names\n>> > > are changed.\n>> >\n>> > That is, I believe, intentional. It's an effective aid to catching\n>> > mistakes in view redefinitions, such as misaligning the new set of\n>> > columns relative to the old. That's particularly important given\n>> > that we allow you to add columns during CREATE OR REPLACE VIEW.\n>> > Consider the oversimplified case where you start with\n>> >\n>> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n>> >\n>> > and you want to add a column z, and you get sloppy and write\n>> >\n>> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n>> >\n>> > If we did not throw an error on this, references that formerly\n>> > pointed to column y would now point to z (as that's still attnum 2),\n>> > which is highly unlikely to be what you wanted.\n>>\n>> This example makes me wonder if the addtion of column by\n>> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n>> That is, it may increase the oppotunity for users' mistake.\n>> I'm thinking the case where users mistakenly added new column\n>> into the view when replacing the view definition. This mistake can\n>> happen because CREATE OR REPLACE VIEW allows new column to\n>> be added. But what's the worse is that, currently there is no way to\n>> drop the column from the view, except recreation of the view.\n>> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\n>> the drop of the column from the view. So, to fix the mistake,\n>> users would need to drop the view itself and recreate it. If there are\n>> some objects depending the view, they also might need to be recreated.\n>> This looks not good. Since the feature has been supported,\n>> it's too late to say that, though...\n>>\n>> At least, the support for ALTER VIEW DROP COLUMN might be\n>> necessary to alleviate that situation.\n>>\n>\n> - Is this intentional not implemented the \"RENAME COLUMN\" statement for\n> VIEW because it is implemented for Materialized View?\n\nNot sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\nsounds reasonable whether we support the rename of columns when replacing\nthe view definition, or not. Attached is the patch that adds support for\nALTER VIEW RENAME COLUMN command.\n\n> I have made just a similar\n> change to view and it works.\n\nYeah, ISTM that we made the same patch at the same time. You changed gram.y,\nbut I added the following changes additionally.\n\n- Update the doc\n- Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the columns\n- Update tab-complete.c\n- Add regression test\nOh, I just sent the patch to ask, good you do that in all the places. \nOne issue I've not addressed yet is about the command tag of\n\"ALTER VIEW RENAME COLUMN\". Currently \"ALTER TABLE\" is returned as the tag\nlike ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \"ALTER VIEW\"\nis better. I started the discussion about the command tag of\n\"ALTER MATERIALIZED VIEW\" at another thread. I will update the patch according\nto the result of that discussion.\nhttps://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\nAttached patch contain small change for ALTER MATERIALIZED VIEW.Hmm, my small change of \"ALTER MATERIALIZED VIEW\" does not work in some cases need more work on that. The AlterObjectTypeCommandTag function just take one parameter, but toshow \"ALTER MATERIALIZED VIEW\" instead of ALTER TABLE we need topass \"relationType = OBJECT_MATVIEW\" along with \"renameType = OBJECT_COLUMN\"and handle that in the function. The \"AlterObjectTypeCommandTag\" function has manycalls. Do you think just for the command tag, we should do all this change? \n \nRegards,\n\n-- \nFujii Masao\n-- Ibrar Ahmed\n-- Ibrar Ahmed\n-- Ibrar Ahmed","msg_date":"Thu, 31 Oct 2019 17:34:22 +0500","msg_from":"Ibrar Ahmed <ibrar.ahmad@gmail.com>","msg_from_op":false,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"Fujii Masao <masao.fujii@gmail.com> writes:\n> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> Fujii Masao <masao.fujii@gmail.com> writes:\n>>> Currently CREATE OR REPLACE VIEW command fails if the column names\n>>> are changed.\n\n>> That is, I believe, intentional. It's an effective aid to catching\n>> mistakes in view redefinitions, such as misaligning the new set of\n>> columns relative to the old. [example]\n\n> This example makes me wonder if the addtion of column by\n> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n> That is, it may increase the oppotunity for users' mistake.\n\nThe idea in CREATE OR REPLACE VIEW is to allow addition of new\ncolumns at the end, the same as you can do with tables. Checking\nthe column name matchups is a way to ensure that you actually do\nadd at the end, rather than insert, which wouldn't act as you\nexpect. Admittedly it's only heuristic.\n\nWe could, perhaps, have insisted that adding a column also requires\nspecial syntax, but we didn't. Consider for example a case where\nthe new column needs to come from an additionally joined table;\nthen you have to be able to edit the underlying view definition along\nwith adding the column. So that seems like kind of a pain in the\nneck to insist on.\n\n> But what's the worse is that, currently there is no way to\n> drop the column from the view, except recreation of the view.\n\nI think this has been discussed, as well. It's not necessarily\nsimple to drop a view output column. For example, if the view\nuses SELECT DISTINCT, removing an output column would have\nsemantic effects on the set of rows that can be returned, since\ndistinct-ness would now mean something else than it did before.\n\nIt's conceivable that we could enumerate all such effects and\nthen allow DROP COLUMN (probably replacing the output column\nwith a null constant) if none of them apply, but I can't get\nterribly excited about it. The field demand for such a feature\nhas not been high. I'd be a bit worried about bugs arising\nfrom failures to check attisdropped for views, too; so that\nthe cost of getting this working might be greater than it seems\non the surface.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 31 Oct 2019 09:54:24 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"On Thu, Oct 31, 2019 at 10:54 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Fujii Masao <masao.fujii@gmail.com> writes:\n> > On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >> Fujii Masao <masao.fujii@gmail.com> writes:\n> >>> Currently CREATE OR REPLACE VIEW command fails if the column names\n> >>> are changed.\n>\n> >> That is, I believe, intentional. It's an effective aid to catching\n> >> mistakes in view redefinitions, such as misaligning the new set of\n> >> columns relative to the old. [example]\n>\n> > This example makes me wonder if the addtion of column by\n> > CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n> > That is, it may increase the oppotunity for users' mistake.\n>\n> The idea in CREATE OR REPLACE VIEW is to allow addition of new\n> columns at the end, the same as you can do with tables. Checking\n> the column name matchups is a way to ensure that you actually do\n> add at the end, rather than insert, which wouldn't act as you\n> expect. Admittedly it's only heuristic.\n>\n> We could, perhaps, have insisted that adding a column also requires\n> special syntax, but we didn't. Consider for example a case where\n> the new column needs to come from an additionally joined table;\n> then you have to be able to edit the underlying view definition along\n> with adding the column. So that seems like kind of a pain in the\n> neck to insist on.\n>\n> > But what's the worse is that, currently there is no way to\n> > drop the column from the view, except recreation of the view.\n>\n> I think this has been discussed, as well. It's not necessarily\n> simple to drop a view output column. For example, if the view\n> uses SELECT DISTINCT, removing an output column would have\n> semantic effects on the set of rows that can be returned, since\n> distinct-ness would now mean something else than it did before.\n>\n> It's conceivable that we could enumerate all such effects and\n> then allow DROP COLUMN (probably replacing the output column\n> with a null constant) if none of them apply, but I can't get\n> terribly excited about it. The field demand for such a feature\n> has not been high. I'd be a bit worried about bugs arising\n> from failures to check attisdropped for views, too; so that\n> the cost of getting this working might be greater than it seems\n> on the surface.\n\nThanks for the explanation! Understood.\n\nRegards,\n\n-- \nFujii Masao\n\n\n","msg_date":"Fri, 1 Nov 2019 12:42:43 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":true,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"2019-10-31 21:01 に Fujii Masao さんは書きました:\n> On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> \n> wrote:\n>> \n>> \n>> \n>> On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> \n>> wrote:\n>>> \n>>> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>> >\n>>> > Fujii Masao <masao.fujii@gmail.com> writes:\n>>> > > Currently CREATE OR REPLACE VIEW command fails if the column names\n>>> > > are changed.\n>>> >\n>>> > That is, I believe, intentional. It's an effective aid to catching\n>>> > mistakes in view redefinitions, such as misaligning the new set of\n>>> > columns relative to the old. That's particularly important given\n>>> > that we allow you to add columns during CREATE OR REPLACE VIEW.\n>>> > Consider the oversimplified case where you start with\n>>> >\n>>> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n>>> >\n>>> > and you want to add a column z, and you get sloppy and write\n>>> >\n>>> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n>>> >\n>>> > If we did not throw an error on this, references that formerly\n>>> > pointed to column y would now point to z (as that's still attnum 2),\n>>> > which is highly unlikely to be what you wanted.\n>>> \n>>> This example makes me wonder if the addtion of column by\n>>> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n>>> That is, it may increase the oppotunity for users' mistake.\n>>> I'm thinking the case where users mistakenly added new column\n>>> into the view when replacing the view definition. This mistake can\n>>> happen because CREATE OR REPLACE VIEW allows new column to\n>>> be added. But what's the worse is that, currently there is no way to\n>>> drop the column from the view, except recreation of the view.\n>>> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\n>>> the drop of the column from the view. So, to fix the mistake,\n>>> users would need to drop the view itself and recreate it. If there \n>>> are\n>>> some objects depending the view, they also might need to be \n>>> recreated.\n>>> This looks not good. Since the feature has been supported,\n>>> it's too late to say that, though...\n>>> \n>>> At least, the support for ALTER VIEW DROP COLUMN might be\n>>> necessary to alleviate that situation.\n>>> \n>> \n>> - Is this intentional not implemented the \"RENAME COLUMN\" statement \n>> for\n>> VIEW because it is implemented for Materialized View?\n> \n> Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\n> sounds reasonable whether we support the rename of columns when \n> replacing\n> the view definition, or not. Attached is the patch that adds support \n> for\n> ALTER VIEW RENAME COLUMN command.\n> \n>> I have made just a similar\n>> change to view and it works.\n> \n> Yeah, ISTM that we made the same patch at the same time. You changed \n> gram.y,\n> but I added the following changes additionally.\n> \n> - Update the doc\n> - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the \n> columns\n> - Update tab-complete.c\n\n\nI review your patch, and then I found that tab complete of \"alter \nmaterialized view\" is also not enough.\nSo, I made a small patch referencing your patch.\n\nRegards,","msg_date":"Wed, 06 Nov 2019 16:14:09 +0900","msg_from":"btfujiitkp <btfujiitkp@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"On Thu, Oct 31, 2019 at 9:34 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n>\n>\n>\n> On Thu, Oct 31, 2019 at 5:28 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n>>\n>>\n>>\n>> On Thu, Oct 31, 2019 at 5:11 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n>>>\n>>>\n>>>\n>>> On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com> wrote:\n>>>>\n>>>> On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n>>>> >\n>>>> >\n>>>> >\n>>>> > On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> wrote:\n>>>> >>\n>>>> >> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>>> >> >\n>>>> >> > Fujii Masao <masao.fujii@gmail.com> writes:\n>>>> >> > > Currently CREATE OR REPLACE VIEW command fails if the column names\n>>>> >> > > are changed.\n>>>> >> >\n>>>> >> > That is, I believe, intentional. It's an effective aid to catching\n>>>> >> > mistakes in view redefinitions, such as misaligning the new set of\n>>>> >> > columns relative to the old. That's particularly important given\n>>>> >> > that we allow you to add columns during CREATE OR REPLACE VIEW.\n>>>> >> > Consider the oversimplified case where you start with\n>>>> >> >\n>>>> >> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n>>>> >> >\n>>>> >> > and you want to add a column z, and you get sloppy and write\n>>>> >> >\n>>>> >> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n>>>> >> >\n>>>> >> > If we did not throw an error on this, references that formerly\n>>>> >> > pointed to column y would now point to z (as that's still attnum 2),\n>>>> >> > which is highly unlikely to be what you wanted.\n>>>> >>\n>>>> >> This example makes me wonder if the addtion of column by\n>>>> >> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n>>>> >> That is, it may increase the oppotunity for users' mistake.\n>>>> >> I'm thinking the case where users mistakenly added new column\n>>>> >> into the view when replacing the view definition. This mistake can\n>>>> >> happen because CREATE OR REPLACE VIEW allows new column to\n>>>> >> be added. But what's the worse is that, currently there is no way to\n>>>> >> drop the column from the view, except recreation of the view.\n>>>> >> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\n>>>> >> the drop of the column from the view. So, to fix the mistake,\n>>>> >> users would need to drop the view itself and recreate it. If there are\n>>>> >> some objects depending the view, they also might need to be recreated.\n>>>> >> This looks not good. Since the feature has been supported,\n>>>> >> it's too late to say that, though...\n>>>> >>\n>>>> >> At least, the support for ALTER VIEW DROP COLUMN might be\n>>>> >> necessary to alleviate that situation.\n>>>> >>\n>>>> >\n>>>> > - Is this intentional not implemented the \"RENAME COLUMN\" statement for\n>>>> > VIEW because it is implemented for Materialized View?\n>>>>\n>>>> Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\n>>>> sounds reasonable whether we support the rename of columns when replacing\n>>>> the view definition, or not. Attached is the patch that adds support for\n>>>> ALTER VIEW RENAME COLUMN command.\n>>>>\n>>>> > I have made just a similar\n>>>> > change to view and it works.\n>>>>\n>>>> Yeah, ISTM that we made the same patch at the same time. You changed gram.y,\n>>>> but I added the following changes additionally.\n>>>>\n>>>> - Update the doc\n>>>> - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the columns\n>>>> - Update tab-complete.c\n>>>> - Add regression test\n>>>>\n>>>\n>>> Oh, I just sent the patch to ask, good you do that in all the places.\n>>>\n>>>> One issue I've not addressed yet is about the command tag of\n>>>> \"ALTER VIEW RENAME COLUMN\". Currently \"ALTER TABLE\" is returned as the tag\n>>>> like ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \"ALTER VIEW\"\n>>>> is better. I started the discussion about the command tag of\n>>>> \"ALTER MATERIALIZED VIEW\" at another thread. I will update the patch according\n>>>> to the result of that discussion.\n>>>> https://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\n>>>>\n>>> Attached patch contain small change for ALTER MATERIALIZED VIEW.\n>>>\n>>\n>> Hmm, my small change of \"ALTER MATERIALIZED VIEW\" does not work in some cases need more work on that.\n>>\n>\n>\n> The AlterObjectTypeCommandTag function just take one parameter, but to\n> show \"ALTER MATERIALIZED VIEW\" instead of ALTER TABLE we need to\n> pass \"relationType = OBJECT_MATVIEW\" along with \"renameType = OBJECT_COLUMN\"\n> and handle that in the function. The \"AlterObjectTypeCommandTag\" function has many\n> calls. Do you think just for the command tag, we should do all this change?\n\nThanks for trying to address the issue!\nAs probably you've already noticed, the commit 979766c0af fixed the issue\nthanks to your review. So we can review the patch that I posted.\n\nRegards,\n\n-- \nFujii Masao\n\n\n","msg_date":"Wed, 6 Nov 2019 19:08:07 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":true,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"On Wed, Nov 6, 2019 at 4:14 PM btfujiitkp <btfujiitkp@oss.nttdata.com> wrote:\n>\n> 2019-10-31 21:01 に Fujii Masao さんは書きました:\n> > On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com>\n> > wrote:\n> >>\n> >>\n> >>\n> >> On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com>\n> >> wrote:\n> >>>\n> >>> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >>> >\n> >>> > Fujii Masao <masao.fujii@gmail.com> writes:\n> >>> > > Currently CREATE OR REPLACE VIEW command fails if the column names\n> >>> > > are changed.\n> >>> >\n> >>> > That is, I believe, intentional. It's an effective aid to catching\n> >>> > mistakes in view redefinitions, such as misaligning the new set of\n> >>> > columns relative to the old. That's particularly important given\n> >>> > that we allow you to add columns during CREATE OR REPLACE VIEW.\n> >>> > Consider the oversimplified case where you start with\n> >>> >\n> >>> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\n> >>> >\n> >>> > and you want to add a column z, and you get sloppy and write\n> >>> >\n> >>> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\n> >>> >\n> >>> > If we did not throw an error on this, references that formerly\n> >>> > pointed to column y would now point to z (as that's still attnum 2),\n> >>> > which is highly unlikely to be what you wanted.\n> >>>\n> >>> This example makes me wonder if the addtion of column by\n> >>> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\n> >>> That is, it may increase the oppotunity for users' mistake.\n> >>> I'm thinking the case where users mistakenly added new column\n> >>> into the view when replacing the view definition. This mistake can\n> >>> happen because CREATE OR REPLACE VIEW allows new column to\n> >>> be added. But what's the worse is that, currently there is no way to\n> >>> drop the column from the view, except recreation of the view.\n> >>> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\n> >>> the drop of the column from the view. So, to fix the mistake,\n> >>> users would need to drop the view itself and recreate it. If there\n> >>> are\n> >>> some objects depending the view, they also might need to be\n> >>> recreated.\n> >>> This looks not good. Since the feature has been supported,\n> >>> it's too late to say that, though...\n> >>>\n> >>> At least, the support for ALTER VIEW DROP COLUMN might be\n> >>> necessary to alleviate that situation.\n> >>>\n> >>\n> >> - Is this intentional not implemented the \"RENAME COLUMN\" statement\n> >> for\n> >> VIEW because it is implemented for Materialized View?\n> >\n> > Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\n> > sounds reasonable whether we support the rename of columns when\n> > replacing\n> > the view definition, or not. Attached is the patch that adds support\n> > for\n> > ALTER VIEW RENAME COLUMN command.\n> >\n> >> I have made just a similar\n> >> change to view and it works.\n> >\n> > Yeah, ISTM that we made the same patch at the same time. You changed\n> > gram.y,\n> > but I added the following changes additionally.\n> >\n> > - Update the doc\n> > - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the\n> > columns\n> > - Update tab-complete.c\n>\n>\n> I review your patch, and then I found that tab complete of \"alter\n> materialized view\" is also not enough.\n> So, I made a small patch referencing your patch.\n\nGood catch! The patch basically looks good to me.\nBut I think that \"ALTER MATERIALIZED VIEW xxx <TAB>\" should output also\nDEPENDS ON EXTENSION, SET TABLESPACE, CLUSTER ON and RESET.\nSo I added such tab-completes to your patch. Patch attached.\n\nBarring any objection, I'm thinking to commit this patch.\n\nRegards,\n\n-- \nFujii Masao","msg_date":"Thu, 14 Nov 2019 23:14:22 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":true,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"> Barring any objection, I'm thinking to commit this patch.\n> \n> Regards,\n\nBuild and All Test has passed .\nLooks good to me .\n\nRegards,\n\n\n","msg_date":"Wed, 20 Nov 2019 13:11:16 +0900","msg_from":"btkimurayuzk <btkimurayuzk@oss.nttdata.com>","msg_from_op":false,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"},{"msg_contents":"On Wed, Nov 20, 2019 at 1:11 PM btkimurayuzk\n<btkimurayuzk@oss.nttdata.com> wrote:\n>\n> > Barring any objection, I'm thinking to commit this patch.\n> >\n> > Regards,\n>\n> Build and All Test has passed .\n> Looks good to me .\n\nThanks for reviewing the patch!\nI committed the following two patches.\n\n- Allow ALTER VIEW command to rename the column in the view.\n- Improve tab-completion for ALTER MATERIALIZED VIEW.\n\nRegards,\n\n-- \nFujii Masao\n\n\n","msg_date":"Thu, 21 Nov 2019 19:58:32 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":true,"msg_subject":"Re: Allow CREATE OR REPLACE VIEW to rename the columns"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nCurrently CREATE OR REPLACE VIEW command fails if the column names\\nare changed. For example,\\n\\n =# CREATE VIEW test AS SELECT 0 AS a;\\n =# CREATE OR REPLACE VIEW test AS SELECT 0 AS x;\\n ERROR: cannot change name of view column \\\"a\\\" to \\\"x\\\"\\n\\nI'd like to propose the attached patch that allows CREATE OR REPLACE VIEW\\nto rename the view columns. Thought?\\n\\nRegards,\\n\\n-- \\nFujii Masao\",\n \"msg_date\": \"Thu, 31 Oct 2019 11:27:12 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"Fujii Masao <masao.fujii@gmail.com> writes:\\n> Currently CREATE OR REPLACE VIEW command fails if the column names\\n> are changed.\\n\\nThat is, I believe, intentional. It's an effective aid to catching\\nmistakes in view redefinitions, such as misaligning the new set of\\ncolumns relative to the old. That's particularly important given\\nthat we allow you to add columns during CREATE OR REPLACE VIEW.\\nConsider the oversimplified case where you start with\\n\\nCREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n\\nand you want to add a column z, and you get sloppy and write\\n\\nCREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n\\nIf we did not throw an error on this, references that formerly\\npointed to column y would now point to z (as that's still attnum 2),\\nwhich is highly unlikely to be what you wanted.\\n\\nThe right way to handle a column rename in a view is to do a separate\\nALTER VIEW RENAME COLUMN, making it totally clear what you intend to\\nhappen. (Right now, we make you spell that \\\"ALTER TABLE RENAME COLUMN\\\",\\nbut it'd be reasonable to add that syntax to ALTER VIEW too.) I don't\\nthink this functionality should be folded into redefinition of the content\\nof the view. It'd add more opportunity for mistakes than anything else.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 00:42:49 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Fujii Masao <masao.fujii@gmail.com> writes:\\n> > Currently CREATE OR REPLACE VIEW command fails if the column names\\n> > are changed.\\n>\\n> That is, I believe, intentional. It's an effective aid to catching\\n> mistakes in view redefinitions, such as misaligning the new set of\\n> columns relative to the old. That's particularly important given\\n> that we allow you to add columns during CREATE OR REPLACE VIEW.\\n> Consider the oversimplified case where you start with\\n>\\n> CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n>\\n> and you want to add a column z, and you get sloppy and write\\n>\\n> CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n>\\n> If we did not throw an error on this, references that formerly\\n> pointed to column y would now point to z (as that's still attnum 2),\\n> which is highly unlikely to be what you wanted.\\n\\nThis example makes me wonder if the addtion of column by\\nCREATE OR REPLACE VIEW also has the same (or even worse) issue.\\nThat is, it may increase the oppotunity for users' mistake.\\nI'm thinking the case where users mistakenly added new column\\ninto the view when replacing the view definition. This mistake can\\nhappen because CREATE OR REPLACE VIEW allows new column to\\nbe added. But what's the worse is that, currently there is no way to\\ndrop the column from the view, except recreation of the view.\\nNeither CREATE OR REPLACE VIEW nor ALTER TABLE support\\nthe drop of the column from the view. So, to fix the mistake,\\nusers would need to drop the view itself and recreate it. If there are\\nsome objects depending the view, they also might need to be recreated.\\nThis looks not good. Since the feature has been supported,\\nit's too late to say that, though...\\n\\nAt least, the support for ALTER VIEW DROP COLUMN might be\\nnecessary to alleviate that situation.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 16:31:52 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> wrote:\\n\\n> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >\\n> > Fujii Masao <masao.fujii@gmail.com> writes:\\n> > > Currently CREATE OR REPLACE VIEW command fails if the column names\\n> > > are changed.\\n> >\\n> > That is, I believe, intentional. It's an effective aid to catching\\n> > mistakes in view redefinitions, such as misaligning the new set of\\n> > columns relative to the old. That's particularly important given\\n> > that we allow you to add columns during CREATE OR REPLACE VIEW.\\n> > Consider the oversimplified case where you start with\\n> >\\n> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n> >\\n> > and you want to add a column z, and you get sloppy and write\\n> >\\n> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n> >\\n> > If we did not throw an error on this, references that formerly\\n> > pointed to column y would now point to z (as that's still attnum 2),\\n> > which is highly unlikely to be what you wanted.\\n>\\n> This example makes me wonder if the addtion of column by\\n> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n> That is, it may increase the oppotunity for users' mistake.\\n> I'm thinking the case where users mistakenly added new column\\n> into the view when replacing the view definition. This mistake can\\n> happen because CREATE OR REPLACE VIEW allows new column to\\n> be added. But what's the worse is that, currently there is no way to\\n> drop the column from the view, except recreation of the view.\\n> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\\n> the drop of the column from the view. So, to fix the mistake,\\n> users would need to drop the view itself and recreate it. If there are\\n> some objects depending the view, they also might need to be recreated.\\n> This looks not good. Since the feature has been supported,\\n> it's too late to say that, though...\\n>\\n> At least, the support for ALTER VIEW DROP COLUMN might be\\n> necessary to alleviate that situation.\\n>\\n>\\n- Is this intentional not implemented the \\\"RENAME COLUMN\\\" statement for\\nVIEW because it is implemented for Materialized View? I have made just a\\nsimilar\\nchange to view and it works.\\n\\nALTER VIEW v RENAME COLUMN d to e;\\n\\n- For \\\"DROP COLUMN\\\" for VIEW is throwing error.\\n\\npostgres=# alter view v drop column e;\\nERROR: \\\"v\\\" is not a table, composite type, or foreign table\\n\\n\\n\\nRegards,\\n>\\n> --\\n> Fujii Masao\\n>\\n>\\n>\\n\\n-- \\nIbrar Ahmed\",\n \"msg_date\": \"Thu, 31 Oct 2019 15:58:49 +0500\",\n \"msg_from\": \"Ibrar Ahmed <ibrar.ahmad@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n>\\n>\\n>\\n> On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> wrote:\\n>>\\n>> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> >\\n>> > Fujii Masao <masao.fujii@gmail.com> writes:\\n>> > > Currently CREATE OR REPLACE VIEW command fails if the column names\\n>> > > are changed.\\n>> >\\n>> > That is, I believe, intentional. It's an effective aid to catching\\n>> > mistakes in view redefinitions, such as misaligning the new set of\\n>> > columns relative to the old. That's particularly important given\\n>> > that we allow you to add columns during CREATE OR REPLACE VIEW.\\n>> > Consider the oversimplified case where you start with\\n>> >\\n>> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n>> >\\n>> > and you want to add a column z, and you get sloppy and write\\n>> >\\n>> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n>> >\\n>> > If we did not throw an error on this, references that formerly\\n>> > pointed to column y would now point to z (as that's still attnum 2),\\n>> > which is highly unlikely to be what you wanted.\\n>>\\n>> This example makes me wonder if the addtion of column by\\n>> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n>> That is, it may increase the oppotunity for users' mistake.\\n>> I'm thinking the case where users mistakenly added new column\\n>> into the view when replacing the view definition. This mistake can\\n>> happen because CREATE OR REPLACE VIEW allows new column to\\n>> be added. But what's the worse is that, currently there is no way to\\n>> drop the column from the view, except recreation of the view.\\n>> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\\n>> the drop of the column from the view. So, to fix the mistake,\\n>> users would need to drop the view itself and recreate it. If there are\\n>> some objects depending the view, they also might need to be recreated.\\n>> This looks not good. Since the feature has been supported,\\n>> it's too late to say that, though...\\n>>\\n>> At least, the support for ALTER VIEW DROP COLUMN might be\\n>> necessary to alleviate that situation.\\n>>\\n>\\n> - Is this intentional not implemented the \\\"RENAME COLUMN\\\" statement for\\n> VIEW because it is implemented for Materialized View?\\n\\nNot sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\\nsounds reasonable whether we support the rename of columns when replacing\\nthe view definition, or not. Attached is the patch that adds support for\\nALTER VIEW RENAME COLUMN command.\\n\\n> I have made just a similar\\n> change to view and it works.\\n\\nYeah, ISTM that we made the same patch at the same time. You changed gram.y,\\nbut I added the following changes additionally.\\n\\n- Update the doc\\n- Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the columns\\n- Update tab-complete.c\\n- Add regression test\\n\\nOne issue I've not addressed yet is about the command tag of\\n\\\"ALTER VIEW RENAME COLUMN\\\". Currently \\\"ALTER TABLE\\\" is returned as the tag\\nlike ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \\\"ALTER VIEW\\\"\\nis better. I started the discussion about the command tag of\\n\\\"ALTER MATERIALIZED VIEW\\\" at another thread. I will update the patch according\\nto the result of that discussion.\\nhttps://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\\n\\nRegards,\\n\\n-- \\nFujii Masao\",\n \"msg_date\": \"Thu, 31 Oct 2019 21:01:02 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com> wrote:\\n\\n> On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n> >\\n> >\\n> >\\n> > On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com>\\n> wrote:\\n> >>\\n> >> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >> >\\n> >> > Fujii Masao <masao.fujii@gmail.com> writes:\\n> >> > > Currently CREATE OR REPLACE VIEW command fails if the column names\\n> >> > > are changed.\\n> >> >\\n> >> > That is, I believe, intentional. It's an effective aid to catching\\n> >> > mistakes in view redefinitions, such as misaligning the new set of\\n> >> > columns relative to the old. That's particularly important given\\n> >> > that we allow you to add columns during CREATE OR REPLACE VIEW.\\n> >> > Consider the oversimplified case where you start with\\n> >> >\\n> >> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n> >> >\\n> >> > and you want to add a column z, and you get sloppy and write\\n> >> >\\n> >> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n> >> >\\n> >> > If we did not throw an error on this, references that formerly\\n> >> > pointed to column y would now point to z (as that's still attnum 2),\\n> >> > which is highly unlikely to be what you wanted.\\n> >>\\n> >> This example makes me wonder if the addtion of column by\\n> >> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n> >> That is, it may increase the oppotunity for users' mistake.\\n> >> I'm thinking the case where users mistakenly added new column\\n> >> into the view when replacing the view definition. This mistake can\\n> >> happen because CREATE OR REPLACE VIEW allows new column to\\n> >> be added. But what's the worse is that, currently there is no way to\\n> >> drop the column from the view, except recreation of the view.\\n> >> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\\n> >> the drop of the column from the view. So, to fix the mistake,\\n> >> users would need to drop the view itself and recreate it. If there are\\n> >> some objects depending the view, they also might need to be recreated.\\n> >> This looks not good. Since the feature has been supported,\\n> >> it's too late to say that, though...\\n> >>\\n> >> At least, the support for ALTER VIEW DROP COLUMN might be\\n> >> necessary to alleviate that situation.\\n> >>\\n> >\\n> > - Is this intentional not implemented the \\\"RENAME COLUMN\\\" statement for\\n> > VIEW because it is implemented for Materialized View?\\n>\\n> Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\\n> sounds reasonable whether we support the rename of columns when replacing\\n> the view definition, or not. Attached is the patch that adds support for\\n> ALTER VIEW RENAME COLUMN command.\\n>\\n> > I have made just a similar\\n> > change to view and it works.\\n>\\n> Yeah, ISTM that we made the same patch at the same time. You changed\\n> gram.y,\\n> but I added the following changes additionally.\\n>\\n> - Update the doc\\n> - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the\\n> columns\\n> - Update tab-complete.c\\n> - Add regression test\\n>\\n>\\nOh, I just sent the patch to ask, good you do that in all the places.\\n\\nOne issue I've not addressed yet is about the command tag of\\n> \\\"ALTER VIEW RENAME COLUMN\\\". Currently \\\"ALTER TABLE\\\" is returned as the tag\\n> like ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \\\"ALTER VIEW\\\"\\n> is better. I started the discussion about the command tag of\\n> \\\"ALTER MATERIALIZED VIEW\\\" at another thread. I will update the patch\\n> according\\n> to the result of that discussion.\\n>\\n> https://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\\n>\\n> Attached patch contain small change for ALTER MATERIALIZED VIEW.\\n\\n\\n\\n> Regards,\\n>\\n> --\\n> Fujii Masao\\n>\\n\\n\\n-- \\nIbrar Ahmed\",\n \"msg_date\": \"Thu, 31 Oct 2019 17:11:57 +0500\",\n \"msg_from\": \"Ibrar Ahmed <ibrar.ahmad@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 5:11 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n\\n>\\n>\\n> On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com> wrote:\\n>\\n>> On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com>\\n>> wrote:\\n>> >\\n>> >\\n>> >\\n>> > On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com>\\n>> wrote:\\n>> >>\\n>> >> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> >> >\\n>> >> > Fujii Masao <masao.fujii@gmail.com> writes:\\n>> >> > > Currently CREATE OR REPLACE VIEW command fails if the column names\\n>> >> > > are changed.\\n>> >> >\\n>> >> > That is, I believe, intentional. It's an effective aid to catching\\n>> >> > mistakes in view redefinitions, such as misaligning the new set of\\n>> >> > columns relative to the old. That's particularly important given\\n>> >> > that we allow you to add columns during CREATE OR REPLACE VIEW.\\n>> >> > Consider the oversimplified case where you start with\\n>> >> >\\n>> >> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n>> >> >\\n>> >> > and you want to add a column z, and you get sloppy and write\\n>> >> >\\n>> >> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n>> >> >\\n>> >> > If we did not throw an error on this, references that formerly\\n>> >> > pointed to column y would now point to z (as that's still attnum 2),\\n>> >> > which is highly unlikely to be what you wanted.\\n>> >>\\n>> >> This example makes me wonder if the addtion of column by\\n>> >> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n>> >> That is, it may increase the oppotunity for users' mistake.\\n>> >> I'm thinking the case where users mistakenly added new column\\n>> >> into the view when replacing the view definition. This mistake can\\n>> >> happen because CREATE OR REPLACE VIEW allows new column to\\n>> >> be added. But what's the worse is that, currently there is no way to\\n>> >> drop the column from the view, except recreation of the view.\\n>> >> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\\n>> >> the drop of the column from the view. So, to fix the mistake,\\n>> >> users would need to drop the view itself and recreate it. If there are\\n>> >> some objects depending the view, they also might need to be recreated.\\n>> >> This looks not good. Since the feature has been supported,\\n>> >> it's too late to say that, though...\\n>> >>\\n>> >> At least, the support for ALTER VIEW DROP COLUMN might be\\n>> >> necessary to alleviate that situation.\\n>> >>\\n>> >\\n>> > - Is this intentional not implemented the \\\"RENAME COLUMN\\\" statement for\\n>> > VIEW because it is implemented for Materialized View?\\n>>\\n>> Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\\n>> sounds reasonable whether we support the rename of columns when replacing\\n>> the view definition, or not. Attached is the patch that adds support for\\n>> ALTER VIEW RENAME COLUMN command.\\n>>\\n>> > I have made just a similar\\n>> > change to view and it works.\\n>>\\n>> Yeah, ISTM that we made the same patch at the same time. You changed\\n>> gram.y,\\n>> but I added the following changes additionally.\\n>>\\n>> - Update the doc\\n>> - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the\\n>> columns\\n>> - Update tab-complete.c\\n>> - Add regression test\\n>>\\n>>\\n> Oh, I just sent the patch to ask, good you do that in all the places.\\n>\\n> One issue I've not addressed yet is about the command tag of\\n>> \\\"ALTER VIEW RENAME COLUMN\\\". Currently \\\"ALTER TABLE\\\" is returned as the tag\\n>> like ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \\\"ALTER VIEW\\\"\\n>> is better. I started the discussion about the command tag of\\n>> \\\"ALTER MATERIALIZED VIEW\\\" at another thread. I will update the patch\\n>> according\\n>> to the result of that discussion.\\n>>\\n>> https://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\\n>>\\n>> Attached patch contain small change for ALTER MATERIALIZED VIEW.\\n>\\n>\\nHmm, my small change of \\\"ALTER MATERIALIZED VIEW\\\" does not work in some\\ncases need more work on that.\\n\\n\\n>\\n>\\n>> Regards,\\n>>\\n>> --\\n>> Fujii Masao\\n>>\\n>\\n>\\n> --\\n> Ibrar Ahmed\\n>\\n\\n\\n-- \\nIbrar Ahmed\\n\\nOn Thu, Oct 31, 2019 at 5:11 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com> wrote:On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n>\\n>\\n>\\n> On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> wrote:\\n>>\\n>> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> >\\n>> > Fujii Masao <masao.fujii@gmail.com> writes:\\n>> > > Currently CREATE OR REPLACE VIEW command fails if the column names\\n>> > > are changed.\\n>> >\\n>> > That is, I believe, intentional. It's an effective aid to catching\\n>> > mistakes in view redefinitions, such as misaligning the new set of\\n>> > columns relative to the old. That's particularly important given\\n>> > that we allow you to add columns during CREATE OR REPLACE VIEW.\\n>> > Consider the oversimplified case where you start with\\n>> >\\n>> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n>> >\\n>> > and you want to add a column z, and you get sloppy and write\\n>> >\\n>> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n>> >\\n>> > If we did not throw an error on this, references that formerly\\n>> > pointed to column y would now point to z (as that's still attnum 2),\\n>> > which is highly unlikely to be what you wanted.\\n>>\\n>> This example makes me wonder if the addtion of column by\\n>> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n>> That is, it may increase the oppotunity for users' mistake.\\n>> I'm thinking the case where users mistakenly added new column\\n>> into the view when replacing the view definition. This mistake can\\n>> happen because CREATE OR REPLACE VIEW allows new column to\\n>> be added. But what's the worse is that, currently there is no way to\\n>> drop the column from the view, except recreation of the view.\\n>> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\\n>> the drop of the column from the view. So, to fix the mistake,\\n>> users would need to drop the view itself and recreate it. If there are\\n>> some objects depending the view, they also might need to be recreated.\\n>> This looks not good. Since the feature has been supported,\\n>> it's too late to say that, though...\\n>>\\n>> At least, the support for ALTER VIEW DROP COLUMN might be\\n>> necessary to alleviate that situation.\\n>>\\n>\\n> - Is this intentional not implemented the \\\"RENAME COLUMN\\\" statement for\\n> VIEW because it is implemented for Materialized View?\\n\\nNot sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\\nsounds reasonable whether we support the rename of columns when replacing\\nthe view definition, or not. Attached is the patch that adds support for\\nALTER VIEW RENAME COLUMN command.\\n\\n> I have made just a similar\\n> change to view and it works.\\n\\nYeah, ISTM that we made the same patch at the same time. You changed gram.y,\\nbut I added the following changes additionally.\\n\\n- Update the doc\\n- Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the columns\\n- Update tab-complete.c\\n- Add regression test\\nOh, I just sent the patch to ask, good you do that in all the places. \\nOne issue I've not addressed yet is about the command tag of\\n\\\"ALTER VIEW RENAME COLUMN\\\". Currently \\\"ALTER TABLE\\\" is returned as the tag\\nlike ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \\\"ALTER VIEW\\\"\\nis better. I started the discussion about the command tag of\\n\\\"ALTER MATERIALIZED VIEW\\\" at another thread. I will update the patch according\\nto the result of that discussion.\\nhttps://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\\nAttached patch contain small change for ALTER MATERIALIZED VIEW.Hmm, my small change of \\\"ALTER MATERIALIZED VIEW\\\" does not work in some cases need more work on that. \\nRegards,\\n\\n-- \\nFujii Masao\\n-- Ibrar Ahmed\\n-- Ibrar Ahmed\",\n \"msg_date\": \"Thu, 31 Oct 2019 17:28:00 +0500\",\n \"msg_from\": \"Ibrar Ahmed <ibrar.ahmad@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 5:28 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n\\n>\\n>\\n> On Thu, Oct 31, 2019 at 5:11 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n>\\n>>\\n>>\\n>> On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com>\\n>> wrote:\\n>>\\n>>> On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com>\\n>>> wrote:\\n>>> >\\n>>> >\\n>>> >\\n>>> > On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com>\\n>>> wrote:\\n>>> >>\\n>>> >> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>> >> >\\n>>> >> > Fujii Masao <masao.fujii@gmail.com> writes:\\n>>> >> > > Currently CREATE OR REPLACE VIEW command fails if the column names\\n>>> >> > > are changed.\\n>>> >> >\\n>>> >> > That is, I believe, intentional. It's an effective aid to catching\\n>>> >> > mistakes in view redefinitions, such as misaligning the new set of\\n>>> >> > columns relative to the old. That's particularly important given\\n>>> >> > that we allow you to add columns during CREATE OR REPLACE VIEW.\\n>>> >> > Consider the oversimplified case where you start with\\n>>> >> >\\n>>> >> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n>>> >> >\\n>>> >> > and you want to add a column z, and you get sloppy and write\\n>>> >> >\\n>>> >> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n>>> >> >\\n>>> >> > If we did not throw an error on this, references that formerly\\n>>> >> > pointed to column y would now point to z (as that's still attnum 2),\\n>>> >> > which is highly unlikely to be what you wanted.\\n>>> >>\\n>>> >> This example makes me wonder if the addtion of column by\\n>>> >> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n>>> >> That is, it may increase the oppotunity for users' mistake.\\n>>> >> I'm thinking the case where users mistakenly added new column\\n>>> >> into the view when replacing the view definition. This mistake can\\n>>> >> happen because CREATE OR REPLACE VIEW allows new column to\\n>>> >> be added. But what's the worse is that, currently there is no way to\\n>>> >> drop the column from the view, except recreation of the view.\\n>>> >> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\\n>>> >> the drop of the column from the view. So, to fix the mistake,\\n>>> >> users would need to drop the view itself and recreate it. If there are\\n>>> >> some objects depending the view, they also might need to be recreated.\\n>>> >> This looks not good. Since the feature has been supported,\\n>>> >> it's too late to say that, though...\\n>>> >>\\n>>> >> At least, the support for ALTER VIEW DROP COLUMN might be\\n>>> >> necessary to alleviate that situation.\\n>>> >>\\n>>> >\\n>>> > - Is this intentional not implemented the \\\"RENAME COLUMN\\\" statement for\\n>>> > VIEW because it is implemented for Materialized View?\\n>>>\\n>>> Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\\n>>> sounds reasonable whether we support the rename of columns when replacing\\n>>> the view definition, or not. Attached is the patch that adds support for\\n>>> ALTER VIEW RENAME COLUMN command.\\n>>>\\n>>> > I have made just a similar\\n>>> > change to view and it works.\\n>>>\\n>>> Yeah, ISTM that we made the same patch at the same time. You changed\\n>>> gram.y,\\n>>> but I added the following changes additionally.\\n>>>\\n>>> - Update the doc\\n>>> - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the\\n>>> columns\\n>>> - Update tab-complete.c\\n>>> - Add regression test\\n>>>\\n>>>\\n>> Oh, I just sent the patch to ask, good you do that in all the places.\\n>>\\n>> One issue I've not addressed yet is about the command tag of\\n>>> \\\"ALTER VIEW RENAME COLUMN\\\". Currently \\\"ALTER TABLE\\\" is returned as the\\n>>> tag\\n>>> like ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \\\"ALTER VIEW\\\"\\n>>> is better. I started the discussion about the command tag of\\n>>> \\\"ALTER MATERIALIZED VIEW\\\" at another thread. I will update the patch\\n>>> according\\n>>> to the result of that discussion.\\n>>>\\n>>> https://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\\n>>>\\n>>> Attached patch contain small change for ALTER MATERIALIZED VIEW.\\n>>\\n>>\\n> Hmm, my small change of \\\"ALTER MATERIALIZED VIEW\\\" does not work in some\\n> cases need more work on that.\\n>\\n>\\n\\nThe AlterObjectTypeCommandTag function just take one parameter, but to\\nshow \\\"ALTER MATERIALIZED VIEW\\\" instead of ALTER TABLE we need to\\npass \\\"relationType = OBJECT_MATVIEW\\\" along with \\\"renameType = OBJECT_COLUMN\\\"\\nand handle that in the function. The \\\"AlterObjectTypeCommandTag\\\" function\\nhas many\\ncalls. Do you think just for the command tag, we should do all this change?\\n\\n\\n>\\n>>\\n>>> Regards,\\n>>>\\n>>> --\\n>>> Fujii Masao\\n>>>\\n>>\\n>>\\n>> --\\n>> Ibrar Ahmed\\n>>\\n>\\n>\\n> --\\n> Ibrar Ahmed\\n>\\n\\n\\n-- \\nIbrar Ahmed\\n\\nOn Thu, Oct 31, 2019 at 5:28 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:On Thu, Oct 31, 2019 at 5:11 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com> wrote:On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n>\\n>\\n>\\n> On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> wrote:\\n>>\\n>> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> >\\n>> > Fujii Masao <masao.fujii@gmail.com> writes:\\n>> > > Currently CREATE OR REPLACE VIEW command fails if the column names\\n>> > > are changed.\\n>> >\\n>> > That is, I believe, intentional. It's an effective aid to catching\\n>> > mistakes in view redefinitions, such as misaligning the new set of\\n>> > columns relative to the old. That's particularly important given\\n>> > that we allow you to add columns during CREATE OR REPLACE VIEW.\\n>> > Consider the oversimplified case where you start with\\n>> >\\n>> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n>> >\\n>> > and you want to add a column z, and you get sloppy and write\\n>> >\\n>> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n>> >\\n>> > If we did not throw an error on this, references that formerly\\n>> > pointed to column y would now point to z (as that's still attnum 2),\\n>> > which is highly unlikely to be what you wanted.\\n>>\\n>> This example makes me wonder if the addtion of column by\\n>> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n>> That is, it may increase the oppotunity for users' mistake.\\n>> I'm thinking the case where users mistakenly added new column\\n>> into the view when replacing the view definition. This mistake can\\n>> happen because CREATE OR REPLACE VIEW allows new column to\\n>> be added. But what's the worse is that, currently there is no way to\\n>> drop the column from the view, except recreation of the view.\\n>> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\\n>> the drop of the column from the view. So, to fix the mistake,\\n>> users would need to drop the view itself and recreate it. If there are\\n>> some objects depending the view, they also might need to be recreated.\\n>> This looks not good. Since the feature has been supported,\\n>> it's too late to say that, though...\\n>>\\n>> At least, the support for ALTER VIEW DROP COLUMN might be\\n>> necessary to alleviate that situation.\\n>>\\n>\\n> - Is this intentional not implemented the \\\"RENAME COLUMN\\\" statement for\\n> VIEW because it is implemented for Materialized View?\\n\\nNot sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\\nsounds reasonable whether we support the rename of columns when replacing\\nthe view definition, or not. Attached is the patch that adds support for\\nALTER VIEW RENAME COLUMN command.\\n\\n> I have made just a similar\\n> change to view and it works.\\n\\nYeah, ISTM that we made the same patch at the same time. You changed gram.y,\\nbut I added the following changes additionally.\\n\\n- Update the doc\\n- Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the columns\\n- Update tab-complete.c\\n- Add regression test\\nOh, I just sent the patch to ask, good you do that in all the places. \\nOne issue I've not addressed yet is about the command tag of\\n\\\"ALTER VIEW RENAME COLUMN\\\". Currently \\\"ALTER TABLE\\\" is returned as the tag\\nlike ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \\\"ALTER VIEW\\\"\\nis better. I started the discussion about the command tag of\\n\\\"ALTER MATERIALIZED VIEW\\\" at another thread. I will update the patch according\\nto the result of that discussion.\\nhttps://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\\nAttached patch contain small change for ALTER MATERIALIZED VIEW.Hmm, my small change of \\\"ALTER MATERIALIZED VIEW\\\" does not work in some cases need more work on that. The AlterObjectTypeCommandTag function just take one parameter, but toshow \\\"ALTER MATERIALIZED VIEW\\\" instead of ALTER TABLE we need topass \\\"relationType = OBJECT_MATVIEW\\\" along with \\\"renameType = OBJECT_COLUMN\\\"and handle that in the function. The \\\"AlterObjectTypeCommandTag\\\" function has manycalls. Do you think just for the command tag, we should do all this change? \\n \\nRegards,\\n\\n-- \\nFujii Masao\\n-- Ibrar Ahmed\\n-- Ibrar Ahmed\\n-- Ibrar Ahmed\",\n \"msg_date\": \"Thu, 31 Oct 2019 17:34:22 +0500\",\n \"msg_from\": \"Ibrar Ahmed <ibrar.ahmad@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"Fujii Masao <masao.fujii@gmail.com> writes:\\n> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Fujii Masao <masao.fujii@gmail.com> writes:\\n>>> Currently CREATE OR REPLACE VIEW command fails if the column names\\n>>> are changed.\\n\\n>> That is, I believe, intentional. It's an effective aid to catching\\n>> mistakes in view redefinitions, such as misaligning the new set of\\n>> columns relative to the old. [example]\\n\\n> This example makes me wonder if the addtion of column by\\n> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n> That is, it may increase the oppotunity for users' mistake.\\n\\nThe idea in CREATE OR REPLACE VIEW is to allow addition of new\\ncolumns at the end, the same as you can do with tables. Checking\\nthe column name matchups is a way to ensure that you actually do\\nadd at the end, rather than insert, which wouldn't act as you\\nexpect. Admittedly it's only heuristic.\\n\\nWe could, perhaps, have insisted that adding a column also requires\\nspecial syntax, but we didn't. Consider for example a case where\\nthe new column needs to come from an additionally joined table;\\nthen you have to be able to edit the underlying view definition along\\nwith adding the column. So that seems like kind of a pain in the\\nneck to insist on.\\n\\n> But what's the worse is that, currently there is no way to\\n> drop the column from the view, except recreation of the view.\\n\\nI think this has been discussed, as well. It's not necessarily\\nsimple to drop a view output column. For example, if the view\\nuses SELECT DISTINCT, removing an output column would have\\nsemantic effects on the set of rows that can be returned, since\\ndistinct-ness would now mean something else than it did before.\\n\\nIt's conceivable that we could enumerate all such effects and\\nthen allow DROP COLUMN (probably replacing the output column\\nwith a null constant) if none of them apply, but I can't get\\nterribly excited about it. The field demand for such a feature\\nhas not been high. I'd be a bit worried about bugs arising\\nfrom failures to check attisdropped for views, too; so that\\nthe cost of getting this working might be greater than it seems\\non the surface.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 09:54:24 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 10:54 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Fujii Masao <masao.fujii@gmail.com> writes:\\n> > On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >> Fujii Masao <masao.fujii@gmail.com> writes:\\n> >>> Currently CREATE OR REPLACE VIEW command fails if the column names\\n> >>> are changed.\\n>\\n> >> That is, I believe, intentional. It's an effective aid to catching\\n> >> mistakes in view redefinitions, such as misaligning the new set of\\n> >> columns relative to the old. [example]\\n>\\n> > This example makes me wonder if the addtion of column by\\n> > CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n> > That is, it may increase the oppotunity for users' mistake.\\n>\\n> The idea in CREATE OR REPLACE VIEW is to allow addition of new\\n> columns at the end, the same as you can do with tables. Checking\\n> the column name matchups is a way to ensure that you actually do\\n> add at the end, rather than insert, which wouldn't act as you\\n> expect. Admittedly it's only heuristic.\\n>\\n> We could, perhaps, have insisted that adding a column also requires\\n> special syntax, but we didn't. Consider for example a case where\\n> the new column needs to come from an additionally joined table;\\n> then you have to be able to edit the underlying view definition along\\n> with adding the column. So that seems like kind of a pain in the\\n> neck to insist on.\\n>\\n> > But what's the worse is that, currently there is no way to\\n> > drop the column from the view, except recreation of the view.\\n>\\n> I think this has been discussed, as well. It's not necessarily\\n> simple to drop a view output column. For example, if the view\\n> uses SELECT DISTINCT, removing an output column would have\\n> semantic effects on the set of rows that can be returned, since\\n> distinct-ness would now mean something else than it did before.\\n>\\n> It's conceivable that we could enumerate all such effects and\\n> then allow DROP COLUMN (probably replacing the output column\\n> with a null constant) if none of them apply, but I can't get\\n> terribly excited about it. The field demand for such a feature\\n> has not been high. I'd be a bit worried about bugs arising\\n> from failures to check attisdropped for views, too; so that\\n> the cost of getting this working might be greater than it seems\\n> on the surface.\\n\\nThanks for the explanation! Understood.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 12:42:43 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"2019-10-31 21:01 に Fujii Masao さんは書きました:\\n> On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> \\n> wrote:\\n>> \\n>> \\n>> \\n>> On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> \\n>> wrote:\\n>>> \\n>>> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>> >\\n>>> > Fujii Masao <masao.fujii@gmail.com> writes:\\n>>> > > Currently CREATE OR REPLACE VIEW command fails if the column names\\n>>> > > are changed.\\n>>> >\\n>>> > That is, I believe, intentional. It's an effective aid to catching\\n>>> > mistakes in view redefinitions, such as misaligning the new set of\\n>>> > columns relative to the old. That's particularly important given\\n>>> > that we allow you to add columns during CREATE OR REPLACE VIEW.\\n>>> > Consider the oversimplified case where you start with\\n>>> >\\n>>> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n>>> >\\n>>> > and you want to add a column z, and you get sloppy and write\\n>>> >\\n>>> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n>>> >\\n>>> > If we did not throw an error on this, references that formerly\\n>>> > pointed to column y would now point to z (as that's still attnum 2),\\n>>> > which is highly unlikely to be what you wanted.\\n>>> \\n>>> This example makes me wonder if the addtion of column by\\n>>> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n>>> That is, it may increase the oppotunity for users' mistake.\\n>>> I'm thinking the case where users mistakenly added new column\\n>>> into the view when replacing the view definition. This mistake can\\n>>> happen because CREATE OR REPLACE VIEW allows new column to\\n>>> be added. But what's the worse is that, currently there is no way to\\n>>> drop the column from the view, except recreation of the view.\\n>>> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\\n>>> the drop of the column from the view. So, to fix the mistake,\\n>>> users would need to drop the view itself and recreate it. If there \\n>>> are\\n>>> some objects depending the view, they also might need to be \\n>>> recreated.\\n>>> This looks not good. Since the feature has been supported,\\n>>> it's too late to say that, though...\\n>>> \\n>>> At least, the support for ALTER VIEW DROP COLUMN might be\\n>>> necessary to alleviate that situation.\\n>>> \\n>> \\n>> - Is this intentional not implemented the \\\"RENAME COLUMN\\\" statement \\n>> for\\n>> VIEW because it is implemented for Materialized View?\\n> \\n> Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\\n> sounds reasonable whether we support the rename of columns when \\n> replacing\\n> the view definition, or not. Attached is the patch that adds support \\n> for\\n> ALTER VIEW RENAME COLUMN command.\\n> \\n>> I have made just a similar\\n>> change to view and it works.\\n> \\n> Yeah, ISTM that we made the same patch at the same time. You changed \\n> gram.y,\\n> but I added the following changes additionally.\\n> \\n> - Update the doc\\n> - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the \\n> columns\\n> - Update tab-complete.c\\n\\n\\nI review your patch, and then I found that tab complete of \\\"alter \\nmaterialized view\\\" is also not enough.\\nSo, I made a small patch referencing your patch.\\n\\nRegards,\",\n \"msg_date\": \"Wed, 06 Nov 2019 16:14:09 +0900\",\n \"msg_from\": \"btfujiitkp <btfujiitkp@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 9:34 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n>\\n>\\n>\\n> On Thu, Oct 31, 2019 at 5:28 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n>>\\n>>\\n>>\\n>> On Thu, Oct 31, 2019 at 5:11 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n>>>\\n>>>\\n>>>\\n>>> On Thu, Oct 31, 2019 at 5:01 PM Fujii Masao <masao.fujii@gmail.com> wrote:\\n>>>>\\n>>>> On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n>>>> >\\n>>>> >\\n>>>> >\\n>>>> > On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com> wrote:\\n>>>> >>\\n>>>> >> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>>> >> >\\n>>>> >> > Fujii Masao <masao.fujii@gmail.com> writes:\\n>>>> >> > > Currently CREATE OR REPLACE VIEW command fails if the column names\\n>>>> >> > > are changed.\\n>>>> >> >\\n>>>> >> > That is, I believe, intentional. It's an effective aid to catching\\n>>>> >> > mistakes in view redefinitions, such as misaligning the new set of\\n>>>> >> > columns relative to the old. That's particularly important given\\n>>>> >> > that we allow you to add columns during CREATE OR REPLACE VIEW.\\n>>>> >> > Consider the oversimplified case where you start with\\n>>>> >> >\\n>>>> >> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n>>>> >> >\\n>>>> >> > and you want to add a column z, and you get sloppy and write\\n>>>> >> >\\n>>>> >> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n>>>> >> >\\n>>>> >> > If we did not throw an error on this, references that formerly\\n>>>> >> > pointed to column y would now point to z (as that's still attnum 2),\\n>>>> >> > which is highly unlikely to be what you wanted.\\n>>>> >>\\n>>>> >> This example makes me wonder if the addtion of column by\\n>>>> >> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n>>>> >> That is, it may increase the oppotunity for users' mistake.\\n>>>> >> I'm thinking the case where users mistakenly added new column\\n>>>> >> into the view when replacing the view definition. This mistake can\\n>>>> >> happen because CREATE OR REPLACE VIEW allows new column to\\n>>>> >> be added. But what's the worse is that, currently there is no way to\\n>>>> >> drop the column from the view, except recreation of the view.\\n>>>> >> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\\n>>>> >> the drop of the column from the view. So, to fix the mistake,\\n>>>> >> users would need to drop the view itself and recreate it. If there are\\n>>>> >> some objects depending the view, they also might need to be recreated.\\n>>>> >> This looks not good. Since the feature has been supported,\\n>>>> >> it's too late to say that, though...\\n>>>> >>\\n>>>> >> At least, the support for ALTER VIEW DROP COLUMN might be\\n>>>> >> necessary to alleviate that situation.\\n>>>> >>\\n>>>> >\\n>>>> > - Is this intentional not implemented the \\\"RENAME COLUMN\\\" statement for\\n>>>> > VIEW because it is implemented for Materialized View?\\n>>>>\\n>>>> Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\\n>>>> sounds reasonable whether we support the rename of columns when replacing\\n>>>> the view definition, or not. Attached is the patch that adds support for\\n>>>> ALTER VIEW RENAME COLUMN command.\\n>>>>\\n>>>> > I have made just a similar\\n>>>> > change to view and it works.\\n>>>>\\n>>>> Yeah, ISTM that we made the same patch at the same time. You changed gram.y,\\n>>>> but I added the following changes additionally.\\n>>>>\\n>>>> - Update the doc\\n>>>> - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the columns\\n>>>> - Update tab-complete.c\\n>>>> - Add regression test\\n>>>>\\n>>>\\n>>> Oh, I just sent the patch to ask, good you do that in all the places.\\n>>>\\n>>>> One issue I've not addressed yet is about the command tag of\\n>>>> \\\"ALTER VIEW RENAME COLUMN\\\". Currently \\\"ALTER TABLE\\\" is returned as the tag\\n>>>> like ALTER MATERIALIZED VIEW RENAME COLUMN, but ISTM that \\\"ALTER VIEW\\\"\\n>>>> is better. I started the discussion about the command tag of\\n>>>> \\\"ALTER MATERIALIZED VIEW\\\" at another thread. I will update the patch according\\n>>>> to the result of that discussion.\\n>>>> https://postgr.es/m/CAHGQGwGUaC03FFdTFoHsCuDrrNvFvNVQ6xyd40==P25WvuBJjg@mail.gmail.com\\n>>>>\\n>>> Attached patch contain small change for ALTER MATERIALIZED VIEW.\\n>>>\\n>>\\n>> Hmm, my small change of \\\"ALTER MATERIALIZED VIEW\\\" does not work in some cases need more work on that.\\n>>\\n>\\n>\\n> The AlterObjectTypeCommandTag function just take one parameter, but to\\n> show \\\"ALTER MATERIALIZED VIEW\\\" instead of ALTER TABLE we need to\\n> pass \\\"relationType = OBJECT_MATVIEW\\\" along with \\\"renameType = OBJECT_COLUMN\\\"\\n> and handle that in the function. The \\\"AlterObjectTypeCommandTag\\\" function has many\\n> calls. Do you think just for the command tag, we should do all this change?\\n\\nThanks for trying to address the issue!\\nAs probably you've already noticed, the commit 979766c0af fixed the issue\\nthanks to your review. So we can review the patch that I posted.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Nov 2019 19:08:07 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 6, 2019 at 4:14 PM btfujiitkp <btfujiitkp@oss.nttdata.com> wrote:\\n>\\n> 2019-10-31 21:01 に Fujii Masao さんは書きました:\\n> > On Thu, Oct 31, 2019 at 7:59 PM Ibrar Ahmed <ibrar.ahmad@gmail.com>\\n> > wrote:\\n> >>\\n> >>\\n> >>\\n> >> On Thu, Oct 31, 2019 at 12:32 PM Fujii Masao <masao.fujii@gmail.com>\\n> >> wrote:\\n> >>>\\n> >>> On Thu, Oct 31, 2019 at 1:42 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >>> >\\n> >>> > Fujii Masao <masao.fujii@gmail.com> writes:\\n> >>> > > Currently CREATE OR REPLACE VIEW command fails if the column names\\n> >>> > > are changed.\\n> >>> >\\n> >>> > That is, I believe, intentional. It's an effective aid to catching\\n> >>> > mistakes in view redefinitions, such as misaligning the new set of\\n> >>> > columns relative to the old. That's particularly important given\\n> >>> > that we allow you to add columns during CREATE OR REPLACE VIEW.\\n> >>> > Consider the oversimplified case where you start with\\n> >>> >\\n> >>> > CREATE VIEW v AS SELECT 1 AS x, 2 AS y;\\n> >>> >\\n> >>> > and you want to add a column z, and you get sloppy and write\\n> >>> >\\n> >>> > CREATE OR REPLACE VIEW v AS SELECT 1 AS x, 3 AS z, 2 AS y;\\n> >>> >\\n> >>> > If we did not throw an error on this, references that formerly\\n> >>> > pointed to column y would now point to z (as that's still attnum 2),\\n> >>> > which is highly unlikely to be what you wanted.\\n> >>>\\n> >>> This example makes me wonder if the addtion of column by\\n> >>> CREATE OR REPLACE VIEW also has the same (or even worse) issue.\\n> >>> That is, it may increase the oppotunity for users' mistake.\\n> >>> I'm thinking the case where users mistakenly added new column\\n> >>> into the view when replacing the view definition. This mistake can\\n> >>> happen because CREATE OR REPLACE VIEW allows new column to\\n> >>> be added. But what's the worse is that, currently there is no way to\\n> >>> drop the column from the view, except recreation of the view.\\n> >>> Neither CREATE OR REPLACE VIEW nor ALTER TABLE support\\n> >>> the drop of the column from the view. So, to fix the mistake,\\n> >>> users would need to drop the view itself and recreate it. If there\\n> >>> are\\n> >>> some objects depending the view, they also might need to be\\n> >>> recreated.\\n> >>> This looks not good. Since the feature has been supported,\\n> >>> it's too late to say that, though...\\n> >>>\\n> >>> At least, the support for ALTER VIEW DROP COLUMN might be\\n> >>> necessary to alleviate that situation.\\n> >>>\\n> >>\\n> >> - Is this intentional not implemented the \\\"RENAME COLUMN\\\" statement\\n> >> for\\n> >> VIEW because it is implemented for Materialized View?\\n> >\\n> > Not sure that, but Tom's suggestion to support ALTER VIEW RENAME COLUMN\\n> > sounds reasonable whether we support the rename of columns when\\n> > replacing\\n> > the view definition, or not. Attached is the patch that adds support\\n> > for\\n> > ALTER VIEW RENAME COLUMN command.\\n> >\\n> >> I have made just a similar\\n> >> change to view and it works.\\n> >\\n> > Yeah, ISTM that we made the same patch at the same time. You changed\\n> > gram.y,\\n> > but I added the following changes additionally.\\n> >\\n> > - Update the doc\\n> > - Add HINT message emit when CRAETE OR REPLACE VIEW fails to rename the\\n> > columns\\n> > - Update tab-complete.c\\n>\\n>\\n> I review your patch, and then I found that tab complete of \\\"alter\\n> materialized view\\\" is also not enough.\\n> So, I made a small patch referencing your patch.\\n\\nGood catch! The patch basically looks good to me.\\nBut I think that \\\"ALTER MATERIALIZED VIEW xxx <TAB>\\\" should output also\\nDEPENDS ON EXTENSION, SET TABLESPACE, CLUSTER ON and RESET.\\nSo I added such tab-completes to your patch. Patch attached.\\n\\nBarring any objection, I'm thinking to commit this patch.\\n\\nRegards,\\n\\n-- \\nFujii Masao\",\n \"msg_date\": \"Thu, 14 Nov 2019 23:14:22 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"> Barring any objection, I'm thinking to commit this patch.\\n> \\n> Regards,\\n\\nBuild and All Test has passed .\\nLooks good to me .\\n\\nRegards,\\n\\n\\n\",\n \"msg_date\": \"Wed, 20 Nov 2019 13:11:16 +0900\",\n \"msg_from\": \"btkimurayuzk <btkimurayuzk@oss.nttdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 20, 2019 at 1:11 PM btkimurayuzk\\n<btkimurayuzk@oss.nttdata.com> wrote:\\n>\\n> > Barring any objection, I'm thinking to commit this patch.\\n> >\\n> > Regards,\\n>\\n> Build and All Test has passed .\\n> Looks good to me .\\n\\nThanks for reviewing the patch!\\nI committed the following two patches.\\n\\n- Allow ALTER VIEW command to rename the column in the view.\\n- Improve tab-completion for ALTER MATERIALIZED VIEW.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\n\\n\\n\",\n \"msg_date\": \"Thu, 21 Nov 2019 19:58:32 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow CREATE OR REPLACE VIEW to rename the columns\"\n }\n]","isConversational":false}}},{"rowIdx":49370,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\n\nI wondered can we have a shortcut somewhat similar to following POC\nin recomputeNamespacePath () when the recomputed path is the same as the\nprevious baseSearchPath/activeSearchPath :\n\n== POC patch ==\ndiff --git a/src/backend/catalog/namespace.c\nb/src/backend/catalog/namespace.c\nindex e251f5a9fdc..b25ef489e47 100644\n--- a/src/backend/catalog/namespace.c\n+++ b/src/backend/catalog/namespace.c\n@@ -3813,6 +3813,9 @@ recomputeNamespacePath(void)\n !list_member_oid(oidlist, myTempNamespace))\n oidlist = lcons_oid(myTempNamespace, oidlist);\n\n+ /* TODO: POC */\n+ if (equal(oidlist, baseSearchPath))\n+ return;\n /*\n * Now that we've successfully built the new list of namespace OIDs,\nsave\n * it in permanent storage.\n== POC patch end ==\n\nIt can have two advantages as:\n\n1. Avoid unnecessary list_copy() in TopMemoryContext context &\n2. Global pointers like activeSearchPath/baseSearchPath will not change if\nsome\n implementation end up with cascaded call to recomputeNamespacePath().\n\nThoughts/Comments?\n\nRegards,\nAmul\n\nHi all,I wondered can we have a shortcut somewhat similar to following POCin recomputeNamespacePath () when the recomputed path is the same as theprevious baseSearchPath/activeSearchPath :== POC patch ==diff --git a/src/backend/catalog/namespace.c b/src/backend/catalog/namespace.cindex e251f5a9fdc..b25ef489e47 100644--- a/src/backend/catalog/namespace.c+++ b/src/backend/catalog/namespace.c@@ -3813,6 +3813,9 @@ recomputeNamespacePath(void) !list_member_oid(oidlist, myTempNamespace)) oidlist = lcons_oid(myTempNamespace, oidlist); + /* TODO: POC */+ if (equal(oidlist, baseSearchPath))+ return; /* * Now that we've successfully built the new list of namespace OIDs, save * it in permanent storage.== POC patch end ==It can have two advantages as:1. Avoid unnecessary list_copy() in TopMemoryContext context &2. Global pointers like activeSearchPath/baseSearchPath will not change if some implementation end up with cascaded call to recomputeNamespacePath().Thoughts/Comments?Regards,Amul","msg_date":"Thu, 31 Oct 2019 14:11:03 +0530","msg_from":"amul sul <sulamul@gmail.com>","msg_from_op":true,"msg_subject":"Can avoid list_copy in recomputeNamespacePath() conditionally?"},{"msg_contents":"amul sul <sulamul@gmail.com> writes:\n> I wondered can we have a shortcut somewhat similar to following POC\n> in recomputeNamespacePath () when the recomputed path is the same as the\n> previous baseSearchPath/activeSearchPath :\n> + /* TODO: POC */\n> + if (equal(oidlist, baseSearchPath))\n> + return;\n\nThere's an awful lot missing from that sketch; all of the remaining\nsteps still need to be done:\n\n\tbaseCreationNamespace = firstNS;\n\tbaseTempCreationPending = temp_missing;\n\n\t/* Mark the path valid. */\n\tbaseSearchPathValid = true;\n\tnamespaceUser = roleid;\n\n\t/* And make it active. */\n\tactiveSearchPath = baseSearchPath;\n\tactiveCreationNamespace = baseCreationNamespace;\n\tactiveTempCreationPending = baseTempCreationPending;\n\n\t/* Clean up. */\n\tpfree(rawname);\n\tlist_free(namelist);\n\tlist_free(oidlist);\n\nMore to the point, I think the onus would be on the patch submitter\nto prove that the extra complexity had some measurable benefit.\nI really doubt that it would, since the list_copy is surely trivial\ncompared to the catalog lookup work we had to do to compute the OID\nlist above here.\n\nIt'd likely be more useful to see if you could reduce the number of\nplaces where we have to invalidate the path in the first place.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 02 Nov 2019 10:31:14 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Can avoid list_copy in recomputeNamespacePath() conditionally?"},{"msg_contents":"On Sat, Nov 2, 2019 at 8:01 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> amul sul <sulamul@gmail.com> writes:\n> > I wondered can we have a shortcut somewhat similar to following POC\n> > in recomputeNamespacePath () when the recomputed path is the same as the\n> > previous baseSearchPath/activeSearchPath :\n> > + /* TODO: POC */\n> > + if (equal(oidlist, baseSearchPath))\n> > + return;\n>\n> There's an awful lot missing from that sketch; all of the remaining\n> steps still need to be done:\n>\n>\nYou are correct, but that was intentionally skipped to avoid longer post\ndescriptions for the initial discussion. Sorry for being little lazy.\n\n\n> baseCreationNamespace = firstNS;\n> baseTempCreationPending = temp_missing;\n>\n> /* Mark the path valid. */\n> baseSearchPathValid = true;\n> namespaceUser = roleid;\n>\n> /* And make it active. */\n> activeSearchPath = baseSearchPath;\n> activeCreationNamespace = baseCreationNamespace;\n> activeTempCreationPending = baseTempCreationPending;\n>\n> /* Clean up. */\n> pfree(rawname);\n> list_free(namelist);\n> list_free(oidlist);\n>\n> More to the point, I think the onus would be on the patch submitter\n> to prove that the extra complexity had some measurable benefit.\n> I really doubt that it would, since the list_copy is surely trivial\n> compared to the catalog lookup work we had to do to compute the OID\n> list above here.\n>\n\nAgree.\n\n\n> It'd likely be more useful to see if you could reduce the number of\n> places where we have to invalidate the path in the first place.\n>\n\nUnderstood, let me check.\n\nRegards,\nAmul\n\nOn Sat, Nov 2, 2019 at 8:01 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:amul sul <sulamul@gmail.com> writes:\n> I wondered can we have a shortcut somewhat similar to following POC\n> in recomputeNamespacePath () when the recomputed path is the same as the\n> previous baseSearchPath/activeSearchPath :\n> + /* TODO: POC */\n> + if (equal(oidlist, baseSearchPath))\n> + return;\n\nThere's an awful lot missing from that sketch; all of the remaining\nsteps still need to be done:\nYou are correct, but that was intentionally skipped to avoid longer postdescriptions for the initial discussion. Sorry for being little lazy. \n baseCreationNamespace = firstNS;\n baseTempCreationPending = temp_missing;\n\n /* Mark the path valid. */\n baseSearchPathValid = true;\n namespaceUser = roleid;\n\n /* And make it active. */\n activeSearchPath = baseSearchPath;\n activeCreationNamespace = baseCreationNamespace;\n activeTempCreationPending = baseTempCreationPending;\n\n /* Clean up. */\n pfree(rawname);\n list_free(namelist);\n list_free(oidlist);\n\nMore to the point, I think the onus would be on the patch submitter\nto prove that the extra complexity had some measurable benefit.\nI really doubt that it would, since the list_copy is surely trivial\ncompared to the catalog lookup work we had to do to compute the OID\nlist above here.Agree. \nIt'd likely be more useful to see if you could reduce the number of\nplaces where we have to invalidate the path in the first place.Understood, let me check.Regards,Amul","msg_date":"Mon, 4 Nov 2019 10:33:25 +0530","msg_from":"amul sul <sulamul@gmail.com>","msg_from_op":true,"msg_subject":"Re: Can avoid list_copy in recomputeNamespacePath() conditionally?"}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\n\\nI wondered can we have a shortcut somewhat similar to following POC\\nin recomputeNamespacePath () when the recomputed path is the same as the\\nprevious baseSearchPath/activeSearchPath :\\n\\n== POC patch ==\\ndiff --git a/src/backend/catalog/namespace.c\\nb/src/backend/catalog/namespace.c\\nindex e251f5a9fdc..b25ef489e47 100644\\n--- a/src/backend/catalog/namespace.c\\n+++ b/src/backend/catalog/namespace.c\\n@@ -3813,6 +3813,9 @@ recomputeNamespacePath(void)\\n !list_member_oid(oidlist, myTempNamespace))\\n oidlist = lcons_oid(myTempNamespace, oidlist);\\n\\n+ /* TODO: POC */\\n+ if (equal(oidlist, baseSearchPath))\\n+ return;\\n /*\\n * Now that we've successfully built the new list of namespace OIDs,\\nsave\\n * it in permanent storage.\\n== POC patch end ==\\n\\nIt can have two advantages as:\\n\\n1. Avoid unnecessary list_copy() in TopMemoryContext context &\\n2. Global pointers like activeSearchPath/baseSearchPath will not change if\\nsome\\n implementation end up with cascaded call to recomputeNamespacePath().\\n\\nThoughts/Comments?\\n\\nRegards,\\nAmul\\n\\nHi all,I wondered can we have a shortcut somewhat similar to following POCin recomputeNamespacePath () when the recomputed path is the same as theprevious baseSearchPath/activeSearchPath :== POC patch ==diff --git a/src/backend/catalog/namespace.c b/src/backend/catalog/namespace.cindex e251f5a9fdc..b25ef489e47 100644--- a/src/backend/catalog/namespace.c+++ b/src/backend/catalog/namespace.c@@ -3813,6 +3813,9 @@ recomputeNamespacePath(void) !list_member_oid(oidlist, myTempNamespace)) oidlist = lcons_oid(myTempNamespace, oidlist); + /* TODO: POC */+ if (equal(oidlist, baseSearchPath))+ return; /* * Now that we've successfully built the new list of namespace OIDs, save * it in permanent storage.== POC patch end ==It can have two advantages as:1. Avoid unnecessary list_copy() in TopMemoryContext context &2. Global pointers like activeSearchPath/baseSearchPath will not change if some implementation end up with cascaded call to recomputeNamespacePath().Thoughts/Comments?Regards,Amul\",\n \"msg_date\": \"Thu, 31 Oct 2019 14:11:03 +0530\",\n \"msg_from\": \"amul sul <sulamul@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Can avoid list_copy in recomputeNamespacePath() conditionally?\"\n },\n {\n \"msg_contents\": \"amul sul <sulamul@gmail.com> writes:\\n> I wondered can we have a shortcut somewhat similar to following POC\\n> in recomputeNamespacePath () when the recomputed path is the same as the\\n> previous baseSearchPath/activeSearchPath :\\n> + /* TODO: POC */\\n> + if (equal(oidlist, baseSearchPath))\\n> + return;\\n\\nThere's an awful lot missing from that sketch; all of the remaining\\nsteps still need to be done:\\n\\n\\tbaseCreationNamespace = firstNS;\\n\\tbaseTempCreationPending = temp_missing;\\n\\n\\t/* Mark the path valid. */\\n\\tbaseSearchPathValid = true;\\n\\tnamespaceUser = roleid;\\n\\n\\t/* And make it active. */\\n\\tactiveSearchPath = baseSearchPath;\\n\\tactiveCreationNamespace = baseCreationNamespace;\\n\\tactiveTempCreationPending = baseTempCreationPending;\\n\\n\\t/* Clean up. */\\n\\tpfree(rawname);\\n\\tlist_free(namelist);\\n\\tlist_free(oidlist);\\n\\nMore to the point, I think the onus would be on the patch submitter\\nto prove that the extra complexity had some measurable benefit.\\nI really doubt that it would, since the list_copy is surely trivial\\ncompared to the catalog lookup work we had to do to compute the OID\\nlist above here.\\n\\nIt'd likely be more useful to see if you could reduce the number of\\nplaces where we have to invalidate the path in the first place.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 02 Nov 2019 10:31:14 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Can avoid list_copy in recomputeNamespacePath() conditionally?\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 2, 2019 at 8:01 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> amul sul <sulamul@gmail.com> writes:\\n> > I wondered can we have a shortcut somewhat similar to following POC\\n> > in recomputeNamespacePath () when the recomputed path is the same as the\\n> > previous baseSearchPath/activeSearchPath :\\n> > + /* TODO: POC */\\n> > + if (equal(oidlist, baseSearchPath))\\n> > + return;\\n>\\n> There's an awful lot missing from that sketch; all of the remaining\\n> steps still need to be done:\\n>\\n>\\nYou are correct, but that was intentionally skipped to avoid longer post\\ndescriptions for the initial discussion. Sorry for being little lazy.\\n\\n\\n> baseCreationNamespace = firstNS;\\n> baseTempCreationPending = temp_missing;\\n>\\n> /* Mark the path valid. */\\n> baseSearchPathValid = true;\\n> namespaceUser = roleid;\\n>\\n> /* And make it active. */\\n> activeSearchPath = baseSearchPath;\\n> activeCreationNamespace = baseCreationNamespace;\\n> activeTempCreationPending = baseTempCreationPending;\\n>\\n> /* Clean up. */\\n> pfree(rawname);\\n> list_free(namelist);\\n> list_free(oidlist);\\n>\\n> More to the point, I think the onus would be on the patch submitter\\n> to prove that the extra complexity had some measurable benefit.\\n> I really doubt that it would, since the list_copy is surely trivial\\n> compared to the catalog lookup work we had to do to compute the OID\\n> list above here.\\n>\\n\\nAgree.\\n\\n\\n> It'd likely be more useful to see if you could reduce the number of\\n> places where we have to invalidate the path in the first place.\\n>\\n\\nUnderstood, let me check.\\n\\nRegards,\\nAmul\\n\\nOn Sat, Nov 2, 2019 at 8:01 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:amul sul <sulamul@gmail.com> writes:\\n> I wondered can we have a shortcut somewhat similar to following POC\\n> in recomputeNamespacePath () when the recomputed path is the same as the\\n> previous baseSearchPath/activeSearchPath :\\n> + /* TODO: POC */\\n> + if (equal(oidlist, baseSearchPath))\\n> + return;\\n\\nThere's an awful lot missing from that sketch; all of the remaining\\nsteps still need to be done:\\nYou are correct, but that was intentionally skipped to avoid longer postdescriptions for the initial discussion. Sorry for being little lazy. \\n baseCreationNamespace = firstNS;\\n baseTempCreationPending = temp_missing;\\n\\n /* Mark the path valid. */\\n baseSearchPathValid = true;\\n namespaceUser = roleid;\\n\\n /* And make it active. */\\n activeSearchPath = baseSearchPath;\\n activeCreationNamespace = baseCreationNamespace;\\n activeTempCreationPending = baseTempCreationPending;\\n\\n /* Clean up. */\\n pfree(rawname);\\n list_free(namelist);\\n list_free(oidlist);\\n\\nMore to the point, I think the onus would be on the patch submitter\\nto prove that the extra complexity had some measurable benefit.\\nI really doubt that it would, since the list_copy is surely trivial\\ncompared to the catalog lookup work we had to do to compute the OID\\nlist above here.Agree. \\nIt'd likely be more useful to see if you could reduce the number of\\nplaces where we have to invalidate the path in the first place.Understood, let me check.Regards,Amul\",\n \"msg_date\": \"Mon, 4 Nov 2019 10:33:25 +0530\",\n \"msg_from\": \"amul sul <sulamul@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Can avoid list_copy in recomputeNamespacePath() conditionally?\"\n }\n]","isConversational":false}}},{"rowIdx":49371,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"AFAICT, these build options were only useful to maintain compatibility \nfor version-0 functions, but those are no longer supported, so these \noptions can be removed. There is a fair amount of code all over the \nplace to support these options, so the cleanup is quite significant.\n\nThe current behavior became the default in PG9.3. Note that this does \nnot affect on-disk storage. The only upgrade issue that I can see is \nthat pg_upgrade refuses to upgrade incompatible clusters if you have \ncontrib/isn installed. But hopefully everyone who is affected by that \nwill have upgraded at least once since PG9.2 already.\n\nfloat4 is now always pass-by-value; the pass-by-reference code path is \ncompletely removed.\n\nfloat8 and related types are now hardcoded to pass-by-value or \npass-by-reference depending on whether the build is 64- or 32-bit, as \nwas previously also the default.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 31 Oct 2019 09:50:16 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Remove configure --disable-float4-byval and --disable-float8-byval"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> float4 is now always pass-by-value; the pass-by-reference code path is \n> completely removed.\n\nI think this is OK.\n\n> float8 and related types are now hardcoded to pass-by-value or \n> pass-by-reference depending on whether the build is 64- or 32-bit, as \n> was previously also the default.\n\nI'm less happy with doing this. It makes it impossible to test the\npass-by-reference code paths without actually firing up a 32-bit\nenvironment. It'd be fine to document --disable-float8-byval as\na developer-only option (it might be so already), but I don't want\nto lose it completely. I fail to see any advantage in getting rid\nof it, anyway, since we do still have to maintain both code paths.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 31 Oct 2019 09:36:30 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Thu, Oct 31, 2019 at 9:36 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > float8 and related types are now hardcoded to pass-by-value or\n> > pass-by-reference depending on whether the build is 64- or 32-bit, as\n> > was previously also the default.\n>\n> I'm less happy with doing this. It makes it impossible to test the\n> pass-by-reference code paths without actually firing up a 32-bit\n> environment. It'd be fine to document --disable-float8-byval as\n> a developer-only option (it might be so already), but I don't want\n> to lose it completely. I fail to see any advantage in getting rid\n> of it, anyway, since we do still have to maintain both code paths.\n\nCould we get around this by making Datum 8 bytes everywhere?\n\nYears ago, we got rid of INT64_IS_BUSTED, so we're relying on 64-bit\ntypes working on all platforms. However, Datum on a system where\npointers are only 32 bits wide is also only 32 bits wide, so we can't\npass 64-bit quantities the same way we do elsewhere. But, why is the\nwidth of a Datum equal to the width of a pointer, anyway? It would\nsurely cost something to widen 1, 2, and 4 byte quantities to 8 bytes\nwhen packing them into datums on 32-bit platforms, but (1) we've long\nsince accepted that cost on 64-bit platforms, (2) it would save\npalloc/pfree cycles when packing 8 byte quantities into 4-byte values,\n(3) 32-bit platforms are now marginal and performance on those\nplatforms is not critical, and (4) it would simplify a lot of code and\nreduce future bugs.\n\nOn a related note, why do we store typbyval in the catalog anyway\ninstead of inferring it from typlen and maybe typalign? It seems like\na bad idea to record on disk the way we pass around values in memory,\nbecause it means that a change to how values are passed around in\nmemory has ramifications for on-disk compatibility.\n\nrhaas=# select typname, typlen, typbyval, typalign from pg_type where\ntyplen in (1,2,4,8) != typbyval;\n typname | typlen | typbyval | typalign\n----------+--------+----------+----------\n macaddr8 | 8 | f | i\n(1 row)\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 1 Nov 2019 10:41:19 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Fri, Nov 1, 2019 at 7:41 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> Could we get around this by making Datum 8 bytes everywhere?\n\nI really like that idea.\n\nEven Raspberry Pi devices (which can cost as little as $35) use 64-bit\nARM processors. It's abundantly clear that 32-bit platforms do not\nmatter enough to justify keeping all the SIZEOF_DATUM crud around.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Fri, 1 Nov 2019 08:22:27 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"Peter Geoghegan <pg@bowt.ie> writes:\n> On Fri, Nov 1, 2019 at 7:41 AM Robert Haas <robertmhaas@gmail.com> wrote:\n>> Could we get around this by making Datum 8 bytes everywhere?\n\n> I really like that idea.\n\n> Even Raspberry Pi devices (which can cost as little as $35) use 64-bit\n> ARM processors. It's abundantly clear that 32-bit platforms do not\n> matter enough to justify keeping all the SIZEOF_DATUM crud around.\n\nThis line of argument seems to me to be the moral equivalent of\n\"let's drop 32-bit support altogether\". I'm not entirely on board\nwith that. Certainly, a lot of the world is 64-bit these days,\nbut people are still building small systems and they might want\na database; preferably one that hasn't been detuned to the extent\nthat it barely manages to run at all on such a platform. Making\na whole lot of internal APIs 64-bit would be a pretty big hit for\na 32-bit platform --- more instructions, more memory consumed for\nthings like Datum arrays, all in a memory space that's not that big.\n\nIt seems especially insane to conclude that we should pull the plug\non such use-cases just to get rid of one obscure configure option.\nIf we were expending any significant devel effort on supporting\n32-bit platforms, I might be ready to drop support, but we're not.\n(Robert's proposal looks to me like it's actually creating new work\nto do, not saving work.)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 01 Nov 2019 14:00:10 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Fri, Nov 1, 2019 at 11:00 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> This line of argument seems to me to be the moral equivalent of\n> \"let's drop 32-bit support altogether\". I'm not entirely on board\n> with that.\n\nI don't think that those two things are equivalent at all. There may\neven be workloads that will benefit when run on 32-bit hardware.\nHaving to palloc() and pfree() with 8 byte integers is probably very\nslow.\n\n> Certainly, a lot of the world is 64-bit these days,\n> but people are still building small systems and they might want\n> a database; preferably one that hasn't been detuned to the extent\n> that it barely manages to run at all on such a platform.\n\nEven the very low end is 64-bit these days. $100 smartphones have\n64-bit CPUs and 4GB of ram. AFAICT, anything still being produced that\nis recognizable as a general purpose CPU (e.g. by having virtual\nmemory) is 64-bit. We're talking about a change that can't affect\nusers until late 2020 at the earliest.\n\nI accept that there are some number of users that still have 32-bit\nPostgres installations, probably because they happened to have a\n32-bit machine close at hand. The economics of running a database\nserver on a 32-bit machine are already awful, though.\n\n> It seems especially insane to conclude that we should pull the plug\n> on such use-cases just to get rid of one obscure configure option.\n> If we were expending any significant devel effort on supporting\n> 32-bit platforms, I might be ready to drop support, but we're not.\n> (Robert's proposal looks to me like it's actually creating new work\n> to do, not saving work.)\n\nThe mental burden of considering \"SIZEOF_DATUM == 4\" and\n\"USE_FLOAT8_BYVAL\" is a real cost for us. We maintain non-trivial\n32-bit only code for numeric abbreviated keys. We also have to worry\nabout pfree()'ing memory when USE_FLOAT8_BYVAL within\nheapam_index_validate_scan(). How confident are we that there isn't\nsome place that leaks memory on !USE_FLOAT8_BYVAL builds because\nsomebody forgot to add a pfree() in an #ifdef block?\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Fri, 1 Nov 2019 12:15:06 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Fri, Nov 1, 2019 at 3:15 PM Peter Geoghegan <pg@bowt.ie> wrote:\n> I don't think that those two things are equivalent at all. There may\n> even be workloads that will benefit when run on 32-bit hardware.\n> Having to palloc() and pfree() with 8 byte integers is probably very\n> slow.\n\nYeah! I mean, users who are using only 4-byte or smaller pass-by-value\nquantities will be harmed, especially in cases where they are storing\na lot of them at the same time (e.g. sorting) and especially if they\ndouble their space consumption and run out of their very limited\nsupply of memory. Those are all worthwhile considerations and perhaps\nstrong arguments against my proposal. But, people using 8-byte\noughta-be-pass-by-value quantities are certainly being harmed by the\npresent system. It's not a black-and-white thing, like, oh, 8-byte\ndatums on 32-bit system is awful all the time. Or at least, I don't\nthink it is.\n\n> The mental burden of considering \"SIZEOF_DATUM == 4\" and\n> \"USE_FLOAT8_BYVAL\" is a real cost for us. We maintain non-trivial\n> 32-bit only code for numeric abbreviated keys. We also have to worry\n> about pfree()'ing memory when USE_FLOAT8_BYVAL within\n> heapam_index_validate_scan(). How confident are we that there isn't\n> some place that leaks memory on !USE_FLOAT8_BYVAL builds because\n> somebody forgot to add a pfree() in an #ifdef block?\n\nYeah, I've had to fight with this multiple times, and so have other\npeople. It's a nuisance, and it causes bugs, certainly in draft\npatches, sometimes in committed ones. It's not \"free.\" If it's a cost\nworth paying, ok, but is it?\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 1 Nov 2019 16:19:12 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Fri, Nov 1, 2019 at 1:19 PM Robert Haas <robertmhaas@gmail.com> wrote:\n> Yeah! I mean, users who are using only 4-byte or smaller pass-by-value\n> quantities will be harmed, especially in cases where they are storing\n> a lot of them at the same time (e.g. sorting) and especially if they\n> double their space consumption and run out of their very limited\n> supply of memory.\n\nI think that you meant treble, not double. You're forgetting about the\npalloc() header overhead. ;-)\n\nDoing even slightly serious work on a 32-bit machine is penny wise and\npound foolish. I also believe that we should support minority\nplatforms reasonably well, including 32-bit platforms, because it's\nalways a good idea to try to meet people where they are. Your proposal\ndoesn't seem like it really gives up on that goal.\n\n> Those are all worthwhile considerations and perhaps\n> strong arguments against my proposal. But, people using 8-byte\n> oughta-be-pass-by-value quantities are certainly being harmed by the\n> present system. It's not a black-and-white thing, like, oh, 8-byte\n> datums on 32-bit system is awful all the time. Or at least, I don't\n> think it is.\n\nRight.\n\n> Yeah, I've had to fight with this multiple times, and so have other\n> people. It's a nuisance, and it causes bugs, certainly in draft\n> patches, sometimes in committed ones. It's not \"free.\" If it's a cost\n> worth paying, ok, but is it?\n\n#ifdef crud is something that we should go out of our way to eliminate\non general principle. All good portable C codebases go to great\nlengths to encapsulate platform differences, if necessary by adding a\ncompatibility layer. One of the worst things about the OpenSSL\ncodebase is that it makes writing portable code everybody's problem.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Fri, 1 Nov 2019 13:49:38 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Fri, Nov 01, 2019 at 02:00:10PM -0400, Tom Lane wrote:\n> Peter Geoghegan <pg@bowt.ie> writes:\n>> Even Raspberry Pi devices (which can cost as little as $35) use 64-bit\n>> ARM processors. It's abundantly clear that 32-bit platforms do not\n>> matter enough to justify keeping all the SIZEOF_DATUM crud around.\n> \n> This line of argument seems to me to be the moral equivalent of\n> \"let's drop 32-bit support altogether\". I'm not entirely on board\n> with that. Certainly, a lot of the world is 64-bit these days,\n> but people are still building small systems and they might want\n> a database; preferably one that hasn't been detuned to the extent\n> that it barely manages to run at all on such a platform. Making\n> a whole lot of internal APIs 64-bit would be a pretty big hit for\n> a 32-bit platform --- more instructions, more memory consumed for\n> things like Datum arrays, all in a memory space that's not that big.\n\nI don't agree as well with the line of arguments to just remove 32b\nsupport. The newest models of PI indeed use 64b ARM processors, but\nthe first model, as well as the PI zero are on 32b if I recall\ncorrectly, and I would like to believe that these are still widely\nused.\n--\nMichael","msg_date":"Sat, 2 Nov 2019 11:47:26 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On 2019-11-02 11:47:26 +0900, Michael Paquier wrote:\n> On Fri, Nov 01, 2019 at 02:00:10PM -0400, Tom Lane wrote:\n> > Peter Geoghegan <pg@bowt.ie> writes:\n> >> Even Raspberry Pi devices (which can cost as little as $35) use 64-bit\n> >> ARM processors. It's abundantly clear that 32-bit platforms do not\n> >> matter enough to justify keeping all the SIZEOF_DATUM crud around.\n> > \n> > This line of argument seems to me to be the moral equivalent of\n> > \"let's drop 32-bit support altogether\". I'm not entirely on board\n> > with that. Certainly, a lot of the world is 64-bit these days,\n> > but people are still building small systems and they might want\n> > a database; preferably one that hasn't been detuned to the extent\n> > that it barely manages to run at all on such a platform. Making\n> > a whole lot of internal APIs 64-bit would be a pretty big hit for\n> > a 32-bit platform --- more instructions, more memory consumed for\n> > things like Datum arrays, all in a memory space that's not that big.\n> \n> I don't agree as well with the line of arguments to just remove 32b\n> support.\n\nNobody is actually proposing that, as far as I can tell? I mean, by all\nmeans argue that the overhead is too high, but just claiming that\nslowing down 32bit systems by a measurable fraction is morally\nequivalent to removing 32bit support seems pointlessly facetious.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 1 Nov 2019 20:14:47 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Fri, Nov 1, 2019 at 7:47 PM Michael Paquier <michael@paquier.xyz> wrote:\n> > This line of argument seems to me to be the moral equivalent of\n> > \"let's drop 32-bit support altogether\". I'm not entirely on board\n> > with that. Certainly, a lot of the world is 64-bit these days,\n> > but people are still building small systems and they might want\n> > a database; preferably one that hasn't been detuned to the extent\n> > that it barely manages to run at all on such a platform. Making\n> > a whole lot of internal APIs 64-bit would be a pretty big hit for\n> > a 32-bit platform --- more instructions, more memory consumed for\n> > things like Datum arrays, all in a memory space that's not that big.\n>\n> I don't agree as well with the line of arguments to just remove 32b\n> support.\n\nClearly you didn't read what I actually wrote, Michael.\n\n--\nPeter Geoghegan\n\n\n","msg_date":"Fri, 1 Nov 2019 21:41:58 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On 2019-10-31 14:36, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n>> float4 is now always pass-by-value; the pass-by-reference code path is\n>> completely removed.\n> \n> I think this is OK.\n\nOK, here is a patch for just this part, and we can continue the \ndiscussion on the rest in the meantime.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Sat, 2 Nov 2019 08:39:16 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On 2019-11-01 15:41, Robert Haas wrote:\n> On a related note, why do we store typbyval in the catalog anyway\n> instead of inferring it from typlen and maybe typalign? It seems like\n> a bad idea to record on disk the way we pass around values in memory,\n> because it means that a change to how values are passed around in\n> memory has ramifications for on-disk compatibility.\n\nThis sounds interesting. It would remove a pg_upgrade hazard (in the \nlong run).\n\nThere is some backward compatibility to be concerned about. This change \nwould require extension authors to change their code to insert #ifdef \nUSE_FLOAT8_BYVAL or similar, where currently their code might only \nsupport one method or the other.\n\n> rhaas=# select typname, typlen, typbyval, typalign from pg_type where\n> typlen in (1,2,4,8) != typbyval;\n\nThere are also typlen=6 types. Who knew. ;-)\n\n> typname | typlen | typbyval | typalign\n> ----------+--------+----------+----------\n> macaddr8 | 8 | f | i\n> (1 row)\n\nThis might be a case of the above issue: It's easier to just make it \npass by reference always than deal with a bunch of #ifdefs.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Sat, 2 Nov 2019 08:46:12 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> On 2019-11-01 15:41, Robert Haas wrote:\n>> On a related note, why do we store typbyval in the catalog anyway\n>> instead of inferring it from typlen and maybe typalign? It seems like\n>> a bad idea to record on disk the way we pass around values in memory,\n>> because it means that a change to how values are passed around in\n>> memory has ramifications for on-disk compatibility.\n\n> This sounds interesting. It would remove a pg_upgrade hazard (in the \n> long run).\n\n> There is some backward compatibility to be concerned about.\n\nYeah. The point here is that typbyval specifies what the C functions\nconcerned with the datatype are expecting. We can't just up and say\n\"we're going to decide that for you\".\n\nI do get the point that supporting two different typbyval options\nfor float8 and related types is a nontrivial cost.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 02 Nov 2019 11:26:59 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"Hi,\n\nOn 2019-11-02 08:46:12 +0100, Peter Eisentraut wrote:\n> On 2019-11-01 15:41, Robert Haas wrote:\n> > On a related note, why do we store typbyval in the catalog anyway\n> > instead of inferring it from typlen and maybe typalign? It seems like\n> > a bad idea to record on disk the way we pass around values in memory,\n> > because it means that a change to how values are passed around in\n> > memory has ramifications for on-disk compatibility.\n>\n> This sounds interesting. It would remove a pg_upgrade hazard (in the long\n> run).\n>\n> There is some backward compatibility to be concerned about. This change\n> would require extension authors to change their code to insert #ifdef\n> USE_FLOAT8_BYVAL or similar, where currently their code might only support\n> one method or the other.\n\nI think we really ought to remove the difference behind macros. That is,\nfor types that could be either, provide macros that fetch function\narguments into local memory, independent of whether the argument is a\nbyval or byref type. I.e. instead of having separate #ifdef\nUSE_FLOAT8_BYVALs for DatumGetFloat8(), DatumGetInt64(), ... we should\nprovide that logic in one centralized set of macros.\n\nThe fact that USE_FLOAT8_BYVAL has to creep into various functions imo\nis the reasons why people are unhappy about it.\n\nOne way to do this would be something roughly like this sketch:\n\n/* allow to force types to be byref, for testing purposes */\n#if USE_FLOAT8_BYVAL\n#define DatumForTypeIsByval(type) (sizeof(type) <= SIZEOF_DATUM)\n#else\n#define DatumForTypeIsByval(type) (sizeof(type) <= sizeof(int))\n#endif\n\n/* yes, I dream of being C++ once grown up */\n#define DefineSmallFixedWidthDatumTypeConversions(type, TypeToDatumName, DatumToTypeName) \\\n static inline type \\\n TypeToDatumName (type value) \\\n { \\\n if (DatumForTypeIsByval(type)) \\\n { \\\n Datum tmp; \\\n \\\n /* ensure correct conversion, consider e.g. float */ \\\n memcpy(&tmp, &value, sizeof(type)); \\\n \\\n return tmp; \\\n } \\\n else \\\n { \\\n type *tmp = (type *) palloc(sizeof(datum)); \\\n\n *tmp = value;\n\n return PointerGetDatum(tmp); \\\n } \\\n } \\\n \\\n static inline type \\\n DatumToTypeName (Datum datum) \\\n { \\\n if (DatumForTypeIsByval(type)) \\\n { \\\n type tmp; \\\n \\\n /* ensure correct conversion */ \\\n memcpy(&tmp, &datum, sizeof(type)); \\\n \\\n return tmp; \\\n } \\\n else \\\n return *(type *) DatumGetPointer(type); \\\n } \\\n\nAnd then have\n\nDefineSmallFixedWidthDatumTypeConversions(\n float8,\n Float8GetDatum,\n DatumGetFloat8);\n\nDefineSmallFixedWidthDatumTypeConversions(\n int64,\n Int64GetDatum,\n DatumGetInt64);\n\nAnd now also\n\nDefineSmallFixedWidthDatumTypeConversions(\n macaddr,\n Macaddr8GetDatum,\n DatumGetMacaddr8);\n\n(there's probably plenty of bugs in the above, it's just a sketch)\n\n\nWe don't have to break types / extensions with inferring byval for fixed\nwidth types. Instead we can change CREATE TYPE's PASSEDBYVALUE to accept\nan optional parameter 'auto', allowing to opt in.\n\n\n> > rhaas=# select typname, typlen, typbyval, typalign from pg_type where\n> > typlen in (1,2,4,8) != typbyval;\n>\n> There are also typlen=6 types. Who knew. ;-)\n\nThere's a recent thread about them :)\n\n\n> > typname | typlen | typbyval | typalign\n> > ----------+--------+----------+----------\n> > macaddr8 | 8 | f | i\n> > (1 row)\n>\n> This might be a case of the above issue: It's easier to just make it pass by\n> reference always than deal with a bunch of #ifdefs.\n\nIndeed. And that's a bad sign imo.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sat, 2 Nov 2019 17:00:17 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Fri, Nov 01, 2019 at 09:41:58PM -0700, Peter Geoghegan wrote:\n> On Fri, Nov 1, 2019 at 7:47 PM Michael Paquier <michael@paquier.xyz> wrote:\n>> I don't agree as well with the line of arguments to just remove 32b\n>> support.\n> \n> Clearly you didn't read what I actually wrote, Michael.\n\nSorry, that was an misunderstanding from my side.\n--\nMichael","msg_date":"Thu, 7 Nov 2019 15:47:30 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Sat, Nov 2, 2019 at 8:00 PM Andres Freund <andres@anarazel.de> wrote:\n> I think we really ought to remove the difference behind macros. That is,\n> for types that could be either, provide macros that fetch function\n> arguments into local memory, independent of whether the argument is a\n> byval or byref type. I.e. instead of having separate #ifdef\n> USE_FLOAT8_BYVALs for DatumGetFloat8(), DatumGetInt64(), ... we should\n> provide that logic in one centralized set of macros.\n>\n> The fact that USE_FLOAT8_BYVAL has to creep into various functions imo\n> is the reasons why people are unhappy about it.\n\nI think I'm *more* unhappy about the fact that it affects a bunch of\nthings that are not about whether float8 is passed byval. I mean, you\nmention DatumGetInt64() above, but why in the world does a setting\nwith \"float8\" in the name affect how we pass int64? The thing is like\nkudzu, getting into all sorts of things that it has no business\naffecting - at least if you judge by the name - and for no really\nclear reason.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 13 Nov 2019 07:36:27 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On 2019-11-02 08:39, Peter Eisentraut wrote:\n> On 2019-10-31 14:36, Tom Lane wrote:\n>> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n>>> float4 is now always pass-by-value; the pass-by-reference code path is\n>>> completely removed.\n>>\n>> I think this is OK.\n> \n> OK, here is a patch for just this part, and we can continue the\n> discussion on the rest in the meantime.\n\nI have committed this part.\n\nI will rebase and continue developing the rest of the patches based on \nthe discussion so far.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 21 Nov 2019 19:20:28 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Thu, Nov 21, 2019 at 07:20:28PM +0100, Peter Eisentraut wrote:\n> I have committed this part.\n> \n> I will rebase and continue developing the rest of the patches based on the\n> discussion so far.\n\nBased on that I am marking the patch as committed in the CF. The rest\nof the patch set could have a new entry.\n--\nMichael","msg_date":"Mon, 25 Nov 2019 11:31:58 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"My revised proposal is to remove --disable-float8-byval as a configure \noption but keep it as an option in pg_config_manual.h. It is no longer \nuseful as a user-facing option, but as was pointed out, it is somewhat \nuseful for developers, so pg_config_manual.h seems like the right place.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Tue, 26 Nov 2019 21:27:40 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> My revised proposal is to remove --disable-float8-byval as a configure \n> option but keep it as an option in pg_config_manual.h. It is no longer \n> useful as a user-facing option, but as was pointed out, it is somewhat \n> useful for developers, so pg_config_manual.h seems like the right place.\n\nOK, works for me.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 26 Nov 2019 15:33:14 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On 2019-11-26 21:33, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n>> My revised proposal is to remove --disable-float8-byval as a configure\n>> option but keep it as an option in pg_config_manual.h. It is no longer\n>> useful as a user-facing option, but as was pointed out, it is somewhat\n>> useful for developers, so pg_config_manual.h seems like the right place.\n> \n> OK, works for me.\n\ndone\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Wed, 27 Nov 2019 13:30:29 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Fri, Nov 1, 2019 at 1:19 PM Robert Haas <robertmhaas@gmail.com> wrote:\n> On Fri, Nov 1, 2019 at 3:15 PM Peter Geoghegan <pg@bowt.ie> wrote:\n> > I don't think that those two things are equivalent at all. There may\n> > even be workloads that will benefit when run on 32-bit hardware.\n> > Having to palloc() and pfree() with 8 byte integers is probably very\n> > slow.\n>\n> Yeah! I mean, users who are using only 4-byte or smaller pass-by-value\n> quantities will be harmed, especially in cases where they are storing\n> a lot of them at the same time (e.g. sorting) and especially if they\n> double their space consumption and run out of their very limited\n> supply of memory.\n\nApparently Linux has almost no upstream resources for testing 32-bit\nx86, and it shows:\n\nhttps://lwn.net/ml/oss-security/CALCETrW1z0gCLFJz-1Jwj_wcT3+axXkP_wOCxY8JkbSLzV80GA@mail.gmail.com/\n\nI think that this kind of thing argues for minimizing the amount of\ncode that can only be tested on a small minority of the computers that\nare in general use today. If no Postgres hacker regularly runs the\ncode, then its chances of having bugs are far higher. Having coverage\nin the buildfarm certainly helps, but it's no substitute.\n\nSticking with the !USE_FLOAT8_BYVAL example, it's easy to imagine\nsomebody forgetting to add a !USE_FLOAT8_BYVAL block, that contains\nthe required pfree(). Now you have a memory leak that only affects a\nsmall minority of platforms. How likely is it that buildfarm coverage\nwill help somebody detect that problem?\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Thu, 12 Dec 2019 14:06:03 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On 2019-12-12 23:06, Peter Geoghegan wrote:\n> Apparently Linux has almost no upstream resources for testing 32-bit\n> x86, and it shows:\n\nBut isn't 32-bit Windows still a thing? Or does that work differently?\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 13 Dec 2019 13:33:21 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"On Fri, Dec 13, 2019 at 6:33 AM Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n> On 2019-12-12 23:06, Peter Geoghegan wrote:\n> > Apparently Linux has almost no upstream resources for testing 32-bit\n> > x86, and it shows:\n>\n> But isn't 32-bit Windows still a thing? Or does that work differently?\n\nWell, again, I think the proposal here is not get rid of 32-bit\nsupport, but to have less code that only gets regularly tested on\n32-bit machines. If we made datums 8 bytes everywhere, we would have\nless such code, and very likely fewer bugs. And as pointed out\nupthread, although some things might perform worse for the remaining\nsupply of 32-bit users, other things might perform better. I'm not\n100% sure that it would work out to a win overall, but I think there's\na good chance, especially when you factor in the reduced bug surface.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 13 Dec 2019 09:43:39 -0600","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> Well, again, I think the proposal here is not get rid of 32-bit\n> support, but to have less code that only gets regularly tested on\n> 32-bit machines.\n\nThat seems like generally a good plan. But as to the specific idea...\n\n> If we made datums 8 bytes everywhere, we would have\n> less such code, and very likely fewer bugs.\n\n... it's not entirely clear to me that it'd be possible to do this\nwithout causing a storm of \"cast from pointer to integer of different\nsize\" (and vice versa) warnings on 32-bit machines. That would be a\ndeal-breaker independently of any performance considerations, IMO.\nSo if anyone wants to pursue this, finding a way around that might be\nthe first thing to look at.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 13 Dec 2019 11:25:37 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Remove configure --disable-float4-byval and\n --disable-float8-byval"}],"string":"[\n {\n \"msg_contents\": \"AFAICT, these build options were only useful to maintain compatibility \\nfor version-0 functions, but those are no longer supported, so these \\noptions can be removed. There is a fair amount of code all over the \\nplace to support these options, so the cleanup is quite significant.\\n\\nThe current behavior became the default in PG9.3. Note that this does \\nnot affect on-disk storage. The only upgrade issue that I can see is \\nthat pg_upgrade refuses to upgrade incompatible clusters if you have \\ncontrib/isn installed. But hopefully everyone who is affected by that \\nwill have upgraded at least once since PG9.2 already.\\n\\nfloat4 is now always pass-by-value; the pass-by-reference code path is \\ncompletely removed.\\n\\nfloat8 and related types are now hardcoded to pass-by-value or \\npass-by-reference depending on whether the build is 64- or 32-bit, as \\nwas previously also the default.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 31 Oct 2019 09:50:16 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Remove configure --disable-float4-byval and --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> float4 is now always pass-by-value; the pass-by-reference code path is \\n> completely removed.\\n\\nI think this is OK.\\n\\n> float8 and related types are now hardcoded to pass-by-value or \\n> pass-by-reference depending on whether the build is 64- or 32-bit, as \\n> was previously also the default.\\n\\nI'm less happy with doing this. It makes it impossible to test the\\npass-by-reference code paths without actually firing up a 32-bit\\nenvironment. It'd be fine to document --disable-float8-byval as\\na developer-only option (it might be so already), but I don't want\\nto lose it completely. I fail to see any advantage in getting rid\\nof it, anyway, since we do still have to maintain both code paths.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 09:36:30 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 9:36 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > float8 and related types are now hardcoded to pass-by-value or\\n> > pass-by-reference depending on whether the build is 64- or 32-bit, as\\n> > was previously also the default.\\n>\\n> I'm less happy with doing this. It makes it impossible to test the\\n> pass-by-reference code paths without actually firing up a 32-bit\\n> environment. It'd be fine to document --disable-float8-byval as\\n> a developer-only option (it might be so already), but I don't want\\n> to lose it completely. I fail to see any advantage in getting rid\\n> of it, anyway, since we do still have to maintain both code paths.\\n\\nCould we get around this by making Datum 8 bytes everywhere?\\n\\nYears ago, we got rid of INT64_IS_BUSTED, so we're relying on 64-bit\\ntypes working on all platforms. However, Datum on a system where\\npointers are only 32 bits wide is also only 32 bits wide, so we can't\\npass 64-bit quantities the same way we do elsewhere. But, why is the\\nwidth of a Datum equal to the width of a pointer, anyway? It would\\nsurely cost something to widen 1, 2, and 4 byte quantities to 8 bytes\\nwhen packing them into datums on 32-bit platforms, but (1) we've long\\nsince accepted that cost on 64-bit platforms, (2) it would save\\npalloc/pfree cycles when packing 8 byte quantities into 4-byte values,\\n(3) 32-bit platforms are now marginal and performance on those\\nplatforms is not critical, and (4) it would simplify a lot of code and\\nreduce future bugs.\\n\\nOn a related note, why do we store typbyval in the catalog anyway\\ninstead of inferring it from typlen and maybe typalign? It seems like\\na bad idea to record on disk the way we pass around values in memory,\\nbecause it means that a change to how values are passed around in\\nmemory has ramifications for on-disk compatibility.\\n\\nrhaas=# select typname, typlen, typbyval, typalign from pg_type where\\ntyplen in (1,2,4,8) != typbyval;\\n typname | typlen | typbyval | typalign\\n----------+--------+----------+----------\\n macaddr8 | 8 | f | i\\n(1 row)\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 10:41:19 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 7:41 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> Could we get around this by making Datum 8 bytes everywhere?\\n\\nI really like that idea.\\n\\nEven Raspberry Pi devices (which can cost as little as $35) use 64-bit\\nARM processors. It's abundantly clear that 32-bit platforms do not\\nmatter enough to justify keeping all the SIZEOF_DATUM crud around.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 08:22:27 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"Peter Geoghegan <pg@bowt.ie> writes:\\n> On Fri, Nov 1, 2019 at 7:41 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n>> Could we get around this by making Datum 8 bytes everywhere?\\n\\n> I really like that idea.\\n\\n> Even Raspberry Pi devices (which can cost as little as $35) use 64-bit\\n> ARM processors. It's abundantly clear that 32-bit platforms do not\\n> matter enough to justify keeping all the SIZEOF_DATUM crud around.\\n\\nThis line of argument seems to me to be the moral equivalent of\\n\\\"let's drop 32-bit support altogether\\\". I'm not entirely on board\\nwith that. Certainly, a lot of the world is 64-bit these days,\\nbut people are still building small systems and they might want\\na database; preferably one that hasn't been detuned to the extent\\nthat it barely manages to run at all on such a platform. Making\\na whole lot of internal APIs 64-bit would be a pretty big hit for\\na 32-bit platform --- more instructions, more memory consumed for\\nthings like Datum arrays, all in a memory space that's not that big.\\n\\nIt seems especially insane to conclude that we should pull the plug\\non such use-cases just to get rid of one obscure configure option.\\nIf we were expending any significant devel effort on supporting\\n32-bit platforms, I might be ready to drop support, but we're not.\\n(Robert's proposal looks to me like it's actually creating new work\\nto do, not saving work.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 01 Nov 2019 14:00:10 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 11:00 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> This line of argument seems to me to be the moral equivalent of\\n> \\\"let's drop 32-bit support altogether\\\". I'm not entirely on board\\n> with that.\\n\\nI don't think that those two things are equivalent at all. There may\\neven be workloads that will benefit when run on 32-bit hardware.\\nHaving to palloc() and pfree() with 8 byte integers is probably very\\nslow.\\n\\n> Certainly, a lot of the world is 64-bit these days,\\n> but people are still building small systems and they might want\\n> a database; preferably one that hasn't been detuned to the extent\\n> that it barely manages to run at all on such a platform.\\n\\nEven the very low end is 64-bit these days. $100 smartphones have\\n64-bit CPUs and 4GB of ram. AFAICT, anything still being produced that\\nis recognizable as a general purpose CPU (e.g. by having virtual\\nmemory) is 64-bit. We're talking about a change that can't affect\\nusers until late 2020 at the earliest.\\n\\nI accept that there are some number of users that still have 32-bit\\nPostgres installations, probably because they happened to have a\\n32-bit machine close at hand. The economics of running a database\\nserver on a 32-bit machine are already awful, though.\\n\\n> It seems especially insane to conclude that we should pull the plug\\n> on such use-cases just to get rid of one obscure configure option.\\n> If we were expending any significant devel effort on supporting\\n> 32-bit platforms, I might be ready to drop support, but we're not.\\n> (Robert's proposal looks to me like it's actually creating new work\\n> to do, not saving work.)\\n\\nThe mental burden of considering \\\"SIZEOF_DATUM == 4\\\" and\\n\\\"USE_FLOAT8_BYVAL\\\" is a real cost for us. We maintain non-trivial\\n32-bit only code for numeric abbreviated keys. We also have to worry\\nabout pfree()'ing memory when USE_FLOAT8_BYVAL within\\nheapam_index_validate_scan(). How confident are we that there isn't\\nsome place that leaks memory on !USE_FLOAT8_BYVAL builds because\\nsomebody forgot to add a pfree() in an #ifdef block?\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 12:15:06 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 3:15 PM Peter Geoghegan <pg@bowt.ie> wrote:\\n> I don't think that those two things are equivalent at all. There may\\n> even be workloads that will benefit when run on 32-bit hardware.\\n> Having to palloc() and pfree() with 8 byte integers is probably very\\n> slow.\\n\\nYeah! I mean, users who are using only 4-byte or smaller pass-by-value\\nquantities will be harmed, especially in cases where they are storing\\na lot of them at the same time (e.g. sorting) and especially if they\\ndouble their space consumption and run out of their very limited\\nsupply of memory. Those are all worthwhile considerations and perhaps\\nstrong arguments against my proposal. But, people using 8-byte\\noughta-be-pass-by-value quantities are certainly being harmed by the\\npresent system. It's not a black-and-white thing, like, oh, 8-byte\\ndatums on 32-bit system is awful all the time. Or at least, I don't\\nthink it is.\\n\\n> The mental burden of considering \\\"SIZEOF_DATUM == 4\\\" and\\n> \\\"USE_FLOAT8_BYVAL\\\" is a real cost for us. We maintain non-trivial\\n> 32-bit only code for numeric abbreviated keys. We also have to worry\\n> about pfree()'ing memory when USE_FLOAT8_BYVAL within\\n> heapam_index_validate_scan(). How confident are we that there isn't\\n> some place that leaks memory on !USE_FLOAT8_BYVAL builds because\\n> somebody forgot to add a pfree() in an #ifdef block?\\n\\nYeah, I've had to fight with this multiple times, and so have other\\npeople. It's a nuisance, and it causes bugs, certainly in draft\\npatches, sometimes in committed ones. It's not \\\"free.\\\" If it's a cost\\nworth paying, ok, but is it?\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 16:19:12 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 1:19 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n> Yeah! I mean, users who are using only 4-byte or smaller pass-by-value\\n> quantities will be harmed, especially in cases where they are storing\\n> a lot of them at the same time (e.g. sorting) and especially if they\\n> double their space consumption and run out of their very limited\\n> supply of memory.\\n\\nI think that you meant treble, not double. You're forgetting about the\\npalloc() header overhead. ;-)\\n\\nDoing even slightly serious work on a 32-bit machine is penny wise and\\npound foolish. I also believe that we should support minority\\nplatforms reasonably well, including 32-bit platforms, because it's\\nalways a good idea to try to meet people where they are. Your proposal\\ndoesn't seem like it really gives up on that goal.\\n\\n> Those are all worthwhile considerations and perhaps\\n> strong arguments against my proposal. But, people using 8-byte\\n> oughta-be-pass-by-value quantities are certainly being harmed by the\\n> present system. It's not a black-and-white thing, like, oh, 8-byte\\n> datums on 32-bit system is awful all the time. Or at least, I don't\\n> think it is.\\n\\nRight.\\n\\n> Yeah, I've had to fight with this multiple times, and so have other\\n> people. It's a nuisance, and it causes bugs, certainly in draft\\n> patches, sometimes in committed ones. It's not \\\"free.\\\" If it's a cost\\n> worth paying, ok, but is it?\\n\\n#ifdef crud is something that we should go out of our way to eliminate\\non general principle. All good portable C codebases go to great\\nlengths to encapsulate platform differences, if necessary by adding a\\ncompatibility layer. One of the worst things about the OpenSSL\\ncodebase is that it makes writing portable code everybody's problem.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 13:49:38 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 01, 2019 at 02:00:10PM -0400, Tom Lane wrote:\\n> Peter Geoghegan <pg@bowt.ie> writes:\\n>> Even Raspberry Pi devices (which can cost as little as $35) use 64-bit\\n>> ARM processors. It's abundantly clear that 32-bit platforms do not\\n>> matter enough to justify keeping all the SIZEOF_DATUM crud around.\\n> \\n> This line of argument seems to me to be the moral equivalent of\\n> \\\"let's drop 32-bit support altogether\\\". I'm not entirely on board\\n> with that. Certainly, a lot of the world is 64-bit these days,\\n> but people are still building small systems and they might want\\n> a database; preferably one that hasn't been detuned to the extent\\n> that it barely manages to run at all on such a platform. Making\\n> a whole lot of internal APIs 64-bit would be a pretty big hit for\\n> a 32-bit platform --- more instructions, more memory consumed for\\n> things like Datum arrays, all in a memory space that's not that big.\\n\\nI don't agree as well with the line of arguments to just remove 32b\\nsupport. The newest models of PI indeed use 64b ARM processors, but\\nthe first model, as well as the PI zero are on 32b if I recall\\ncorrectly, and I would like to believe that these are still widely\\nused.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 2 Nov 2019 11:47:26 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On 2019-11-02 11:47:26 +0900, Michael Paquier wrote:\\n> On Fri, Nov 01, 2019 at 02:00:10PM -0400, Tom Lane wrote:\\n> > Peter Geoghegan <pg@bowt.ie> writes:\\n> >> Even Raspberry Pi devices (which can cost as little as $35) use 64-bit\\n> >> ARM processors. It's abundantly clear that 32-bit platforms do not\\n> >> matter enough to justify keeping all the SIZEOF_DATUM crud around.\\n> > \\n> > This line of argument seems to me to be the moral equivalent of\\n> > \\\"let's drop 32-bit support altogether\\\". I'm not entirely on board\\n> > with that. Certainly, a lot of the world is 64-bit these days,\\n> > but people are still building small systems and they might want\\n> > a database; preferably one that hasn't been detuned to the extent\\n> > that it barely manages to run at all on such a platform. Making\\n> > a whole lot of internal APIs 64-bit would be a pretty big hit for\\n> > a 32-bit platform --- more instructions, more memory consumed for\\n> > things like Datum arrays, all in a memory space that's not that big.\\n> \\n> I don't agree as well with the line of arguments to just remove 32b\\n> support.\\n\\nNobody is actually proposing that, as far as I can tell? I mean, by all\\nmeans argue that the overhead is too high, but just claiming that\\nslowing down 32bit systems by a measurable fraction is morally\\nequivalent to removing 32bit support seems pointlessly facetious.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 20:14:47 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 7:47 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> > This line of argument seems to me to be the moral equivalent of\\n> > \\\"let's drop 32-bit support altogether\\\". I'm not entirely on board\\n> > with that. Certainly, a lot of the world is 64-bit these days,\\n> > but people are still building small systems and they might want\\n> > a database; preferably one that hasn't been detuned to the extent\\n> > that it barely manages to run at all on such a platform. Making\\n> > a whole lot of internal APIs 64-bit would be a pretty big hit for\\n> > a 32-bit platform --- more instructions, more memory consumed for\\n> > things like Datum arrays, all in a memory space that's not that big.\\n>\\n> I don't agree as well with the line of arguments to just remove 32b\\n> support.\\n\\nClearly you didn't read what I actually wrote, Michael.\\n\\n--\\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 21:41:58 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On 2019-10-31 14:36, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n>> float4 is now always pass-by-value; the pass-by-reference code path is\\n>> completely removed.\\n> \\n> I think this is OK.\\n\\nOK, here is a patch for just this part, and we can continue the \\ndiscussion on the rest in the meantime.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Sat, 2 Nov 2019 08:39:16 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On 2019-11-01 15:41, Robert Haas wrote:\\n> On a related note, why do we store typbyval in the catalog anyway\\n> instead of inferring it from typlen and maybe typalign? It seems like\\n> a bad idea to record on disk the way we pass around values in memory,\\n> because it means that a change to how values are passed around in\\n> memory has ramifications for on-disk compatibility.\\n\\nThis sounds interesting. It would remove a pg_upgrade hazard (in the \\nlong run).\\n\\nThere is some backward compatibility to be concerned about. This change \\nwould require extension authors to change their code to insert #ifdef \\nUSE_FLOAT8_BYVAL or similar, where currently their code might only \\nsupport one method or the other.\\n\\n> rhaas=# select typname, typlen, typbyval, typalign from pg_type where\\n> typlen in (1,2,4,8) != typbyval;\\n\\nThere are also typlen=6 types. Who knew. ;-)\\n\\n> typname | typlen | typbyval | typalign\\n> ----------+--------+----------+----------\\n> macaddr8 | 8 | f | i\\n> (1 row)\\n\\nThis might be a case of the above issue: It's easier to just make it \\npass by reference always than deal with a bunch of #ifdefs.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Sat, 2 Nov 2019 08:46:12 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> On 2019-11-01 15:41, Robert Haas wrote:\\n>> On a related note, why do we store typbyval in the catalog anyway\\n>> instead of inferring it from typlen and maybe typalign? It seems like\\n>> a bad idea to record on disk the way we pass around values in memory,\\n>> because it means that a change to how values are passed around in\\n>> memory has ramifications for on-disk compatibility.\\n\\n> This sounds interesting. It would remove a pg_upgrade hazard (in the \\n> long run).\\n\\n> There is some backward compatibility to be concerned about.\\n\\nYeah. The point here is that typbyval specifies what the C functions\\nconcerned with the datatype are expecting. We can't just up and say\\n\\\"we're going to decide that for you\\\".\\n\\nI do get the point that supporting two different typbyval options\\nfor float8 and related types is a nontrivial cost.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 02 Nov 2019 11:26:59 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-11-02 08:46:12 +0100, Peter Eisentraut wrote:\\n> On 2019-11-01 15:41, Robert Haas wrote:\\n> > On a related note, why do we store typbyval in the catalog anyway\\n> > instead of inferring it from typlen and maybe typalign? It seems like\\n> > a bad idea to record on disk the way we pass around values in memory,\\n> > because it means that a change to how values are passed around in\\n> > memory has ramifications for on-disk compatibility.\\n>\\n> This sounds interesting. It would remove a pg_upgrade hazard (in the long\\n> run).\\n>\\n> There is some backward compatibility to be concerned about. This change\\n> would require extension authors to change their code to insert #ifdef\\n> USE_FLOAT8_BYVAL or similar, where currently their code might only support\\n> one method or the other.\\n\\nI think we really ought to remove the difference behind macros. That is,\\nfor types that could be either, provide macros that fetch function\\narguments into local memory, independent of whether the argument is a\\nbyval or byref type. I.e. instead of having separate #ifdef\\nUSE_FLOAT8_BYVALs for DatumGetFloat8(), DatumGetInt64(), ... we should\\nprovide that logic in one centralized set of macros.\\n\\nThe fact that USE_FLOAT8_BYVAL has to creep into various functions imo\\nis the reasons why people are unhappy about it.\\n\\nOne way to do this would be something roughly like this sketch:\\n\\n/* allow to force types to be byref, for testing purposes */\\n#if USE_FLOAT8_BYVAL\\n#define DatumForTypeIsByval(type) (sizeof(type) <= SIZEOF_DATUM)\\n#else\\n#define DatumForTypeIsByval(type) (sizeof(type) <= sizeof(int))\\n#endif\\n\\n/* yes, I dream of being C++ once grown up */\\n#define DefineSmallFixedWidthDatumTypeConversions(type, TypeToDatumName, DatumToTypeName) \\\\\\n static inline type \\\\\\n TypeToDatumName (type value) \\\\\\n { \\\\\\n if (DatumForTypeIsByval(type)) \\\\\\n { \\\\\\n Datum tmp; \\\\\\n \\\\\\n /* ensure correct conversion, consider e.g. float */ \\\\\\n memcpy(&tmp, &value, sizeof(type)); \\\\\\n \\\\\\n return tmp; \\\\\\n } \\\\\\n else \\\\\\n { \\\\\\n type *tmp = (type *) palloc(sizeof(datum)); \\\\\\n\\n *tmp = value;\\n\\n return PointerGetDatum(tmp); \\\\\\n } \\\\\\n } \\\\\\n \\\\\\n static inline type \\\\\\n DatumToTypeName (Datum datum) \\\\\\n { \\\\\\n if (DatumForTypeIsByval(type)) \\\\\\n { \\\\\\n type tmp; \\\\\\n \\\\\\n /* ensure correct conversion */ \\\\\\n memcpy(&tmp, &datum, sizeof(type)); \\\\\\n \\\\\\n return tmp; \\\\\\n } \\\\\\n else \\\\\\n return *(type *) DatumGetPointer(type); \\\\\\n } \\\\\\n\\nAnd then have\\n\\nDefineSmallFixedWidthDatumTypeConversions(\\n float8,\\n Float8GetDatum,\\n DatumGetFloat8);\\n\\nDefineSmallFixedWidthDatumTypeConversions(\\n int64,\\n Int64GetDatum,\\n DatumGetInt64);\\n\\nAnd now also\\n\\nDefineSmallFixedWidthDatumTypeConversions(\\n macaddr,\\n Macaddr8GetDatum,\\n DatumGetMacaddr8);\\n\\n(there's probably plenty of bugs in the above, it's just a sketch)\\n\\n\\nWe don't have to break types / extensions with inferring byval for fixed\\nwidth types. Instead we can change CREATE TYPE's PASSEDBYVALUE to accept\\nan optional parameter 'auto', allowing to opt in.\\n\\n\\n> > rhaas=# select typname, typlen, typbyval, typalign from pg_type where\\n> > typlen in (1,2,4,8) != typbyval;\\n>\\n> There are also typlen=6 types. Who knew. ;-)\\n\\nThere's a recent thread about them :)\\n\\n\\n> > typname | typlen | typbyval | typalign\\n> > ----------+--------+----------+----------\\n> > macaddr8 | 8 | f | i\\n> > (1 row)\\n>\\n> This might be a case of the above issue: It's easier to just make it pass by\\n> reference always than deal with a bunch of #ifdefs.\\n\\nIndeed. And that's a bad sign imo.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sat, 2 Nov 2019 17:00:17 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 01, 2019 at 09:41:58PM -0700, Peter Geoghegan wrote:\\n> On Fri, Nov 1, 2019 at 7:47 PM Michael Paquier <michael@paquier.xyz> wrote:\\n>> I don't agree as well with the line of arguments to just remove 32b\\n>> support.\\n> \\n> Clearly you didn't read what I actually wrote, Michael.\\n\\nSorry, that was an misunderstanding from my side.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 7 Nov 2019 15:47:30 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 2, 2019 at 8:00 PM Andres Freund <andres@anarazel.de> wrote:\\n> I think we really ought to remove the difference behind macros. That is,\\n> for types that could be either, provide macros that fetch function\\n> arguments into local memory, independent of whether the argument is a\\n> byval or byref type. I.e. instead of having separate #ifdef\\n> USE_FLOAT8_BYVALs for DatumGetFloat8(), DatumGetInt64(), ... we should\\n> provide that logic in one centralized set of macros.\\n>\\n> The fact that USE_FLOAT8_BYVAL has to creep into various functions imo\\n> is the reasons why people are unhappy about it.\\n\\nI think I'm *more* unhappy about the fact that it affects a bunch of\\nthings that are not about whether float8 is passed byval. I mean, you\\nmention DatumGetInt64() above, but why in the world does a setting\\nwith \\\"float8\\\" in the name affect how we pass int64? The thing is like\\nkudzu, getting into all sorts of things that it has no business\\naffecting - at least if you judge by the name - and for no really\\nclear reason.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Nov 2019 07:36:27 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On 2019-11-02 08:39, Peter Eisentraut wrote:\\n> On 2019-10-31 14:36, Tom Lane wrote:\\n>> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n>>> float4 is now always pass-by-value; the pass-by-reference code path is\\n>>> completely removed.\\n>>\\n>> I think this is OK.\\n> \\n> OK, here is a patch for just this part, and we can continue the\\n> discussion on the rest in the meantime.\\n\\nI have committed this part.\\n\\nI will rebase and continue developing the rest of the patches based on \\nthe discussion so far.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 21 Nov 2019 19:20:28 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 21, 2019 at 07:20:28PM +0100, Peter Eisentraut wrote:\\n> I have committed this part.\\n> \\n> I will rebase and continue developing the rest of the patches based on the\\n> discussion so far.\\n\\nBased on that I am marking the patch as committed in the CF. The rest\\nof the patch set could have a new entry.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 25 Nov 2019 11:31:58 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"My revised proposal is to remove --disable-float8-byval as a configure \\noption but keep it as an option in pg_config_manual.h. It is no longer \\nuseful as a user-facing option, but as was pointed out, it is somewhat \\nuseful for developers, so pg_config_manual.h seems like the right place.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Tue, 26 Nov 2019 21:27:40 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> My revised proposal is to remove --disable-float8-byval as a configure \\n> option but keep it as an option in pg_config_manual.h. It is no longer \\n> useful as a user-facing option, but as was pointed out, it is somewhat \\n> useful for developers, so pg_config_manual.h seems like the right place.\\n\\nOK, works for me.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Nov 2019 15:33:14 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On 2019-11-26 21:33, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n>> My revised proposal is to remove --disable-float8-byval as a configure\\n>> option but keep it as an option in pg_config_manual.h. It is no longer\\n>> useful as a user-facing option, but as was pointed out, it is somewhat\\n>> useful for developers, so pg_config_manual.h seems like the right place.\\n> \\n> OK, works for me.\\n\\ndone\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Wed, 27 Nov 2019 13:30:29 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 1:19 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Fri, Nov 1, 2019 at 3:15 PM Peter Geoghegan <pg@bowt.ie> wrote:\\n> > I don't think that those two things are equivalent at all. There may\\n> > even be workloads that will benefit when run on 32-bit hardware.\\n> > Having to palloc() and pfree() with 8 byte integers is probably very\\n> > slow.\\n>\\n> Yeah! I mean, users who are using only 4-byte or smaller pass-by-value\\n> quantities will be harmed, especially in cases where they are storing\\n> a lot of them at the same time (e.g. sorting) and especially if they\\n> double their space consumption and run out of their very limited\\n> supply of memory.\\n\\nApparently Linux has almost no upstream resources for testing 32-bit\\nx86, and it shows:\\n\\nhttps://lwn.net/ml/oss-security/CALCETrW1z0gCLFJz-1Jwj_wcT3+axXkP_wOCxY8JkbSLzV80GA@mail.gmail.com/\\n\\nI think that this kind of thing argues for minimizing the amount of\\ncode that can only be tested on a small minority of the computers that\\nare in general use today. If no Postgres hacker regularly runs the\\ncode, then its chances of having bugs are far higher. Having coverage\\nin the buildfarm certainly helps, but it's no substitute.\\n\\nSticking with the !USE_FLOAT8_BYVAL example, it's easy to imagine\\nsomebody forgetting to add a !USE_FLOAT8_BYVAL block, that contains\\nthe required pfree(). Now you have a memory leak that only affects a\\nsmall minority of platforms. How likely is it that buildfarm coverage\\nwill help somebody detect that problem?\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Thu, 12 Dec 2019 14:06:03 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On 2019-12-12 23:06, Peter Geoghegan wrote:\\n> Apparently Linux has almost no upstream resources for testing 32-bit\\n> x86, and it shows:\\n\\nBut isn't 32-bit Windows still a thing? Or does that work differently?\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 13 Dec 2019 13:33:21 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 13, 2019 at 6:33 AM Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n> On 2019-12-12 23:06, Peter Geoghegan wrote:\\n> > Apparently Linux has almost no upstream resources for testing 32-bit\\n> > x86, and it shows:\\n>\\n> But isn't 32-bit Windows still a thing? Or does that work differently?\\n\\nWell, again, I think the proposal here is not get rid of 32-bit\\nsupport, but to have less code that only gets regularly tested on\\n32-bit machines. If we made datums 8 bytes everywhere, we would have\\nless such code, and very likely fewer bugs. And as pointed out\\nupthread, although some things might perform worse for the remaining\\nsupply of 32-bit users, other things might perform better. I'm not\\n100% sure that it would work out to a win overall, but I think there's\\na good chance, especially when you factor in the reduced bug surface.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 13 Dec 2019 09:43:39 -0600\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> Well, again, I think the proposal here is not get rid of 32-bit\\n> support, but to have less code that only gets regularly tested on\\n> 32-bit machines.\\n\\nThat seems like generally a good plan. But as to the specific idea...\\n\\n> If we made datums 8 bytes everywhere, we would have\\n> less such code, and very likely fewer bugs.\\n\\n... it's not entirely clear to me that it'd be possible to do this\\nwithout causing a storm of \\\"cast from pointer to integer of different\\nsize\\\" (and vice versa) warnings on 32-bit machines. That would be a\\ndeal-breaker independently of any performance considerations, IMO.\\nSo if anyone wants to pursue this, finding a way around that might be\\nthe first thing to look at.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 13 Dec 2019 11:25:37 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Remove configure --disable-float4-byval and\\n --disable-float8-byval\"\n }\n]","isConversational":false}}},{"rowIdx":49372,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI want to know how postgres stores catalog relations in cache in-depth. Is\nthere any documentation for that?\n\nHi,I want to know how postgres stores catalog relations in cache in-depth. Is there any documentation for that?","msg_date":"Thu, 31 Oct 2019 15:19:23 +0530","msg_from":"Natarajan R <nataraj3098@gmail.com>","msg_from_op":true,"msg_subject":"Postgres cache"},{"msg_contents":"On Thu, Oct 31, 2019 at 03:19:23PM +0530, Natarajan R wrote:\n>Hi,\n>\n>I want to know how postgres stores catalog relations in cache in-depth. Is\n>there any documentation for that?\n\nNot sure what exactly you mean by \"cache\" - whether shared buffers (as a\nshared general database cache) or syscache/catcache, i.e. the special\ncache of catalog records each backend maintains privately.\n\nI'm not aware of exhaustive developer docs explaining these parts, but\nfor shared buffers you might want to look at\n\n src/backend/storage/buffer/README\n\nwhile for catcache/syscache you probably need to look at the code and\ncomments in the related files, particularly in \n\n src/backend/utils/cache/syscache.c\n src/backend/utils/cache/relcache.c\n\nNot sure if there's a better source of information :-(\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Thu, 31 Oct 2019 22:24:54 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Postgres cache"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI want to know how postgres stores catalog relations in cache in-depth. Is\\nthere any documentation for that?\\n\\nHi,I want to know how postgres stores catalog relations in cache in-depth. Is there any documentation for that?\",\n \"msg_date\": \"Thu, 31 Oct 2019 15:19:23 +0530\",\n \"msg_from\": \"Natarajan R <nataraj3098@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Postgres cache\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 03:19:23PM +0530, Natarajan R wrote:\\n>Hi,\\n>\\n>I want to know how postgres stores catalog relations in cache in-depth. Is\\n>there any documentation for that?\\n\\nNot sure what exactly you mean by \\\"cache\\\" - whether shared buffers (as a\\nshared general database cache) or syscache/catcache, i.e. the special\\ncache of catalog records each backend maintains privately.\\n\\nI'm not aware of exhaustive developer docs explaining these parts, but\\nfor shared buffers you might want to look at\\n\\n src/backend/storage/buffer/README\\n\\nwhile for catcache/syscache you probably need to look at the code and\\ncomments in the related files, particularly in \\n\\n src/backend/utils/cache/syscache.c\\n src/backend/utils/cache/relcache.c\\n\\nNot sure if there's a better source of information :-(\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 22:24:54 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Postgres cache\"\n }\n]","isConversational":false}}},{"rowIdx":49373,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nThe command tag of ALTER MATERIALIZED VIEW is basically\n\"ALTER MATERIALIZED VIEW\". For example,\n\n =# ALTER MATERIALIZED VIEW test ALTER COLUMN j SET STATISTICS 100;\n ALTER MATERIALIZED VIEW\n =# ALTER MATERIALIZED VIEW test OWNER TO CURRENT_USER;\n ALTER MATERIALIZED VIEW\n =# ALTER MATERIALIZED VIEW test RENAME TO hoge;\n ALTER MATERIALIZED VIEW\n\nThis is ok and looks intuitive to users. But I found that the command tag of\nALTER MATERIALIZED VIEW RENAME COLUMN is \"ALTER TABLE\", not \"ALTER VIEW\".\n\n =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\n ALTER TABLE\n\nIs this intentional? Or bug?\n\nRegards,\n\n-- \nFujii Masao\n\n\n","msg_date":"Thu, 31 Oct 2019 19:38:35 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":true,"msg_subject":"The command tag of \"ALTER MATERIALIZED VIEW RENAME COLUMN\""},{"msg_contents":"Fujii Masao <masao.fujii@gmail.com> writes:\n> ... I found that the command tag of\n> ALTER MATERIALIZED VIEW RENAME COLUMN is \"ALTER TABLE\", not \"ALTER VIEW\".\n\n> =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\n> ALTER TABLE\n\n> Is this intentional? Or bug?\n\nSeems like an oversight.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 31 Oct 2019 09:56:17 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: The command tag of \"ALTER MATERIALIZED VIEW RENAME COLUMN\""},{"msg_contents":"On Thu, Oct 31, 2019 at 6:56 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> Fujii Masao <masao.fujii@gmail.com> writes:\n> > ... I found that the command tag of\n> > ALTER MATERIALIZED VIEW RENAME COLUMN is \"ALTER TABLE\", not \"ALTER VIEW\".\n>\n> > =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\n> > ALTER TABLE\n>\n> > Is this intentional? Or bug?\n>\n> Seems like an oversight.\n>\n> regards, tom lane\n>\n>\n>\nThe same issue is with ALTER FOREIGN TABLE\n\n# ALTER FOREIGN TABLE ft RENAME COLUMN a to t;\n\nALTER TABLE\n\n\n# ALTER MATERIALIZED VIEW mv RENAME COLUMN a to r;\n\nALTER TABLE\n\n\n\nAttached patch fixes that for ALTER VIEW , ALTER MATERIALIZED VIEW and\nALTER FOREIGN TABLE\n\n\n# ALTER MATERIALIZED VIEW mv RENAME COLUMN a to r;\n\nALTER MATERIALIZED VIEW\n\n\n# ALTER FOREIGN TABLE ft RENAME COLUMN a to t;\n\nALTER FOREIGN TABLE\n\n\n-- \nIbrar Ahmed","msg_date":"Fri, 1 Nov 2019 02:33:53 +0500","msg_from":"Ibrar Ahmed <ibrar.ahmad@gmail.com>","msg_from_op":false,"msg_subject":"Re: The command tag of \"ALTER MATERIALIZED VIEW RENAME COLUMN\""},{"msg_contents":"On Fri, Nov 1, 2019 at 6:34 AM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n>\n>\n>\n> On Thu, Oct 31, 2019 at 6:56 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>\n>> Fujii Masao <masao.fujii@gmail.com> writes:\n>> > ... I found that the command tag of\n>> > ALTER MATERIALIZED VIEW RENAME COLUMN is \"ALTER TABLE\", not \"ALTER VIEW\".\n>>\n>> > =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\n>> > ALTER TABLE\n>>\n>> > Is this intentional? Or bug?\n>>\n>> Seems like an oversight.\n\nThanks for the check!\n\n> The same issue is with ALTER FOREIGN TABLE\n\nYes.\n\n> Attached patch fixes that for ALTER VIEW , ALTER MATERIALIZED VIEW and ALTER FOREIGN TABLE\n\nYou introduced subtype in your patch, but I think it's better and simpler\nto just give relationType to AlterObjectTypeCommandTag()\nif renaming the columns (i.e., renameType = OBJECT_COLUMN).\n\nTo avoid this kind of oversight about command tag, I'd like to add regression\ntests to make sure that SQL returns valid and correct command tag.\nBut currently there seems no mechanism for such test, in regression\ntest. Right??\nMaybe we will need that mechanism.\n\nRegards,\n\n-- \nFujii Masao","msg_date":"Fri, 1 Nov 2019 12:00:31 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":true,"msg_subject":"Re: The command tag of \"ALTER MATERIALIZED VIEW RENAME COLUMN\""},{"msg_contents":"On Fri, Nov 1, 2019 at 8:00 AM Fujii Masao <masao.fujii@gmail.com> wrote:\n\n> On Fri, Nov 1, 2019 at 6:34 AM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n> >\n> >\n> >\n> > On Thu, Oct 31, 2019 at 6:56 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >>\n> >> Fujii Masao <masao.fujii@gmail.com> writes:\n> >> > ... I found that the command tag of\n> >> > ALTER MATERIALIZED VIEW RENAME COLUMN is \"ALTER TABLE\", not \"ALTER\n> VIEW\".\n> >>\n> >> > =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\n> >> > ALTER TABLE\n> >>\n> >> > Is this intentional? Or bug?\n> >>\n> >> Seems like an oversight.\n>\n> Thanks for the check!\n>\n> > The same issue is with ALTER FOREIGN TABLE\n>\n> Yes.\n>\n> > Attached patch fixes that for ALTER VIEW , ALTER MATERIALIZED VIEW and\n> ALTER FOREIGN TABLE\n>\n> You introduced subtype in your patch, but I think it's better and simpler\n> to just give relationType to AlterObjectTypeCommandTag()\n> if renaming the columns (i.e., renameType = OBJECT_COLUMN).\n>\n> That's works perfectly along with future oversight about the command tag.\n\n\n> To avoid this kind of oversight about command tag, I'd like to add\n> regression\n> tests to make sure that SQL returns valid and correct command tag.\n> But currently there seems no mechanism for such test, in regression\n> test. Right??\n>\n\nDo we really need a regression test cases for such small oversights?\n\n\n> Maybe we will need that mechanism.\n>\n> Regards,\n>\n> --\n> Fujii Masao\n>\n\n\n-- \nIbrar Ahmed\n\nOn Fri, Nov 1, 2019 at 8:00 AM Fujii Masao <masao.fujii@gmail.com> wrote:On Fri, Nov 1, 2019 at 6:34 AM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\n>\n>\n>\n> On Thu, Oct 31, 2019 at 6:56 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>\n>> Fujii Masao <masao.fujii@gmail.com> writes:\n>> > ... I found that the command tag of\n>> > ALTER MATERIALIZED VIEW RENAME COLUMN is \"ALTER TABLE\", not \"ALTER VIEW\".\n>>\n>> > =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\n>> > ALTER TABLE\n>>\n>> > Is this intentional? Or bug?\n>>\n>> Seems like an oversight.\n\nThanks for the check!\n\n> The same issue is with ALTER FOREIGN TABLE\n\nYes.\n\n> Attached patch fixes that for ALTER VIEW , ALTER MATERIALIZED VIEW and ALTER FOREIGN TABLE\n\nYou introduced subtype in your patch, but I think it's better and simpler\nto just give relationType to AlterObjectTypeCommandTag()\nif renaming the columns (i.e., renameType = OBJECT_COLUMN).\nThat's works perfectly along with future oversight about the command tag. \nTo avoid this kind of oversight about command tag, I'd like to add regression\ntests to make sure that SQL returns valid and correct command tag.\nBut currently there seems no mechanism for such test, in regression\ntest. Right?? Do we really need a regression test cases for such small oversights? \nMaybe we will need that mechanism.\n\nRegards,\n\n-- \nFujii Masao\n-- Ibrar Ahmed","msg_date":"Fri, 1 Nov 2019 14:17:03 +0500","msg_from":"Ibrar Ahmed <ibrar.ahmad@gmail.com>","msg_from_op":false,"msg_subject":"Re: The command tag of \"ALTER MATERIALIZED VIEW RENAME COLUMN\""},{"msg_contents":"On Fri, Nov 01, 2019 at 02:17:03PM +0500, Ibrar Ahmed wrote:\n> Do we really need a regression test cases for such small oversights?\n\nIt is possible to get the command tags using an event trigger... But\nthat sounds hack-ish.\n--\nMichael","msg_date":"Sat, 2 Nov 2019 16:40:35 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: The command tag of \"ALTER MATERIALIZED VIEW RENAME COLUMN\""},{"msg_contents":"On Sat, Nov 2, 2019 at 4:40 PM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Fri, Nov 01, 2019 at 02:17:03PM +0500, Ibrar Ahmed wrote:\n> > Do we really need a regression test cases for such small oversights?\n>\n> It is possible to get the command tags using an event trigger... But\n> that sounds hack-ish.\n\nYes, so if simple test mechanism to check command tag exists,\nit would be helpful.\n\nI'm thinking to commit the patch. But I have one question; is it ok to\nback-patch? Since the patch changes the command tags for some commands,\nfor example, which might break the existing event trigger functions\nusing TG_TAG if we back-patch it. Or we should guarantee the compatibility of\ncommand tag within the same major version?\n\nRegards,\n\n-- \nFujii Masao\n\n\n","msg_date":"Tue, 5 Nov 2019 14:39:51 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":true,"msg_subject":"Re: The command tag of \"ALTER MATERIALIZED VIEW RENAME COLUMN\""},{"msg_contents":"Fujii Masao <masao.fujii@gmail.com> writes:\n> I'm thinking to commit the patch. But I have one question; is it ok to\n> back-patch? Since the patch changes the command tags for some commands,\n> for example, which might break the existing event trigger functions\n> using TG_TAG if we back-patch it. Or we should guarantee the compatibility of\n> command tag within the same major version?\n\nI would not back-patch this. I don't think it's enough of a bug\nto justify taking any compatibility risks for.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 05 Nov 2019 09:19:20 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: The command tag of \"ALTER MATERIALIZED VIEW RENAME COLUMN\""},{"msg_contents":"On Tue, Nov 5, 2019 at 11:19 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Fujii Masao <masao.fujii@gmail.com> writes:\n> > I'm thinking to commit the patch. But I have one question; is it ok to\n> > back-patch? Since the patch changes the command tags for some commands,\n> > for example, which might break the existing event trigger functions\n> > using TG_TAG if we back-patch it. Or we should guarantee the compatibility of\n> > command tag within the same major version?\n>\n> I would not back-patch this. I don't think it's enough of a bug\n> to justify taking any compatibility risks for.\n\n+1\nI committed the patch only to the master. Thanks!\n\nRegards,\n\n-- \nFujii Masao\n\n\n","msg_date":"Wed, 6 Nov 2019 12:56:44 +0900","msg_from":"Fujii Masao <masao.fujii@gmail.com>","msg_from_op":true,"msg_subject":"Re: The command tag of \"ALTER MATERIALIZED VIEW RENAME COLUMN\""}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nThe command tag of ALTER MATERIALIZED VIEW is basically\\n\\\"ALTER MATERIALIZED VIEW\\\". For example,\\n\\n =# ALTER MATERIALIZED VIEW test ALTER COLUMN j SET STATISTICS 100;\\n ALTER MATERIALIZED VIEW\\n =# ALTER MATERIALIZED VIEW test OWNER TO CURRENT_USER;\\n ALTER MATERIALIZED VIEW\\n =# ALTER MATERIALIZED VIEW test RENAME TO hoge;\\n ALTER MATERIALIZED VIEW\\n\\nThis is ok and looks intuitive to users. But I found that the command tag of\\nALTER MATERIALIZED VIEW RENAME COLUMN is \\\"ALTER TABLE\\\", not \\\"ALTER VIEW\\\".\\n\\n =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\\n ALTER TABLE\\n\\nIs this intentional? Or bug?\\n\\nRegards,\\n\\n-- \\nFujii Masao\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 19:38:35 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"The command tag of \\\"ALTER MATERIALIZED VIEW RENAME COLUMN\\\"\"\n },\n {\n \"msg_contents\": \"Fujii Masao <masao.fujii@gmail.com> writes:\\n> ... I found that the command tag of\\n> ALTER MATERIALIZED VIEW RENAME COLUMN is \\\"ALTER TABLE\\\", not \\\"ALTER VIEW\\\".\\n\\n> =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\\n> ALTER TABLE\\n\\n> Is this intentional? Or bug?\\n\\nSeems like an oversight.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 09:56:17 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: The command tag of \\\"ALTER MATERIALIZED VIEW RENAME COLUMN\\\"\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 6:56 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> Fujii Masao <masao.fujii@gmail.com> writes:\\n> > ... I found that the command tag of\\n> > ALTER MATERIALIZED VIEW RENAME COLUMN is \\\"ALTER TABLE\\\", not \\\"ALTER VIEW\\\".\\n>\\n> > =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\\n> > ALTER TABLE\\n>\\n> > Is this intentional? Or bug?\\n>\\n> Seems like an oversight.\\n>\\n> regards, tom lane\\n>\\n>\\n>\\nThe same issue is with ALTER FOREIGN TABLE\\n\\n# ALTER FOREIGN TABLE ft RENAME COLUMN a to t;\\n\\nALTER TABLE\\n\\n\\n# ALTER MATERIALIZED VIEW mv RENAME COLUMN a to r;\\n\\nALTER TABLE\\n\\n\\n\\nAttached patch fixes that for ALTER VIEW , ALTER MATERIALIZED VIEW and\\nALTER FOREIGN TABLE\\n\\n\\n# ALTER MATERIALIZED VIEW mv RENAME COLUMN a to r;\\n\\nALTER MATERIALIZED VIEW\\n\\n\\n# ALTER FOREIGN TABLE ft RENAME COLUMN a to t;\\n\\nALTER FOREIGN TABLE\\n\\n\\n-- \\nIbrar Ahmed\",\n \"msg_date\": \"Fri, 1 Nov 2019 02:33:53 +0500\",\n \"msg_from\": \"Ibrar Ahmed <ibrar.ahmad@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: The command tag of \\\"ALTER MATERIALIZED VIEW RENAME COLUMN\\\"\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 6:34 AM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n>\\n>\\n>\\n> On Thu, Oct 31, 2019 at 6:56 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>\\n>> Fujii Masao <masao.fujii@gmail.com> writes:\\n>> > ... I found that the command tag of\\n>> > ALTER MATERIALIZED VIEW RENAME COLUMN is \\\"ALTER TABLE\\\", not \\\"ALTER VIEW\\\".\\n>>\\n>> > =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\\n>> > ALTER TABLE\\n>>\\n>> > Is this intentional? Or bug?\\n>>\\n>> Seems like an oversight.\\n\\nThanks for the check!\\n\\n> The same issue is with ALTER FOREIGN TABLE\\n\\nYes.\\n\\n> Attached patch fixes that for ALTER VIEW , ALTER MATERIALIZED VIEW and ALTER FOREIGN TABLE\\n\\nYou introduced subtype in your patch, but I think it's better and simpler\\nto just give relationType to AlterObjectTypeCommandTag()\\nif renaming the columns (i.e., renameType = OBJECT_COLUMN).\\n\\nTo avoid this kind of oversight about command tag, I'd like to add regression\\ntests to make sure that SQL returns valid and correct command tag.\\nBut currently there seems no mechanism for such test, in regression\\ntest. Right??\\nMaybe we will need that mechanism.\\n\\nRegards,\\n\\n-- \\nFujii Masao\",\n \"msg_date\": \"Fri, 1 Nov 2019 12:00:31 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: The command tag of \\\"ALTER MATERIALIZED VIEW RENAME COLUMN\\\"\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 8:00 AM Fujii Masao <masao.fujii@gmail.com> wrote:\\n\\n> On Fri, Nov 1, 2019 at 6:34 AM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n> >\\n> >\\n> >\\n> > On Thu, Oct 31, 2019 at 6:56 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >>\\n> >> Fujii Masao <masao.fujii@gmail.com> writes:\\n> >> > ... I found that the command tag of\\n> >> > ALTER MATERIALIZED VIEW RENAME COLUMN is \\\"ALTER TABLE\\\", not \\\"ALTER\\n> VIEW\\\".\\n> >>\\n> >> > =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\\n> >> > ALTER TABLE\\n> >>\\n> >> > Is this intentional? Or bug?\\n> >>\\n> >> Seems like an oversight.\\n>\\n> Thanks for the check!\\n>\\n> > The same issue is with ALTER FOREIGN TABLE\\n>\\n> Yes.\\n>\\n> > Attached patch fixes that for ALTER VIEW , ALTER MATERIALIZED VIEW and\\n> ALTER FOREIGN TABLE\\n>\\n> You introduced subtype in your patch, but I think it's better and simpler\\n> to just give relationType to AlterObjectTypeCommandTag()\\n> if renaming the columns (i.e., renameType = OBJECT_COLUMN).\\n>\\n> That's works perfectly along with future oversight about the command tag.\\n\\n\\n> To avoid this kind of oversight about command tag, I'd like to add\\n> regression\\n> tests to make sure that SQL returns valid and correct command tag.\\n> But currently there seems no mechanism for such test, in regression\\n> test. Right??\\n>\\n\\nDo we really need a regression test cases for such small oversights?\\n\\n\\n> Maybe we will need that mechanism.\\n>\\n> Regards,\\n>\\n> --\\n> Fujii Masao\\n>\\n\\n\\n-- \\nIbrar Ahmed\\n\\nOn Fri, Nov 1, 2019 at 8:00 AM Fujii Masao <masao.fujii@gmail.com> wrote:On Fri, Nov 1, 2019 at 6:34 AM Ibrar Ahmed <ibrar.ahmad@gmail.com> wrote:\\n>\\n>\\n>\\n> On Thu, Oct 31, 2019 at 6:56 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>\\n>> Fujii Masao <masao.fujii@gmail.com> writes:\\n>> > ... I found that the command tag of\\n>> > ALTER MATERIALIZED VIEW RENAME COLUMN is \\\"ALTER TABLE\\\", not \\\"ALTER VIEW\\\".\\n>>\\n>> > =# ALTER MATERIALIZED VIEW hoge RENAME COLUMN j TO x;\\n>> > ALTER TABLE\\n>>\\n>> > Is this intentional? Or bug?\\n>>\\n>> Seems like an oversight.\\n\\nThanks for the check!\\n\\n> The same issue is with ALTER FOREIGN TABLE\\n\\nYes.\\n\\n> Attached patch fixes that for ALTER VIEW , ALTER MATERIALIZED VIEW and ALTER FOREIGN TABLE\\n\\nYou introduced subtype in your patch, but I think it's better and simpler\\nto just give relationType to AlterObjectTypeCommandTag()\\nif renaming the columns (i.e., renameType = OBJECT_COLUMN).\\nThat's works perfectly along with future oversight about the command tag. \\nTo avoid this kind of oversight about command tag, I'd like to add regression\\ntests to make sure that SQL returns valid and correct command tag.\\nBut currently there seems no mechanism for such test, in regression\\ntest. Right?? Do we really need a regression test cases for such small oversights? \\nMaybe we will need that mechanism.\\n\\nRegards,\\n\\n-- \\nFujii Masao\\n-- Ibrar Ahmed\",\n \"msg_date\": \"Fri, 1 Nov 2019 14:17:03 +0500\",\n \"msg_from\": \"Ibrar Ahmed <ibrar.ahmad@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: The command tag of \\\"ALTER MATERIALIZED VIEW RENAME COLUMN\\\"\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 01, 2019 at 02:17:03PM +0500, Ibrar Ahmed wrote:\\n> Do we really need a regression test cases for such small oversights?\\n\\nIt is possible to get the command tags using an event trigger... But\\nthat sounds hack-ish.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 2 Nov 2019 16:40:35 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: The command tag of \\\"ALTER MATERIALIZED VIEW RENAME COLUMN\\\"\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 2, 2019 at 4:40 PM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Fri, Nov 01, 2019 at 02:17:03PM +0500, Ibrar Ahmed wrote:\\n> > Do we really need a regression test cases for such small oversights?\\n>\\n> It is possible to get the command tags using an event trigger... But\\n> that sounds hack-ish.\\n\\nYes, so if simple test mechanism to check command tag exists,\\nit would be helpful.\\n\\nI'm thinking to commit the patch. But I have one question; is it ok to\\nback-patch? Since the patch changes the command tags for some commands,\\nfor example, which might break the existing event trigger functions\\nusing TG_TAG if we back-patch it. Or we should guarantee the compatibility of\\ncommand tag within the same major version?\\n\\nRegards,\\n\\n-- \\nFujii Masao\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Nov 2019 14:39:51 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: The command tag of \\\"ALTER MATERIALIZED VIEW RENAME COLUMN\\\"\"\n },\n {\n \"msg_contents\": \"Fujii Masao <masao.fujii@gmail.com> writes:\\n> I'm thinking to commit the patch. But I have one question; is it ok to\\n> back-patch? Since the patch changes the command tags for some commands,\\n> for example, which might break the existing event trigger functions\\n> using TG_TAG if we back-patch it. Or we should guarantee the compatibility of\\n> command tag within the same major version?\\n\\nI would not back-patch this. I don't think it's enough of a bug\\nto justify taking any compatibility risks for.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 05 Nov 2019 09:19:20 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: The command tag of \\\"ALTER MATERIALIZED VIEW RENAME COLUMN\\\"\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 5, 2019 at 11:19 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Fujii Masao <masao.fujii@gmail.com> writes:\\n> > I'm thinking to commit the patch. But I have one question; is it ok to\\n> > back-patch? Since the patch changes the command tags for some commands,\\n> > for example, which might break the existing event trigger functions\\n> > using TG_TAG if we back-patch it. Or we should guarantee the compatibility of\\n> > command tag within the same major version?\\n>\\n> I would not back-patch this. I don't think it's enough of a bug\\n> to justify taking any compatibility risks for.\\n\\n+1\\nI committed the patch only to the master. Thanks!\\n\\nRegards,\\n\\n-- \\nFujii Masao\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Nov 2019 12:56:44 +0900\",\n \"msg_from\": \"Fujii Masao <masao.fujii@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: The command tag of \\\"ALTER MATERIALIZED VIEW RENAME COLUMN\\\"\"\n }\n]","isConversational":false}}},{"rowIdx":49374,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Dear hackers,\n\nAs declared last month, I propose again the new ECPG grammar, DECLARE STATEMENT.\nThis had been committed once, but it removed from PG12 because of\nsome problems. \nIn this mail, I want to report some problems that previous implementation has,\nproduce a new solution, and attach a WIP patch.\n\n[Basic function, Grammar, and Use case]\nThis statement will be used for the purpose of designating a connection easily.\nPlease see below:\nhttps://www.postgresql.org/message-id/flat/4E72940DA2BF16479384A86D54D0988A4D80D3C9@G01JPEXMBKW04\nThe Oracle's manual will also help your understanding:\nhttps://docs.oracle.com/en/database/oracle/oracle-database/19/lnpcc/embedded-SQL-statements-and-directives.html#GUID-0A30B7B4-BD91-42EA-AACE-2E9CBF7E9C1A\n\n[Issues]\nThat's why this feature has been reverted.\n1. The namespace of the identifier was not clear. If you use a same identifier for other SQL statements,\n these interfered each other and statements might be executed at the unexpected connection.\n2. Declaring at the outside of functions was not allowed. This specification is quite different from the other \n declarative statements, so some users might be confused.\n For instance, the following example was rejected.\n```\nEXEC SQL DECLARE stmt STATEMENT;\n\nint\nmain()\n{\n...\n\tEXEC SQL DECLARE cur CURSOR FOR stmt;\n...\n}\n```\n3. These specifications were not compatible with other DBMSs.\n\n[Solutions]\nThe namespace is set to be a file unit. This follows other DBMSs.\nWhen the DECLARE SATATEMENT statement is read, the name, identifier\nand the related connection are recorded.\nAnd if you use the declared identifier in order to prepare or declare cursor,\nthe fourth argument for ECPGdo(it represents the connection) will be overwritten.\nThis declaration is enabled only the precompile phase.\n\n [Limitations]\nThe declaration must be appeared before using it.\nThis also follows Pro*C precompiler.\n\nA WIP patch is attached. Confirm that all ECPG tests have passed,\nhowever, some documents are not included.\nThey will be added later.\nI applied the pgindent as a test, but it might be failed because this is the\nfirst time for me.\n\nBest regards\n\nHayato Kuroda\nFUJITSU LIMITED\nE-Mail:kuroda.hayato@fujitsu.com","msg_date":"Thu, 31 Oct 2019 12:29:30 +0000","msg_from":"\"kuroda.hayato@fujitsu.com\" <kuroda.hayato@fujitsu.com>","msg_from_op":true,"msg_subject":"ECPG: proposal for new DECLARE STATEMENT"},{"msg_contents":"Hi,\n\nOn Thu, Oct 31, 2019 at 12:29:30PM +0000, kuroda.hayato@fujitsu.com wrote:\n>Dear hackers,\n>\n>As declared last month, I propose again the new ECPG grammar, DECLARE STATEMENT.\n>This had been committed once, but it removed from PG12 because of\n>some problems.\n>In this mail, I want to report some problems that previous implementation has,\n>produce a new solution, and attach a WIP patch.\n>\n>[Basic function, Grammar, and Use case]\n>This statement will be used for the purpose of designating a connection easily.\n>Please see below:\n>https://www.postgresql.org/message-id/flat/4E72940DA2BF16479384A86D54D0988A4D80D3C9@G01JPEXMBKW04\n>The Oracle's manual will also help your understanding:\n>https://docs.oracle.com/en/database/oracle/oracle-database/19/lnpcc/embedded-SQL-statements-and-directives.html#GUID-0A30B7B4-BD91-42EA-AACE-2E9CBF7E9C1A\n>\n>[Issues]\n>That's why this feature has been reverted.\n>1. The namespace of the identifier was not clear. If you use a same identifier for other SQL statements,\n> these interfered each other and statements might be executed at the unexpected connection.\n>2. Declaring at the outside of functions was not allowed. This specification is quite different from the other\n> declarative statements, so some users might be confused.\n> For instance, the following example was rejected.\n>```\n>EXEC SQL DECLARE stmt STATEMENT;\n>\n>int\n>main()\n>{\n>...\n>\tEXEC SQL DECLARE cur CURSOR FOR stmt;\n>...\n>}\n>```\n>3. These specifications were not compatible with other DBMSs.\n>\n>[Solutions]\n>The namespace is set to be a file unit. This follows other DBMSs.\n>When the DECLARE SATATEMENT statement is read, the name, identifier\n>and the related connection are recorded.\n>And if you use the declared identifier in order to prepare or declare cursor,\n>the fourth argument for ECPGdo(it represents the connection) will be overwritten.\n>This declaration is enabled only the precompile phase.\n>\n> [Limitations]\n>The declaration must be appeared before using it.\n>This also follows Pro*C precompiler.\n>\n>A WIP patch is attached. Confirm that all ECPG tests have passed,\n>however, some documents are not included.\n>They will be added later.\n>I applied the pgindent as a test, but it might be failed because this is the\n>first time for me.\n>\n\nI see there were no reviews of this new patch, with the feature\nreimplemented after it was reverted from PG12 in September :-(\n\nI'm not an ecpg expert (in fact I've never even used it), so my review\nis pretty superficial, but I only found a couple of minor whitespace\nissues (adding/removing a line/tab) - see the attached file.\n\nKuroda-san, you mentioned the patch is WIP. What other bits you think\nare missing / need improvement? I see you mentioned some documentation\nis missing - I suppose that's one of the missing pieces?\n\n\nFor the record, there were two threads discussing the implementation [1]\nand then the revert [2].\n\n[1] https://www.postgresql.org/message-id/flat/1F66B161998C704BABF8989B8A2AC0A313AA41%40G01JPEXMBYT05\n[2] https://www.postgresql.org/message-id/flat/TY2PR01MB2443EC8286995378AEB7D9F8F5B10%40TY2PR01MB2443.jpnprd01.prod.outlook.com\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Sun, 12 Jan 2020 03:52:48 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ECPG: proposal for new DECLARE STATEMENT"},{"msg_contents":"On Sun, Jan 12, 2020 at 03:52:48AM +0100, Tomas Vondra wrote:\n> ...\n>\n>I'm not an ecpg expert (in fact I've never even used it), so my review\n>is pretty superficial, but I only found a couple of minor whitespace\n>issues (adding/removing a line/tab) - see the attached file.\n>\n\nMeh, forgot to attach the file ...\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Sun, 12 Jan 2020 04:41:03 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ECPG: proposal for new DECLARE STATEMENT"},{"msg_contents":"On 1/11/20 10:41 PM, Tomas Vondra wrote:\n> On Sun, Jan 12, 2020 at 03:52:48AM +0100, Tomas Vondra wrote:\n>> ...\n>>\n>> I'm not an ecpg expert (in fact I've never even used it), so my review\n>> is pretty superficial, but I only found a couple of minor whitespace\n>> issues (adding/removing a line/tab) - see the attached file.\n>>\n> \n> Meh, forgot to attach the file ...\n\nAny thoughts on Tomas' comments?\n\nA big part of moving a patch forward is keeping the thread active and \nanswering comments/review.\n\nRegards,\n-- \n-David\ndavid@pgmasters.net\n\n\n","msg_date":"Mon, 30 Mar 2020 12:53:29 -0400","msg_from":"David Steele <david@pgmasters.net>","msg_from_op":false,"msg_subject":"Re: ECPG: proposal for new DECLARE STATEMENT"},{"msg_contents":"> On 30 Mar 2020, at 18:53, David Steele <david@pgmasters.net> wrote:\n> \n> On 1/11/20 10:41 PM, Tomas Vondra wrote:\n>> On Sun, Jan 12, 2020 at 03:52:48AM +0100, Tomas Vondra wrote:\n>>> ...\n>>> \n>>> I'm not an ecpg expert (in fact I've never even used it), so my review\n>>> is pretty superficial, but I only found a couple of minor whitespace\n>>> issues (adding/removing a line/tab) - see the attached file.\n>>> \n>> Meh, forgot to attach the file ...\n> \n> Any thoughts on Tomas' comments?\n> \n> A big part of moving a patch forward is keeping the thread active and answering comments/review.\n\nThis patch has now been silent for quite a while, unless someone is interested\nenough to bring it forward it seems about time to close it.\n\ncheers ./daniel\n\n","msg_date":"Tue, 15 Sep 2020 12:07:29 +0200","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: ECPG: proposal for new DECLARE STATEMENT"},{"msg_contents":"> This patch has now been silent for quite a while, unless someone is\n> interested\n> enough to bring it forward it seems about time to close it.\n\nI am interested but still short on time. I will definitely look into it\nas soon as I find some spare minutes.\n\nMichael\n-- \nMichael Meskes\nMichael at Fam-Meskes dot De\nMichael at Meskes dot (De|Com|Net|Org)\nMeskes at (Debian|Postgresql) dot Org\n\n\n\n","msg_date":"Tue, 15 Sep 2020 12:31:57 +0200","msg_from":"Michael Meskes <meskes@postgresql.org>","msg_from_op":false,"msg_subject":"Re: ECPG: proposal for new DECLARE STATEMENT"},{"msg_contents":"Dear Tomas, Daniel, Michael, \r\n\r\nI missed your e-mails, and I apologize the very late reply. \r\nI want you to thank keeping the thread.\r\n\r\n> I'm not an ecpg expert (in fact I've never even used it), so my review\r\n> is pretty superficial, but I only found a couple of minor whitespace\r\n> issues (adding/removing a line/tab) - see the attached file.\r\n\r\nThanks, I fixed it.\r\n\r\n> Kuroda-san, you mentioned the patch is WIP. What other bits you think\r\n> are missing / need improvement? I see you mentioned some documentation\r\n> is missing - I suppose that's one of the missing pieces?\r\n\r\nAll functionalities I expect has been already implemented in the previous patch, \r\nand I thought that only doc and reviews were needed.\r\n\r\nFinally I attach new patch. This patch contains source changes, a test code,\r\nand documentation changes. This one is not WIP.\r\n\r\nI will try to review other topics on the next Commitfest.\r\n\r\nBest regards,\r\nHayato Kuroda\r\nFUJITSU LIMITED\r\n\r\n-----Original Message-----\r\nFrom: Michael Meskes <meskes@postgresql.org> \r\nSent: Tuesday, September 15, 2020 7:32 PM\r\nTo: pgsql-hackers@lists.postgresql.org\r\nSubject: Re: ECPG: proposal for new DECLARE STATEMENT\r\n\r\n> This patch has now been silent for quite a while, unless someone is\r\n> interested\r\n> enough to bring it forward it seems about time to close it.\r\n\r\nI am interested but still short on time. I will definitely look into it\r\nas soon as I find some spare minutes.\r\n\r\nMichael\r\n-- \r\nMichael Meskes\r\nMichael at Fam-Meskes dot De\r\nMichael at Meskes dot (De|Com|Net|Org)\r\nMeskes at (Debian|Postgresql) dot Org","msg_date":"Fri, 23 Oct 2020 06:25:25 +0000","msg_from":"\"kuroda.hayato@fujitsu.com\" <kuroda.hayato@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: ECPG: proposal for new DECLARE STATEMENT"},{"msg_contents":"The following review has been posted through the commitfest application:\nmake installcheck-world: tested, passed\nImplements feature: tested, passed\nSpec compliant: tested, passed\nDocumentation: tested, passed\n\nRecently I have been doing some work on ecpg. So I review this patch. No problem was found.\n\nThe new status of this patch is: Ready for Committer\n","msg_date":"Mon, 16 Nov 2020 12:52:40 +0000","msg_from":"Shawn Wang <shawn.wang.pg@gmail.com>","msg_from_op":false,"msg_subject":"Re: ECPG: proposal for new DECLARE STATEMENT"},{"msg_contents":"Dear Hackers,\r\n\r\nI know I'm asking a big favor, but could you review(and commit) the patch?\r\nThe status has become RFC last Nov., but no one checked this after that.\r\nMaybe Meskes is quite busy and have no time to review it.\r\n\r\nThe main part of the patch is about 200 lines(It means not so long), and maybe\r\nI have reviewed other patches more than it.\r\n\r\nI will review more, so I'm happy if this commits until the end of next CF.\r\n\r\nBest Regards,\r\nHayato Kuroda\r\nFUJITSU LIMITED\r\n","msg_date":"Wed, 27 Jan 2021 09:18:28 +0000","msg_from":"\"kuroda.hayato@fujitsu.com\" <kuroda.hayato@fujitsu.com>","msg_from_op":true,"msg_subject":"RE: ECPG: proposal for new DECLARE STATEMENT"}],"string":"[\n {\n \"msg_contents\": \"Dear hackers,\\n\\nAs declared last month, I propose again the new ECPG grammar, DECLARE STATEMENT.\\nThis had been committed once, but it removed from PG12 because of\\nsome problems. \\nIn this mail, I want to report some problems that previous implementation has,\\nproduce a new solution, and attach a WIP patch.\\n\\n[Basic function, Grammar, and Use case]\\nThis statement will be used for the purpose of designating a connection easily.\\nPlease see below:\\nhttps://www.postgresql.org/message-id/flat/4E72940DA2BF16479384A86D54D0988A4D80D3C9@G01JPEXMBKW04\\nThe Oracle's manual will also help your understanding:\\nhttps://docs.oracle.com/en/database/oracle/oracle-database/19/lnpcc/embedded-SQL-statements-and-directives.html#GUID-0A30B7B4-BD91-42EA-AACE-2E9CBF7E9C1A\\n\\n[Issues]\\nThat's why this feature has been reverted.\\n1. The namespace of the identifier was not clear. If you use a same identifier for other SQL statements,\\n these interfered each other and statements might be executed at the unexpected connection.\\n2. Declaring at the outside of functions was not allowed. This specification is quite different from the other \\n declarative statements, so some users might be confused.\\n For instance, the following example was rejected.\\n```\\nEXEC SQL DECLARE stmt STATEMENT;\\n\\nint\\nmain()\\n{\\n...\\n\\tEXEC SQL DECLARE cur CURSOR FOR stmt;\\n...\\n}\\n```\\n3. These specifications were not compatible with other DBMSs.\\n\\n[Solutions]\\nThe namespace is set to be a file unit. This follows other DBMSs.\\nWhen the DECLARE SATATEMENT statement is read, the name, identifier\\nand the related connection are recorded.\\nAnd if you use the declared identifier in order to prepare or declare cursor,\\nthe fourth argument for ECPGdo(it represents the connection) will be overwritten.\\nThis declaration is enabled only the precompile phase.\\n\\n [Limitations]\\nThe declaration must be appeared before using it.\\nThis also follows Pro*C precompiler.\\n\\nA WIP patch is attached. Confirm that all ECPG tests have passed,\\nhowever, some documents are not included.\\nThey will be added later.\\nI applied the pgindent as a test, but it might be failed because this is the\\nfirst time for me.\\n\\nBest regards\\n\\nHayato Kuroda\\nFUJITSU LIMITED\\nE-Mail:kuroda.hayato@fujitsu.com\",\n \"msg_date\": \"Thu, 31 Oct 2019 12:29:30 +0000\",\n \"msg_from\": \"\\\"kuroda.hayato@fujitsu.com\\\" <kuroda.hayato@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"ECPG: proposal for new DECLARE STATEMENT\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Thu, Oct 31, 2019 at 12:29:30PM +0000, kuroda.hayato@fujitsu.com wrote:\\n>Dear hackers,\\n>\\n>As declared last month, I propose again the new ECPG grammar, DECLARE STATEMENT.\\n>This had been committed once, but it removed from PG12 because of\\n>some problems.\\n>In this mail, I want to report some problems that previous implementation has,\\n>produce a new solution, and attach a WIP patch.\\n>\\n>[Basic function, Grammar, and Use case]\\n>This statement will be used for the purpose of designating a connection easily.\\n>Please see below:\\n>https://www.postgresql.org/message-id/flat/4E72940DA2BF16479384A86D54D0988A4D80D3C9@G01JPEXMBKW04\\n>The Oracle's manual will also help your understanding:\\n>https://docs.oracle.com/en/database/oracle/oracle-database/19/lnpcc/embedded-SQL-statements-and-directives.html#GUID-0A30B7B4-BD91-42EA-AACE-2E9CBF7E9C1A\\n>\\n>[Issues]\\n>That's why this feature has been reverted.\\n>1. The namespace of the identifier was not clear. If you use a same identifier for other SQL statements,\\n> these interfered each other and statements might be executed at the unexpected connection.\\n>2. Declaring at the outside of functions was not allowed. This specification is quite different from the other\\n> declarative statements, so some users might be confused.\\n> For instance, the following example was rejected.\\n>```\\n>EXEC SQL DECLARE stmt STATEMENT;\\n>\\n>int\\n>main()\\n>{\\n>...\\n>\\tEXEC SQL DECLARE cur CURSOR FOR stmt;\\n>...\\n>}\\n>```\\n>3. These specifications were not compatible with other DBMSs.\\n>\\n>[Solutions]\\n>The namespace is set to be a file unit. This follows other DBMSs.\\n>When the DECLARE SATATEMENT statement is read, the name, identifier\\n>and the related connection are recorded.\\n>And if you use the declared identifier in order to prepare or declare cursor,\\n>the fourth argument for ECPGdo(it represents the connection) will be overwritten.\\n>This declaration is enabled only the precompile phase.\\n>\\n> [Limitations]\\n>The declaration must be appeared before using it.\\n>This also follows Pro*C precompiler.\\n>\\n>A WIP patch is attached. Confirm that all ECPG tests have passed,\\n>however, some documents are not included.\\n>They will be added later.\\n>I applied the pgindent as a test, but it might be failed because this is the\\n>first time for me.\\n>\\n\\nI see there were no reviews of this new patch, with the feature\\nreimplemented after it was reverted from PG12 in September :-(\\n\\nI'm not an ecpg expert (in fact I've never even used it), so my review\\nis pretty superficial, but I only found a couple of minor whitespace\\nissues (adding/removing a line/tab) - see the attached file.\\n\\nKuroda-san, you mentioned the patch is WIP. What other bits you think\\nare missing / need improvement? I see you mentioned some documentation\\nis missing - I suppose that's one of the missing pieces?\\n\\n\\nFor the record, there were two threads discussing the implementation [1]\\nand then the revert [2].\\n\\n[1] https://www.postgresql.org/message-id/flat/1F66B161998C704BABF8989B8A2AC0A313AA41%40G01JPEXMBYT05\\n[2] https://www.postgresql.org/message-id/flat/TY2PR01MB2443EC8286995378AEB7D9F8F5B10%40TY2PR01MB2443.jpnprd01.prod.outlook.com\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Sun, 12 Jan 2020 03:52:48 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ECPG: proposal for new DECLARE STATEMENT\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 12, 2020 at 03:52:48AM +0100, Tomas Vondra wrote:\\n> ...\\n>\\n>I'm not an ecpg expert (in fact I've never even used it), so my review\\n>is pretty superficial, but I only found a couple of minor whitespace\\n>issues (adding/removing a line/tab) - see the attached file.\\n>\\n\\nMeh, forgot to attach the file ...\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Sun, 12 Jan 2020 04:41:03 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ECPG: proposal for new DECLARE STATEMENT\"\n },\n {\n \"msg_contents\": \"On 1/11/20 10:41 PM, Tomas Vondra wrote:\\n> On Sun, Jan 12, 2020 at 03:52:48AM +0100, Tomas Vondra wrote:\\n>> ...\\n>>\\n>> I'm not an ecpg expert (in fact I've never even used it), so my review\\n>> is pretty superficial, but I only found a couple of minor whitespace\\n>> issues (adding/removing a line/tab) - see the attached file.\\n>>\\n> \\n> Meh, forgot to attach the file ...\\n\\nAny thoughts on Tomas' comments?\\n\\nA big part of moving a patch forward is keeping the thread active and \\nanswering comments/review.\\n\\nRegards,\\n-- \\n-David\\ndavid@pgmasters.net\\n\\n\\n\",\n \"msg_date\": \"Mon, 30 Mar 2020 12:53:29 -0400\",\n \"msg_from\": \"David Steele <david@pgmasters.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ECPG: proposal for new DECLARE STATEMENT\"\n },\n {\n \"msg_contents\": \"> On 30 Mar 2020, at 18:53, David Steele <david@pgmasters.net> wrote:\\n> \\n> On 1/11/20 10:41 PM, Tomas Vondra wrote:\\n>> On Sun, Jan 12, 2020 at 03:52:48AM +0100, Tomas Vondra wrote:\\n>>> ...\\n>>> \\n>>> I'm not an ecpg expert (in fact I've never even used it), so my review\\n>>> is pretty superficial, but I only found a couple of minor whitespace\\n>>> issues (adding/removing a line/tab) - see the attached file.\\n>>> \\n>> Meh, forgot to attach the file ...\\n> \\n> Any thoughts on Tomas' comments?\\n> \\n> A big part of moving a patch forward is keeping the thread active and answering comments/review.\\n\\nThis patch has now been silent for quite a while, unless someone is interested\\nenough to bring it forward it seems about time to close it.\\n\\ncheers ./daniel\\n\\n\",\n \"msg_date\": \"Tue, 15 Sep 2020 12:07:29 +0200\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ECPG: proposal for new DECLARE STATEMENT\"\n },\n {\n \"msg_contents\": \"> This patch has now been silent for quite a while, unless someone is\\n> interested\\n> enough to bring it forward it seems about time to close it.\\n\\nI am interested but still short on time. I will definitely look into it\\nas soon as I find some spare minutes.\\n\\nMichael\\n-- \\nMichael Meskes\\nMichael at Fam-Meskes dot De\\nMichael at Meskes dot (De|Com|Net|Org)\\nMeskes at (Debian|Postgresql) dot Org\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 15 Sep 2020 12:31:57 +0200\",\n \"msg_from\": \"Michael Meskes <meskes@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ECPG: proposal for new DECLARE STATEMENT\"\n },\n {\n \"msg_contents\": \"Dear Tomas, Daniel, Michael, \\r\\n\\r\\nI missed your e-mails, and I apologize the very late reply. \\r\\nI want you to thank keeping the thread.\\r\\n\\r\\n> I'm not an ecpg expert (in fact I've never even used it), so my review\\r\\n> is pretty superficial, but I only found a couple of minor whitespace\\r\\n> issues (adding/removing a line/tab) - see the attached file.\\r\\n\\r\\nThanks, I fixed it.\\r\\n\\r\\n> Kuroda-san, you mentioned the patch is WIP. What other bits you think\\r\\n> are missing / need improvement? I see you mentioned some documentation\\r\\n> is missing - I suppose that's one of the missing pieces?\\r\\n\\r\\nAll functionalities I expect has been already implemented in the previous patch, \\r\\nand I thought that only doc and reviews were needed.\\r\\n\\r\\nFinally I attach new patch. This patch contains source changes, a test code,\\r\\nand documentation changes. This one is not WIP.\\r\\n\\r\\nI will try to review other topics on the next Commitfest.\\r\\n\\r\\nBest regards,\\r\\nHayato Kuroda\\r\\nFUJITSU LIMITED\\r\\n\\r\\n-----Original Message-----\\r\\nFrom: Michael Meskes <meskes@postgresql.org> \\r\\nSent: Tuesday, September 15, 2020 7:32 PM\\r\\nTo: pgsql-hackers@lists.postgresql.org\\r\\nSubject: Re: ECPG: proposal for new DECLARE STATEMENT\\r\\n\\r\\n> This patch has now been silent for quite a while, unless someone is\\r\\n> interested\\r\\n> enough to bring it forward it seems about time to close it.\\r\\n\\r\\nI am interested but still short on time. I will definitely look into it\\r\\nas soon as I find some spare minutes.\\r\\n\\r\\nMichael\\r\\n-- \\r\\nMichael Meskes\\r\\nMichael at Fam-Meskes dot De\\r\\nMichael at Meskes dot (De|Com|Net|Org)\\r\\nMeskes at (Debian|Postgresql) dot Org\",\n \"msg_date\": \"Fri, 23 Oct 2020 06:25:25 +0000\",\n \"msg_from\": \"\\\"kuroda.hayato@fujitsu.com\\\" <kuroda.hayato@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: ECPG: proposal for new DECLARE STATEMENT\"\n },\n {\n \"msg_contents\": \"The following review has been posted through the commitfest application:\\nmake installcheck-world: tested, passed\\nImplements feature: tested, passed\\nSpec compliant: tested, passed\\nDocumentation: tested, passed\\n\\nRecently I have been doing some work on ecpg. So I review this patch. No problem was found.\\n\\nThe new status of this patch is: Ready for Committer\\n\",\n \"msg_date\": \"Mon, 16 Nov 2020 12:52:40 +0000\",\n \"msg_from\": \"Shawn Wang <shawn.wang.pg@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ECPG: proposal for new DECLARE STATEMENT\"\n },\n {\n \"msg_contents\": \"Dear Hackers,\\r\\n\\r\\nI know I'm asking a big favor, but could you review(and commit) the patch?\\r\\nThe status has become RFC last Nov., but no one checked this after that.\\r\\nMaybe Meskes is quite busy and have no time to review it.\\r\\n\\r\\nThe main part of the patch is about 200 lines(It means not so long), and maybe\\r\\nI have reviewed other patches more than it.\\r\\n\\r\\nI will review more, so I'm happy if this commits until the end of next CF.\\r\\n\\r\\nBest Regards,\\r\\nHayato Kuroda\\r\\nFUJITSU LIMITED\\r\\n\",\n \"msg_date\": \"Wed, 27 Jan 2021 09:18:28 +0000\",\n \"msg_from\": \"\\\"kuroda.hayato@fujitsu.com\\\" <kuroda.hayato@fujitsu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"RE: ECPG: proposal for new DECLARE STATEMENT\"\n }\n]","isConversational":false}}},{"rowIdx":49375,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nThe attached patch adds an -a / --appname command line switch to \npg_basebackup, pg_receivewal and pg_recvlogical.\n\nThis is useful when f.ex. pg_receivewal needs to connect as a \nsynchronous client (synchronous_standby_names),\n\n pg_receivewal -h myhost -p 5432 -S replica1 -a replica1 --synchronous \n-D /wal\n\nI'll add the patch to the CommitFest for discussion, as there is overlap \nwith the -d switch.\n\nBest regards,\n Jesper","msg_date":"Thu, 31 Oct 2019 08:52:58 -0400","msg_from":"Jesper Pedersen <jesper.pedersen@redhat.com>","msg_from_op":true,"msg_subject":"Application name for pg_basebackup and friends"},{"msg_contents":"On 31/10/2019 14:52, Jesper Pedersen wrote:\n> Hi,\n> \n> The attached patch adds an -a / --appname command line switch to\n> pg_basebackup, pg_receivewal and pg_recvlogical.\n> \n> This is useful when f.ex. pg_receivewal needs to connect as a\n> synchronous client (synchronous_standby_names),\n> \n> pg_receivewal -h myhost -p 5432 -S replica1 -a replica1 --synchronous\n> -D /wal\n> \n> I'll add the patch to the CommitFest for discussion, as there is overlap\n> with the -d switch.\n\nYou can already set application name with the environment variable or on \nthe database connections string:\n\npg_receivewal -D /wal -d \"host=myhost application_name=myreceiver\"\n\nI don't think we need a new comand line switch for it.\n\n- Heikki\n\n\n","msg_date":"Thu, 31 Oct 2019 15:10:35 +0200","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":false,"msg_subject":"Re: Application name for pg_basebackup and friends"},{"msg_contents":"On Thu, Oct 31, 2019 at 03:10:35PM +0200, Heikki Linnakangas wrote:\n> You can already set application name with the environment variable or on the\n> database connections string:\n> \n> pg_receivewal -D /wal -d \"host=myhost application_name=myreceiver\"\n> \n> I don't think we need a new comand line switch for it.\n\n+1.\n--\nMichael","msg_date":"Sat, 2 Nov 2019 16:45:40 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Application name for pg_basebackup and friends"},{"msg_contents":"On Sat, Nov 02, 2019 at 04:45:40PM +0900, Michael Paquier wrote:\n> On Thu, Oct 31, 2019 at 03:10:35PM +0200, Heikki Linnakangas wrote:\n>> You can already set application name with the environment variable or on the\n>> database connections string:\n>> \n>> pg_receivewal -D /wal -d \"host=myhost application_name=myreceiver\"\n>> \n>> I don't think we need a new comand line switch for it.\n> \n> +1.\n\nPlease note that I have marked this patch as rejected in the CF app,\nper the arguments upthread.\n--\nMichael","msg_date":"Thu, 7 Nov 2019 15:51:59 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Application name for pg_basebackup and friends"},{"msg_contents":"On 11/7/19 1:51 AM, Michael Paquier wrote:\n>>> I don't think we need a new comand line switch for it.\n>>\n>> +1.\n> \n> Please note that I have marked this patch as rejected in the CF app,\n> per the arguments upthread.\n\nOk, agreed.\n\nThanks for the feedback !\n\nBest regards,\n Jesper\n\n\n\n","msg_date":"Thu, 7 Nov 2019 12:36:42 -0500","msg_from":"Jesper Pedersen <jesper.pedersen@redhat.com>","msg_from_op":true,"msg_subject":"Re: Application name for pg_basebackup and friends"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nThe attached patch adds an -a / --appname command line switch to \\npg_basebackup, pg_receivewal and pg_recvlogical.\\n\\nThis is useful when f.ex. pg_receivewal needs to connect as a \\nsynchronous client (synchronous_standby_names),\\n\\n pg_receivewal -h myhost -p 5432 -S replica1 -a replica1 --synchronous \\n-D /wal\\n\\nI'll add the patch to the CommitFest for discussion, as there is overlap \\nwith the -d switch.\\n\\nBest regards,\\n Jesper\",\n \"msg_date\": \"Thu, 31 Oct 2019 08:52:58 -0400\",\n \"msg_from\": \"Jesper Pedersen <jesper.pedersen@redhat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Application name for pg_basebackup and friends\"\n },\n {\n \"msg_contents\": \"On 31/10/2019 14:52, Jesper Pedersen wrote:\\n> Hi,\\n> \\n> The attached patch adds an -a / --appname command line switch to\\n> pg_basebackup, pg_receivewal and pg_recvlogical.\\n> \\n> This is useful when f.ex. pg_receivewal needs to connect as a\\n> synchronous client (synchronous_standby_names),\\n> \\n> pg_receivewal -h myhost -p 5432 -S replica1 -a replica1 --synchronous\\n> -D /wal\\n> \\n> I'll add the patch to the CommitFest for discussion, as there is overlap\\n> with the -d switch.\\n\\nYou can already set application name with the environment variable or on \\nthe database connections string:\\n\\npg_receivewal -D /wal -d \\\"host=myhost application_name=myreceiver\\\"\\n\\nI don't think we need a new comand line switch for it.\\n\\n- Heikki\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 15:10:35 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Application name for pg_basebackup and friends\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 03:10:35PM +0200, Heikki Linnakangas wrote:\\n> You can already set application name with the environment variable or on the\\n> database connections string:\\n> \\n> pg_receivewal -D /wal -d \\\"host=myhost application_name=myreceiver\\\"\\n> \\n> I don't think we need a new comand line switch for it.\\n\\n+1.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 2 Nov 2019 16:45:40 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Application name for pg_basebackup and friends\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 02, 2019 at 04:45:40PM +0900, Michael Paquier wrote:\\n> On Thu, Oct 31, 2019 at 03:10:35PM +0200, Heikki Linnakangas wrote:\\n>> You can already set application name with the environment variable or on the\\n>> database connections string:\\n>> \\n>> pg_receivewal -D /wal -d \\\"host=myhost application_name=myreceiver\\\"\\n>> \\n>> I don't think we need a new comand line switch for it.\\n> \\n> +1.\\n\\nPlease note that I have marked this patch as rejected in the CF app,\\nper the arguments upthread.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 7 Nov 2019 15:51:59 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Application name for pg_basebackup and friends\"\n },\n {\n \"msg_contents\": \"On 11/7/19 1:51 AM, Michael Paquier wrote:\\n>>> I don't think we need a new comand line switch for it.\\n>>\\n>> +1.\\n> \\n> Please note that I have marked this patch as rejected in the CF app,\\n> per the arguments upthread.\\n\\nOk, agreed.\\n\\nThanks for the feedback !\\n\\nBest regards,\\n Jesper\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Nov 2019 12:36:42 -0500\",\n \"msg_from\": \"Jesper Pedersen <jesper.pedersen@redhat.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Application name for pg_basebackup and friends\"\n }\n]","isConversational":false}}},{"rowIdx":49376,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi\n\nI almost finished patch optimizing non volatile function calls.\n\nselect f(t.n) from t where f(t.n) > 10 and f(t.n) < 100; needs 3 calls of\nf() for each tuple,\nafter applying patch only 1.\n\nAny pros and cons ?\n\nHiI almost finished patch optimizing non volatile function calls.select f(t.n) from t where f(t.n) > 10 and f(t.n) < 100; needs 3 calls of f() for each tuple,after applying patch only 1.Any pros and cons ?","msg_date":"Thu, 31 Oct 2019 15:06:13 +0100","msg_from":"Andrzej Barszcz <abusinf@gmail.com>","msg_from_op":true,"msg_subject":"function calls optimization"},{"msg_contents":"Hi, \n\nOn October 31, 2019 7:06:13 AM PDT, Andrzej Barszcz <abusinf@gmail.com> wrote:\n>Hi\n>\n>I almost finished patch optimizing non volatile function calls.\n>\n>select f(t.n) from t where f(t.n) > 10 and f(t.n) < 100; needs 3 calls\n>of\n>f() for each tuple,\n>after applying patch only 1.\n>\n>Any pros and cons ?\n\nDepends on the actual way of implementing this proposal. Think we need more details than what you idea here.\n\nAndres\n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.\n\n\n","msg_date":"Thu, 31 Oct 2019 07:25:06 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On October 31, 2019 7:06:13 AM PDT, Andrzej Barszcz <abusinf@gmail.com> wrote:\n>> I almost finished patch optimizing non volatile function calls.\n>> \n>> select f(t.n) from t where f(t.n) > 10 and f(t.n) < 100; needs 3 calls\n>> of\n>> f() for each tuple,\n>> after applying patch only 1.\n>> \n>> Any pros and cons ?\n\n> Depends on the actual way of implementing this proposal. Think we need more details than what you idea here.\n\nWe've typically supposed that the cost of searching for duplicate\nsubexpressions would outweigh the benefits of sometimes finding them.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 31 Oct 2019 10:45:26 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"Hi, \n\nOn October 31, 2019 7:45:26 AM PDT, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>Andres Freund <andres@anarazel.de> writes:\n>> On October 31, 2019 7:06:13 AM PDT, Andrzej Barszcz\n><abusinf@gmail.com> wrote:\n>>> I almost finished patch optimizing non volatile function calls.\n>>> \n>>> select f(t.n) from t where f(t.n) > 10 and f(t.n) < 100; needs 3\n>calls\n>>> of\n>>> f() for each tuple,\n>>> after applying patch only 1.\n>>> \n>>> Any pros and cons ?\n>\n>> Depends on the actual way of implementing this proposal. Think we\n>need more details than what you idea here.\n>\n>We've typically supposed that the cost of searching for duplicate\n>subexpressions would outweigh the benefits of sometimes finding them.\n\nBased on profiles I've seen I'm not sure that's the right choice. Both for when the calls are expensive (say postgis stuff), and for when a lot of rows are processed.\n\nI think one part of doing this in a realistic manner is an efficient search for redundant expressions. The other, also non trivial, is how to even represent references to the results of expressions in other parts of the expression tree / other expressions.\n\nAndres\n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.\n\n\n","msg_date":"Thu, 31 Oct 2019 07:53:20 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"Hi\n\nOn October 31, 2019 7:53:20 AM PDT, Andres Freund <andres@anarazel.de> wrote:\n>On October 31, 2019 7:45:26 AM PDT, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>We've typically supposed that the cost of searching for duplicate\n>>subexpressions would outweigh the benefits of sometimes finding them.\n>\n>Based on profiles I've seen I'm not sure that's the right choice. Both\n>for when the calls are expensive (say postgis stuff), and for when a\n>lot of rows are processed.\n>\n>I think one part of doing this in a realistic manner is an efficient\n>search for redundant expressions. \n\nOne way to improve the situation - which is applicable in a lot of situations, e.g. setrefs.c etc - would be to compute hashes for (sub-) expression trees. Which makes it a lot easier to bail out early when trees are not the same, and also easier to build an efficient way to find redundant expressions. There's some complexity in invalidating such hashes once computed, and when to first compute them, obviously.\n\nAndres\n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.\n\n\n","msg_date":"Thu, 31 Oct 2019 08:02:48 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"On 10/31/19 3:45 PM, Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n>> On October 31, 2019 7:06:13 AM PDT, Andrzej Barszcz <abusinf@gmail.com> wrote:\n>>> Any pros and cons ?\n> \n>> Depends on the actual way of implementing this proposal. Think we need more details than what you idea here.\n> \n> We've typically supposed that the cost of searching for duplicate\n> subexpressions would outweigh the benefits of sometimes finding them.\n\nThat is an important concern, but given how SQL does not make it \nconvenient to re-use partial results of calculations I think such \nqueries are quite common in real world workloads.\n\nSo if we can make it cheap enough I think that it is a worthwhile \noptimization.\n\nAndreas\n\n\n","msg_date":"Thu, 31 Oct 2019 16:05:28 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On October 31, 2019 7:45:26 AM PDT, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> We've typically supposed that the cost of searching for duplicate\n>> subexpressions would outweigh the benefits of sometimes finding them.\n\n> Based on profiles I've seen I'm not sure that's the right choice. Both for when the calls are expensive (say postgis stuff), and for when a lot of rows are processed.\n\nYeah, if your mental model of a function call is some remarkably expensive\nPostGIS geometry manipulation, it's easy to justify doing a lot of work\nto try to detect duplicates. But most functions in most queries are\nmore like int4pl or btint8cmp, and it's going to be extremely remarkable\nif you can make back the planner costs of checking for duplicate usages\nof those.\n\nPossibly this could be finessed by only trying to find duplicates of\nfunctions that have high cost estimates. Not sure how high is high\nenough.\n\n> I think one part of doing this in a realistic manner is an efficient\n> search for redundant expressions. The other, also non trivial, is how to\n> even represent re eferences to the results of expressions in other parts of the expression tree / other expressions.\n\nYup, both of those would be critical to do right.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 31 Oct 2019 11:06:50 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"Hi, \n\nOn October 31, 2019 8:06:50 AM PDT, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>Andres Freund <andres@anarazel.de> writes:\n>> On October 31, 2019 7:45:26 AM PDT, Tom Lane <tgl@sss.pgh.pa.us>\n>wrote:\n>>> We've typically supposed that the cost of searching for duplicate\n>>> subexpressions would outweigh the benefits of sometimes finding\n>them.\n>\n>> Based on profiles I've seen I'm not sure that's the right choice.\n>Both for when the calls are expensive (say postgis stuff), and for when\n>a lot of rows are processed.\n>\n>Yeah, if your mental model of a function call is some remarkably\n>expensive\n>PostGIS geometry manipulation, it's easy to justify doing a lot of work\n>to try to detect duplicates. But most functions in most queries are\n>more like int4pl or btint8cmp, and it's going to be extremely\n>remarkable\n>if you can make back the planner costs of checking for duplicate usages\n>of those.\n\nWell, if it's an expression containing those individuals cheap calls on a seqscan on a large table below an aggregate, it'd likely still be a win. But we don't, to my knowledge, really have a good way to model optimizations like this that should only be done if either expensive or have a high loop count.\n\nI guess one ugly way to deal with this would be to eliminate redundancies very late, e.g. during setrefs (where a better data structure for matching expressions would be good anyway), as we already know all the costs. \n\nBut ideally we would want to do be able to take such savings into account earlier, when considering different paths. I suspect that it might be a good enough vehicle to tackle the rest of the work however, at least initially.\n\nWe could also \"just\" do such elimination during expression \"compilation\", but it'd be better to not have to do something as complicated as this for every execution of a prepared statement.\n\n\n>> I think one part of doing this in a realistic manner is an efficient\n>> search for redundant expressions. The other, also non trivial, is how\n>to\n>> even represent re eferences to the results of expressions in other\n>parts of the expression tree / other expressions.\n>\n>Yup, both of those would be critical to do right.\n\nPotentially related note: for nodes like seqscan, combining the qual and projection processing into one expression seems to be a noticable win (at least when taking care do emit two different sets of deform expression steps). Wonder if something like that would take care of avoiding the need for cross expression value passing in enough places.\n\nAndres\n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.\n\n\n","msg_date":"Thu, 31 Oct 2019 08:20:39 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> Potentially related note: for nodes like seqscan, combining the qual and projection processing into one expression seems to be a noticable win (at least when taking care do emit two different sets of deform expression steps).\n\nThere's just one problem: that violates SQL semantics, and not in\na small way.\n\n\tselect 1/x from tab where x <> 0\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 31 Oct 2019 11:45:26 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"Hi, \n\nOn October 31, 2019 8:45:26 AM PDT, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>Andres Freund <andres@anarazel.de> writes:\n>> Potentially related note: for nodes like seqscan, combining the qual\n>and projection processing into one expression seems to be a noticable\n>win (at least when taking care do emit two different sets of deform\n>expression steps).\n>\n>There's just one problem: that violates SQL semantics, and not in\n>a small way.\n>\n>\tselect 1/x from tab where x <> 0\n\nThe expression would obviously have to return early, before projecting, when not matching the qual? I'm basically just thinking of first executing the steps for the qual, and in the success case execute the projection steps before returning the quals positive result. \n\nAndres\n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.\n\n\n","msg_date":"Thu, 31 Oct 2019 08:49:46 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"x <> 0 is evaluated first, 1/x only when x <> 0, not ?\n\nczw., 31 paź 2019 o 16:45 Tom Lane <tgl@sss.pgh.pa.us> napisał(a):\n\n> Andres Freund <andres@anarazel.de> writes:\n> > Potentially related note: for nodes like seqscan, combining the qual and\n> projection processing into one expression seems to be a noticable win (at\n> least when taking care do emit two different sets of deform expression\n> steps).\n>\n> There's just one problem: that violates SQL semantics, and not in\n> a small way.\n>\n> select 1/x from tab where x <> 0\n>\n> regards, tom lane\n>\n\nx <> 0 is evaluated first, 1/x only when x <> 0, not ?czw., 31 paź 2019 o 16:45 Tom Lane <tgl@sss.pgh.pa.us> napisał(a):Andres Freund <andres@anarazel.de> writes:\n> Potentially related note: for nodes like seqscan, combining the qual and projection processing into one expression seems to be a noticable win (at least when taking care do emit two different sets of deform expression steps).\n\nThere's just one problem: that violates SQL semantics, and not in\na small way.\n\n select 1/x from tab where x <> 0\n\n regards, tom lane","msg_date":"Thu, 31 Oct 2019 16:51:11 +0100","msg_from":"Andrzej Barszcz <abusinf@gmail.com>","msg_from_op":true,"msg_subject":"Re: function calls optimization"},{"msg_contents":"Hi, \n\nOn October 31, 2019 8:51:11 AM PDT, Andrzej Barszcz <abusinf@gmail.com> wrote:\n>x <> 0 is evaluated first, 1/x only when x <> 0, not ?\n>\n>czw., 31 paź 2019 o 16:45 Tom Lane <tgl@sss.pgh.pa.us> napisał(a):\n>\n>> Andres Freund <andres@anarazel.de> writes:\n>> > Potentially related note: for nodes like seqscan, combining the\n>qual and\n>> projection processing into one expression seems to be a noticable win\n>(at\n>> least when taking care do emit two different sets of deform\n>expression\n>> steps).\n>>\n>> There's just one problem: that violates SQL semantics, and not in\n>> a small way.\n>>\n>> select 1/x from tab where x <> 0\n\nOn postgres lists the policy is to reply below the quoted bit, and to trim the quote appropriately.\n\nAndres\n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.\n\n\n","msg_date":"Thu, 31 Oct 2019 08:52:53 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"I recently implemented something closely related to this. Combining and\nmigrating expensive STABLE user-defined functions to the FROM clause, where\nthe function is evaluated as a lateral join (or \"cross apply\"). I'm\ndefining expensive as 50x times more expensive than the default function\ncost.\n\nFor functions that return multiple outputs and where the query uses (...).*\nnotation, this will, for example, consolidate all of the calls in the *\nexpansion into a single call. It also looks in the WHERE clause and HAVING\nclause, and combines those references, too. Currently it requires the\nfunction to be in a top-level AND condition, if it appears in a predicate.\n\nI think I can get permission for contributing it back. If there's an\ninterest in it, let me know.\n\n\n\n-----\nJim Finnerty, AWS, Amazon Aurora PostgreSQL\n--\nSent from: https://www.postgresql-archive.org/PostgreSQL-hackers-f1928748.html\n\n\n","msg_date":"Thu, 31 Oct 2019 08:56:22 -0700 (MST)","msg_from":"Jim Finnerty <jfinnert@amazon.com>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"Hi\n\nI need advice.\nResetExprContext(econtext) is defined as\nMemoryContextReset((econtext)->ecxt_per_tuple_memory).\nI can register callback in MemoryContext but it is always cleaned on every\ncall to MemoryContextReset().\nHow to reset some fields of ExprContext ( living in per_query_memory ) when\nResetExprContext is called ?\n\nczw., 31 paź 2019 o 16:52 Andres Freund <andres@anarazel.de> napisał(a):\n\n> Hi,\n>\n> On October 31, 2019 8:51:11 AM PDT, Andrzej Barszcz <abusinf@gmail.com>\n> wrote:\n> >x <> 0 is evaluated first, 1/x only when x <> 0, not ?\n> >\n> >czw., 31 paź 2019 o 16:45 Tom Lane <tgl@sss.pgh.pa.us> napisał(a):\n> >\n> >> Andres Freund <andres@anarazel.de> writes:\n> >> > Potentially related note: for nodes like seqscan, combining the\n> >qual and\n> >> projection processing into one expression seems to be a noticable win\n> >(at\n> >> least when taking care do emit two different sets of deform\n> >expression\n> >> steps).\n> >>\n> >> There's just one problem: that violates SQL semantics, and not in\n> >> a small way.\n> >>\n> >> select 1/x from tab where x <> 0\n>\n> On postgres lists the policy is to reply below the quoted bit, and to trim\n> the quote appropriately.\n>\n> Andres\n> --\n> Sent from my Android device with K-9 Mail. Please excuse my brevity.\n>\n\nHiI need advice. ResetExprContext(econtext) is defined as MemoryContextReset((econtext)->ecxt_per_tuple_memory).I can register callback in MemoryContext but it is always cleaned on every call to MemoryContextReset().How to reset some fields of ExprContext ( living in per_query_memory ) when ResetExprContext is called ?czw., 31 paź 2019 o 16:52 Andres Freund <andres@anarazel.de> napisał(a):Hi, \n\nOn October 31, 2019 8:51:11 AM PDT, Andrzej Barszcz <abusinf@gmail.com> wrote:\n>x <> 0 is evaluated first, 1/x only when x <> 0, not ?\n>\n>czw., 31 paź 2019 o 16:45 Tom Lane <tgl@sss.pgh.pa.us> napisał(a):\n>\n>> Andres Freund <andres@anarazel.de> writes:\n>> > Potentially related note: for nodes like seqscan, combining the\n>qual and\n>> projection processing into one expression seems to be a noticable win\n>(at\n>> least when taking care do emit two different sets of deform\n>expression\n>> steps).\n>>\n>> There's just one problem: that violates SQL semantics, and not in\n>> a small way.\n>>\n>> select 1/x from tab where x <> 0\n\nOn postgres lists the policy is to reply below the quoted bit, and to trim the quote appropriately.\n\nAndres\n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.","msg_date":"Mon, 18 Nov 2019 15:20:48 +0100","msg_from":"Andrzej Barszcz <abusinf@gmail.com>","msg_from_op":true,"msg_subject":"Re: function calls optimization"},{"msg_contents":"On Thu, Oct 31, 2019 at 11:07 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n>\n>\n> Possibly this could be finessed by only trying to find duplicates of\n> functions that have high cost estimates. Not sure how high is high\n> enough.\n\n\ncan we just add a flag on pg_proc to show if the cost is high or not, if\nuser are not happy with that, they can change it by updating the value?\nbased on that most of the function call cost are low, this way may be\nhelpful for the searching of duplicate expressions.\n\nOn Thu, Oct 31, 2019 at 11:07 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\nPossibly this could be finessed by only trying to find duplicates of\nfunctions that have high cost estimates. Not sure how high is high\nenough. can we just add a flag on pg_proc to show if the cost is high or not, if user are not happy with that, they can change it by updating the value? based on that most of the function call cost are low, this way may be helpful for the searching of duplicate expressions.","msg_date":"Thu, 21 Nov 2019 09:05:14 +0800","msg_from":"Andy Fan <zhihui.fan1213@gmail.com>","msg_from_op":false,"msg_subject":"Re: function calls optimization"},{"msg_contents":"I think your first thought was good.\nHow high ? I think it's a matter of convention, certainly more than default\n100.\n\n\n\nczw., 21 lis 2019 o 02:05 Andy Fan <zhihui.fan1213@gmail.com> napisał(a):\n\n>\n>\n> On Thu, Oct 31, 2019 at 11:07 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n>>\n>>\n>> Possibly this could be finessed by only trying to find duplicates of\n>> functions that have high cost estimates. Not sure how high is high\n>> enough.\n>\n>\n> can we just add a flag on pg_proc to show if the cost is high or not, if\n> user are not happy with that, they can change it by updating the value?\n> based on that most of the function call cost are low, this way may be\n> helpful for the searching of duplicate expressions.\n>\n\nI think your first thought was good. How high ? I think it's a matter of convention, certainly more than default 100. czw., 21 lis 2019 o 02:05 Andy Fan <zhihui.fan1213@gmail.com> napisał(a):On Thu, Oct 31, 2019 at 11:07 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\nPossibly this could be finessed by only trying to find duplicates of\nfunctions that have high cost estimates. Not sure how high is high\nenough. can we just add a flag on pg_proc to show if the cost is high or not, if user are not happy with that, they can change it by updating the value? based on that most of the function call cost are low, this way may be helpful for the searching of duplicate expressions.","msg_date":"Thu, 21 Nov 2019 09:37:46 +0100","msg_from":"Andrzej Barszcz <abusinf@gmail.com>","msg_from_op":true,"msg_subject":"Re: function calls optimization"}],"string":"[\n {\n \"msg_contents\": \"Hi\\n\\nI almost finished patch optimizing non volatile function calls.\\n\\nselect f(t.n) from t where f(t.n) > 10 and f(t.n) < 100; needs 3 calls of\\nf() for each tuple,\\nafter applying patch only 1.\\n\\nAny pros and cons ?\\n\\nHiI almost finished patch optimizing non volatile function calls.select f(t.n) from t where f(t.n) > 10 and f(t.n) < 100; needs 3 calls of f() for each tuple,after applying patch only 1.Any pros and cons ?\",\n \"msg_date\": \"Thu, 31 Oct 2019 15:06:13 +0100\",\n \"msg_from\": \"Andrzej Barszcz <abusinf@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"function calls optimization\"\n },\n {\n \"msg_contents\": \"Hi, \\n\\nOn October 31, 2019 7:06:13 AM PDT, Andrzej Barszcz <abusinf@gmail.com> wrote:\\n>Hi\\n>\\n>I almost finished patch optimizing non volatile function calls.\\n>\\n>select f(t.n) from t where f(t.n) > 10 and f(t.n) < 100; needs 3 calls\\n>of\\n>f() for each tuple,\\n>after applying patch only 1.\\n>\\n>Any pros and cons ?\\n\\nDepends on the actual way of implementing this proposal. Think we need more details than what you idea here.\\n\\nAndres\\n-- \\nSent from my Android device with K-9 Mail. Please excuse my brevity.\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 07:25:06 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On October 31, 2019 7:06:13 AM PDT, Andrzej Barszcz <abusinf@gmail.com> wrote:\\n>> I almost finished patch optimizing non volatile function calls.\\n>> \\n>> select f(t.n) from t where f(t.n) > 10 and f(t.n) < 100; needs 3 calls\\n>> of\\n>> f() for each tuple,\\n>> after applying patch only 1.\\n>> \\n>> Any pros and cons ?\\n\\n> Depends on the actual way of implementing this proposal. Think we need more details than what you idea here.\\n\\nWe've typically supposed that the cost of searching for duplicate\\nsubexpressions would outweigh the benefits of sometimes finding them.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 10:45:26 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"Hi, \\n\\nOn October 31, 2019 7:45:26 AM PDT, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>Andres Freund <andres@anarazel.de> writes:\\n>> On October 31, 2019 7:06:13 AM PDT, Andrzej Barszcz\\n><abusinf@gmail.com> wrote:\\n>>> I almost finished patch optimizing non volatile function calls.\\n>>> \\n>>> select f(t.n) from t where f(t.n) > 10 and f(t.n) < 100; needs 3\\n>calls\\n>>> of\\n>>> f() for each tuple,\\n>>> after applying patch only 1.\\n>>> \\n>>> Any pros and cons ?\\n>\\n>> Depends on the actual way of implementing this proposal. Think we\\n>need more details than what you idea here.\\n>\\n>We've typically supposed that the cost of searching for duplicate\\n>subexpressions would outweigh the benefits of sometimes finding them.\\n\\nBased on profiles I've seen I'm not sure that's the right choice. Both for when the calls are expensive (say postgis stuff), and for when a lot of rows are processed.\\n\\nI think one part of doing this in a realistic manner is an efficient search for redundant expressions. The other, also non trivial, is how to even represent references to the results of expressions in other parts of the expression tree / other expressions.\\n\\nAndres\\n-- \\nSent from my Android device with K-9 Mail. Please excuse my brevity.\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 07:53:20 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"Hi\\n\\nOn October 31, 2019 7:53:20 AM PDT, Andres Freund <andres@anarazel.de> wrote:\\n>On October 31, 2019 7:45:26 AM PDT, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>We've typically supposed that the cost of searching for duplicate\\n>>subexpressions would outweigh the benefits of sometimes finding them.\\n>\\n>Based on profiles I've seen I'm not sure that's the right choice. Both\\n>for when the calls are expensive (say postgis stuff), and for when a\\n>lot of rows are processed.\\n>\\n>I think one part of doing this in a realistic manner is an efficient\\n>search for redundant expressions. \\n\\nOne way to improve the situation - which is applicable in a lot of situations, e.g. setrefs.c etc - would be to compute hashes for (sub-) expression trees. Which makes it a lot easier to bail out early when trees are not the same, and also easier to build an efficient way to find redundant expressions. There's some complexity in invalidating such hashes once computed, and when to first compute them, obviously.\\n\\nAndres\\n-- \\nSent from my Android device with K-9 Mail. Please excuse my brevity.\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 08:02:48 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"On 10/31/19 3:45 PM, Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n>> On October 31, 2019 7:06:13 AM PDT, Andrzej Barszcz <abusinf@gmail.com> wrote:\\n>>> Any pros and cons ?\\n> \\n>> Depends on the actual way of implementing this proposal. Think we need more details than what you idea here.\\n> \\n> We've typically supposed that the cost of searching for duplicate\\n> subexpressions would outweigh the benefits of sometimes finding them.\\n\\nThat is an important concern, but given how SQL does not make it \\nconvenient to re-use partial results of calculations I think such \\nqueries are quite common in real world workloads.\\n\\nSo if we can make it cheap enough I think that it is a worthwhile \\noptimization.\\n\\nAndreas\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 16:05:28 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On October 31, 2019 7:45:26 AM PDT, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> We've typically supposed that the cost of searching for duplicate\\n>> subexpressions would outweigh the benefits of sometimes finding them.\\n\\n> Based on profiles I've seen I'm not sure that's the right choice. Both for when the calls are expensive (say postgis stuff), and for when a lot of rows are processed.\\n\\nYeah, if your mental model of a function call is some remarkably expensive\\nPostGIS geometry manipulation, it's easy to justify doing a lot of work\\nto try to detect duplicates. But most functions in most queries are\\nmore like int4pl or btint8cmp, and it's going to be extremely remarkable\\nif you can make back the planner costs of checking for duplicate usages\\nof those.\\n\\nPossibly this could be finessed by only trying to find duplicates of\\nfunctions that have high cost estimates. Not sure how high is high\\nenough.\\n\\n> I think one part of doing this in a realistic manner is an efficient\\n> search for redundant expressions. The other, also non trivial, is how to\\n> even represent re eferences to the results of expressions in other parts of the expression tree / other expressions.\\n\\nYup, both of those would be critical to do right.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 11:06:50 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"Hi, \\n\\nOn October 31, 2019 8:06:50 AM PDT, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>Andres Freund <andres@anarazel.de> writes:\\n>> On October 31, 2019 7:45:26 AM PDT, Tom Lane <tgl@sss.pgh.pa.us>\\n>wrote:\\n>>> We've typically supposed that the cost of searching for duplicate\\n>>> subexpressions would outweigh the benefits of sometimes finding\\n>them.\\n>\\n>> Based on profiles I've seen I'm not sure that's the right choice.\\n>Both for when the calls are expensive (say postgis stuff), and for when\\n>a lot of rows are processed.\\n>\\n>Yeah, if your mental model of a function call is some remarkably\\n>expensive\\n>PostGIS geometry manipulation, it's easy to justify doing a lot of work\\n>to try to detect duplicates. But most functions in most queries are\\n>more like int4pl or btint8cmp, and it's going to be extremely\\n>remarkable\\n>if you can make back the planner costs of checking for duplicate usages\\n>of those.\\n\\nWell, if it's an expression containing those individuals cheap calls on a seqscan on a large table below an aggregate, it'd likely still be a win. But we don't, to my knowledge, really have a good way to model optimizations like this that should only be done if either expensive or have a high loop count.\\n\\nI guess one ugly way to deal with this would be to eliminate redundancies very late, e.g. during setrefs (where a better data structure for matching expressions would be good anyway), as we already know all the costs. \\n\\nBut ideally we would want to do be able to take such savings into account earlier, when considering different paths. I suspect that it might be a good enough vehicle to tackle the rest of the work however, at least initially.\\n\\nWe could also \\\"just\\\" do such elimination during expression \\\"compilation\\\", but it'd be better to not have to do something as complicated as this for every execution of a prepared statement.\\n\\n\\n>> I think one part of doing this in a realistic manner is an efficient\\n>> search for redundant expressions. The other, also non trivial, is how\\n>to\\n>> even represent re eferences to the results of expressions in other\\n>parts of the expression tree / other expressions.\\n>\\n>Yup, both of those would be critical to do right.\\n\\nPotentially related note: for nodes like seqscan, combining the qual and projection processing into one expression seems to be a noticable win (at least when taking care do emit two different sets of deform expression steps). Wonder if something like that would take care of avoiding the need for cross expression value passing in enough places.\\n\\nAndres\\n-- \\nSent from my Android device with K-9 Mail. Please excuse my brevity.\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 08:20:39 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> Potentially related note: for nodes like seqscan, combining the qual and projection processing into one expression seems to be a noticable win (at least when taking care do emit two different sets of deform expression steps).\\n\\nThere's just one problem: that violates SQL semantics, and not in\\na small way.\\n\\n\\tselect 1/x from tab where x <> 0\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 11:45:26 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"Hi, \\n\\nOn October 31, 2019 8:45:26 AM PDT, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>Andres Freund <andres@anarazel.de> writes:\\n>> Potentially related note: for nodes like seqscan, combining the qual\\n>and projection processing into one expression seems to be a noticable\\n>win (at least when taking care do emit two different sets of deform\\n>expression steps).\\n>\\n>There's just one problem: that violates SQL semantics, and not in\\n>a small way.\\n>\\n>\\tselect 1/x from tab where x <> 0\\n\\nThe expression would obviously have to return early, before projecting, when not matching the qual? I'm basically just thinking of first executing the steps for the qual, and in the success case execute the projection steps before returning the quals positive result. \\n\\nAndres\\n-- \\nSent from my Android device with K-9 Mail. Please excuse my brevity.\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 08:49:46 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"x <> 0 is evaluated first, 1/x only when x <> 0, not ?\\n\\nczw., 31 paź 2019 o 16:45 Tom Lane <tgl@sss.pgh.pa.us> napisał(a):\\n\\n> Andres Freund <andres@anarazel.de> writes:\\n> > Potentially related note: for nodes like seqscan, combining the qual and\\n> projection processing into one expression seems to be a noticable win (at\\n> least when taking care do emit two different sets of deform expression\\n> steps).\\n>\\n> There's just one problem: that violates SQL semantics, and not in\\n> a small way.\\n>\\n> select 1/x from tab where x <> 0\\n>\\n> regards, tom lane\\n>\\n\\nx <> 0 is evaluated first, 1/x only when x <> 0, not ?czw., 31 paź 2019 o 16:45 Tom Lane <tgl@sss.pgh.pa.us> napisał(a):Andres Freund <andres@anarazel.de> writes:\\n> Potentially related note: for nodes like seqscan, combining the qual and projection processing into one expression seems to be a noticable win (at least when taking care do emit two different sets of deform expression steps).\\n\\nThere's just one problem: that violates SQL semantics, and not in\\na small way.\\n\\n select 1/x from tab where x <> 0\\n\\n regards, tom lane\",\n \"msg_date\": \"Thu, 31 Oct 2019 16:51:11 +0100\",\n \"msg_from\": \"Andrzej Barszcz <abusinf@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"Hi, \\n\\nOn October 31, 2019 8:51:11 AM PDT, Andrzej Barszcz <abusinf@gmail.com> wrote:\\n>x <> 0 is evaluated first, 1/x only when x <> 0, not ?\\n>\\n>czw., 31 paź 2019 o 16:45 Tom Lane <tgl@sss.pgh.pa.us> napisał(a):\\n>\\n>> Andres Freund <andres@anarazel.de> writes:\\n>> > Potentially related note: for nodes like seqscan, combining the\\n>qual and\\n>> projection processing into one expression seems to be a noticable win\\n>(at\\n>> least when taking care do emit two different sets of deform\\n>expression\\n>> steps).\\n>>\\n>> There's just one problem: that violates SQL semantics, and not in\\n>> a small way.\\n>>\\n>> select 1/x from tab where x <> 0\\n\\nOn postgres lists the policy is to reply below the quoted bit, and to trim the quote appropriately.\\n\\nAndres\\n-- \\nSent from my Android device with K-9 Mail. Please excuse my brevity.\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 08:52:53 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"I recently implemented something closely related to this. Combining and\\nmigrating expensive STABLE user-defined functions to the FROM clause, where\\nthe function is evaluated as a lateral join (or \\\"cross apply\\\"). I'm\\ndefining expensive as 50x times more expensive than the default function\\ncost.\\n\\nFor functions that return multiple outputs and where the query uses (...).*\\nnotation, this will, for example, consolidate all of the calls in the *\\nexpansion into a single call. It also looks in the WHERE clause and HAVING\\nclause, and combines those references, too. Currently it requires the\\nfunction to be in a top-level AND condition, if it appears in a predicate.\\n\\nI think I can get permission for contributing it back. If there's an\\ninterest in it, let me know.\\n\\n\\n\\n-----\\nJim Finnerty, AWS, Amazon Aurora PostgreSQL\\n--\\nSent from: https://www.postgresql-archive.org/PostgreSQL-hackers-f1928748.html\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 08:56:22 -0700 (MST)\",\n \"msg_from\": \"Jim Finnerty <jfinnert@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"Hi\\n\\nI need advice.\\nResetExprContext(econtext) is defined as\\nMemoryContextReset((econtext)->ecxt_per_tuple_memory).\\nI can register callback in MemoryContext but it is always cleaned on every\\ncall to MemoryContextReset().\\nHow to reset some fields of ExprContext ( living in per_query_memory ) when\\nResetExprContext is called ?\\n\\nczw., 31 paź 2019 o 16:52 Andres Freund <andres@anarazel.de> napisał(a):\\n\\n> Hi,\\n>\\n> On October 31, 2019 8:51:11 AM PDT, Andrzej Barszcz <abusinf@gmail.com>\\n> wrote:\\n> >x <> 0 is evaluated first, 1/x only when x <> 0, not ?\\n> >\\n> >czw., 31 paź 2019 o 16:45 Tom Lane <tgl@sss.pgh.pa.us> napisał(a):\\n> >\\n> >> Andres Freund <andres@anarazel.de> writes:\\n> >> > Potentially related note: for nodes like seqscan, combining the\\n> >qual and\\n> >> projection processing into one expression seems to be a noticable win\\n> >(at\\n> >> least when taking care do emit two different sets of deform\\n> >expression\\n> >> steps).\\n> >>\\n> >> There's just one problem: that violates SQL semantics, and not in\\n> >> a small way.\\n> >>\\n> >> select 1/x from tab where x <> 0\\n>\\n> On postgres lists the policy is to reply below the quoted bit, and to trim\\n> the quote appropriately.\\n>\\n> Andres\\n> --\\n> Sent from my Android device with K-9 Mail. Please excuse my brevity.\\n>\\n\\nHiI need advice. ResetExprContext(econtext) is defined as MemoryContextReset((econtext)->ecxt_per_tuple_memory).I can register callback in MemoryContext but it is always cleaned on every call to MemoryContextReset().How to reset some fields of ExprContext ( living in per_query_memory ) when ResetExprContext is called ?czw., 31 paź 2019 o 16:52 Andres Freund <andres@anarazel.de> napisał(a):Hi, \\n\\nOn October 31, 2019 8:51:11 AM PDT, Andrzej Barszcz <abusinf@gmail.com> wrote:\\n>x <> 0 is evaluated first, 1/x only when x <> 0, not ?\\n>\\n>czw., 31 paź 2019 o 16:45 Tom Lane <tgl@sss.pgh.pa.us> napisał(a):\\n>\\n>> Andres Freund <andres@anarazel.de> writes:\\n>> > Potentially related note: for nodes like seqscan, combining the\\n>qual and\\n>> projection processing into one expression seems to be a noticable win\\n>(at\\n>> least when taking care do emit two different sets of deform\\n>expression\\n>> steps).\\n>>\\n>> There's just one problem: that violates SQL semantics, and not in\\n>> a small way.\\n>>\\n>> select 1/x from tab where x <> 0\\n\\nOn postgres lists the policy is to reply below the quoted bit, and to trim the quote appropriately.\\n\\nAndres\\n-- \\nSent from my Android device with K-9 Mail. Please excuse my brevity.\",\n \"msg_date\": \"Mon, 18 Nov 2019 15:20:48 +0100\",\n \"msg_from\": \"Andrzej Barszcz <abusinf@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 11:07 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n>\\n>\\n> Possibly this could be finessed by only trying to find duplicates of\\n> functions that have high cost estimates. Not sure how high is high\\n> enough.\\n\\n\\ncan we just add a flag on pg_proc to show if the cost is high or not, if\\nuser are not happy with that, they can change it by updating the value?\\nbased on that most of the function call cost are low, this way may be\\nhelpful for the searching of duplicate expressions.\\n\\nOn Thu, Oct 31, 2019 at 11:07 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\nPossibly this could be finessed by only trying to find duplicates of\\nfunctions that have high cost estimates. Not sure how high is high\\nenough. can we just add a flag on pg_proc to show if the cost is high or not, if user are not happy with that, they can change it by updating the value? based on that most of the function call cost are low, this way may be helpful for the searching of duplicate expressions.\",\n \"msg_date\": \"Thu, 21 Nov 2019 09:05:14 +0800\",\n \"msg_from\": \"Andy Fan <zhihui.fan1213@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: function calls optimization\"\n },\n {\n \"msg_contents\": \"I think your first thought was good.\\nHow high ? I think it's a matter of convention, certainly more than default\\n100.\\n\\n\\n\\nczw., 21 lis 2019 o 02:05 Andy Fan <zhihui.fan1213@gmail.com> napisał(a):\\n\\n>\\n>\\n> On Thu, Oct 31, 2019 at 11:07 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n>>\\n>>\\n>> Possibly this could be finessed by only trying to find duplicates of\\n>> functions that have high cost estimates. Not sure how high is high\\n>> enough.\\n>\\n>\\n> can we just add a flag on pg_proc to show if the cost is high or not, if\\n> user are not happy with that, they can change it by updating the value?\\n> based on that most of the function call cost are low, this way may be\\n> helpful for the searching of duplicate expressions.\\n>\\n\\nI think your first thought was good. How high ? I think it's a matter of convention, certainly more than default 100. czw., 21 lis 2019 o 02:05 Andy Fan <zhihui.fan1213@gmail.com> napisał(a):On Thu, Oct 31, 2019 at 11:07 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\nPossibly this could be finessed by only trying to find duplicates of\\nfunctions that have high cost estimates. Not sure how high is high\\nenough. can we just add a flag on pg_proc to show if the cost is high or not, if user are not happy with that, they can change it by updating the value? based on that most of the function call cost are low, this way may be helpful for the searching of duplicate expressions.\",\n \"msg_date\": \"Thu, 21 Nov 2019 09:37:46 +0100\",\n \"msg_from\": \"Andrzej Barszcz <abusinf@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: function calls optimization\"\n }\n]","isConversational":false}}},{"rowIdx":49377,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"This is the first of a number of patches to enhance SSL functionality,\nparticularly w.r.t. passphrases.\n\n\nThis patch provides a hook for a function that can supply an SSL\npassphrase. The hook can be filled in by a shared preloadable module. In\norder for that to be effective, the startup order is modified slightly.\nThere is a test attached that builds and uses one trivial\nimplementation, which just takes a configuration setting and rot13's it\nbefore supplying the result as the passphrase.\n\n\ncheers\n\n\nandrew\n\n\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 31 Oct 2019 11:37:04 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"ssl passphrase callback"},{"msg_contents":"On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\n<andrew.dunstan@2ndquadrant.com> wrote:\n> This patch provides a hook for a function that can supply an SSL\n> passphrase. The hook can be filled in by a shared preloadable module. In\n> order for that to be effective, the startup order is modified slightly.\n> There is a test attached that builds and uses one trivial\n> implementation, which just takes a configuration setting and rot13's it\n> before supplying the result as the passphrase.\n\nIt seems to me that it would be a lot better to have an example in\ncontrib that does something which might be of actual use to users,\nsuch as running a shell command and reading the passphrase from\nstdout.\n\nFeatures that are only accessible by writing C code are, in general,\nnot as desirable as features which can be accessed via SQL or\nconfiguration.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 1 Nov 2019 11:01:15 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"\nOn 11/1/19 11:01 AM, Robert Haas wrote:\n> On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\n> <andrew.dunstan@2ndquadrant.com> wrote:\n>> This patch provides a hook for a function that can supply an SSL\n>> passphrase. The hook can be filled in by a shared preloadable module. In\n>> order for that to be effective, the startup order is modified slightly.\n>> There is a test attached that builds and uses one trivial\n>> implementation, which just takes a configuration setting and rot13's it\n>> before supplying the result as the passphrase.\n> It seems to me that it would be a lot better to have an example in\n> contrib that does something which might be of actual use to users,\n> such as running a shell command and reading the passphrase from\n> stdout.\n>\n> Features that are only accessible by writing C code are, in general,\n> not as desirable as features which can be accessed via SQL or\n> configuration.\n>\n\n\nWell, I tried to provide the most trivial and simple test I could come\nup with. Running a shell command can already be accomplished via the\nssl_passphrase_command setting.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Fri, 1 Nov 2019 13:57:29 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Sat, Nov 2, 2019 at 6:57 AM Andrew Dunstan\n<andrew.dunstan@2ndquadrant.com> wrote:\n> On 11/1/19 11:01 AM, Robert Haas wrote:\n> > On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\n> > <andrew.dunstan@2ndquadrant.com> wrote:\n> >> This patch provides a hook for a function that can supply an SSL\n> >> passphrase. The hook can be filled in by a shared preloadable module. In\n> >> order for that to be effective, the startup order is modified slightly.\n> >> There is a test attached that builds and uses one trivial\n> >> implementation, which just takes a configuration setting and rot13's it\n> >> before supplying the result as the passphrase.\n> > It seems to me that it would be a lot better to have an example in\n> > contrib that does something which might be of actual use to users,\n> > such as running a shell command and reading the passphrase from\n> > stdout.\n> >\n> > Features that are only accessible by writing C code are, in general,\n> > not as desirable as features which can be accessed via SQL or\n> > configuration.\n>\n> Well, I tried to provide the most trivial and simple test I could come\n> up with. Running a shell command can already be accomplished via the\n> ssl_passphrase_command setting.\n\nIt looks like the new declarations in libpq-be.h are ifdef'd out in a\nnon-USE_SSL build, but then we still try to build the new test module\nand it fails:\n\nhttps://ci.appveyor.com/project/postgresql-cfbot/postgresql/build/1.0.64071\n\n\n","msg_date":"Tue, 5 Nov 2019 10:43:45 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Fri, Nov 1, 2019 at 01:57:29PM -0400, Andrew Dunstan wrote:\n> \n> On 11/1/19 11:01 AM, Robert Haas wrote:\n> > On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\n> > <andrew.dunstan@2ndquadrant.com> wrote:\n> >> This patch provides a hook for a function that can supply an SSL\n> >> passphrase. The hook can be filled in by a shared preloadable module. In\n> >> order for that to be effective, the startup order is modified slightly.\n> >> There is a test attached that builds and uses one trivial\n> >> implementation, which just takes a configuration setting and rot13's it\n> >> before supplying the result as the passphrase.\n> > It seems to me that it would be a lot better to have an example in\n> > contrib that does something which might be of actual use to users,\n> > such as running a shell command and reading the passphrase from\n> > stdout.\n> >\n> > Features that are only accessible by writing C code are, in general,\n> > not as desirable as features which can be accessed via SQL or\n> > configuration.\n> >\n> \n> \n> Well, I tried to provide the most trivial and simple test I could come\n> up with. Running a shell command can already be accomplished via the\n> ssl_passphrase_command setting.\n\nWhat is the value of a shared library over a shell command? We had this\ndiscussion in relation to archive_command years ago, and decided on a\nshell command as the best API.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Wed, 6 Nov 2019 20:23:56 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On 11/4/19 4:43 PM, Thomas Munro wrote:\n>\n> It looks like the new declarations in libpq-be.h are ifdef'd out in a\n> non-USE_SSL build, but then we still try to build the new test module\n> and it fails:\n>\n> https://ci.appveyor.com/project/postgresql-cfbot/postgresql/build/1.0.64071\n\n\n\nI think this updated patch should fix things.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 7 Nov 2019 12:30:50 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Thu, 7 Nov 2019 at 10:24, Bruce Momjian <bruce@momjian.us> wrote:\n\n\n> What is the value of a shared library over a shell command? We had this\n> discussion in relation to archive_command years ago, and decided on a\n> shell command as the best API.\n>\n\nI don't recall such a discussion, but I can give perspective:\n\n* shell command offered the widest and simplest API for integration, which\nwas the most important consideration for a backup API. That choice caused\ndifficulty with the need to pass information to the external command, e.g.\n%f %p\n\n* shared library is more appropriate for a security-related module, so\nusers can't see how it is configured, as well as being more\ntightly integrated so it can be better tailored to various uses\n\nSummary is that the choice is not random, nor mere preference\n\n-- \nSimon Riggs http://www.2ndQuadrant.com/\n<http://www.2ndquadrant.com/>\nPostgreSQL Solutions for the Enterprise\n\nOn Thu, 7 Nov 2019 at 10:24, Bruce Momjian <bruce@momjian.us> wrote: What is the value of a shared library over a shell command? We had this\ndiscussion in relation to archive_command years ago, and decided on a\nshell command as the best API.I don't recall such a discussion, but I can give perspective:* shell command offered the widest and simplest API for integration, which was the most important consideration for a backup API. That choice caused difficulty with the need to pass information to the external command, e.g. %f %p* shared library is more appropriate for a security-related module, so users can't see how it is configured, as well as being more tightly integrated so it can be better tailored to various usesSummary is that the choice is not random, nor mere preference-- Simon Riggs http://www.2ndQuadrant.com/PostgreSQL Solutions for the Enterprise","msg_date":"Fri, 8 Nov 2019 23:12:08 +0900","msg_from":"Simon Riggs <simon@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Fri, Nov 08, 2019 at 11:12:08PM +0900, Simon Riggs wrote:\n>On Thu, 7 Nov 2019 at 10:24, Bruce Momjian <bruce@momjian.us> wrote:\n>\n>\n>> What is the value of a shared library over a shell command? We had\n>> this discussion in relation to archive_command years ago, and decided\n>> on a shell command as the best API.\n>>\n>\n>I don't recall such a discussion, but I can give perspective:\n>\n>* shell command offered the widest and simplest API for integration,\n>which was the most important consideration for a backup API. That\n>choice caused difficulty with the need to pass information to the\n>external command, e.g. %f %p\n>\n\nIt's not clear to me why simple API for integration would be less\nvaluable for this feature. Also, I'm sure passing data to/from shell\ncommand may be tricky, but presumably we have figured how to do that.\n\n>* shared library is more appropriate for a security-related module, so\n>users can't see how it is configured, as well as being more\n>tightly integrated so it can be better tailored to various uses\n>\n\nI don't follow. Why would there be a significant difference between a\nshell command/script and shared library in this respect? If you don't\nwant the users to see the config, just store it in a separate file and\nit's about the same as storing it in the .so library.\n\nIs there something that can be done with a .so library but can't be done\nwith a shell command (which may just call a binary, with all the config\nincluded, making it equal to the .so solution)?\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Sat, 9 Nov 2019 12:52:25 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\n\n> On Fri, Nov 1, 2019 at 01:57:29PM -0400, Andrew Dunstan wrote:\n> >\n> > On 11/1/19 11:01 AM, Robert Haas wrote:\n> > > On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\n> > > <andrew.dunstan@2ndquadrant.com> wrote:\n> > >> This patch provides a hook for a function that can supply an SSL\n> > >> passphrase. The hook can be filled in by a shared preloadable module.\n> In\n> > >> order for that to be effective, the startup order is modified\n> slightly.\n> > >> There is a test attached that builds and uses one trivial\n> > >> implementation, which just takes a configuration setting and rot13's\n> it\n> > >> before supplying the result as the passphrase.\n> > > It seems to me that it would be a lot better to have an example in\n> > > contrib that does something which might be of actual use to users,\n> > > such as running a shell command and reading the passphrase from\n> > > stdout.\n> > >\n> > > Features that are only accessible by writing C code are, in general,\n> > > not as desirable as features which can be accessed via SQL or\n> > > configuration.\n> > >\n> >\n> >\n> > Well, I tried to provide the most trivial and simple test I could come\n> > up with. Running a shell command can already be accomplished via the\n> > ssl_passphrase_command setting.\n>\n> What is the value of a shared library over a shell command?\n\n\nFor one, platforms where shell commands are a lot less convenient, such as\nWindows.\n\n\n\n> We had this\n> discussion in relation to archive_command years ago, and decided on a\n> shell command as the best API.\n>\n>\nI don't recall that from back then, but that was a long time ago.\n\nBut it's interesting that you mention it, given the number of people I have\nbeen discussing that with recently, under the topic of changing it from a\nshell command into a shared library API (with there being a shell command\nas one possible implementation of course).\n\nOne of the main reasons there being to be easily able to transfer more\nstate and give results other than just an exit code, no need to deal with\nparameter escaping etc. Which probably wouldn't matter as much to an SSL\npassphrase command, but still.\n\n//Magnus\n\nOn Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:On Fri, Nov 1, 2019 at 01:57:29PM -0400, Andrew Dunstan wrote:\n> \n> On 11/1/19 11:01 AM, Robert Haas wrote:\n> > On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\n> > <andrew.dunstan@2ndquadrant.com> wrote:\n> >> This patch provides a hook for a function that can supply an SSL\n> >> passphrase. The hook can be filled in by a shared preloadable module. In\n> >> order for that to be effective, the startup order is modified slightly.\n> >> There is a test attached that builds and uses one trivial\n> >> implementation, which just takes a configuration setting and rot13's it\n> >> before supplying the result as the passphrase.\n> > It seems to me that it would be a lot better to have an example in\n> > contrib that does something which might be of actual use to users,\n> > such as running a shell command and reading the passphrase from\n> > stdout.\n> >\n> > Features that are only accessible by writing C code are, in general,\n> > not as desirable as features which can be accessed via SQL or\n> > configuration.\n> >\n> \n> \n> Well, I tried to provide the most trivial and simple test I could come\n> up with. Running a shell command can already be accomplished via the\n> ssl_passphrase_command setting.\n\nWhat is the value of a shared library over a shell command?For one, platforms where shell commands are a lot less convenient, such as Windows. We had this\ndiscussion in relation to archive_command years ago, and decided on a\nshell command as the best API.I don't recall that from back then, but that was a long time ago.But it's interesting that you mention it, given the number of people I have been discussing that with recently, under the topic of changing it from a shell command into a shared library API (with there being a shell command as one possible implementation of course). One of the main reasons there being to be easily able to transfer more state and give results other than just an exit code, no need to deal with parameter escaping etc. Which probably wouldn't matter as much to an SSL passphrase command, but still.//Magnus","msg_date":"Sun, 10 Nov 2019 13:01:17 -0600","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On 11/7/19 12:30 PM, Andrew Dunstan wrote:\n> On 11/4/19 4:43 PM, Thomas Munro wrote:\n>> It looks like the new declarations in libpq-be.h are ifdef'd out in a\n>> non-USE_SSL build, but then we still try to build the new test module\n>> and it fails:\n>>\n>> https://ci.appveyor.com/project/postgresql-cfbot/postgresql/build/1.0.64071\n>\n>\n> I think this updated patch should fix things.\n>\n>\n\n\n\nThis time with a typo fixed to keep the cfbot happy.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Mon, 11 Nov 2019 14:19:44 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\n> On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\n> � We had this\n> discussion in relation to archive_command years ago, and decided on a\n> shell command as the best API.\n>\n> I don't recall that from back then, but that was a long time ago.\n> \n> But it's interesting that you mention it, given the number of people I have\n> been discussing that with recently, under the topic of changing it from a shell\n> command into a shared library API (with there being a shell command as one\n> possible implementation of course).�\n> \n> One of the main reasons there being to be easily able to transfer more state\n> and give results other than just an exit code, no need to deal with parameter\n> escaping etc. Which probably wouldn't matter as much to an SSL passphrase\n> command, but still.\n\nI get the callback-is-easier issue with shared objects, but are we\nexpecting to pass in more information here than we do for\narchive_command? I would think not. What I am saying is that if we\ndon't think passing things in works, we should fix all these external\ncommands, or something. I don't see why ssl_passphrase_command is\ndifferent, except that it is new. Or is it related to _securely_\npassing something?\n\nAlso, why was this patch posted without any discussion of these issues?\nShouldn't we ideally discuss the API first?\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Tue, 12 Nov 2019 21:51:33 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\n> On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\n> > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\n> > � We had this\n> > discussion in relation to archive_command years ago, and decided on a\n> > shell command as the best API.\n> >\n> > I don't recall that from back then, but that was a long time ago.\n> > \n> > But it's interesting that you mention it, given the number of people I have\n> > been discussing that with recently, under the topic of changing it from a shell\n> > command into a shared library API (with there being a shell command as one\n> > possible implementation of course).�\n> > \n> > One of the main reasons there being to be easily able to transfer more state\n> > and give results other than just an exit code, no need to deal with parameter\n> > escaping etc. Which probably wouldn't matter as much to an SSL passphrase\n> > command, but still.\n> \n> I get the callback-is-easier issue with shared objects, but are we\n> expecting to pass in more information here than we do for\n> archive_command? I would think not. What I am saying is that if we\n> don't think passing things in works, we should fix all these external\n> commands, or something. I don't see why ssl_passphrase_command is\n> different, except that it is new. Or is it related to _securely_\n> passing something?\n> \n> Also, why was this patch posted without any discussion of these issues?\n> Shouldn't we ideally discuss the API first?\n\nI wonder if every GUC that takes an OS command should allow a shared\nobject to be specified --- maybe control that if the command string\nstarts with a # or something.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Wed, 13 Nov 2019 08:08:23 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Wed, 13 Nov 2019 at 13:08, Bruce Momjian <bruce@momjian.us> wrote:\n\n> On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\n> > On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\n> > > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\n>\n> > > One of the main reasons there being to be easily able to transfer more\n> state\n> > > and give results other than just an exit code, no need to deal with\n> parameter\n> > > escaping etc. Which probably wouldn't matter as much to an SSL\n> passphrase\n> > > command, but still.\n> >\n> > I get the callback-is-easier issue with shared objects, but are we\n> > expecting to pass in more information here than we do for\n> > archive_command? I would think not. What I am saying is that if we\n> > don't think passing things in works, we should fix all these external\n> > commands, or something. I don't see why ssl_passphrase_command is\n> > different, except that it is new.\n\n\n\n> Or is it related to _securely_passing something?\n>\n\nYes\n\n\n> > Also, why was this patch posted without any discussion of these issues?\n> > Shouldn't we ideally discuss the API first?\n>\n> I wonder if every GUC that takes an OS command should allow a shared\n> object to be specified --- maybe control that if the command string\n> starts with a # or something.\n>\n\nVery good idea\n\n-- \nSimon Riggs http://www.2ndQuadrant.com/\n<http://www.2ndquadrant.com/>\nPostgreSQL Solutions for the Enterprise\n\nOn Wed, 13 Nov 2019 at 13:08, Bruce Momjian <bruce@momjian.us> wrote:On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\n> On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\n> > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\n> > One of the main reasons there being to be easily able to transfer more state\n> > and give results other than just an exit code, no need to deal with parameter\n> > escaping etc. Which probably wouldn't matter as much to an SSL passphrase\n> > command, but still.\n> \n> I get the callback-is-easier issue with shared objects, but are we\n> expecting to pass in more information here than we do for\n> archive_command? I would think not. What I am saying is that if we\n> don't think passing things in works, we should fix all these external\n> commands, or something. I don't see why ssl_passphrase_command is\n> different, except that it is new. Or is it related to _securely_passing something?Yes \n> Also, why was this patch posted without any discussion of these issues?\n> Shouldn't we ideally discuss the API first?\n\nI wonder if every GUC that takes an OS command should allow a shared\nobject to be specified --- maybe control that if the command string\nstarts with a # or something.Very good idea -- Simon Riggs http://www.2ndQuadrant.com/PostgreSQL Solutions for the Enterprise","msg_date":"Wed, 13 Nov 2019 13:20:43 +0000","msg_from":"Simon Riggs <simon@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"\nOn 11/13/19 8:08 AM, Bruce Momjian wrote:\n>\n>>\n>> Also, why was this patch posted without any discussion of these issues?\n>> Shouldn't we ideally discuss the API first?\n\n\nThis is a very tiny patch. I don't think it's unusual to post such\nthings without prior discussion. I would agree with your point if it\nwere thousands of lines instead of 20 or so lines of core code.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Wed, 13 Nov 2019 14:48:01 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Wed, Nov 13, 2019 at 01:20:43PM +0000, Simon Riggs wrote:\n>On Wed, 13 Nov 2019 at 13:08, Bruce Momjian <bruce@momjian.us> wrote:\n>\n>> On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\n>> > On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\n>> > > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\n>>\n>> > > One of the main reasons there being to be easily able to transfer more\n>> state\n>> > > and give results other than just an exit code, no need to deal with\n>> parameter\n>> > > escaping etc. Which probably wouldn't matter as much to an SSL\n>> passphrase\n>> > > command, but still.\n>> >\n>> > I get the callback-is-easier issue with shared objects, but are we\n>> > expecting to pass in more information here than we do for\n>> > archive_command? I would think not. What I am saying is that if we\n>> > don't think passing things in works, we should fix all these external\n>> > commands, or something. I don't see why ssl_passphrase_command is\n>> > different, except that it is new.\n>\n>\n>\n>> Or is it related to _securely_passing something?\n>>\n>\n>Yes\n>\n\nI think it would be beneficial to explain why shared object is more\nsecure than an OS command. Perhaps it's common knowledge, but it's not\nquite obvious to me.\n\n>\n>> > Also, why was this patch posted without any discussion of these issues?\n>> > Shouldn't we ideally discuss the API first?\n>>\n>> I wonder if every GUC that takes an OS command should allow a shared\n>> object to be specified --- maybe control that if the command string\n>> starts with a # or something.\n>>\n>\n>Very good idea\n>\n\nIf it's about securely passing sensitive information (i.e. passphrase)\nas was suggested, then I think that only applies to fairly small number\nof GUCs.\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Wed, 13 Nov 2019 21:23:23 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Wed, Nov 13, 2019 at 02:48:01PM -0500, Andrew Dunstan wrote:\n>\n>On 11/13/19 8:08 AM, Bruce Momjian wrote:\n>>\n>>>\n>>> Also, why was this patch posted without any discussion of these issues?\n>>> Shouldn't we ideally discuss the API first?\n>\n>\n>This is a very tiny patch. I don't think it's unusual to post such\n>things without prior discussion. I would agree with your point if it\n>were thousands of lines instead of 20 or so lines of core code.\n>\n\nNot sure that's really true. I think patches should provide some basic\nexplanation why the feature is desirable, no matter the number of lines.\n\nAlso, there were vague references to issues with passing parameters to\narchive_command. A link to details, past discussion, or brief\nexplanation would be nice.\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Wed, 13 Nov 2019 21:34:38 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\nwrote:\n\n> On Wed, Nov 13, 2019 at 01:20:43PM +0000, Simon Riggs wrote:\n> >On Wed, 13 Nov 2019 at 13:08, Bruce Momjian <bruce@momjian.us> wrote:\n> >\n> >> On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\n> >> > On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\n> >> > > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us>\n> wrote:\n> >>\n> >> > > One of the main reasons there being to be easily able to transfer\n> more\n> >> state\n> >> > > and give results other than just an exit code, no need to deal with\n> >> parameter\n> >> > > escaping etc. Which probably wouldn't matter as much to an SSL\n> >> passphrase\n> >> > > command, but still.\n> >> >\n> >> > I get the callback-is-easier issue with shared objects, but are we\n> >> > expecting to pass in more information here than we do for\n> >> > archive_command? I would think not. What I am saying is that if we\n> >> > don't think passing things in works, we should fix all these external\n> >> > commands, or something. I don't see why ssl_passphrase_command is\n> >> > different, except that it is new.\n> >\n> >\n> >\n> >> Or is it related to _securely_passing something?\n> >>\n> >\n> >Yes\n> >\n>\n> I think it would be beneficial to explain why shared object is more\n> secure than an OS command. Perhaps it's common knowledge, but it's not\n> quite obvious to me.\n>\n\nYeah, that probably wouldn't hurt. It's also securely passing from more\nthan one perspective -- both from the \"cannot be eavesdropped\" (like\nputting the password on the commandline for example) and the requirement\nfor escaping.\n\n\n>\n> >\n> >> > Also, why was this patch posted without any discussion of these\n> issues?\n> >> > Shouldn't we ideally discuss the API first?\n> >>\n> >> I wonder if every GUC that takes an OS command should allow a shared\n> >> object to be specified --- maybe control that if the command string\n> >> starts with a # or something.\n> >>\n> >\n> >Very good idea\n> >\n>\n> If it's about securely passing sensitive information (i.e. passphrase)\n> as was suggested, then I think that only applies to fairly small number\n> of GUCs.\n>\n\nThere aren't exactly a large number of GUCs that take OS commands in total.\nConsistency itself certainly has some value.\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 Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:On Wed, Nov 13, 2019 at 01:20:43PM +0000, Simon Riggs wrote:\n>On Wed, 13 Nov 2019 at 13:08, Bruce Momjian <bruce@momjian.us> wrote:\n>\n>> On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\n>> > On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\n>> > > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\n>>\n>> > > One of the main reasons there being to be easily able to transfer more\n>> state\n>> > > and give results other than just an exit code, no need to deal with\n>> parameter\n>> > > escaping etc. Which probably wouldn't matter as much to an SSL\n>> passphrase\n>> > > command, but still.\n>> >\n>> > I get the callback-is-easier issue with shared objects, but are we\n>> > expecting to pass in more information here than we do for\n>> > archive_command? I would think not. What I am saying is that if we\n>> > don't think passing things in works, we should fix all these external\n>> > commands, or something. I don't see why ssl_passphrase_command is\n>> > different, except that it is new.\n>\n>\n>\n>> Or is it related to _securely_passing something?\n>>\n>\n>Yes\n>\n\nI think it would be beneficial to explain why shared object is more\nsecure than an OS command. Perhaps it's common knowledge, but it's not\nquite obvious to me.Yeah, that probably wouldn't hurt. It's also securely passing from more than one perspective -- both from the \"cannot be eavesdropped\" (like putting the password on the commandline for example) and the requirement for escaping. \n\n>\n>> > Also, why was this patch posted without any discussion of these issues?\n>> > Shouldn't we ideally discuss the API first?\n>>\n>> I wonder if every GUC that takes an OS command should allow a shared\n>> object to be specified --- maybe control that if the command string\n>> starts with a # or something.\n>>\n>\n>Very good idea\n>\n\nIf it's about securely passing sensitive information (i.e. passphrase)\nas was suggested, then I think that only applies to fairly small number\nof GUCs.There aren't exactly a large number of GUCs that take OS commands in total. Consistency itself certainly has some value. -- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/","msg_date":"Thu, 14 Nov 2019 11:42:05 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Wed, Nov 13, 2019 at 3:23 PM Tomas Vondra\n<tomas.vondra@2ndquadrant.com> wrote:\n> I think it would be beneficial to explain why shared object is more\n> secure than an OS command. Perhaps it's common knowledge, but it's not\n> quite obvious to me.\n\nExternal command args can be viewed by other OS users (not just the\npostgres user). For non-sensitive arguments (ex: WAL filename?) that's\nnot an issue but if you plan on passing in something potentially\nsecret value from the backend to the external OS command, that value\nwould be exposed:\n\nTerminal 1 (run a command as some other user):\n$ sudo -u nobody sleep 5\n\nTerminal 2 (view command args as a different non-super user):\n$ ps -u nobody -o args\nCOMMAND\nsleep 5\n\nA shared library would not have this problem as the backend directly\nexecutes the library in the same process.\n\nHas the idea of using environment variables (rather than command line\nargs) for external commands been brought up before? I couldn't find\nanything in the mailing list archives.\n\nEnvironment variables have the advantage of only being readable by the\nprocess owner and super user. They also naturally have a \"name\" and do\nnot have escaping or quoting issues.\n\nFor example, archive_command %p could be exposed as\nPG_ARG_ARCHIVE_PATH and %f could be exposed as\nPG_ARG_ARCHIVE_FILENAME. Then you could have a script use them via:\n\n#!/usr/bin/env bash\nset -euo pipefail\nmain () {\n local archive_dir=\"/path/to/archive_dir/\"\n local archive_file=\"${archive_dir}${PG_ARG_ARCHIVE_FILENAME}\"\n test ! -f \"${archive_file}\" && cp -- \"${PG_ARG_ARCHIVE_PATH}\"\n\"${archive_file}\"\n}\nmain \"$@\"\n\nIt's not particularly useful for that basic archive case but if\nthere's something like PG_ARG_SUPER_SECRET then the executed command\ncould receive that value without it being exposed. That'd be useful\nfor something like a callout to an external KMS (key management\nsystem).\n\nNothing stops something like this with being used in tandem with\nstring substitution to create the full commands. That'd give backwards\ncompatibility too. The main limitation compared to a shared library is\nthat you'd still have to explicitly pick and name the exposed argument\nenvironment variables (i.e. like picking the set of %? substitutions).\nIf there's a generic shared-library-for-external-commands approach\nthen there could be a built-in library that provides this type of\n\"expose as env vars\" functionality.\n\nRegards,\n-- Sehrope Sarkuni\nFounder & CEO | JackDB, Inc. | https://www.jackdb.com/\n\n\n","msg_date":"Thu, 14 Nov 2019 08:54:27 -0500","msg_from":"Sehrope Sarkuni <sehrope@jackdb.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\n> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\n> I think it would be beneficial to explain why shared object is more\n> secure than an OS command. Perhaps it's common knowledge, but it's not\n> quite obvious to me.\n> \n> \n> Yeah, that probably wouldn't hurt. It's also securely passing from more than\n> one perspective -- both from the \"cannot be eavesdropped\" (like putting the\n> password on the commandline for example) and the requirement for escaping.\n\nI think a bigger issue is that if you want to give people the option of\nusing a shell command or a shared object, and if you use two commands to\ncontrol it, it isn't clear what happens if both are defined. By using\nsome character prefix to control if a shared object is used, you can use\na single variable and there is no confusion over having two variables\nand their conflicting behavior.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Thu, 14 Nov 2019 11:07:52 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"\nOn 11/14/19 11:07 AM, Bruce Momjian wrote:\n> On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\n>> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\n>> I think it would be beneficial to explain why shared object is more\n>> secure than an OS command. Perhaps it's common knowledge, but it's not\n>> quite obvious to me.\n>>\n>>\n>> Yeah, that probably wouldn't hurt. It's also securely passing from more than\n>> one perspective -- both from the \"cannot be eavesdropped\" (like putting the\n>> password on the commandline for example) and the requirement for escaping.\n> I think a bigger issue is that if you want to give people the option of\n> using a shell command or a shared object, and if you use two commands to\n> control it, it isn't clear what happens if both are defined. By using\n> some character prefix to control if a shared object is used, you can use\n> a single variable and there is no confusion over having two variables\n> and their conflicting behavior.\n>\n\n\nI'm not sure how that would work in the present instance. The shared\npreloaded module installs a function and defines the params it wants. If\nwe somehow unify the params with ssl_passphrase_command that could look\nicky, and the module would have to parse the settings string. That's not\na problem for the sample module which only needs one param, but it will\nbe for other more complex implementations.\n\nI'm quite open to suggestions, but I want things to be tolerably clean.\n\n\ncheers\n\n\nandrew\n\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Thu, 14 Nov 2019 11:34:24 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Thu, Nov 14, 2019 at 11:34:24AM -0500, Andrew Dunstan wrote:\n>\n>On 11/14/19 11:07 AM, Bruce Momjian wrote:\n>> On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\n>>> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\n>>> I think it would be beneficial to explain why shared object is more\n>>> secure than an OS command. Perhaps it's common knowledge, but it's not\n>>> quite obvious to me.\n>>>\n>>>\n>>> Yeah, that probably wouldn't hurt. It's also securely passing from more than\n>>> one perspective -- both from the \"cannot be eavesdropped\" (like putting the\n>>> password on the commandline for example) and the requirement for escaping.\n>> I think a bigger issue is that if you want to give people the option of\n>> using a shell command or a shared object, and if you use two commands to\n>> control it, it isn't clear what happens if both are defined. By using\n>> some character prefix to control if a shared object is used, you can use\n>> a single variable and there is no confusion over having two variables\n>> and their conflicting behavior.\n>>\n>\n>\n>I'm� not sure how that would work in the present instance. The shared\n>preloaded module installs a function and defines the params it wants. If\n>we somehow unify the params with ssl_passphrase_command that could look\n>icky, and the module would have to parse the settings string. That's not\n>a problem for the sample module which only needs one param, but it will\n>be for other more complex implementations.\n>\n>I'm quite open to suggestions, but I want things to be tolerably clean.\n>\n\nI agree it's better to have two separate GUCs - one for command, one for\nshared object, and documented order of precedence. I suppose we may log\na warning when both are specified, or perhaps \"reset\" the value with\nlower order of precedence.\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 14 Nov 2019 17:52:50 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Thu, Nov 14, 2019 at 11:34:24AM -0500, Andrew Dunstan wrote:\n> \n> On 11/14/19 11:07 AM, Bruce Momjian wrote:\n> > On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\n> >> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\n> >> I think it would be beneficial to explain why shared object is more\n> >> secure than an OS command. Perhaps it's common knowledge, but it's not\n> >> quite obvious to me.\n> >>\n> >>\n> >> Yeah, that probably wouldn't hurt. It's also securely passing from more than\n> >> one perspective -- both from the \"cannot be eavesdropped\" (like putting the\n> >> password on the commandline for example) and the requirement for escaping.\n> > I think a bigger issue is that if you want to give people the option of\n> > using a shell command or a shared object, and if you use two commands to\n> > control it, it isn't clear what happens if both are defined. By using\n> > some character prefix to control if a shared object is used, you can use\n> > a single variable and there is no confusion over having two variables\n> > and their conflicting behavior.\n> >\n> \n> \n> I'm� not sure how that would work in the present instance. The shared\n> preloaded module installs a function and defines the params it wants. If\n> we somehow unify the params with ssl_passphrase_command that could look\n> icky, and the module would have to parse the settings string. That's not\n> a problem for the sample module which only needs one param, but it will\n> be for other more complex implementations.\n> \n> I'm quite open to suggestions, but I want things to be tolerably clean.\n\nI was assuming if the variable starts with a #, it is a shared object,\nif not, it is a shell command:\n\n\tssl_passphrase_command='#/lib/x.so'\n\tssl_passphrase_command='my_command a b c'\n\nCan you show what you are talking about?\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Thu, 14 Nov 2019 11:53:09 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On 2019-Nov-14, Bruce Momjian wrote:\n\n> I was assuming if the variable starts with a #, it is a shared object,\n> if not, it is a shell command:\n> \n> \tssl_passphrase_command='#/lib/x.so'\n> \tssl_passphrase_command='my_command a b c'\n\nNote that the proposed patch doesn't use a separate GUC -- it just uses\nshared_preload_libraries, and then it is the library that's in charge of\nsetting up the function. We probably wouldn't like to have multiple\nsettings that all do the same thing, such as recovery target (which\nseems to be a plentiful source of confusion).\n\nChanging the interface so that the user has to specify the function name\n(not the library name) in ssl_passphrase_command closes that ambiguity\nhole.\n\nNote that if you specify only the library name, it becomes redundant\nw.r.t. shared_preload_libraries; you could have more than one library\nsetting the function callback and it's hard to see which one wins.\n\nI think something like this would do it:\n ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\n\nThis way, the library can still create any custom GUCs it pleases/needs,\nbut there's no possible confusion as to the function that's going to be\ncalled.\n\n-- \n�lvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 14 Nov 2019 14:07:58 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Thu, Nov 14, 2019 at 02:07:58PM -0300, Alvaro Herrera wrote:\n> On 2019-Nov-14, Bruce Momjian wrote:\n> \n> > I was assuming if the variable starts with a #, it is a shared object,\n> > if not, it is a shell command:\n> > \n> > \tssl_passphrase_command='#/lib/x.so'\n> > \tssl_passphrase_command='my_command a b c'\n> \n> Note that the proposed patch doesn't use a separate GUC -- it just uses\n> shared_preload_libraries, and then it is the library that's in charge of\n> setting up the function. We probably wouldn't like to have multiple\n> settings that all do the same thing, such as recovery target (which\n> seems to be a plentiful source of confusion).\n> \n> Changing the interface so that the user has to specify the function name\n> (not the library name) in ssl_passphrase_command closes that ambiguity\n> hole.\n> \n> Note that if you specify only the library name, it becomes redundant\n> w.r.t. shared_preload_libraries; you could have more than one library\n> setting the function callback and it's hard to see which one wins.\n> \n> I think something like this would do it:\n> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\n> \n> This way, the library can still create any custom GUCs it pleases/needs,\n> but there's no possible confusion as to the function that's going to be\n> called.\n\nYeah, I was unclear how the function name would be specified. I thought\nit would just be hard-coded, but I like the above better. I am still\nunclear how parameters are passed.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Thu, 14 Nov 2019 12:15:44 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"\nOn 11/14/19 12:07 PM, Alvaro Herrera wrote:\n> On 2019-Nov-14, Bruce Momjian wrote:\n>\n>> I was assuming if the variable starts with a #, it is a shared object,\n>> if not, it is a shell command:\n>>\n>> \tssl_passphrase_command='#/lib/x.so'\n>> \tssl_passphrase_command='my_command a b c'\n> Note that the proposed patch doesn't use a separate GUC -- it just uses\n> shared_preload_libraries, and then it is the library that's in charge of\n> setting up the function. We probably wouldn't like to have multiple\n> settings that all do the same thing, such as recovery target (which\n> seems to be a plentiful source of confusion).\n>\n> Changing the interface so that the user has to specify the function name\n> (not the library name) in ssl_passphrase_command closes that ambiguity\n> hole.\n>\n> Note that if you specify only the library name, it becomes redundant\n> w.r.t. shared_preload_libraries; you could have more than one library\n> setting the function callback and it's hard to see which one wins.\n>\n> I think something like this would do it:\n> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\n>\n> This way, the library can still create any custom GUCs it pleases/needs,\n> but there's no possible confusion as to the function that's going to be\n> called.\n\n\nI guess this would work. There would have to be a deal of code to load\nthe library and lookup the symbol. Do we really think it's worth it?\nLeveraging shared_preload_libraries makes this comparatively simple.\n\n\nAlso, calling this 'ssl_passphrase_command' seems a little odd.\n\n\nA simpler way to handle it might be simply to error out and refuse to\nstart if both ssl_passphrase_function is set and ssl_passphrase_command\nis set.\n\n\ncheers\n\n\nandrew\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Thu, 14 Nov 2019 14:29:23 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On 2019-Nov-14, Andrew Dunstan wrote:\n\n> I guess this would work. There would have to be a deal of code to load\n> the library and lookup the symbol. Do we really think it's worth it?\n> Leveraging shared_preload_libraries makes this comparatively simple.\n\nUsing the generic interface has the drawback that the user can make more\nmistakes. I think that's part of Bruce's issue with it (although I may\nmisinterpret.)\n\nI think if you add most of it as a new entry point in dfmgr.c (where you\ncan leverage internal_library_load) and returns a function pointer to\nthe user specified function, it's all that much additional code.\n\n(I don't think you can use load_external_function as is, because it\nassumes fmgr V1 calling convention, which I'm not sure serves your case.\nBut then maybe it does. And if not, then those 10 lines should be very\nsimilar to the code you'd need to add.)\n\n> A simpler way to handle it might be simply to error out and refuse to\n> start if both ssl_passphrase_function is set and ssl_passphrase_command\n> is set.\n\nYeah, you can do that too I guess, but I'm not sure I see that as simpler.\n\n> Also, calling this 'ssl_passphrase_command' seems a little odd.\n\nWe could just rename ssl_passphrase_command to something more\ngeneric, and add the existing name to map_old_guc_names to preserve\ncompatibility with pg12. Maybe the new name could be simply\nssl_passphrase or perhaps ssl_passphrase_{reader,getter,pinentry}.\n\n-- \n�lvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 14 Nov 2019 17:21:51 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"\nOn 11/14/19 3:21 PM, Alvaro Herrera wrote:\n> On 2019-Nov-14, Andrew Dunstan wrote:\n>\n>> I guess this would work. There would have to be a deal of code to load\n>> the library and lookup the symbol. Do we really think it's worth it?\n>> Leveraging shared_preload_libraries makes this comparatively simple.\n> Using the generic interface has the drawback that the user can make more\n> mistakes. I think that's part of Bruce's issue with it (although I may\n> misinterpret.)\n>\n> I think if you add most of it as a new entry point in dfmgr.c (where you\n> can leverage internal_library_load) and returns a function pointer to\n> the user specified function, it's all that much additional code.\n>\n> (I don't think you can use load_external_function as is, because it\n> assumes fmgr V1 calling convention, which I'm not sure serves your case.\n> But then maybe it does. And if not, then those 10 lines should be very\n> similar to the code you'd need to add.)\n\n\n\nIn the absence of further comment I will try to code up something along\nthese lines.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Fri, 15 Nov 2019 08:59:45 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Fri, 15 Nov 2019 at 00:34, Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\nwrote:\n\n>\n> On 11/14/19 11:07 AM, Bruce Momjian wrote:\n> > On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\n> >> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <\n> tomas.vondra@2ndquadrant.com>\n> >> I think it would be beneficial to explain why shared object is more\n> >> secure than an OS command. Perhaps it's common knowledge, but it's\n> not\n> >> quite obvious to me.\n> >>\n> >>\n> >> Yeah, that probably wouldn't hurt. It's also securely passing from more\n> than\n> >> one perspective -- both from the \"cannot be eavesdropped\" (like putting\n> the\n> >> password on the commandline for example) and the requirement for\n> escaping.\n> > I think a bigger issue is that if you want to give people the option of\n> > using a shell command or a shared object, and if you use two commands to\n> > control it, it isn't clear what happens if both are defined. By using\n> > some character prefix to control if a shared object is used, you can use\n> > a single variable and there is no confusion over having two variables\n> > and their conflicting behavior.\n> >\n>\n>\n> I'm not sure how that would work in the present instance. The shared\n> preloaded module installs a function and defines the params it wants. If\n> we somehow unify the params with ssl_passphrase_command that could look\n> icky, and the module would have to parse the settings string. That's not\n> a problem for the sample module which only needs one param, but it will\n> be for other more complex implementations.\n>\n> I'm quite open to suggestions, but I want things to be tolerably clean.\n\n\nIf someone wants a shell command wrapper, they can load a contrib that\ndelegates the hook to a shell command. So we can just ship a contrib, which\nacts both as test coverage for the feature, and a shell-command-support\nwrapper for anyone who desires that.\n\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\n\nOn Fri, 15 Nov 2019 at 00:34, Andrew Dunstan <andrew.dunstan@2ndquadrant.com> wrote:\nOn 11/14/19 11:07 AM, Bruce Momjian wrote:\n> On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\n>> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\n>> I think it would be beneficial to explain why shared object is more\n>> secure than an OS command. Perhaps it's common knowledge, but it's not\n>> quite obvious to me.\n>>\n>>\n>> Yeah, that probably wouldn't hurt. It's also securely passing from more than\n>> one perspective -- both from the \"cannot be eavesdropped\" (like putting the\n>> password on the commandline for example) and the requirement for escaping.\n> I think a bigger issue is that if you want to give people the option of\n> using a shell command or a shared object, and if you use two commands to\n> control it, it isn't clear what happens if both are defined. By using\n> some character prefix to control if a shared object is used, you can use\n> a single variable and there is no confusion over having two variables\n> and their conflicting behavior.\n>\n\n\nI'm not sure how that would work in the present instance. The shared\npreloaded module installs a function and defines the params it wants. If\nwe somehow unify the params with ssl_passphrase_command that could look\nicky, and the module would have to parse the settings string. That's not\na problem for the sample module which only needs one param, but it will\nbe for other more complex implementations.\n\nI'm quite open to suggestions, but I want things to be tolerably clean.If someone wants a shell command wrapper, they can load a contrib that delegates the hook to a shell command. So we can just ship a contrib, which acts both as test coverage for the feature, and a shell-command-support wrapper for anyone who desires that. -- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise","msg_date":"Fri, 22 Nov 2019 13:19:24 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"\nOn 11/15/19 8:59 AM, Andrew Dunstan wrote:\n> On 11/14/19 3:21 PM, Alvaro Herrera wrote:\n>> On 2019-Nov-14, Andrew Dunstan wrote:\n>>\n>>> I guess this would work. There would have to be a deal of code to load\n>>> the library and lookup the symbol. Do we really think it's worth it?\n>>> Leveraging shared_preload_libraries makes this comparatively simple.\n>> Using the generic interface has the drawback that the user can make more\n>> mistakes. I think that's part of Bruce's issue with it (although I may\n>> misinterpret.)\n>>\n>> I think if you add most of it as a new entry point in dfmgr.c (where you\n>> can leverage internal_library_load) and returns a function pointer to\n>> the user specified function, it's all that much additional code.\n>>\n>> (I don't think you can use load_external_function as is, because it\n>> assumes fmgr V1 calling convention, which I'm not sure serves your case.\n>> But then maybe it does. And if not, then those 10 lines should be very\n>> similar to the code you'd need to add.)\n\n\nI've just been looking at that. load_external_function() doesn't\nactually do anything V1-ish with the value, it just looks up the symbol\nusing dlsym and returns it cast to a PGFunction. Is there any reason I\ncan't just use that and cast it again to the callback function type?\n\n\ncheers\n\n\nandrew\n\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Fri, 6 Dec 2019 17:21:58 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n> I've just been looking at that. load_external_function() doesn't\n> actually do anything V1-ish with the value, it just looks up the symbol\n> using dlsym and returns it cast to a PGFunction. Is there any reason I\n> can't just use that and cast it again to the callback function type?\n\nTBH, I think this entire discussion has gone seriously off into the\nweeds. The original design where we just let a shared_preload_library\nfunction get into a hook is far superior to any of the overcomplicated\nkluges that are being discussed now. Something like this, for instance:\n\n>>> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\n\nmakes me positively ill. It introduces problems that we don't need,\nlike how to parse out the sub-parts of the string, and the\nquoting/escaping issues that will come along with that; while from\nthe user's perspective it replaces a simple and intellectually-coherent\nvariable definition with an unintelligible mess.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 06 Dec 2019 18:20:54 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"\nOn 12/6/19 6:20 PM, Tom Lane wrote:\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n>> I've just been looking at that. load_external_function() doesn't\n>> actually do anything V1-ish with the value, it just looks up the symbol\n>> using dlsym and returns it cast to a PGFunction. Is there any reason I\n>> can't just use that and cast it again to the callback function type?\n> TBH, I think this entire discussion has gone seriously off into the\n> weeds. The original design where we just let a shared_preload_library\n> function get into a hook is far superior to any of the overcomplicated\n> kluges that are being discussed now. Something like this, for instance:\n>\n>>>> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\n> makes me positively ill. It introduces problems that we don't need,\n> like how to parse out the sub-parts of the string, and the\n> quoting/escaping issues that will come along with that; while from\n> the user's perspective it replaces a simple and intellectually-coherent\n> variable definition with an unintelligible mess.\n>\n> \t\t\t\n\n\n\nYeah, you have a point.\n\n\nBruce was worried about what would happen if we defined both\nssl_passphrase_command and ssl_passphrase_callback. The submitted patch\nlet's the callback have precedence, but it might be cleaner to error out\nwith such a config. OTOH, that wouldn't be so nice on a reload, so it\nmight be better just to document the behaviour.\n\n\nHe was also worried that multiple shared libraries might try to provide\nthe hook. I think that's fairly fanciful, TBH. It comes into the\ncategory of \"Don't do that.\"\n\n\ncheers\n\n\nandrew\n\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Fri, 6 Dec 2019 19:32:32 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n> Bruce was worried about what would happen if we defined both\n> ssl_passphrase_command and ssl_passphrase_callback. The submitted patch\n> let's the callback have precedence, but it might be cleaner to error out\n> with such a config. OTOH, that wouldn't be so nice on a reload, so it\n> might be better just to document the behaviour.\n\nI think it would be up to the extension that's using the hook to\ndecide what to do if ssl_passphrase_command is set. It would not\nbe our choice, and it would certainly not fall to us to document it.\n\n> He was also worried that multiple shared libraries might try to provide\n> the hook. I think that's fairly fanciful, TBH. It comes into the\n> category of \"Don't do that.\"\n\nAgain, it's somebody else's problem. We have plenty of hooks that\nare of dubious use for multiple extensions, so why should this one be\nheld to a higher standard?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 07 Dec 2019 12:16:17 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"\nOn 12/7/19 12:16 PM, Tom Lane wrote:\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n>> Bruce was worried about what would happen if we defined both\n>> ssl_passphrase_command and ssl_passphrase_callback. The submitted patch\n>> let's the callback have precedence, but it might be cleaner to error out\n>> with such a config. OTOH, that wouldn't be so nice on a reload, so it\n>> might be better just to document the behaviour.\n> I think it would be up to the extension that's using the hook to\n> decide what to do if ssl_passphrase_command is set. It would not\n> be our choice, and it would certainly not fall to us to document it.\n>\n>> He was also worried that multiple shared libraries might try to provide\n>> the hook. I think that's fairly fanciful, TBH. It comes into the\n>> category of \"Don't do that.\"\n> Again, it's somebody else's problem. We have plenty of hooks that\n> are of dubious use for multiple extensions, so why should this one be\n> held to a higher standard?\n>\n> \t\t\t\n\n\nWell that pretty much brings us back to the patch as submitted :-)\n\n\ncheers\n\n\nandrew\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Sat, 7 Dec 2019 16:03:27 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n> Well that pretty much brings us back to the patch as submitted :-)\n\nYeah, pretty nearly. Taking a quick look over the v3 patch, my\nonly quibble is that it doesn't provide any convenient way for the\nexternal module to make decisions about how to interact with\nssl_passphrase_command --- in particular, if it would like to allow\nthat to take precedence, it can't because there's no way for it to\ninvoke the static function ssl_external_passwd_cb.\n\nBut rather than expose that globally, maybe the theory ought to be\n\"set up the state as we'd normally do, then let loadable modules\nchoose to override it\". So I'm tempted to propose a hook function\nwith the signature\n\nvoid openssl_tls_init_hook(SSL_CTX *context, bool isServerStart);\n\nand invoke that somewhere in be_tls_init --- maybe fairly late,\nso that it can override other settings if it wants, not only the\nSSL_CTX_set_default_passwd_cb setting.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 07 Dec 2019 17:32:01 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Sat, 7 Dec 2019 at 07:21, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n> > I've just been looking at that. load_external_function() doesn't\n> > actually do anything V1-ish with the value, it just looks up the symbol\n> > using dlsym and returns it cast to a PGFunction. Is there any reason I\n> > can't just use that and cast it again to the callback function type?\n>\n> TBH, I think this entire discussion has gone seriously off into the\n> weeds. The original design where we just let a shared_preload_library\n> function get into a hook is far superior to any of the overcomplicated\n> kluges that are being discussed now. Something like this, for instance:\n>\n> >>> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\n>\n> makes me positively ill. It introduces problems that we don't need,\n> like how to parse out the sub-parts of the string, and the\n> quoting/escaping issues that will come along with that; while from\n> the user's perspective it replaces a simple and intellectually-coherent\n> variable definition with an unintelligible mess.\n>\n\n+1000 from me on that.\n\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\n\nOn Sat, 7 Dec 2019 at 07:21, Tom Lane <tgl@sss.pgh.pa.us> wrote:Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n> I've just been looking at that. load_external_function() doesn't\n> actually do anything V1-ish with the value, it just looks up the symbol\n> using dlsym and returns it cast to a PGFunction. Is there any reason I\n> can't just use that and cast it again to the callback function type?\n\nTBH, I think this entire discussion has gone seriously off into the\nweeds. The original design where we just let a shared_preload_library\nfunction get into a hook is far superior to any of the overcomplicated\nkluges that are being discussed now. Something like this, for instance:\n\n>>> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\n\nmakes me positively ill. It introduces problems that we don't need,\nlike how to parse out the sub-parts of the string, and the\nquoting/escaping issues that will come along with that; while from\nthe user's perspective it replaces a simple and intellectually-coherent\nvariable definition with an unintelligible mess.+1000 from me on that. -- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise","msg_date":"Mon, 9 Dec 2019 10:22:12 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Sun, Dec 8, 2019 at 9:02 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n> > Well that pretty much brings us back to the patch as submitted :-)\n>\n> Yeah, pretty nearly. Taking a quick look over the v3 patch, my\n> only quibble is that it doesn't provide any convenient way for the\n> external module to make decisions about how to interact with\n> ssl_passphrase_command --- in particular, if it would like to allow\n> that to take precedence, it can't because there's no way for it to\n> invoke the static function ssl_external_passwd_cb.\n>\n> But rather than expose that globally, maybe the theory ought to be\n> \"set up the state as we'd normally do, then let loadable modules\n> choose to override it\". So I'm tempted to propose a hook function\n> with the signature\n>\n> void openssl_tls_init_hook(SSL_CTX *context, bool isServerStart);\n>\n> and invoke that somewhere in be_tls_init --- maybe fairly late,\n> so that it can override other settings if it wants, not only the\n> SSL_CTX_set_default_passwd_cb setting.\n>\n\n\nNot sure if the placement is what you want, but maybe something like this?\n\ncheers\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Wed, 22 Jan 2020 17:32:01 +1030","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Thu, Nov 14, 2019 at 8:54 AM Sehrope Sarkuni <sehrope@jackdb.com> wrote:\n> Has the idea of using environment variables (rather than command line\n> args) for external commands been brought up before? I couldn't find\n> anything in the mailing list archives.\n\nPassing data through environment variables isn't secure. Try 'ps -E'\non MacOS, or something like 'ps axe' on Linux.\n\nIf we want to pass data securely to child processes, the way to do it\nis via stdin. Data sent back and forth via file descriptors can't\neasily be snooped by other users on the system.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 23 Jan 2020 12:30:36 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Wed, Jan 22, 2020 at 8:02 PM Andrew Dunstan\n<andrew.dunstan@2ndquadrant.com> wrote:\n> Not sure if the placement is what you want, but maybe something like this?\n\nHi Andrew, FYI this failed here:\n\nt/001_testfunc.pl .. Bailout called. Further testing stopped: pg_ctl\nstart failed\nFAILED--Further testing stopped: pg_ctl start failed\nMakefile:23: recipe for target 'prove-check' failed\n\nUnfortunately my robot is poorly trained and does not dump any of the\ninteresting logs for this case, but it looks like it's failing that\nway every time.\n\nhttps://travis-ci.org/postgresql-cfbot/postgresql/builds/651756191\n\n\n","msg_date":"Tue, 18 Feb 2020 16:30:32 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Tue, Feb 18, 2020 at 2:01 PM Thomas Munro <thomas.munro@gmail.com> wrote:\n>\n> On Wed, Jan 22, 2020 at 8:02 PM Andrew Dunstan\n> <andrew.dunstan@2ndquadrant.com> wrote:\n> > Not sure if the placement is what you want, but maybe something like this?\n>\n> Hi Andrew, FYI this failed here:\n>\n> t/001_testfunc.pl .. Bailout called. Further testing stopped: pg_ctl\n> start failed\n> FAILED--Further testing stopped: pg_ctl start failed\n> Makefile:23: recipe for target 'prove-check' failed\n>\n> Unfortunately my robot is poorly trained and does not dump any of the\n> interesting logs for this case, but it looks like it's failing that\n> way every time.\n>\n> https://travis-ci.org/postgresql-cfbot/postgresql/builds/651756191\n\n\nThanks for letting me know, I will investigate.\n\ncheers\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Wed, 19 Feb 2020 07:10:20 +1030","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On Wed, Feb 19, 2020 at 7:10 AM Andrew Dunstan\n<andrew.dunstan@2ndquadrant.com> wrote:\n>\n> On Tue, Feb 18, 2020 at 2:01 PM Thomas Munro <thomas.munro@gmail.com> wrote:\n> >\n> > On Wed, Jan 22, 2020 at 8:02 PM Andrew Dunstan\n> > <andrew.dunstan@2ndquadrant.com> wrote:\n> > > Not sure if the placement is what you want, but maybe something like this?\n> >\n> > Hi Andrew, FYI this failed here:\n> >\n> > t/001_testfunc.pl .. Bailout called. Further testing stopped: pg_ctl\n> > start failed\n> > FAILED--Further testing stopped: pg_ctl start failed\n> > Makefile:23: recipe for target 'prove-check' failed\n> >\n> > Unfortunately my robot is poorly trained and does not dump any of the\n> > interesting logs for this case, but it looks like it's failing that\n> > way every time.\n> >\n> > https://travis-ci.org/postgresql-cfbot/postgresql/builds/651756191\n>\n>\n> Thanks for letting me know, I will investigate.\n>\n\n\nThis should fix the issue, it happened when I switched to using a\npre-generated cert/key.\n\ncheers\n\nandrew\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Wed, 19 Feb 2020 09:09:20 +1030","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On 2/18/20 11:39 PM, Andrew Dunstan wrote:\n> This should fix the issue, it happened when I switched to using a\n> pre-generated cert/key.\n\n# Review\n\nThe patch still applies and passes the test suite, both with openssl \nenabled and with it disabled.\n\nAs for the feature I agree that it is nice to expose this callback to \nextension writers and I agree with the approach taken. The other \nproposals up-thread seem over engineered to me. Maybe if it was a \ngeneral feature used in many places those proposals would be worth it, \nbut I am still skeptical even then. This approach is so much simpler.\n\nThe only real risk I see is that if people install multiple libraries \nfor this they will overwrite the hook for each other but we have other \ncases like that already so I think that is fine.\n\nThe patch moves secure_initialize() to after \nprocess_shared_preload_libraries() which in theory could break some \nextension but it does not seem like a likely thing for extensions to \nrely on. Or is it?\n\nAn idea would be to have the code in ssl_passphrase_func.c to warn if \nthe ssl_passphrase_command GUC is set to make it more useful as example \ncode for people implementing this hook.\n\n# Nitpicking\n\nThe certificate expires in 2030 while all other certificates used in \ntests expires in 2046. Should we be consistent?\n\nThere is text in server.crt and server.key, while other certificates and \nkeys used in the tests do not have this. Again, should we be consistent?\n\nEmpty first line in \nsrc/test/modules/ssl_passphrase_callback/t/001_testfunc.pl which should \nprobably just be removed or replaced with a shebang.\n\nThere is an extra space between the parentheses in the line below. Does \nthat follow our code style for Perl?\n\n+unless ( ($ENV{with_openssl} || 'no') eq 'yes')\n\nMissing space after comma in:\n\n+ok(-e \"$ddir/postmaster.pid\",\"postgres started\");\n\nMissing space after comma in:\n\n+ok(! -e \"$ddir/postmaster.pid\",\"postgres not started with bad passphrase\");\n\nAndreas\n\n\n","msg_date":"Mon, 16 Mar 2020 03:14:57 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"Hello Andrew,\r\n\r\nFrom: Andreas Karlsson <andreas@proxel.se>\r\n> # Nitpicking\r\n> \r\n> The certificate expires in 2030 while all other certificates used in\r\n> tests expires in 2046. Should we be consistent?\r\n> \r\n> There is text in server.crt and server.key, while other certificates and\r\n> keys used in the tests do not have this. Again, should we be consistent?\r\n> \r\n> Empty first line in\r\n> src/test/modules/ssl_passphrase_callback/t/001_testfunc.pl which should\r\n> probably just be removed or replaced with a shebang.\r\n> \r\n> There is an extra space between the parentheses in the line below. Does\r\n> that follow our code style for Perl?\r\n> \r\n> +unless ( ($ENV{with_openssl} || 'no') eq 'yes')\r\n> \r\n> Missing space after comma in:\r\n> \r\n> +ok(-e \"$ddir/postmaster.pid\",\"postgres started\");\r\n> \r\n> Missing space after comma in:\r\n> \r\n> +ok(! -e \"$ddir/postmaster.pid\",\"postgres not started with bad passphrase\");\r\n> \r\n> Andreas\r\n> \r\n\r\nTrailing space:\r\n\r\n220 + X509v3 Subject Key Identifier:\r\n222 + X509v3 Authority Key Identifier:\r\n\r\nMissing \"d\"(password?):\r\n\r\n121 +/* init hook for SSL, the default sets the passwor callback if appropriate */\r\n\r\nRegards,\r\n\r\n--\r\nTakanori Asaba\r\n\r\n\r\n","msg_date":"Thu, 19 Mar 2020 08:10:58 +0000","msg_from":"\"asaba.takanori@fujitsu.com\" <asaba.takanori@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: ssl passphrase callback"},{"msg_contents":"\nOn 3/15/20 10:14 PM, Andreas Karlsson wrote:\n> On 2/18/20 11:39 PM, Andrew Dunstan wrote:\n>> This should fix the issue, it happened when I switched to using a\n>> pre-generated cert/key.\n>\n> # Review\n>\n> The patch still applies and passes the test suite, both with openssl\n> enabled and with it disabled.\n>\n> As for the feature I agree that it is nice to expose this callback to\n> extension writers and I agree with the approach taken. The other\n> proposals up-thread seem over engineered to me. Maybe if it was a\n> general feature used in many places those proposals would be worth it,\n> but I am still skeptical even then. This approach is so much simpler.\n>\n> The only real risk I see is that if people install multiple libraries\n> for this they will overwrite the hook for each other but we have other\n> cases like that already so I think that is fine.\n\n\nRight, me too.\n\n\n>\n> The patch moves secure_initialize() to after\n> process_shared_preload_libraries() which in theory could break some\n> extension but it does not seem like a likely thing for extensions to\n> rely on. Or is it?\n\n\nI don't think so.\n\n\n>\n> An idea would be to have the code in ssl_passphrase_func.c to warn if\n> the ssl_passphrase_command GUC is set to make it more useful as\n> example code for people implementing this hook.\n\n\nI'll look at that. Should be possible.\n\n\n>\n> # Nitpicking\n>\n> The certificate expires in 2030 while all other certificates used in\n> tests expires in 2046. Should we be consistent?\n\n\nSure. will fix.\n\n\n>\n> There is text in server.crt and server.key, while other certificates\n> and keys used in the tests do not have this. Again, should we be\n> consistent?\n\n\nNot in server.key, but I will suppress it for the crt file.\n\n\n\n>\n> Empty first line in\n> src/test/modules/ssl_passphrase_callback/t/001_testfunc.pl which\n> should probably just be removed or replaced with a shebang.\n\n\nOK\n\n\n>\n> There is an extra space between the parentheses in the line below.\n> Does that follow our code style for Perl?\n>\n> +unless ( ($ENV{with_openssl} || 'no') eq 'yes')\n>\n> Missing space after comma in:\n>\n> +ok(-e \"$ddir/postmaster.pid\",\"postgres started\");\n>\n> Missing space after comma in:\n>\n> +ok(! -e \"$ddir/postmaster.pid\",\"postgres not started with bad\n> passphrase\");\n\n\nI'll make sure to run it through our perl indenter.\n\n\nThanks for the review.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Sat, 21 Mar 2020 09:15:53 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"\nOn 3/19/20 4:10 AM, asaba.takanori@fujitsu.com wrote:\n\n\n\n> Trailing space:\n>\n> 220 + X509v3 Subject Key Identifier:\n> 222 + X509v3 Authority Key Identifier:\n\n\nWe're going to remove all the text, so this becomes moot.\n\n\n>\n> Missing \"d\"(password?):\n>\n> 121 +/* init hook for SSL, the default sets the passwor callback if appropriate */\n>\n\nWill fix, thanks.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Sat, 21 Mar 2020 09:18:26 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On 3/21/20 9:18 AM, Andrew Dunstan wrote:\n> On 3/19/20 4:10 AM, asaba.takanori@fujitsu.com wrote:\n>\n>\n>\n>> Trailing space:\n>>\n>> 220 + X509v3 Subject Key Identifier:\n>> 222 + X509v3 Authority Key Identifier:\n>\n> We're going to remove all the text, so this becomes moot.\n>\n>\n>> Missing \"d\"(password?):\n>>\n>> 121 +/* init hook for SSL, the default sets the passwor callback if appropriate */\n>>\n> Will fix, thanks.\n>\n>\n\n\nLatest patch attached, I think all comments have been addressed. I\npropose to push this later this coming week if there are no more comments.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Sat, 21 Mar 2020 20:08:19 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"On 3/22/20 1:08 AM, Andrew Dunstan wrote:\n> Latest patch attached, I think all comments have been addressed. I\n> propose to push this later this coming week if there are no more comments.\n\nI do not have any objections.\n\nAndreas\n\n\n\n","msg_date":"Mon, 23 Mar 2020 01:58:14 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"Andreas Karlsson <andreas@proxel.se> writes:\n> On 3/22/20 1:08 AM, Andrew Dunstan wrote:\n>> Latest patch attached, I think all comments have been addressed. I\n>> propose to push this later this coming week if there are no more comments.\n\n> I do not have any objections.\n\nThis CF entry is still open, should it not be closed as committed?\n\nhttps://commitfest.postgresql.org/27/2338/\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 28 Mar 2020 13:45:00 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ssl passphrase callback"},{"msg_contents":"\nOn 3/28/20 1:45 PM, Tom Lane wrote:\n> Andreas Karlsson <andreas@proxel.se> writes:\n>> On 3/22/20 1:08 AM, Andrew Dunstan wrote:\n>>> Latest patch attached, I think all comments have been addressed. I\n>>> propose to push this later this coming week if there are no more comments.\n>> I do not have any objections.\n> This CF entry is still open, should it not be closed as committed?\n>\n> https://commitfest.postgresql.org/27/2338/\n>\n> \t\t\t\n\n\nDone, thanks for the reminder.\n\n\ncheers\n\n\nandrew\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Sat, 28 Mar 2020 15:35:07 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: ssl passphrase callback"}],"string":"[\n {\n \"msg_contents\": \"This is the first of a number of patches to enhance SSL functionality,\\nparticularly w.r.t. passphrases.\\n\\n\\nThis patch provides a hook for a function that can supply an SSL\\npassphrase. The hook can be filled in by a shared preloadable module. In\\norder for that to be effective, the startup order is modified slightly.\\nThere is a test attached that builds and uses one trivial\\nimplementation, which just takes a configuration setting and rot13's it\\nbefore supplying the result as the passphrase.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 31 Oct 2019 11:37:04 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\\n<andrew.dunstan@2ndquadrant.com> wrote:\\n> This patch provides a hook for a function that can supply an SSL\\n> passphrase. The hook can be filled in by a shared preloadable module. In\\n> order for that to be effective, the startup order is modified slightly.\\n> There is a test attached that builds and uses one trivial\\n> implementation, which just takes a configuration setting and rot13's it\\n> before supplying the result as the passphrase.\\n\\nIt seems to me that it would be a lot better to have an example in\\ncontrib that does something which might be of actual use to users,\\nsuch as running a shell command and reading the passphrase from\\nstdout.\\n\\nFeatures that are only accessible by writing C code are, in general,\\nnot as desirable as features which can be accessed via SQL or\\nconfiguration.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 11:01:15 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"\\nOn 11/1/19 11:01 AM, Robert Haas wrote:\\n> On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\\n> <andrew.dunstan@2ndquadrant.com> wrote:\\n>> This patch provides a hook for a function that can supply an SSL\\n>> passphrase. The hook can be filled in by a shared preloadable module. In\\n>> order for that to be effective, the startup order is modified slightly.\\n>> There is a test attached that builds and uses one trivial\\n>> implementation, which just takes a configuration setting and rot13's it\\n>> before supplying the result as the passphrase.\\n> It seems to me that it would be a lot better to have an example in\\n> contrib that does something which might be of actual use to users,\\n> such as running a shell command and reading the passphrase from\\n> stdout.\\n>\\n> Features that are only accessible by writing C code are, in general,\\n> not as desirable as features which can be accessed via SQL or\\n> configuration.\\n>\\n\\n\\nWell, I tried to provide the most trivial and simple test I could come\\nup with. Running a shell command can already be accomplished via the\\nssl_passphrase_command setting.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 13:57:29 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 2, 2019 at 6:57 AM Andrew Dunstan\\n<andrew.dunstan@2ndquadrant.com> wrote:\\n> On 11/1/19 11:01 AM, Robert Haas wrote:\\n> > On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\\n> > <andrew.dunstan@2ndquadrant.com> wrote:\\n> >> This patch provides a hook for a function that can supply an SSL\\n> >> passphrase. The hook can be filled in by a shared preloadable module. In\\n> >> order for that to be effective, the startup order is modified slightly.\\n> >> There is a test attached that builds and uses one trivial\\n> >> implementation, which just takes a configuration setting and rot13's it\\n> >> before supplying the result as the passphrase.\\n> > It seems to me that it would be a lot better to have an example in\\n> > contrib that does something which might be of actual use to users,\\n> > such as running a shell command and reading the passphrase from\\n> > stdout.\\n> >\\n> > Features that are only accessible by writing C code are, in general,\\n> > not as desirable as features which can be accessed via SQL or\\n> > configuration.\\n>\\n> Well, I tried to provide the most trivial and simple test I could come\\n> up with. Running a shell command can already be accomplished via the\\n> ssl_passphrase_command setting.\\n\\nIt looks like the new declarations in libpq-be.h are ifdef'd out in a\\nnon-USE_SSL build, but then we still try to build the new test module\\nand it fails:\\n\\nhttps://ci.appveyor.com/project/postgresql-cfbot/postgresql/build/1.0.64071\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Nov 2019 10:43:45 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 01:57:29PM -0400, Andrew Dunstan wrote:\\n> \\n> On 11/1/19 11:01 AM, Robert Haas wrote:\\n> > On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\\n> > <andrew.dunstan@2ndquadrant.com> wrote:\\n> >> This patch provides a hook for a function that can supply an SSL\\n> >> passphrase. The hook can be filled in by a shared preloadable module. In\\n> >> order for that to be effective, the startup order is modified slightly.\\n> >> There is a test attached that builds and uses one trivial\\n> >> implementation, which just takes a configuration setting and rot13's it\\n> >> before supplying the result as the passphrase.\\n> > It seems to me that it would be a lot better to have an example in\\n> > contrib that does something which might be of actual use to users,\\n> > such as running a shell command and reading the passphrase from\\n> > stdout.\\n> >\\n> > Features that are only accessible by writing C code are, in general,\\n> > not as desirable as features which can be accessed via SQL or\\n> > configuration.\\n> >\\n> \\n> \\n> Well, I tried to provide the most trivial and simple test I could come\\n> up with. Running a shell command can already be accomplished via the\\n> ssl_passphrase_command setting.\\n\\nWhat is the value of a shared library over a shell command? We had this\\ndiscussion in relation to archive_command years ago, and decided on a\\nshell command as the best API.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Nov 2019 20:23:56 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On 11/4/19 4:43 PM, Thomas Munro wrote:\\n>\\n> It looks like the new declarations in libpq-be.h are ifdef'd out in a\\n> non-USE_SSL build, but then we still try to build the new test module\\n> and it fails:\\n>\\n> https://ci.appveyor.com/project/postgresql-cfbot/postgresql/build/1.0.64071\\n\\n\\n\\nI think this updated patch should fix things.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 7 Nov 2019 12:30:50 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Thu, 7 Nov 2019 at 10:24, Bruce Momjian <bruce@momjian.us> wrote:\\n\\n\\n> What is the value of a shared library over a shell command? We had this\\n> discussion in relation to archive_command years ago, and decided on a\\n> shell command as the best API.\\n>\\n\\nI don't recall such a discussion, but I can give perspective:\\n\\n* shell command offered the widest and simplest API for integration, which\\nwas the most important consideration for a backup API. That choice caused\\ndifficulty with the need to pass information to the external command, e.g.\\n%f %p\\n\\n* shared library is more appropriate for a security-related module, so\\nusers can't see how it is configured, as well as being more\\ntightly integrated so it can be better tailored to various uses\\n\\nSummary is that the choice is not random, nor mere preference\\n\\n-- \\nSimon Riggs http://www.2ndQuadrant.com/\\n<http://www.2ndquadrant.com/>\\nPostgreSQL Solutions for the Enterprise\\n\\nOn Thu, 7 Nov 2019 at 10:24, Bruce Momjian <bruce@momjian.us> wrote: What is the value of a shared library over a shell command? We had this\\ndiscussion in relation to archive_command years ago, and decided on a\\nshell command as the best API.I don't recall such a discussion, but I can give perspective:* shell command offered the widest and simplest API for integration, which was the most important consideration for a backup API. That choice caused difficulty with the need to pass information to the external command, e.g. %f %p* shared library is more appropriate for a security-related module, so users can't see how it is configured, as well as being more tightly integrated so it can be better tailored to various usesSummary is that the choice is not random, nor mere preference-- Simon Riggs http://www.2ndQuadrant.com/PostgreSQL Solutions for the Enterprise\",\n \"msg_date\": \"Fri, 8 Nov 2019 23:12:08 +0900\",\n \"msg_from\": \"Simon Riggs <simon@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 08, 2019 at 11:12:08PM +0900, Simon Riggs wrote:\\n>On Thu, 7 Nov 2019 at 10:24, Bruce Momjian <bruce@momjian.us> wrote:\\n>\\n>\\n>> What is the value of a shared library over a shell command? We had\\n>> this discussion in relation to archive_command years ago, and decided\\n>> on a shell command as the best API.\\n>>\\n>\\n>I don't recall such a discussion, but I can give perspective:\\n>\\n>* shell command offered the widest and simplest API for integration,\\n>which was the most important consideration for a backup API. That\\n>choice caused difficulty with the need to pass information to the\\n>external command, e.g. %f %p\\n>\\n\\nIt's not clear to me why simple API for integration would be less\\nvaluable for this feature. Also, I'm sure passing data to/from shell\\ncommand may be tricky, but presumably we have figured how to do that.\\n\\n>* shared library is more appropriate for a security-related module, so\\n>users can't see how it is configured, as well as being more\\n>tightly integrated so it can be better tailored to various uses\\n>\\n\\nI don't follow. Why would there be a significant difference between a\\nshell command/script and shared library in this respect? If you don't\\nwant the users to see the config, just store it in a separate file and\\nit's about the same as storing it in the .so library.\\n\\nIs there something that can be done with a .so library but can't be done\\nwith a shell command (which may just call a binary, with all the config\\nincluded, making it equal to the .so solution)?\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Sat, 9 Nov 2019 12:52:25 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\\n\\n> On Fri, Nov 1, 2019 at 01:57:29PM -0400, Andrew Dunstan wrote:\\n> >\\n> > On 11/1/19 11:01 AM, Robert Haas wrote:\\n> > > On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\\n> > > <andrew.dunstan@2ndquadrant.com> wrote:\\n> > >> This patch provides a hook for a function that can supply an SSL\\n> > >> passphrase. The hook can be filled in by a shared preloadable module.\\n> In\\n> > >> order for that to be effective, the startup order is modified\\n> slightly.\\n> > >> There is a test attached that builds and uses one trivial\\n> > >> implementation, which just takes a configuration setting and rot13's\\n> it\\n> > >> before supplying the result as the passphrase.\\n> > > It seems to me that it would be a lot better to have an example in\\n> > > contrib that does something which might be of actual use to users,\\n> > > such as running a shell command and reading the passphrase from\\n> > > stdout.\\n> > >\\n> > > Features that are only accessible by writing C code are, in general,\\n> > > not as desirable as features which can be accessed via SQL or\\n> > > configuration.\\n> > >\\n> >\\n> >\\n> > Well, I tried to provide the most trivial and simple test I could come\\n> > up with. Running a shell command can already be accomplished via the\\n> > ssl_passphrase_command setting.\\n>\\n> What is the value of a shared library over a shell command?\\n\\n\\nFor one, platforms where shell commands are a lot less convenient, such as\\nWindows.\\n\\n\\n\\n> We had this\\n> discussion in relation to archive_command years ago, and decided on a\\n> shell command as the best API.\\n>\\n>\\nI don't recall that from back then, but that was a long time ago.\\n\\nBut it's interesting that you mention it, given the number of people I have\\nbeen discussing that with recently, under the topic of changing it from a\\nshell command into a shared library API (with there being a shell command\\nas one possible implementation of course).\\n\\nOne of the main reasons there being to be easily able to transfer more\\nstate and give results other than just an exit code, no need to deal with\\nparameter escaping etc. Which probably wouldn't matter as much to an SSL\\npassphrase command, but still.\\n\\n//Magnus\\n\\nOn Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:On Fri, Nov 1, 2019 at 01:57:29PM -0400, Andrew Dunstan wrote:\\n> \\n> On 11/1/19 11:01 AM, Robert Haas wrote:\\n> > On Thu, Oct 31, 2019 at 11:37 AM Andrew Dunstan\\n> > <andrew.dunstan@2ndquadrant.com> wrote:\\n> >> This patch provides a hook for a function that can supply an SSL\\n> >> passphrase. The hook can be filled in by a shared preloadable module. In\\n> >> order for that to be effective, the startup order is modified slightly.\\n> >> There is a test attached that builds and uses one trivial\\n> >> implementation, which just takes a configuration setting and rot13's it\\n> >> before supplying the result as the passphrase.\\n> > It seems to me that it would be a lot better to have an example in\\n> > contrib that does something which might be of actual use to users,\\n> > such as running a shell command and reading the passphrase from\\n> > stdout.\\n> >\\n> > Features that are only accessible by writing C code are, in general,\\n> > not as desirable as features which can be accessed via SQL or\\n> > configuration.\\n> >\\n> \\n> \\n> Well, I tried to provide the most trivial and simple test I could come\\n> up with. Running a shell command can already be accomplished via the\\n> ssl_passphrase_command setting.\\n\\nWhat is the value of a shared library over a shell command?For one, platforms where shell commands are a lot less convenient, such as Windows. We had this\\ndiscussion in relation to archive_command years ago, and decided on a\\nshell command as the best API.I don't recall that from back then, but that was a long time ago.But it's interesting that you mention it, given the number of people I have been discussing that with recently, under the topic of changing it from a shell command into a shared library API (with there being a shell command as one possible implementation of course). One of the main reasons there being to be easily able to transfer more state and give results other than just an exit code, no need to deal with parameter escaping etc. Which probably wouldn't matter as much to an SSL passphrase command, but still.//Magnus\",\n \"msg_date\": \"Sun, 10 Nov 2019 13:01:17 -0600\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On 11/7/19 12:30 PM, Andrew Dunstan wrote:\\n> On 11/4/19 4:43 PM, Thomas Munro wrote:\\n>> It looks like the new declarations in libpq-be.h are ifdef'd out in a\\n>> non-USE_SSL build, but then we still try to build the new test module\\n>> and it fails:\\n>>\\n>> https://ci.appveyor.com/project/postgresql-cfbot/postgresql/build/1.0.64071\\n>\\n>\\n> I think this updated patch should fix things.\\n>\\n>\\n\\n\\n\\nThis time with a typo fixed to keep the cfbot happy.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Mon, 11 Nov 2019 14:19:44 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\\n> On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\\n> � We had this\\n> discussion in relation to archive_command years ago, and decided on a\\n> shell command as the best API.\\n>\\n> I don't recall that from back then, but that was a long time ago.\\n> \\n> But it's interesting that you mention it, given the number of people I have\\n> been discussing that with recently, under the topic of changing it from a shell\\n> command into a shared library API (with there being a shell command as one\\n> possible implementation of course).�\\n> \\n> One of the main reasons there being to be easily able to transfer more state\\n> and give results other than just an exit code, no need to deal with parameter\\n> escaping etc. Which probably wouldn't matter as much to an SSL passphrase\\n> command, but still.\\n\\nI get the callback-is-easier issue with shared objects, but are we\\nexpecting to pass in more information here than we do for\\narchive_command? I would think not. What I am saying is that if we\\ndon't think passing things in works, we should fix all these external\\ncommands, or something. I don't see why ssl_passphrase_command is\\ndifferent, except that it is new. Or is it related to _securely_\\npassing something?\\n\\nAlso, why was this patch posted without any discussion of these issues?\\nShouldn't we ideally discuss the API first?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Nov 2019 21:51:33 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\\n> On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\\n> > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\\n> > � We had this\\n> > discussion in relation to archive_command years ago, and decided on a\\n> > shell command as the best API.\\n> >\\n> > I don't recall that from back then, but that was a long time ago.\\n> > \\n> > But it's interesting that you mention it, given the number of people I have\\n> > been discussing that with recently, under the topic of changing it from a shell\\n> > command into a shared library API (with there being a shell command as one\\n> > possible implementation of course).�\\n> > \\n> > One of the main reasons there being to be easily able to transfer more state\\n> > and give results other than just an exit code, no need to deal with parameter\\n> > escaping etc. Which probably wouldn't matter as much to an SSL passphrase\\n> > command, but still.\\n> \\n> I get the callback-is-easier issue with shared objects, but are we\\n> expecting to pass in more information here than we do for\\n> archive_command? I would think not. What I am saying is that if we\\n> don't think passing things in works, we should fix all these external\\n> commands, or something. I don't see why ssl_passphrase_command is\\n> different, except that it is new. Or is it related to _securely_\\n> passing something?\\n> \\n> Also, why was this patch posted without any discussion of these issues?\\n> Shouldn't we ideally discuss the API first?\\n\\nI wonder if every GUC that takes an OS command should allow a shared\\nobject to be specified --- maybe control that if the command string\\nstarts with a # or something.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Nov 2019 08:08:23 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Wed, 13 Nov 2019 at 13:08, Bruce Momjian <bruce@momjian.us> wrote:\\n\\n> On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\\n> > On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\\n> > > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\\n>\\n> > > One of the main reasons there being to be easily able to transfer more\\n> state\\n> > > and give results other than just an exit code, no need to deal with\\n> parameter\\n> > > escaping etc. Which probably wouldn't matter as much to an SSL\\n> passphrase\\n> > > command, but still.\\n> >\\n> > I get the callback-is-easier issue with shared objects, but are we\\n> > expecting to pass in more information here than we do for\\n> > archive_command? I would think not. What I am saying is that if we\\n> > don't think passing things in works, we should fix all these external\\n> > commands, or something. I don't see why ssl_passphrase_command is\\n> > different, except that it is new.\\n\\n\\n\\n> Or is it related to _securely_passing something?\\n>\\n\\nYes\\n\\n\\n> > Also, why was this patch posted without any discussion of these issues?\\n> > Shouldn't we ideally discuss the API first?\\n>\\n> I wonder if every GUC that takes an OS command should allow a shared\\n> object to be specified --- maybe control that if the command string\\n> starts with a # or something.\\n>\\n\\nVery good idea\\n\\n-- \\nSimon Riggs http://www.2ndQuadrant.com/\\n<http://www.2ndquadrant.com/>\\nPostgreSQL Solutions for the Enterprise\\n\\nOn Wed, 13 Nov 2019 at 13:08, Bruce Momjian <bruce@momjian.us> wrote:On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\\n> On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\\n> > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\\n> > One of the main reasons there being to be easily able to transfer more state\\n> > and give results other than just an exit code, no need to deal with parameter\\n> > escaping etc. Which probably wouldn't matter as much to an SSL passphrase\\n> > command, but still.\\n> \\n> I get the callback-is-easier issue with shared objects, but are we\\n> expecting to pass in more information here than we do for\\n> archive_command? I would think not. What I am saying is that if we\\n> don't think passing things in works, we should fix all these external\\n> commands, or something. I don't see why ssl_passphrase_command is\\n> different, except that it is new. Or is it related to _securely_passing something?Yes \\n> Also, why was this patch posted without any discussion of these issues?\\n> Shouldn't we ideally discuss the API first?\\n\\nI wonder if every GUC that takes an OS command should allow a shared\\nobject to be specified --- maybe control that if the command string\\nstarts with a # or something.Very good idea -- Simon Riggs http://www.2ndQuadrant.com/PostgreSQL Solutions for the Enterprise\",\n \"msg_date\": \"Wed, 13 Nov 2019 13:20:43 +0000\",\n \"msg_from\": \"Simon Riggs <simon@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"\\nOn 11/13/19 8:08 AM, Bruce Momjian wrote:\\n>\\n>>\\n>> Also, why was this patch posted without any discussion of these issues?\\n>> Shouldn't we ideally discuss the API first?\\n\\n\\nThis is a very tiny patch. I don't think it's unusual to post such\\nthings without prior discussion. I would agree with your point if it\\nwere thousands of lines instead of 20 or so lines of core code.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Nov 2019 14:48:01 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 13, 2019 at 01:20:43PM +0000, Simon Riggs wrote:\\n>On Wed, 13 Nov 2019 at 13:08, Bruce Momjian <bruce@momjian.us> wrote:\\n>\\n>> On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\\n>> > On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\\n>> > > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\\n>>\\n>> > > One of the main reasons there being to be easily able to transfer more\\n>> state\\n>> > > and give results other than just an exit code, no need to deal with\\n>> parameter\\n>> > > escaping etc. Which probably wouldn't matter as much to an SSL\\n>> passphrase\\n>> > > command, but still.\\n>> >\\n>> > I get the callback-is-easier issue with shared objects, but are we\\n>> > expecting to pass in more information here than we do for\\n>> > archive_command? I would think not. What I am saying is that if we\\n>> > don't think passing things in works, we should fix all these external\\n>> > commands, or something. I don't see why ssl_passphrase_command is\\n>> > different, except that it is new.\\n>\\n>\\n>\\n>> Or is it related to _securely_passing something?\\n>>\\n>\\n>Yes\\n>\\n\\nI think it would be beneficial to explain why shared object is more\\nsecure than an OS command. Perhaps it's common knowledge, but it's not\\nquite obvious to me.\\n\\n>\\n>> > Also, why was this patch posted without any discussion of these issues?\\n>> > Shouldn't we ideally discuss the API first?\\n>>\\n>> I wonder if every GUC that takes an OS command should allow a shared\\n>> object to be specified --- maybe control that if the command string\\n>> starts with a # or something.\\n>>\\n>\\n>Very good idea\\n>\\n\\nIf it's about securely passing sensitive information (i.e. passphrase)\\nas was suggested, then I think that only applies to fairly small number\\nof GUCs.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Nov 2019 21:23:23 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 13, 2019 at 02:48:01PM -0500, Andrew Dunstan wrote:\\n>\\n>On 11/13/19 8:08 AM, Bruce Momjian wrote:\\n>>\\n>>>\\n>>> Also, why was this patch posted without any discussion of these issues?\\n>>> Shouldn't we ideally discuss the API first?\\n>\\n>\\n>This is a very tiny patch. I don't think it's unusual to post such\\n>things without prior discussion. I would agree with your point if it\\n>were thousands of lines instead of 20 or so lines of core code.\\n>\\n\\nNot sure that's really true. I think patches should provide some basic\\nexplanation why the feature is desirable, no matter the number of lines.\\n\\nAlso, there were vague references to issues with passing parameters to\\narchive_command. A link to details, past discussion, or brief\\nexplanation would be nice.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Nov 2019 21:34:38 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\\nwrote:\\n\\n> On Wed, Nov 13, 2019 at 01:20:43PM +0000, Simon Riggs wrote:\\n> >On Wed, 13 Nov 2019 at 13:08, Bruce Momjian <bruce@momjian.us> wrote:\\n> >\\n> >> On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\\n> >> > On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\\n> >> > > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us>\\n> wrote:\\n> >>\\n> >> > > One of the main reasons there being to be easily able to transfer\\n> more\\n> >> state\\n> >> > > and give results other than just an exit code, no need to deal with\\n> >> parameter\\n> >> > > escaping etc. Which probably wouldn't matter as much to an SSL\\n> >> passphrase\\n> >> > > command, but still.\\n> >> >\\n> >> > I get the callback-is-easier issue with shared objects, but are we\\n> >> > expecting to pass in more information here than we do for\\n> >> > archive_command? I would think not. What I am saying is that if we\\n> >> > don't think passing things in works, we should fix all these external\\n> >> > commands, or something. I don't see why ssl_passphrase_command is\\n> >> > different, except that it is new.\\n> >\\n> >\\n> >\\n> >> Or is it related to _securely_passing something?\\n> >>\\n> >\\n> >Yes\\n> >\\n>\\n> I think it would be beneficial to explain why shared object is more\\n> secure than an OS command. Perhaps it's common knowledge, but it's not\\n> quite obvious to me.\\n>\\n\\nYeah, that probably wouldn't hurt. It's also securely passing from more\\nthan one perspective -- both from the \\\"cannot be eavesdropped\\\" (like\\nputting the password on the commandline for example) and the requirement\\nfor escaping.\\n\\n\\n>\\n> >\\n> >> > Also, why was this patch posted without any discussion of these\\n> issues?\\n> >> > Shouldn't we ideally discuss the API first?\\n> >>\\n> >> I wonder if every GUC that takes an OS command should allow a shared\\n> >> object to be specified --- maybe control that if the command string\\n> >> starts with a # or something.\\n> >>\\n> >\\n> >Very good idea\\n> >\\n>\\n> If it's about securely passing sensitive information (i.e. passphrase)\\n> as was suggested, then I think that only applies to fairly small number\\n> of GUCs.\\n>\\n\\nThere aren't exactly a large number of GUCs that take OS commands in total.\\nConsistency itself certainly has some value.\\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 Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:On Wed, Nov 13, 2019 at 01:20:43PM +0000, Simon Riggs wrote:\\n>On Wed, 13 Nov 2019 at 13:08, Bruce Momjian <bruce@momjian.us> wrote:\\n>\\n>> On Tue, Nov 12, 2019 at 09:51:33PM -0500, Bruce Momjian wrote:\\n>> > On Sun, Nov 10, 2019 at 01:01:17PM -0600, Magnus Hagander wrote:\\n>> > > On Wed, Nov 6, 2019 at 7:24 PM Bruce Momjian <bruce@momjian.us> wrote:\\n>>\\n>> > > One of the main reasons there being to be easily able to transfer more\\n>> state\\n>> > > and give results other than just an exit code, no need to deal with\\n>> parameter\\n>> > > escaping etc. Which probably wouldn't matter as much to an SSL\\n>> passphrase\\n>> > > command, but still.\\n>> >\\n>> > I get the callback-is-easier issue with shared objects, but are we\\n>> > expecting to pass in more information here than we do for\\n>> > archive_command? I would think not. What I am saying is that if we\\n>> > don't think passing things in works, we should fix all these external\\n>> > commands, or something. I don't see why ssl_passphrase_command is\\n>> > different, except that it is new.\\n>\\n>\\n>\\n>> Or is it related to _securely_passing something?\\n>>\\n>\\n>Yes\\n>\\n\\nI think it would be beneficial to explain why shared object is more\\nsecure than an OS command. Perhaps it's common knowledge, but it's not\\nquite obvious to me.Yeah, that probably wouldn't hurt. It's also securely passing from more than one perspective -- both from the \\\"cannot be eavesdropped\\\" (like putting the password on the commandline for example) and the requirement for escaping. \\n\\n>\\n>> > Also, why was this patch posted without any discussion of these issues?\\n>> > Shouldn't we ideally discuss the API first?\\n>>\\n>> I wonder if every GUC that takes an OS command should allow a shared\\n>> object to be specified --- maybe control that if the command string\\n>> starts with a # or something.\\n>>\\n>\\n>Very good idea\\n>\\n\\nIf it's about securely passing sensitive information (i.e. passphrase)\\nas was suggested, then I think that only applies to fairly small number\\nof GUCs.There aren't exactly a large number of GUCs that take OS commands in total. Consistency itself certainly has some value. -- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/\",\n \"msg_date\": \"Thu, 14 Nov 2019 11:42:05 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 13, 2019 at 3:23 PM Tomas Vondra\\n<tomas.vondra@2ndquadrant.com> wrote:\\n> I think it would be beneficial to explain why shared object is more\\n> secure than an OS command. Perhaps it's common knowledge, but it's not\\n> quite obvious to me.\\n\\nExternal command args can be viewed by other OS users (not just the\\npostgres user). For non-sensitive arguments (ex: WAL filename?) that's\\nnot an issue but if you plan on passing in something potentially\\nsecret value from the backend to the external OS command, that value\\nwould be exposed:\\n\\nTerminal 1 (run a command as some other user):\\n$ sudo -u nobody sleep 5\\n\\nTerminal 2 (view command args as a different non-super user):\\n$ ps -u nobody -o args\\nCOMMAND\\nsleep 5\\n\\nA shared library would not have this problem as the backend directly\\nexecutes the library in the same process.\\n\\nHas the idea of using environment variables (rather than command line\\nargs) for external commands been brought up before? I couldn't find\\nanything in the mailing list archives.\\n\\nEnvironment variables have the advantage of only being readable by the\\nprocess owner and super user. They also naturally have a \\\"name\\\" and do\\nnot have escaping or quoting issues.\\n\\nFor example, archive_command %p could be exposed as\\nPG_ARG_ARCHIVE_PATH and %f could be exposed as\\nPG_ARG_ARCHIVE_FILENAME. Then you could have a script use them via:\\n\\n#!/usr/bin/env bash\\nset -euo pipefail\\nmain () {\\n local archive_dir=\\\"/path/to/archive_dir/\\\"\\n local archive_file=\\\"${archive_dir}${PG_ARG_ARCHIVE_FILENAME}\\\"\\n test ! -f \\\"${archive_file}\\\" && cp -- \\\"${PG_ARG_ARCHIVE_PATH}\\\"\\n\\\"${archive_file}\\\"\\n}\\nmain \\\"$@\\\"\\n\\nIt's not particularly useful for that basic archive case but if\\nthere's something like PG_ARG_SUPER_SECRET then the executed command\\ncould receive that value without it being exposed. That'd be useful\\nfor something like a callout to an external KMS (key management\\nsystem).\\n\\nNothing stops something like this with being used in tandem with\\nstring substitution to create the full commands. That'd give backwards\\ncompatibility too. The main limitation compared to a shared library is\\nthat you'd still have to explicitly pick and name the exposed argument\\nenvironment variables (i.e. like picking the set of %? substitutions).\\nIf there's a generic shared-library-for-external-commands approach\\nthen there could be a built-in library that provides this type of\\n\\\"expose as env vars\\\" functionality.\\n\\nRegards,\\n-- Sehrope Sarkuni\\nFounder & CEO | JackDB, Inc. | https://www.jackdb.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Nov 2019 08:54:27 -0500\",\n \"msg_from\": \"Sehrope Sarkuni <sehrope@jackdb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\\n> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\\n> I think it would be beneficial to explain why shared object is more\\n> secure than an OS command. Perhaps it's common knowledge, but it's not\\n> quite obvious to me.\\n> \\n> \\n> Yeah, that probably wouldn't hurt. It's also securely passing from more than\\n> one perspective -- both from the \\\"cannot be eavesdropped\\\" (like putting the\\n> password on the commandline for example) and the requirement for escaping.\\n\\nI think a bigger issue is that if you want to give people the option of\\nusing a shell command or a shared object, and if you use two commands to\\ncontrol it, it isn't clear what happens if both are defined. By using\\nsome character prefix to control if a shared object is used, you can use\\na single variable and there is no confusion over having two variables\\nand their conflicting behavior.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Nov 2019 11:07:52 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"\\nOn 11/14/19 11:07 AM, Bruce Momjian wrote:\\n> On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\\n>> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\\n>> I think it would be beneficial to explain why shared object is more\\n>> secure than an OS command. Perhaps it's common knowledge, but it's not\\n>> quite obvious to me.\\n>>\\n>>\\n>> Yeah, that probably wouldn't hurt. It's also securely passing from more than\\n>> one perspective -- both from the \\\"cannot be eavesdropped\\\" (like putting the\\n>> password on the commandline for example) and the requirement for escaping.\\n> I think a bigger issue is that if you want to give people the option of\\n> using a shell command or a shared object, and if you use two commands to\\n> control it, it isn't clear what happens if both are defined. By using\\n> some character prefix to control if a shared object is used, you can use\\n> a single variable and there is no confusion over having two variables\\n> and their conflicting behavior.\\n>\\n\\n\\nI'm not sure how that would work in the present instance. The shared\\npreloaded module installs a function and defines the params it wants. If\\nwe somehow unify the params with ssl_passphrase_command that could look\\nicky, and the module would have to parse the settings string. That's not\\na problem for the sample module which only needs one param, but it will\\nbe for other more complex implementations.\\n\\nI'm quite open to suggestions, but I want things to be tolerably clean.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Nov 2019 11:34:24 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 14, 2019 at 11:34:24AM -0500, Andrew Dunstan wrote:\\n>\\n>On 11/14/19 11:07 AM, Bruce Momjian wrote:\\n>> On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\\n>>> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\\n>>> I think it would be beneficial to explain why shared object is more\\n>>> secure than an OS command. Perhaps it's common knowledge, but it's not\\n>>> quite obvious to me.\\n>>>\\n>>>\\n>>> Yeah, that probably wouldn't hurt. It's also securely passing from more than\\n>>> one perspective -- both from the \\\"cannot be eavesdropped\\\" (like putting the\\n>>> password on the commandline for example) and the requirement for escaping.\\n>> I think a bigger issue is that if you want to give people the option of\\n>> using a shell command or a shared object, and if you use two commands to\\n>> control it, it isn't clear what happens if both are defined. By using\\n>> some character prefix to control if a shared object is used, you can use\\n>> a single variable and there is no confusion over having two variables\\n>> and their conflicting behavior.\\n>>\\n>\\n>\\n>I'm� not sure how that would work in the present instance. The shared\\n>preloaded module installs a function and defines the params it wants. If\\n>we somehow unify the params with ssl_passphrase_command that could look\\n>icky, and the module would have to parse the settings string. That's not\\n>a problem for the sample module which only needs one param, but it will\\n>be for other more complex implementations.\\n>\\n>I'm quite open to suggestions, but I want things to be tolerably clean.\\n>\\n\\nI agree it's better to have two separate GUCs - one for command, one for\\nshared object, and documented order of precedence. I suppose we may log\\na warning when both are specified, or perhaps \\\"reset\\\" the value with\\nlower order of precedence.\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Nov 2019 17:52:50 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 14, 2019 at 11:34:24AM -0500, Andrew Dunstan wrote:\\n> \\n> On 11/14/19 11:07 AM, Bruce Momjian wrote:\\n> > On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\\n> >> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\\n> >> I think it would be beneficial to explain why shared object is more\\n> >> secure than an OS command. Perhaps it's common knowledge, but it's not\\n> >> quite obvious to me.\\n> >>\\n> >>\\n> >> Yeah, that probably wouldn't hurt. It's also securely passing from more than\\n> >> one perspective -- both from the \\\"cannot be eavesdropped\\\" (like putting the\\n> >> password on the commandline for example) and the requirement for escaping.\\n> > I think a bigger issue is that if you want to give people the option of\\n> > using a shell command or a shared object, and if you use two commands to\\n> > control it, it isn't clear what happens if both are defined. By using\\n> > some character prefix to control if a shared object is used, you can use\\n> > a single variable and there is no confusion over having two variables\\n> > and their conflicting behavior.\\n> >\\n> \\n> \\n> I'm� not sure how that would work in the present instance. The shared\\n> preloaded module installs a function and defines the params it wants. If\\n> we somehow unify the params with ssl_passphrase_command that could look\\n> icky, and the module would have to parse the settings string. That's not\\n> a problem for the sample module which only needs one param, but it will\\n> be for other more complex implementations.\\n> \\n> I'm quite open to suggestions, but I want things to be tolerably clean.\\n\\nI was assuming if the variable starts with a #, it is a shared object,\\nif not, it is a shell command:\\n\\n\\tssl_passphrase_command='#/lib/x.so'\\n\\tssl_passphrase_command='my_command a b c'\\n\\nCan you show what you are talking about?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Nov 2019 11:53:09 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On 2019-Nov-14, Bruce Momjian wrote:\\n\\n> I was assuming if the variable starts with a #, it is a shared object,\\n> if not, it is a shell command:\\n> \\n> \\tssl_passphrase_command='#/lib/x.so'\\n> \\tssl_passphrase_command='my_command a b c'\\n\\nNote that the proposed patch doesn't use a separate GUC -- it just uses\\nshared_preload_libraries, and then it is the library that's in charge of\\nsetting up the function. We probably wouldn't like to have multiple\\nsettings that all do the same thing, such as recovery target (which\\nseems to be a plentiful source of confusion).\\n\\nChanging the interface so that the user has to specify the function name\\n(not the library name) in ssl_passphrase_command closes that ambiguity\\nhole.\\n\\nNote that if you specify only the library name, it becomes redundant\\nw.r.t. shared_preload_libraries; you could have more than one library\\nsetting the function callback and it's hard to see which one wins.\\n\\nI think something like this would do it:\\n ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\\n\\nThis way, the library can still create any custom GUCs it pleases/needs,\\nbut there's no possible confusion as to the function that's going to be\\ncalled.\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Nov 2019 14:07:58 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 14, 2019 at 02:07:58PM -0300, Alvaro Herrera wrote:\\n> On 2019-Nov-14, Bruce Momjian wrote:\\n> \\n> > I was assuming if the variable starts with a #, it is a shared object,\\n> > if not, it is a shell command:\\n> > \\n> > \\tssl_passphrase_command='#/lib/x.so'\\n> > \\tssl_passphrase_command='my_command a b c'\\n> \\n> Note that the proposed patch doesn't use a separate GUC -- it just uses\\n> shared_preload_libraries, and then it is the library that's in charge of\\n> setting up the function. We probably wouldn't like to have multiple\\n> settings that all do the same thing, such as recovery target (which\\n> seems to be a plentiful source of confusion).\\n> \\n> Changing the interface so that the user has to specify the function name\\n> (not the library name) in ssl_passphrase_command closes that ambiguity\\n> hole.\\n> \\n> Note that if you specify only the library name, it becomes redundant\\n> w.r.t. shared_preload_libraries; you could have more than one library\\n> setting the function callback and it's hard to see which one wins.\\n> \\n> I think something like this would do it:\\n> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\\n> \\n> This way, the library can still create any custom GUCs it pleases/needs,\\n> but there's no possible confusion as to the function that's going to be\\n> called.\\n\\nYeah, I was unclear how the function name would be specified. I thought\\nit would just be hard-coded, but I like the above better. I am still\\nunclear how parameters are passed.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Nov 2019 12:15:44 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"\\nOn 11/14/19 12:07 PM, Alvaro Herrera wrote:\\n> On 2019-Nov-14, Bruce Momjian wrote:\\n>\\n>> I was assuming if the variable starts with a #, it is a shared object,\\n>> if not, it is a shell command:\\n>>\\n>> \\tssl_passphrase_command='#/lib/x.so'\\n>> \\tssl_passphrase_command='my_command a b c'\\n> Note that the proposed patch doesn't use a separate GUC -- it just uses\\n> shared_preload_libraries, and then it is the library that's in charge of\\n> setting up the function. We probably wouldn't like to have multiple\\n> settings that all do the same thing, such as recovery target (which\\n> seems to be a plentiful source of confusion).\\n>\\n> Changing the interface so that the user has to specify the function name\\n> (not the library name) in ssl_passphrase_command closes that ambiguity\\n> hole.\\n>\\n> Note that if you specify only the library name, it becomes redundant\\n> w.r.t. shared_preload_libraries; you could have more than one library\\n> setting the function callback and it's hard to see which one wins.\\n>\\n> I think something like this would do it:\\n> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\\n>\\n> This way, the library can still create any custom GUCs it pleases/needs,\\n> but there's no possible confusion as to the function that's going to be\\n> called.\\n\\n\\nI guess this would work. There would have to be a deal of code to load\\nthe library and lookup the symbol. Do we really think it's worth it?\\nLeveraging shared_preload_libraries makes this comparatively simple.\\n\\n\\nAlso, calling this 'ssl_passphrase_command' seems a little odd.\\n\\n\\nA simpler way to handle it might be simply to error out and refuse to\\nstart if both ssl_passphrase_function is set and ssl_passphrase_command\\nis set.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Nov 2019 14:29:23 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On 2019-Nov-14, Andrew Dunstan wrote:\\n\\n> I guess this would work. There would have to be a deal of code to load\\n> the library and lookup the symbol. Do we really think it's worth it?\\n> Leveraging shared_preload_libraries makes this comparatively simple.\\n\\nUsing the generic interface has the drawback that the user can make more\\nmistakes. I think that's part of Bruce's issue with it (although I may\\nmisinterpret.)\\n\\nI think if you add most of it as a new entry point in dfmgr.c (where you\\ncan leverage internal_library_load) and returns a function pointer to\\nthe user specified function, it's all that much additional code.\\n\\n(I don't think you can use load_external_function as is, because it\\nassumes fmgr V1 calling convention, which I'm not sure serves your case.\\nBut then maybe it does. And if not, then those 10 lines should be very\\nsimilar to the code you'd need to add.)\\n\\n> A simpler way to handle it might be simply to error out and refuse to\\n> start if both ssl_passphrase_function is set and ssl_passphrase_command\\n> is set.\\n\\nYeah, you can do that too I guess, but I'm not sure I see that as simpler.\\n\\n> Also, calling this 'ssl_passphrase_command' seems a little odd.\\n\\nWe could just rename ssl_passphrase_command to something more\\ngeneric, and add the existing name to map_old_guc_names to preserve\\ncompatibility with pg12. Maybe the new name could be simply\\nssl_passphrase or perhaps ssl_passphrase_{reader,getter,pinentry}.\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Nov 2019 17:21:51 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"\\nOn 11/14/19 3:21 PM, Alvaro Herrera wrote:\\n> On 2019-Nov-14, Andrew Dunstan wrote:\\n>\\n>> I guess this would work. There would have to be a deal of code to load\\n>> the library and lookup the symbol. Do we really think it's worth it?\\n>> Leveraging shared_preload_libraries makes this comparatively simple.\\n> Using the generic interface has the drawback that the user can make more\\n> mistakes. I think that's part of Bruce's issue with it (although I may\\n> misinterpret.)\\n>\\n> I think if you add most of it as a new entry point in dfmgr.c (where you\\n> can leverage internal_library_load) and returns a function pointer to\\n> the user specified function, it's all that much additional code.\\n>\\n> (I don't think you can use load_external_function as is, because it\\n> assumes fmgr V1 calling convention, which I'm not sure serves your case.\\n> But then maybe it does. And if not, then those 10 lines should be very\\n> similar to the code you'd need to add.)\\n\\n\\n\\nIn the absence of further comment I will try to code up something along\\nthese lines.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 15 Nov 2019 08:59:45 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Fri, 15 Nov 2019 at 00:34, Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\\nwrote:\\n\\n>\\n> On 11/14/19 11:07 AM, Bruce Momjian wrote:\\n> > On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\\n> >> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <\\n> tomas.vondra@2ndquadrant.com>\\n> >> I think it would be beneficial to explain why shared object is more\\n> >> secure than an OS command. Perhaps it's common knowledge, but it's\\n> not\\n> >> quite obvious to me.\\n> >>\\n> >>\\n> >> Yeah, that probably wouldn't hurt. It's also securely passing from more\\n> than\\n> >> one perspective -- both from the \\\"cannot be eavesdropped\\\" (like putting\\n> the\\n> >> password on the commandline for example) and the requirement for\\n> escaping.\\n> > I think a bigger issue is that if you want to give people the option of\\n> > using a shell command or a shared object, and if you use two commands to\\n> > control it, it isn't clear what happens if both are defined. By using\\n> > some character prefix to control if a shared object is used, you can use\\n> > a single variable and there is no confusion over having two variables\\n> > and their conflicting behavior.\\n> >\\n>\\n>\\n> I'm not sure how that would work in the present instance. The shared\\n> preloaded module installs a function and defines the params it wants. If\\n> we somehow unify the params with ssl_passphrase_command that could look\\n> icky, and the module would have to parse the settings string. That's not\\n> a problem for the sample module which only needs one param, but it will\\n> be for other more complex implementations.\\n>\\n> I'm quite open to suggestions, but I want things to be tolerably clean.\\n\\n\\nIf someone wants a shell command wrapper, they can load a contrib that\\ndelegates the hook to a shell command. So we can just ship a contrib, which\\nacts both as test coverage for the feature, and a shell-command-support\\nwrapper for anyone who desires that.\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\\n\\nOn Fri, 15 Nov 2019 at 00:34, Andrew Dunstan <andrew.dunstan@2ndquadrant.com> wrote:\\nOn 11/14/19 11:07 AM, Bruce Momjian wrote:\\n> On Thu, Nov 14, 2019 at 11:42:05AM +0100, Magnus Hagander wrote:\\n>> On Wed, Nov 13, 2019 at 9:23 PM Tomas Vondra <tomas.vondra@2ndquadrant.com>\\n>> I think it would be beneficial to explain why shared object is more\\n>> secure than an OS command. Perhaps it's common knowledge, but it's not\\n>> quite obvious to me.\\n>>\\n>>\\n>> Yeah, that probably wouldn't hurt. It's also securely passing from more than\\n>> one perspective -- both from the \\\"cannot be eavesdropped\\\" (like putting the\\n>> password on the commandline for example) and the requirement for escaping.\\n> I think a bigger issue is that if you want to give people the option of\\n> using a shell command or a shared object, and if you use two commands to\\n> control it, it isn't clear what happens if both are defined. By using\\n> some character prefix to control if a shared object is used, you can use\\n> a single variable and there is no confusion over having two variables\\n> and their conflicting behavior.\\n>\\n\\n\\nI'm not sure how that would work in the present instance. The shared\\npreloaded module installs a function and defines the params it wants. If\\nwe somehow unify the params with ssl_passphrase_command that could look\\nicky, and the module would have to parse the settings string. That's not\\na problem for the sample module which only needs one param, but it will\\nbe for other more complex implementations.\\n\\nI'm quite open to suggestions, but I want things to be tolerably clean.If someone wants a shell command wrapper, they can load a contrib that delegates the hook to a shell command. So we can just ship a contrib, which acts both as test coverage for the feature, and a shell-command-support wrapper for anyone who desires that. -- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise\",\n \"msg_date\": \"Fri, 22 Nov 2019 13:19:24 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"\\nOn 11/15/19 8:59 AM, Andrew Dunstan wrote:\\n> On 11/14/19 3:21 PM, Alvaro Herrera wrote:\\n>> On 2019-Nov-14, Andrew Dunstan wrote:\\n>>\\n>>> I guess this would work. There would have to be a deal of code to load\\n>>> the library and lookup the symbol. Do we really think it's worth it?\\n>>> Leveraging shared_preload_libraries makes this comparatively simple.\\n>> Using the generic interface has the drawback that the user can make more\\n>> mistakes. I think that's part of Bruce's issue with it (although I may\\n>> misinterpret.)\\n>>\\n>> I think if you add most of it as a new entry point in dfmgr.c (where you\\n>> can leverage internal_library_load) and returns a function pointer to\\n>> the user specified function, it's all that much additional code.\\n>>\\n>> (I don't think you can use load_external_function as is, because it\\n>> assumes fmgr V1 calling convention, which I'm not sure serves your case.\\n>> But then maybe it does. And if not, then those 10 lines should be very\\n>> similar to the code you'd need to add.)\\n\\n\\nI've just been looking at that. load_external_function() doesn't\\nactually do anything V1-ish with the value, it just looks up the symbol\\nusing dlsym and returns it cast to a PGFunction. Is there any reason I\\ncan't just use that and cast it again to the callback function type?\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 6 Dec 2019 17:21:58 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n> I've just been looking at that. load_external_function() doesn't\\n> actually do anything V1-ish with the value, it just looks up the symbol\\n> using dlsym and returns it cast to a PGFunction. Is there any reason I\\n> can't just use that and cast it again to the callback function type?\\n\\nTBH, I think this entire discussion has gone seriously off into the\\nweeds. The original design where we just let a shared_preload_library\\nfunction get into a hook is far superior to any of the overcomplicated\\nkluges that are being discussed now. Something like this, for instance:\\n\\n>>> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\\n\\nmakes me positively ill. It introduces problems that we don't need,\\nlike how to parse out the sub-parts of the string, and the\\nquoting/escaping issues that will come along with that; while from\\nthe user's perspective it replaces a simple and intellectually-coherent\\nvariable definition with an unintelligible mess.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 06 Dec 2019 18:20:54 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"\\nOn 12/6/19 6:20 PM, Tom Lane wrote:\\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n>> I've just been looking at that. load_external_function() doesn't\\n>> actually do anything V1-ish with the value, it just looks up the symbol\\n>> using dlsym and returns it cast to a PGFunction. Is there any reason I\\n>> can't just use that and cast it again to the callback function type?\\n> TBH, I think this entire discussion has gone seriously off into the\\n> weeds. The original design where we just let a shared_preload_library\\n> function get into a hook is far superior to any of the overcomplicated\\n> kluges that are being discussed now. Something like this, for instance:\\n>\\n>>>> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\\n> makes me positively ill. It introduces problems that we don't need,\\n> like how to parse out the sub-parts of the string, and the\\n> quoting/escaping issues that will come along with that; while from\\n> the user's perspective it replaces a simple and intellectually-coherent\\n> variable definition with an unintelligible mess.\\n>\\n> \\t\\t\\t\\n\\n\\n\\nYeah, you have a point.\\n\\n\\nBruce was worried about what would happen if we defined both\\nssl_passphrase_command and ssl_passphrase_callback. The submitted patch\\nlet's the callback have precedence, but it might be cleaner to error out\\nwith such a config. OTOH, that wouldn't be so nice on a reload, so it\\nmight be better just to document the behaviour.\\n\\n\\nHe was also worried that multiple shared libraries might try to provide\\nthe hook. I think that's fairly fanciful, TBH. It comes into the\\ncategory of \\\"Don't do that.\\\"\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 6 Dec 2019 19:32:32 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n> Bruce was worried about what would happen if we defined both\\n> ssl_passphrase_command and ssl_passphrase_callback. The submitted patch\\n> let's the callback have precedence, but it might be cleaner to error out\\n> with such a config. OTOH, that wouldn't be so nice on a reload, so it\\n> might be better just to document the behaviour.\\n\\nI think it would be up to the extension that's using the hook to\\ndecide what to do if ssl_passphrase_command is set. It would not\\nbe our choice, and it would certainly not fall to us to document it.\\n\\n> He was also worried that multiple shared libraries might try to provide\\n> the hook. I think that's fairly fanciful, TBH. It comes into the\\n> category of \\\"Don't do that.\\\"\\n\\nAgain, it's somebody else's problem. We have plenty of hooks that\\nare of dubious use for multiple extensions, so why should this one be\\nheld to a higher standard?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 07 Dec 2019 12:16:17 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"\\nOn 12/7/19 12:16 PM, Tom Lane wrote:\\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n>> Bruce was worried about what would happen if we defined both\\n>> ssl_passphrase_command and ssl_passphrase_callback. The submitted patch\\n>> let's the callback have precedence, but it might be cleaner to error out\\n>> with such a config. OTOH, that wouldn't be so nice on a reload, so it\\n>> might be better just to document the behaviour.\\n> I think it would be up to the extension that's using the hook to\\n> decide what to do if ssl_passphrase_command is set. It would not\\n> be our choice, and it would certainly not fall to us to document it.\\n>\\n>> He was also worried that multiple shared libraries might try to provide\\n>> the hook. I think that's fairly fanciful, TBH. It comes into the\\n>> category of \\\"Don't do that.\\\"\\n> Again, it's somebody else's problem. We have plenty of hooks that\\n> are of dubious use for multiple extensions, so why should this one be\\n> held to a higher standard?\\n>\\n> \\t\\t\\t\\n\\n\\nWell that pretty much brings us back to the patch as submitted :-)\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 7 Dec 2019 16:03:27 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n> Well that pretty much brings us back to the patch as submitted :-)\\n\\nYeah, pretty nearly. Taking a quick look over the v3 patch, my\\nonly quibble is that it doesn't provide any convenient way for the\\nexternal module to make decisions about how to interact with\\nssl_passphrase_command --- in particular, if it would like to allow\\nthat to take precedence, it can't because there's no way for it to\\ninvoke the static function ssl_external_passwd_cb.\\n\\nBut rather than expose that globally, maybe the theory ought to be\\n\\\"set up the state as we'd normally do, then let loadable modules\\nchoose to override it\\\". So I'm tempted to propose a hook function\\nwith the signature\\n\\nvoid openssl_tls_init_hook(SSL_CTX *context, bool isServerStart);\\n\\nand invoke that somewhere in be_tls_init --- maybe fairly late,\\nso that it can override other settings if it wants, not only the\\nSSL_CTX_set_default_passwd_cb setting.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 07 Dec 2019 17:32:01 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Sat, 7 Dec 2019 at 07:21, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n> > I've just been looking at that. load_external_function() doesn't\\n> > actually do anything V1-ish with the value, it just looks up the symbol\\n> > using dlsym and returns it cast to a PGFunction. Is there any reason I\\n> > can't just use that and cast it again to the callback function type?\\n>\\n> TBH, I think this entire discussion has gone seriously off into the\\n> weeds. The original design where we just let a shared_preload_library\\n> function get into a hook is far superior to any of the overcomplicated\\n> kluges that are being discussed now. Something like this, for instance:\\n>\\n> >>> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\\n>\\n> makes me positively ill. It introduces problems that we don't need,\\n> like how to parse out the sub-parts of the string, and the\\n> quoting/escaping issues that will come along with that; while from\\n> the user's perspective it replaces a simple and intellectually-coherent\\n> variable definition with an unintelligible mess.\\n>\\n\\n+1000 from me on that.\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\\n\\nOn Sat, 7 Dec 2019 at 07:21, Tom Lane <tgl@sss.pgh.pa.us> wrote:Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n> I've just been looking at that. load_external_function() doesn't\\n> actually do anything V1-ish with the value, it just looks up the symbol\\n> using dlsym and returns it cast to a PGFunction. Is there any reason I\\n> can't just use that and cast it again to the callback function type?\\n\\nTBH, I think this entire discussion has gone seriously off into the\\nweeds. The original design where we just let a shared_preload_library\\nfunction get into a hook is far superior to any of the overcomplicated\\nkluges that are being discussed now. Something like this, for instance:\\n\\n>>> ssl_passphrase_command='#superlib.so,my_rot13_passphrase'\\n\\nmakes me positively ill. It introduces problems that we don't need,\\nlike how to parse out the sub-parts of the string, and the\\nquoting/escaping issues that will come along with that; while from\\nthe user's perspective it replaces a simple and intellectually-coherent\\nvariable definition with an unintelligible mess.+1000 from me on that. -- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise\",\n \"msg_date\": \"Mon, 9 Dec 2019 10:22:12 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Sun, Dec 8, 2019 at 9:02 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n> > Well that pretty much brings us back to the patch as submitted :-)\\n>\\n> Yeah, pretty nearly. Taking a quick look over the v3 patch, my\\n> only quibble is that it doesn't provide any convenient way for the\\n> external module to make decisions about how to interact with\\n> ssl_passphrase_command --- in particular, if it would like to allow\\n> that to take precedence, it can't because there's no way for it to\\n> invoke the static function ssl_external_passwd_cb.\\n>\\n> But rather than expose that globally, maybe the theory ought to be\\n> \\\"set up the state as we'd normally do, then let loadable modules\\n> choose to override it\\\". So I'm tempted to propose a hook function\\n> with the signature\\n>\\n> void openssl_tls_init_hook(SSL_CTX *context, bool isServerStart);\\n>\\n> and invoke that somewhere in be_tls_init --- maybe fairly late,\\n> so that it can override other settings if it wants, not only the\\n> SSL_CTX_set_default_passwd_cb setting.\\n>\\n\\n\\nNot sure if the placement is what you want, but maybe something like this?\\n\\ncheers\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 22 Jan 2020 17:32:01 +1030\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 14, 2019 at 8:54 AM Sehrope Sarkuni <sehrope@jackdb.com> wrote:\\n> Has the idea of using environment variables (rather than command line\\n> args) for external commands been brought up before? I couldn't find\\n> anything in the mailing list archives.\\n\\nPassing data through environment variables isn't secure. Try 'ps -E'\\non MacOS, or something like 'ps axe' on Linux.\\n\\nIf we want to pass data securely to child processes, the way to do it\\nis via stdin. Data sent back and forth via file descriptors can't\\neasily be snooped by other users on the system.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 23 Jan 2020 12:30:36 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 22, 2020 at 8:02 PM Andrew Dunstan\\n<andrew.dunstan@2ndquadrant.com> wrote:\\n> Not sure if the placement is what you want, but maybe something like this?\\n\\nHi Andrew, FYI this failed here:\\n\\nt/001_testfunc.pl .. Bailout called. Further testing stopped: pg_ctl\\nstart failed\\nFAILED--Further testing stopped: pg_ctl start failed\\nMakefile:23: recipe for target 'prove-check' failed\\n\\nUnfortunately my robot is poorly trained and does not dump any of the\\ninteresting logs for this case, but it looks like it's failing that\\nway every time.\\n\\nhttps://travis-ci.org/postgresql-cfbot/postgresql/builds/651756191\\n\\n\\n\",\n \"msg_date\": \"Tue, 18 Feb 2020 16:30:32 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Tue, Feb 18, 2020 at 2:01 PM Thomas Munro <thomas.munro@gmail.com> wrote:\\n>\\n> On Wed, Jan 22, 2020 at 8:02 PM Andrew Dunstan\\n> <andrew.dunstan@2ndquadrant.com> wrote:\\n> > Not sure if the placement is what you want, but maybe something like this?\\n>\\n> Hi Andrew, FYI this failed here:\\n>\\n> t/001_testfunc.pl .. Bailout called. Further testing stopped: pg_ctl\\n> start failed\\n> FAILED--Further testing stopped: pg_ctl start failed\\n> Makefile:23: recipe for target 'prove-check' failed\\n>\\n> Unfortunately my robot is poorly trained and does not dump any of the\\n> interesting logs for this case, but it looks like it's failing that\\n> way every time.\\n>\\n> https://travis-ci.org/postgresql-cfbot/postgresql/builds/651756191\\n\\n\\nThanks for letting me know, I will investigate.\\n\\ncheers\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Wed, 19 Feb 2020 07:10:20 +1030\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 19, 2020 at 7:10 AM Andrew Dunstan\\n<andrew.dunstan@2ndquadrant.com> wrote:\\n>\\n> On Tue, Feb 18, 2020 at 2:01 PM Thomas Munro <thomas.munro@gmail.com> wrote:\\n> >\\n> > On Wed, Jan 22, 2020 at 8:02 PM Andrew Dunstan\\n> > <andrew.dunstan@2ndquadrant.com> wrote:\\n> > > Not sure if the placement is what you want, but maybe something like this?\\n> >\\n> > Hi Andrew, FYI this failed here:\\n> >\\n> > t/001_testfunc.pl .. Bailout called. Further testing stopped: pg_ctl\\n> > start failed\\n> > FAILED--Further testing stopped: pg_ctl start failed\\n> > Makefile:23: recipe for target 'prove-check' failed\\n> >\\n> > Unfortunately my robot is poorly trained and does not dump any of the\\n> > interesting logs for this case, but it looks like it's failing that\\n> > way every time.\\n> >\\n> > https://travis-ci.org/postgresql-cfbot/postgresql/builds/651756191\\n>\\n>\\n> Thanks for letting me know, I will investigate.\\n>\\n\\n\\nThis should fix the issue, it happened when I switched to using a\\npre-generated cert/key.\\n\\ncheers\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 19 Feb 2020 09:09:20 +1030\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On 2/18/20 11:39 PM, Andrew Dunstan wrote:\\n> This should fix the issue, it happened when I switched to using a\\n> pre-generated cert/key.\\n\\n# Review\\n\\nThe patch still applies and passes the test suite, both with openssl \\nenabled and with it disabled.\\n\\nAs for the feature I agree that it is nice to expose this callback to \\nextension writers and I agree with the approach taken. The other \\nproposals up-thread seem over engineered to me. Maybe if it was a \\ngeneral feature used in many places those proposals would be worth it, \\nbut I am still skeptical even then. This approach is so much simpler.\\n\\nThe only real risk I see is that if people install multiple libraries \\nfor this they will overwrite the hook for each other but we have other \\ncases like that already so I think that is fine.\\n\\nThe patch moves secure_initialize() to after \\nprocess_shared_preload_libraries() which in theory could break some \\nextension but it does not seem like a likely thing for extensions to \\nrely on. Or is it?\\n\\nAn idea would be to have the code in ssl_passphrase_func.c to warn if \\nthe ssl_passphrase_command GUC is set to make it more useful as example \\ncode for people implementing this hook.\\n\\n# Nitpicking\\n\\nThe certificate expires in 2030 while all other certificates used in \\ntests expires in 2046. Should we be consistent?\\n\\nThere is text in server.crt and server.key, while other certificates and \\nkeys used in the tests do not have this. Again, should we be consistent?\\n\\nEmpty first line in \\nsrc/test/modules/ssl_passphrase_callback/t/001_testfunc.pl which should \\nprobably just be removed or replaced with a shebang.\\n\\nThere is an extra space between the parentheses in the line below. Does \\nthat follow our code style for Perl?\\n\\n+unless ( ($ENV{with_openssl} || 'no') eq 'yes')\\n\\nMissing space after comma in:\\n\\n+ok(-e \\\"$ddir/postmaster.pid\\\",\\\"postgres started\\\");\\n\\nMissing space after comma in:\\n\\n+ok(! -e \\\"$ddir/postmaster.pid\\\",\\\"postgres not started with bad passphrase\\\");\\n\\nAndreas\\n\\n\\n\",\n \"msg_date\": \"Mon, 16 Mar 2020 03:14:57 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"Hello Andrew,\\r\\n\\r\\nFrom: Andreas Karlsson <andreas@proxel.se>\\r\\n> # Nitpicking\\r\\n> \\r\\n> The certificate expires in 2030 while all other certificates used in\\r\\n> tests expires in 2046. Should we be consistent?\\r\\n> \\r\\n> There is text in server.crt and server.key, while other certificates and\\r\\n> keys used in the tests do not have this. Again, should we be consistent?\\r\\n> \\r\\n> Empty first line in\\r\\n> src/test/modules/ssl_passphrase_callback/t/001_testfunc.pl which should\\r\\n> probably just be removed or replaced with a shebang.\\r\\n> \\r\\n> There is an extra space between the parentheses in the line below. Does\\r\\n> that follow our code style for Perl?\\r\\n> \\r\\n> +unless ( ($ENV{with_openssl} || 'no') eq 'yes')\\r\\n> \\r\\n> Missing space after comma in:\\r\\n> \\r\\n> +ok(-e \\\"$ddir/postmaster.pid\\\",\\\"postgres started\\\");\\r\\n> \\r\\n> Missing space after comma in:\\r\\n> \\r\\n> +ok(! -e \\\"$ddir/postmaster.pid\\\",\\\"postgres not started with bad passphrase\\\");\\r\\n> \\r\\n> Andreas\\r\\n> \\r\\n\\r\\nTrailing space:\\r\\n\\r\\n220 + X509v3 Subject Key Identifier:\\r\\n222 + X509v3 Authority Key Identifier:\\r\\n\\r\\nMissing \\\"d\\\"(password?):\\r\\n\\r\\n121 +/* init hook for SSL, the default sets the passwor callback if appropriate */\\r\\n\\r\\nRegards,\\r\\n\\r\\n--\\r\\nTakanori Asaba\\r\\n\\r\\n\\r\\n\",\n \"msg_date\": \"Thu, 19 Mar 2020 08:10:58 +0000\",\n \"msg_from\": \"\\\"asaba.takanori@fujitsu.com\\\" <asaba.takanori@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"\\nOn 3/15/20 10:14 PM, Andreas Karlsson wrote:\\n> On 2/18/20 11:39 PM, Andrew Dunstan wrote:\\n>> This should fix the issue, it happened when I switched to using a\\n>> pre-generated cert/key.\\n>\\n> # Review\\n>\\n> The patch still applies and passes the test suite, both with openssl\\n> enabled and with it disabled.\\n>\\n> As for the feature I agree that it is nice to expose this callback to\\n> extension writers and I agree with the approach taken. The other\\n> proposals up-thread seem over engineered to me. Maybe if it was a\\n> general feature used in many places those proposals would be worth it,\\n> but I am still skeptical even then. This approach is so much simpler.\\n>\\n> The only real risk I see is that if people install multiple libraries\\n> for this they will overwrite the hook for each other but we have other\\n> cases like that already so I think that is fine.\\n\\n\\nRight, me too.\\n\\n\\n>\\n> The patch moves secure_initialize() to after\\n> process_shared_preload_libraries() which in theory could break some\\n> extension but it does not seem like a likely thing for extensions to\\n> rely on. Or is it?\\n\\n\\nI don't think so.\\n\\n\\n>\\n> An idea would be to have the code in ssl_passphrase_func.c to warn if\\n> the ssl_passphrase_command GUC is set to make it more useful as\\n> example code for people implementing this hook.\\n\\n\\nI'll look at that. Should be possible.\\n\\n\\n>\\n> # Nitpicking\\n>\\n> The certificate expires in 2030 while all other certificates used in\\n> tests expires in 2046. Should we be consistent?\\n\\n\\nSure. will fix.\\n\\n\\n>\\n> There is text in server.crt and server.key, while other certificates\\n> and keys used in the tests do not have this. Again, should we be\\n> consistent?\\n\\n\\nNot in server.key, but I will suppress it for the crt file.\\n\\n\\n\\n>\\n> Empty first line in\\n> src/test/modules/ssl_passphrase_callback/t/001_testfunc.pl which\\n> should probably just be removed or replaced with a shebang.\\n\\n\\nOK\\n\\n\\n>\\n> There is an extra space between the parentheses in the line below.\\n> Does that follow our code style for Perl?\\n>\\n> +unless ( ($ENV{with_openssl} || 'no') eq 'yes')\\n>\\n> Missing space after comma in:\\n>\\n> +ok(-e \\\"$ddir/postmaster.pid\\\",\\\"postgres started\\\");\\n>\\n> Missing space after comma in:\\n>\\n> +ok(! -e \\\"$ddir/postmaster.pid\\\",\\\"postgres not started with bad\\n> passphrase\\\");\\n\\n\\nI'll make sure to run it through our perl indenter.\\n\\n\\nThanks for the review.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 21 Mar 2020 09:15:53 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"\\nOn 3/19/20 4:10 AM, asaba.takanori@fujitsu.com wrote:\\n\\n\\n\\n> Trailing space:\\n>\\n> 220 + X509v3 Subject Key Identifier:\\n> 222 + X509v3 Authority Key Identifier:\\n\\n\\nWe're going to remove all the text, so this becomes moot.\\n\\n\\n>\\n> Missing \\\"d\\\"(password?):\\n>\\n> 121 +/* init hook for SSL, the default sets the passwor callback if appropriate */\\n>\\n\\nWill fix, thanks.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 21 Mar 2020 09:18:26 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On 3/21/20 9:18 AM, Andrew Dunstan wrote:\\n> On 3/19/20 4:10 AM, asaba.takanori@fujitsu.com wrote:\\n>\\n>\\n>\\n>> Trailing space:\\n>>\\n>> 220 + X509v3 Subject Key Identifier:\\n>> 222 + X509v3 Authority Key Identifier:\\n>\\n> We're going to remove all the text, so this becomes moot.\\n>\\n>\\n>> Missing \\\"d\\\"(password?):\\n>>\\n>> 121 +/* init hook for SSL, the default sets the passwor callback if appropriate */\\n>>\\n> Will fix, thanks.\\n>\\n>\\n\\n\\nLatest patch attached, I think all comments have been addressed. I\\npropose to push this later this coming week if there are no more comments.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Sat, 21 Mar 2020 20:08:19 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"On 3/22/20 1:08 AM, Andrew Dunstan wrote:\\n> Latest patch attached, I think all comments have been addressed. I\\n> propose to push this later this coming week if there are no more comments.\\n\\nI do not have any objections.\\n\\nAndreas\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 23 Mar 2020 01:58:14 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"Andreas Karlsson <andreas@proxel.se> writes:\\n> On 3/22/20 1:08 AM, Andrew Dunstan wrote:\\n>> Latest patch attached, I think all comments have been addressed. I\\n>> propose to push this later this coming week if there are no more comments.\\n\\n> I do not have any objections.\\n\\nThis CF entry is still open, should it not be closed as committed?\\n\\nhttps://commitfest.postgresql.org/27/2338/\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 28 Mar 2020 13:45:00 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n },\n {\n \"msg_contents\": \"\\nOn 3/28/20 1:45 PM, Tom Lane wrote:\\n> Andreas Karlsson <andreas@proxel.se> writes:\\n>> On 3/22/20 1:08 AM, Andrew Dunstan wrote:\\n>>> Latest patch attached, I think all comments have been addressed. I\\n>>> propose to push this later this coming week if there are no more comments.\\n>> I do not have any objections.\\n> This CF entry is still open, should it not be closed as committed?\\n>\\n> https://commitfest.postgresql.org/27/2338/\\n>\\n> \\t\\t\\t\\n\\n\\nDone, thanks for the reminder.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 28 Mar 2020 15:35:07 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ssl passphrase callback\"\n }\n]","isConversational":false}}},{"rowIdx":49378,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi everyone,\n\nI was taking a look at pg_stat_statements module and noticed that it does\nnot collect any percentile metrics. I believe that It would be really handy\nto have those available and I'd love to contribute with this feature.\n\nThe basic idea is to accumulate the the query execution times using an\napproximation structure like q-digest or t-digest and add those results to\nthe pg_stat_statements table as fixed columns. Something like this\n\np90_time:\np95_time:\np99_time:\np70_time:\n...\n\nAnother solution is to persist de digest structure in a binary column and\nuse a function to extract the desired quantile ilke this SELECT\napprox_quantile(digest_times, 0.99) FROM pg_stat_statements\n\nWhat do you guys think?\nCheers,\n\nHi everyone,I was taking a look at pg_stat_statements module and noticed that it does not collect any percentile metrics. I believe that It would be really handy to have those available and I'd love to contribute with this feature. The basic idea is to accumulate the the query execution times using an approximation structure like q-digest or t-digest and add those results to the pg_stat_statements table as fixed columns. Something like thisp90_time:p95_time:p99_time:p70_time:...Another solution is to persist de digest structure in a binary column and use a function to extract the desired quantile ilke this SELECT approx_quantile(digest_times, 0.99) FROM pg_stat_statementsWhat do you guys think? Cheers,","msg_date":"Thu, 31 Oct 2019 12:51:17 -0300","msg_from":"Igor Calabria <igor.calabria@gmail.com>","msg_from_op":true,"msg_subject":"Adding percentile metrics to pg_stat_statements module"},{"msg_contents":"čt 31. 10. 2019 v 16:51 odesílatel Igor Calabria <igor.calabria@gmail.com>\nnapsal:\n\n> Hi everyone,\n>\n> I was taking a look at pg_stat_statements module and noticed that it does\n> not collect any percentile metrics. I believe that It would be really handy\n> to have those available and I'd love to contribute with this feature.\n>\n> The basic idea is to accumulate the the query execution times using an\n> approximation structure like q-digest or t-digest and add those results to\n> the pg_stat_statements table as fixed columns. Something like this\n>\n> p90_time:\n> p95_time:\n> p99_time:\n> p70_time:\n> ...\n>\n> Another solution is to persist de digest structure in a binary column and\n> use a function to extract the desired quantile ilke this SELECT\n> approx_quantile(digest_times, 0.99) FROM pg_stat_statements\n>\n> What do you guys think?\n>\n\n+ the idea\n\nBut I am not sure about performance and memory overhead\n\nPavel\n\n> Cheers,\n>\n>\n\nčt 31. 10. 2019 v 16:51 odesílatel Igor Calabria <igor.calabria@gmail.com> napsal:Hi everyone,I was taking a look at pg_stat_statements module and noticed that it does not collect any percentile metrics. I believe that It would be really handy to have those available and I'd love to contribute with this feature. The basic idea is to accumulate the the query execution times using an approximation structure like q-digest or t-digest and add those results to the pg_stat_statements table as fixed columns. Something like thisp90_time:p95_time:p99_time:p70_time:...Another solution is to persist de digest structure in a binary column and use a function to extract the desired quantile ilke this SELECT approx_quantile(digest_times, 0.99) FROM pg_stat_statementsWhat do you guys think? + the ideaBut I am not sure about performance and memory overheadPavelCheers,","msg_date":"Thu, 31 Oct 2019 17:36:27 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Adding percentile metrics to pg_stat_statements module"},{"msg_contents":"On Thu, Oct 31, 2019 at 12:51:17PM -0300, Igor Calabria wrote:\n>Hi everyone,\n>\n>I was taking a look at pg_stat_statements module and noticed that it does\n>not collect any percentile metrics. I believe that It would be really handy\n>to have those available and I'd love to contribute with this feature.\n>\n>The basic idea is to accumulate the the query execution times using an\n>approximation structure like q-digest or t-digest and add those results to\n>the pg_stat_statements table as fixed columns. Something like this\n>\n>p90_time:\n>p95_time:\n>p99_time:\n>p70_time:\n>...\n>\n>Another solution is to persist de digest structure in a binary column and\n>use a function to extract the desired quantile ilke this SELECT\n>approx_quantile(digest_times, 0.99) FROM pg_stat_statements\n>\n\nIMO having some sort of CDF approximation (being a q-digest or t-digest)\nwould be useful, although it'd probably need to be optional (mostly\nbecuase of memory consumption).\n\nI don't see why we would not store the digests themselves. Storing just\nsome selected percentiles would be pretty problematic due to losing a\nlot of information on restart. Also, pg_stat_statements is not a table\nbut a view on in-memory hash table.\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 31 Oct 2019 20:32:47 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Adding percentile metrics to pg_stat_statements module"},{"msg_contents":"Yeah, I agree that there's no reason to store the digests themselves and I\nreally liked the idea of it being optional.\nIf it turns out that memory consumption on real workloads is small enough,\nit could eventually be turned on by default.\n\nI'll start working on patch\n\nEm qui, 31 de out de 2019 às 16:32, Tomas Vondra <\ntomas.vondra@2ndquadrant.com> escreveu:\n\n> On Thu, Oct 31, 2019 at 12:51:17PM -0300, Igor Calabria wrote:\n> >Hi everyone,\n> >\n> >I was taking a look at pg_stat_statements module and noticed that it does\n> >not collect any percentile metrics. I believe that It would be really\n> handy\n> >to have those available and I'd love to contribute with this feature.\n> >\n> >The basic idea is to accumulate the the query execution times using an\n> >approximation structure like q-digest or t-digest and add those results to\n> >the pg_stat_statements table as fixed columns. Something like this\n> >\n> >p90_time:\n> >p95_time:\n> >p99_time:\n> >p70_time:\n> >...\n> >\n> >Another solution is to persist de digest structure in a binary column and\n> >use a function to extract the desired quantile ilke this SELECT\n> >approx_quantile(digest_times, 0.99) FROM pg_stat_statements\n> >\n>\n> IMO having some sort of CDF approximation (being a q-digest or t-digest)\n> would be useful, although it'd probably need to be optional (mostly\n> becuase of memory consumption).\n>\n> I don't see why we would not store the digests themselves. Storing just\n> some selected percentiles would be pretty problematic due to losing a\n> lot of information on restart. Also, pg_stat_statements is not a table\n> but a view on in-memory hash table.\n>\n>\n> regards\n>\n> --\n> Tomas Vondra http://www.2ndQuadrant.com\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\nYeah, I agree that there's no reason to store the digests themselves and I really liked the idea of it being optional. If it turns out that memory consumption on real workloads is small enough, it could eventually be turned on by default. I'll start working on patch Em qui, 31 de out de 2019 às 16:32, Tomas Vondra <tomas.vondra@2ndquadrant.com> escreveu:On Thu, Oct 31, 2019 at 12:51:17PM -0300, Igor Calabria wrote:\n>Hi everyone,\n>\n>I was taking a look at pg_stat_statements module and noticed that it does\n>not collect any percentile metrics. I believe that It would be really handy\n>to have those available and I'd love to contribute with this feature.\n>\n>The basic idea is to accumulate the the query execution times using an\n>approximation structure like q-digest or t-digest and add those results to\n>the pg_stat_statements table as fixed columns. Something like this\n>\n>p90_time:\n>p95_time:\n>p99_time:\n>p70_time:\n>...\n>\n>Another solution is to persist de digest structure in a binary column and\n>use a function to extract the desired quantile ilke this SELECT\n>approx_quantile(digest_times, 0.99) FROM pg_stat_statements\n>\n\nIMO having some sort of CDF approximation (being a q-digest or t-digest)\nwould be useful, although it'd probably need to be optional (mostly\nbecuase of memory consumption).\n\nI don't see why we would not store the digests themselves. Storing just\nsome selected percentiles would be pretty problematic due to losing a\nlot of information on restart. Also, pg_stat_statements is not a table\nbut a view on in-memory hash table.\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Fri, 1 Nov 2019 11:11:13 -0300","msg_from":"Igor Calabria <igor.calabria@gmail.com>","msg_from_op":true,"msg_subject":"Re: Adding percentile metrics to pg_stat_statements module"},{"msg_contents":"On Fri, Nov 01, 2019 at 11:11:13AM -0300, Igor Calabria wrote:\n>Yeah, I agree that there's no reason to store the digests themselves and I\n>really liked the idea of it being optional.\n\nThat's not what I wrote. My point was that we *should* store the digests\nthemselves, otherwise we just introduce additional errors into the\nestimates, because it discards the weights/frequencies.\n\n>If it turns out that memory consumption on real workloads is small enough,\n>it could eventually be turned on by default.\n>\n\nMaybe, but it's not just about memory consumption. CPU matters too.\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Fri, 1 Nov 2019 15:17:51 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Adding percentile metrics to pg_stat_statements module"},{"msg_contents":">\n> That's not what I wrote. My point was that we *should* store the digests\n> themselves, otherwise we just introduce additional errors into the\n> estimates, because it discards the weights/frequencies.\n\n\nSorry. I meant to write \"no reason to *not* store the digests\"\n\n\nEm sex, 1 de nov de 2019 às 11:17, Tomas Vondra <\ntomas.vondra@2ndquadrant.com> escreveu:\n\n> On Fri, Nov 01, 2019 at 11:11:13AM -0300, Igor Calabria wrote:\n> >Yeah, I agree that there's no reason to store the digests themselves and I\n> >really liked the idea of it being optional.\n>\n> That's not what I wrote. My point was that we *should* store the digests\n> themselves, otherwise we just introduce additional errors into the\n> estimates, because it discards the weights/frequencies.\n>\n> >If it turns out that memory consumption on real workloads is small enough,\n> >it could eventually be turned on by default.\n> >\n>\n> Maybe, but it's not just about memory consumption. CPU matters too.\n>\n> regards\n>\n> --\n> Tomas Vondra http://www.2ndQuadrant.com\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\nThat's not what I wrote. My point was that we *should* store the digests\nthemselves, otherwise we just introduce additional errors into the\nestimates, because it discards the weights/frequencies.Sorry. I meant to write \"no reason to not store the digests\" Em sex, 1 de nov de 2019 às 11:17, Tomas Vondra <tomas.vondra@2ndquadrant.com> escreveu:On Fri, Nov 01, 2019 at 11:11:13AM -0300, Igor Calabria wrote:\n>Yeah, I agree that there's no reason to store the digests themselves and I\n>really liked the idea of it being optional.\n\nThat's not what I wrote. My point was that we *should* store the digests\nthemselves, otherwise we just introduce additional errors into the\nestimates, because it discards the weights/frequencies.\n\n>If it turns out that memory consumption on real workloads is small enough,\n>it could eventually be turned on by default.\n>\n\nMaybe, but it's not just about memory consumption. CPU matters too.\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Fri, 1 Nov 2019 13:05:02 -0300","msg_from":"Igor Calabria <igor.calabria@gmail.com>","msg_from_op":true,"msg_subject":"Re: Adding percentile metrics to pg_stat_statements module"},{"msg_contents":"On 10/31/19 8:32 PM, Tomas Vondra wrote:\n> IMO having some sort of CDF approximation (being a q-digest or t-digest)\n> would be useful, although it'd probably need to be optional (mostly\n> becuase of memory consumption).\n\n+1, I like this idea. If we are afraid of CPU cost we could imagine some kind of\nsampling or add the possibility to collect only for a specific queryid.\n\nI dreamed of this kind of feature for PoWA. Thus, it could make possible to\ncompare CDF between two days for example, before and after introducing a change.\n\nRegards,\n\n-- \nAdrien NAYRAT\n\n\n\n","msg_date":"Sat, 2 Nov 2019 10:23:49 +0100","msg_from":"Adrien Nayrat <adrien.nayrat@anayrat.info>","msg_from_op":false,"msg_subject":"Re: Adding percentile metrics to pg_stat_statements module"},{"msg_contents":"Hello\n\n\nWhile looking at slow queries on pg_stat_statements. I was looking for percentile fields..\n\n\nIf we are worried about CPU cost, maybe it could be useful to turn in on when you have a high stddev_exec_time for the query ?\n\n\nRegards,\n\n________________________________\nDe : Adrien Nayrat <adrien.nayrat@anayrat.info>\nEnvoyé : samedi 2 novembre 2019 10:23:49\nÀ : Tomas Vondra; Igor Calabria\nCc : pgsql-hackers@postgresql.org\nObjet : Re: Adding percentile metrics to pg_stat_statements module\n\nOn 10/31/19 8:32 PM, Tomas Vondra wrote:\n> IMO having some sort of CDF approximation (being a q-digest or t-digest)\n> would be useful, although it'd probably need to be optional (mostly\n> becuase of memory consumption).\n\n+1, I like this idea. If we are afraid of CPU cost we could imagine some kind of\nsampling or add the possibility to collect only for a specific queryid.\n\nI dreamed of this kind of feature for PoWA. Thus, it could make possible to\ncompare CDF between two days for example, before and after introducing a change.\n\nRegards,\n\n--\nAdrien NAYRAT\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nHello\n\n\nWhile looking at slow queries on pg_stat_statements. I was looking for percentile fields.. \n\n\nIf we are worried about CPU cost, maybe it could be useful to turn in on when you have a high stddev_exec_time for the query ?\n\n\nRegards,\n\n\nDe : Adrien Nayrat <adrien.nayrat@anayrat.info>\nEnvoyé : samedi 2 novembre 2019 10:23:49\nÀ : Tomas Vondra; Igor Calabria\nCc : pgsql-hackers@postgresql.org\nObjet : Re: Adding percentile metrics to pg_stat_statements module\n \n\n\n\nOn 10/31/19 8:32 PM, Tomas Vondra wrote:\n> IMO having some sort of CDF approximation (being a q-digest or t-digest)\n> would be useful, although it'd probably need to be optional (mostly\n> becuase of memory consumption).\n\n+1, I like this idea. If we are afraid of CPU cost we could imagine some kind of\nsampling or add the possibility to collect only for a specific queryid.\n\nI dreamed of this kind of feature for PoWA. Thus, it could make possible to\ncompare CDF between two days for example, before and after introducing a change.\n\nRegards,\n\n-- \nAdrien NAYRAT","msg_date":"Mon, 5 Jun 2023 08:05:10 +0000","msg_from":"benoit <benoit@hopsandfork.com>","msg_from_op":false,"msg_subject":"RE: Adding percentile metrics to pg_stat_statements module"}],"string":"[\n {\n \"msg_contents\": \"Hi everyone,\\n\\nI was taking a look at pg_stat_statements module and noticed that it does\\nnot collect any percentile metrics. I believe that It would be really handy\\nto have those available and I'd love to contribute with this feature.\\n\\nThe basic idea is to accumulate the the query execution times using an\\napproximation structure like q-digest or t-digest and add those results to\\nthe pg_stat_statements table as fixed columns. Something like this\\n\\np90_time:\\np95_time:\\np99_time:\\np70_time:\\n...\\n\\nAnother solution is to persist de digest structure in a binary column and\\nuse a function to extract the desired quantile ilke this SELECT\\napprox_quantile(digest_times, 0.99) FROM pg_stat_statements\\n\\nWhat do you guys think?\\nCheers,\\n\\nHi everyone,I was taking a look at pg_stat_statements module and noticed that it does not collect any percentile metrics. I believe that It would be really handy to have those available and I'd love to contribute with this feature. The basic idea is to accumulate the the query execution times using an approximation structure like q-digest or t-digest and add those results to the pg_stat_statements table as fixed columns. Something like thisp90_time:p95_time:p99_time:p70_time:...Another solution is to persist de digest structure in a binary column and use a function to extract the desired quantile ilke this SELECT approx_quantile(digest_times, 0.99) FROM pg_stat_statementsWhat do you guys think? Cheers,\",\n \"msg_date\": \"Thu, 31 Oct 2019 12:51:17 -0300\",\n \"msg_from\": \"Igor Calabria <igor.calabria@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Adding percentile metrics to pg_stat_statements module\"\n },\n {\n \"msg_contents\": \"čt 31. 10. 2019 v 16:51 odesílatel Igor Calabria <igor.calabria@gmail.com>\\nnapsal:\\n\\n> Hi everyone,\\n>\\n> I was taking a look at pg_stat_statements module and noticed that it does\\n> not collect any percentile metrics. I believe that It would be really handy\\n> to have those available and I'd love to contribute with this feature.\\n>\\n> The basic idea is to accumulate the the query execution times using an\\n> approximation structure like q-digest or t-digest and add those results to\\n> the pg_stat_statements table as fixed columns. Something like this\\n>\\n> p90_time:\\n> p95_time:\\n> p99_time:\\n> p70_time:\\n> ...\\n>\\n> Another solution is to persist de digest structure in a binary column and\\n> use a function to extract the desired quantile ilke this SELECT\\n> approx_quantile(digest_times, 0.99) FROM pg_stat_statements\\n>\\n> What do you guys think?\\n>\\n\\n+ the idea\\n\\nBut I am not sure about performance and memory overhead\\n\\nPavel\\n\\n> Cheers,\\n>\\n>\\n\\nčt 31. 10. 2019 v 16:51 odesílatel Igor Calabria <igor.calabria@gmail.com> napsal:Hi everyone,I was taking a look at pg_stat_statements module and noticed that it does not collect any percentile metrics. I believe that It would be really handy to have those available and I'd love to contribute with this feature. The basic idea is to accumulate the the query execution times using an approximation structure like q-digest or t-digest and add those results to the pg_stat_statements table as fixed columns. Something like thisp90_time:p95_time:p99_time:p70_time:...Another solution is to persist de digest structure in a binary column and use a function to extract the desired quantile ilke this SELECT approx_quantile(digest_times, 0.99) FROM pg_stat_statementsWhat do you guys think? + the ideaBut I am not sure about performance and memory overheadPavelCheers,\",\n \"msg_date\": \"Thu, 31 Oct 2019 17:36:27 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Adding percentile metrics to pg_stat_statements module\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 12:51:17PM -0300, Igor Calabria wrote:\\n>Hi everyone,\\n>\\n>I was taking a look at pg_stat_statements module and noticed that it does\\n>not collect any percentile metrics. I believe that It would be really handy\\n>to have those available and I'd love to contribute with this feature.\\n>\\n>The basic idea is to accumulate the the query execution times using an\\n>approximation structure like q-digest or t-digest and add those results to\\n>the pg_stat_statements table as fixed columns. Something like this\\n>\\n>p90_time:\\n>p95_time:\\n>p99_time:\\n>p70_time:\\n>...\\n>\\n>Another solution is to persist de digest structure in a binary column and\\n>use a function to extract the desired quantile ilke this SELECT\\n>approx_quantile(digest_times, 0.99) FROM pg_stat_statements\\n>\\n\\nIMO having some sort of CDF approximation (being a q-digest or t-digest)\\nwould be useful, although it'd probably need to be optional (mostly\\nbecuase of memory consumption).\\n\\nI don't see why we would not store the digests themselves. Storing just\\nsome selected percentiles would be pretty problematic due to losing a\\nlot of information on restart. Also, pg_stat_statements is not a table\\nbut a view on in-memory hash table.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 20:32:47 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Adding percentile metrics to pg_stat_statements module\"\n },\n {\n \"msg_contents\": \"Yeah, I agree that there's no reason to store the digests themselves and I\\nreally liked the idea of it being optional.\\nIf it turns out that memory consumption on real workloads is small enough,\\nit could eventually be turned on by default.\\n\\nI'll start working on patch\\n\\nEm qui, 31 de out de 2019 às 16:32, Tomas Vondra <\\ntomas.vondra@2ndquadrant.com> escreveu:\\n\\n> On Thu, Oct 31, 2019 at 12:51:17PM -0300, Igor Calabria wrote:\\n> >Hi everyone,\\n> >\\n> >I was taking a look at pg_stat_statements module and noticed that it does\\n> >not collect any percentile metrics. I believe that It would be really\\n> handy\\n> >to have those available and I'd love to contribute with this feature.\\n> >\\n> >The basic idea is to accumulate the the query execution times using an\\n> >approximation structure like q-digest or t-digest and add those results to\\n> >the pg_stat_statements table as fixed columns. Something like this\\n> >\\n> >p90_time:\\n> >p95_time:\\n> >p99_time:\\n> >p70_time:\\n> >...\\n> >\\n> >Another solution is to persist de digest structure in a binary column and\\n> >use a function to extract the desired quantile ilke this SELECT\\n> >approx_quantile(digest_times, 0.99) FROM pg_stat_statements\\n> >\\n>\\n> IMO having some sort of CDF approximation (being a q-digest or t-digest)\\n> would be useful, although it'd probably need to be optional (mostly\\n> becuase of memory consumption).\\n>\\n> I don't see why we would not store the digests themselves. Storing just\\n> some selected percentiles would be pretty problematic due to losing a\\n> lot of information on restart. Also, pg_stat_statements is not a table\\n> but a view on in-memory hash table.\\n>\\n>\\n> regards\\n>\\n> --\\n> Tomas Vondra http://www.2ndQuadrant.com\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\nYeah, I agree that there's no reason to store the digests themselves and I really liked the idea of it being optional. If it turns out that memory consumption on real workloads is small enough, it could eventually be turned on by default. I'll start working on patch Em qui, 31 de out de 2019 às 16:32, Tomas Vondra <tomas.vondra@2ndquadrant.com> escreveu:On Thu, Oct 31, 2019 at 12:51:17PM -0300, Igor Calabria wrote:\\n>Hi everyone,\\n>\\n>I was taking a look at pg_stat_statements module and noticed that it does\\n>not collect any percentile metrics. I believe that It would be really handy\\n>to have those available and I'd love to contribute with this feature.\\n>\\n>The basic idea is to accumulate the the query execution times using an\\n>approximation structure like q-digest or t-digest and add those results to\\n>the pg_stat_statements table as fixed columns. Something like this\\n>\\n>p90_time:\\n>p95_time:\\n>p99_time:\\n>p70_time:\\n>...\\n>\\n>Another solution is to persist de digest structure in a binary column and\\n>use a function to extract the desired quantile ilke this SELECT\\n>approx_quantile(digest_times, 0.99) FROM pg_stat_statements\\n>\\n\\nIMO having some sort of CDF approximation (being a q-digest or t-digest)\\nwould be useful, although it'd probably need to be optional (mostly\\nbecuase of memory consumption).\\n\\nI don't see why we would not store the digests themselves. Storing just\\nsome selected percentiles would be pretty problematic due to losing a\\nlot of information on restart. Also, pg_stat_statements is not a table\\nbut a view on in-memory hash table.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 1 Nov 2019 11:11:13 -0300\",\n \"msg_from\": \"Igor Calabria <igor.calabria@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Adding percentile metrics to pg_stat_statements module\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 01, 2019 at 11:11:13AM -0300, Igor Calabria wrote:\\n>Yeah, I agree that there's no reason to store the digests themselves and I\\n>really liked the idea of it being optional.\\n\\nThat's not what I wrote. My point was that we *should* store the digests\\nthemselves, otherwise we just introduce additional errors into the\\nestimates, because it discards the weights/frequencies.\\n\\n>If it turns out that memory consumption on real workloads is small enough,\\n>it could eventually be turned on by default.\\n>\\n\\nMaybe, but it's not just about memory consumption. CPU matters too.\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 15:17:51 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Adding percentile metrics to pg_stat_statements module\"\n },\n {\n \"msg_contents\": \">\\n> That's not what I wrote. My point was that we *should* store the digests\\n> themselves, otherwise we just introduce additional errors into the\\n> estimates, because it discards the weights/frequencies.\\n\\n\\nSorry. I meant to write \\\"no reason to *not* store the digests\\\"\\n\\n\\nEm sex, 1 de nov de 2019 às 11:17, Tomas Vondra <\\ntomas.vondra@2ndquadrant.com> escreveu:\\n\\n> On Fri, Nov 01, 2019 at 11:11:13AM -0300, Igor Calabria wrote:\\n> >Yeah, I agree that there's no reason to store the digests themselves and I\\n> >really liked the idea of it being optional.\\n>\\n> That's not what I wrote. My point was that we *should* store the digests\\n> themselves, otherwise we just introduce additional errors into the\\n> estimates, because it discards the weights/frequencies.\\n>\\n> >If it turns out that memory consumption on real workloads is small enough,\\n> >it could eventually be turned on by default.\\n> >\\n>\\n> Maybe, but it's not just about memory consumption. CPU matters too.\\n>\\n> regards\\n>\\n> --\\n> Tomas Vondra http://www.2ndQuadrant.com\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\nThat's not what I wrote. My point was that we *should* store the digests\\nthemselves, otherwise we just introduce additional errors into the\\nestimates, because it discards the weights/frequencies.Sorry. I meant to write \\\"no reason to not store the digests\\\" Em sex, 1 de nov de 2019 às 11:17, Tomas Vondra <tomas.vondra@2ndquadrant.com> escreveu:On Fri, Nov 01, 2019 at 11:11:13AM -0300, Igor Calabria wrote:\\n>Yeah, I agree that there's no reason to store the digests themselves and I\\n>really liked the idea of it being optional.\\n\\nThat's not what I wrote. My point was that we *should* store the digests\\nthemselves, otherwise we just introduce additional errors into the\\nestimates, because it discards the weights/frequencies.\\n\\n>If it turns out that memory consumption on real workloads is small enough,\\n>it could eventually be turned on by default.\\n>\\n\\nMaybe, but it's not just about memory consumption. CPU matters too.\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 1 Nov 2019 13:05:02 -0300\",\n \"msg_from\": \"Igor Calabria <igor.calabria@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Adding percentile metrics to pg_stat_statements module\"\n },\n {\n \"msg_contents\": \"On 10/31/19 8:32 PM, Tomas Vondra wrote:\\n> IMO having some sort of CDF approximation (being a q-digest or t-digest)\\n> would be useful, although it'd probably need to be optional (mostly\\n> becuase of memory consumption).\\n\\n+1, I like this idea. If we are afraid of CPU cost we could imagine some kind of\\nsampling or add the possibility to collect only for a specific queryid.\\n\\nI dreamed of this kind of feature for PoWA. Thus, it could make possible to\\ncompare CDF between two days for example, before and after introducing a change.\\n\\nRegards,\\n\\n-- \\nAdrien NAYRAT\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 2 Nov 2019 10:23:49 +0100\",\n \"msg_from\": \"Adrien Nayrat <adrien.nayrat@anayrat.info>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Adding percentile metrics to pg_stat_statements module\"\n },\n {\n \"msg_contents\": \"Hello\\n\\n\\nWhile looking at slow queries on pg_stat_statements. I was looking for percentile fields..\\n\\n\\nIf we are worried about CPU cost, maybe it could be useful to turn in on when you have a high stddev_exec_time for the query ?\\n\\n\\nRegards,\\n\\n________________________________\\nDe : Adrien Nayrat <adrien.nayrat@anayrat.info>\\nEnvoyé : samedi 2 novembre 2019 10:23:49\\nÀ : Tomas Vondra; Igor Calabria\\nCc : pgsql-hackers@postgresql.org\\nObjet : Re: Adding percentile metrics to pg_stat_statements module\\n\\nOn 10/31/19 8:32 PM, Tomas Vondra wrote:\\n> IMO having some sort of CDF approximation (being a q-digest or t-digest)\\n> would be useful, although it'd probably need to be optional (mostly\\n> becuase of memory consumption).\\n\\n+1, I like this idea. If we are afraid of CPU cost we could imagine some kind of\\nsampling or add the possibility to collect only for a specific queryid.\\n\\nI dreamed of this kind of feature for PoWA. Thus, it could make possible to\\ncompare CDF between two days for example, before and after introducing a change.\\n\\nRegards,\\n\\n--\\nAdrien NAYRAT\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nHello\\n\\n\\nWhile looking at slow queries on pg_stat_statements. I was looking for percentile fields.. \\n\\n\\nIf we are worried about CPU cost, maybe it could be useful to turn in on when you have a high stddev_exec_time for the query ?\\n\\n\\nRegards,\\n\\n\\nDe : Adrien Nayrat <adrien.nayrat@anayrat.info>\\nEnvoyé : samedi 2 novembre 2019 10:23:49\\nÀ : Tomas Vondra; Igor Calabria\\nCc : pgsql-hackers@postgresql.org\\nObjet : Re: Adding percentile metrics to pg_stat_statements module\\n \\n\\n\\n\\nOn 10/31/19 8:32 PM, Tomas Vondra wrote:\\n> IMO having some sort of CDF approximation (being a q-digest or t-digest)\\n> would be useful, although it'd probably need to be optional (mostly\\n> becuase of memory consumption).\\n\\n+1, I like this idea. If we are afraid of CPU cost we could imagine some kind of\\nsampling or add the possibility to collect only for a specific queryid.\\n\\nI dreamed of this kind of feature for PoWA. Thus, it could make possible to\\ncompare CDF between two days for example, before and after introducing a change.\\n\\nRegards,\\n\\n-- \\nAdrien NAYRAT\",\n \"msg_date\": \"Mon, 5 Jun 2023 08:05:10 +0000\",\n \"msg_from\": \"benoit <benoit@hopsandfork.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Adding percentile metrics to pg_stat_statements module\"\n }\n]","isConversational":false}}},{"rowIdx":49379,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi\n\nlong time we are think how to allow add some custom commands in psql. I had\na following idea\n\n1. psql can has special buffer for custom queries. This buffer can be\nfilled by special command \\gdefq. This command will have two parameters -\nname and number of arguments.\n\nsome like\n\nselect * from pg_class where relname = :'_first' \\gdefcq m1 1\nselect * from pg_class where relnamespace = :_first::regnamespace and\nrename = :'_second' \\gdefcq m1 2\n\nthe custom queries can be executed via doubled backslash like\n\n\\\\m1 pg_proc\n\\\\m1 pg_catalog pg_proc\n\nthe runtime will count number of parameters and chose variant with selected\nname and same number of arguments. Next, it save parameters to variables\nlike _first, _second. Last step is query execution.\n\nWhat do you think about this?\n\nRegards\n\nPavel\n\nHilong time we are think how to allow add some custom commands in psql. I had a following idea1. psql can has special buffer for custom queries. This buffer can be filled by special command \\gdefq. This command will have two parameters - name and number of arguments.some like select * from pg_class where relname = :'_first' \\gdefcq m1 1select * from pg_class where relnamespace = :_first::regnamespace and rename = :'_second' \\gdefcq m1 2the custom queries can be executed via doubled backslash like\\\\m1 pg_proc\\\\m1 pg_catalog pg_procthe runtime will count number of parameters and chose variant with selected name and same number of arguments. Next, it save parameters to variables like _first, _second. Last step is query execution.What do you think about this?RegardsPavel","msg_date":"Thu, 31 Oct 2019 17:48:22 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":true,"msg_subject":"idea - proposal - defining own psql commands"}],"string":"[\n {\n \"msg_contents\": \"Hi\\n\\nlong time we are think how to allow add some custom commands in psql. I had\\na following idea\\n\\n1. psql can has special buffer for custom queries. This buffer can be\\nfilled by special command \\\\gdefq. This command will have two parameters -\\nname and number of arguments.\\n\\nsome like\\n\\nselect * from pg_class where relname = :'_first' \\\\gdefcq m1 1\\nselect * from pg_class where relnamespace = :_first::regnamespace and\\nrename = :'_second' \\\\gdefcq m1 2\\n\\nthe custom queries can be executed via doubled backslash like\\n\\n\\\\\\\\m1 pg_proc\\n\\\\\\\\m1 pg_catalog pg_proc\\n\\nthe runtime will count number of parameters and chose variant with selected\\nname and same number of arguments. Next, it save parameters to variables\\nlike _first, _second. Last step is query execution.\\n\\nWhat do you think about this?\\n\\nRegards\\n\\nPavel\\n\\nHilong time we are think how to allow add some custom commands in psql. I had a following idea1. psql can has special buffer for custom queries. This buffer can be filled by special command \\\\gdefq. This command will have two parameters - name and number of arguments.some like select * from pg_class where relname = :'_first' \\\\gdefcq m1 1select * from pg_class where relnamespace = :_first::regnamespace and rename = :'_second' \\\\gdefcq m1 2the custom queries can be executed via doubled backslash like\\\\\\\\m1 pg_proc\\\\\\\\m1 pg_catalog pg_procthe runtime will count number of parameters and chose variant with selected name and same number of arguments. Next, it save parameters to variables like _first, _second. Last step is query execution.What do you think about this?RegardsPavel\",\n \"msg_date\": \"Thu, 31 Oct 2019 17:48:22 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"idea - proposal - defining own psql commands\"\n }\n]","isConversational":false}}},{"rowIdx":49380,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"This small patch authored by my colleague Craig Ringer enhances\nTestlib's command_fails_like by allowing the passing of extra keyword\ntype arguments. The keyword initially recognized is 'extra_ipcrun_opts'.\nThe value for this keyword needs to be an array, and is passed through\nto the call to IPC::Run.\n\nSome patches I will be submitting shortly rely on this enhancement.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 31 Oct 2019 13:02:58 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"TestLib::command_fails_like enhancement"},{"msg_contents":"On Thu, Oct 31, 2019 at 01:02:58PM -0400, Andrew Dunstan wrote:\n> This small patch authored by my colleague Craig Ringer enhances\n> Testlib's command_fails_like by allowing the passing of extra keyword\n> type arguments. The keyword initially recognized is 'extra_ipcrun_opts'.\n> The value for this keyword needs to be an array, and is passed through\n> to the call to IPC::Run.\n\nWhy not.\n\n> Some patches I will be submitting shortly rely on this enhancement.\n\nAnything submitted yet or any examples? I was just wondering in which\ncase it mattered.\n--\nMichael","msg_date":"Sat, 2 Nov 2019 16:44:40 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\n\nOn 10/31/19 10:02 AM, Andrew Dunstan wrote:\n> \n> This small patch authored by my colleague Craig Ringer enhances\n> Testlib's command_fails_like by allowing the passing of extra keyword\n> type arguments. The keyword initially recognized is 'extra_ipcrun_opts'.\n> The value for this keyword needs to be an array, and is passed through\n> to the call to IPC::Run.\n\nHi Andrew, a few code review comments:\n\nThe POD documentation for this function should be updated to include a \ndescription of the %kwargs argument list.\n\nSince command_fails_like is patterned on command_like, perhaps you \nshould make this change to both of them, even if you only originally \nintend to use the new functionality in command_fails_like. I'm not sure \nof this, though, since I haven't seen any example usage yet.\n\nI'm vaguely bothered by having %kwargs gobble up the remaining function \narguments, not because it isn't a perl-ish thing to do, but because none \nof the other functions in this module do anything similar. The function \ncheck_mode_recursive takes an optional $ignore_list array reference as \nits last argument. Perhaps command_fails_like could follow that pattern \nby taking an optional $kwargs hash reference.\n\n-- \nMark Dilger\n\n\n","msg_date":"Thu, 7 Nov 2019 14:28:24 -0800","msg_from":"Mark Dilger <hornschnorter@gmail.com>","msg_from_op":false,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"On Fri, 8 Nov 2019 at 06:28, Mark Dilger <hornschnorter@gmail.com> wrote:\n\n>\n>\n> On 10/31/19 10:02 AM, Andrew Dunstan wrote:\n> >\n> > This small patch authored by my colleague Craig Ringer enhances\n> > Testlib's command_fails_like by allowing the passing of extra keyword\n> > type arguments. The keyword initially recognized is 'extra_ipcrun_opts'.\n> > The value for this keyword needs to be an array, and is passed through\n> > to the call to IPC::Run.\n>\n> Hi Andrew, a few code review comments:\n>\n> The POD documentation for this function should be updated to include a\n> description of the %kwargs argument list.\n>\n> Since command_fails_like is patterned on command_like, perhaps you\n> should make this change to both of them, even if you only originally\n> intend to use the new functionality in command_fails_like. I'm not sure\n> of this, though, since I haven't seen any example usage yet.\n>\n> I'm vaguely bothered by having %kwargs gobble up the remaining function\n> arguments, not because it isn't a perl-ish thing to do, but because none\n> of the other functions in this module do anything similar. The function\n> check_mode_recursive takes an optional $ignore_list array reference as\n> its last argument. Perhaps command_fails_like could follow that pattern\n> by taking an optional $kwargs hash reference.\n>\n\nYeah, that's probably sensible.\n\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\n\nOn Fri, 8 Nov 2019 at 06:28, Mark Dilger <hornschnorter@gmail.com> wrote:\n\nOn 10/31/19 10:02 AM, Andrew Dunstan wrote:\n> \n> This small patch authored by my colleague Craig Ringer enhances\n> Testlib's command_fails_like by allowing the passing of extra keyword\n> type arguments. The keyword initially recognized is 'extra_ipcrun_opts'.\n> The value for this keyword needs to be an array, and is passed through\n> to the call to IPC::Run.\n\nHi Andrew, a few code review comments:\n\nThe POD documentation for this function should be updated to include a \ndescription of the %kwargs argument list.\n\nSince command_fails_like is patterned on command_like, perhaps you \nshould make this change to both of them, even if you only originally \nintend to use the new functionality in command_fails_like. I'm not sure \nof this, though, since I haven't seen any example usage yet.\n\nI'm vaguely bothered by having %kwargs gobble up the remaining function \narguments, not because it isn't a perl-ish thing to do, but because none \nof the other functions in this module do anything similar. The function \ncheck_mode_recursive takes an optional $ignore_list array reference as \nits last argument. Perhaps command_fails_like could follow that pattern \nby taking an optional $kwargs hash reference.Yeah, that's probably sensible. -- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise","msg_date":"Fri, 8 Nov 2019 14:16:28 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\nOn 11/8/19 1:16 AM, Craig Ringer wrote:\n> On Fri, 8 Nov 2019 at 06:28, Mark Dilger <hornschnorter@gmail.com\n> <mailto:hornschnorter@gmail.com>> wrote:\n>\n>\n>\n> On 10/31/19 10:02 AM, Andrew Dunstan wrote:\n> >\n> > This small patch authored by my colleague Craig Ringer enhances\n> > Testlib's command_fails_like by allowing the passing of extra\n> keyword\n> > type arguments. The keyword initially recognized is\n> 'extra_ipcrun_opts'.\n> > The value for this keyword needs to be an array, and is passed\n> through\n> > to the call to IPC::Run.\n>\n> Hi Andrew, a few code review comments:\n>\n> The POD documentation for this function should be updated to\n> include a\n> description of the %kwargs argument list.\n>\n> Since command_fails_like is patterned on command_like, perhaps you\n> should make this change to both of them, even if you only originally\n> intend to use the new functionality in command_fails_like. I'm\n> not sure\n> of this, though, since I haven't seen any example usage yet.\n>\n> I'm vaguely bothered by having %kwargs gobble up the remaining\n> function\n> arguments, not because it isn't a perl-ish thing to do, but\n> because none\n> of the other functions in this module do anything similar. The\n> function\n> check_mode_recursive takes an optional $ignore_list array\n> reference as\n> its last argument. Perhaps command_fails_like could follow that\n> pattern\n> by taking an optional $kwargs hash reference.\n>\n>\n> Yeah, that's probably sensible. \n>\n>\n>\n\n\nOK, I will rework it taking these comments into account. Thanks for the\ncomments Mark.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Fri, 8 Nov 2019 09:33:20 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\n\nOn 11/8/19 6:33 AM, Andrew Dunstan wrote:\n> \n> On 11/8/19 1:16 AM, Craig Ringer wrote:\n>> On Fri, 8 Nov 2019 at 06:28, Mark Dilger <hornschnorter@gmail.com\n>> <mailto:hornschnorter@gmail.com>> wrote:\n>>\n>>\n>>\n>> On 10/31/19 10:02 AM, Andrew Dunstan wrote:\n>> >\n>> > This small patch authored by my colleague Craig Ringer enhances\n>> > Testlib's command_fails_like by allowing the passing of extra\n>> keyword\n>> > type arguments. The keyword initially recognized is\n>> 'extra_ipcrun_opts'.\n>> > The value for this keyword needs to be an array, and is passed\n>> through\n>> > to the call to IPC::Run.\n>>\n>> Hi Andrew, a few code review comments:\n>>\n>> The POD documentation for this function should be updated to\n>> include a\n>> description of the %kwargs argument list.\n>>\n>> Since command_fails_like is patterned on command_like, perhaps you\n>> should make this change to both of them, even if you only originally\n>> intend to use the new functionality in command_fails_like. I'm\n>> not sure\n>> of this, though, since I haven't seen any example usage yet.\n>>\n>> I'm vaguely bothered by having %kwargs gobble up the remaining\n>> function\n>> arguments, not because it isn't a perl-ish thing to do, but\n>> because none\n>> of the other functions in this module do anything similar. The\n>> function\n>> check_mode_recursive takes an optional $ignore_list array\n>> reference as\n>> its last argument. Perhaps command_fails_like could follow that\n>> pattern\n>> by taking an optional $kwargs hash reference.\n>>\n>>\n>> Yeah, that's probably sensible.\n>>\n>>\n>>\n> \n> \n> OK, I will rework it taking these comments into account. Thanks for the\n> comments Mark.\n\nI'd be happy to see the regression tests you are writing sooner than \nthat, if you don't mind posting them. It's hard to do a proper review \nfor you without a better sense of where you are going with these changes.\n\n-- \nMark Dilger\n\n\n","msg_date":"Fri, 8 Nov 2019 08:25:50 -0800","msg_from":"Mark Dilger <hornschnorter@gmail.com>","msg_from_op":false,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\nOn 11/8/19 11:25 AM, Mark Dilger wrote:\n>\n>\n> On 11/8/19 6:33 AM, Andrew Dunstan wrote:\n>>\n>> On 11/8/19 1:16 AM, Craig Ringer wrote:\n>>> On Fri, 8 Nov 2019 at 06:28, Mark Dilger <hornschnorter@gmail.com\n>>> <mailto:hornschnorter@gmail.com>> wrote:\n>>>\n>>>\n>>>\n>>> On 10/31/19 10:02 AM, Andrew Dunstan wrote:\n>>> >\n>>> > This small patch authored by my colleague Craig Ringer enhances\n>>> > Testlib's command_fails_like by allowing the passing of extra\n>>> keyword\n>>> > type arguments. The keyword initially recognized is\n>>> 'extra_ipcrun_opts'.\n>>> > The value for this keyword needs to be an array, and is passed\n>>> through\n>>> > to the call to IPC::Run.\n>>>\n>>> Hi Andrew, a few code review comments:\n>>>\n>>> The POD documentation for this function should be updated to\n>>> include a\n>>> description of the %kwargs argument list.\n>>>\n>>> Since command_fails_like is patterned on command_like, perhaps you\n>>> should make this change to both of them, even if you only\n>>> originally\n>>> intend to use the new functionality in command_fails_like. I'm\n>>> not sure\n>>> of this, though, since I haven't seen any example usage yet.\n>>>\n>>> I'm vaguely bothered by having %kwargs gobble up the remaining\n>>> function\n>>> arguments, not because it isn't a perl-ish thing to do, but\n>>> because none\n>>> of the other functions in this module do anything similar. The\n>>> function\n>>> check_mode_recursive takes an optional $ignore_list array\n>>> reference as\n>>> its last argument. Perhaps command_fails_like could follow that\n>>> pattern\n>>> by taking an optional $kwargs hash reference.\n>>>\n>>>\n>>> Yeah, that's probably sensible.\n>>>\n>>>\n>>>\n>>\n>>\n>> OK, I will rework it taking these comments into account. Thanks for the\n>> comments Mark.\n>\n> I'd be happy to see the regression tests you are writing sooner than\n> that, if you don't mind posting them. It's hard to do a proper review\n> for you without a better sense of where you are going with these changes.\n\n\nThis will need to be rewritten in light of the above, but see\n<https://www.postgresql.org/message-id/87b1e36b-e36a-add5-1a9b-9fa34914a256@2ndQuadrant.com>\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Fri, 8 Nov 2019 12:22:05 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\n\nOn 11/8/19 9:22 AM, Andrew Dunstan wrote:\n...\n> This will need to be rewritten in light of the above, but see\n> <https://www.postgresql.org/message-id/87b1e36b-e36a-add5-1a9b-9fa34914a256@2ndQuadrant.com>\n\nThanks for the reference. Having read your motivating example, this new \nreview reverses some of what I suggested in the prior review.\n\n\nIn the existing TestLib.pm, there are eight occurrences of nearly \nidentical usages of IPC::Run scattered through similar functions:\n\nrun_command:\n my $result = IPC::Run::run $cmd, '>', \\$stdout, '2>', \\$stderr;\n\ncheck_pg_config:\n my $result = IPC::Run::run [ 'pg_config', '--includedir' ], '>',\n \\$stdout, '2>', \\$stderr\n or die \"could not execute pg_config\";\n\nprogram_help_ok:\n my $result = IPC::Run::run [ $cmd, '--help' ], '>', \\$stdout, '2>',\n \\$stderr;\n\nprogram_version_ok:\n my $result = IPC::Run::run [ $cmd, '--version' ], '>', \\$stdout, '2>',\n \\$stderr;\n\nprogram_options_handling_ok:\n my $result = IPC::Run::run [ $cmd, '--not-a-valid-option' ], '>',\n \\$stdout,\n '2>', \\$stderr;\n\ncommand_like:\n my $result = IPC::Run::run $cmd, '>', \\$stdout, '2>', \\$stderr;\n\ncommand_like_safe:\n my $result = IPC::Run::run $cmd, '>', $stdoutfile, '2>', $stderrfile;\n\ncommand_fails_like:\n my $result = IPC::Run::run $cmd, '>', \\$stdout, '2>', \\$stderr,\n @extra_ipcrun_opts;\n\nOne rational motive for designing TestLib with so much code duplication \nis to make the tests that use the library easier to read:\n\n command_like_safe(foo);\n command_like(bar);\n command_fails_like(baz);\n\nwhich is easier to understand than:\n\n command_like(foo, failure_mode => safe);\n command_like(bar);\n command_like(baz, failure => expected);\n\nand so forth.\n\nIn line with that thinking, perhaps you should just create:\n\n command_fails_without_tty_like(foo)\n\nand be done, or perhaps:\n\n command_fails_like(foo, tty => 'closed')\n\nand still preserve some of the test readability. Will anyone like the \nreadability of your tests if you have:\n\n command_fails_like(foo, extra_ipcrun_opts => ['<pty<', \\$eof_in]) ?\n\nAdmittedly, \"foo\", \"bar\", and \"baz\" above are shorthand notation for \nthings in practice that are already somewhat hard to read, as in:\n\n command_fails_like(\n [ 'pg_dump', 'qqq', 'abc' ],\n qr/\\Qpg_dump: error: too many command-line arguments (first is \n\"abc\")\\E/,\n 'pg_dump: too many command-line arguments');\n\nbut adding more to that cruft just makes it worse. Right?\n\n-- \nMark Dilger\n\n\n","msg_date":"Fri, 8 Nov 2019 13:40:06 -0800","msg_from":"Mark Dilger <hornschnorter@gmail.com>","msg_from_op":false,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\nOn 11/8/19 4:40 PM, Mark Dilger wrote:\n>\n>\n> On 11/8/19 9:22 AM, Andrew Dunstan wrote:\n> ...\n>> This will need to be rewritten in light of the above, but see\n>> <https://www.postgresql.org/message-id/87b1e36b-e36a-add5-1a9b-9fa34914a256@2ndQuadrant.com>\n>>\n>\n> Thanks for the reference. Having read your motivating example, this\n> new review reverses some of what I suggested in the prior review.\n>\n>\n> In the existing TestLib.pm, there are eight occurrences of nearly\n> identical usages of IPC::Run scattered through similar functions:\n>\n>\n[snip]\n\n\n>\n> One rational motive for designing TestLib with so much code\n> duplication is to make the tests that use the library easier to read:\n>\n> command_like_safe(foo);\n> command_like(bar);\n> command_fails_like(baz);\n>\n> which is easier to understand than:\n>\n> command_like(foo, failure_mode => safe);\n> command_like(bar);\n> command_like(baz, failure => expected);\n>\n> and so forth.\n>\n> In line with that thinking, perhaps you should just create:\n>\n> command_fails_without_tty_like(foo)\n>\n> and be done, or perhaps:\n>\n> command_fails_like(foo, tty => 'closed')\n>\n> and still preserve some of the test readability. Will anyone like the\n> readability of your tests if you have:\n>\n> command_fails_like(foo, extra_ipcrun_opts => ['<pty<', \\$eof_in]) ?\n>\n> Admittedly, \"foo\", \"bar\", and \"baz\" above are shorthand notation for\n> things in practice that are already somewhat hard to read, as in:\n>\n> command_fails_like(\n> [ 'pg_dump', 'qqq', 'abc' ],\n> qr/\\Qpg_dump: error: too many command-line arguments (first is\n> \"abc\")\\E/,\n> 'pg_dump: too many command-line arguments');\n>\n> but adding more to that cruft just makes it worse. Right?\n>\n\nOK, I agree that we're getting rather baroque here. I could go with your\nsuggestion of YA function, or possibly a solution that simple passes any\nextra arguments straight through to IPC::Run::run(), e.g.\n\ncommand_fails_like(\n [ 'pg_dump', 'qqq', 'abc' ],\n qr/\\Qpg_dump: error: too many command-line arguments (first is\n\"abc\")\\E/,\n 'pg_dump: too many command-line arguments',\n '<pty<', \\$eof_in);\n\nThat means we're not future-proofing the function - we'll never be able\nto add more arguments to it, but I'm not really certain that matters\nanyway. I should note that perlcritic whines about subroutines with too\nmany arguments, so making provision for more seems unnecessary anyway.\n\nI don't think this is worth spending a huge amount of time on, we've\nalready spent more time discussing it than it would take to implement\neither solution.\n\n\ncheers\n\n\nandrew\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Sat, 9 Nov 2019 08:25:10 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\nOn 11/9/19 8:25 AM, Andrew Dunstan wrote:\n> OK, I agree that we're getting rather baroque here. I could go with your\n> suggestion of YA function, or possibly a solution that simple passes any\n> extra arguments straight through to IPC::Run::run(), e.g.\n>\n> command_fails_like(\n> [ 'pg_dump', 'qqq', 'abc' ],\n> qr/\\Qpg_dump: error: too many command-line arguments (first is\n> \"abc\")\\E/,\n> 'pg_dump: too many command-line arguments',\n> '<pty<', \\$eof_in);\n>\n> That means we're not future-proofing the function - we'll never be able\n> to add more arguments to it, but I'm not really certain that matters\n> anyway. I should note that perlcritic whines about subroutines with too\n> many arguments, so making provision for more seems unnecessary anyway.\n>\n> I don't think this is worth spending a huge amount of time on, we've\n> already spent more time discussing it than it would take to implement\n> either solution.\n>\n>\n\nOn further consideration, I'm wondering why we don't just\nunconditionally use a closed input pty for all these functions (except\nrun_log). None of them use any input, and making the client worry about\nwhether or not to close it seems something of an abstraction break.\nThere would be no API change at all involved in this case, just a bit of\nextra cleanliness. Would need testing on Windows, I'll go and do that now.\n\n\nThoughts?\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Mon, 11 Nov 2019 11:48:25 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\n\nOn 11/11/19 8:48 AM, Andrew Dunstan wrote:\n> \n> On 11/9/19 8:25 AM, Andrew Dunstan wrote:\n>> OK, I agree that we're getting rather baroque here. I could go with your\n>> suggestion of YA function, or possibly a solution that simple passes any\n>> extra arguments straight through to IPC::Run::run(), e.g.\n>>\n>> command_fails_like(\n>> [ 'pg_dump', 'qqq', 'abc' ],\n>> qr/\\Qpg_dump: error: too many command-line arguments (first is\n>> \"abc\")\\E/,\n>> 'pg_dump: too many command-line arguments',\n>> '<pty<', \\$eof_in);\n>>\n>> That means we're not future-proofing the function - we'll never be able\n>> to add more arguments to it, but I'm not really certain that matters\n>> anyway. I should note that perlcritic whines about subroutines with too\n>> many arguments, so making provision for more seems unnecessary anyway.\n>>\n>> I don't think this is worth spending a huge amount of time on, we've\n>> already spent more time discussing it than it would take to implement\n>> either solution.\n>>\n>>\n> \n> On further consideration, I'm wondering why we don't just\n> unconditionally use a closed input pty for all these functions (except\n> run_log). None of them use any input, and making the client worry about\n> whether or not to close it seems something of an abstraction break.\n> There would be no API change at all involved in this case, just a bit of\n> extra cleanliness. Would need testing on Windows, I'll go and do that now.\n> \n> \n> Thoughts?\n\nThat sounds a lot better than your previous patch.\n\nPostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you change \nall the invocations in TestLib to close input pty, should you do the \nsame for PostgresNode? I don't have a strong argument for doing so, but \nit seems cleaner to have both libraries invoking commands under \nidentical conditions, such that if commands were borrowed from one \nlibrary and called from the other they would behave the same.\n\nPostgresNode already uses TestLib, so perhaps setting up the environment \ncan be abstracted into something, perhaps TestLib::run, and then used \neverywhere that IPC::Run::run currently is used.\n\n\n-- \nMark Dilger\n\n\n","msg_date":"Mon, 11 Nov 2019 10:27:35 -0800","msg_from":"Mark Dilger <hornschnorter@gmail.com>","msg_from_op":false,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\nOn 11/11/19 1:27 PM, Mark Dilger wrote:\n>\n>\n> On 11/11/19 8:48 AM, Andrew Dunstan wrote:\n>>\n>> On 11/9/19 8:25 AM, Andrew Dunstan wrote:\n>>> OK, I agree that we're getting rather baroque here. I could go with\n>>> your\n>>> suggestion of YA function, or possibly a solution that simple passes\n>>> any\n>>> extra arguments straight through to IPC::Run::run(), e.g.\n>>>\n>>> command_fails_like(\n>>> [ 'pg_dump', 'qqq', 'abc' ],\n>>> qr/\\Qpg_dump: error: too many command-line arguments (first is\n>>> \"abc\")\\E/,\n>>> 'pg_dump: too many command-line arguments',\n>>> '<pty<', \\$eof_in);\n>>>\n>>> That means we're not future-proofing the function - we'll never be able\n>>> to add more arguments to it, but I'm not really certain that matters\n>>> anyway. I should note that perlcritic whines about subroutines with too\n>>> many arguments, so making provision for more seems unnecessary anyway.\n>>>\n>>> I don't think this is worth spending a huge amount of time on, we've\n>>> already spent more time discussing it than it would take to implement\n>>> either solution.\n>>>\n>>>\n>>\n>> On further consideration, I'm wondering why we don't just\n>> unconditionally use a closed input pty for all these functions (except\n>> run_log). None of them use any input, and making the client worry about\n>> whether or not to close it seems something of an abstraction break.\n>> There would be no API change at all involved in this case, just a bit of\n>> extra cleanliness. Would need testing on Windows, I'll go and do that\n>> now.\n>>\n>>\n>> Thoughts?\n>\n> That sounds a lot better than your previous patch.\n>\n> PostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you\n> change all the invocations in TestLib to close input pty, should you\n> do the same for PostgresNode? I don't have a strong argument for\n> doing so, but it seems cleaner to have both libraries invoking\n> commands under identical conditions, such that if commands were\n> borrowed from one library and called from the other they would behave\n> the same.\n>\n> PostgresNode already uses TestLib, so perhaps setting up the\n> environment can be abstracted into something, perhaps TestLib::run,\n> and then used everywhere that IPC::Run::run currently is used.\n\n\n\nI don't think we need to do that. In the case of the PostgresNode.pm\nuses we know what the executable is, unlike the the TestLib.pm cases.\nThey are our own executables and we don't expect them to be doing\nanything funky with /dev/tty.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Mon, 11 Nov 2019 14:28:23 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\n\nOn 11/11/19 11:28 AM, Andrew Dunstan wrote:\n> \n> On 11/11/19 1:27 PM, Mark Dilger wrote:\n>>\n>>\n>> On 11/11/19 8:48 AM, Andrew Dunstan wrote:\n>>>\n>>> On 11/9/19 8:25 AM, Andrew Dunstan wrote:\n>>>> OK, I agree that we're getting rather baroque here. I could go with\n>>>> your\n>>>> suggestion of YA function, or possibly a solution that simple passes\n>>>> any\n>>>> extra arguments straight through to IPC::Run::run(), e.g.\n>>>>\n>>>> command_fails_like(\n>>>> [ 'pg_dump', 'qqq', 'abc' ],\n>>>> qr/\\Qpg_dump: error: too many command-line arguments (first is\n>>>> \"abc\")\\E/,\n>>>> 'pg_dump: too many command-line arguments',\n>>>> '<pty<', \\$eof_in);\n>>>>\n>>>> That means we're not future-proofing the function - we'll never be able\n>>>> to add more arguments to it, but I'm not really certain that matters\n>>>> anyway. I should note that perlcritic whines about subroutines with too\n>>>> many arguments, so making provision for more seems unnecessary anyway.\n>>>>\n>>>> I don't think this is worth spending a huge amount of time on, we've\n>>>> already spent more time discussing it than it would take to implement\n>>>> either solution.\n>>>>\n>>>>\n>>>\n>>> On further consideration, I'm wondering why we don't just\n>>> unconditionally use a closed input pty for all these functions (except\n>>> run_log). None of them use any input, and making the client worry about\n>>> whether or not to close it seems something of an abstraction break.\n>>> There would be no API change at all involved in this case, just a bit of\n>>> extra cleanliness. Would need testing on Windows, I'll go and do that\n>>> now.\n>>>\n>>>\n>>> Thoughts?\n>>\n>> That sounds a lot better than your previous patch.\n>>\n>> PostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you\n>> change all the invocations in TestLib to close input pty, should you\n>> do the same for PostgresNode? I don't have a strong argument for\n>> doing so, but it seems cleaner to have both libraries invoking\n>> commands under identical conditions, such that if commands were\n>> borrowed from one library and called from the other they would behave\n>> the same.\n>>\n>> PostgresNode already uses TestLib, so perhaps setting up the\n>> environment can be abstracted into something, perhaps TestLib::run,\n>> and then used everywhere that IPC::Run::run currently is used.\n> \n> \n> \n> I don't think we need to do that. In the case of the PostgresNode.pm\n> uses we know what the executable is, unlike the the TestLib.pm cases.\n> They are our own executables and we don't expect them to be doing\n> anything funky with /dev/tty.\n\nOk. I think your proposal sounds fine.\n\n-- \nMark Dilger\n\n\n","msg_date":"Mon, 11 Nov 2019 13:28:37 -0800","msg_from":"Mark Dilger <hornschnorter@gmail.com>","msg_from_op":false,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"On 11/11/19 4:28 PM, Mark Dilger wrote:\n>\n>\n>>>>>\n>>>>\n>>>> On further consideration, I'm wondering why we don't just\n>>>> unconditionally use a closed input pty for all these functions (except\n>>>> run_log). None of them use any input, and making the client worry\n>>>> about\n>>>> whether or not to close it seems something of an abstraction break.\n>>>> There would be no API change at all involved in this case, just a\n>>>> bit of\n>>>> extra cleanliness. Would need testing on Windows, I'll go and do that\n>>>> now.\n>>>>\n>>>>\n>>>> Thoughts?\n>>>\n>>> That sounds a lot better than your previous patch.\n>>>\n>>> PostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you\n>>> change all the invocations in TestLib to close input pty, should you\n>>> do the same for PostgresNode? I don't have a strong argument for\n>>> doing so, but it seems cleaner to have both libraries invoking\n>>> commands under identical conditions, such that if commands were\n>>> borrowed from one library and called from the other they would behave\n>>> the same.\n>>>\n>>> PostgresNode already uses TestLib, so perhaps setting up the\n>>> environment can be abstracted into something, perhaps TestLib::run,\n>>> and then used everywhere that IPC::Run::run currently is used.\n>>\n>>\n>>\n>> I don't think we need to do that. In the case of the PostgresNode.pm\n>> uses we know what the executable is, unlike the the TestLib.pm cases.\n>> They are our own executables and we don't expect them to be doing\n>> anything funky with /dev/tty.\n>\n> Ok. I think your proposal sounds fine.\n\n\n\nHere's a patch for that. The pty stuff crashes and burns on my Windows\ntest box, so there I just set stdin to an empty string via the usual\npipe mechanism.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Mon, 25 Nov 2019 08:08:56 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\n\nOn 11/25/19 5:08 AM, Andrew Dunstan wrote:\n> \n> On 11/11/19 4:28 PM, Mark Dilger wrote:\n>>\n>>\n>>>>>>\n>>>>>\n>>>>> On further consideration, I'm wondering why we don't just\n>>>>> unconditionally use a closed input pty for all these functions (except\n>>>>> run_log). None of them use any input, and making the client worry\n>>>>> about\n>>>>> whether or not to close it seems something of an abstraction break.\n>>>>> There would be no API change at all involved in this case, just a\n>>>>> bit of\n>>>>> extra cleanliness. Would need testing on Windows, I'll go and do that\n>>>>> now.\n>>>>>\n>>>>>\n>>>>> Thoughts?\n>>>>\n>>>> That sounds a lot better than your previous patch.\n>>>>\n>>>> PostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you\n>>>> change all the invocations in TestLib to close input pty, should you\n>>>> do the same for PostgresNode? I don't have a strong argument for\n>>>> doing so, but it seems cleaner to have both libraries invoking\n>>>> commands under identical conditions, such that if commands were\n>>>> borrowed from one library and called from the other they would behave\n>>>> the same.\n>>>>\n>>>> PostgresNode already uses TestLib, so perhaps setting up the\n>>>> environment can be abstracted into something, perhaps TestLib::run,\n>>>> and then used everywhere that IPC::Run::run currently is used.\n>>>\n>>>\n>>>\n>>> I don't think we need to do that. In the case of the PostgresNode.pm\n>>> uses we know what the executable is, unlike the the TestLib.pm cases.\n>>> They are our own executables and we don't expect them to be doing\n>>> anything funky with /dev/tty.\n>>\n>> Ok. I think your proposal sounds fine.\n> \n> \n> \n> Here's a patch for that. The pty stuff crashes and burns on my Windows\n> test box, so there I just set stdin to an empty string via the usual\n> pipe mechanism.\n\nOk, I've reviewed and tested this. It works fine for me on Linux. I\nam not set up to test it on Windows. I think it is ready to commit.\n\nI have one remaining comment about the code, and this is just FYI. I\nwon't quibble with you committing your patch as it currently stands.\n\nYou might consider changing the '\\x04' literal to use a named control\ncharacter, both for readability and portability, as here:\n\n+ use charnames ':full';\n+ @no_stdin = ('<pty<', \\\"\\N{END OF TRANSMISSION}\");\n\nThe only character set I can find where this matters is EBCDIC, in\nwhich the EOT character is 55 rather than 4. Since EBCDIC does not\noccur in the list of supported character sets for postgres, per the\ndocs section 23.3.1, I don't suppose it matters too much. Nor can\nI test how this works on EBCDIC, so I'm mostly guessing that perl\nwould do the right thing there. But, at least to my eyes, it is\nmore immediately clear what the code is doing when the control\ncharacter name is spelled out.\n\n\n-- \nMark Dilger\n\n\n","msg_date":"Mon, 25 Nov 2019 10:56:38 -0800","msg_from":"Mark Dilger <hornschnorter@gmail.com>","msg_from_op":false,"msg_subject":"Re: TestLib::command_fails_like enhancement"},{"msg_contents":"\nOn 11/25/19 1:56 PM, Mark Dilger wrote:\n>\n>\n> On 11/25/19 5:08 AM, Andrew Dunstan wrote:\n>>\n>> On 11/11/19 4:28 PM, Mark Dilger wrote:\n>>>\n>>>\n>>>>>>>\n>>>>>>\n>>>>>> On further consideration, I'm wondering why we don't just\n>>>>>> unconditionally use a closed input pty for all these functions\n>>>>>> (except\n>>>>>> run_log). None of them use any input, and making the client worry\n>>>>>> about\n>>>>>> whether or not to close it seems something of an abstraction break.\n>>>>>> There would be no API change at all involved in this case, just a\n>>>>>> bit of\n>>>>>> extra cleanliness. Would need testing on Windows, I'll go and do\n>>>>>> that\n>>>>>> now.\n>>>>>>\n>>>>>>\n>>>>>> Thoughts?\n>>>>>\n>>>>> That sounds a lot better than your previous patch.\n>>>>>\n>>>>> PostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you\n>>>>> change all the invocations in TestLib to close input pty, should you\n>>>>> do the same for PostgresNode? I don't have a strong argument for\n>>>>> doing so, but it seems cleaner to have both libraries invoking\n>>>>> commands under identical conditions, such that if commands were\n>>>>> borrowed from one library and called from the other they would behave\n>>>>> the same.\n>>>>>\n>>>>> PostgresNode already uses TestLib, so perhaps setting up the\n>>>>> environment can be abstracted into something, perhaps TestLib::run,\n>>>>> and then used everywhere that IPC::Run::run currently is used.\n>>>>\n>>>>\n>>>>\n>>>> I don't think we need to do that. In the case of the PostgresNode.pm\n>>>> uses we know what the executable is, unlike the the TestLib.pm cases.\n>>>> They are our own executables and we don't expect them to be doing\n>>>> anything funky with /dev/tty.\n>>>\n>>> Ok. I think your proposal sounds fine.\n>>\n>>\n>>\n>> Here's a patch for that. The pty stuff crashes and burns on my Windows\n>> test box, so there I just set stdin to an empty string via the usual\n>> pipe mechanism.\n>\n> Ok, I've reviewed and tested this. It works fine for me on Linux. I\n> am not set up to test it on Windows. I think it is ready to commit.\n>\n> I have one remaining comment about the code, and this is just FYI. I\n> won't quibble with you committing your patch as it currently stands.\n>\n> You might consider changing the '\\x04' literal to use a named control\n> character, both for readability and portability, as here:\n>\n> + use charnames ':full';\n> + @no_stdin = ('<pty<', \\\"\\N{END OF TRANSMISSION}\");\n>\n> The only character set I can find where this matters is EBCDIC, in\n> which the EOT character is 55 rather than 4. Since EBCDIC does not\n> occur in the list of supported character sets for postgres, per the\n> docs section 23.3.1, I don't suppose it matters too much. Nor can\n> I test how this works on EBCDIC, so I'm mostly guessing that perl\n> would do the right thing there. But, at least to my eyes, it is\n> more immediately clear what the code is doing when the control\n> character name is spelled out.\n>\n>\n\n\nAgreed, I'll do it that way. This is quite timely, as I just finished\nreworking the patch that relies on it. Thanks for the review.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Mon, 25 Nov 2019 15:07:10 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: TestLib::command_fails_like enhancement"}],"string":"[\n {\n \"msg_contents\": \"This small patch authored by my colleague Craig Ringer enhances\\nTestlib's command_fails_like by allowing the passing of extra keyword\\ntype arguments. The keyword initially recognized is 'extra_ipcrun_opts'.\\nThe value for this keyword needs to be an array, and is passed through\\nto the call to IPC::Run.\\n\\nSome patches I will be submitting shortly rely on this enhancement.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 31 Oct 2019 13:02:58 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 01:02:58PM -0400, Andrew Dunstan wrote:\\n> This small patch authored by my colleague Craig Ringer enhances\\n> Testlib's command_fails_like by allowing the passing of extra keyword\\n> type arguments. The keyword initially recognized is 'extra_ipcrun_opts'.\\n> The value for this keyword needs to be an array, and is passed through\\n> to the call to IPC::Run.\\n\\nWhy not.\\n\\n> Some patches I will be submitting shortly rely on this enhancement.\\n\\nAnything submitted yet or any examples? I was just wondering in which\\ncase it mattered.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 2 Nov 2019 16:44:40 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 10/31/19 10:02 AM, Andrew Dunstan wrote:\\n> \\n> This small patch authored by my colleague Craig Ringer enhances\\n> Testlib's command_fails_like by allowing the passing of extra keyword\\n> type arguments. The keyword initially recognized is 'extra_ipcrun_opts'.\\n> The value for this keyword needs to be an array, and is passed through\\n> to the call to IPC::Run.\\n\\nHi Andrew, a few code review comments:\\n\\nThe POD documentation for this function should be updated to include a \\ndescription of the %kwargs argument list.\\n\\nSince command_fails_like is patterned on command_like, perhaps you \\nshould make this change to both of them, even if you only originally \\nintend to use the new functionality in command_fails_like. I'm not sure \\nof this, though, since I haven't seen any example usage yet.\\n\\nI'm vaguely bothered by having %kwargs gobble up the remaining function \\narguments, not because it isn't a perl-ish thing to do, but because none \\nof the other functions in this module do anything similar. The function \\ncheck_mode_recursive takes an optional $ignore_list array reference as \\nits last argument. Perhaps command_fails_like could follow that pattern \\nby taking an optional $kwargs hash reference.\\n\\n-- \\nMark Dilger\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Nov 2019 14:28:24 -0800\",\n \"msg_from\": \"Mark Dilger <hornschnorter@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"On Fri, 8 Nov 2019 at 06:28, Mark Dilger <hornschnorter@gmail.com> wrote:\\n\\n>\\n>\\n> On 10/31/19 10:02 AM, Andrew Dunstan wrote:\\n> >\\n> > This small patch authored by my colleague Craig Ringer enhances\\n> > Testlib's command_fails_like by allowing the passing of extra keyword\\n> > type arguments. The keyword initially recognized is 'extra_ipcrun_opts'.\\n> > The value for this keyword needs to be an array, and is passed through\\n> > to the call to IPC::Run.\\n>\\n> Hi Andrew, a few code review comments:\\n>\\n> The POD documentation for this function should be updated to include a\\n> description of the %kwargs argument list.\\n>\\n> Since command_fails_like is patterned on command_like, perhaps you\\n> should make this change to both of them, even if you only originally\\n> intend to use the new functionality in command_fails_like. I'm not sure\\n> of this, though, since I haven't seen any example usage yet.\\n>\\n> I'm vaguely bothered by having %kwargs gobble up the remaining function\\n> arguments, not because it isn't a perl-ish thing to do, but because none\\n> of the other functions in this module do anything similar. The function\\n> check_mode_recursive takes an optional $ignore_list array reference as\\n> its last argument. Perhaps command_fails_like could follow that pattern\\n> by taking an optional $kwargs hash reference.\\n>\\n\\nYeah, that's probably sensible.\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\\n\\nOn Fri, 8 Nov 2019 at 06:28, Mark Dilger <hornschnorter@gmail.com> wrote:\\n\\nOn 10/31/19 10:02 AM, Andrew Dunstan wrote:\\n> \\n> This small patch authored by my colleague Craig Ringer enhances\\n> Testlib's command_fails_like by allowing the passing of extra keyword\\n> type arguments. The keyword initially recognized is 'extra_ipcrun_opts'.\\n> The value for this keyword needs to be an array, and is passed through\\n> to the call to IPC::Run.\\n\\nHi Andrew, a few code review comments:\\n\\nThe POD documentation for this function should be updated to include a \\ndescription of the %kwargs argument list.\\n\\nSince command_fails_like is patterned on command_like, perhaps you \\nshould make this change to both of them, even if you only originally \\nintend to use the new functionality in command_fails_like. I'm not sure \\nof this, though, since I haven't seen any example usage yet.\\n\\nI'm vaguely bothered by having %kwargs gobble up the remaining function \\narguments, not because it isn't a perl-ish thing to do, but because none \\nof the other functions in this module do anything similar. The function \\ncheck_mode_recursive takes an optional $ignore_list array reference as \\nits last argument. Perhaps command_fails_like could follow that pattern \\nby taking an optional $kwargs hash reference.Yeah, that's probably sensible. -- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise\",\n \"msg_date\": \"Fri, 8 Nov 2019 14:16:28 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\nOn 11/8/19 1:16 AM, Craig Ringer wrote:\\n> On Fri, 8 Nov 2019 at 06:28, Mark Dilger <hornschnorter@gmail.com\\n> <mailto:hornschnorter@gmail.com>> wrote:\\n>\\n>\\n>\\n> On 10/31/19 10:02 AM, Andrew Dunstan wrote:\\n> >\\n> > This small patch authored by my colleague Craig Ringer enhances\\n> > Testlib's command_fails_like by allowing the passing of extra\\n> keyword\\n> > type arguments. The keyword initially recognized is\\n> 'extra_ipcrun_opts'.\\n> > The value for this keyword needs to be an array, and is passed\\n> through\\n> > to the call to IPC::Run.\\n>\\n> Hi Andrew, a few code review comments:\\n>\\n> The POD documentation for this function should be updated to\\n> include a\\n> description of the %kwargs argument list.\\n>\\n> Since command_fails_like is patterned on command_like, perhaps you\\n> should make this change to both of them, even if you only originally\\n> intend to use the new functionality in command_fails_like. I'm\\n> not sure\\n> of this, though, since I haven't seen any example usage yet.\\n>\\n> I'm vaguely bothered by having %kwargs gobble up the remaining\\n> function\\n> arguments, not because it isn't a perl-ish thing to do, but\\n> because none\\n> of the other functions in this module do anything similar. The\\n> function\\n> check_mode_recursive takes an optional $ignore_list array\\n> reference as\\n> its last argument. Perhaps command_fails_like could follow that\\n> pattern\\n> by taking an optional $kwargs hash reference.\\n>\\n>\\n> Yeah, that's probably sensible. \\n>\\n>\\n>\\n\\n\\nOK, I will rework it taking these comments into account. Thanks for the\\ncomments Mark.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 09:33:20 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 11/8/19 6:33 AM, Andrew Dunstan wrote:\\n> \\n> On 11/8/19 1:16 AM, Craig Ringer wrote:\\n>> On Fri, 8 Nov 2019 at 06:28, Mark Dilger <hornschnorter@gmail.com\\n>> <mailto:hornschnorter@gmail.com>> wrote:\\n>>\\n>>\\n>>\\n>> On 10/31/19 10:02 AM, Andrew Dunstan wrote:\\n>> >\\n>> > This small patch authored by my colleague Craig Ringer enhances\\n>> > Testlib's command_fails_like by allowing the passing of extra\\n>> keyword\\n>> > type arguments. The keyword initially recognized is\\n>> 'extra_ipcrun_opts'.\\n>> > The value for this keyword needs to be an array, and is passed\\n>> through\\n>> > to the call to IPC::Run.\\n>>\\n>> Hi Andrew, a few code review comments:\\n>>\\n>> The POD documentation for this function should be updated to\\n>> include a\\n>> description of the %kwargs argument list.\\n>>\\n>> Since command_fails_like is patterned on command_like, perhaps you\\n>> should make this change to both of them, even if you only originally\\n>> intend to use the new functionality in command_fails_like. I'm\\n>> not sure\\n>> of this, though, since I haven't seen any example usage yet.\\n>>\\n>> I'm vaguely bothered by having %kwargs gobble up the remaining\\n>> function\\n>> arguments, not because it isn't a perl-ish thing to do, but\\n>> because none\\n>> of the other functions in this module do anything similar. The\\n>> function\\n>> check_mode_recursive takes an optional $ignore_list array\\n>> reference as\\n>> its last argument. Perhaps command_fails_like could follow that\\n>> pattern\\n>> by taking an optional $kwargs hash reference.\\n>>\\n>>\\n>> Yeah, that's probably sensible.\\n>>\\n>>\\n>>\\n> \\n> \\n> OK, I will rework it taking these comments into account. Thanks for the\\n> comments Mark.\\n\\nI'd be happy to see the regression tests you are writing sooner than \\nthat, if you don't mind posting them. It's hard to do a proper review \\nfor you without a better sense of where you are going with these changes.\\n\\n-- \\nMark Dilger\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 08:25:50 -0800\",\n \"msg_from\": \"Mark Dilger <hornschnorter@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\nOn 11/8/19 11:25 AM, Mark Dilger wrote:\\n>\\n>\\n> On 11/8/19 6:33 AM, Andrew Dunstan wrote:\\n>>\\n>> On 11/8/19 1:16 AM, Craig Ringer wrote:\\n>>> On Fri, 8 Nov 2019 at 06:28, Mark Dilger <hornschnorter@gmail.com\\n>>> <mailto:hornschnorter@gmail.com>> wrote:\\n>>>\\n>>>\\n>>>\\n>>> On 10/31/19 10:02 AM, Andrew Dunstan wrote:\\n>>> >\\n>>> > This small patch authored by my colleague Craig Ringer enhances\\n>>> > Testlib's command_fails_like by allowing the passing of extra\\n>>> keyword\\n>>> > type arguments. The keyword initially recognized is\\n>>> 'extra_ipcrun_opts'.\\n>>> > The value for this keyword needs to be an array, and is passed\\n>>> through\\n>>> > to the call to IPC::Run.\\n>>>\\n>>> Hi Andrew, a few code review comments:\\n>>>\\n>>> The POD documentation for this function should be updated to\\n>>> include a\\n>>> description of the %kwargs argument list.\\n>>>\\n>>> Since command_fails_like is patterned on command_like, perhaps you\\n>>> should make this change to both of them, even if you only\\n>>> originally\\n>>> intend to use the new functionality in command_fails_like. I'm\\n>>> not sure\\n>>> of this, though, since I haven't seen any example usage yet.\\n>>>\\n>>> I'm vaguely bothered by having %kwargs gobble up the remaining\\n>>> function\\n>>> arguments, not because it isn't a perl-ish thing to do, but\\n>>> because none\\n>>> of the other functions in this module do anything similar. The\\n>>> function\\n>>> check_mode_recursive takes an optional $ignore_list array\\n>>> reference as\\n>>> its last argument. Perhaps command_fails_like could follow that\\n>>> pattern\\n>>> by taking an optional $kwargs hash reference.\\n>>>\\n>>>\\n>>> Yeah, that's probably sensible.\\n>>>\\n>>>\\n>>>\\n>>\\n>>\\n>> OK, I will rework it taking these comments into account. Thanks for the\\n>> comments Mark.\\n>\\n> I'd be happy to see the regression tests you are writing sooner than\\n> that, if you don't mind posting them. It's hard to do a proper review\\n> for you without a better sense of where you are going with these changes.\\n\\n\\nThis will need to be rewritten in light of the above, but see\\n<https://www.postgresql.org/message-id/87b1e36b-e36a-add5-1a9b-9fa34914a256@2ndQuadrant.com>\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 12:22:05 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 11/8/19 9:22 AM, Andrew Dunstan wrote:\\n...\\n> This will need to be rewritten in light of the above, but see\\n> <https://www.postgresql.org/message-id/87b1e36b-e36a-add5-1a9b-9fa34914a256@2ndQuadrant.com>\\n\\nThanks for the reference. Having read your motivating example, this new \\nreview reverses some of what I suggested in the prior review.\\n\\n\\nIn the existing TestLib.pm, there are eight occurrences of nearly \\nidentical usages of IPC::Run scattered through similar functions:\\n\\nrun_command:\\n my $result = IPC::Run::run $cmd, '>', \\\\$stdout, '2>', \\\\$stderr;\\n\\ncheck_pg_config:\\n my $result = IPC::Run::run [ 'pg_config', '--includedir' ], '>',\\n \\\\$stdout, '2>', \\\\$stderr\\n or die \\\"could not execute pg_config\\\";\\n\\nprogram_help_ok:\\n my $result = IPC::Run::run [ $cmd, '--help' ], '>', \\\\$stdout, '2>',\\n \\\\$stderr;\\n\\nprogram_version_ok:\\n my $result = IPC::Run::run [ $cmd, '--version' ], '>', \\\\$stdout, '2>',\\n \\\\$stderr;\\n\\nprogram_options_handling_ok:\\n my $result = IPC::Run::run [ $cmd, '--not-a-valid-option' ], '>',\\n \\\\$stdout,\\n '2>', \\\\$stderr;\\n\\ncommand_like:\\n my $result = IPC::Run::run $cmd, '>', \\\\$stdout, '2>', \\\\$stderr;\\n\\ncommand_like_safe:\\n my $result = IPC::Run::run $cmd, '>', $stdoutfile, '2>', $stderrfile;\\n\\ncommand_fails_like:\\n my $result = IPC::Run::run $cmd, '>', \\\\$stdout, '2>', \\\\$stderr,\\n @extra_ipcrun_opts;\\n\\nOne rational motive for designing TestLib with so much code duplication \\nis to make the tests that use the library easier to read:\\n\\n command_like_safe(foo);\\n command_like(bar);\\n command_fails_like(baz);\\n\\nwhich is easier to understand than:\\n\\n command_like(foo, failure_mode => safe);\\n command_like(bar);\\n command_like(baz, failure => expected);\\n\\nand so forth.\\n\\nIn line with that thinking, perhaps you should just create:\\n\\n command_fails_without_tty_like(foo)\\n\\nand be done, or perhaps:\\n\\n command_fails_like(foo, tty => 'closed')\\n\\nand still preserve some of the test readability. Will anyone like the \\nreadability of your tests if you have:\\n\\n command_fails_like(foo, extra_ipcrun_opts => ['<pty<', \\\\$eof_in]) ?\\n\\nAdmittedly, \\\"foo\\\", \\\"bar\\\", and \\\"baz\\\" above are shorthand notation for \\nthings in practice that are already somewhat hard to read, as in:\\n\\n command_fails_like(\\n [ 'pg_dump', 'qqq', 'abc' ],\\n qr/\\\\Qpg_dump: error: too many command-line arguments (first is \\n\\\"abc\\\")\\\\E/,\\n 'pg_dump: too many command-line arguments');\\n\\nbut adding more to that cruft just makes it worse. Right?\\n\\n-- \\nMark Dilger\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 13:40:06 -0800\",\n \"msg_from\": \"Mark Dilger <hornschnorter@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\nOn 11/8/19 4:40 PM, Mark Dilger wrote:\\n>\\n>\\n> On 11/8/19 9:22 AM, Andrew Dunstan wrote:\\n> ...\\n>> This will need to be rewritten in light of the above, but see\\n>> <https://www.postgresql.org/message-id/87b1e36b-e36a-add5-1a9b-9fa34914a256@2ndQuadrant.com>\\n>>\\n>\\n> Thanks for the reference. Having read your motivating example, this\\n> new review reverses some of what I suggested in the prior review.\\n>\\n>\\n> In the existing TestLib.pm, there are eight occurrences of nearly\\n> identical usages of IPC::Run scattered through similar functions:\\n>\\n>\\n[snip]\\n\\n\\n>\\n> One rational motive for designing TestLib with so much code\\n> duplication is to make the tests that use the library easier to read:\\n>\\n> command_like_safe(foo);\\n> command_like(bar);\\n> command_fails_like(baz);\\n>\\n> which is easier to understand than:\\n>\\n> command_like(foo, failure_mode => safe);\\n> command_like(bar);\\n> command_like(baz, failure => expected);\\n>\\n> and so forth.\\n>\\n> In line with that thinking, perhaps you should just create:\\n>\\n> command_fails_without_tty_like(foo)\\n>\\n> and be done, or perhaps:\\n>\\n> command_fails_like(foo, tty => 'closed')\\n>\\n> and still preserve some of the test readability. Will anyone like the\\n> readability of your tests if you have:\\n>\\n> command_fails_like(foo, extra_ipcrun_opts => ['<pty<', \\\\$eof_in]) ?\\n>\\n> Admittedly, \\\"foo\\\", \\\"bar\\\", and \\\"baz\\\" above are shorthand notation for\\n> things in practice that are already somewhat hard to read, as in:\\n>\\n> command_fails_like(\\n> [ 'pg_dump', 'qqq', 'abc' ],\\n> qr/\\\\Qpg_dump: error: too many command-line arguments (first is\\n> \\\"abc\\\")\\\\E/,\\n> 'pg_dump: too many command-line arguments');\\n>\\n> but adding more to that cruft just makes it worse. Right?\\n>\\n\\nOK, I agree that we're getting rather baroque here. I could go with your\\nsuggestion of YA function, or possibly a solution that simple passes any\\nextra arguments straight through to IPC::Run::run(), e.g.\\n\\ncommand_fails_like(\\n [ 'pg_dump', 'qqq', 'abc' ],\\n qr/\\\\Qpg_dump: error: too many command-line arguments (first is\\n\\\"abc\\\")\\\\E/,\\n 'pg_dump: too many command-line arguments',\\n '<pty<', \\\\$eof_in);\\n\\nThat means we're not future-proofing the function - we'll never be able\\nto add more arguments to it, but I'm not really certain that matters\\nanyway. I should note that perlcritic whines about subroutines with too\\nmany arguments, so making provision for more seems unnecessary anyway.\\n\\nI don't think this is worth spending a huge amount of time on, we've\\nalready spent more time discussing it than it would take to implement\\neither solution.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 9 Nov 2019 08:25:10 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\nOn 11/9/19 8:25 AM, Andrew Dunstan wrote:\\n> OK, I agree that we're getting rather baroque here. I could go with your\\n> suggestion of YA function, or possibly a solution that simple passes any\\n> extra arguments straight through to IPC::Run::run(), e.g.\\n>\\n> command_fails_like(\\n> [ 'pg_dump', 'qqq', 'abc' ],\\n> qr/\\\\Qpg_dump: error: too many command-line arguments (first is\\n> \\\"abc\\\")\\\\E/,\\n> 'pg_dump: too many command-line arguments',\\n> '<pty<', \\\\$eof_in);\\n>\\n> That means we're not future-proofing the function - we'll never be able\\n> to add more arguments to it, but I'm not really certain that matters\\n> anyway. I should note that perlcritic whines about subroutines with too\\n> many arguments, so making provision for more seems unnecessary anyway.\\n>\\n> I don't think this is worth spending a huge amount of time on, we've\\n> already spent more time discussing it than it would take to implement\\n> either solution.\\n>\\n>\\n\\nOn further consideration, I'm wondering why we don't just\\nunconditionally use a closed input pty for all these functions (except\\nrun_log). None of them use any input, and making the client worry about\\nwhether or not to close it seems something of an abstraction break.\\nThere would be no API change at all involved in this case, just a bit of\\nextra cleanliness. Would need testing on Windows, I'll go and do that now.\\n\\n\\nThoughts?\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Nov 2019 11:48:25 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 11/11/19 8:48 AM, Andrew Dunstan wrote:\\n> \\n> On 11/9/19 8:25 AM, Andrew Dunstan wrote:\\n>> OK, I agree that we're getting rather baroque here. I could go with your\\n>> suggestion of YA function, or possibly a solution that simple passes any\\n>> extra arguments straight through to IPC::Run::run(), e.g.\\n>>\\n>> command_fails_like(\\n>> [ 'pg_dump', 'qqq', 'abc' ],\\n>> qr/\\\\Qpg_dump: error: too many command-line arguments (first is\\n>> \\\"abc\\\")\\\\E/,\\n>> 'pg_dump: too many command-line arguments',\\n>> '<pty<', \\\\$eof_in);\\n>>\\n>> That means we're not future-proofing the function - we'll never be able\\n>> to add more arguments to it, but I'm not really certain that matters\\n>> anyway. I should note that perlcritic whines about subroutines with too\\n>> many arguments, so making provision for more seems unnecessary anyway.\\n>>\\n>> I don't think this is worth spending a huge amount of time on, we've\\n>> already spent more time discussing it than it would take to implement\\n>> either solution.\\n>>\\n>>\\n> \\n> On further consideration, I'm wondering why we don't just\\n> unconditionally use a closed input pty for all these functions (except\\n> run_log). None of them use any input, and making the client worry about\\n> whether or not to close it seems something of an abstraction break.\\n> There would be no API change at all involved in this case, just a bit of\\n> extra cleanliness. Would need testing on Windows, I'll go and do that now.\\n> \\n> \\n> Thoughts?\\n\\nThat sounds a lot better than your previous patch.\\n\\nPostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you change \\nall the invocations in TestLib to close input pty, should you do the \\nsame for PostgresNode? I don't have a strong argument for doing so, but \\nit seems cleaner to have both libraries invoking commands under \\nidentical conditions, such that if commands were borrowed from one \\nlibrary and called from the other they would behave the same.\\n\\nPostgresNode already uses TestLib, so perhaps setting up the environment \\ncan be abstracted into something, perhaps TestLib::run, and then used \\neverywhere that IPC::Run::run currently is used.\\n\\n\\n-- \\nMark Dilger\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Nov 2019 10:27:35 -0800\",\n \"msg_from\": \"Mark Dilger <hornschnorter@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\nOn 11/11/19 1:27 PM, Mark Dilger wrote:\\n>\\n>\\n> On 11/11/19 8:48 AM, Andrew Dunstan wrote:\\n>>\\n>> On 11/9/19 8:25 AM, Andrew Dunstan wrote:\\n>>> OK, I agree that we're getting rather baroque here. I could go with\\n>>> your\\n>>> suggestion of YA function, or possibly a solution that simple passes\\n>>> any\\n>>> extra arguments straight through to IPC::Run::run(), e.g.\\n>>>\\n>>> command_fails_like(\\n>>> [ 'pg_dump', 'qqq', 'abc' ],\\n>>> qr/\\\\Qpg_dump: error: too many command-line arguments (first is\\n>>> \\\"abc\\\")\\\\E/,\\n>>> 'pg_dump: too many command-line arguments',\\n>>> '<pty<', \\\\$eof_in);\\n>>>\\n>>> That means we're not future-proofing the function - we'll never be able\\n>>> to add more arguments to it, but I'm not really certain that matters\\n>>> anyway. I should note that perlcritic whines about subroutines with too\\n>>> many arguments, so making provision for more seems unnecessary anyway.\\n>>>\\n>>> I don't think this is worth spending a huge amount of time on, we've\\n>>> already spent more time discussing it than it would take to implement\\n>>> either solution.\\n>>>\\n>>>\\n>>\\n>> On further consideration, I'm wondering why we don't just\\n>> unconditionally use a closed input pty for all these functions (except\\n>> run_log). None of them use any input, and making the client worry about\\n>> whether or not to close it seems something of an abstraction break.\\n>> There would be no API change at all involved in this case, just a bit of\\n>> extra cleanliness. Would need testing on Windows, I'll go and do that\\n>> now.\\n>>\\n>>\\n>> Thoughts?\\n>\\n> That sounds a lot better than your previous patch.\\n>\\n> PostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you\\n> change all the invocations in TestLib to close input pty, should you\\n> do the same for PostgresNode? I don't have a strong argument for\\n> doing so, but it seems cleaner to have both libraries invoking\\n> commands under identical conditions, such that if commands were\\n> borrowed from one library and called from the other they would behave\\n> the same.\\n>\\n> PostgresNode already uses TestLib, so perhaps setting up the\\n> environment can be abstracted into something, perhaps TestLib::run,\\n> and then used everywhere that IPC::Run::run currently is used.\\n\\n\\n\\nI don't think we need to do that. In the case of the PostgresNode.pm\\nuses we know what the executable is, unlike the the TestLib.pm cases.\\nThey are our own executables and we don't expect them to be doing\\nanything funky with /dev/tty.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Nov 2019 14:28:23 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 11/11/19 11:28 AM, Andrew Dunstan wrote:\\n> \\n> On 11/11/19 1:27 PM, Mark Dilger wrote:\\n>>\\n>>\\n>> On 11/11/19 8:48 AM, Andrew Dunstan wrote:\\n>>>\\n>>> On 11/9/19 8:25 AM, Andrew Dunstan wrote:\\n>>>> OK, I agree that we're getting rather baroque here. I could go with\\n>>>> your\\n>>>> suggestion of YA function, or possibly a solution that simple passes\\n>>>> any\\n>>>> extra arguments straight through to IPC::Run::run(), e.g.\\n>>>>\\n>>>> command_fails_like(\\n>>>> [ 'pg_dump', 'qqq', 'abc' ],\\n>>>> qr/\\\\Qpg_dump: error: too many command-line arguments (first is\\n>>>> \\\"abc\\\")\\\\E/,\\n>>>> 'pg_dump: too many command-line arguments',\\n>>>> '<pty<', \\\\$eof_in);\\n>>>>\\n>>>> That means we're not future-proofing the function - we'll never be able\\n>>>> to add more arguments to it, but I'm not really certain that matters\\n>>>> anyway. I should note that perlcritic whines about subroutines with too\\n>>>> many arguments, so making provision for more seems unnecessary anyway.\\n>>>>\\n>>>> I don't think this is worth spending a huge amount of time on, we've\\n>>>> already spent more time discussing it than it would take to implement\\n>>>> either solution.\\n>>>>\\n>>>>\\n>>>\\n>>> On further consideration, I'm wondering why we don't just\\n>>> unconditionally use a closed input pty for all these functions (except\\n>>> run_log). None of them use any input, and making the client worry about\\n>>> whether or not to close it seems something of an abstraction break.\\n>>> There would be no API change at all involved in this case, just a bit of\\n>>> extra cleanliness. Would need testing on Windows, I'll go and do that\\n>>> now.\\n>>>\\n>>>\\n>>> Thoughts?\\n>>\\n>> That sounds a lot better than your previous patch.\\n>>\\n>> PostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you\\n>> change all the invocations in TestLib to close input pty, should you\\n>> do the same for PostgresNode? I don't have a strong argument for\\n>> doing so, but it seems cleaner to have both libraries invoking\\n>> commands under identical conditions, such that if commands were\\n>> borrowed from one library and called from the other they would behave\\n>> the same.\\n>>\\n>> PostgresNode already uses TestLib, so perhaps setting up the\\n>> environment can be abstracted into something, perhaps TestLib::run,\\n>> and then used everywhere that IPC::Run::run currently is used.\\n> \\n> \\n> \\n> I don't think we need to do that. In the case of the PostgresNode.pm\\n> uses we know what the executable is, unlike the the TestLib.pm cases.\\n> They are our own executables and we don't expect them to be doing\\n> anything funky with /dev/tty.\\n\\nOk. I think your proposal sounds fine.\\n\\n-- \\nMark Dilger\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Nov 2019 13:28:37 -0800\",\n \"msg_from\": \"Mark Dilger <hornschnorter@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"On 11/11/19 4:28 PM, Mark Dilger wrote:\\n>\\n>\\n>>>>>\\n>>>>\\n>>>> On further consideration, I'm wondering why we don't just\\n>>>> unconditionally use a closed input pty for all these functions (except\\n>>>> run_log). None of them use any input, and making the client worry\\n>>>> about\\n>>>> whether or not to close it seems something of an abstraction break.\\n>>>> There would be no API change at all involved in this case, just a\\n>>>> bit of\\n>>>> extra cleanliness. Would need testing on Windows, I'll go and do that\\n>>>> now.\\n>>>>\\n>>>>\\n>>>> Thoughts?\\n>>>\\n>>> That sounds a lot better than your previous patch.\\n>>>\\n>>> PostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you\\n>>> change all the invocations in TestLib to close input pty, should you\\n>>> do the same for PostgresNode? I don't have a strong argument for\\n>>> doing so, but it seems cleaner to have both libraries invoking\\n>>> commands under identical conditions, such that if commands were\\n>>> borrowed from one library and called from the other they would behave\\n>>> the same.\\n>>>\\n>>> PostgresNode already uses TestLib, so perhaps setting up the\\n>>> environment can be abstracted into something, perhaps TestLib::run,\\n>>> and then used everywhere that IPC::Run::run currently is used.\\n>>\\n>>\\n>>\\n>> I don't think we need to do that. In the case of the PostgresNode.pm\\n>> uses we know what the executable is, unlike the the TestLib.pm cases.\\n>> They are our own executables and we don't expect them to be doing\\n>> anything funky with /dev/tty.\\n>\\n> Ok. I think your proposal sounds fine.\\n\\n\\n\\nHere's a patch for that. The pty stuff crashes and burns on my Windows\\ntest box, so there I just set stdin to an empty string via the usual\\npipe mechanism.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Mon, 25 Nov 2019 08:08:56 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 11/25/19 5:08 AM, Andrew Dunstan wrote:\\n> \\n> On 11/11/19 4:28 PM, Mark Dilger wrote:\\n>>\\n>>\\n>>>>>>\\n>>>>>\\n>>>>> On further consideration, I'm wondering why we don't just\\n>>>>> unconditionally use a closed input pty for all these functions (except\\n>>>>> run_log). None of them use any input, and making the client worry\\n>>>>> about\\n>>>>> whether or not to close it seems something of an abstraction break.\\n>>>>> There would be no API change at all involved in this case, just a\\n>>>>> bit of\\n>>>>> extra cleanliness. Would need testing on Windows, I'll go and do that\\n>>>>> now.\\n>>>>>\\n>>>>>\\n>>>>> Thoughts?\\n>>>>\\n>>>> That sounds a lot better than your previous patch.\\n>>>>\\n>>>> PostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you\\n>>>> change all the invocations in TestLib to close input pty, should you\\n>>>> do the same for PostgresNode? I don't have a strong argument for\\n>>>> doing so, but it seems cleaner to have both libraries invoking\\n>>>> commands under identical conditions, such that if commands were\\n>>>> borrowed from one library and called from the other they would behave\\n>>>> the same.\\n>>>>\\n>>>> PostgresNode already uses TestLib, so perhaps setting up the\\n>>>> environment can be abstracted into something, perhaps TestLib::run,\\n>>>> and then used everywhere that IPC::Run::run currently is used.\\n>>>\\n>>>\\n>>>\\n>>> I don't think we need to do that. In the case of the PostgresNode.pm\\n>>> uses we know what the executable is, unlike the the TestLib.pm cases.\\n>>> They are our own executables and we don't expect them to be doing\\n>>> anything funky with /dev/tty.\\n>>\\n>> Ok. I think your proposal sounds fine.\\n> \\n> \\n> \\n> Here's a patch for that. The pty stuff crashes and burns on my Windows\\n> test box, so there I just set stdin to an empty string via the usual\\n> pipe mechanism.\\n\\nOk, I've reviewed and tested this. It works fine for me on Linux. I\\nam not set up to test it on Windows. I think it is ready to commit.\\n\\nI have one remaining comment about the code, and this is just FYI. I\\nwon't quibble with you committing your patch as it currently stands.\\n\\nYou might consider changing the '\\\\x04' literal to use a named control\\ncharacter, both for readability and portability, as here:\\n\\n+ use charnames ':full';\\n+ @no_stdin = ('<pty<', \\\\\\\"\\\\N{END OF TRANSMISSION}\\\");\\n\\nThe only character set I can find where this matters is EBCDIC, in\\nwhich the EOT character is 55 rather than 4. Since EBCDIC does not\\noccur in the list of supported character sets for postgres, per the\\ndocs section 23.3.1, I don't suppose it matters too much. Nor can\\nI test how this works on EBCDIC, so I'm mostly guessing that perl\\nwould do the right thing there. But, at least to my eyes, it is\\nmore immediately clear what the code is doing when the control\\ncharacter name is spelled out.\\n\\n\\n-- \\nMark Dilger\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Nov 2019 10:56:38 -0800\",\n \"msg_from\": \"Mark Dilger <hornschnorter@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n },\n {\n \"msg_contents\": \"\\nOn 11/25/19 1:56 PM, Mark Dilger wrote:\\n>\\n>\\n> On 11/25/19 5:08 AM, Andrew Dunstan wrote:\\n>>\\n>> On 11/11/19 4:28 PM, Mark Dilger wrote:\\n>>>\\n>>>\\n>>>>>>>\\n>>>>>>\\n>>>>>> On further consideration, I'm wondering why we don't just\\n>>>>>> unconditionally use a closed input pty for all these functions\\n>>>>>> (except\\n>>>>>> run_log). None of them use any input, and making the client worry\\n>>>>>> about\\n>>>>>> whether or not to close it seems something of an abstraction break.\\n>>>>>> There would be no API change at all involved in this case, just a\\n>>>>>> bit of\\n>>>>>> extra cleanliness. Would need testing on Windows, I'll go and do\\n>>>>>> that\\n>>>>>> now.\\n>>>>>>\\n>>>>>>\\n>>>>>> Thoughts?\\n>>>>>\\n>>>>> That sounds a lot better than your previous patch.\\n>>>>>\\n>>>>> PostgresNode.pm and TestLib.pm both invoke IPC::Run::run. If you\\n>>>>> change all the invocations in TestLib to close input pty, should you\\n>>>>> do the same for PostgresNode? I don't have a strong argument for\\n>>>>> doing so, but it seems cleaner to have both libraries invoking\\n>>>>> commands under identical conditions, such that if commands were\\n>>>>> borrowed from one library and called from the other they would behave\\n>>>>> the same.\\n>>>>>\\n>>>>> PostgresNode already uses TestLib, so perhaps setting up the\\n>>>>> environment can be abstracted into something, perhaps TestLib::run,\\n>>>>> and then used everywhere that IPC::Run::run currently is used.\\n>>>>\\n>>>>\\n>>>>\\n>>>> I don't think we need to do that. In the case of the PostgresNode.pm\\n>>>> uses we know what the executable is, unlike the the TestLib.pm cases.\\n>>>> They are our own executables and we don't expect them to be doing\\n>>>> anything funky with /dev/tty.\\n>>>\\n>>> Ok. I think your proposal sounds fine.\\n>>\\n>>\\n>>\\n>> Here's a patch for that. The pty stuff crashes and burns on my Windows\\n>> test box, so there I just set stdin to an empty string via the usual\\n>> pipe mechanism.\\n>\\n> Ok, I've reviewed and tested this. It works fine for me on Linux. I\\n> am not set up to test it on Windows. I think it is ready to commit.\\n>\\n> I have one remaining comment about the code, and this is just FYI. I\\n> won't quibble with you committing your patch as it currently stands.\\n>\\n> You might consider changing the '\\\\x04' literal to use a named control\\n> character, both for readability and portability, as here:\\n>\\n> + use charnames ':full';\\n> + @no_stdin = ('<pty<', \\\\\\\"\\\\N{END OF TRANSMISSION}\\\");\\n>\\n> The only character set I can find where this matters is EBCDIC, in\\n> which the EOT character is 55 rather than 4. Since EBCDIC does not\\n> occur in the list of supported character sets for postgres, per the\\n> docs section 23.3.1, I don't suppose it matters too much. Nor can\\n> I test how this works on EBCDIC, so I'm mostly guessing that perl\\n> would do the right thing there. But, at least to my eyes, it is\\n> more immediately clear what the code is doing when the control\\n> character name is spelled out.\\n>\\n>\\n\\n\\nAgreed, I'll do it that way. This is quite timely, as I just finished\\nreworking the patch that relies on it. Thanks for the review.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Nov 2019 15:07:10 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: TestLib::command_fails_like enhancement\"\n }\n]","isConversational":false}}},{"rowIdx":49381,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello Devs,\n\nThis patch moves duplicated query cancellation code code from psql & \nscripts to fe-utils, so that it is shared and may be used by other \ncommands.\n\nThis is because Masao-san suggested to add a query cancellation feature to \npgbench for long queries (server-side data generation being discussed, but \npossibly pk and fk could use that as well).\n\n-- \nFabien.","msg_date":"Thu, 31 Oct 2019 19:43:36 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":true,"msg_subject":"fe-utils - share query cancellation code"},{"msg_contents":"On Thu, Oct 31, 2019 at 11:43 PM Fabien COELHO <coelho@cri.ensmp.fr> wrote:\n\n>\n> Hello Devs,\n>\n> This patch moves duplicated query cancellation code code from psql &\n> scripts to fe-utils, so that it is shared and may be used by other\n> commands.\n>\n> This is because Masao-san suggested to add a query cancellation feature to\n> pgbench for long queries (server-side data generation being discussed, but\n> possibly pk and fk could use that as well).\n>\n> --\n> Fabien.\n\n\nI give a quick look and I think we can\n\nvoid\npsql_setup_cancel_handler(void)\n{\n#ifndef WIN32\n setup_cancel_handler(psql_sigint_callback);\n#else\n setup_cancel_handler();\n#endif /* WIN32 */\n}\n\nto\n\nvoid\npsql_setup_cancel_handler(void)\n{\n setup_cancel_handler(psql_sigint_callback);\n}\n\nBecause it does not matter for setup_cancel_handler what we passed\nbecause it is ignoring that in case of windows.\n\nHmm, need to remove the assert in the function\n\"setup_cancel_handler\"\n\n-- Ibrar Ahmed\n\nOn Thu, Oct 31, 2019 at 11:43 PM Fabien COELHO <coelho@cri.ensmp.fr> wrote:\nHello Devs,\n\nThis patch moves duplicated query cancellation code code from psql & \nscripts to fe-utils, so that it is shared and may be used by other \ncommands.\n\nThis is because Masao-san suggested to add a query cancellation feature to \npgbench for long queries (server-side data generation being discussed, but \npossibly pk and fk could use that as well).\n\n-- \nFabien.I give a quick look and I think we can voidpsql_setup_cancel_handler(void){#ifndef WIN32 setup_cancel_handler(psql_sigint_callback);#else setup_cancel_handler();#endif /* WIN32 */}to voidpsql_setup_cancel_handler(void){ setup_cancel_handler(psql_sigint_callback);}Because it does not matter for setup_cancel_handler what we passedbecause it is ignoring that in case of windows.Hmm, need to remove the assert in the function\"setup_cancel_handler\"-- Ibrar Ahmed","msg_date":"Fri, 1 Nov 2019 00:50:40 +0500","msg_from":"Ibrar Ahmed <ibrar.ahmad@gmail.com>","msg_from_op":false,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"Hello,\n\n> I give a quick look and I think we can\n>\n> void\n> psql_setup_cancel_handler(void)\n> {\n> setup_cancel_handler(psql_sigint_callback);\n> }\n>\n> Because it does not matter for setup_cancel_handler what we passed\n> because it is ignoring that in case of windows.\n\nThe \"psql_sigint_callback\" function is not defined under WIN32.\n\nI've fixed a missing NULL argument in the section you pointed out, though.\n\nI've used the shared infrastructure in pgbench.\n\nI've noticed yet another instance of the cancelation stuff in \n\"src/bin/pg_dump/parallel.c\", but it seems somehow different from the two \nothers, so I have not tried to used the shared version.\n\n-- \nFabien.","msg_date":"Fri, 1 Nov 2019 10:19:04 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":true,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"On 2019-Nov-01, Fabien COELHO wrote:\n\n> diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c\n> index 03bcd22996..389b4d7bcd 100644\n> --- a/src/bin/pgbench/pgbench.c\n> +++ b/src/bin/pgbench/pgbench.c\n> @@ -59,9 +59,10 @@\n> \n> #include \"common/int.h\"\n> #include \"common/logging.h\"\n> -#include \"fe_utils/conditional.h\"\n> #include \"getopt_long.h\"\n> #include \"libpq-fe.h\"\n> +#include \"fe_utils/conditional.h\"\n> +#include \"fe_utils/cancel.h\"\n> #include \"pgbench.h\"\n> #include \"portability/instr_time.h\"\n\nwtf?\n\n-- \n�lvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 1 Nov 2019 10:30:19 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"\nHello Alvaro,\n\n>> #include \"common/int.h\"\n>> #include \"common/logging.h\"\n>> -#include \"fe_utils/conditional.h\"\n>> #include \"getopt_long.h\"\n>> #include \"libpq-fe.h\"\n>> +#include \"fe_utils/conditional.h\"\n>> +#include \"fe_utils/cancel.h\"\n>> #include \"pgbench.h\"\n>> #include \"portability/instr_time.h\"\n>\n> wtf?\n\nI understand that you are unhappy about something, but where the issue is \nfails me, the \"wtf\" 3 characters are not enough to point me in the right \ndirection. Feel free to elaborate a little bit more:-)\n\n-- \nFabien.\n\n\n","msg_date":"Fri, 1 Nov 2019 16:26:42 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":true,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"On 2019-Nov-01, Fabien COELHO wrote:\n\n> > > #include \"common/int.h\"\n> > > #include \"common/logging.h\"\n> > > -#include \"fe_utils/conditional.h\"\n> > > #include \"getopt_long.h\"\n> > > #include \"libpq-fe.h\"\n> > > +#include \"fe_utils/conditional.h\"\n> > > +#include \"fe_utils/cancel.h\"\n> > > #include \"pgbench.h\"\n> > > #include \"portability/instr_time.h\"\n> > \n> > wtf?\n> \n> I understand that you are unhappy about something, but where the issue is\n> fails me, the \"wtf\" 3 characters are not enough to point me in the right\n> direction. Feel free to elaborate a little bit more:-)\n\nI don't see why you move the \"conditional.h\" line out of its correct\nalphabetical position (where it is now), and then add \"cancel.h\" next to\nit also out of its correct alphabetical position.\n\n-- \n�lvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 1 Nov 2019 12:30:34 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":">> I understand that you are unhappy about something, but where the issue is\n>> fails me, the \"wtf\" 3 characters are not enough to point me in the right\n>> direction. Feel free to elaborate a little bit more:-)\n>\n> I don't see why you move the \"conditional.h\" line out of its correct\n> alphabetical position (where it is now), and then add \"cancel.h\" next to\n> it also out of its correct alphabetical position.\n\nBecause \"cancel.h\" requires PGconn declaration, thus must appear after \n\"libpq-fe.h\", and once I put it after that letting \"conditional.h\" above & \nalone looked a little bit silly. I put cancel after conditional because it \nwas the new addition, which is somehow logical, although not alphabetical.\n\nNow I can put cancel before conditional, sure.\n\nPatch v3 attached does that.\n\n-- \nFabien.","msg_date":"Fri, 1 Nov 2019 18:41:52 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":true,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"On 2019-Nov-01, Fabien COELHO wrote:\n\n> Because \"cancel.h\" requires PGconn declaration, thus must appear after\n> \"libpq-fe.h\",\n\nThen you need to add #include libpq-fe.h in cancel.h. Our policy is\nthat headers compile standalone (c.h / postgres_fe.h / postgres.h\nexcluded).\n\n-- \n�lvaro Herrera https://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 1 Nov 2019 16:17:42 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"> Then you need to add #include libpq-fe.h in cancel.h. Our policy is\n> that headers compile standalone (c.h / postgres_fe.h / postgres.h\n> excluded).\n\nOk. I do not make a habit of adding headers in postgres, so I did not \nnotice there was an alphabetical logic to that.\n\nAttached patch v4 does it.\n\n-- \nFabien.","msg_date":"Fri, 1 Nov 2019 22:38:09 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":true,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"On Sat, Nov 2, 2019 at 10:38 AM Fabien COELHO <coelho@cri.ensmp.fr> wrote:\n> Attached patch v4 does it.\n\nHi Fabien,\n\nIt looks like don't define sigint_interrupt_jmp and\nsigint_interrupt_enabled on Windows, yet they are still declared and\nreferenced?\n\ncommand.obj : error LNK2001: unresolved external symbol\nsigint_interrupt_enabled [C:\\projects\\postgresql\\psql.vcxproj]\ncopy.obj : error LNK2001: unresolved external symbol\nsigint_interrupt_enabled [C:\\projects\\postgresql\\psql.vcxproj]\ninput.obj : error LNK2001: unresolved external symbol\nsigint_interrupt_enabled [C:\\projects\\postgresql\\psql.vcxproj]\nmainloop.obj : error LNK2001: unresolved external symbol\nsigint_interrupt_enabled [C:\\projects\\postgresql\\psql.vcxproj]\ncommand.obj : error LNK2001: unresolved external symbol\nsigint_interrupt_jmp [C:\\projects\\postgresql\\psql.vcxproj]\ncopy.obj : error LNK2001: unresolved external symbol\nsigint_interrupt_jmp [C:\\projects\\postgresql\\psql.vcxproj]\nmainloop.obj : error LNK2001: unresolved external symbol\nsigint_interrupt_jmp [C:\\projects\\postgresql\\psql.vcxproj]\n.\\Release\\psql\\psql.exe : fatal error LNK1120: 2 unresolved externals\n[C:\\projects\\postgresql\\psql.vcxproj]\n0 Warning(s)\n\nhttps://ci.appveyor.com/project/postgresql-cfbot/postgresql/build/1.0.64074\n\n\n","msg_date":"Mon, 4 Nov 2019 14:28:36 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"> It looks like don't define sigint_interrupt_jmp and \n> sigint_interrupt_enabled on Windows, yet they are still declared and \n> referenced?\n\nIndeed, I put it on the wrong side of a \"#ifndef WIN32\".\n\nBasically it is a false constant under WIN32, which it seems does not have \nsigint handler, but the code checks whether the non existent handler is \nenabled anyway.\n\nPatch v5 attached fixes that, hopefully.\n\n-- \nFabien.","msg_date":"Wed, 6 Nov 2019 10:41:39 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":true,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"On Wed, Nov 06, 2019 at 10:41:39AM +0100, Fabien COELHO wrote:\n> Indeed, I put it on the wrong side of a \"#ifndef WIN32\".\n> \n> Basically it is a false constant under WIN32, which it seems does not have\n> sigint handler, but the code checks whether the non existent handler is\n> enabled anyway.\n> \n> Patch v5 attached fixes that, hopefully.\n\nI have looked at this one, and found a couple of issues, most of them\nsmall-ish.\n\ns/cancelation/cancellation/ in fe_utils/cancel.h.\n\nThen, the format of the new file headers was not really consistent\nwith the rest, and the order of the headers included in most of the\nfiles was incorrect. That would break the recent flow of commits done\nby Amit K.\n\nThe query cancellation added to pgbench is different than the actual\nrefactoring, and it is a result of the refactoring, so I would rather\nsplit that into two different commits for clarity. The split is easy\nenough, so that's fine not to send two different patches.\n\nCompilation of the patch fails for me on Windows for psql:\nunresolved external symbol sigint_interrupt_jmp \nPlease note that Mr Robot complains as well at build time:\nhttp://commitfest.cputube.org/fabien-coelho.html\n\nVisibly the problem here is that sigint_interrupt_jmp is declared in\ncommon.h, but you have moved it to a non-WIN32 section of the code in\npsql/common.c. And actually, note that copy.c and mainloop.c make use\nof it...\n\nI would not worry much about SIGTERM as you mentioned in the comments,\nquery cancellations are associated to SIGINT now. There could be an\nargument possible later to allow passing down a custom signal though.\n\nAttached is an updated patch with a couple of edits I have done,\nincluding the removal of some noise diffs and the previous edits. I\nam switching the patch as waiting on author, bumping it to next CF at\nthe same time. Could you please fix the Windows issue?\n--\nMichael","msg_date":"Thu, 28 Nov 2019 16:52:08 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"Bonjour Michaï¿œl,\n\n> The query cancellation added to pgbench is different than the actual\n> refactoring, and it is a result of the refactoring, so I would rather\n> split that into two different commits for clarity. The split is easy\n> enough, so that's fine not to send two different patches.\n\nYep, different file set.\n\n> Compilation of the patch fails for me on Windows for psql:\n> unresolved external symbol sigint_interrupt_jmp\n> Please note that Mr Robot complains as well at build time:\n> http://commitfest.cputube.org/fabien-coelho.html\n>\n> Visibly the problem here is that sigint_interrupt_jmp is declared in\n> common.h, but you have moved it to a non-WIN32 section of the code in\n> psql/common.c. And actually, note that copy.c and mainloop.c make use\n> of it...\n\nIndeed.\n\n> I would not worry much about SIGTERM as you mentioned in the comments,\n> query cancellations are associated to SIGINT now. There could be an\n> argument possible later to allow passing down a custom signal though.\n\nOk.\n\n> Attached is an updated patch with a couple of edits I have done,\n> including the removal of some noise diffs and the previous edits.\n\nThanks!\n\n> I am switching the patch as waiting on author, bumping it to next CF at \n> the same time. Could you please fix the Windows issue?\n\nI do not have a Windows host, so I can only do blind tests. I just moved \nthe declaration out of !WIN32 scope in attached v7, which might solve the \nresolution error. All other issues pointed out above seem fixed in the v6 \nyou sent.\n\n-- \nFabien.","msg_date":"Fri, 29 Nov 2019 08:44:25 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":true,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"On Fri, Nov 29, 2019 at 08:44:25AM +0100, Fabien COELHO wrote:\n> I do not have a Windows host, so I can only do blind tests. I just moved the\n> declaration out of !WIN32 scope in attached v7, which might solve the\n> resolution error. All other issues pointed out above seem fixed in the v6\n> you sent.\n\nCommitted the patch after splitting things into two commits and after\ntesting things from Linux and from a Windows console: the actual\nrefactoring and the pgbench changes. While polishing the code, I have\nfound the upthread argument of Ibrar quite appealing because there are\nuse cases where a callback can be interesting on Windows, like simply\nbeing able to log the cancel event to a different source. So I have\nremoved the callback restriction and the assertion, making the\ncallback of psql a no-op on Windows. A second thing is that two large\ncomments originally in psql had better be moved to cancel.c because\nthe logic with libpq cancel routines applies only there.\n--\nMichael","msg_date":"Mon, 2 Dec 2019 11:54:02 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"On Mon, Dec 02, 2019 at 11:54:02AM +0900, Michael Paquier wrote:\n> Committed the patch after splitting things into two commits and after\n> testing things from Linux and from a Windows console: the actual\n> refactoring and the pgbench changes.\n\nI have found that we have a useless declaration of CancelRequested in\ncommon.h, which is already part of cancel.h. On top of that I think\nthat we need to rework a bit the header inclusions of bin/scripts/, as\nper the attached. A small set of issues, still these are issues.\nSorry for having missed these.\n--\nMichael","msg_date":"Tue, 3 Dec 2019 19:16:25 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"Bonjour Michaï¿œl,\n\n>> Committed the patch after splitting things into two commits and after \n>> testing things from Linux and from a Windows console: the actual \n>> refactoring and the pgbench changes.\n>\n> I have found that we have a useless declaration of CancelRequested in \n> common.h, which is already part of cancel.h.\n\nOk.\n\n> On top of that I think that we need to rework a bit the header \n> inclusions of bin/scripts/, as per the attached.\n\nLooks fine to me: patch applies, compiles, runs.\n\n-- \nFabien.","msg_date":"Tue, 3 Dec 2019 13:11:27 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":true,"msg_subject":"Re: fe-utils - share query cancellation code"},{"msg_contents":"On Tue, Dec 03, 2019 at 01:11:27PM +0100, Fabien COELHO wrote:\n> Looks fine to me: patch applies, compiles, runs.\n\nThanks for double-checking. Done.\n--\nMichael","msg_date":"Wed, 4 Dec 2019 10:10:12 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: fe-utils - share query cancellation code"}],"string":"[\n {\n \"msg_contents\": \"Hello Devs,\\n\\nThis patch moves duplicated query cancellation code code from psql & \\nscripts to fe-utils, so that it is shared and may be used by other \\ncommands.\\n\\nThis is because Masao-san suggested to add a query cancellation feature to \\npgbench for long queries (server-side data generation being discussed, but \\npossibly pk and fk could use that as well).\\n\\n-- \\nFabien.\",\n \"msg_date\": \"Thu, 31 Oct 2019 19:43:36 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 11:43 PM Fabien COELHO <coelho@cri.ensmp.fr> wrote:\\n\\n>\\n> Hello Devs,\\n>\\n> This patch moves duplicated query cancellation code code from psql &\\n> scripts to fe-utils, so that it is shared and may be used by other\\n> commands.\\n>\\n> This is because Masao-san suggested to add a query cancellation feature to\\n> pgbench for long queries (server-side data generation being discussed, but\\n> possibly pk and fk could use that as well).\\n>\\n> --\\n> Fabien.\\n\\n\\nI give a quick look and I think we can\\n\\nvoid\\npsql_setup_cancel_handler(void)\\n{\\n#ifndef WIN32\\n setup_cancel_handler(psql_sigint_callback);\\n#else\\n setup_cancel_handler();\\n#endif /* WIN32 */\\n}\\n\\nto\\n\\nvoid\\npsql_setup_cancel_handler(void)\\n{\\n setup_cancel_handler(psql_sigint_callback);\\n}\\n\\nBecause it does not matter for setup_cancel_handler what we passed\\nbecause it is ignoring that in case of windows.\\n\\nHmm, need to remove the assert in the function\\n\\\"setup_cancel_handler\\\"\\n\\n-- Ibrar Ahmed\\n\\nOn Thu, Oct 31, 2019 at 11:43 PM Fabien COELHO <coelho@cri.ensmp.fr> wrote:\\nHello Devs,\\n\\nThis patch moves duplicated query cancellation code code from psql & \\nscripts to fe-utils, so that it is shared and may be used by other \\ncommands.\\n\\nThis is because Masao-san suggested to add a query cancellation feature to \\npgbench for long queries (server-side data generation being discussed, but \\npossibly pk and fk could use that as well).\\n\\n-- \\nFabien.I give a quick look and I think we can voidpsql_setup_cancel_handler(void){#ifndef WIN32 setup_cancel_handler(psql_sigint_callback);#else setup_cancel_handler();#endif /* WIN32 */}to voidpsql_setup_cancel_handler(void){ setup_cancel_handler(psql_sigint_callback);}Because it does not matter for setup_cancel_handler what we passedbecause it is ignoring that in case of windows.Hmm, need to remove the assert in the function\\\"setup_cancel_handler\\\"-- Ibrar Ahmed\",\n \"msg_date\": \"Fri, 1 Nov 2019 00:50:40 +0500\",\n \"msg_from\": \"Ibrar Ahmed <ibrar.ahmad@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"Hello,\\n\\n> I give a quick look and I think we can\\n>\\n> void\\n> psql_setup_cancel_handler(void)\\n> {\\n> setup_cancel_handler(psql_sigint_callback);\\n> }\\n>\\n> Because it does not matter for setup_cancel_handler what we passed\\n> because it is ignoring that in case of windows.\\n\\nThe \\\"psql_sigint_callback\\\" function is not defined under WIN32.\\n\\nI've fixed a missing NULL argument in the section you pointed out, though.\\n\\nI've used the shared infrastructure in pgbench.\\n\\nI've noticed yet another instance of the cancelation stuff in \\n\\\"src/bin/pg_dump/parallel.c\\\", but it seems somehow different from the two \\nothers, so I have not tried to used the shared version.\\n\\n-- \\nFabien.\",\n \"msg_date\": \"Fri, 1 Nov 2019 10:19:04 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"On 2019-Nov-01, Fabien COELHO wrote:\\n\\n> diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c\\n> index 03bcd22996..389b4d7bcd 100644\\n> --- a/src/bin/pgbench/pgbench.c\\n> +++ b/src/bin/pgbench/pgbench.c\\n> @@ -59,9 +59,10 @@\\n> \\n> #include \\\"common/int.h\\\"\\n> #include \\\"common/logging.h\\\"\\n> -#include \\\"fe_utils/conditional.h\\\"\\n> #include \\\"getopt_long.h\\\"\\n> #include \\\"libpq-fe.h\\\"\\n> +#include \\\"fe_utils/conditional.h\\\"\\n> +#include \\\"fe_utils/cancel.h\\\"\\n> #include \\\"pgbench.h\\\"\\n> #include \\\"portability/instr_time.h\\\"\\n\\nwtf?\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 10:30:19 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"\\nHello Alvaro,\\n\\n>> #include \\\"common/int.h\\\"\\n>> #include \\\"common/logging.h\\\"\\n>> -#include \\\"fe_utils/conditional.h\\\"\\n>> #include \\\"getopt_long.h\\\"\\n>> #include \\\"libpq-fe.h\\\"\\n>> +#include \\\"fe_utils/conditional.h\\\"\\n>> +#include \\\"fe_utils/cancel.h\\\"\\n>> #include \\\"pgbench.h\\\"\\n>> #include \\\"portability/instr_time.h\\\"\\n>\\n> wtf?\\n\\nI understand that you are unhappy about something, but where the issue is \\nfails me, the \\\"wtf\\\" 3 characters are not enough to point me in the right \\ndirection. Feel free to elaborate a little bit more:-)\\n\\n-- \\nFabien.\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 16:26:42 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"On 2019-Nov-01, Fabien COELHO wrote:\\n\\n> > > #include \\\"common/int.h\\\"\\n> > > #include \\\"common/logging.h\\\"\\n> > > -#include \\\"fe_utils/conditional.h\\\"\\n> > > #include \\\"getopt_long.h\\\"\\n> > > #include \\\"libpq-fe.h\\\"\\n> > > +#include \\\"fe_utils/conditional.h\\\"\\n> > > +#include \\\"fe_utils/cancel.h\\\"\\n> > > #include \\\"pgbench.h\\\"\\n> > > #include \\\"portability/instr_time.h\\\"\\n> > \\n> > wtf?\\n> \\n> I understand that you are unhappy about something, but where the issue is\\n> fails me, the \\\"wtf\\\" 3 characters are not enough to point me in the right\\n> direction. Feel free to elaborate a little bit more:-)\\n\\nI don't see why you move the \\\"conditional.h\\\" line out of its correct\\nalphabetical position (where it is now), and then add \\\"cancel.h\\\" next to\\nit also out of its correct alphabetical position.\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 12:30:34 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \">> I understand that you are unhappy about something, but where the issue is\\n>> fails me, the \\\"wtf\\\" 3 characters are not enough to point me in the right\\n>> direction. Feel free to elaborate a little bit more:-)\\n>\\n> I don't see why you move the \\\"conditional.h\\\" line out of its correct\\n> alphabetical position (where it is now), and then add \\\"cancel.h\\\" next to\\n> it also out of its correct alphabetical position.\\n\\nBecause \\\"cancel.h\\\" requires PGconn declaration, thus must appear after \\n\\\"libpq-fe.h\\\", and once I put it after that letting \\\"conditional.h\\\" above & \\nalone looked a little bit silly. I put cancel after conditional because it \\nwas the new addition, which is somehow logical, although not alphabetical.\\n\\nNow I can put cancel before conditional, sure.\\n\\nPatch v3 attached does that.\\n\\n-- \\nFabien.\",\n \"msg_date\": \"Fri, 1 Nov 2019 18:41:52 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"On 2019-Nov-01, Fabien COELHO wrote:\\n\\n> Because \\\"cancel.h\\\" requires PGconn declaration, thus must appear after\\n> \\\"libpq-fe.h\\\",\\n\\nThen you need to add #include libpq-fe.h in cancel.h. Our policy is\\nthat headers compile standalone (c.h / postgres_fe.h / postgres.h\\nexcluded).\\n\\n-- \\n�lvaro Herrera https://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 16:17:42 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"> Then you need to add #include libpq-fe.h in cancel.h. Our policy is\\n> that headers compile standalone (c.h / postgres_fe.h / postgres.h\\n> excluded).\\n\\nOk. I do not make a habit of adding headers in postgres, so I did not \\nnotice there was an alphabetical logic to that.\\n\\nAttached patch v4 does it.\\n\\n-- \\nFabien.\",\n \"msg_date\": \"Fri, 1 Nov 2019 22:38:09 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 2, 2019 at 10:38 AM Fabien COELHO <coelho@cri.ensmp.fr> wrote:\\n> Attached patch v4 does it.\\n\\nHi Fabien,\\n\\nIt looks like don't define sigint_interrupt_jmp and\\nsigint_interrupt_enabled on Windows, yet they are still declared and\\nreferenced?\\n\\ncommand.obj : error LNK2001: unresolved external symbol\\nsigint_interrupt_enabled [C:\\\\projects\\\\postgresql\\\\psql.vcxproj]\\ncopy.obj : error LNK2001: unresolved external symbol\\nsigint_interrupt_enabled [C:\\\\projects\\\\postgresql\\\\psql.vcxproj]\\ninput.obj : error LNK2001: unresolved external symbol\\nsigint_interrupt_enabled [C:\\\\projects\\\\postgresql\\\\psql.vcxproj]\\nmainloop.obj : error LNK2001: unresolved external symbol\\nsigint_interrupt_enabled [C:\\\\projects\\\\postgresql\\\\psql.vcxproj]\\ncommand.obj : error LNK2001: unresolved external symbol\\nsigint_interrupt_jmp [C:\\\\projects\\\\postgresql\\\\psql.vcxproj]\\ncopy.obj : error LNK2001: unresolved external symbol\\nsigint_interrupt_jmp [C:\\\\projects\\\\postgresql\\\\psql.vcxproj]\\nmainloop.obj : error LNK2001: unresolved external symbol\\nsigint_interrupt_jmp [C:\\\\projects\\\\postgresql\\\\psql.vcxproj]\\n.\\\\Release\\\\psql\\\\psql.exe : fatal error LNK1120: 2 unresolved externals\\n[C:\\\\projects\\\\postgresql\\\\psql.vcxproj]\\n0 Warning(s)\\n\\nhttps://ci.appveyor.com/project/postgresql-cfbot/postgresql/build/1.0.64074\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Nov 2019 14:28:36 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"> It looks like don't define sigint_interrupt_jmp and \\n> sigint_interrupt_enabled on Windows, yet they are still declared and \\n> referenced?\\n\\nIndeed, I put it on the wrong side of a \\\"#ifndef WIN32\\\".\\n\\nBasically it is a false constant under WIN32, which it seems does not have \\nsigint handler, but the code checks whether the non existent handler is \\nenabled anyway.\\n\\nPatch v5 attached fixes that, hopefully.\\n\\n-- \\nFabien.\",\n \"msg_date\": \"Wed, 6 Nov 2019 10:41:39 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 06, 2019 at 10:41:39AM +0100, Fabien COELHO wrote:\\n> Indeed, I put it on the wrong side of a \\\"#ifndef WIN32\\\".\\n> \\n> Basically it is a false constant under WIN32, which it seems does not have\\n> sigint handler, but the code checks whether the non existent handler is\\n> enabled anyway.\\n> \\n> Patch v5 attached fixes that, hopefully.\\n\\nI have looked at this one, and found a couple of issues, most of them\\nsmall-ish.\\n\\ns/cancelation/cancellation/ in fe_utils/cancel.h.\\n\\nThen, the format of the new file headers was not really consistent\\nwith the rest, and the order of the headers included in most of the\\nfiles was incorrect. That would break the recent flow of commits done\\nby Amit K.\\n\\nThe query cancellation added to pgbench is different than the actual\\nrefactoring, and it is a result of the refactoring, so I would rather\\nsplit that into two different commits for clarity. The split is easy\\nenough, so that's fine not to send two different patches.\\n\\nCompilation of the patch fails for me on Windows for psql:\\nunresolved external symbol sigint_interrupt_jmp \\nPlease note that Mr Robot complains as well at build time:\\nhttp://commitfest.cputube.org/fabien-coelho.html\\n\\nVisibly the problem here is that sigint_interrupt_jmp is declared in\\ncommon.h, but you have moved it to a non-WIN32 section of the code in\\npsql/common.c. And actually, note that copy.c and mainloop.c make use\\nof it...\\n\\nI would not worry much about SIGTERM as you mentioned in the comments,\\nquery cancellations are associated to SIGINT now. There could be an\\nargument possible later to allow passing down a custom signal though.\\n\\nAttached is an updated patch with a couple of edits I have done,\\nincluding the removal of some noise diffs and the previous edits. I\\nam switching the patch as waiting on author, bumping it to next CF at\\nthe same time. Could you please fix the Windows issue?\\n--\\nMichael\",\n \"msg_date\": \"Thu, 28 Nov 2019 16:52:08 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"Bonjour Michaï¿œl,\\n\\n> The query cancellation added to pgbench is different than the actual\\n> refactoring, and it is a result of the refactoring, so I would rather\\n> split that into two different commits for clarity. The split is easy\\n> enough, so that's fine not to send two different patches.\\n\\nYep, different file set.\\n\\n> Compilation of the patch fails for me on Windows for psql:\\n> unresolved external symbol sigint_interrupt_jmp\\n> Please note that Mr Robot complains as well at build time:\\n> http://commitfest.cputube.org/fabien-coelho.html\\n>\\n> Visibly the problem here is that sigint_interrupt_jmp is declared in\\n> common.h, but you have moved it to a non-WIN32 section of the code in\\n> psql/common.c. And actually, note that copy.c and mainloop.c make use\\n> of it...\\n\\nIndeed.\\n\\n> I would not worry much about SIGTERM as you mentioned in the comments,\\n> query cancellations are associated to SIGINT now. There could be an\\n> argument possible later to allow passing down a custom signal though.\\n\\nOk.\\n\\n> Attached is an updated patch with a couple of edits I have done,\\n> including the removal of some noise diffs and the previous edits.\\n\\nThanks!\\n\\n> I am switching the patch as waiting on author, bumping it to next CF at \\n> the same time. Could you please fix the Windows issue?\\n\\nI do not have a Windows host, so I can only do blind tests. I just moved \\nthe declaration out of !WIN32 scope in attached v7, which might solve the \\nresolution error. All other issues pointed out above seem fixed in the v6 \\nyou sent.\\n\\n-- \\nFabien.\",\n \"msg_date\": \"Fri, 29 Nov 2019 08:44:25 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 29, 2019 at 08:44:25AM +0100, Fabien COELHO wrote:\\n> I do not have a Windows host, so I can only do blind tests. I just moved the\\n> declaration out of !WIN32 scope in attached v7, which might solve the\\n> resolution error. All other issues pointed out above seem fixed in the v6\\n> you sent.\\n\\nCommitted the patch after splitting things into two commits and after\\ntesting things from Linux and from a Windows console: the actual\\nrefactoring and the pgbench changes. While polishing the code, I have\\nfound the upthread argument of Ibrar quite appealing because there are\\nuse cases where a callback can be interesting on Windows, like simply\\nbeing able to log the cancel event to a different source. So I have\\nremoved the callback restriction and the assertion, making the\\ncallback of psql a no-op on Windows. A second thing is that two large\\ncomments originally in psql had better be moved to cancel.c because\\nthe logic with libpq cancel routines applies only there.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 2 Dec 2019 11:54:02 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 02, 2019 at 11:54:02AM +0900, Michael Paquier wrote:\\n> Committed the patch after splitting things into two commits and after\\n> testing things from Linux and from a Windows console: the actual\\n> refactoring and the pgbench changes.\\n\\nI have found that we have a useless declaration of CancelRequested in\\ncommon.h, which is already part of cancel.h. On top of that I think\\nthat we need to rework a bit the header inclusions of bin/scripts/, as\\nper the attached. A small set of issues, still these are issues.\\nSorry for having missed these.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 3 Dec 2019 19:16:25 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"Bonjour Michaï¿œl,\\n\\n>> Committed the patch after splitting things into two commits and after \\n>> testing things from Linux and from a Windows console: the actual \\n>> refactoring and the pgbench changes.\\n>\\n> I have found that we have a useless declaration of CancelRequested in \\n> common.h, which is already part of cancel.h.\\n\\nOk.\\n\\n> On top of that I think that we need to rework a bit the header \\n> inclusions of bin/scripts/, as per the attached.\\n\\nLooks fine to me: patch applies, compiles, runs.\\n\\n-- \\nFabien.\",\n \"msg_date\": \"Tue, 3 Dec 2019 13:11:27 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 03, 2019 at 01:11:27PM +0100, Fabien COELHO wrote:\\n> Looks fine to me: patch applies, compiles, runs.\\n\\nThanks for double-checking. Done.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 4 Dec 2019 10:10:12 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: fe-utils - share query cancellation code\"\n }\n]","isConversational":false}}},{"rowIdx":49382,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nWe currently align byval types such as int4/8, float4/8, timestamp *,\ndate etc, even though we mostly don't need to. When tuples are deformed,\nall byval types are copied out from the tuple data into the\ncorresponding Datum array, therefore the original alignment in the tuple\ndata doesn't matter. This is different from byref types, where the\nDatum formed will often be a pointer into the tuple data.\n\nWhile there are some older systems where it could be a bit slower to\ncopy data out from unaligned positions into the datum array, this is\nmore than bought back by the next point:\n\n\nThe fact that these types are aligned has substantial costs:\n\nFor one, we often waste substantial amounts of space inside tables with\nalignment padding. It's not uncommon to see about 30% or more of space\nwasted (especially when taking alignment of the first column into\naccount).\n\nFor another, and this I think is less obvious, we actually waste\nsubstantial amounts of CPU maintaining the alignment. This is primarily\nthe case because we have to perform to align the pointer to the next\nfield during tuple [de]forming. Those instructions [1] have to be\nexecuted taking time, but what's worse, they also reduce the ability of\nout-of-order execution to hide latencies. There's a hard dependency on\nknowing the offset to the next column to be able to continue with the\nnext column.\n\n\nThere's two reasons why we can't just set the alignment for these types\nto 'c'.\n1) pg_upgrade, for fairly obvious reasons\n2) We map catalog table rows to structs, in a *lot* of places.\n\n\nIt seems to me that, despite the above, it's still worth trying to\nimprove upon the current state, to benefit from reduced space and CPU\nusage.\n\nAs it turns out we already separate out the alignment for a type, and a\ncolumn, between pg_type.typalign and pg_attribute.attalign. One way to\ntackle this would be to allow to specify, for byval types only, at\ncolumn creation time whether a column uses a 'struct-mappable' alignment\nor not. When set, set attalign to pg_type.typalign for alignment, when\nnot, to 'c'. By changing pg_dump in binary upgrade mode to emit the\nnecessary options, and by adding such options during bki processing,\nwe'd solve 1) and 2), but otherwise gain the benefits.\n\nAlternatively we could declare such a propert on the table level, but\nthat seems more restrictive, without a corresponding upside.\n\n\nIt's possible that we should do something related with a few varlena\ndatatypes. We currently use intalign for types like text, json, and as\nfar as I can tell that does not make all that much sense. They're not\nstruct mappable *anyway* (and if they were, they'd need to be 8 byte\naligned on common platforms, __alignof__(void*) is 8). We'd have to take\na bit of care to treat the varlena header as unaligned - but we need to\ndo so anyway, because of 1byte varlenas. Short varlenas seems to make it\nless crucial to pursue this, as the average datum that'd benefit is long\nenough to make padding a non-issue. So I don't think it'd make sense to\ntackle this project at the same time.\n\n\nTo fully benefit from the increased tuple deforming speed, it might be\nbeneficial to branch very early between two different versions within\nslot_deform_heap_tuple, having determined whether there's any byval\ncolumns with alignment requirements at slot creation /\nExecSetSlotDescriptor() time (or even set a different callback\ngetsomeattrs callback, but that's a bit more complicated).\n\n\nThoughts?\n\n\nIndependent of the above, I think it might make sense to replace\npg_attribute.attalign with a smallint or such. It's a bit absurd that we\nneed code like\n#define att_align_nominal(cur_offset, attalign) \\\n( \\\n\t((attalign) == 'i') ? INTALIGN(cur_offset) : \\\n\t (((attalign) == 'c') ? (uintptr_t) (cur_offset) : \\\n\t (((attalign) == 'd') ? DOUBLEALIGN(cur_offset) : \\\n\t ( \\\n\t\t\tAssertMacro((attalign) == 's'), \\\n\t\t\tSHORTALIGN(cur_offset) \\\n\t ))) \\\n)\n\ninstead of just using TYPEALIGN(). There's no need to adapt CREATE TYPE,\nor pg_type - we should just store the number in\npg_attribute.attalign. That keeps CREATE TYPE 32/64bit/arch independent,\ndoesn't require reconstructing c/s/i/d in pg_dump, simplifies the\ngenerated code [1], and would also \"just\" work for what I described\nearlier in this email.\n\nGreetings,\n\nAndres Freund\n\n\n[1] E.g. as a function of (void *ptr, char attalign) this ends up with assembly\nlike\n\t.globl\talignme\n\t.type\talignme, @function\nalignme:\n.LFB210:\n\t.cfi_startproc\n\tmovq\t%rdi, %rax\n\tcmpb\t$105, %sil\n\tje\t.L496\n\tcmpb\t$99, %sil\n\tje\t.L492\n\taddq\t$7, %rax\n\tleaq\t1(%rdi), %rdi\n\tandq\t$-8, %rax\n\tandq\t$-2, %rdi\n\tcmpb\t$100, %sil\n\tcmovne\t%rdi, %rax\n.L492:\n\tret\n\t.p2align 4,,10\n\t.p2align 3\n.L496:\n\taddq\t$3, %rax\n\tandq\t$-4, %rax\n\tret\n\t.cfi_endproc\n\nusing (void *ptr, int8 attalign) instead yields:\n\t.globl\talignme2\n\t.type\talignme2, @function\nalignme2:\n.LFB211:\n\t.cfi_startproc\n\tmovzbl\t%sil, %esi\n\tleal\t-1(%rsi), %eax\n\tcltq\n\tnegl\t%esi\n\taddq\t%rax, %rdi\n\tmovslq\t%esi, %rax\n\tandq\t%rdi, %rax\n\tret\n\t.cfi_endproc\n\n\n","msg_date":"Thu, 31 Oct 2019 11:48:21 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Removing alignment padding for byval types"},{"msg_contents":"On Thu, Oct 31, 2019 at 11:48:21AM -0700, Andres Freund wrote:\n>Hi,\n>\n>We currently align byval types such as int4/8, float4/8, timestamp *,\n>date etc, even though we mostly don't need to. When tuples are deformed,\n>all byval types are copied out from the tuple data into the\n>corresponding Datum array, therefore the original alignment in the tuple\n>data doesn't matter. This is different from byref types, where the\n>Datum formed will often be a pointer into the tuple data.\n>\n>While there are some older systems where it could be a bit slower to\n>copy data out from unaligned positions into the datum array, this is\n>more than bought back by the next point:\n>\n>\n>The fact that these types are aligned has substantial costs:\n>\n>For one, we often waste substantial amounts of space inside tables with\n>alignment padding. It's not uncommon to see about 30% or more of space\n>wasted (especially when taking alignment of the first column into\n>account).\n>\n>For another, and this I think is less obvious, we actually waste\n>substantial amounts of CPU maintaining the alignment. This is primarily\n>the case because we have to perform to align the pointer to the next\n>field during tuple [de]forming. Those instructions [1] have to be\n>executed taking time, but what's worse, they also reduce the ability of\n>out-of-order execution to hide latencies. There's a hard dependency on\n>knowing the offset to the next column to be able to continue with the\n>next column.\n>\n\nRight. Reducing this overhead was one of the goals to allow \"logical\nordering\" of columns in a table (while arbitrarily reordering the\nphysical ones), but that patch got out of hand pretty quickly. Also,\nit'd still keep some of the overhead, because it was not removing the\nalignment/padding entirely.\n\n>\n>There's two reasons why we can't just set the alignment for these types\n>to 'c'.\n>1) pg_upgrade, for fairly obvious reasons\n>2) We map catalog table rows to structs, in a *lot* of places.\n>\n>\n>It seems to me that, despite the above, it's still worth trying to\n>improve upon the current state, to benefit from reduced space and CPU\n>usage.\n>\n>As it turns out we already separate out the alignment for a type, and a\n>column, between pg_type.typalign and pg_attribute.attalign. One way to\n>tackle this would be to allow to specify, for byval types only, at\n>column creation time whether a column uses a 'struct-mappable' alignment\n>or not. When set, set attalign to pg_type.typalign for alignment, when\n>not, to 'c'. By changing pg_dump in binary upgrade mode to emit the\n>necessary options, and by adding such options during bki processing,\n>we'd solve 1) and 2), but otherwise gain the benefits.\n>\n>Alternatively we could declare such a propert on the table level, but\n>that seems more restrictive, without a corresponding upside.\n>\n\nI don't know, but it seems entirely sufficient specifying this at the\ntable level, no? What would be the use case for removing padding for\nonly some of the columns? I don't see the use case for that.\n\n>\n>It's possible that we should do something related with a few varlena\n>datatypes. We currently use intalign for types like text, json, and as\n>far as I can tell that does not make all that much sense. They're not\n>struct mappable *anyway* (and if they were, they'd need to be 8 byte\n>aligned on common platforms, __alignof__(void*) is 8). We'd have to take\n>a bit of care to treat the varlena header as unaligned - but we need to\n>do so anyway, because of 1byte varlenas. Short varlenas seems to make it\n>less crucial to pursue this, as the average datum that'd benefit is long\n>enough to make padding a non-issue. So I don't think it'd make sense to\n>tackle this project at the same time.\n>\n\nNot sure, but how come it's not failing on the picky platforms, then? On\nx86 it's probably OK because it's pretty permissive, but I'd expect some\nplatforms (s390, parisc, itanium, powerpc, ...) to be much pickier.\n\n>\n>To fully benefit from the increased tuple deforming speed, it might be\n>beneficial to branch very early between two different versions within\n>slot_deform_heap_tuple, having determined whether there's any byval\n>columns with alignment requirements at slot creation /\n>ExecSetSlotDescriptor() time (or even set a different callback\n>getsomeattrs callback, but that's a bit more complicated).\n>\n>\n>Thoughts?\n>\n\nSeems reasonable. I certainly agree this padding is pretty annoying, so\nif we can get rid of it without causing issues, that'd be nice. \n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 31 Oct 2019 20:15:12 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Removing alignment padding for byval types"},{"msg_contents":"Hi,\n\nOn 2019-10-31 20:15:12 +0100, Tomas Vondra wrote:\n> On Thu, Oct 31, 2019 at 11:48:21AM -0700, Andres Freund wrote:\n> > We currently align byval types such as int4/8, float4/8, timestamp *,\n> > date etc, even though we mostly don't need to. When tuples are deformed,\n> > all byval types are copied out from the tuple data into the\n> > corresponding Datum array, therefore the original alignment in the tuple\n> > data doesn't matter. This is different from byref types, where the\n> > Datum formed will often be a pointer into the tuple data.\n> > \n> > While there are some older systems where it could be a bit slower to\n> > copy data out from unaligned positions into the datum array, this is\n> > more than bought back by the next point:\n> > \n> > \n> > The fact that these types are aligned has substantial costs:\n> > \n> > For one, we often waste substantial amounts of space inside tables with\n> > alignment padding. It's not uncommon to see about 30% or more of space\n> > wasted (especially when taking alignment of the first column into\n> > account).\n> > \n> > For another, and this I think is less obvious, we actually waste\n> > substantial amounts of CPU maintaining the alignment. This is primarily\n> > the case because we have to perform to align the pointer to the next\n> > field during tuple [de]forming. Those instructions [1] have to be\n> > executed taking time, but what's worse, they also reduce the ability of\n> > out-of-order execution to hide latencies. There's a hard dependency on\n> > knowing the offset to the next column to be able to continue with the\n> > next column.\n> > \n> \n> Right. Reducing this overhead was one of the goals to allow \"logical\n> ordering\" of columns in a table (while arbitrarily reordering the\n> physical ones), but that patch got out of hand pretty quickly. Also,\n> it'd still keep some of the overhead, because it was not removing the\n> alignment/padding entirely.\n\nYea. It'd keep just about all the CPU overhead, because we'd still need\nto align as soon as there is a preceding nulled or varlena colum.\n\nThere's still some benefit for logical column order, as grouping NOT\nNULL fixed-length columns at the start is beneficial. And it's also\nbeneficial to have frequently accessed columns at the start. But I think\nthis proposal gains a lot of the space related benefits, at a much lower\ncomplexity, together with a lot of other benefits.\n\n\n> > There's two reasons why we can't just set the alignment for these types\n> > to 'c'.\n> > 1) pg_upgrade, for fairly obvious reasons\n> > 2) We map catalog table rows to structs, in a *lot* of places.\n> > \n> > \n> > It seems to me that, despite the above, it's still worth trying to\n> > improve upon the current state, to benefit from reduced space and CPU\n> > usage.\n> > \n> > As it turns out we already separate out the alignment for a type, and a\n> > column, between pg_type.typalign and pg_attribute.attalign. One way to\n> > tackle this would be to allow to specify, for byval types only, at\n> > column creation time whether a column uses a 'struct-mappable' alignment\n> > or not. When set, set attalign to pg_type.typalign for alignment, when\n> > not, to 'c'. By changing pg_dump in binary upgrade mode to emit the\n> > necessary options, and by adding such options during bki processing,\n> > we'd solve 1) and 2), but otherwise gain the benefits.\n> > \n> > Alternatively we could declare such a propert on the table level, but\n> > that seems more restrictive, without a corresponding upside.\n\n> I don't know, but it seems entirely sufficient specifying this at the\n> table level, no? What would be the use case for removing padding for\n> only some of the columns? I don't see the use case for that.\n\nWell, if we had it on a per-table level, we'd also align\na) catalog table columns that follow the first varlena column - which we don't need\n to align, as they can't be accessed via mapping\nb) columns in pg_upgraded tables that have been added after the upgrade\n\n\n> > It's possible that we should do something related with a few varlena\n> > datatypes. We currently use intalign for types like text, json, and as\n> > far as I can tell that does not make all that much sense. They're not\n> > struct mappable *anyway* (and if they were, they'd need to be 8 byte\n> > aligned on common platforms, __alignof__(void*) is 8). We'd have to take\n> > a bit of care to treat the varlena header as unaligned - but we need to\n> > do so anyway, because of 1byte varlenas. Short varlenas seems to make it\n> > less crucial to pursue this, as the average datum that'd benefit is long\n> > enough to make padding a non-issue. So I don't think it'd make sense to\n> > tackle this project at the same time.\n> > \n> \n> Not sure, but how come it's not failing on the picky platforms, then? On\n> x86 it's probably OK because it's pretty permissive, but I'd expect some\n> platforms (s390, parisc, itanium, powerpc, ...) to be much pickier.\n\nI'm not quite following? As I said, we already need to use alignment\naware code due to caring for short varlenas. And these types aren't\nactually struct mappable.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 31 Oct 2019 12:24:33 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: Removing alignment padding for byval types"},{"msg_contents":"On Thu, Oct 31, 2019 at 12:24:33PM -0700, Andres Freund wrote:\n>Hi,\n>\n>On 2019-10-31 20:15:12 +0100, Tomas Vondra wrote:\n>> On Thu, Oct 31, 2019 at 11:48:21AM -0700, Andres Freund wrote:\n>> > We currently align byval types such as int4/8, float4/8, timestamp *,\n>> > date etc, even though we mostly don't need to. When tuples are deformed,\n>> > all byval types are copied out from the tuple data into the\n>> > corresponding Datum array, therefore the original alignment in the tuple\n>> > data doesn't matter. This is different from byref types, where the\n>> > Datum formed will often be a pointer into the tuple data.\n>> >\n>> > While there are some older systems where it could be a bit slower to\n>> > copy data out from unaligned positions into the datum array, this is\n>> > more than bought back by the next point:\n>> >\n>> >\n>> > The fact that these types are aligned has substantial costs:\n>> >\n>> > For one, we often waste substantial amounts of space inside tables with\n>> > alignment padding. It's not uncommon to see about 30% or more of space\n>> > wasted (especially when taking alignment of the first column into\n>> > account).\n>> >\n>> > For another, and this I think is less obvious, we actually waste\n>> > substantial amounts of CPU maintaining the alignment. This is primarily\n>> > the case because we have to perform to align the pointer to the next\n>> > field during tuple [de]forming. Those instructions [1] have to be\n>> > executed taking time, but what's worse, they also reduce the ability of\n>> > out-of-order execution to hide latencies. There's a hard dependency on\n>> > knowing the offset to the next column to be able to continue with the\n>> > next column.\n>> >\n>>\n>> Right. Reducing this overhead was one of the goals to allow \"logical\n>> ordering\" of columns in a table (while arbitrarily reordering the\n>> physical ones), but that patch got out of hand pretty quickly. Also,\n>> it'd still keep some of the overhead, because it was not removing the\n>> alignment/padding entirely.\n>\n>Yea. It'd keep just about all the CPU overhead, because we'd still need\n>to align as soon as there is a preceding nulled or varlena colum.\n>\n>There's still some benefit for logical column order, as grouping NOT\n>NULL fixed-length columns at the start is beneficial. And it's also\n>beneficial to have frequently accessed columns at the start. But I think\n>this proposal gains a lot of the space related benefits, at a much lower\n>complexity, together with a lot of other benefits.\n>\n\n+1\n\n>\n>> > There's two reasons why we can't just set the alignment for these types\n>> > to 'c'.\n>> > 1) pg_upgrade, for fairly obvious reasons\n>> > 2) We map catalog table rows to structs, in a *lot* of places.\n>> >\n>> >\n>> > It seems to me that, despite the above, it's still worth trying to\n>> > improve upon the current state, to benefit from reduced space and CPU\n>> > usage.\n>> >\n>> > As it turns out we already separate out the alignment for a type, and a\n>> > column, between pg_type.typalign and pg_attribute.attalign. One way to\n>> > tackle this would be to allow to specify, for byval types only, at\n>> > column creation time whether a column uses a 'struct-mappable' alignment\n>> > or not. When set, set attalign to pg_type.typalign for alignment, when\n>> > not, to 'c'. By changing pg_dump in binary upgrade mode to emit the\n>> > necessary options, and by adding such options during bki processing,\n>> > we'd solve 1) and 2), but otherwise gain the benefits.\n>> >\n>> > Alternatively we could declare such a propert on the table level, but\n>> > that seems more restrictive, without a corresponding upside.\n>\n>> I don't know, but it seems entirely sufficient specifying this at the\n>> table level, no? What would be the use case for removing padding for\n>> only some of the columns? I don't see the use case for that.\n>\n>Well, if we had it on a per-table level, we'd also align\n>a) catalog table columns that follow the first varlena column - which we don't need\n> to align, as they can't be accessed via mapping\n>b) columns in pg_upgraded tables that have been added after the upgrade\n>\n\nHmm, OK. I think the question is whether it's worth the extra\ncomplexity. I'd say it's not, but perhaps I'm wrong.\n\n>\n>> > It's possible that we should do something related with a few varlena\n>> > datatypes. We currently use intalign for types like text, json, and as\n>> > far as I can tell that does not make all that much sense. They're not\n>> > struct mappable *anyway* (and if they were, they'd need to be 8 byte\n>> > aligned on common platforms, __alignof__(void*) is 8). We'd have to take\n>> > a bit of care to treat the varlena header as unaligned - but we need to\n>> > do so anyway, because of 1byte varlenas. Short varlenas seems to make it\n>> > less crucial to pursue this, as the average datum that'd benefit is long\n>> > enough to make padding a non-issue. So I don't think it'd make sense to\n>> > tackle this project at the same time.\n>> >\n>>\n>> Not sure, but how come it's not failing on the picky platforms, then? On\n>> x86 it's probably OK because it's pretty permissive, but I'd expect some\n>> platforms (s390, parisc, itanium, powerpc, ...) to be much pickier.\n>\n>I'm not quite following? As I said, we already need to use alignment\n>aware code due to caring for short varlenas. And these types aren't\n>actually struct mappable.\n>\n\nSorry, I misread/misunderstood what you wrote.\n\n\ncheers\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 31 Oct 2019 20:45:34 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Removing alignment padding for byval types"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nWe currently align byval types such as int4/8, float4/8, timestamp *,\\ndate etc, even though we mostly don't need to. When tuples are deformed,\\nall byval types are copied out from the tuple data into the\\ncorresponding Datum array, therefore the original alignment in the tuple\\ndata doesn't matter. This is different from byref types, where the\\nDatum formed will often be a pointer into the tuple data.\\n\\nWhile there are some older systems where it could be a bit slower to\\ncopy data out from unaligned positions into the datum array, this is\\nmore than bought back by the next point:\\n\\n\\nThe fact that these types are aligned has substantial costs:\\n\\nFor one, we often waste substantial amounts of space inside tables with\\nalignment padding. It's not uncommon to see about 30% or more of space\\nwasted (especially when taking alignment of the first column into\\naccount).\\n\\nFor another, and this I think is less obvious, we actually waste\\nsubstantial amounts of CPU maintaining the alignment. This is primarily\\nthe case because we have to perform to align the pointer to the next\\nfield during tuple [de]forming. Those instructions [1] have to be\\nexecuted taking time, but what's worse, they also reduce the ability of\\nout-of-order execution to hide latencies. There's a hard dependency on\\nknowing the offset to the next column to be able to continue with the\\nnext column.\\n\\n\\nThere's two reasons why we can't just set the alignment for these types\\nto 'c'.\\n1) pg_upgrade, for fairly obvious reasons\\n2) We map catalog table rows to structs, in a *lot* of places.\\n\\n\\nIt seems to me that, despite the above, it's still worth trying to\\nimprove upon the current state, to benefit from reduced space and CPU\\nusage.\\n\\nAs it turns out we already separate out the alignment for a type, and a\\ncolumn, between pg_type.typalign and pg_attribute.attalign. One way to\\ntackle this would be to allow to specify, for byval types only, at\\ncolumn creation time whether a column uses a 'struct-mappable' alignment\\nor not. When set, set attalign to pg_type.typalign for alignment, when\\nnot, to 'c'. By changing pg_dump in binary upgrade mode to emit the\\nnecessary options, and by adding such options during bki processing,\\nwe'd solve 1) and 2), but otherwise gain the benefits.\\n\\nAlternatively we could declare such a propert on the table level, but\\nthat seems more restrictive, without a corresponding upside.\\n\\n\\nIt's possible that we should do something related with a few varlena\\ndatatypes. We currently use intalign for types like text, json, and as\\nfar as I can tell that does not make all that much sense. They're not\\nstruct mappable *anyway* (and if they were, they'd need to be 8 byte\\naligned on common platforms, __alignof__(void*) is 8). We'd have to take\\na bit of care to treat the varlena header as unaligned - but we need to\\ndo so anyway, because of 1byte varlenas. Short varlenas seems to make it\\nless crucial to pursue this, as the average datum that'd benefit is long\\nenough to make padding a non-issue. So I don't think it'd make sense to\\ntackle this project at the same time.\\n\\n\\nTo fully benefit from the increased tuple deforming speed, it might be\\nbeneficial to branch very early between two different versions within\\nslot_deform_heap_tuple, having determined whether there's any byval\\ncolumns with alignment requirements at slot creation /\\nExecSetSlotDescriptor() time (or even set a different callback\\ngetsomeattrs callback, but that's a bit more complicated).\\n\\n\\nThoughts?\\n\\n\\nIndependent of the above, I think it might make sense to replace\\npg_attribute.attalign with a smallint or such. It's a bit absurd that we\\nneed code like\\n#define att_align_nominal(cur_offset, attalign) \\\\\\n( \\\\\\n\\t((attalign) == 'i') ? INTALIGN(cur_offset) : \\\\\\n\\t (((attalign) == 'c') ? (uintptr_t) (cur_offset) : \\\\\\n\\t (((attalign) == 'd') ? DOUBLEALIGN(cur_offset) : \\\\\\n\\t ( \\\\\\n\\t\\t\\tAssertMacro((attalign) == 's'), \\\\\\n\\t\\t\\tSHORTALIGN(cur_offset) \\\\\\n\\t ))) \\\\\\n)\\n\\ninstead of just using TYPEALIGN(). There's no need to adapt CREATE TYPE,\\nor pg_type - we should just store the number in\\npg_attribute.attalign. That keeps CREATE TYPE 32/64bit/arch independent,\\ndoesn't require reconstructing c/s/i/d in pg_dump, simplifies the\\ngenerated code [1], and would also \\\"just\\\" work for what I described\\nearlier in this email.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n[1] E.g. as a function of (void *ptr, char attalign) this ends up with assembly\\nlike\\n\\t.globl\\talignme\\n\\t.type\\talignme, @function\\nalignme:\\n.LFB210:\\n\\t.cfi_startproc\\n\\tmovq\\t%rdi, %rax\\n\\tcmpb\\t$105, %sil\\n\\tje\\t.L496\\n\\tcmpb\\t$99, %sil\\n\\tje\\t.L492\\n\\taddq\\t$7, %rax\\n\\tleaq\\t1(%rdi), %rdi\\n\\tandq\\t$-8, %rax\\n\\tandq\\t$-2, %rdi\\n\\tcmpb\\t$100, %sil\\n\\tcmovne\\t%rdi, %rax\\n.L492:\\n\\tret\\n\\t.p2align 4,,10\\n\\t.p2align 3\\n.L496:\\n\\taddq\\t$3, %rax\\n\\tandq\\t$-4, %rax\\n\\tret\\n\\t.cfi_endproc\\n\\nusing (void *ptr, int8 attalign) instead yields:\\n\\t.globl\\talignme2\\n\\t.type\\talignme2, @function\\nalignme2:\\n.LFB211:\\n\\t.cfi_startproc\\n\\tmovzbl\\t%sil, %esi\\n\\tleal\\t-1(%rsi), %eax\\n\\tcltq\\n\\tnegl\\t%esi\\n\\taddq\\t%rax, %rdi\\n\\tmovslq\\t%esi, %rax\\n\\tandq\\t%rdi, %rax\\n\\tret\\n\\t.cfi_endproc\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 11:48:21 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Removing alignment padding for byval types\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 11:48:21AM -0700, Andres Freund wrote:\\n>Hi,\\n>\\n>We currently align byval types such as int4/8, float4/8, timestamp *,\\n>date etc, even though we mostly don't need to. When tuples are deformed,\\n>all byval types are copied out from the tuple data into the\\n>corresponding Datum array, therefore the original alignment in the tuple\\n>data doesn't matter. This is different from byref types, where the\\n>Datum formed will often be a pointer into the tuple data.\\n>\\n>While there are some older systems where it could be a bit slower to\\n>copy data out from unaligned positions into the datum array, this is\\n>more than bought back by the next point:\\n>\\n>\\n>The fact that these types are aligned has substantial costs:\\n>\\n>For one, we often waste substantial amounts of space inside tables with\\n>alignment padding. It's not uncommon to see about 30% or more of space\\n>wasted (especially when taking alignment of the first column into\\n>account).\\n>\\n>For another, and this I think is less obvious, we actually waste\\n>substantial amounts of CPU maintaining the alignment. This is primarily\\n>the case because we have to perform to align the pointer to the next\\n>field during tuple [de]forming. Those instructions [1] have to be\\n>executed taking time, but what's worse, they also reduce the ability of\\n>out-of-order execution to hide latencies. There's a hard dependency on\\n>knowing the offset to the next column to be able to continue with the\\n>next column.\\n>\\n\\nRight. Reducing this overhead was one of the goals to allow \\\"logical\\nordering\\\" of columns in a table (while arbitrarily reordering the\\nphysical ones), but that patch got out of hand pretty quickly. Also,\\nit'd still keep some of the overhead, because it was not removing the\\nalignment/padding entirely.\\n\\n>\\n>There's two reasons why we can't just set the alignment for these types\\n>to 'c'.\\n>1) pg_upgrade, for fairly obvious reasons\\n>2) We map catalog table rows to structs, in a *lot* of places.\\n>\\n>\\n>It seems to me that, despite the above, it's still worth trying to\\n>improve upon the current state, to benefit from reduced space and CPU\\n>usage.\\n>\\n>As it turns out we already separate out the alignment for a type, and a\\n>column, between pg_type.typalign and pg_attribute.attalign. One way to\\n>tackle this would be to allow to specify, for byval types only, at\\n>column creation time whether a column uses a 'struct-mappable' alignment\\n>or not. When set, set attalign to pg_type.typalign for alignment, when\\n>not, to 'c'. By changing pg_dump in binary upgrade mode to emit the\\n>necessary options, and by adding such options during bki processing,\\n>we'd solve 1) and 2), but otherwise gain the benefits.\\n>\\n>Alternatively we could declare such a propert on the table level, but\\n>that seems more restrictive, without a corresponding upside.\\n>\\n\\nI don't know, but it seems entirely sufficient specifying this at the\\ntable level, no? What would be the use case for removing padding for\\nonly some of the columns? I don't see the use case for that.\\n\\n>\\n>It's possible that we should do something related with a few varlena\\n>datatypes. We currently use intalign for types like text, json, and as\\n>far as I can tell that does not make all that much sense. They're not\\n>struct mappable *anyway* (and if they were, they'd need to be 8 byte\\n>aligned on common platforms, __alignof__(void*) is 8). We'd have to take\\n>a bit of care to treat the varlena header as unaligned - but we need to\\n>do so anyway, because of 1byte varlenas. Short varlenas seems to make it\\n>less crucial to pursue this, as the average datum that'd benefit is long\\n>enough to make padding a non-issue. So I don't think it'd make sense to\\n>tackle this project at the same time.\\n>\\n\\nNot sure, but how come it's not failing on the picky platforms, then? On\\nx86 it's probably OK because it's pretty permissive, but I'd expect some\\nplatforms (s390, parisc, itanium, powerpc, ...) to be much pickier.\\n\\n>\\n>To fully benefit from the increased tuple deforming speed, it might be\\n>beneficial to branch very early between two different versions within\\n>slot_deform_heap_tuple, having determined whether there's any byval\\n>columns with alignment requirements at slot creation /\\n>ExecSetSlotDescriptor() time (or even set a different callback\\n>getsomeattrs callback, but that's a bit more complicated).\\n>\\n>\\n>Thoughts?\\n>\\n\\nSeems reasonable. I certainly agree this padding is pretty annoying, so\\nif we can get rid of it without causing issues, that'd be nice. \\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 20:15:12 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Removing alignment padding for byval types\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-10-31 20:15:12 +0100, Tomas Vondra wrote:\\n> On Thu, Oct 31, 2019 at 11:48:21AM -0700, Andres Freund wrote:\\n> > We currently align byval types such as int4/8, float4/8, timestamp *,\\n> > date etc, even though we mostly don't need to. When tuples are deformed,\\n> > all byval types are copied out from the tuple data into the\\n> > corresponding Datum array, therefore the original alignment in the tuple\\n> > data doesn't matter. This is different from byref types, where the\\n> > Datum formed will often be a pointer into the tuple data.\\n> > \\n> > While there are some older systems where it could be a bit slower to\\n> > copy data out from unaligned positions into the datum array, this is\\n> > more than bought back by the next point:\\n> > \\n> > \\n> > The fact that these types are aligned has substantial costs:\\n> > \\n> > For one, we often waste substantial amounts of space inside tables with\\n> > alignment padding. It's not uncommon to see about 30% or more of space\\n> > wasted (especially when taking alignment of the first column into\\n> > account).\\n> > \\n> > For another, and this I think is less obvious, we actually waste\\n> > substantial amounts of CPU maintaining the alignment. This is primarily\\n> > the case because we have to perform to align the pointer to the next\\n> > field during tuple [de]forming. Those instructions [1] have to be\\n> > executed taking time, but what's worse, they also reduce the ability of\\n> > out-of-order execution to hide latencies. There's a hard dependency on\\n> > knowing the offset to the next column to be able to continue with the\\n> > next column.\\n> > \\n> \\n> Right. Reducing this overhead was one of the goals to allow \\\"logical\\n> ordering\\\" of columns in a table (while arbitrarily reordering the\\n> physical ones), but that patch got out of hand pretty quickly. Also,\\n> it'd still keep some of the overhead, because it was not removing the\\n> alignment/padding entirely.\\n\\nYea. It'd keep just about all the CPU overhead, because we'd still need\\nto align as soon as there is a preceding nulled or varlena colum.\\n\\nThere's still some benefit for logical column order, as grouping NOT\\nNULL fixed-length columns at the start is beneficial. And it's also\\nbeneficial to have frequently accessed columns at the start. But I think\\nthis proposal gains a lot of the space related benefits, at a much lower\\ncomplexity, together with a lot of other benefits.\\n\\n\\n> > There's two reasons why we can't just set the alignment for these types\\n> > to 'c'.\\n> > 1) pg_upgrade, for fairly obvious reasons\\n> > 2) We map catalog table rows to structs, in a *lot* of places.\\n> > \\n> > \\n> > It seems to me that, despite the above, it's still worth trying to\\n> > improve upon the current state, to benefit from reduced space and CPU\\n> > usage.\\n> > \\n> > As it turns out we already separate out the alignment for a type, and a\\n> > column, between pg_type.typalign and pg_attribute.attalign. One way to\\n> > tackle this would be to allow to specify, for byval types only, at\\n> > column creation time whether a column uses a 'struct-mappable' alignment\\n> > or not. When set, set attalign to pg_type.typalign for alignment, when\\n> > not, to 'c'. By changing pg_dump in binary upgrade mode to emit the\\n> > necessary options, and by adding such options during bki processing,\\n> > we'd solve 1) and 2), but otherwise gain the benefits.\\n> > \\n> > Alternatively we could declare such a propert on the table level, but\\n> > that seems more restrictive, without a corresponding upside.\\n\\n> I don't know, but it seems entirely sufficient specifying this at the\\n> table level, no? What would be the use case for removing padding for\\n> only some of the columns? I don't see the use case for that.\\n\\nWell, if we had it on a per-table level, we'd also align\\na) catalog table columns that follow the first varlena column - which we don't need\\n to align, as they can't be accessed via mapping\\nb) columns in pg_upgraded tables that have been added after the upgrade\\n\\n\\n> > It's possible that we should do something related with a few varlena\\n> > datatypes. We currently use intalign for types like text, json, and as\\n> > far as I can tell that does not make all that much sense. They're not\\n> > struct mappable *anyway* (and if they were, they'd need to be 8 byte\\n> > aligned on common platforms, __alignof__(void*) is 8). We'd have to take\\n> > a bit of care to treat the varlena header as unaligned - but we need to\\n> > do so anyway, because of 1byte varlenas. Short varlenas seems to make it\\n> > less crucial to pursue this, as the average datum that'd benefit is long\\n> > enough to make padding a non-issue. So I don't think it'd make sense to\\n> > tackle this project at the same time.\\n> > \\n> \\n> Not sure, but how come it's not failing on the picky platforms, then? On\\n> x86 it's probably OK because it's pretty permissive, but I'd expect some\\n> platforms (s390, parisc, itanium, powerpc, ...) to be much pickier.\\n\\nI'm not quite following? As I said, we already need to use alignment\\naware code due to caring for short varlenas. And these types aren't\\nactually struct mappable.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 12:24:33 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Removing alignment padding for byval types\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 12:24:33PM -0700, Andres Freund wrote:\\n>Hi,\\n>\\n>On 2019-10-31 20:15:12 +0100, Tomas Vondra wrote:\\n>> On Thu, Oct 31, 2019 at 11:48:21AM -0700, Andres Freund wrote:\\n>> > We currently align byval types such as int4/8, float4/8, timestamp *,\\n>> > date etc, even though we mostly don't need to. When tuples are deformed,\\n>> > all byval types are copied out from the tuple data into the\\n>> > corresponding Datum array, therefore the original alignment in the tuple\\n>> > data doesn't matter. This is different from byref types, where the\\n>> > Datum formed will often be a pointer into the tuple data.\\n>> >\\n>> > While there are some older systems where it could be a bit slower to\\n>> > copy data out from unaligned positions into the datum array, this is\\n>> > more than bought back by the next point:\\n>> >\\n>> >\\n>> > The fact that these types are aligned has substantial costs:\\n>> >\\n>> > For one, we often waste substantial amounts of space inside tables with\\n>> > alignment padding. It's not uncommon to see about 30% or more of space\\n>> > wasted (especially when taking alignment of the first column into\\n>> > account).\\n>> >\\n>> > For another, and this I think is less obvious, we actually waste\\n>> > substantial amounts of CPU maintaining the alignment. This is primarily\\n>> > the case because we have to perform to align the pointer to the next\\n>> > field during tuple [de]forming. Those instructions [1] have to be\\n>> > executed taking time, but what's worse, they also reduce the ability of\\n>> > out-of-order execution to hide latencies. There's a hard dependency on\\n>> > knowing the offset to the next column to be able to continue with the\\n>> > next column.\\n>> >\\n>>\\n>> Right. Reducing this overhead was one of the goals to allow \\\"logical\\n>> ordering\\\" of columns in a table (while arbitrarily reordering the\\n>> physical ones), but that patch got out of hand pretty quickly. Also,\\n>> it'd still keep some of the overhead, because it was not removing the\\n>> alignment/padding entirely.\\n>\\n>Yea. It'd keep just about all the CPU overhead, because we'd still need\\n>to align as soon as there is a preceding nulled or varlena colum.\\n>\\n>There's still some benefit for logical column order, as grouping NOT\\n>NULL fixed-length columns at the start is beneficial. And it's also\\n>beneficial to have frequently accessed columns at the start. But I think\\n>this proposal gains a lot of the space related benefits, at a much lower\\n>complexity, together with a lot of other benefits.\\n>\\n\\n+1\\n\\n>\\n>> > There's two reasons why we can't just set the alignment for these types\\n>> > to 'c'.\\n>> > 1) pg_upgrade, for fairly obvious reasons\\n>> > 2) We map catalog table rows to structs, in a *lot* of places.\\n>> >\\n>> >\\n>> > It seems to me that, despite the above, it's still worth trying to\\n>> > improve upon the current state, to benefit from reduced space and CPU\\n>> > usage.\\n>> >\\n>> > As it turns out we already separate out the alignment for a type, and a\\n>> > column, between pg_type.typalign and pg_attribute.attalign. One way to\\n>> > tackle this would be to allow to specify, for byval types only, at\\n>> > column creation time whether a column uses a 'struct-mappable' alignment\\n>> > or not. When set, set attalign to pg_type.typalign for alignment, when\\n>> > not, to 'c'. By changing pg_dump in binary upgrade mode to emit the\\n>> > necessary options, and by adding such options during bki processing,\\n>> > we'd solve 1) and 2), but otherwise gain the benefits.\\n>> >\\n>> > Alternatively we could declare such a propert on the table level, but\\n>> > that seems more restrictive, without a corresponding upside.\\n>\\n>> I don't know, but it seems entirely sufficient specifying this at the\\n>> table level, no? What would be the use case for removing padding for\\n>> only some of the columns? I don't see the use case for that.\\n>\\n>Well, if we had it on a per-table level, we'd also align\\n>a) catalog table columns that follow the first varlena column - which we don't need\\n> to align, as they can't be accessed via mapping\\n>b) columns in pg_upgraded tables that have been added after the upgrade\\n>\\n\\nHmm, OK. I think the question is whether it's worth the extra\\ncomplexity. I'd say it's not, but perhaps I'm wrong.\\n\\n>\\n>> > It's possible that we should do something related with a few varlena\\n>> > datatypes. We currently use intalign for types like text, json, and as\\n>> > far as I can tell that does not make all that much sense. They're not\\n>> > struct mappable *anyway* (and if they were, they'd need to be 8 byte\\n>> > aligned on common platforms, __alignof__(void*) is 8). We'd have to take\\n>> > a bit of care to treat the varlena header as unaligned - but we need to\\n>> > do so anyway, because of 1byte varlenas. Short varlenas seems to make it\\n>> > less crucial to pursue this, as the average datum that'd benefit is long\\n>> > enough to make padding a non-issue. So I don't think it'd make sense to\\n>> > tackle this project at the same time.\\n>> >\\n>>\\n>> Not sure, but how come it's not failing on the picky platforms, then? On\\n>> x86 it's probably OK because it's pretty permissive, but I'd expect some\\n>> platforms (s390, parisc, itanium, powerpc, ...) to be much pickier.\\n>\\n>I'm not quite following? As I said, we already need to use alignment\\n>aware code due to caring for short varlenas. And these types aren't\\n>actually struct mappable.\\n>\\n\\nSorry, I misread/misunderstood what you wrote.\\n\\n\\ncheers\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 20:45:34 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Removing alignment padding for byval types\"\n }\n]","isConversational":false}}},{"rowIdx":49383,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"This patch allows the superuser to grant passwordless connection rights\nin postgres_fdw user mappings.\n\n\nThe patch is authored by my colleague Craig Ringer, with slight bitrot\nfixed by me.\n\n\nOne use case for this is with passphrase-protected client certificates,\na patch for which will follow shortly.\n\n\nHere are Craig's remarks on the patch:\n\n \n postgres_fdw denies a non-superuser the ability to establish a\nconnection that\n doesn't have a password in the connection string, or one that fails\nto actually\n use the password in authentication. This is to stop the unprivileged\nuser using\n OS-level authentication as the postgres server (peer, ident, trust).\nIt also\n stops unauthorized use of local credentials like .pgpass, a service\nfile,\n client certificate files, etc.\n \n Add the ability for a superuser to create user mappings that\noverride this\n behaviour by setting the passwordless_ok attribute to true in a user\nmapping\n for a non-superuser. The non-superuser gains the ability to use the\nFDW the\n mapping applies to even if there's no password in their mapping or\nin the\n connection string.\n \n This is only safe if the superuser has established that the local\nserver is\n configured safely. It must be configured not to allow\n trust/peer/ident/sspi/gssapi auth to allow the OS user the postgres\nserver runs\n as to log in to postgres as a superuser. Client certificate keys can\nbe used\n too, if accessible. But the superuser can already GRANT superrole TO\n normalrole, so it's not any sort of new power.\n \n\ncheers\n\n\nandrew\n\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 31 Oct 2019 16:58:20 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Allow superuser to grant passwordless connection rights on\n postgres_fdw"},{"msg_contents":"On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\n<andrew.dunstan@2ndquadrant.com> wrote:\n> This patch allows the superuser to grant passwordless connection rights\n> in postgres_fdw user mappings.\n\nThis is clearly something that we need, as the current code seems\nwoefully ignorant of the fact that passwords are not the only\nauthentication method supported by PostgreSQL, nor even the most\nsecure.\n\nBut, I do wonder a bit if we ought to think harder about the overall\nauthentication model for FDW. Like, maybe we'd take a different view\nof how to solve this particular piece of the problem if we were\nthinking about how FDWs could do LDAP authentication, SSL\nauthentication, credentials forwarding...\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 1 Nov 2019 12:58:51 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Allow superuser to grant passwordless connection rights on\n postgres_fdw"},{"msg_contents":"\nOn 11/1/19 12:58 PM, Robert Haas wrote:\n> On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\n> <andrew.dunstan@2ndquadrant.com> wrote:\n>> This patch allows the superuser to grant passwordless connection rights\n>> in postgres_fdw user mappings.\n> This is clearly something that we need, as the current code seems\n> woefully ignorant of the fact that passwords are not the only\n> authentication method supported by PostgreSQL, nor even the most\n> secure.\n>\n> But, I do wonder a bit if we ought to think harder about the overall\n> authentication model for FDW. Like, maybe we'd take a different view\n> of how to solve this particular piece of the problem if we were\n> thinking about how FDWs could do LDAP authentication, SSL\n> authentication, credentials forwarding...\n>\n\n\nI'm certainly open to alternatives.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Fri, 1 Nov 2019 14:00:27 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Allow superuser to grant passwordless connection rights on\n postgres_fdw"},{"msg_contents":"Greetings,\n\n* Andrew Dunstan (andrew.dunstan@2ndquadrant.com) wrote:\n> On 11/1/19 12:58 PM, Robert Haas wrote:\n> > On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\n> > <andrew.dunstan@2ndquadrant.com> wrote:\n> >> This patch allows the superuser to grant passwordless connection rights\n> >> in postgres_fdw user mappings.\n> > This is clearly something that we need, as the current code seems\n> > woefully ignorant of the fact that passwords are not the only\n> > authentication method supported by PostgreSQL, nor even the most\n> > secure.\n> >\n> > But, I do wonder a bit if we ought to think harder about the overall\n> > authentication model for FDW. Like, maybe we'd take a different view\n> > of how to solve this particular piece of the problem if we were\n> > thinking about how FDWs could do LDAP authentication, SSL\n> > authentication, credentials forwarding...\n> \n> I'm certainly open to alternatives.\n\nI've long felt that the way to handle this kind of requirement is to\nhave a \"trusted remote server\" kind of option- where the local server\nauthenticates to the remote server as a *server* and then says \"this is\nthe user on this server, and this is the user that this user wishes to\nbe\" and the remote server is then able to decide if they accept that, or\nnot.\n\nTo be specific, there would be some kind of 'trust' established between\nthe servers and only if there is some kind of server-level\nauthentication, eg: dual TLS auth, or dual GSSAPI auth; and then, a\nmapping is defined for that server, which specifies what remote user is\nallowed to log in as what local user.\n\nThis would be a server-to-server auth arrangement, and is quite\ndifferent from credential forwarding, or similar. I am certainly also a\nhuge fan of the idea that we support Kerberos/GSSAPI credential\nforwarding / delegation, where a client willingly forwards to the PG\nserver a set of credentials which then allow the PG server to\nauthenticate as that user to another system (eg: through an FDW to\nanother PG server).\n\nOf course, as long as we're talking pie-in-the-sky ideas, I would\ncertainly be entirely for supporting both. ;)\n\nThanks,\n\nStephen","msg_date":"Sun, 3 Nov 2019 23:20:05 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Allow superuser to grant passwordless connection rights on\n postgres_fdw"},{"msg_contents":"On Mon, 4 Nov 2019 at 12:20, Stephen Frost <sfrost@snowman.net> wrote:\n\n> Greetings,\n>\n> * Andrew Dunstan (andrew.dunstan@2ndquadrant.com) wrote:\n> > On 11/1/19 12:58 PM, Robert Haas wrote:\n> > > On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\n> > > <andrew.dunstan@2ndquadrant.com> wrote:\n> > >> This patch allows the superuser to grant passwordless connection\n> rights\n> > >> in postgres_fdw user mappings.\n> > > This is clearly something that we need, as the current code seems\n> > > woefully ignorant of the fact that passwords are not the only\n> > > authentication method supported by PostgreSQL, nor even the most\n> > > secure.\n> > >\n> > > But, I do wonder a bit if we ought to think harder about the overall\n> > > authentication model for FDW. Like, maybe we'd take a different view\n> > > of how to solve this particular piece of the problem if we were\n> > > thinking about how FDWs could do LDAP authentication, SSL\n> > > authentication, credentials forwarding...\n> >\n> > I'm certainly open to alternatives.\n>\n> I've long felt that the way to handle this kind of requirement is to\n> have a \"trusted remote server\" kind of option- where the local server\n> authenticates to the remote server as a *server* and then says \"this is\n> the user on this server, and this is the user that this user wishes to\n> be\" and the remote server is then able to decide if they accept that, or\n> not.\n>\n\nThe original use case for the patch was to allow FDWs to use SSL/TLS client\ncertificates. Each user-mapping has its own certificate - there's a\nseparate patch to allow that. So there's no delegation of trust via\nKerberos etc in that particular case.\n\nI can see value in using Kerberos etc for that too though, as it separates\nauthorization and authentication in the same manner as most sensible\nsystems. You can say \"user postgres@foo is trusted to vet users so you can\nsafely hand out tickets for any bar@foo that postgres@foo says is legit\".\n\nI would strongly discourage allowing all users on host A to authenticate as\nuser postgres on host B. But with appropriate user-mappings support, we\ncould likely support that sort of model for both SSPI and Kerberos.\n\nA necessary prerequisite is that Pg be able to cope with passwordless\nuser-mappings though. Hence this patch.\n\n\n\n>\n> To be specific, there would be some kind of 'trust' established between\n> the servers and only if there is some kind of server-level\n> authentication, eg: dual TLS auth, or dual GSSAPI auth; and then, a\n> mapping is defined for that server, which specifies what remote user is\n> allowed to log in as what local user.\n>\n> This would be a server-to-server auth arrangement, and is quite\n> different from credential forwarding, or similar. I am certainly also a\n> huge fan of the idea that we support Kerberos/GSSAPI credential\n> forwarding / delegation, where a client willingly forwards to the PG\n> server a set of credentials which then allow the PG server to\n> authenticate as that user to another system (eg: through an FDW to\n> another PG server).\n>\n> Of course, as long as we're talking pie-in-the-sky ideas, I would\n> certainly be entirely for supporting both. ;)\n>\n> Thanks,\n>\n> Stephen\n>\n\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\n\nOn Mon, 4 Nov 2019 at 12:20, Stephen Frost <sfrost@snowman.net> wrote:Greetings,\n\n* Andrew Dunstan (andrew.dunstan@2ndquadrant.com) wrote:\n> On 11/1/19 12:58 PM, Robert Haas wrote:\n> > On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\n> > <andrew.dunstan@2ndquadrant.com> wrote:\n> >> This patch allows the superuser to grant passwordless connection rights\n> >> in postgres_fdw user mappings.\n> > This is clearly something that we need, as the current code seems\n> > woefully ignorant of the fact that passwords are not the only\n> > authentication method supported by PostgreSQL, nor even the most\n> > secure.\n> >\n> > But, I do wonder a bit if we ought to think harder about the overall\n> > authentication model for FDW. Like, maybe we'd take a different view\n> > of how to solve this particular piece of the problem if we were\n> > thinking about how FDWs could do LDAP authentication, SSL\n> > authentication, credentials forwarding...\n> \n> I'm certainly open to alternatives.\n\nI've long felt that the way to handle this kind of requirement is to\nhave a \"trusted remote server\" kind of option- where the local server\nauthenticates to the remote server as a *server* and then says \"this is\nthe user on this server, and this is the user that this user wishes to\nbe\" and the remote server is then able to decide if they accept that, or\nnot.The original use case for the patch was to allow FDWs to use SSL/TLS client certificates. Each user-mapping has its own certificate - there's a separate patch to allow that. So there's no delegation of trust via Kerberos etc in that particular case.I can see value in using Kerberos etc for that too though, as it separates authorization and authentication in the same manner as most sensible systems. You can say \"user postgres@foo is trusted to vet users so you can safely hand out tickets for any bar@foo that postgres@foo says is legit\".I would strongly discourage allowing all users on host A to authenticate as user postgres on host B. But with appropriate user-mappings support, we could likely support that sort of model for both SSPI and Kerberos.A necessary prerequisite is that Pg be able to cope with passwordless user-mappings though. Hence this patch. \n\nTo be specific, there would be some kind of 'trust' established between\nthe servers and only if there is some kind of server-level\nauthentication, eg: dual TLS auth, or dual GSSAPI auth; and then, a\nmapping is defined for that server, which specifies what remote user is\nallowed to log in as what local user.\n\nThis would be a server-to-server auth arrangement, and is quite\ndifferent from credential forwarding, or similar. I am certainly also a\nhuge fan of the idea that we support Kerberos/GSSAPI credential\nforwarding / delegation, where a client willingly forwards to the PG\nserver a set of credentials which then allow the PG server to\nauthenticate as that user to another system (eg: through an FDW to\nanother PG server).\n\nOf course, as long as we're talking pie-in-the-sky ideas, I would\ncertainly be entirely for supporting both. ;)\n\nThanks,\n\nStephen\n-- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise","msg_date":"Sun, 10 Nov 2019 17:35:36 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Allow superuser to grant passwordless connection rights on\n postgres_fdw"},{"msg_contents":"On Sun, Nov 10, 2019 at 4:35 AM Craig Ringer <craig@2ndquadrant.com> wrote:\n>\n> On Mon, 4 Nov 2019 at 12:20, Stephen Frost <sfrost@snowman.net> wrote:\n>>\n>> Greetings,\n>>\n>> * Andrew Dunstan (andrew.dunstan@2ndquadrant.com) wrote:\n>> > On 11/1/19 12:58 PM, Robert Haas wrote:\n>> > > On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\n>> > > <andrew.dunstan@2ndquadrant.com> wrote:\n>> > >> This patch allows the superuser to grant passwordless connection rights\n>> > >> in postgres_fdw user mappings.\n>> > > This is clearly something that we need, as the current code seems\n>> > > woefully ignorant of the fact that passwords are not the only\n>> > > authentication method supported by PostgreSQL, nor even the most\n>> > > secure.\n>> > >\n>> > > But, I do wonder a bit if we ought to think harder about the overall\n>> > > authentication model for FDW. Like, maybe we'd take a different view\n>> > > of how to solve this particular piece of the problem if we were\n>> > > thinking about how FDWs could do LDAP authentication, SSL\n>> > > authentication, credentials forwarding...\n>> >\n>> > I'm certainly open to alternatives.\n>>\n>> I've long felt that the way to handle this kind of requirement is to\n>> have a \"trusted remote server\" kind of option- where the local server\n>> authenticates to the remote server as a *server* and then says \"this is\n>> the user on this server, and this is the user that this user wishes to\n>> be\" and the remote server is then able to decide if they accept that, or\n>> not.\n>\n>\n> The original use case for the patch was to allow FDWs to use SSL/TLS client certificates. Each user-mapping has its own certificate - there's a separate patch to allow that. So there's no delegation of trust via Kerberos etc in that particular case.\n>\n> I can see value in using Kerberos etc for that too though, as it separates authorization and authentication in the same manner as most sensible systems. You can say \"user postgres@foo is trusted to vet users so you can safely hand out tickets for any bar@foo that postgres@foo says is legit\".\n>\n> I would strongly discourage allowing all users on host A to authenticate as user postgres on host B. But with appropriate user-mappings support, we could likely support that sort of model for both SSPI and Kerberos.\n>\n> A necessary prerequisite is that Pg be able to cope with passwordless user-mappings though. Hence this patch.\n>\n>\n\n\nYeah, I agree. Does anyone else want to weigh in on this? If nobody\nobjects I'd like to tidy this up and get it committed so we can add\nsupport for client certs in postgres_fdw, which is the real business\nat hand, and which I know from various offline comments a number of\npeople are keen to have.\n\ncheers\n\nandrew\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Mon, 25 Nov 2019 16:56:24 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Allow superuser to grant passwordless connection rights on\n postgres_fdw"},{"msg_contents":"Greetings,\n\n* Craig Ringer (craig@2ndquadrant.com) wrote:\n> On Mon, 4 Nov 2019 at 12:20, Stephen Frost <sfrost@snowman.net> wrote:\n> > I've long felt that the way to handle this kind of requirement is to\n> > have a \"trusted remote server\" kind of option- where the local server\n> > authenticates to the remote server as a *server* and then says \"this is\n> > the user on this server, and this is the user that this user wishes to\n> > be\" and the remote server is then able to decide if they accept that, or\n> > not.\n> \n> The original use case for the patch was to allow FDWs to use SSL/TLS client\n> certificates. Each user-mapping has its own certificate - there's a\n> separate patch to allow that. So there's no delegation of trust via\n> Kerberos etc in that particular case.\n> \n> I can see value in using Kerberos etc for that too though, as it separates\n> authorization and authentication in the same manner as most sensible\n> systems. You can say \"user postgres@foo is trusted to vet users so you can\n> safely hand out tickets for any bar@foo that postgres@foo says is legit\".\n\nSo, just to be clear, the way this *actually* works is a bit different\nfrom the way being described above, last time I looked into Kerberos\ndelegations anyway.\n\nEssentially, the KDC can be set up to allow 'bar@foo' to request a\nticket to delegate to 'postgres@foo', which then allows 'postgres@foo'\nto connect as if they are 'bar@foo' to some other service (and in some\nimplementations, I believe it's further possible to say that the ticket\nfor 'bar@foo' which is delegated to 'postgres@foo' is only allowed to\nrequest tickets for certain specific services, such as 'postgres2@foo'\nor what-have-you).\n\nNote that setting this up with an MIT KDC requires configuring it with\nan LDAP backend as the traditional KDC database doesn't support this\nkind of complex delegation control (again, last time I checked anyway).\n\n> I would strongly discourage allowing all users on host A to authenticate as\n> user postgres on host B. But with appropriate user-mappings support, we\n> could likely support that sort of model for both SSPI and Kerberos.\n\nIdeally, both sides would get a 'vote' regarding what's allowed, I would\nthink. That is, the connecting side would have to have a user mapping\nthat says \"this authenticated user is allowed to connect to this remote\nserver as this user\", and the remote server would have something like\n\"this server that's connecting, validated by the certificate presented\nby the server, is allowed to authenticate as this user\". I feel like\nwe're mostly there by allowing the connecting server to use a\ncertificate to connect to the remote server, while it's also checking\nthe user mapping, and the remote server's pg_hba.conf being configured\nto allow cert-based auth with a CN mapping from the CN of the connecting\nserver's certificate to authenticate to whatever users the remote server\nwants to allow. Is that more-or-less the idea here..?\n\n> A necessary prerequisite is that Pg be able to cope with passwordless\n> user-mappings though. Hence this patch.\n\nSure, that part seems like it makes sense to me (and perhaps has now\nbeen done, just catching up on things after travel and holidays and such\nhere in the US).\n\nThanks!\n\nStephen","msg_date":"Tue, 3 Dec 2019 09:36:23 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Allow superuser to grant passwordless connection rights on\n postgres_fdw"},{"msg_contents":"On Tue, Dec 3, 2019 at 9:36 AM Stephen Frost <sfrost@snowman.net> wrote:\n\n\n>\n> > A necessary prerequisite is that Pg be able to cope with passwordless\n> > user-mappings though. Hence this patch.\n>\n> Sure, that part seems like it makes sense to me (and perhaps has now\n> been done, just catching up on things after travel and holidays and such\n> here in the US).\n>\n\n\nIt hasn't been done, but I now propose to commit it shortly so other\nwork can proceed.\n\ncheers\n\nandrew\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Wed, 4 Dec 2019 11:28:33 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Allow superuser to grant passwordless connection rights on\n postgres_fdw"}],"string":"[\n {\n \"msg_contents\": \"This patch allows the superuser to grant passwordless connection rights\\nin postgres_fdw user mappings.\\n\\n\\nThe patch is authored by my colleague Craig Ringer, with slight bitrot\\nfixed by me.\\n\\n\\nOne use case for this is with passphrase-protected client certificates,\\na patch for which will follow shortly.\\n\\n\\nHere are Craig's remarks on the patch:\\n\\n \\n postgres_fdw denies a non-superuser the ability to establish a\\nconnection that\\n doesn't have a password in the connection string, or one that fails\\nto actually\\n use the password in authentication. This is to stop the unprivileged\\nuser using\\n OS-level authentication as the postgres server (peer, ident, trust).\\nIt also\\n stops unauthorized use of local credentials like .pgpass, a service\\nfile,\\n client certificate files, etc.\\n \\n Add the ability for a superuser to create user mappings that\\noverride this\\n behaviour by setting the passwordless_ok attribute to true in a user\\nmapping\\n for a non-superuser. The non-superuser gains the ability to use the\\nFDW the\\n mapping applies to even if there's no password in their mapping or\\nin the\\n connection string.\\n \\n This is only safe if the superuser has established that the local\\nserver is\\n configured safely. It must be configured not to allow\\n trust/peer/ident/sspi/gssapi auth to allow the OS user the postgres\\nserver runs\\n as to log in to postgres as a superuser. Client certificate keys can\\nbe used\\n too, if accessible. But the superuser can already GRANT superrole TO\\n normalrole, so it's not any sort of new power.\\n \\n\\ncheers\\n\\n\\nandrew\\n\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 31 Oct 2019 16:58:20 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Allow superuser to grant passwordless connection rights on\\n postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\\n<andrew.dunstan@2ndquadrant.com> wrote:\\n> This patch allows the superuser to grant passwordless connection rights\\n> in postgres_fdw user mappings.\\n\\nThis is clearly something that we need, as the current code seems\\nwoefully ignorant of the fact that passwords are not the only\\nauthentication method supported by PostgreSQL, nor even the most\\nsecure.\\n\\nBut, I do wonder a bit if we ought to think harder about the overall\\nauthentication model for FDW. Like, maybe we'd take a different view\\nof how to solve this particular piece of the problem if we were\\nthinking about how FDWs could do LDAP authentication, SSL\\nauthentication, credentials forwarding...\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 12:58:51 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow superuser to grant passwordless connection rights on\\n postgres_fdw\"\n },\n {\n \"msg_contents\": \"\\nOn 11/1/19 12:58 PM, Robert Haas wrote:\\n> On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\\n> <andrew.dunstan@2ndquadrant.com> wrote:\\n>> This patch allows the superuser to grant passwordless connection rights\\n>> in postgres_fdw user mappings.\\n> This is clearly something that we need, as the current code seems\\n> woefully ignorant of the fact that passwords are not the only\\n> authentication method supported by PostgreSQL, nor even the most\\n> secure.\\n>\\n> But, I do wonder a bit if we ought to think harder about the overall\\n> authentication model for FDW. Like, maybe we'd take a different view\\n> of how to solve this particular piece of the problem if we were\\n> thinking about how FDWs could do LDAP authentication, SSL\\n> authentication, credentials forwarding...\\n>\\n\\n\\nI'm certainly open to alternatives.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 14:00:27 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow superuser to grant passwordless connection rights on\\n postgres_fdw\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\n* Andrew Dunstan (andrew.dunstan@2ndquadrant.com) wrote:\\n> On 11/1/19 12:58 PM, Robert Haas wrote:\\n> > On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\\n> > <andrew.dunstan@2ndquadrant.com> wrote:\\n> >> This patch allows the superuser to grant passwordless connection rights\\n> >> in postgres_fdw user mappings.\\n> > This is clearly something that we need, as the current code seems\\n> > woefully ignorant of the fact that passwords are not the only\\n> > authentication method supported by PostgreSQL, nor even the most\\n> > secure.\\n> >\\n> > But, I do wonder a bit if we ought to think harder about the overall\\n> > authentication model for FDW. Like, maybe we'd take a different view\\n> > of how to solve this particular piece of the problem if we were\\n> > thinking about how FDWs could do LDAP authentication, SSL\\n> > authentication, credentials forwarding...\\n> \\n> I'm certainly open to alternatives.\\n\\nI've long felt that the way to handle this kind of requirement is to\\nhave a \\\"trusted remote server\\\" kind of option- where the local server\\nauthenticates to the remote server as a *server* and then says \\\"this is\\nthe user on this server, and this is the user that this user wishes to\\nbe\\\" and the remote server is then able to decide if they accept that, or\\nnot.\\n\\nTo be specific, there would be some kind of 'trust' established between\\nthe servers and only if there is some kind of server-level\\nauthentication, eg: dual TLS auth, or dual GSSAPI auth; and then, a\\nmapping is defined for that server, which specifies what remote user is\\nallowed to log in as what local user.\\n\\nThis would be a server-to-server auth arrangement, and is quite\\ndifferent from credential forwarding, or similar. I am certainly also a\\nhuge fan of the idea that we support Kerberos/GSSAPI credential\\nforwarding / delegation, where a client willingly forwards to the PG\\nserver a set of credentials which then allow the PG server to\\nauthenticate as that user to another system (eg: through an FDW to\\nanother PG server).\\n\\nOf course, as long as we're talking pie-in-the-sky ideas, I would\\ncertainly be entirely for supporting both. ;)\\n\\nThanks,\\n\\nStephen\",\n \"msg_date\": \"Sun, 3 Nov 2019 23:20:05 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow superuser to grant passwordless connection rights on\\n postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Mon, 4 Nov 2019 at 12:20, Stephen Frost <sfrost@snowman.net> wrote:\\n\\n> Greetings,\\n>\\n> * Andrew Dunstan (andrew.dunstan@2ndquadrant.com) wrote:\\n> > On 11/1/19 12:58 PM, Robert Haas wrote:\\n> > > On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\\n> > > <andrew.dunstan@2ndquadrant.com> wrote:\\n> > >> This patch allows the superuser to grant passwordless connection\\n> rights\\n> > >> in postgres_fdw user mappings.\\n> > > This is clearly something that we need, as the current code seems\\n> > > woefully ignorant of the fact that passwords are not the only\\n> > > authentication method supported by PostgreSQL, nor even the most\\n> > > secure.\\n> > >\\n> > > But, I do wonder a bit if we ought to think harder about the overall\\n> > > authentication model for FDW. Like, maybe we'd take a different view\\n> > > of how to solve this particular piece of the problem if we were\\n> > > thinking about how FDWs could do LDAP authentication, SSL\\n> > > authentication, credentials forwarding...\\n> >\\n> > I'm certainly open to alternatives.\\n>\\n> I've long felt that the way to handle this kind of requirement is to\\n> have a \\\"trusted remote server\\\" kind of option- where the local server\\n> authenticates to the remote server as a *server* and then says \\\"this is\\n> the user on this server, and this is the user that this user wishes to\\n> be\\\" and the remote server is then able to decide if they accept that, or\\n> not.\\n>\\n\\nThe original use case for the patch was to allow FDWs to use SSL/TLS client\\ncertificates. Each user-mapping has its own certificate - there's a\\nseparate patch to allow that. So there's no delegation of trust via\\nKerberos etc in that particular case.\\n\\nI can see value in using Kerberos etc for that too though, as it separates\\nauthorization and authentication in the same manner as most sensible\\nsystems. You can say \\\"user postgres@foo is trusted to vet users so you can\\nsafely hand out tickets for any bar@foo that postgres@foo says is legit\\\".\\n\\nI would strongly discourage allowing all users on host A to authenticate as\\nuser postgres on host B. But with appropriate user-mappings support, we\\ncould likely support that sort of model for both SSPI and Kerberos.\\n\\nA necessary prerequisite is that Pg be able to cope with passwordless\\nuser-mappings though. Hence this patch.\\n\\n\\n\\n>\\n> To be specific, there would be some kind of 'trust' established between\\n> the servers and only if there is some kind of server-level\\n> authentication, eg: dual TLS auth, or dual GSSAPI auth; and then, a\\n> mapping is defined for that server, which specifies what remote user is\\n> allowed to log in as what local user.\\n>\\n> This would be a server-to-server auth arrangement, and is quite\\n> different from credential forwarding, or similar. I am certainly also a\\n> huge fan of the idea that we support Kerberos/GSSAPI credential\\n> forwarding / delegation, where a client willingly forwards to the PG\\n> server a set of credentials which then allow the PG server to\\n> authenticate as that user to another system (eg: through an FDW to\\n> another PG server).\\n>\\n> Of course, as long as we're talking pie-in-the-sky ideas, I would\\n> certainly be entirely for supporting both. ;)\\n>\\n> Thanks,\\n>\\n> Stephen\\n>\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\\n\\nOn Mon, 4 Nov 2019 at 12:20, Stephen Frost <sfrost@snowman.net> wrote:Greetings,\\n\\n* Andrew Dunstan (andrew.dunstan@2ndquadrant.com) wrote:\\n> On 11/1/19 12:58 PM, Robert Haas wrote:\\n> > On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\\n> > <andrew.dunstan@2ndquadrant.com> wrote:\\n> >> This patch allows the superuser to grant passwordless connection rights\\n> >> in postgres_fdw user mappings.\\n> > This is clearly something that we need, as the current code seems\\n> > woefully ignorant of the fact that passwords are not the only\\n> > authentication method supported by PostgreSQL, nor even the most\\n> > secure.\\n> >\\n> > But, I do wonder a bit if we ought to think harder about the overall\\n> > authentication model for FDW. Like, maybe we'd take a different view\\n> > of how to solve this particular piece of the problem if we were\\n> > thinking about how FDWs could do LDAP authentication, SSL\\n> > authentication, credentials forwarding...\\n> \\n> I'm certainly open to alternatives.\\n\\nI've long felt that the way to handle this kind of requirement is to\\nhave a \\\"trusted remote server\\\" kind of option- where the local server\\nauthenticates to the remote server as a *server* and then says \\\"this is\\nthe user on this server, and this is the user that this user wishes to\\nbe\\\" and the remote server is then able to decide if they accept that, or\\nnot.The original use case for the patch was to allow FDWs to use SSL/TLS client certificates. Each user-mapping has its own certificate - there's a separate patch to allow that. So there's no delegation of trust via Kerberos etc in that particular case.I can see value in using Kerberos etc for that too though, as it separates authorization and authentication in the same manner as most sensible systems. You can say \\\"user postgres@foo is trusted to vet users so you can safely hand out tickets for any bar@foo that postgres@foo says is legit\\\".I would strongly discourage allowing all users on host A to authenticate as user postgres on host B. But with appropriate user-mappings support, we could likely support that sort of model for both SSPI and Kerberos.A necessary prerequisite is that Pg be able to cope with passwordless user-mappings though. Hence this patch. \\n\\nTo be specific, there would be some kind of 'trust' established between\\nthe servers and only if there is some kind of server-level\\nauthentication, eg: dual TLS auth, or dual GSSAPI auth; and then, a\\nmapping is defined for that server, which specifies what remote user is\\nallowed to log in as what local user.\\n\\nThis would be a server-to-server auth arrangement, and is quite\\ndifferent from credential forwarding, or similar. I am certainly also a\\nhuge fan of the idea that we support Kerberos/GSSAPI credential\\nforwarding / delegation, where a client willingly forwards to the PG\\nserver a set of credentials which then allow the PG server to\\nauthenticate as that user to another system (eg: through an FDW to\\nanother PG server).\\n\\nOf course, as long as we're talking pie-in-the-sky ideas, I would\\ncertainly be entirely for supporting both. ;)\\n\\nThanks,\\n\\nStephen\\n-- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise\",\n \"msg_date\": \"Sun, 10 Nov 2019 17:35:36 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow superuser to grant passwordless connection rights on\\n postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 10, 2019 at 4:35 AM Craig Ringer <craig@2ndquadrant.com> wrote:\\n>\\n> On Mon, 4 Nov 2019 at 12:20, Stephen Frost <sfrost@snowman.net> wrote:\\n>>\\n>> Greetings,\\n>>\\n>> * Andrew Dunstan (andrew.dunstan@2ndquadrant.com) wrote:\\n>> > On 11/1/19 12:58 PM, Robert Haas wrote:\\n>> > > On Thu, Oct 31, 2019 at 4:58 PM Andrew Dunstan\\n>> > > <andrew.dunstan@2ndquadrant.com> wrote:\\n>> > >> This patch allows the superuser to grant passwordless connection rights\\n>> > >> in postgres_fdw user mappings.\\n>> > > This is clearly something that we need, as the current code seems\\n>> > > woefully ignorant of the fact that passwords are not the only\\n>> > > authentication method supported by PostgreSQL, nor even the most\\n>> > > secure.\\n>> > >\\n>> > > But, I do wonder a bit if we ought to think harder about the overall\\n>> > > authentication model for FDW. Like, maybe we'd take a different view\\n>> > > of how to solve this particular piece of the problem if we were\\n>> > > thinking about how FDWs could do LDAP authentication, SSL\\n>> > > authentication, credentials forwarding...\\n>> >\\n>> > I'm certainly open to alternatives.\\n>>\\n>> I've long felt that the way to handle this kind of requirement is to\\n>> have a \\\"trusted remote server\\\" kind of option- where the local server\\n>> authenticates to the remote server as a *server* and then says \\\"this is\\n>> the user on this server, and this is the user that this user wishes to\\n>> be\\\" and the remote server is then able to decide if they accept that, or\\n>> not.\\n>\\n>\\n> The original use case for the patch was to allow FDWs to use SSL/TLS client certificates. Each user-mapping has its own certificate - there's a separate patch to allow that. So there's no delegation of trust via Kerberos etc in that particular case.\\n>\\n> I can see value in using Kerberos etc for that too though, as it separates authorization and authentication in the same manner as most sensible systems. You can say \\\"user postgres@foo is trusted to vet users so you can safely hand out tickets for any bar@foo that postgres@foo says is legit\\\".\\n>\\n> I would strongly discourage allowing all users on host A to authenticate as user postgres on host B. But with appropriate user-mappings support, we could likely support that sort of model for both SSPI and Kerberos.\\n>\\n> A necessary prerequisite is that Pg be able to cope with passwordless user-mappings though. Hence this patch.\\n>\\n>\\n\\n\\nYeah, I agree. Does anyone else want to weigh in on this? If nobody\\nobjects I'd like to tidy this up and get it committed so we can add\\nsupport for client certs in postgres_fdw, which is the real business\\nat hand, and which I know from various offline comments a number of\\npeople are keen to have.\\n\\ncheers\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Nov 2019 16:56:24 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow superuser to grant passwordless connection rights on\\n postgres_fdw\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\n* Craig Ringer (craig@2ndquadrant.com) wrote:\\n> On Mon, 4 Nov 2019 at 12:20, Stephen Frost <sfrost@snowman.net> wrote:\\n> > I've long felt that the way to handle this kind of requirement is to\\n> > have a \\\"trusted remote server\\\" kind of option- where the local server\\n> > authenticates to the remote server as a *server* and then says \\\"this is\\n> > the user on this server, and this is the user that this user wishes to\\n> > be\\\" and the remote server is then able to decide if they accept that, or\\n> > not.\\n> \\n> The original use case for the patch was to allow FDWs to use SSL/TLS client\\n> certificates. Each user-mapping has its own certificate - there's a\\n> separate patch to allow that. So there's no delegation of trust via\\n> Kerberos etc in that particular case.\\n> \\n> I can see value in using Kerberos etc for that too though, as it separates\\n> authorization and authentication in the same manner as most sensible\\n> systems. You can say \\\"user postgres@foo is trusted to vet users so you can\\n> safely hand out tickets for any bar@foo that postgres@foo says is legit\\\".\\n\\nSo, just to be clear, the way this *actually* works is a bit different\\nfrom the way being described above, last time I looked into Kerberos\\ndelegations anyway.\\n\\nEssentially, the KDC can be set up to allow 'bar@foo' to request a\\nticket to delegate to 'postgres@foo', which then allows 'postgres@foo'\\nto connect as if they are 'bar@foo' to some other service (and in some\\nimplementations, I believe it's further possible to say that the ticket\\nfor 'bar@foo' which is delegated to 'postgres@foo' is only allowed to\\nrequest tickets for certain specific services, such as 'postgres2@foo'\\nor what-have-you).\\n\\nNote that setting this up with an MIT KDC requires configuring it with\\nan LDAP backend as the traditional KDC database doesn't support this\\nkind of complex delegation control (again, last time I checked anyway).\\n\\n> I would strongly discourage allowing all users on host A to authenticate as\\n> user postgres on host B. But with appropriate user-mappings support, we\\n> could likely support that sort of model for both SSPI and Kerberos.\\n\\nIdeally, both sides would get a 'vote' regarding what's allowed, I would\\nthink. That is, the connecting side would have to have a user mapping\\nthat says \\\"this authenticated user is allowed to connect to this remote\\nserver as this user\\\", and the remote server would have something like\\n\\\"this server that's connecting, validated by the certificate presented\\nby the server, is allowed to authenticate as this user\\\". I feel like\\nwe're mostly there by allowing the connecting server to use a\\ncertificate to connect to the remote server, while it's also checking\\nthe user mapping, and the remote server's pg_hba.conf being configured\\nto allow cert-based auth with a CN mapping from the CN of the connecting\\nserver's certificate to authenticate to whatever users the remote server\\nwants to allow. Is that more-or-less the idea here..?\\n\\n> A necessary prerequisite is that Pg be able to cope with passwordless\\n> user-mappings though. Hence this patch.\\n\\nSure, that part seems like it makes sense to me (and perhaps has now\\nbeen done, just catching up on things after travel and holidays and such\\nhere in the US).\\n\\nThanks!\\n\\nStephen\",\n \"msg_date\": \"Tue, 3 Dec 2019 09:36:23 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow superuser to grant passwordless connection rights on\\n postgres_fdw\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 3, 2019 at 9:36 AM Stephen Frost <sfrost@snowman.net> wrote:\\n\\n\\n>\\n> > A necessary prerequisite is that Pg be able to cope with passwordless\\n> > user-mappings though. Hence this patch.\\n>\\n> Sure, that part seems like it makes sense to me (and perhaps has now\\n> been done, just catching up on things after travel and holidays and such\\n> here in the US).\\n>\\n\\n\\nIt hasn't been done, but I now propose to commit it shortly so other\\nwork can proceed.\\n\\ncheers\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Wed, 4 Dec 2019 11:28:33 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow superuser to grant passwordless connection rights on\\n postgres_fdw\"\n }\n]","isConversational":false}}},{"rowIdx":49384,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"\nThis patch provides for an sslpassword parameter for libpq, and a hook\nthat a client can fill in for a callback function to set the password.\n\n\nThis provides similar facilities to those already available in the JDBC\ndriver.\n\n\nThere is also a function to fetch the sslpassword from the connection\nparameters, in the same way that other settings can be fetched.\n\n\nThis is mostly the excellent work of my colleague Craig Ringer, with a\nfew embellishments from me.\n\n\nHere are his notes:\n\n\n Allow libpq to non-interactively decrypt client certificates that\nare stored\n encrypted by adding a new \"sslpassword\" connection option.\n \n The sslpassword option offers a middle ground between a cleartext\nkey and\n setting up advanced key mangement via openssl engines, PKCS#11, USB\ncrypto\n offload and key escrow, etc.\n \n Previously use of encrypted client certificate keys only worked if\nthe user\n could enter the key's password interactively on stdin, in response\nto openssl's\n default prompt callback:\n \n Enter PEM passhprase:\n \n That's infesible in many situations, especially things like use from\n postgres_fdw.\n \n This change also allows admins to prevent libpq from ever prompting\nfor a\n password by calling:\n \n PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\n \n which is useful since OpenSSL likes to open /dev/tty to prompt for a\npassword,\n so even closing stdin won't stop it blocking if there's no user\ninput available.\n Applications may also override or extend SSL password fetching with\ntheir own\n callback.\n \n There is deliberately no environment variable equivalent for the\nsslpassword\n option.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Thu, 31 Oct 2019 18:33:12 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"libpq sslpassword parameter and callback function"},{"msg_contents":"This time with attachment.\n\n\nOn 10/31/19 6:33 PM, Andrew Dunstan wrote:\n> This patch provides for an sslpassword parameter for libpq, and a hook\n> that a client can fill in for a callback function to set the password.\n>\n>\n> This provides similar facilities to those already available in the JDBC\n> driver.\n>\n>\n> There is also a function to fetch the sslpassword from the connection\n> parameters, in the same way that other settings can be fetched.\n>\n>\n> This is mostly the excellent work of my colleague Craig Ringer, with a\n> few embellishments from me.\n>\n>\n> Here are his notes:\n>\n>\n> Allow libpq to non-interactively decrypt client certificates that\n> are stored\n> encrypted by adding a new \"sslpassword\" connection option.\n> \n> The sslpassword option offers a middle ground between a cleartext\n> key and\n> setting up advanced key mangement via openssl engines, PKCS#11, USB\n> crypto\n> offload and key escrow, etc.\n> \n> Previously use of encrypted client certificate keys only worked if\n> the user\n> could enter the key's password interactively on stdin, in response\n> to openssl's\n> default prompt callback:\n> \n> Enter PEM passhprase:\n> \n> That's infesible in many situations, especially things like use from\n> postgres_fdw.\n> \n> This change also allows admins to prevent libpq from ever prompting\n> for a\n> password by calling:\n> \n> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\n> \n> which is useful since OpenSSL likes to open /dev/tty to prompt for a\n> password,\n> so even closing stdin won't stop it blocking if there's no user\n> input available.\n> Applications may also override or extend SSL password fetching with\n> their own\n> callback.\n> \n> There is deliberately no environment variable equivalent for the\n> sslpassword\n> option.\n>\n>\n> cheers\n>\n>\n> andrew\n>\n>\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 31 Oct 2019 18:34:32 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: libpq sslpassword parameter and callback function"},{"msg_contents":"\nOn 10/31/19 6:34 PM, Andrew Dunstan wrote:\n> This time with attachment.\n>\n>\n> On 10/31/19 6:33 PM, Andrew Dunstan wrote:\n>> This patch provides for an sslpassword parameter for libpq, and a hook\n>> that a client can fill in for a callback function to set the password.\n>>\n>>\n>> This provides similar facilities to those already available in the JDBC\n>> driver.\n>>\n>>\n>> There is also a function to fetch the sslpassword from the connection\n>> parameters, in the same way that other settings can be fetched.\n>>\n>>\n>> This is mostly the excellent work of my colleague Craig Ringer, with a\n>> few embellishments from me.\n>>\n>>\n>> Here are his notes:\n>>\n>>\n>> Allow libpq to non-interactively decrypt client certificates that\n>> are stored\n>> encrypted by adding a new \"sslpassword\" connection option.\n>> \n>> The sslpassword option offers a middle ground between a cleartext\n>> key and\n>> setting up advanced key mangement via openssl engines, PKCS#11, USB\n>> crypto\n>> offload and key escrow, etc.\n>> \n>> Previously use of encrypted client certificate keys only worked if\n>> the user\n>> could enter the key's password interactively on stdin, in response\n>> to openssl's\n>> default prompt callback:\n>> \n>> Enter PEM passhprase:\n>> \n>> That's infesible in many situations, especially things like use from\n>> postgres_fdw.\n>> \n>> This change also allows admins to prevent libpq from ever prompting\n>> for a\n>> password by calling:\n>> \n>> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\n>> \n>> which is useful since OpenSSL likes to open /dev/tty to prompt for a\n>> password,\n>> so even closing stdin won't stop it blocking if there's no user\n>> input available.\n>> Applications may also override or extend SSL password fetching with\n>> their own\n>> callback.\n>> \n>> There is deliberately no environment variable equivalent for the\n>> sslpassword\n>> option.\n>>\n>>\n\nI should also mention that this patch provides for support for DER\nformat certificates and keys.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Thu, 31 Oct 2019 19:27:52 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: libpq sslpassword parameter and callback function"},{"msg_contents":"On Fri, 1 Nov 2019 at 07:27, Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\nwrote:\n\n>\n> On 10/31/19 6:34 PM, Andrew Dunstan wrote:\n> > This time with attachment.\n> >\n> >\n> > On 10/31/19 6:33 PM, Andrew Dunstan wrote:\n> >> This patch provides for an sslpassword parameter for libpq, and a hook\n> >> that a client can fill in for a callback function to set the password.\n> >>\n> >>\n> >> This provides similar facilities to those already available in the JDBC\n> >> driver.\n> >>\n> >>\n> >> There is also a function to fetch the sslpassword from the connection\n> >> parameters, in the same way that other settings can be fetched.\n> >>\n> >>\n> >> This is mostly the excellent work of my colleague Craig Ringer, with a\n> >> few embellishments from me.\n> >>\n> >>\n> >> Here are his notes:\n> >>\n> >>\n> >> Allow libpq to non-interactively decrypt client certificates that\n> >> are stored\n> >> encrypted by adding a new \"sslpassword\" connection option.\n> >>\n> >> The sslpassword option offers a middle ground between a cleartext\n> >> key and\n> >> setting up advanced key mangement via openssl engines, PKCS#11, USB\n> >> crypto\n> >> offload and key escrow, etc.\n> >>\n> >> Previously use of encrypted client certificate keys only worked if\n> >> the user\n> >> could enter the key's password interactively on stdin, in response\n> >> to openssl's\n> >> default prompt callback:\n> >>\n> >> Enter PEM passhprase:\n> >>\n> >> That's infesible in many situations, especially things like use from\n> >> postgres_fdw.\n> >>\n> >> This change also allows admins to prevent libpq from ever prompting\n> >> for a\n> >> password by calling:\n> >>\n> >> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\n> >>\n> >> which is useful since OpenSSL likes to open /dev/tty to prompt for a\n> >> password,\n> >> so even closing stdin won't stop it blocking if there's no user\n> >> input available.\n> >> Applications may also override or extend SSL password fetching with\n> >> their own\n> >> callback.\n> >>\n> >> There is deliberately no environment variable equivalent for the\n> >> sslpassword\n> >> option.\n> >>\n> >>\n>\n> I should also mention that this patch provides for support for DER\n> format certificates and keys.\n>\n>\nYep, that was a trivial change I rolled into it.\n\nFWIW, this is related to two other patches: the patch to allow passwordless\nfdw connections with explicit superuser approval, and the patch to allow\nsslkey/sslpassword to be set as user mapping options in postgres_fdw .\nTogether all three patches make it possible to use SSL client certificates\nto manage authentication in postgres_fdw user mappings.\n\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\n\nOn Fri, 1 Nov 2019 at 07:27, Andrew Dunstan <andrew.dunstan@2ndquadrant.com> wrote:\nOn 10/31/19 6:34 PM, Andrew Dunstan wrote:\n> This time with attachment.\n>\n>\n> On 10/31/19 6:33 PM, Andrew Dunstan wrote:\n>> This patch provides for an sslpassword parameter for libpq, and a hook\n>> that a client can fill in for a callback function to set the password.\n>>\n>>\n>> This provides similar facilities to those already available in the JDBC\n>> driver.\n>>\n>>\n>> There is also a function to fetch the sslpassword from the connection\n>> parameters, in the same way that other settings can be fetched.\n>>\n>>\n>> This is mostly the excellent work of my colleague Craig Ringer, with a\n>> few embellishments from me.\n>>\n>>\n>> Here are his notes:\n>>\n>>\n>> Allow libpq to non-interactively decrypt client certificates that\n>> are stored\n>> encrypted by adding a new \"sslpassword\" connection option.\n>> \n>> The sslpassword option offers a middle ground between a cleartext\n>> key and\n>> setting up advanced key mangement via openssl engines, PKCS#11, USB\n>> crypto\n>> offload and key escrow, etc.\n>> \n>> Previously use of encrypted client certificate keys only worked if\n>> the user\n>> could enter the key's password interactively on stdin, in response\n>> to openssl's\n>> default prompt callback:\n>> \n>> Enter PEM passhprase:\n>> \n>> That's infesible in many situations, especially things like use from\n>> postgres_fdw.\n>> \n>> This change also allows admins to prevent libpq from ever prompting\n>> for a\n>> password by calling:\n>> \n>> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\n>> \n>> which is useful since OpenSSL likes to open /dev/tty to prompt for a\n>> password,\n>> so even closing stdin won't stop it blocking if there's no user\n>> input available.\n>> Applications may also override or extend SSL password fetching with\n>> their own\n>> callback.\n>> \n>> There is deliberately no environment variable equivalent for the\n>> sslpassword\n>> option.\n>>\n>>\n\nI should also mention that this patch provides for support for DER\nformat certificates and keys.Yep, that was a trivial change I rolled into it.FWIW, this is related to two other patches: the patch to allow passwordless fdw connections with explicit superuser approval, and the patch to allow sslkey/sslpassword to be set as user mapping options in postgres_fdw . Together all three patches make it possible to use SSL client certificates to manage authentication in postgres_fdw user mappings.-- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise","msg_date":"Sun, 10 Nov 2019 17:47:24 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: libpq sslpassword parameter and callback function"},{"msg_contents":"On Sun, Nov 10, 2019 at 05:47:24PM +0800, Craig Ringer wrote:\n> Yep, that was a trivial change I rolled into it.\n> \n> FWIW, this is related to two other patches: the patch to allow passwordless fdw\n> connections with explicit superuser approval, and the patch to allow sslkey/\n> sslpassword to be set as user mapping options in postgres_fdw . Together all\n> three patches make it possible to use SSL client certificates to manage\n> authentication in postgres_fdw user mappings.\n\nOh, nice, greatly needed!\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Tue, 12 Nov 2019 22:10:38 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: libpq sslpassword parameter and callback function"},{"msg_contents":"On 10/31/19 7:27 PM, Andrew Dunstan wrote:\n> On 10/31/19 6:34 PM, Andrew Dunstan wrote:\n>> This time with attachment.\n>>\n>>\n>> On 10/31/19 6:33 PM, Andrew Dunstan wrote:\n>>> This patch provides for an sslpassword parameter for libpq, and a hook\n>>> that a client can fill in for a callback function to set the password.\n>>>\n>>>\n>>> This provides similar facilities to those already available in the JDBC\n>>> driver.\n>>>\n>>>\n>>> There is also a function to fetch the sslpassword from the connection\n>>> parameters, in the same way that other settings can be fetched.\n>>>\n>>>\n>>> This is mostly the excellent work of my colleague Craig Ringer, with a\n>>> few embellishments from me.\n>>>\n>>>\n>>> Here are his notes:\n>>>\n>>>\n>>> Allow libpq to non-interactively decrypt client certificates that\n>>> are stored\n>>> encrypted by adding a new \"sslpassword\" connection option.\n>>> \n>>> The sslpassword option offers a middle ground between a cleartext\n>>> key and\n>>> setting up advanced key mangement via openssl engines, PKCS#11, USB\n>>> crypto\n>>> offload and key escrow, etc.\n>>> \n>>> Previously use of encrypted client certificate keys only worked if\n>>> the user\n>>> could enter the key's password interactively on stdin, in response\n>>> to openssl's\n>>> default prompt callback:\n>>> \n>>> Enter PEM passhprase:\n>>> \n>>> That's infesible in many situations, especially things like use from\n>>> postgres_fdw.\n>>> \n>>> This change also allows admins to prevent libpq from ever prompting\n>>> for a\n>>> password by calling:\n>>> \n>>> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\n>>> \n>>> which is useful since OpenSSL likes to open /dev/tty to prompt for a\n>>> password,\n>>> so even closing stdin won't stop it blocking if there's no user\n>>> input available.\n>>> Applications may also override or extend SSL password fetching with\n>>> their own\n>>> callback.\n>>> \n>>> There is deliberately no environment variable equivalent for the\n>>> sslpassword\n>>> option.\n>>>\n>>>\n> I should also mention that this patch provides for support for DER\n> format certificates and keys.\n>\n>\n\n\nHere's an updated version of the patch, adjusted to the now committed\nchanges to TestLib.pm.\n\n\ncheers\n\n\nandrew\n\n\n-- \n\nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Mon, 25 Nov 2019 16:09:08 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: libpq sslpassword parameter and callback function"},{"msg_contents":"On 11/25/19 4:09 PM, Andrew Dunstan wrote:\n> On 10/31/19 7:27 PM, Andrew Dunstan wrote:\n>> On 10/31/19 6:34 PM, Andrew Dunstan wrote:\n>>> This time with attachment.\n>>>\n>>>\n>>> On 10/31/19 6:33 PM, Andrew Dunstan wrote:\n>>>> This patch provides for an sslpassword parameter for libpq, and a hook\n>>>> that a client can fill in for a callback function to set the password.\n>>>>\n>>>>\n>>>> This provides similar facilities to those already available in the JDBC\n>>>> driver.\n>>>>\n>>>>\n>>>> There is also a function to fetch the sslpassword from the connection\n>>>> parameters, in the same way that other settings can be fetched.\n>>>>\n>>>>\n>>>> This is mostly the excellent work of my colleague Craig Ringer, with a\n>>>> few embellishments from me.\n>>>>\n>>>>\n>>>> Here are his notes:\n>>>>\n>>>>\n>>>> Allow libpq to non-interactively decrypt client certificates that\n>>>> are stored\n>>>> encrypted by adding a new \"sslpassword\" connection option.\n>>>> \n>>>> The sslpassword option offers a middle ground between a cleartext\n>>>> key and\n>>>> setting up advanced key mangement via openssl engines, PKCS#11, USB\n>>>> crypto\n>>>> offload and key escrow, etc.\n>>>> \n>>>> Previously use of encrypted client certificate keys only worked if\n>>>> the user\n>>>> could enter the key's password interactively on stdin, in response\n>>>> to openssl's\n>>>> default prompt callback:\n>>>> \n>>>> Enter PEM passhprase:\n>>>> \n>>>> That's infesible in many situations, especially things like use from\n>>>> postgres_fdw.\n>>>> \n>>>> This change also allows admins to prevent libpq from ever prompting\n>>>> for a\n>>>> password by calling:\n>>>> \n>>>> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\n>>>> \n>>>> which is useful since OpenSSL likes to open /dev/tty to prompt for a\n>>>> password,\n>>>> so even closing stdin won't stop it blocking if there's no user\n>>>> input available.\n>>>> Applications may also override or extend SSL password fetching with\n>>>> their own\n>>>> callback.\n>>>> \n>>>> There is deliberately no environment variable equivalent for the\n>>>> sslpassword\n>>>> option.\n>>>>\n>>>>\n>> I should also mention that this patch provides for support for DER\n>> format certificates and keys.\n>>\n>>\n>\n> Here's an updated version of the patch, adjusted to the now committed\n> changes to TestLib.pm.\n>\n>\n\n\nHere's an update now we have backed out the TestLib changes. The tests\nthat need a pty are skipped.\n\n\ncheers\n\n\nandrew\n\n\n-- \n\nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Wed, 27 Nov 2019 19:06:10 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: libpq sslpassword parameter and callback function"},{"msg_contents":"The following review has been posted through the commitfest application:\nmake installcheck-world: tested, passed\nImplements feature: tested, passed\nSpec compliant: tested, passed\nDocumentation: tested, failed\n\nHi Andrew,\r\n\r\nI've reviewed your \"libpq sslpassword parameter and callback function\" patch (0001-libpq-sslpassword-der-support.patch), and only found a few minor things (otherwise it looked good to me):\r\n\r\n1) There's a few trailing white-space warnings on patch application (from where you modified to skip 2 of the tests):\r\ngit apply 0001-libpq-sslpassword-der-support.patch\r\n0001-libpq-sslpassword-der-support.patch:649: trailing whitespace.\r\n\t# so they don't hang. For now they are not performed. \r\n0001-libpq-sslpassword-der-support.patch:659: trailing whitespace.\r\n\t\r\nwarning: 2 lines add whitespace errors.\r\n\r\n\r\n2) src/interfaces/libpq/libpq-fe.h\r\nThe following portion of the comment should be removed.\r\n\r\n+ * 2ndQPostgres extension. If you need to be compatible with unpatched libpq\r\n+ * you must dlsym() these.\r\n\r\n3) Documentation for the \"PQsslpassword\" function should be added to the libpq \"33.2 Connection Status Functions\" section.\r\n\r\n\r\nI made the following notes about how/what I reviewed and tested:\r\n\r\n- Applied patch and built Postgres (--with-openssl --enable-tap-tests), checked build output\r\n- Checked patch code modifications (format, logic, memory usage, efficiency, corner cases etc.)\r\n- Built documentation and checked updated portions (format, grammar, details, completeness etc.)\r\n- Checked test updates \r\n- Ran updated contrib/dblink tests - confirmed all passed\r\n- Ran updated SSL (TAP) tests - confirmed all passed\r\n- Ran \"make installcheck-world\", as per review requirements\r\n- Wrote small libpq-based app to test:\r\n - new APIs (PQsslpassword, PQsetSSLKeyPassHook, PQgetSSLKeyPassHook, PQdefaultSSLKeyPassHook)\r\n - passphrase-protected key with/without patch\r\n - patch with/without new key password callack\r\n - patch and certificate with/without pass phrase protection on key\r\n - default callback, callback delegation\r\n - PEM/DER keys\r\n\r\n\r\nRegards,\r\nGreg","msg_date":"Fri, 29 Nov 2019 03:25:45 +0000","msg_from":"Greg Nancarrow <gregn4422@gmail.com>","msg_from_op":false,"msg_subject":"Re: libpq sslpassword parameter and callback function"},{"msg_contents":"On 11/28/19 10:25 PM, Greg Nancarrow wrote:\n> The following review has been posted through the commitfest application:\n> make installcheck-world: tested, passed\n> Implements feature: tested, passed\n> Spec compliant: tested, passed\n> Documentation: tested, failed\n>\n> Hi Andrew,\n>\n> I've reviewed your \"libpq sslpassword parameter and callback function\" patch (0001-libpq-sslpassword-der-support.patch), and only found a few minor things (otherwise it looked good to me):\n>\n> 1) There's a few trailing white-space warnings on patch application (from where you modified to skip 2 of the tests):\n> git apply 0001-libpq-sslpassword-der-support.patch\n> 0001-libpq-sslpassword-der-support.patch:649: trailing whitespace.\n> \t# so they don't hang. For now they are not performed. \n> 0001-libpq-sslpassword-der-support.patch:659: trailing whitespace.\n> \t\n> warning: 2 lines add whitespace errors.\n>\n>\n> 2) src/interfaces/libpq/libpq-fe.h\n> The following portion of the comment should be removed.\n>\n> + * 2ndQPostgres extension. If you need to be compatible with unpatched libpq\n> + * you must dlsym() these.\n>\n> 3) Documentation for the \"PQsslpassword\" function should be added to the libpq \"33.2 Connection Status Functions\" section.\n>\n>\n> I made the following notes about how/what I reviewed and tested:\n>\n> - Applied patch and built Postgres (--with-openssl --enable-tap-tests), checked build output\n> - Checked patch code modifications (format, logic, memory usage, efficiency, corner cases etc.)\n> - Built documentation and checked updated portions (format, grammar, details, completeness etc.)\n> - Checked test updates \n> - Ran updated contrib/dblink tests - confirmed all passed\n> - Ran updated SSL (TAP) tests - confirmed all passed\n> - Ran \"make installcheck-world\", as per review requirements\n> - Wrote small libpq-based app to test:\n> - new APIs (PQsslpassword, PQsetSSLKeyPassHook, PQgetSSLKeyPassHook, PQdefaultSSLKeyPassHook)\n> - passphrase-protected key with/without patch\n> - patch with/without new key password callack\n> - patch and certificate with/without pass phrase protection on key\n> - default callback, callback delegation\n> - PEM/DER keys\n>\n>\n\n\nThanks, nice thorough review.\n\n\nHere's an updated patch that I think fixes all the things you mentioned.\nI plan to commit this tomorrow.\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Fri, 29 Nov 2019 09:27:02 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: libpq sslpassword parameter and callback function"},{"msg_contents":"On Fri, Nov 29, 2019 at 09:27:02AM -0500, Andrew Dunstan wrote:\n> On 11/28/19 10:25 PM, Greg Nancarrow wrote:\n> > 3) Documentation for the \"PQsslpassword\" function should be added to the libpq \"33.2 Connection Status Functions\" section.\n\n> I plan to commit this tomorrow.\n\nThe PQsslpassword() function retrieves a connection parameter value, which one\ncan retrieve with PQconninfo(). Since introducing PQconninfo(), we have not\nadded any per-parameter accessor functions. Would you remove PQsslpassword()?\n\n\n","msg_date":"Fri, 6 Dec 2019 07:57:15 +0000","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":false,"msg_subject":"Re: libpq sslpassword parameter and callback function"},{"msg_contents":"\nOn 12/6/19 2:57 AM, Noah Misch wrote:\n> On Fri, Nov 29, 2019 at 09:27:02AM -0500, Andrew Dunstan wrote:\n>> On 11/28/19 10:25 PM, Greg Nancarrow wrote:\n>>> 3) Documentation for the \"PQsslpassword\" function should be added to the libpq \"33.2 Connection Status Functions\" section.\n>> I plan to commit this tomorrow.\n> The PQsslpassword() function retrieves a connection parameter value, which one\n> can retrieve with PQconninfo(). Since introducing PQconninfo(), we have not\n> added any per-parameter accessor functions. Would you remove PQsslpassword()?\n\n\n\n*sigh* ok\n\n\ncheers\n\n\nandrew\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Sat, 7 Dec 2019 08:51:20 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: libpq sslpassword parameter and callback function"}],"string":"[\n {\n \"msg_contents\": \"\\nThis patch provides for an sslpassword parameter for libpq, and a hook\\nthat a client can fill in for a callback function to set the password.\\n\\n\\nThis provides similar facilities to those already available in the JDBC\\ndriver.\\n\\n\\nThere is also a function to fetch the sslpassword from the connection\\nparameters, in the same way that other settings can be fetched.\\n\\n\\nThis is mostly the excellent work of my colleague Craig Ringer, with a\\nfew embellishments from me.\\n\\n\\nHere are his notes:\\n\\n\\n Allow libpq to non-interactively decrypt client certificates that\\nare stored\\n encrypted by adding a new \\\"sslpassword\\\" connection option.\\n \\n The sslpassword option offers a middle ground between a cleartext\\nkey and\\n setting up advanced key mangement via openssl engines, PKCS#11, USB\\ncrypto\\n offload and key escrow, etc.\\n \\n Previously use of encrypted client certificate keys only worked if\\nthe user\\n could enter the key's password interactively on stdin, in response\\nto openssl's\\n default prompt callback:\\n \\n Enter PEM passhprase:\\n \\n That's infesible in many situations, especially things like use from\\n postgres_fdw.\\n \\n This change also allows admins to prevent libpq from ever prompting\\nfor a\\n password by calling:\\n \\n PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\\n \\n which is useful since OpenSSL likes to open /dev/tty to prompt for a\\npassword,\\n so even closing stdin won't stop it blocking if there's no user\\ninput available.\\n Applications may also override or extend SSL password fetching with\\ntheir own\\n callback.\\n \\n There is deliberately no environment variable equivalent for the\\nsslpassword\\n option.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 18:33:12 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"libpq sslpassword parameter and callback function\"\n },\n {\n \"msg_contents\": \"This time with attachment.\\n\\n\\nOn 10/31/19 6:33 PM, Andrew Dunstan wrote:\\n> This patch provides for an sslpassword parameter for libpq, and a hook\\n> that a client can fill in for a callback function to set the password.\\n>\\n>\\n> This provides similar facilities to those already available in the JDBC\\n> driver.\\n>\\n>\\n> There is also a function to fetch the sslpassword from the connection\\n> parameters, in the same way that other settings can be fetched.\\n>\\n>\\n> This is mostly the excellent work of my colleague Craig Ringer, with a\\n> few embellishments from me.\\n>\\n>\\n> Here are his notes:\\n>\\n>\\n> Allow libpq to non-interactively decrypt client certificates that\\n> are stored\\n> encrypted by adding a new \\\"sslpassword\\\" connection option.\\n> \\n> The sslpassword option offers a middle ground between a cleartext\\n> key and\\n> setting up advanced key mangement via openssl engines, PKCS#11, USB\\n> crypto\\n> offload and key escrow, etc.\\n> \\n> Previously use of encrypted client certificate keys only worked if\\n> the user\\n> could enter the key's password interactively on stdin, in response\\n> to openssl's\\n> default prompt callback:\\n> \\n> Enter PEM passhprase:\\n> \\n> That's infesible in many situations, especially things like use from\\n> postgres_fdw.\\n> \\n> This change also allows admins to prevent libpq from ever prompting\\n> for a\\n> password by calling:\\n> \\n> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\\n> \\n> which is useful since OpenSSL likes to open /dev/tty to prompt for a\\n> password,\\n> so even closing stdin won't stop it blocking if there's no user\\n> input available.\\n> Applications may also override or extend SSL password fetching with\\n> their own\\n> callback.\\n> \\n> There is deliberately no environment variable equivalent for the\\n> sslpassword\\n> option.\\n>\\n>\\n> cheers\\n>\\n>\\n> andrew\\n>\\n>\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 31 Oct 2019 18:34:32 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: libpq sslpassword parameter and callback function\"\n },\n {\n \"msg_contents\": \"\\nOn 10/31/19 6:34 PM, Andrew Dunstan wrote:\\n> This time with attachment.\\n>\\n>\\n> On 10/31/19 6:33 PM, Andrew Dunstan wrote:\\n>> This patch provides for an sslpassword parameter for libpq, and a hook\\n>> that a client can fill in for a callback function to set the password.\\n>>\\n>>\\n>> This provides similar facilities to those already available in the JDBC\\n>> driver.\\n>>\\n>>\\n>> There is also a function to fetch the sslpassword from the connection\\n>> parameters, in the same way that other settings can be fetched.\\n>>\\n>>\\n>> This is mostly the excellent work of my colleague Craig Ringer, with a\\n>> few embellishments from me.\\n>>\\n>>\\n>> Here are his notes:\\n>>\\n>>\\n>> Allow libpq to non-interactively decrypt client certificates that\\n>> are stored\\n>> encrypted by adding a new \\\"sslpassword\\\" connection option.\\n>> \\n>> The sslpassword option offers a middle ground between a cleartext\\n>> key and\\n>> setting up advanced key mangement via openssl engines, PKCS#11, USB\\n>> crypto\\n>> offload and key escrow, etc.\\n>> \\n>> Previously use of encrypted client certificate keys only worked if\\n>> the user\\n>> could enter the key's password interactively on stdin, in response\\n>> to openssl's\\n>> default prompt callback:\\n>> \\n>> Enter PEM passhprase:\\n>> \\n>> That's infesible in many situations, especially things like use from\\n>> postgres_fdw.\\n>> \\n>> This change also allows admins to prevent libpq from ever prompting\\n>> for a\\n>> password by calling:\\n>> \\n>> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\\n>> \\n>> which is useful since OpenSSL likes to open /dev/tty to prompt for a\\n>> password,\\n>> so even closing stdin won't stop it blocking if there's no user\\n>> input available.\\n>> Applications may also override or extend SSL password fetching with\\n>> their own\\n>> callback.\\n>> \\n>> There is deliberately no environment variable equivalent for the\\n>> sslpassword\\n>> option.\\n>>\\n>>\\n\\nI should also mention that this patch provides for support for DER\\nformat certificates and keys.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 19:27:52 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: libpq sslpassword parameter and callback function\"\n },\n {\n \"msg_contents\": \"On Fri, 1 Nov 2019 at 07:27, Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\\nwrote:\\n\\n>\\n> On 10/31/19 6:34 PM, Andrew Dunstan wrote:\\n> > This time with attachment.\\n> >\\n> >\\n> > On 10/31/19 6:33 PM, Andrew Dunstan wrote:\\n> >> This patch provides for an sslpassword parameter for libpq, and a hook\\n> >> that a client can fill in for a callback function to set the password.\\n> >>\\n> >>\\n> >> This provides similar facilities to those already available in the JDBC\\n> >> driver.\\n> >>\\n> >>\\n> >> There is also a function to fetch the sslpassword from the connection\\n> >> parameters, in the same way that other settings can be fetched.\\n> >>\\n> >>\\n> >> This is mostly the excellent work of my colleague Craig Ringer, with a\\n> >> few embellishments from me.\\n> >>\\n> >>\\n> >> Here are his notes:\\n> >>\\n> >>\\n> >> Allow libpq to non-interactively decrypt client certificates that\\n> >> are stored\\n> >> encrypted by adding a new \\\"sslpassword\\\" connection option.\\n> >>\\n> >> The sslpassword option offers a middle ground between a cleartext\\n> >> key and\\n> >> setting up advanced key mangement via openssl engines, PKCS#11, USB\\n> >> crypto\\n> >> offload and key escrow, etc.\\n> >>\\n> >> Previously use of encrypted client certificate keys only worked if\\n> >> the user\\n> >> could enter the key's password interactively on stdin, in response\\n> >> to openssl's\\n> >> default prompt callback:\\n> >>\\n> >> Enter PEM passhprase:\\n> >>\\n> >> That's infesible in many situations, especially things like use from\\n> >> postgres_fdw.\\n> >>\\n> >> This change also allows admins to prevent libpq from ever prompting\\n> >> for a\\n> >> password by calling:\\n> >>\\n> >> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\\n> >>\\n> >> which is useful since OpenSSL likes to open /dev/tty to prompt for a\\n> >> password,\\n> >> so even closing stdin won't stop it blocking if there's no user\\n> >> input available.\\n> >> Applications may also override or extend SSL password fetching with\\n> >> their own\\n> >> callback.\\n> >>\\n> >> There is deliberately no environment variable equivalent for the\\n> >> sslpassword\\n> >> option.\\n> >>\\n> >>\\n>\\n> I should also mention that this patch provides for support for DER\\n> format certificates and keys.\\n>\\n>\\nYep, that was a trivial change I rolled into it.\\n\\nFWIW, this is related to two other patches: the patch to allow passwordless\\nfdw connections with explicit superuser approval, and the patch to allow\\nsslkey/sslpassword to be set as user mapping options in postgres_fdw .\\nTogether all three patches make it possible to use SSL client certificates\\nto manage authentication in postgres_fdw user mappings.\\n\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\\n\\nOn Fri, 1 Nov 2019 at 07:27, Andrew Dunstan <andrew.dunstan@2ndquadrant.com> wrote:\\nOn 10/31/19 6:34 PM, Andrew Dunstan wrote:\\n> This time with attachment.\\n>\\n>\\n> On 10/31/19 6:33 PM, Andrew Dunstan wrote:\\n>> This patch provides for an sslpassword parameter for libpq, and a hook\\n>> that a client can fill in for a callback function to set the password.\\n>>\\n>>\\n>> This provides similar facilities to those already available in the JDBC\\n>> driver.\\n>>\\n>>\\n>> There is also a function to fetch the sslpassword from the connection\\n>> parameters, in the same way that other settings can be fetched.\\n>>\\n>>\\n>> This is mostly the excellent work of my colleague Craig Ringer, with a\\n>> few embellishments from me.\\n>>\\n>>\\n>> Here are his notes:\\n>>\\n>>\\n>> Allow libpq to non-interactively decrypt client certificates that\\n>> are stored\\n>> encrypted by adding a new \\\"sslpassword\\\" connection option.\\n>> \\n>> The sslpassword option offers a middle ground between a cleartext\\n>> key and\\n>> setting up advanced key mangement via openssl engines, PKCS#11, USB\\n>> crypto\\n>> offload and key escrow, etc.\\n>> \\n>> Previously use of encrypted client certificate keys only worked if\\n>> the user\\n>> could enter the key's password interactively on stdin, in response\\n>> to openssl's\\n>> default prompt callback:\\n>> \\n>> Enter PEM passhprase:\\n>> \\n>> That's infesible in many situations, especially things like use from\\n>> postgres_fdw.\\n>> \\n>> This change also allows admins to prevent libpq from ever prompting\\n>> for a\\n>> password by calling:\\n>> \\n>> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\\n>> \\n>> which is useful since OpenSSL likes to open /dev/tty to prompt for a\\n>> password,\\n>> so even closing stdin won't stop it blocking if there's no user\\n>> input available.\\n>> Applications may also override or extend SSL password fetching with\\n>> their own\\n>> callback.\\n>> \\n>> There is deliberately no environment variable equivalent for the\\n>> sslpassword\\n>> option.\\n>>\\n>>\\n\\nI should also mention that this patch provides for support for DER\\nformat certificates and keys.Yep, that was a trivial change I rolled into it.FWIW, this is related to two other patches: the patch to allow passwordless fdw connections with explicit superuser approval, and the patch to allow sslkey/sslpassword to be set as user mapping options in postgres_fdw . Together all three patches make it possible to use SSL client certificates to manage authentication in postgres_fdw user mappings.-- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise\",\n \"msg_date\": \"Sun, 10 Nov 2019 17:47:24 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: libpq sslpassword parameter and callback function\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 10, 2019 at 05:47:24PM +0800, Craig Ringer wrote:\\n> Yep, that was a trivial change I rolled into it.\\n> \\n> FWIW, this is related to two other patches: the patch to allow passwordless fdw\\n> connections with explicit superuser approval, and the patch to allow sslkey/\\n> sslpassword to be set as user mapping options in postgres_fdw . Together all\\n> three patches make it possible to use SSL client certificates to manage\\n> authentication in postgres_fdw user mappings.\\n\\nOh, nice, greatly needed!\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Nov 2019 22:10:38 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: libpq sslpassword parameter and callback function\"\n },\n {\n \"msg_contents\": \"On 10/31/19 7:27 PM, Andrew Dunstan wrote:\\n> On 10/31/19 6:34 PM, Andrew Dunstan wrote:\\n>> This time with attachment.\\n>>\\n>>\\n>> On 10/31/19 6:33 PM, Andrew Dunstan wrote:\\n>>> This patch provides for an sslpassword parameter for libpq, and a hook\\n>>> that a client can fill in for a callback function to set the password.\\n>>>\\n>>>\\n>>> This provides similar facilities to those already available in the JDBC\\n>>> driver.\\n>>>\\n>>>\\n>>> There is also a function to fetch the sslpassword from the connection\\n>>> parameters, in the same way that other settings can be fetched.\\n>>>\\n>>>\\n>>> This is mostly the excellent work of my colleague Craig Ringer, with a\\n>>> few embellishments from me.\\n>>>\\n>>>\\n>>> Here are his notes:\\n>>>\\n>>>\\n>>> Allow libpq to non-interactively decrypt client certificates that\\n>>> are stored\\n>>> encrypted by adding a new \\\"sslpassword\\\" connection option.\\n>>> \\n>>> The sslpassword option offers a middle ground between a cleartext\\n>>> key and\\n>>> setting up advanced key mangement via openssl engines, PKCS#11, USB\\n>>> crypto\\n>>> offload and key escrow, etc.\\n>>> \\n>>> Previously use of encrypted client certificate keys only worked if\\n>>> the user\\n>>> could enter the key's password interactively on stdin, in response\\n>>> to openssl's\\n>>> default prompt callback:\\n>>> \\n>>> Enter PEM passhprase:\\n>>> \\n>>> That's infesible in many situations, especially things like use from\\n>>> postgres_fdw.\\n>>> \\n>>> This change also allows admins to prevent libpq from ever prompting\\n>>> for a\\n>>> password by calling:\\n>>> \\n>>> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\\n>>> \\n>>> which is useful since OpenSSL likes to open /dev/tty to prompt for a\\n>>> password,\\n>>> so even closing stdin won't stop it blocking if there's no user\\n>>> input available.\\n>>> Applications may also override or extend SSL password fetching with\\n>>> their own\\n>>> callback.\\n>>> \\n>>> There is deliberately no environment variable equivalent for the\\n>>> sslpassword\\n>>> option.\\n>>>\\n>>>\\n> I should also mention that this patch provides for support for DER\\n> format certificates and keys.\\n>\\n>\\n\\n\\nHere's an updated version of the patch, adjusted to the now committed\\nchanges to TestLib.pm.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\n\\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Mon, 25 Nov 2019 16:09:08 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: libpq sslpassword parameter and callback function\"\n },\n {\n \"msg_contents\": \"On 11/25/19 4:09 PM, Andrew Dunstan wrote:\\n> On 10/31/19 7:27 PM, Andrew Dunstan wrote:\\n>> On 10/31/19 6:34 PM, Andrew Dunstan wrote:\\n>>> This time with attachment.\\n>>>\\n>>>\\n>>> On 10/31/19 6:33 PM, Andrew Dunstan wrote:\\n>>>> This patch provides for an sslpassword parameter for libpq, and a hook\\n>>>> that a client can fill in for a callback function to set the password.\\n>>>>\\n>>>>\\n>>>> This provides similar facilities to those already available in the JDBC\\n>>>> driver.\\n>>>>\\n>>>>\\n>>>> There is also a function to fetch the sslpassword from the connection\\n>>>> parameters, in the same way that other settings can be fetched.\\n>>>>\\n>>>>\\n>>>> This is mostly the excellent work of my colleague Craig Ringer, with a\\n>>>> few embellishments from me.\\n>>>>\\n>>>>\\n>>>> Here are his notes:\\n>>>>\\n>>>>\\n>>>> Allow libpq to non-interactively decrypt client certificates that\\n>>>> are stored\\n>>>> encrypted by adding a new \\\"sslpassword\\\" connection option.\\n>>>> \\n>>>> The sslpassword option offers a middle ground between a cleartext\\n>>>> key and\\n>>>> setting up advanced key mangement via openssl engines, PKCS#11, USB\\n>>>> crypto\\n>>>> offload and key escrow, etc.\\n>>>> \\n>>>> Previously use of encrypted client certificate keys only worked if\\n>>>> the user\\n>>>> could enter the key's password interactively on stdin, in response\\n>>>> to openssl's\\n>>>> default prompt callback:\\n>>>> \\n>>>> Enter PEM passhprase:\\n>>>> \\n>>>> That's infesible in many situations, especially things like use from\\n>>>> postgres_fdw.\\n>>>> \\n>>>> This change also allows admins to prevent libpq from ever prompting\\n>>>> for a\\n>>>> password by calling:\\n>>>> \\n>>>> PQsetSSLKeyPassHook(PQdefaultSSLKeyPassHook);\\n>>>> \\n>>>> which is useful since OpenSSL likes to open /dev/tty to prompt for a\\n>>>> password,\\n>>>> so even closing stdin won't stop it blocking if there's no user\\n>>>> input available.\\n>>>> Applications may also override or extend SSL password fetching with\\n>>>> their own\\n>>>> callback.\\n>>>> \\n>>>> There is deliberately no environment variable equivalent for the\\n>>>> sslpassword\\n>>>> option.\\n>>>>\\n>>>>\\n>> I should also mention that this patch provides for support for DER\\n>> format certificates and keys.\\n>>\\n>>\\n>\\n> Here's an updated version of the patch, adjusted to the now committed\\n> changes to TestLib.pm.\\n>\\n>\\n\\n\\nHere's an update now we have backed out the TestLib changes. The tests\\nthat need a pty are skipped.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\n\\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Wed, 27 Nov 2019 19:06:10 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: libpq sslpassword parameter and callback function\"\n },\n {\n \"msg_contents\": \"The following review has been posted through the commitfest application:\\nmake installcheck-world: tested, passed\\nImplements feature: tested, passed\\nSpec compliant: tested, passed\\nDocumentation: tested, failed\\n\\nHi Andrew,\\r\\n\\r\\nI've reviewed your \\\"libpq sslpassword parameter and callback function\\\" patch (0001-libpq-sslpassword-der-support.patch), and only found a few minor things (otherwise it looked good to me):\\r\\n\\r\\n1) There's a few trailing white-space warnings on patch application (from where you modified to skip 2 of the tests):\\r\\ngit apply 0001-libpq-sslpassword-der-support.patch\\r\\n0001-libpq-sslpassword-der-support.patch:649: trailing whitespace.\\r\\n\\t# so they don't hang. For now they are not performed. \\r\\n0001-libpq-sslpassword-der-support.patch:659: trailing whitespace.\\r\\n\\t\\r\\nwarning: 2 lines add whitespace errors.\\r\\n\\r\\n\\r\\n2) src/interfaces/libpq/libpq-fe.h\\r\\nThe following portion of the comment should be removed.\\r\\n\\r\\n+ * 2ndQPostgres extension. If you need to be compatible with unpatched libpq\\r\\n+ * you must dlsym() these.\\r\\n\\r\\n3) Documentation for the \\\"PQsslpassword\\\" function should be added to the libpq \\\"33.2 Connection Status Functions\\\" section.\\r\\n\\r\\n\\r\\nI made the following notes about how/what I reviewed and tested:\\r\\n\\r\\n- Applied patch and built Postgres (--with-openssl --enable-tap-tests), checked build output\\r\\n- Checked patch code modifications (format, logic, memory usage, efficiency, corner cases etc.)\\r\\n- Built documentation and checked updated portions (format, grammar, details, completeness etc.)\\r\\n- Checked test updates \\r\\n- Ran updated contrib/dblink tests - confirmed all passed\\r\\n- Ran updated SSL (TAP) tests - confirmed all passed\\r\\n- Ran \\\"make installcheck-world\\\", as per review requirements\\r\\n- Wrote small libpq-based app to test:\\r\\n - new APIs (PQsslpassword, PQsetSSLKeyPassHook, PQgetSSLKeyPassHook, PQdefaultSSLKeyPassHook)\\r\\n - passphrase-protected key with/without patch\\r\\n - patch with/without new key password callack\\r\\n - patch and certificate with/without pass phrase protection on key\\r\\n - default callback, callback delegation\\r\\n - PEM/DER keys\\r\\n\\r\\n\\r\\nRegards,\\r\\nGreg\",\n \"msg_date\": \"Fri, 29 Nov 2019 03:25:45 +0000\",\n \"msg_from\": \"Greg Nancarrow <gregn4422@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: libpq sslpassword parameter and callback function\"\n },\n {\n \"msg_contents\": \"On 11/28/19 10:25 PM, Greg Nancarrow wrote:\\n> The following review has been posted through the commitfest application:\\n> make installcheck-world: tested, passed\\n> Implements feature: tested, passed\\n> Spec compliant: tested, passed\\n> Documentation: tested, failed\\n>\\n> Hi Andrew,\\n>\\n> I've reviewed your \\\"libpq sslpassword parameter and callback function\\\" patch (0001-libpq-sslpassword-der-support.patch), and only found a few minor things (otherwise it looked good to me):\\n>\\n> 1) There's a few trailing white-space warnings on patch application (from where you modified to skip 2 of the tests):\\n> git apply 0001-libpq-sslpassword-der-support.patch\\n> 0001-libpq-sslpassword-der-support.patch:649: trailing whitespace.\\n> \\t# so they don't hang. For now they are not performed. \\n> 0001-libpq-sslpassword-der-support.patch:659: trailing whitespace.\\n> \\t\\n> warning: 2 lines add whitespace errors.\\n>\\n>\\n> 2) src/interfaces/libpq/libpq-fe.h\\n> The following portion of the comment should be removed.\\n>\\n> + * 2ndQPostgres extension. If you need to be compatible with unpatched libpq\\n> + * you must dlsym() these.\\n>\\n> 3) Documentation for the \\\"PQsslpassword\\\" function should be added to the libpq \\\"33.2 Connection Status Functions\\\" section.\\n>\\n>\\n> I made the following notes about how/what I reviewed and tested:\\n>\\n> - Applied patch and built Postgres (--with-openssl --enable-tap-tests), checked build output\\n> - Checked patch code modifications (format, logic, memory usage, efficiency, corner cases etc.)\\n> - Built documentation and checked updated portions (format, grammar, details, completeness etc.)\\n> - Checked test updates \\n> - Ran updated contrib/dblink tests - confirmed all passed\\n> - Ran updated SSL (TAP) tests - confirmed all passed\\n> - Ran \\\"make installcheck-world\\\", as per review requirements\\n> - Wrote small libpq-based app to test:\\n> - new APIs (PQsslpassword, PQsetSSLKeyPassHook, PQgetSSLKeyPassHook, PQdefaultSSLKeyPassHook)\\n> - passphrase-protected key with/without patch\\n> - patch with/without new key password callack\\n> - patch and certificate with/without pass phrase protection on key\\n> - default callback, callback delegation\\n> - PEM/DER keys\\n>\\n>\\n\\n\\nThanks, nice thorough review.\\n\\n\\nHere's an updated patch that I think fixes all the things you mentioned.\\nI plan to commit this tomorrow.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 29 Nov 2019 09:27:02 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: libpq sslpassword parameter and callback function\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 29, 2019 at 09:27:02AM -0500, Andrew Dunstan wrote:\\n> On 11/28/19 10:25 PM, Greg Nancarrow wrote:\\n> > 3) Documentation for the \\\"PQsslpassword\\\" function should be added to the libpq \\\"33.2 Connection Status Functions\\\" section.\\n\\n> I plan to commit this tomorrow.\\n\\nThe PQsslpassword() function retrieves a connection parameter value, which one\\ncan retrieve with PQconninfo(). Since introducing PQconninfo(), we have not\\nadded any per-parameter accessor functions. Would you remove PQsslpassword()?\\n\\n\\n\",\n \"msg_date\": \"Fri, 6 Dec 2019 07:57:15 +0000\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: libpq sslpassword parameter and callback function\"\n },\n {\n \"msg_contents\": \"\\nOn 12/6/19 2:57 AM, Noah Misch wrote:\\n> On Fri, Nov 29, 2019 at 09:27:02AM -0500, Andrew Dunstan wrote:\\n>> On 11/28/19 10:25 PM, Greg Nancarrow wrote:\\n>>> 3) Documentation for the \\\"PQsslpassword\\\" function should be added to the libpq \\\"33.2 Connection Status Functions\\\" section.\\n>> I plan to commit this tomorrow.\\n> The PQsslpassword() function retrieves a connection parameter value, which one\\n> can retrieve with PQconninfo(). Since introducing PQconninfo(), we have not\\n> added any per-parameter accessor functions. Would you remove PQsslpassword()?\\n\\n\\n\\n*sigh* ok\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 7 Dec 2019 08:51:20 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: libpq sslpassword parameter and callback function\"\n }\n]","isConversational":false}}},{"rowIdx":49385,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello hackers,\n\nPlease feel free to edit this new page, which I'd like to use to keep\ntrack of observations, ideas and threads relating to hash joins.\n\nhttps://wiki.postgresql.org/wiki/Hash_Join\n\n\n","msg_date":"Fri, 1 Nov 2019 10:16:35 +1100","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":true,"msg_subject":"A wiki page to track hash join projects and ideas"}],"string":"[\n {\n \"msg_contents\": \"Hello hackers,\\n\\nPlease feel free to edit this new page, which I'd like to use to keep\\ntrack of observations, ideas and threads relating to hash joins.\\n\\nhttps://wiki.postgresql.org/wiki/Hash_Join\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 10:16:35 +1100\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"A wiki page to track hash join projects and ideas\"\n }\n]","isConversational":false}}},{"rowIdx":49386,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi!\n\nOur customer faced with issue, when index is invisible after creation.\nThe reproducible case is following.\n\n $ psql db2\n # begin;\n # select txid_current();\n$ psql db1\n# select i as id, 0 as v into t from generate_series(1, 100000) i;\n# create unique index idx on t (id);\n# update t set v = v + 1 where id = 10000;\n# update t set v = v + 1 where id = 10000;\n# update t set v = v + 1 where id = 10000;\n# update t set v = v + 1 where id = 10000;\n# update t set v = v + 1 where id = 10000;\n# drop index idx;\n# create unique index idx on t (id);\n# explain analyze select v from t where id = 10000;\n\nThere is no issue if there is no parallel session in database db2.\nThe fact that index visibility depends on open transaction in\ndifferent database is ridiculous for users.\n\nThis happens so, because we're checking that there is no broken HOT\nchains after index creation by comparison pg_index.xmin and\nTransactionXmin. So, we check that pg_index.xmin is in the past for\ncurrent transaction in lossy way by comparison just xmins. Attached\npatch changes this check to XidInMVCCSnapshot().\n\nWith patch the issue is gone. My doubt about this patch is that it\nchanges check with TransactionXmin to check with GetActiveSnapshot(),\nwhich might be more recent. However, query shouldn't be executer with\nolder snapshot than one it was planned with.\n\nAny thoughts?\n\n------\nAlexander Korotkov\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company","msg_date":"Fri, 1 Nov 2019 02:50:39 +0300","msg_from":"Alexander Korotkov <a.korotkov@postgrespro.ru>","msg_from_op":true,"msg_subject":"Improve checking for pg_index.xmin"},{"msg_contents":"On 01/11/2019 01:50, Alexander Korotkov wrote:\n> Hi!\n> \n> Our customer faced with issue, when index is invisible after creation.\n> The reproducible case is following.\n> \n> $ psql db2\n> # begin;\n> # select txid_current();\n> $ psql db1\n> # select i as id, 0 as v into t from generate_series(1, 100000) i;\n> # create unique index idx on t (id);\n> # update t set v = v + 1 where id = 10000;\n> # update t set v = v + 1 where id = 10000;\n> # update t set v = v + 1 where id = 10000;\n> # update t set v = v + 1 where id = 10000;\n> # update t set v = v + 1 where id = 10000;\n> # drop index idx;\n> # create unique index idx on t (id);\n> # explain analyze select v from t where id = 10000;\n> \n> There is no issue if there is no parallel session in database db2.\n> The fact that index visibility depends on open transaction in\n> different database is ridiculous for users.\n> \n> This happens so, because we're checking that there is no broken HOT\n> chains after index creation by comparison pg_index.xmin and\n> TransactionXmin. So, we check that pg_index.xmin is in the past for\n> current transaction in lossy way by comparison just xmins. Attached\n> patch changes this check to XidInMVCCSnapshot().\n> \n> With patch the issue is gone. My doubt about this patch is that it\n> changes check with TransactionXmin to check with GetActiveSnapshot(),\n> which might be more recent. However, query shouldn't be executer with\n> older snapshot than one it was planned with.\n\nHmm. Maybe you could construct a case like that with a creative mix of \nstable and volatile functions? Using GetOldestSnapshot() would be safer.\n\n- Heikki\n\n\n","msg_date":"Wed, 8 Jan 2020 11:28:49 +0200","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":false,"msg_subject":"Re: Improve checking for pg_index.xmin"},{"msg_contents":"Heikki Linnakangas <hlinnaka@iki.fi> writes:\n> On 01/11/2019 01:50, Alexander Korotkov wrote:\n>> This happens so, because we're checking that there is no broken HOT\n>> chains after index creation by comparison pg_index.xmin and\n>> TransactionXmin. So, we check that pg_index.xmin is in the past for\n>> current transaction in lossy way by comparison just xmins. Attached\n>> patch changes this check to XidInMVCCSnapshot().\n>> With patch the issue is gone. My doubt about this patch is that it\n>> changes check with TransactionXmin to check with GetActiveSnapshot(),\n>> which might be more recent. However, query shouldn't be executer with\n>> older snapshot than one it was planned with.\n\n> Hmm. Maybe you could construct a case like that with a creative mix of \n> stable and volatile functions? Using GetOldestSnapshot() would be safer.\n\nI really wonder if this is safe at all.\n\n(1) Can we assume that the query will be executed with same-or-newer\nsnapshot as what was used by the planner? There's no such constraint\nin the plancache, I'm pretty sure.\n\n(2) Is committed-ness of the index-creating transaction actually\nsufficient to ensure that none of the broken HOT chains it saw are\na problem for the onlooker transaction? This is, at best, really\nun-obvious. Some of those HOT chains could involve xacts that were\nstill not committed when the index build finished, I believe.\n\n(3) What if the index was made with CREATE INDEX CONCURRENTLY ---\nwhich xid is actually on the pg_index row, and how does that factor\ninto (1) and (2)?\n\nOn the whole I don't find the risk/reward tradeoff of this looking\npromising. Even if it works reliably, I think the situations where\nit'll help a lot are a bit artificial.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 08 Jan 2020 08:37:28 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Improve checking for pg_index.xmin"},{"msg_contents":"On Wed, Jan 8, 2020 at 4:37 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Heikki Linnakangas <hlinnaka@iki.fi> writes:\n> > On 01/11/2019 01:50, Alexander Korotkov wrote:\n> >> This happens so, because we're checking that there is no broken HOT\n> >> chains after index creation by comparison pg_index.xmin and\n> >> TransactionXmin. So, we check that pg_index.xmin is in the past for\n> >> current transaction in lossy way by comparison just xmins. Attached\n> >> patch changes this check to XidInMVCCSnapshot().\n> >> With patch the issue is gone. My doubt about this patch is that it\n> >> changes check with TransactionXmin to check with GetActiveSnapshot(),\n> >> which might be more recent. However, query shouldn't be executer with\n> >> older snapshot than one it was planned with.\n>\n> > Hmm. Maybe you could construct a case like that with a creative mix of\n> > stable and volatile functions? Using GetOldestSnapshot() would be safer.\n>\n> I really wonder if this is safe at all.\n>\n> (1) Can we assume that the query will be executed with same-or-newer\n> snapshot as what was used by the planner? There's no such constraint\n> in the plancache, I'm pretty sure.\n>\n> (2) Is committed-ness of the index-creating transaction actually\n> sufficient to ensure that none of the broken HOT chains it saw are\n> a problem for the onlooker transaction? This is, at best, really\n> un-obvious. Some of those HOT chains could involve xacts that were\n> still not committed when the index build finished, I believe.\n>\n> (3) What if the index was made with CREATE INDEX CONCURRENTLY ---\n> which xid is actually on the pg_index row, and how does that factor\n> into (1) and (2)?\n\nThank you for pointing. I'll investigate these issues in detail.\n\n> On the whole I don't find the risk/reward tradeoff of this looking\n> promising. Even if it works reliably, I think the situations where\n> it'll help a lot are a bit artificial.\n\nI can't agree that these situations are artificial. For me, it seems\nnatural that user expects index to be visible once it's created.\nAlso, we always teach users that long-running transactions are evil,\nbut nevertheless they are frequent in real life. So, it doesn't seem\nunlikely that one expects index to become visible, while long\ntransaction is running in parallel. This particular case was reported\nby our customer. After investigation I was surprised how rare such\ncases are reported...\n\n------\nAlexander Korotkov\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company\n\n\n","msg_date":"Sun, 12 Jan 2020 01:02:50 +0300","msg_from":"Alexander Korotkov <a.korotkov@postgrespro.ru>","msg_from_op":true,"msg_subject":"Re: Improve checking for pg_index.xmin"},{"msg_contents":"On Sun, Jan 12, 2020 at 3:33 AM Alexander Korotkov\n<a.korotkov@postgrespro.ru> wrote:\n>\n> On Wed, Jan 8, 2020 at 4:37 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > Heikki Linnakangas <hlinnaka@iki.fi> writes:\n> > > On 01/11/2019 01:50, Alexander Korotkov wrote:\n> > >> This happens so, because we're checking that there is no broken HOT\n> > >> chains after index creation by comparison pg_index.xmin and\n> > >> TransactionXmin. So, we check that pg_index.xmin is in the past for\n> > >> current transaction in lossy way by comparison just xmins. Attached\n> > >> patch changes this check to XidInMVCCSnapshot().\n> > >> With patch the issue is gone. My doubt about this patch is that it\n> > >> changes check with TransactionXmin to check with GetActiveSnapshot(),\n> > >> which might be more recent. However, query shouldn't be executer with\n> > >> older snapshot than one it was planned with.\n> >\n> > > Hmm. Maybe you could construct a case like that with a creative mix of\n> > > stable and volatile functions? Using GetOldestSnapshot() would be safer.\n> >\n> > I really wonder if this is safe at all.\n> >\n> > (1) Can we assume that the query will be executed with same-or-newer\n> > snapshot as what was used by the planner? There's no such constraint\n> > in the plancache, I'm pretty sure.\n> >\n> > (2) Is committed-ness of the index-creating transaction actually\n> > sufficient to ensure that none of the broken HOT chains it saw are\n> > a problem for the onlooker transaction? This is, at best, really\n> > un-obvious. Some of those HOT chains could involve xacts that were\n> > still not committed when the index build finished, I believe.\n> >\n> > (3) What if the index was made with CREATE INDEX CONCURRENTLY ---\n> > which xid is actually on the pg_index row, and how does that factor\n> > into (1) and (2)?\n>\n> Thank you for pointing. I'll investigate these issues in detail.\n>\n\nAre you planning to work on this patch [1] for current CF? If not,\nthen I think it is better to either move this to the next CF or mark\nit as RWF.\n\n[1] - https://commitfest.postgresql.org/27/2337/\n\n-- \nWith Regards,\nAmit Kapila.\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 24 Mar 2020 18:08:29 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Improve checking for pg_index.xmin"},{"msg_contents":"On Tue, Mar 24, 2020 at 3:38 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> On Sun, Jan 12, 2020 at 3:33 AM Alexander Korotkov\n> <a.korotkov@postgrespro.ru> wrote:\n> >\n> > On Wed, Jan 8, 2020 at 4:37 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > > Heikki Linnakangas <hlinnaka@iki.fi> writes:\n> > > > On 01/11/2019 01:50, Alexander Korotkov wrote:\n> > > >> This happens so, because we're checking that there is no broken HOT\n> > > >> chains after index creation by comparison pg_index.xmin and\n> > > >> TransactionXmin. So, we check that pg_index.xmin is in the past for\n> > > >> current transaction in lossy way by comparison just xmins. Attached\n> > > >> patch changes this check to XidInMVCCSnapshot().\n> > > >> With patch the issue is gone. My doubt about this patch is that it\n> > > >> changes check with TransactionXmin to check with GetActiveSnapshot(),\n> > > >> which might be more recent. However, query shouldn't be executer with\n> > > >> older snapshot than one it was planned with.\n> > >\n> > > > Hmm. Maybe you could construct a case like that with a creative mix of\n> > > > stable and volatile functions? Using GetOldestSnapshot() would be safer.\n> > >\n> > > I really wonder if this is safe at all.\n> > >\n> > > (1) Can we assume that the query will be executed with same-or-newer\n> > > snapshot as what was used by the planner? There's no such constraint\n> > > in the plancache, I'm pretty sure.\n> > >\n> > > (2) Is committed-ness of the index-creating transaction actually\n> > > sufficient to ensure that none of the broken HOT chains it saw are\n> > > a problem for the onlooker transaction? This is, at best, really\n> > > un-obvious. Some of those HOT chains could involve xacts that were\n> > > still not committed when the index build finished, I believe.\n> > >\n> > > (3) What if the index was made with CREATE INDEX CONCURRENTLY ---\n> > > which xid is actually on the pg_index row, and how does that factor\n> > > into (1) and (2)?\n> >\n> > Thank you for pointing. I'll investigate these issues in detail.\n> >\n>\n> Are you planning to work on this patch [1] for current CF? If not,\n> then I think it is better to either move this to the next CF or mark\n> it as RWF.\n\nI didn't manage to investigate this subject and provide new patch\nversion. I'm marking this RWF.\n\n------\nAlexander Korotkov\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company\n\n\n","msg_date":"Wed, 25 Mar 2020 01:27:22 +0300","msg_from":"Alexander Korotkov <a.korotkov@postgrespro.ru>","msg_from_op":true,"msg_subject":"Re: Improve checking for pg_index.xmin"}],"string":"[\n {\n \"msg_contents\": \"Hi!\\n\\nOur customer faced with issue, when index is invisible after creation.\\nThe reproducible case is following.\\n\\n $ psql db2\\n # begin;\\n # select txid_current();\\n$ psql db1\\n# select i as id, 0 as v into t from generate_series(1, 100000) i;\\n# create unique index idx on t (id);\\n# update t set v = v + 1 where id = 10000;\\n# update t set v = v + 1 where id = 10000;\\n# update t set v = v + 1 where id = 10000;\\n# update t set v = v + 1 where id = 10000;\\n# update t set v = v + 1 where id = 10000;\\n# drop index idx;\\n# create unique index idx on t (id);\\n# explain analyze select v from t where id = 10000;\\n\\nThere is no issue if there is no parallel session in database db2.\\nThe fact that index visibility depends on open transaction in\\ndifferent database is ridiculous for users.\\n\\nThis happens so, because we're checking that there is no broken HOT\\nchains after index creation by comparison pg_index.xmin and\\nTransactionXmin. So, we check that pg_index.xmin is in the past for\\ncurrent transaction in lossy way by comparison just xmins. Attached\\npatch changes this check to XidInMVCCSnapshot().\\n\\nWith patch the issue is gone. My doubt about this patch is that it\\nchanges check with TransactionXmin to check with GetActiveSnapshot(),\\nwhich might be more recent. However, query shouldn't be executer with\\nolder snapshot than one it was planned with.\\n\\nAny thoughts?\\n\\n------\\nAlexander Korotkov\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\",\n \"msg_date\": \"Fri, 1 Nov 2019 02:50:39 +0300\",\n \"msg_from\": \"Alexander Korotkov <a.korotkov@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Improve checking for pg_index.xmin\"\n },\n {\n \"msg_contents\": \"On 01/11/2019 01:50, Alexander Korotkov wrote:\\n> Hi!\\n> \\n> Our customer faced with issue, when index is invisible after creation.\\n> The reproducible case is following.\\n> \\n> $ psql db2\\n> # begin;\\n> # select txid_current();\\n> $ psql db1\\n> # select i as id, 0 as v into t from generate_series(1, 100000) i;\\n> # create unique index idx on t (id);\\n> # update t set v = v + 1 where id = 10000;\\n> # update t set v = v + 1 where id = 10000;\\n> # update t set v = v + 1 where id = 10000;\\n> # update t set v = v + 1 where id = 10000;\\n> # update t set v = v + 1 where id = 10000;\\n> # drop index idx;\\n> # create unique index idx on t (id);\\n> # explain analyze select v from t where id = 10000;\\n> \\n> There is no issue if there is no parallel session in database db2.\\n> The fact that index visibility depends on open transaction in\\n> different database is ridiculous for users.\\n> \\n> This happens so, because we're checking that there is no broken HOT\\n> chains after index creation by comparison pg_index.xmin and\\n> TransactionXmin. So, we check that pg_index.xmin is in the past for\\n> current transaction in lossy way by comparison just xmins. Attached\\n> patch changes this check to XidInMVCCSnapshot().\\n> \\n> With patch the issue is gone. My doubt about this patch is that it\\n> changes check with TransactionXmin to check with GetActiveSnapshot(),\\n> which might be more recent. However, query shouldn't be executer with\\n> older snapshot than one it was planned with.\\n\\nHmm. Maybe you could construct a case like that with a creative mix of \\nstable and volatile functions? Using GetOldestSnapshot() would be safer.\\n\\n- Heikki\\n\\n\\n\",\n \"msg_date\": \"Wed, 8 Jan 2020 11:28:49 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Improve checking for pg_index.xmin\"\n },\n {\n \"msg_contents\": \"Heikki Linnakangas <hlinnaka@iki.fi> writes:\\n> On 01/11/2019 01:50, Alexander Korotkov wrote:\\n>> This happens so, because we're checking that there is no broken HOT\\n>> chains after index creation by comparison pg_index.xmin and\\n>> TransactionXmin. So, we check that pg_index.xmin is in the past for\\n>> current transaction in lossy way by comparison just xmins. Attached\\n>> patch changes this check to XidInMVCCSnapshot().\\n>> With patch the issue is gone. My doubt about this patch is that it\\n>> changes check with TransactionXmin to check with GetActiveSnapshot(),\\n>> which might be more recent. However, query shouldn't be executer with\\n>> older snapshot than one it was planned with.\\n\\n> Hmm. Maybe you could construct a case like that with a creative mix of \\n> stable and volatile functions? Using GetOldestSnapshot() would be safer.\\n\\nI really wonder if this is safe at all.\\n\\n(1) Can we assume that the query will be executed with same-or-newer\\nsnapshot as what was used by the planner? There's no such constraint\\nin the plancache, I'm pretty sure.\\n\\n(2) Is committed-ness of the index-creating transaction actually\\nsufficient to ensure that none of the broken HOT chains it saw are\\na problem for the onlooker transaction? This is, at best, really\\nun-obvious. Some of those HOT chains could involve xacts that were\\nstill not committed when the index build finished, I believe.\\n\\n(3) What if the index was made with CREATE INDEX CONCURRENTLY ---\\nwhich xid is actually on the pg_index row, and how does that factor\\ninto (1) and (2)?\\n\\nOn the whole I don't find the risk/reward tradeoff of this looking\\npromising. Even if it works reliably, I think the situations where\\nit'll help a lot are a bit artificial.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 08 Jan 2020 08:37:28 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Improve checking for pg_index.xmin\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 8, 2020 at 4:37 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Heikki Linnakangas <hlinnaka@iki.fi> writes:\\n> > On 01/11/2019 01:50, Alexander Korotkov wrote:\\n> >> This happens so, because we're checking that there is no broken HOT\\n> >> chains after index creation by comparison pg_index.xmin and\\n> >> TransactionXmin. So, we check that pg_index.xmin is in the past for\\n> >> current transaction in lossy way by comparison just xmins. Attached\\n> >> patch changes this check to XidInMVCCSnapshot().\\n> >> With patch the issue is gone. My doubt about this patch is that it\\n> >> changes check with TransactionXmin to check with GetActiveSnapshot(),\\n> >> which might be more recent. However, query shouldn't be executer with\\n> >> older snapshot than one it was planned with.\\n>\\n> > Hmm. Maybe you could construct a case like that with a creative mix of\\n> > stable and volatile functions? Using GetOldestSnapshot() would be safer.\\n>\\n> I really wonder if this is safe at all.\\n>\\n> (1) Can we assume that the query will be executed with same-or-newer\\n> snapshot as what was used by the planner? There's no such constraint\\n> in the plancache, I'm pretty sure.\\n>\\n> (2) Is committed-ness of the index-creating transaction actually\\n> sufficient to ensure that none of the broken HOT chains it saw are\\n> a problem for the onlooker transaction? This is, at best, really\\n> un-obvious. Some of those HOT chains could involve xacts that were\\n> still not committed when the index build finished, I believe.\\n>\\n> (3) What if the index was made with CREATE INDEX CONCURRENTLY ---\\n> which xid is actually on the pg_index row, and how does that factor\\n> into (1) and (2)?\\n\\nThank you for pointing. I'll investigate these issues in detail.\\n\\n> On the whole I don't find the risk/reward tradeoff of this looking\\n> promising. Even if it works reliably, I think the situations where\\n> it'll help a lot are a bit artificial.\\n\\nI can't agree that these situations are artificial. For me, it seems\\nnatural that user expects index to be visible once it's created.\\nAlso, we always teach users that long-running transactions are evil,\\nbut nevertheless they are frequent in real life. So, it doesn't seem\\nunlikely that one expects index to become visible, while long\\ntransaction is running in parallel. This particular case was reported\\nby our customer. After investigation I was surprised how rare such\\ncases are reported...\\n\\n------\\nAlexander Korotkov\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\\n\\n\\n\",\n \"msg_date\": \"Sun, 12 Jan 2020 01:02:50 +0300\",\n \"msg_from\": \"Alexander Korotkov <a.korotkov@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Improve checking for pg_index.xmin\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 12, 2020 at 3:33 AM Alexander Korotkov\\n<a.korotkov@postgrespro.ru> wrote:\\n>\\n> On Wed, Jan 8, 2020 at 4:37 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > Heikki Linnakangas <hlinnaka@iki.fi> writes:\\n> > > On 01/11/2019 01:50, Alexander Korotkov wrote:\\n> > >> This happens so, because we're checking that there is no broken HOT\\n> > >> chains after index creation by comparison pg_index.xmin and\\n> > >> TransactionXmin. So, we check that pg_index.xmin is in the past for\\n> > >> current transaction in lossy way by comparison just xmins. Attached\\n> > >> patch changes this check to XidInMVCCSnapshot().\\n> > >> With patch the issue is gone. My doubt about this patch is that it\\n> > >> changes check with TransactionXmin to check with GetActiveSnapshot(),\\n> > >> which might be more recent. However, query shouldn't be executer with\\n> > >> older snapshot than one it was planned with.\\n> >\\n> > > Hmm. Maybe you could construct a case like that with a creative mix of\\n> > > stable and volatile functions? Using GetOldestSnapshot() would be safer.\\n> >\\n> > I really wonder if this is safe at all.\\n> >\\n> > (1) Can we assume that the query will be executed with same-or-newer\\n> > snapshot as what was used by the planner? There's no such constraint\\n> > in the plancache, I'm pretty sure.\\n> >\\n> > (2) Is committed-ness of the index-creating transaction actually\\n> > sufficient to ensure that none of the broken HOT chains it saw are\\n> > a problem for the onlooker transaction? This is, at best, really\\n> > un-obvious. Some of those HOT chains could involve xacts that were\\n> > still not committed when the index build finished, I believe.\\n> >\\n> > (3) What if the index was made with CREATE INDEX CONCURRENTLY ---\\n> > which xid is actually on the pg_index row, and how does that factor\\n> > into (1) and (2)?\\n>\\n> Thank you for pointing. I'll investigate these issues in detail.\\n>\\n\\nAre you planning to work on this patch [1] for current CF? If not,\\nthen I think it is better to either move this to the next CF or mark\\nit as RWF.\\n\\n[1] - https://commitfest.postgresql.org/27/2337/\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 24 Mar 2020 18:08:29 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Improve checking for pg_index.xmin\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 24, 2020 at 3:38 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> On Sun, Jan 12, 2020 at 3:33 AM Alexander Korotkov\\n> <a.korotkov@postgrespro.ru> wrote:\\n> >\\n> > On Wed, Jan 8, 2020 at 4:37 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > > Heikki Linnakangas <hlinnaka@iki.fi> writes:\\n> > > > On 01/11/2019 01:50, Alexander Korotkov wrote:\\n> > > >> This happens so, because we're checking that there is no broken HOT\\n> > > >> chains after index creation by comparison pg_index.xmin and\\n> > > >> TransactionXmin. So, we check that pg_index.xmin is in the past for\\n> > > >> current transaction in lossy way by comparison just xmins. Attached\\n> > > >> patch changes this check to XidInMVCCSnapshot().\\n> > > >> With patch the issue is gone. My doubt about this patch is that it\\n> > > >> changes check with TransactionXmin to check with GetActiveSnapshot(),\\n> > > >> which might be more recent. However, query shouldn't be executer with\\n> > > >> older snapshot than one it was planned with.\\n> > >\\n> > > > Hmm. Maybe you could construct a case like that with a creative mix of\\n> > > > stable and volatile functions? Using GetOldestSnapshot() would be safer.\\n> > >\\n> > > I really wonder if this is safe at all.\\n> > >\\n> > > (1) Can we assume that the query will be executed with same-or-newer\\n> > > snapshot as what was used by the planner? There's no such constraint\\n> > > in the plancache, I'm pretty sure.\\n> > >\\n> > > (2) Is committed-ness of the index-creating transaction actually\\n> > > sufficient to ensure that none of the broken HOT chains it saw are\\n> > > a problem for the onlooker transaction? This is, at best, really\\n> > > un-obvious. Some of those HOT chains could involve xacts that were\\n> > > still not committed when the index build finished, I believe.\\n> > >\\n> > > (3) What if the index was made with CREATE INDEX CONCURRENTLY ---\\n> > > which xid is actually on the pg_index row, and how does that factor\\n> > > into (1) and (2)?\\n> >\\n> > Thank you for pointing. I'll investigate these issues in detail.\\n> >\\n>\\n> Are you planning to work on this patch [1] for current CF? If not,\\n> then I think it is better to either move this to the next CF or mark\\n> it as RWF.\\n\\nI didn't manage to investigate this subject and provide new patch\\nversion. I'm marking this RWF.\\n\\n------\\nAlexander Korotkov\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 25 Mar 2020 01:27:22 +0300\",\n \"msg_from\": \"Alexander Korotkov <a.korotkov@postgrespro.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Improve checking for pg_index.xmin\"\n }\n]","isConversational":false}}},{"rowIdx":49387,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"This patch achieves $SUBJECT and also provides some testing of the\nsslpassword setting.\n\n\ncheers\n\n\nandrew\n\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Thu, 31 Oct 2019 19:54:41 -0400","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"On Thu, Oct 31, 2019 at 07:54:41PM -0400, Andrew Dunstan wrote:\n> This patch achieves $SUBJECT and also provides some testing of the\n> sslpassword setting.\n\nThe patch does not apply anymore, so a rebase is needed. As it has\nnot been reviewed, I am moving it to next CF, waiting on author.\n--\nMichael","msg_date":"Sun, 1 Dec 2019 10:48:57 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"\nOn 11/30/19 8:48 PM, Michael Paquier wrote:\n> On Thu, Oct 31, 2019 at 07:54:41PM -0400, Andrew Dunstan wrote:\n>> This patch achieves $SUBJECT and also provides some testing of the\n>> sslpassword setting.\n> The patch does not apply anymore, so a rebase is needed. As it has\n> not been reviewed, I am moving it to next CF, waiting on author.\n\n\n\nThat's OK. This patch is dependent, as it always has been, on the patch\nto allow passwordless user mappings for postgres_fdw. I hope to commit\nthat soon, but I'd prefer some signoff from prominent hackers, as I\ndon't want to go too far down this road and then encounter a bunch of\nobjections.\n\n\ncheers\n\n\nandrew\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n","msg_date":"Sun, 1 Dec 2019 18:12:14 -0500","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"On 2019-12-02 00:12, Andrew Dunstan wrote:\n> On 11/30/19 8:48 PM, Michael Paquier wrote:\n>> On Thu, Oct 31, 2019 at 07:54:41PM -0400, Andrew Dunstan wrote:\n>>> This patch achieves $SUBJECT and also provides some testing of the\n>>> sslpassword setting.\n>> The patch does not apply anymore, so a rebase is needed. As it has\n>> not been reviewed, I am moving it to next CF, waiting on author.\n> \n> That's OK. This patch is dependent, as it always has been, on the patch\n> to allow passwordless user mappings for postgres_fdw. I hope to commit\n> that soon, but I'd prefer some signoff from prominent hackers, as I\n> don't want to go too far down this road and then encounter a bunch of\n> objections.\n\nThe prerequisite patch has been committed, so please see about getting \nthis patch moving forward.\n\nThe patch is very small, of course, but I don't understand the \"bit of a \nhack\" comment. Could you explain that?\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Wed, 8 Jan 2020 10:06:02 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"On Wed, Jan 8, 2020 at 7:36 PM Peter Eisentraut\n<peter.eisentraut@2ndquadrant.com> wrote:\n>\n> On 2019-12-02 00:12, Andrew Dunstan wrote:\n> > On 11/30/19 8:48 PM, Michael Paquier wrote:\n> >> On Thu, Oct 31, 2019 at 07:54:41PM -0400, Andrew Dunstan wrote:\n> >>> This patch achieves $SUBJECT and also provides some testing of the\n> >>> sslpassword setting.\n> >> The patch does not apply anymore, so a rebase is needed. As it has\n> >> not been reviewed, I am moving it to next CF, waiting on author.\n> >\n> > That's OK. This patch is dependent, as it always has been, on the patch\n> > to allow passwordless user mappings for postgres_fdw. I hope to commit\n> > that soon, but I'd prefer some signoff from prominent hackers, as I\n> > don't want to go too far down this road and then encounter a bunch of\n> > objections.\n>\n> The prerequisite patch has been committed, so please see about getting\n> this patch moving forward.\n>\n> The patch is very small, of course, but I don't understand the \"bit of a\n> hack\" comment. Could you explain that?\n>\n\n\nI have rewritten the comment, and committed the feature.\n\ncheers\n\nandrew\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Thu, 9 Jan 2020 18:44:33 +1030","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"Re: Andrew Dunstan 2019-11-01 <f941b95e-27ad-cb5c-2495-13c44f90b1bc@2ndQuadrant.com>\n> \t\t{\"password_required\", UserMappingRelationId, false},\n> +\t\t/*\n> +\t\t * Extra room for the user mapping copies of sslcert and sslkey. These\n> +\t\t * are really libpq options but we repeat them here to allow them to\n> +\t\t * appear in both foreign server context (when we generate libpq\n> +\t\t * options) and user mapping context (from here). Bit of a hack\n> +\t\t * putting this in \"non_libpq_options\".\n> +\t\t */\n> +\t\t{\"sslcert\", UserMappingRelationId, true},\n> +\t\t{\"sslkey\", UserMappingRelationId, true},\n\nNice feature, we were actually looking for exactly this yesterday.\n\nI have some concerns about security, though. It's true that the\nsslcert/sslkey options can only be set/modified by superusers when\n\"password_required\" is set. But when password_required is not set, any\nuser and create user mappings that reference arbitrary files on the\nserver filesystem. I believe the options are still used in that case\nfor creating connections, even when that means the remote server isn't\nset up for cert auth, which needs password_required=false to succeed.\n\nIn short, I believe these options need explicit superuser checks.\n\nChristoph\n\n\n","msg_date":"Thu, 9 Jan 2020 11:30:14 +0100","msg_from":"Christoph Berg <myon@debian.org>","msg_from_op":false,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"Re: To Andrew Dunstan 2020-01-09 <20200109103014.GA4192@msg.df7cb.de>\n> sslcert/sslkey options can only be set/modified by superusers when\n> \"password_required\" is set. But when password_required is not set, any\n> user and create user mappings that reference arbitrary files on the\n> server filesystem.\n\n(A nice addition here which would avoid the problems would be the\npossibility to pass in the certificates as strings, but that needs\nsupport in libpq.)\n\nChristoph\n\n\n","msg_date":"Thu, 9 Jan 2020 11:45:11 +0100","msg_from":"Christoph Berg <myon@debian.org>","msg_from_op":false,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"Re: To Andrew Dunstan 2020-01-09 <20200109103014.GA4192@msg.df7cb.de>\n> I believe the options are still used in that case\n> for creating connections, even when that means the remote server isn't\n> set up for cert auth, which needs password_required=false to succeed.\n\nThey are indeed:\n\nstat(\"/var/lib/postgresql/.postgresql/root.crt\", 0x7ffcff3e2bb0) = -1 ENOENT (Datei oder Verzeichnis nicht gefunden)\nstat(\"/foo\", 0x7ffcff3e2bb0) = -1 ENOENT (Datei oder Verzeichnis nicht gefunden)\n ^^^^ sslcert\n\nI'm not sure if that could be exploited in any way, but let's just\nforbid it.\n\nChristoph\n\n\n","msg_date":"Thu, 9 Jan 2020 13:48:55 +0100","msg_from":"Christoph Berg <myon@debian.org>","msg_from_op":false,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"On Thu, Jan 9, 2020 at 5:30 AM Christoph Berg <myon@debian.org> wrote:\n> I have some concerns about security, though. It's true that the\n> sslcert/sslkey options can only be set/modified by superusers when\n> \"password_required\" is set. But when password_required is not set, any\n> user and create user mappings that reference arbitrary files on the\n> server filesystem. I believe the options are still used in that case\n> for creating connections, even when that means the remote server isn't\n> set up for cert auth, which needs password_required=false to succeed.\n>\n> In short, I believe these options need explicit superuser checks.\n\nI share the concern about the security issue here. I can't testify to\nwhether Christoph's whole analysis is here, but as a general point,\nnon-superusers can't be allowed to do things that cause the server to\naccess arbitrary local files.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 9 Jan 2020 09:51:32 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"On Fri, Jan 10, 2020 at 1:21 AM Robert Haas <robertmhaas@gmail.com> wrote:\n>\n> On Thu, Jan 9, 2020 at 5:30 AM Christoph Berg <myon@debian.org> wrote:\n> > I have some concerns about security, though. It's true that the\n> > sslcert/sslkey options can only be set/modified by superusers when\n> > \"password_required\" is set. But when password_required is not set, any\n> > user and create user mappings that reference arbitrary files on the\n> > server filesystem. I believe the options are still used in that case\n> > for creating connections, even when that means the remote server isn't\n> > set up for cert auth, which needs password_required=false to succeed.\n> >\n> > In short, I believe these options need explicit superuser checks.\n>\n> I share the concern about the security issue here. I can't testify to\n> whether Christoph's whole analysis is here, but as a general point,\n> non-superusers can't be allowed to do things that cause the server to\n> access arbitrary local files.\n\n\nIt's probably fairly easy to do (c.f. 6136e94dcb). I'm not (yet)\nconvinced that there is any significant security threat here. This\ndoesn't give the user or indeed any postgres code any access to the\ncontents of these files. But if there is a consensus to restrict this\nI'll do it.\n\ncheers\n\nandrew\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 10 Jan 2020 08:08:42 +1030","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"> On 9 Jan 2020, at 22:38, Andrew Dunstan <andrew.dunstan@2ndquadrant.com> wrote:\n\n> I'm not (yet)\n> convinced that there is any significant security threat here. This\n> doesn't give the user or indeed any postgres code any access to the\n> contents of these files. But if there is a consensus to restrict this\n> I'll do it.\n\nI've seen successful exploits made from parts that I in my wildest imagination\ncouldn't think be useful, so FWIW +1 for adding belts to suspenders and\nrestricting this.\n\ncheers ./daniel\n\n","msg_date":"Thu, 9 Jan 2020 23:00:59 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n> On Fri, Jan 10, 2020 at 1:21 AM Robert Haas <robertmhaas@gmail.com> wrote:\n>> I share the concern about the security issue here. I can't testify to\n>> whether Christoph's whole analysis is here, but as a general point,\n>> non-superusers can't be allowed to do things that cause the server to\n>> access arbitrary local files.\n\n> It's probably fairly easy to do (c.f. 6136e94dcb). I'm not (yet)\n> convinced that there is any significant security threat here. This\n> doesn't give the user or indeed any postgres code any access to the\n> contents of these files. But if there is a consensus to restrict this\n> I'll do it.\n\nWell, even without access to the file contents, the mere ability to\nprobe the existence of a file is something we don't want unprivileged\nusers to have. And (I suppose) this is enough for that, by looking\nat what error you get back from trying it.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 09 Jan 2020 17:02:18 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"On Fri, Jan 10, 2020 at 8:32 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n> > On Fri, Jan 10, 2020 at 1:21 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> >> I share the concern about the security issue here. I can't testify to\n> >> whether Christoph's whole analysis is here, but as a general point,\n> >> non-superusers can't be allowed to do things that cause the server to\n> >> access arbitrary local files.\n>\n> > It's probably fairly easy to do (c.f. 6136e94dcb). I'm not (yet)\n> > convinced that there is any significant security threat here. This\n> > doesn't give the user or indeed any postgres code any access to the\n> > contents of these files. But if there is a consensus to restrict this\n> > I'll do it.\n>\n> Well, even without access to the file contents, the mere ability to\n> probe the existence of a file is something we don't want unprivileged\n> users to have. And (I suppose) this is enough for that, by looking\n> at what error you get back from trying it.\n>\n\n\nOK, that's convincing enough. Will do it before long.\n\ncheers\n\nandrew\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 10 Jan 2020 08:46:11 +1030","msg_from":"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"},{"msg_contents":"On Fri, 10 Jan 2020 at 06:16, Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\nwrote:\n\n> On Fri, Jan 10, 2020 at 8:32 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >\n> > Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n> > > On Fri, Jan 10, 2020 at 1:21 AM Robert Haas <robertmhaas@gmail.com>\n> wrote:\n> > >> I share the concern about the security issue here. I can't testify to\n> > >> whether Christoph's whole analysis is here, but as a general point,\n> > >> non-superusers can't be allowed to do things that cause the server to\n> > >> access arbitrary local files.\n> >\n> > > It's probably fairly easy to do (c.f. 6136e94dcb). I'm not (yet)\n> > > convinced that there is any significant security threat here. This\n> > > doesn't give the user or indeed any postgres code any access to the\n> > > contents of these files. But if there is a consensus to restrict this\n> > > I'll do it.\n> >\n> > Well, even without access to the file contents, the mere ability to\n> > probe the existence of a file is something we don't want unprivileged\n> > users to have. And (I suppose) this is enough for that, by looking\n> > at what error you get back from trying it.\n> >\n>\n>\n> OK, that's convincing enough. Will do it before long.\n\n\nThanks. I'm 100% convinced the superuser restriction should be imposed. I\ncan imagine there being a risk of leaking file contents in error output\nsuch as parse errors from OpenSSL that we pass on for example. Tricking Pg\ninto reading from a fifo could be problematic too.\n\nI should've applied that restriction from the start, the same way as\npasswordless connections are restricted.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\n\nOn Fri, 10 Jan 2020 at 06:16, Andrew Dunstan <andrew.dunstan@2ndquadrant.com> wrote:On Fri, Jan 10, 2020 at 8:32 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\n> > On Fri, Jan 10, 2020 at 1:21 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> >> I share the concern about the security issue here. I can't testify to\n> >> whether Christoph's whole analysis is here, but as a general point,\n> >> non-superusers can't be allowed to do things that cause the server to\n> >> access arbitrary local files.\n>\n> > It's probably fairly easy to do (c.f. 6136e94dcb). I'm not (yet)\n> > convinced that there is any significant security threat here. This\n> > doesn't give the user or indeed any postgres code any access to the\n> > contents of these files. But if there is a consensus to restrict this\n> > I'll do it.\n>\n> Well, even without access to the file contents, the mere ability to\n> probe the existence of a file is something we don't want unprivileged\n> users to have. And (I suppose) this is enough for that, by looking\n> at what error you get back from trying it.\n>\n\n\nOK, that's convincing enough. Will do it before long.Thanks. I'm 100% convinced the superuser restriction should be imposed. I can imagine there being a risk of leaking file contents in error output such as parse errors from OpenSSL that we pass on for example. Tricking Pg into reading from a fifo could be problematic too.I should've applied that restriction from the start, the same way as passwordless connections are restricted.-- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise","msg_date":"Mon, 20 Jan 2020 16:09:26 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings"}],"string":"[\n {\n \"msg_contents\": \"This patch achieves $SUBJECT and also provides some testing of the\\nsslpassword setting.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Thu, 31 Oct 2019 19:54:41 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 31, 2019 at 07:54:41PM -0400, Andrew Dunstan wrote:\\n> This patch achieves $SUBJECT and also provides some testing of the\\n> sslpassword setting.\\n\\nThe patch does not apply anymore, so a rebase is needed. As it has\\nnot been reviewed, I am moving it to next CF, waiting on author.\\n--\\nMichael\",\n \"msg_date\": \"Sun, 1 Dec 2019 10:48:57 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"\\nOn 11/30/19 8:48 PM, Michael Paquier wrote:\\n> On Thu, Oct 31, 2019 at 07:54:41PM -0400, Andrew Dunstan wrote:\\n>> This patch achieves $SUBJECT and also provides some testing of the\\n>> sslpassword setting.\\n> The patch does not apply anymore, so a rebase is needed. As it has\\n> not been reviewed, I am moving it to next CF, waiting on author.\\n\\n\\n\\nThat's OK. This patch is dependent, as it always has been, on the patch\\nto allow passwordless user mappings for postgres_fdw. I hope to commit\\nthat soon, but I'd prefer some signoff from prominent hackers, as I\\ndon't want to go too far down this road and then encounter a bunch of\\nobjections.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n\",\n \"msg_date\": \"Sun, 1 Dec 2019 18:12:14 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"On 2019-12-02 00:12, Andrew Dunstan wrote:\\n> On 11/30/19 8:48 PM, Michael Paquier wrote:\\n>> On Thu, Oct 31, 2019 at 07:54:41PM -0400, Andrew Dunstan wrote:\\n>>> This patch achieves $SUBJECT and also provides some testing of the\\n>>> sslpassword setting.\\n>> The patch does not apply anymore, so a rebase is needed. As it has\\n>> not been reviewed, I am moving it to next CF, waiting on author.\\n> \\n> That's OK. This patch is dependent, as it always has been, on the patch\\n> to allow passwordless user mappings for postgres_fdw. I hope to commit\\n> that soon, but I'd prefer some signoff from prominent hackers, as I\\n> don't want to go too far down this road and then encounter a bunch of\\n> objections.\\n\\nThe prerequisite patch has been committed, so please see about getting \\nthis patch moving forward.\\n\\nThe patch is very small, of course, but I don't understand the \\\"bit of a \\nhack\\\" comment. Could you explain that?\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Wed, 8 Jan 2020 10:06:02 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 8, 2020 at 7:36 PM Peter Eisentraut\\n<peter.eisentraut@2ndquadrant.com> wrote:\\n>\\n> On 2019-12-02 00:12, Andrew Dunstan wrote:\\n> > On 11/30/19 8:48 PM, Michael Paquier wrote:\\n> >> On Thu, Oct 31, 2019 at 07:54:41PM -0400, Andrew Dunstan wrote:\\n> >>> This patch achieves $SUBJECT and also provides some testing of the\\n> >>> sslpassword setting.\\n> >> The patch does not apply anymore, so a rebase is needed. As it has\\n> >> not been reviewed, I am moving it to next CF, waiting on author.\\n> >\\n> > That's OK. This patch is dependent, as it always has been, on the patch\\n> > to allow passwordless user mappings for postgres_fdw. I hope to commit\\n> > that soon, but I'd prefer some signoff from prominent hackers, as I\\n> > don't want to go too far down this road and then encounter a bunch of\\n> > objections.\\n>\\n> The prerequisite patch has been committed, so please see about getting\\n> this patch moving forward.\\n>\\n> The patch is very small, of course, but I don't understand the \\\"bit of a\\n> hack\\\" comment. Could you explain that?\\n>\\n\\n\\nI have rewritten the comment, and committed the feature.\\n\\ncheers\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Jan 2020 18:44:33 +1030\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"Re: Andrew Dunstan 2019-11-01 <f941b95e-27ad-cb5c-2495-13c44f90b1bc@2ndQuadrant.com>\\n> \\t\\t{\\\"password_required\\\", UserMappingRelationId, false},\\n> +\\t\\t/*\\n> +\\t\\t * Extra room for the user mapping copies of sslcert and sslkey. These\\n> +\\t\\t * are really libpq options but we repeat them here to allow them to\\n> +\\t\\t * appear in both foreign server context (when we generate libpq\\n> +\\t\\t * options) and user mapping context (from here). Bit of a hack\\n> +\\t\\t * putting this in \\\"non_libpq_options\\\".\\n> +\\t\\t */\\n> +\\t\\t{\\\"sslcert\\\", UserMappingRelationId, true},\\n> +\\t\\t{\\\"sslkey\\\", UserMappingRelationId, true},\\n\\nNice feature, we were actually looking for exactly this yesterday.\\n\\nI have some concerns about security, though. It's true that the\\nsslcert/sslkey options can only be set/modified by superusers when\\n\\\"password_required\\\" is set. But when password_required is not set, any\\nuser and create user mappings that reference arbitrary files on the\\nserver filesystem. I believe the options are still used in that case\\nfor creating connections, even when that means the remote server isn't\\nset up for cert auth, which needs password_required=false to succeed.\\n\\nIn short, I believe these options need explicit superuser checks.\\n\\nChristoph\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Jan 2020 11:30:14 +0100\",\n \"msg_from\": \"Christoph Berg <myon@debian.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"Re: To Andrew Dunstan 2020-01-09 <20200109103014.GA4192@msg.df7cb.de>\\n> sslcert/sslkey options can only be set/modified by superusers when\\n> \\\"password_required\\\" is set. But when password_required is not set, any\\n> user and create user mappings that reference arbitrary files on the\\n> server filesystem.\\n\\n(A nice addition here which would avoid the problems would be the\\npossibility to pass in the certificates as strings, but that needs\\nsupport in libpq.)\\n\\nChristoph\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Jan 2020 11:45:11 +0100\",\n \"msg_from\": \"Christoph Berg <myon@debian.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"Re: To Andrew Dunstan 2020-01-09 <20200109103014.GA4192@msg.df7cb.de>\\n> I believe the options are still used in that case\\n> for creating connections, even when that means the remote server isn't\\n> set up for cert auth, which needs password_required=false to succeed.\\n\\nThey are indeed:\\n\\nstat(\\\"/var/lib/postgresql/.postgresql/root.crt\\\", 0x7ffcff3e2bb0) = -1 ENOENT (Datei oder Verzeichnis nicht gefunden)\\nstat(\\\"/foo\\\", 0x7ffcff3e2bb0) = -1 ENOENT (Datei oder Verzeichnis nicht gefunden)\\n ^^^^ sslcert\\n\\nI'm not sure if that could be exploited in any way, but let's just\\nforbid it.\\n\\nChristoph\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Jan 2020 13:48:55 +0100\",\n \"msg_from\": \"Christoph Berg <myon@debian.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 9, 2020 at 5:30 AM Christoph Berg <myon@debian.org> wrote:\\n> I have some concerns about security, though. It's true that the\\n> sslcert/sslkey options can only be set/modified by superusers when\\n> \\\"password_required\\\" is set. But when password_required is not set, any\\n> user and create user mappings that reference arbitrary files on the\\n> server filesystem. I believe the options are still used in that case\\n> for creating connections, even when that means the remote server isn't\\n> set up for cert auth, which needs password_required=false to succeed.\\n>\\n> In short, I believe these options need explicit superuser checks.\\n\\nI share the concern about the security issue here. I can't testify to\\nwhether Christoph's whole analysis is here, but as a general point,\\nnon-superusers can't be allowed to do things that cause the server to\\naccess arbitrary local files.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Jan 2020 09:51:32 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 10, 2020 at 1:21 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> On Thu, Jan 9, 2020 at 5:30 AM Christoph Berg <myon@debian.org> wrote:\\n> > I have some concerns about security, though. It's true that the\\n> > sslcert/sslkey options can only be set/modified by superusers when\\n> > \\\"password_required\\\" is set. But when password_required is not set, any\\n> > user and create user mappings that reference arbitrary files on the\\n> > server filesystem. I believe the options are still used in that case\\n> > for creating connections, even when that means the remote server isn't\\n> > set up for cert auth, which needs password_required=false to succeed.\\n> >\\n> > In short, I believe these options need explicit superuser checks.\\n>\\n> I share the concern about the security issue here. I can't testify to\\n> whether Christoph's whole analysis is here, but as a general point,\\n> non-superusers can't be allowed to do things that cause the server to\\n> access arbitrary local files.\\n\\n\\nIt's probably fairly easy to do (c.f. 6136e94dcb). I'm not (yet)\\nconvinced that there is any significant security threat here. This\\ndoesn't give the user or indeed any postgres code any access to the\\ncontents of these files. But if there is a consensus to restrict this\\nI'll do it.\\n\\ncheers\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 10 Jan 2020 08:08:42 +1030\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"> On 9 Jan 2020, at 22:38, Andrew Dunstan <andrew.dunstan@2ndquadrant.com> wrote:\\n\\n> I'm not (yet)\\n> convinced that there is any significant security threat here. This\\n> doesn't give the user or indeed any postgres code any access to the\\n> contents of these files. But if there is a consensus to restrict this\\n> I'll do it.\\n\\nI've seen successful exploits made from parts that I in my wildest imagination\\ncouldn't think be useful, so FWIW +1 for adding belts to suspenders and\\nrestricting this.\\n\\ncheers ./daniel\\n\\n\",\n \"msg_date\": \"Thu, 9 Jan 2020 23:00:59 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n> On Fri, Jan 10, 2020 at 1:21 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n>> I share the concern about the security issue here. I can't testify to\\n>> whether Christoph's whole analysis is here, but as a general point,\\n>> non-superusers can't be allowed to do things that cause the server to\\n>> access arbitrary local files.\\n\\n> It's probably fairly easy to do (c.f. 6136e94dcb). I'm not (yet)\\n> convinced that there is any significant security threat here. This\\n> doesn't give the user or indeed any postgres code any access to the\\n> contents of these files. But if there is a consensus to restrict this\\n> I'll do it.\\n\\nWell, even without access to the file contents, the mere ability to\\nprobe the existence of a file is something we don't want unprivileged\\nusers to have. And (I suppose) this is enough for that, by looking\\nat what error you get back from trying it.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 09 Jan 2020 17:02:18 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 10, 2020 at 8:32 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n> > On Fri, Jan 10, 2020 at 1:21 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> >> I share the concern about the security issue here. I can't testify to\\n> >> whether Christoph's whole analysis is here, but as a general point,\\n> >> non-superusers can't be allowed to do things that cause the server to\\n> >> access arbitrary local files.\\n>\\n> > It's probably fairly easy to do (c.f. 6136e94dcb). I'm not (yet)\\n> > convinced that there is any significant security threat here. This\\n> > doesn't give the user or indeed any postgres code any access to the\\n> > contents of these files. But if there is a consensus to restrict this\\n> > I'll do it.\\n>\\n> Well, even without access to the file contents, the mere ability to\\n> probe the existence of a file is something we don't want unprivileged\\n> users to have. And (I suppose) this is enough for that, by looking\\n> at what error you get back from trying it.\\n>\\n\\n\\nOK, that's convincing enough. Will do it before long.\\n\\ncheers\\n\\nandrew\\n\\n-- \\nAndrew Dunstan https://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 10 Jan 2020 08:46:11 +1030\",\n \"msg_from\": \"Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n },\n {\n \"msg_contents\": \"On Fri, 10 Jan 2020 at 06:16, Andrew Dunstan <andrew.dunstan@2ndquadrant.com>\\nwrote:\\n\\n> On Fri, Jan 10, 2020 at 8:32 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >\\n> > Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n> > > On Fri, Jan 10, 2020 at 1:21 AM Robert Haas <robertmhaas@gmail.com>\\n> wrote:\\n> > >> I share the concern about the security issue here. I can't testify to\\n> > >> whether Christoph's whole analysis is here, but as a general point,\\n> > >> non-superusers can't be allowed to do things that cause the server to\\n> > >> access arbitrary local files.\\n> >\\n> > > It's probably fairly easy to do (c.f. 6136e94dcb). I'm not (yet)\\n> > > convinced that there is any significant security threat here. This\\n> > > doesn't give the user or indeed any postgres code any access to the\\n> > > contents of these files. But if there is a consensus to restrict this\\n> > > I'll do it.\\n> >\\n> > Well, even without access to the file contents, the mere ability to\\n> > probe the existence of a file is something we don't want unprivileged\\n> > users to have. And (I suppose) this is enough for that, by looking\\n> > at what error you get back from trying it.\\n> >\\n>\\n>\\n> OK, that's convincing enough. Will do it before long.\\n\\n\\nThanks. I'm 100% convinced the superuser restriction should be imposed. I\\ncan imagine there being a risk of leaking file contents in error output\\nsuch as parse errors from OpenSSL that we pass on for example. Tricking Pg\\ninto reading from a fifo could be problematic too.\\n\\nI should've applied that restriction from the start, the same way as\\npasswordless connections are restricted.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n 2ndQuadrant - PostgreSQL Solutions for the Enterprise\\n\\nOn Fri, 10 Jan 2020 at 06:16, Andrew Dunstan <andrew.dunstan@2ndquadrant.com> wrote:On Fri, Jan 10, 2020 at 8:32 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:\\n> > On Fri, Jan 10, 2020 at 1:21 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> >> I share the concern about the security issue here. I can't testify to\\n> >> whether Christoph's whole analysis is here, but as a general point,\\n> >> non-superusers can't be allowed to do things that cause the server to\\n> >> access arbitrary local files.\\n>\\n> > It's probably fairly easy to do (c.f. 6136e94dcb). I'm not (yet)\\n> > convinced that there is any significant security threat here. This\\n> > doesn't give the user or indeed any postgres code any access to the\\n> > contents of these files. But if there is a consensus to restrict this\\n> > I'll do it.\\n>\\n> Well, even without access to the file contents, the mere ability to\\n> probe the existence of a file is something we don't want unprivileged\\n> users to have. And (I suppose) this is enough for that, by looking\\n> at what error you get back from trying it.\\n>\\n\\n\\nOK, that's convincing enough. Will do it before long.Thanks. I'm 100% convinced the superuser restriction should be imposed. I can imagine there being a risk of leaking file contents in error output such as parse errors from OpenSSL that we pass on for example. Tricking Pg into reading from a fifo could be problematic too.I should've applied that restriction from the start, the same way as passwordless connections are restricted.-- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise\",\n \"msg_date\": \"Mon, 20 Jan 2020 16:09:26 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow 'sslkey' and 'sslcert' in postgres_fdw user mappings\"\n }\n]","isConversational":false}}},{"rowIdx":49388,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nSometimes you want to answer if a difference between two timestamps is\nlesser than x minutes but you are not sure which timestamp is greater\nthan the other one (to obtain a positive result -- it is not always\npossible). However, if you cannot obtain the absolute value of\nsubtraction, you have to add two conditions.\n\nThe attached patch implements abs function and @ operator for\nintervals. The following example illustrates the use case:\n\npostgres=# create table xpto (a timestamp, b timestamp);\nCREATE TABLE\npostgres=# insert into xpto (a, b) values(now(), now() - interval '1\nday'),(now() - interval '5 hour', now()),(now() + '3 hour', now());\nINSERT 0 3\npostgres=# select *, a - b as t from xpto;\n a | b | t\n----------------------------+----------------------------+-----------\n 2019-10-31 22:43:30.601861 | 2019-10-30 22:43:30.601861 | 1 day\n 2019-10-31 17:43:30.601861 | 2019-10-31 22:43:30.601861 | -05:00:00\n 2019-11-01 01:43:30.601861 | 2019-10-31 22:43:30.601861 | 03:00:00\n(3 rows)\n\npostgres=# select *, a - b as i from xpto where abs(a - b) < interval '12 hour';\n a | b | i\n----------------------------+----------------------------+-----------\n 2019-10-31 17:43:30.601861 | 2019-10-31 22:43:30.601861 | -05:00:00\n 2019-11-01 01:43:30.601861 | 2019-10-31 22:43:30.601861 | 03:00:00\n(2 rows)\n\npostgres=# select @ interval '1 years -2 months 3 days 4 hours -5\nminutes 6.789 seconds' as t;\n t\n-----------------------------\n 10 mons 3 days 03:55:06.789\n(1 row)\n\n\n-- \n Euler Taveira Timbira -\nhttp://www.timbira.com.br/\n PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento","msg_date":"Thu, 31 Oct 2019 23:20:07 -0300","msg_from":"Euler Taveira <euler@timbira.com.br>","msg_from_op":true,"msg_subject":"abs function for interval"},{"msg_contents":"Hi,\n\nOn 2019-10-31 23:20:07 -0300, Euler Taveira wrote:\n> diff --git a/src/backend/utils/adt/timestamp.c b/src/backend/utils/adt/timestamp.c\n> index 1dc4c820de..a6b8b8c221 100644\n> --- a/src/backend/utils/adt/timestamp.c\n> +++ b/src/backend/utils/adt/timestamp.c\n> @@ -2435,6 +2435,23 @@ interval_cmp(PG_FUNCTION_ARGS)\n> \tPG_RETURN_INT32(interval_cmp_internal(interval1, interval2));\n> }\n>\n> +Datum\n> +interval_abs(PG_FUNCTION_ARGS)\n> +{\n> +\tInterval *interval = PG_GETARG_INTERVAL_P(0);\n> +\tInterval *result;\n> +\n> +\tresult = palloc(sizeof(Interval));\n> +\t*result = *interval;\n> +\n> +\t/* convert all struct Interval members to absolute values */\n> +\tresult->month = (interval->month < 0) ? (-1 * interval->month) : interval->month;\n> +\tresult->day = (interval->day < 0) ? (-1 * interval->day) : interval->day;\n> +\tresult->time = (interval->time < 0) ? (-1 * interval->time) : interval->time;\n> +\n> +\tPG_RETURN_INTERVAL_P(result);\n> +}\n> +\n\nSeveral points:\n\n1) I don't think you can do the < 0 check on an elementwise basis. Your\n code would e.g. make a hash out of abs('1 day -1 second'), by\n inverting the second, but not the day (whereas nothing should be\n done).\n\n It'd probably be easiest to implement this by comparing with a 0\n interval using inteval_lt() or interval_cmp_internal().\n\n2) This will not correctly handle overflows, I believe. What happens if you\n do SELECT abs('-2147483648 days'::interval)? You probably should\n reuse interval_um() for this.\n\n\n> --- a/src/test/regress/expected/interval.out\n> +++ b/src/test/regress/expected/interval.out\n> @@ -927,3 +927,11 @@ select make_interval(secs := 7e12);\n> @ 1944444444 hours 26 mins 40 secs\n> (1 row)\n>\n> +-- test absolute operator\n> +set IntervalStyle to postgres;\n> +select @ interval '1 years -2 months 3 days 4 hours -5 minutes 6.789 seconds' as t;\n> + t\n> +-----------------------------\n> + 10 mons 3 days 03:55:06.789\n> +(1 row)\n> +\n> diff --git a/src/test/regress/sql/interval.sql b/src/test/regress/sql/interval.sql\n> index bc5537d1b9..8f9a2bda29 100644\n> --- a/src/test/regress/sql/interval.sql\n> +++ b/src/test/regress/sql/interval.sql\n\n\n> @@ -308,3 +308,7 @@ select make_interval(months := 'NaN'::float::int);\n> select make_interval(secs := 'inf');\n> select make_interval(secs := 'NaN');\n> select make_interval(secs := 7e12);\n> +\n> +-- test absolute operator\n> +set IntervalStyle to postgres;\n> +select @ interval '1 years -2 months 3 days 4 hours -5 minutes 6.789 seconds' as t;\n> --\n> 2.11.0\n\nThis is not even remotely close to enough tests. In your only test abs()\ndoes not change the value, as there's no negative component (the 1 year\n-2 month result in a positive 10 months, and the hours, minutes and\nseconds get folded together too).\n\nAt the very least a few boundary conditions need to be tested (see b)\nabove), a few more complicated cases with different components being\nof different signs, and you need to show the values with and without\napplying abs().\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 31 Oct 2019 19:45:24 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: abs function for interval"},{"msg_contents":"Em qui, 31 de out de 2019 às 23:45, Andres Freund <andres@anarazel.de> escreveu:\n>\n> 1) I don't think you can do the < 0 check on an elementwise basis. Your\n> code would e.g. make a hash out of abs('1 day -1 second'), by\n> inverting the second, but not the day (whereas nothing should be\n> done).\n>\n> It'd probably be easiest to implement this by comparing with a 0\n> interval using inteval_lt() or interval_cmp_internal().\n>\nHmm. Good idea. Let me try it.\n\n\n-- \n Euler Taveira Timbira -\nhttp://www.timbira.com.br/\n PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento\n\n\n","msg_date":"Fri, 1 Nov 2019 00:48:50 -0300","msg_from":"Euler Taveira <euler@timbira.com.br>","msg_from_op":true,"msg_subject":"Re: abs function for interval"},{"msg_contents":"On Fri, Nov 01, 2019 at 12:48:50AM -0300, Euler Taveira wrote:\n> Hmm. Good idea. Let me try it.\n\nMarked as RwF, as this has not been updated in four weeks. Please\nfeel free to resubmit later once you have an updated version.\n--\nMichael","msg_date":"Thu, 28 Nov 2019 13:17:38 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: abs function for interval"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nSometimes you want to answer if a difference between two timestamps is\\nlesser than x minutes but you are not sure which timestamp is greater\\nthan the other one (to obtain a positive result -- it is not always\\npossible). However, if you cannot obtain the absolute value of\\nsubtraction, you have to add two conditions.\\n\\nThe attached patch implements abs function and @ operator for\\nintervals. The following example illustrates the use case:\\n\\npostgres=# create table xpto (a timestamp, b timestamp);\\nCREATE TABLE\\npostgres=# insert into xpto (a, b) values(now(), now() - interval '1\\nday'),(now() - interval '5 hour', now()),(now() + '3 hour', now());\\nINSERT 0 3\\npostgres=# select *, a - b as t from xpto;\\n a | b | t\\n----------------------------+----------------------------+-----------\\n 2019-10-31 22:43:30.601861 | 2019-10-30 22:43:30.601861 | 1 day\\n 2019-10-31 17:43:30.601861 | 2019-10-31 22:43:30.601861 | -05:00:00\\n 2019-11-01 01:43:30.601861 | 2019-10-31 22:43:30.601861 | 03:00:00\\n(3 rows)\\n\\npostgres=# select *, a - b as i from xpto where abs(a - b) < interval '12 hour';\\n a | b | i\\n----------------------------+----------------------------+-----------\\n 2019-10-31 17:43:30.601861 | 2019-10-31 22:43:30.601861 | -05:00:00\\n 2019-11-01 01:43:30.601861 | 2019-10-31 22:43:30.601861 | 03:00:00\\n(2 rows)\\n\\npostgres=# select @ interval '1 years -2 months 3 days 4 hours -5\\nminutes 6.789 seconds' as t;\\n t\\n-----------------------------\\n 10 mons 3 days 03:55:06.789\\n(1 row)\\n\\n\\n-- \\n Euler Taveira Timbira -\\nhttp://www.timbira.com.br/\\n PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento\",\n \"msg_date\": \"Thu, 31 Oct 2019 23:20:07 -0300\",\n \"msg_from\": \"Euler Taveira <euler@timbira.com.br>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"abs function for interval\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-10-31 23:20:07 -0300, Euler Taveira wrote:\\n> diff --git a/src/backend/utils/adt/timestamp.c b/src/backend/utils/adt/timestamp.c\\n> index 1dc4c820de..a6b8b8c221 100644\\n> --- a/src/backend/utils/adt/timestamp.c\\n> +++ b/src/backend/utils/adt/timestamp.c\\n> @@ -2435,6 +2435,23 @@ interval_cmp(PG_FUNCTION_ARGS)\\n> \\tPG_RETURN_INT32(interval_cmp_internal(interval1, interval2));\\n> }\\n>\\n> +Datum\\n> +interval_abs(PG_FUNCTION_ARGS)\\n> +{\\n> +\\tInterval *interval = PG_GETARG_INTERVAL_P(0);\\n> +\\tInterval *result;\\n> +\\n> +\\tresult = palloc(sizeof(Interval));\\n> +\\t*result = *interval;\\n> +\\n> +\\t/* convert all struct Interval members to absolute values */\\n> +\\tresult->month = (interval->month < 0) ? (-1 * interval->month) : interval->month;\\n> +\\tresult->day = (interval->day < 0) ? (-1 * interval->day) : interval->day;\\n> +\\tresult->time = (interval->time < 0) ? (-1 * interval->time) : interval->time;\\n> +\\n> +\\tPG_RETURN_INTERVAL_P(result);\\n> +}\\n> +\\n\\nSeveral points:\\n\\n1) I don't think you can do the < 0 check on an elementwise basis. Your\\n code would e.g. make a hash out of abs('1 day -1 second'), by\\n inverting the second, but not the day (whereas nothing should be\\n done).\\n\\n It'd probably be easiest to implement this by comparing with a 0\\n interval using inteval_lt() or interval_cmp_internal().\\n\\n2) This will not correctly handle overflows, I believe. What happens if you\\n do SELECT abs('-2147483648 days'::interval)? You probably should\\n reuse interval_um() for this.\\n\\n\\n> --- a/src/test/regress/expected/interval.out\\n> +++ b/src/test/regress/expected/interval.out\\n> @@ -927,3 +927,11 @@ select make_interval(secs := 7e12);\\n> @ 1944444444 hours 26 mins 40 secs\\n> (1 row)\\n>\\n> +-- test absolute operator\\n> +set IntervalStyle to postgres;\\n> +select @ interval '1 years -2 months 3 days 4 hours -5 minutes 6.789 seconds' as t;\\n> + t\\n> +-----------------------------\\n> + 10 mons 3 days 03:55:06.789\\n> +(1 row)\\n> +\\n> diff --git a/src/test/regress/sql/interval.sql b/src/test/regress/sql/interval.sql\\n> index bc5537d1b9..8f9a2bda29 100644\\n> --- a/src/test/regress/sql/interval.sql\\n> +++ b/src/test/regress/sql/interval.sql\\n\\n\\n> @@ -308,3 +308,7 @@ select make_interval(months := 'NaN'::float::int);\\n> select make_interval(secs := 'inf');\\n> select make_interval(secs := 'NaN');\\n> select make_interval(secs := 7e12);\\n> +\\n> +-- test absolute operator\\n> +set IntervalStyle to postgres;\\n> +select @ interval '1 years -2 months 3 days 4 hours -5 minutes 6.789 seconds' as t;\\n> --\\n> 2.11.0\\n\\nThis is not even remotely close to enough tests. In your only test abs()\\ndoes not change the value, as there's no negative component (the 1 year\\n-2 month result in a positive 10 months, and the hours, minutes and\\nseconds get folded together too).\\n\\nAt the very least a few boundary conditions need to be tested (see b)\\nabove), a few more complicated cases with different components being\\nof different signs, and you need to show the values with and without\\napplying abs().\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Oct 2019 19:45:24 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: abs function for interval\"\n },\n {\n \"msg_contents\": \"Em qui, 31 de out de 2019 às 23:45, Andres Freund <andres@anarazel.de> escreveu:\\n>\\n> 1) I don't think you can do the < 0 check on an elementwise basis. Your\\n> code would e.g. make a hash out of abs('1 day -1 second'), by\\n> inverting the second, but not the day (whereas nothing should be\\n> done).\\n>\\n> It'd probably be easiest to implement this by comparing with a 0\\n> interval using inteval_lt() or interval_cmp_internal().\\n>\\nHmm. Good idea. Let me try it.\\n\\n\\n-- \\n Euler Taveira Timbira -\\nhttp://www.timbira.com.br/\\n PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 00:48:50 -0300\",\n \"msg_from\": \"Euler Taveira <euler@timbira.com.br>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: abs function for interval\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 01, 2019 at 12:48:50AM -0300, Euler Taveira wrote:\\n> Hmm. Good idea. Let me try it.\\n\\nMarked as RwF, as this has not been updated in four weeks. Please\\nfeel free to resubmit later once you have an updated version.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 28 Nov 2019 13:17:38 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: abs function for interval\"\n }\n]","isConversational":false}}},{"rowIdx":49389,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\n Postgres has a global variable `disable_cost`. It is set the value\n1.0e10.\n\n This value will be added to the cost of path if related GUC is set off.\nFor example,\n if enable_nestloop is set off, when planner trys to add nestloop join\npath, it continues\n to add such path but with a huge cost `disable_cost`.\n\n But 1.0e10 may not be large enough. I encounter this issue in\nGreenplum(based on postgres).\nHeikki tolds me that someone also encountered the same issue on Postgres.\nSo I send it here to\nhave a discussion.\n\n My issue: I did some spikes and tests on TPCDS 1TB Bytes data. For\nquery 104, it generates\n nestloop join even with enable_nestloop set off. And the final plan's\ntotal cost is very huge (about 1e24). But If I enlarge the disable_cost to\n1e30, then, planner will generate hash join.\n\n So I guess that disable_cost is not large enough for huge amount of\ndata.\n\n It is tricky to set disable_cost a huge number. Can we come up with\nbetter solution?\n\n The following thoughts are from Heikki:\n\n> Aside from not having a large enough disable cost, there's also the\n> fact that the high cost might affect the rest of the plan, if we have to\n> use a plan type that's disabled. For example, if a table doesn't have any\n> indexes, but enable_seqscan is off, we might put the unavoidable Seq Scan\n> on different side of a join than we we would with enable_seqscan=on,\n> because of the high cost estimate.\n\n\n\n> I think a more robust way to disable forbidden plan types would be to\n> handle the disabling in add_path(). Instead of having a high disable cost\n> on the Path itself, the comparison add_path() would always consider\n> disabled paths as more expensive than others, regardless of the cost.\n\n\n Any thoughts or ideas on the problem? Thanks!\n\nBest Regards,\nZhenghua Lyu\n\nHi, Postgres has a global variable `disable_cost`. It is set the value 1.0e10. This value will be added to the cost of path if related GUC is set off. For example, if enable_nestloop is set off, when planner trys to add nestloop join path, it continues to add such path but with a huge cost `disable_cost`. But 1.0e10 may not be large enough. I encounter this issue in Greenplum(based on postgres).Heikki tolds me that someone also encountered the same issue on Postgres. So I send it here tohave a discussion. My issue: I did some spikes and tests on TPCDS 1TB Bytes data. For query 104, it generates nestloop join even with enable_nestloop set off. And the final plan's total cost is very huge (about 1e24). But If I enlarge the disable_cost to 1e30, then, planner will generate hash join. So I guess that disable_cost is not large enough for huge amount of data. It is tricky to set disable_cost a huge number. Can we come up with better solution? The following thoughts are from Heikki: Aside from not having a large enough disable cost, there's also the fact that the high cost might affect the rest of the plan, if we have to use a plan type that's disabled. For example, if a table doesn't have any indexes, but enable_seqscan is off, we might put the unavoidable Seq Scan on different side of a join than we we would with enable_seqscan=on, because of the high cost estimate. I think a more robust way to disable forbidden plan types would be to handle the disabling in add_path(). Instead of having a high disable cost on the Path itself, the comparison add_path() would always consider disabled paths as more expensive than others, regardless of the cost. Any thoughts or ideas on the problem? Thanks!Best Regards,Zhenghua Lyu","msg_date":"Fri, 1 Nov 2019 14:42:25 +0800","msg_from":"Zhenghua Lyu <zlv@pivotal.io>","msg_from_op":true,"msg_subject":"On disable_cost"},{"msg_contents":"On Fri, Nov 1, 2019 at 7:42 PM Zhenghua Lyu <zlv@pivotal.io> wrote:\n> It is tricky to set disable_cost a huge number. Can we come up with better solution?\n\nWhat happens if you use DBL_MAX?\n\n\n","msg_date":"Fri, 1 Nov 2019 19:58:04 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Em sex, 1 de nov de 2019 às 03:42, Zhenghua Lyu <zlv@pivotal.io> escreveu:\n>\n> My issue: I did some spikes and tests on TPCDS 1TB Bytes data. For query 104, it generates\n> nestloop join even with enable_nestloop set off. And the final plan's total cost is very huge (about 1e24). But If I enlarge the disable_cost to 1e30, then, planner will generate hash join.\n>\n> So I guess that disable_cost is not large enough for huge amount of data.\n>\n> It is tricky to set disable_cost a huge number. Can we come up with better solution?\n>\nIsn't it a case for a GUC disable_cost? As Thomas suggested, DBL_MAX\nupper limit should be sufficient.\n\n> The following thoughts are from Heikki:\n>>\n>> Aside from not having a large enough disable cost, there's also the fact that the high cost might affect the rest of the plan, if we have to use a plan type that's disabled. For example, if a table doesn't have any indexes, but enable_seqscan is off, we might put the unavoidable Seq Scan on different side of a join than we we would with enable_seqscan=on, because of the high cost estimate.\n>\n>\n>>\n>> I think a more robust way to disable forbidden plan types would be to handle the disabling in add_path(). Instead of having a high disable cost on the Path itself, the comparison add_path() would always consider disabled paths as more expensive than others, regardless of the cost.\n>\nI'm afraid it is not as cheap as using diable_cost as a node cost. Are\nyou proposing to add a new boolean variable in Path struct to handle\nthose cases in compare_path_costs_fuzzily?\n\n\n-- \n Euler Taveira Timbira -\nhttp://www.timbira.com.br/\n PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento\n\n\n","msg_date":"Fri, 1 Nov 2019 11:48:29 -0300","msg_from":"Euler Taveira <euler@timbira.com.br>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Hi,\n\nOn 2019-11-01 19:58:04 +1300, Thomas Munro wrote:\n> On Fri, Nov 1, 2019 at 7:42 PM Zhenghua Lyu <zlv@pivotal.io> wrote:\n> > It is tricky to set disable_cost a huge number. Can we come up with better solution?\n> \n> What happens if you use DBL_MAX?\n\nThat seems like a bad idea - we add the cost multiple times. And we\nstill want to compare plans that potentially involve that cost, if\nthere's no other way to plan the query.\n\n- Andres\n\n\n","msg_date":"Fri, 1 Nov 2019 09:00:38 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Nov 1, 2019 at 12:00 PM Andres Freund <andres@anarazel.de> wrote:\n> That seems like a bad idea - we add the cost multiple times. And we\n> still want to compare plans that potentially involve that cost, if\n> there's no other way to plan the query.\n\nYeah. I kind of wonder if we shouldn't instead (a) skip adding paths\nthat use methods which are disabled and then (b) if we don't end up\nwith any paths for that reloptinfo, try again, ignoring disabling\nGUCs.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 1 Nov 2019 12:22:06 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"re: coping with adding disable_cost more than once\n\nAnother option would be to have a 2-part Cost structure. If disable_cost is\never added to the Cost, then you set a flag recording this. If any plans\nexist that have no disable_costs added to them, then the planner chooses the\nminimum cost among those, otherwise you choose the minimum cost path.\n\n\n\n-----\nJim Finnerty, AWS, Amazon Aurora PostgreSQL\n--\nSent from: https://www.postgresql-archive.org/PostgreSQL-hackers-f1928748.html\n\n\n","msg_date":"Fri, 1 Nov 2019 09:30:52 -0700 (MST)","msg_from":"Jim Finnerty <jfinnert@amazon.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Hi,\n\nOn 2019-11-01 12:22:06 -0400, Robert Haas wrote:\n> On Fri, Nov 1, 2019 at 12:00 PM Andres Freund <andres@anarazel.de> wrote:\n> > That seems like a bad idea - we add the cost multiple times. And we\n> > still want to compare plans that potentially involve that cost, if\n> > there's no other way to plan the query.\n> \n> Yeah. I kind of wonder if we shouldn't instead (a) skip adding paths\n> that use methods which are disabled and then (b) if we don't end up\n> with any paths for that reloptinfo, try again, ignoring disabling\n> GUCs.\n\nHm. That seems complicated. Is it clear that we'd always notice that we\nhave no plan early enough to know which paths to reconsider? I think\nthere's cases where that'd only happen a few levels up.\n\nAs a first step I'd be inclined to \"just\" adjust disable_cost up to\nsomething like 1.0e12. Unfortunately much higher and and we're getting\ninto the area where the loss of precision starts to be significant\nenough that I'm not sure that we're always careful enough to perform\nmath in the right order (e.g. 1.0e16 + 1 being 1.0e16, and 1e+20 + 1000\nbeing 1e+20). I've seen queries with costs above 1e10 where that costing\nwasn't insane.\n\nAnd then, in a larger patch, go for something like Heikki's proposal\nquoted by Zhenghua Lyu upthread, where we treat 'forbidden' as a\nseparate factor in comparisons of path costs, rather than fudging the\ncost upwards. But there's some care to be taken to make sure we don't\nregress performance too much due to the additional logic in\ncompare_path_cost et al.\n\nI'd also be curious to see if there's some other problem with cost\ncalculation here - some of the quoted final costs seem high enough to be\nsuspicious. I'd be curious to see a plan...\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 1 Nov 2019 09:43:14 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Nov 1, 2019 at 12:43 PM Andres Freund <andres@anarazel.de> wrote:\n> Hm. That seems complicated. Is it clear that we'd always notice that we\n> have no plan early enough to know which paths to reconsider? I think\n> there's cases where that'd only happen a few levels up.\n\nYeah, there could be problems of that kind. I think if a baserel has\nno paths, then we know right away that we've got a problem, but for\njoinrels it might be more complicated.\n\n> As a first step I'd be inclined to \"just\" adjust disable_cost up to\n> something like 1.0e12. Unfortunately much higher and and we're getting\n> into the area where the loss of precision starts to be significant\n> enough that I'm not sure that we're always careful enough to perform\n> math in the right order (e.g. 1.0e16 + 1 being 1.0e16, and 1e+20 + 1000\n> being 1e+20). I've seen queries with costs above 1e10 where that costing\n> wasn't insane.\n\nWe've done that before and we can do it again. But we're going to need\nto have something better eventually, I think, not just keep kicking\nthe can down the road.\n\nAnother point to consider here is that in some cases we could really\njust skip generating certain paths altogether. We already do this for\nhash joins: if we're planning a join and enable_hashjoin is disabled,\nwe just don't generate hash joins paths at all, except for full joins,\nwhere there might be no other legal method. As this example shows,\nthis cannot be applied in all cases, but maybe we could do it more\nwidely than we do today. I'm not sure how beneficial that technique\nwould be, though, because it doesn't seem like it's quite enough to\nsolve this problem by itself.\n\nYet another approach would be to divide the cost into two parts, a\n\"cost\" component and a \"violations\" component. If two paths are\ncompared, the one with fewer violations always wins; if it's a tie,\nthey compare on cost. A path's violation count is the total of its\nchildren, plus one for itself if it does something that's disabled.\nThis would be more principled than the current approach, but maybe\nit's too costly.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 1 Nov 2019 12:56:30 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Nov 01, 2019 at 09:30:52AM -0700, Jim Finnerty wrote:\n>re: coping with adding disable_cost more than once\n>\n>Another option would be to have a 2-part Cost structure. If disable_cost is\n>ever added to the Cost, then you set a flag recording this. If any plans\n>exist that have no disable_costs added to them, then the planner chooses the\n>minimum cost among those, otherwise you choose the minimum cost path.\n>\n\nYeah, I agree having is_disabled flag, and treat all paths with 'true'\nas more expensive than paths with 'false' (and when both paths have the\nsame value then actually compare the cost) is probably the way forward.\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Fri, 1 Nov 2019 18:04:06 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On 2019-11-01 12:56:30 -0400, Robert Haas wrote:\n> On Fri, Nov 1, 2019 at 12:43 PM Andres Freund <andres@anarazel.de> wrote:\n> > As a first step I'd be inclined to \"just\" adjust disable_cost up to\n> > something like 1.0e12. Unfortunately much higher and and we're getting\n> > into the area where the loss of precision starts to be significant\n> > enough that I'm not sure that we're always careful enough to perform\n> > math in the right order (e.g. 1.0e16 + 1 being 1.0e16, and 1e+20 + 1000\n> > being 1e+20). I've seen queries with costs above 1e10 where that costing\n> > wasn't insane.\n> \n> We've done that before and we can do it again. But we're going to need\n> to have something better eventually, I think, not just keep kicking\n> the can down the road.\n\nYea, that's why I continued on to describe what we should do afterwards\n;)\n\n\n> Yet another approach would be to divide the cost into two parts, a\n> \"cost\" component and a \"violations\" component. If two paths are\n> compared, the one with fewer violations always wins; if it's a tie,\n> they compare on cost. A path's violation count is the total of its\n> children, plus one for itself if it does something that's disabled.\n> This would be more principled than the current approach, but maybe\n> it's too costly.\n\nNamely go for something like this. I think we probably get away with the\nadditional comparison, especially if we were to store the violations as\nan integer and did it like if (unlikely(path1->nviolations !=\npath2->nviolations)) or such - that ought to be very well predicted in\nnearly all cases.\n\nI wonder how much we'd need to reformulate\ncompare_path_costs/compare_path_costs_fuzzily to allow the compiler to\nauto-vectorize. Might not be worth caring...\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 1 Nov 2019 10:34:10 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Zhenghua Lyu <zlv@pivotal.io> writes:\n>> I think a more robust way to disable forbidden plan types would be to\n>> handle the disabling in add_path(). Instead of having a high disable cost\n>> on the Path itself, the comparison add_path() would always consider\n>> disabled paths as more expensive than others, regardless of the cost.\n\nGetting rid of disable_cost would be a nice thing to do, but I would\nrather not do it by adding still more complexity to add_path(), not\nto mention having to bloat Paths with a separate \"disabled\" marker.\n\nThe idea that I've been thinking about is to not generate disabled\nPaths in the first place, thus not only fixing the problem but saving\nsome cycles. While this seems easy enough for \"optional\" paths,\nwe have to reserve the ability to generate certain path types regardless,\nif there's no other way to implement the query. This is a bit of a\nstumbling block :-(. At the base relation level, we could do something\nlike generating seqscan last, and only if no other path has been\nsuccessfully generated. But I'm not sure how to scale that up to\njoins. In particular, imagine that we consider joining A to B, and\nfind that the only way is a nestloop, so we generate a nestloop join\ndespite that being nominally disabled. The next join level would\nthen see that as an available path, and it might decide that\n((A nestjoin B) join C) is the cheapest choice, even though there\nmight have been a way to do, say, ((A join C) join B) with no use of\nnestloops. Users would find this surprising.\n\nMaybe the only way to do this is a separate number-of-uses-of-\ndisabled-plan-types cost figure in Paths, but I still don't want\nto go there. The number of cases where disable_cost's shortcomings\nreally matter is too small to justify that, IMHO.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 02 Nov 2019 10:57:33 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Tomas Vondra <tomas.vondra@2ndquadrant.com> writes:\n> On Fri, Nov 01, 2019 at 09:30:52AM -0700, Jim Finnerty wrote:\n>> re: coping with adding disable_cost more than once\n>> \n>> Another option would be to have a 2-part Cost structure. If disable_cost is\n>> ever added to the Cost, then you set a flag recording this. If any plans\n>> exist that have no disable_costs added to them, then the planner chooses the\n>> minimum cost among those, otherwise you choose the minimum cost path.\n\n> Yeah, I agree having is_disabled flag, and treat all paths with 'true'\n> as more expensive than paths with 'false' (and when both paths have the\n> same value then actually compare the cost) is probably the way forward.\n\nIt would have to be a count, not a boolean --- for example, you want to\nprefer a path that uses one disabled SeqScan over a path that uses two.\n\nI'm with Andres in being pretty worried about the extra burden imposed\non add_path comparisons.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 02 Nov 2019 11:04:58 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"As a proof of concept, I hacked around a bit today to re-purpose one of the\nbits of the Cost structure to mean \"is_disabled\" so that we can distinguish\n'disabled' from 'non-disabled' paths without making the Cost structure any\nbigger. In fact, it's still a valid double. The obvious choice would have\nbeen to re-purpose the sign bit, but I've had occasion to exploit negative\ncosts before so for this POC I used the high-order bit of the fractional\nbits of the double. (see Wikipedia for double precision floating point for\nthe layout).\n\nThe idea is to set a special bit when disable_cost is added to a cost. \nDedicating multiple bits instead of just 1 would be easily done, but as it\nis we can accumulate many disable_costs without overflowing, so just\ncomparing the cost suffices.\n\nThe patch is not fully debugged and fails on a couple of tests in the serial\ntest suite. It seems to fail on Cartesian products, and maybe in one other\nnon-CP case. I wasn't able to debug it before the day came to an end.\n\nIn one place the core code subtracts off the disable_cost. I left the\n\"disabled\" bit set in this case, which might be wrong.\n\nI don't see an option to attach the patch as an attachment, so here is the\npatch inline (it is based on PG11). The more interesting part is in a small\nnumber of lines in costsize.c. Other changes just add functions that assign\na disable_cost and set the bit, or that compare costs such that a\nnon-disabled cost always compares less than a disabled cost.\n\n------------------\n\ndiff --git a/src/backend/optimizer/path/costsize.c\nb/src/backend/optimizer/path/costsize.c\nindex 4e86458672..3718639330 100644\n--- a/src/backend/optimizer/path/costsize.c\n+++ b/src/backend/optimizer/path/costsize.c\n@@ -123,6 +123,8 @@ double\t\tparallel_setup_cost =\nDEFAULT_PARALLEL_SETUP_COST;\n int\t\t\teffective_cache_size = DEFAULT_EFFECTIVE_CACHE_SIZE;\n \n Cost\t\tdisable_cost = 1.0e10;\n+uint64 disabled_mask = 0x8000000000000;\n+#define IS_DISABLED(cost) (((uint64) cost) & disabled_mask)\n \n int\t\t\tmax_parallel_workers_per_gather = 2;\n \n@@ -205,6 +207,53 @@ clamp_row_est(double nrows)\n \treturn nrows;\n }\n \n+Cost\n+add_cost(Cost cost, Cost delta_cost)\n+{\n+\tuint64 mask = (delta_cost == disable_cost) ? disabled_mask : 0;\n+\tCost max_cost = disabled_mask - disable_cost;\n+\t\n+\tif (cost + delta_cost < max_cost)\n+\t\treturn ((Cost) ((uint64)(cost + delta_cost) | mask));\n+\telse\n+\t\treturn ((Cost) ((uint64)(max_cost) | mask));\n+}\n+\n+bool\n+is_lower_cost(Cost cost1, Cost cost2)\n+{\n+\tif ((uint64)cost1 & disabled_mask && !((uint64)cost2 & disabled_mask))\n+\t\treturn false;\n+\t\n+\tif (!((uint64)cost1 & disabled_mask) && (uint64)cost2 & disabled_mask)\n+\t\treturn true;\n+\t\n+\treturn (cost1 < cost2);\n+}\n+\n+bool\n+is_greater_cost(Cost cost1, Cost cost2)\n+{\n+\tif ((uint64)cost1 & disabled_mask && !((uint64)cost2 & disabled_mask))\n+\t\treturn true;\n+\t\n+\tif (!((uint64)cost1 & disabled_mask) && (uint64)cost2 & disabled_mask)\n+\t\treturn false;\n+\t\n+\treturn (cost1 > cost2);\n+}\n+\n+bool\n+is_geq_cost(Cost cost1, Cost cost2)\n+{\n+\tif ((uint64)cost1 & disabled_mask && !((uint64)cost2 & disabled_mask))\n+\t\treturn true;\n+\t\n+\tif (!((uint64)cost1 & disabled_mask) && (uint64)cost2 & disabled_mask)\n+\t\treturn false;\n+\t\n+\treturn (cost1 >= cost2);\n+}\n \n /*\n * cost_seqscan\n@@ -235,7 +284,7 @@ cost_seqscan(Path *path, PlannerInfo *root,\n \t\tpath->rows = baserel->rows;\n \n \tif (!enable_seqscan)\n-\t\tstartup_cost += disable_cost;\n+\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n \n \t/* fetch estimated page cost for tablespace containing table */\n \tget_tablespace_page_costs(baserel->reltablespace,\n@@ -424,7 +473,7 @@ cost_gather_merge(GatherMergePath *path, PlannerInfo\n*root,\n \t\tpath->path.rows = rel->rows;\n \n \tif (!enable_gathermerge)\n-\t\tstartup_cost += disable_cost;\n+\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n \n \t/*\n \t * Add one to the number of workers to account for the leader. This might\n@@ -538,7 +587,7 @@ cost_index(IndexPath *path, PlannerInfo *root, double\nloop_count,\n \t}\n \n \tif (!enable_indexscan)\n-\t\tstartup_cost += disable_cost;\n+\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n \t/* we don't need to check enable_indexonlyscan; indxpath.c does that */\n \n \t/*\n@@ -976,7 +1025,7 @@ cost_bitmap_heap_scan(Path *path, PlannerInfo *root,\nRelOptInfo *baserel,\n \t\tpath->rows = baserel->rows;\n \n \tif (!enable_bitmapscan)\n-\t\tstartup_cost += disable_cost;\n+\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n \n \tpages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,\n \t\t\t\t\t\t\t\t\t\t loop_count, &indexTotalCost,\n@@ -1242,10 +1291,10 @@ cost_tidscan(Path *path, PlannerInfo *root,\n \tif (isCurrentOf)\n \t{\n \t\tAssert(baserel->baserestrictcost.startup >= disable_cost);\n-\t\tstartup_cost -= disable_cost;\n+\t\tstartup_cost -= disable_cost; /* but do not un-set the disabled mark */\n \t}\n \telse if (!enable_tidscan)\n-\t\tstartup_cost += disable_cost;\n+\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n \n \t/*\n \t * The TID qual expressions will be computed once, any other baserestrict\n@@ -1676,7 +1725,7 @@ cost_sort(Path *path, PlannerInfo *root,\n \tlong\t\tsort_mem_bytes = sort_mem * 1024L;\n \n \tif (!enable_sort)\n-\t\tstartup_cost += disable_cost;\n+\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n \n \tpath->rows = tuples;\n \n@@ -2121,8 +2170,8 @@ cost_agg(Path *path, PlannerInfo *root,\n \t\ttotal_cost = input_total_cost;\n \t\tif (aggstrategy == AGG_MIXED && !enable_hashagg)\n \t\t{\n-\t\t\tstartup_cost += disable_cost;\n-\t\t\ttotal_cost += disable_cost;\n+\t\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n+\t\t\ttotal_cost = add_cost(total_cost, disable_cost);\n \t\t}\n \t\t/* calcs phrased this way to match HASHED case, see note above */\n \t\ttotal_cost += aggcosts->transCost.startup;\n@@ -2137,7 +2186,7 @@ cost_agg(Path *path, PlannerInfo *root,\n \t\t/* must be AGG_HASHED */\n \t\tstartup_cost = input_total_cost;\n \t\tif (!enable_hashagg)\n-\t\t\tstartup_cost += disable_cost;\n+\t\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n \t\tstartup_cost += aggcosts->transCost.startup;\n \t\tstartup_cost += aggcosts->transCost.per_tuple * input_tuples;\n \t\tstartup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;\n@@ -2436,7 +2485,7 @@ final_cost_nestloop(PlannerInfo *root, NestPath *path,\n \t * disabled, which doesn't seem like the way to bet.\n \t */\n \tif (!enable_nestloop)\n-\t\tstartup_cost += disable_cost;\n+\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n \n \t/* cost of inner-relation source data (we already dealt with outer rel) */\n \n@@ -2882,7 +2931,7 @@ final_cost_mergejoin(PlannerInfo *root, MergePath\n*path,\n \t * disabled, which doesn't seem like the way to bet.\n \t */\n \tif (!enable_mergejoin)\n-\t\tstartup_cost += disable_cost;\n+\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n \n \t/*\n \t * Compute cost of the mergequals and qpquals (other restriction clauses)\n@@ -3312,7 +3361,7 @@ final_cost_hashjoin(PlannerInfo *root, HashPath *path,\n \t * disabled, which doesn't seem like the way to bet.\n \t */\n \tif (!enable_hashjoin)\n-\t\tstartup_cost += disable_cost;\n+\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n \n \t/* mark the path with estimated # of batches */\n \tpath->num_batches = numbatches;\n@@ -3410,7 +3459,7 @@ final_cost_hashjoin(PlannerInfo *root, HashPath *path,\n \tif (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),\n \t\t\t\t\t\t inner_path->pathtarget->width) >\n \t\t(work_mem * 1024L))\n-\t\tstartup_cost += disable_cost;\n+\t\tstartup_cost = add_cost(startup_cost, disable_cost);\n \n \t/*\n \t * Compute cost of the hashquals and qpquals (other restriction clauses)\n@@ -3930,7 +3979,7 @@ cost_qual_eval_walker(Node *node,\ncost_qual_eval_context *context)\n \telse if (IsA(node, CurrentOfExpr))\n \t{\n \t\t/* Report high cost to prevent selection of anything but TID scan */\n-\t\tcontext->total.startup += disable_cost;\n+\t\tcontext->total.startup = add_cost(context->total.startup, disable_cost);\n \t}\n \telse if (IsA(node, SubLink))\n \t{\ndiff --git a/src/backend/optimizer/util/pathnode.c\nb/src/backend/optimizer/util/pathnode.c\nindex 4736d84a83..fd746a06bc 100644\n--- a/src/backend/optimizer/util/pathnode.c\n+++ b/src/backend/optimizer/util/pathnode.c\n@@ -72,33 +72,33 @@ compare_path_costs(Path *path1, Path *path2,\nCostSelector criterion)\n {\n \tif (criterion == STARTUP_COST)\n \t{\n-\t\tif (path1->startup_cost < path2->startup_cost)\n+\t\tif (is_lower_cost(path1->startup_cost, path2->startup_cost))\n \t\t\treturn -1;\n-\t\tif (path1->startup_cost > path2->startup_cost)\n+\t\tif (is_greater_cost(path1->startup_cost, path2->startup_cost))\n \t\t\treturn +1;\n \n \t\t/*\n \t\t * If paths have the same startup cost (not at all unlikely), order\n \t\t * them by total cost.\n \t\t */\n-\t\tif (path1->total_cost < path2->total_cost)\n+\t\tif (is_lower_cost(path1->total_cost, path2->total_cost))\n \t\t\treturn -1;\n-\t\tif (path1->total_cost > path2->total_cost)\n+\t\tif (is_greater_cost(path1->total_cost, path2->total_cost))\n \t\t\treturn +1;\n \t}\n \telse\n \t{\n-\t\tif (path1->total_cost < path2->total_cost)\n+\t\tif (is_lower_cost(path1->total_cost, path2->total_cost))\n \t\t\treturn -1;\n-\t\tif (path1->total_cost > path2->total_cost)\n+\t\tif (is_greater_cost(path1->total_cost, path2->total_cost))\n \t\t\treturn +1;\n \n \t\t/*\n \t\t * If paths have the same total cost, order them by startup cost.\n \t\t */\n-\t\tif (path1->startup_cost < path2->startup_cost)\n+\t\tif (is_lower_cost(path1->startup_cost, path2->startup_cost))\n \t\t\treturn -1;\n-\t\tif (path1->startup_cost > path2->startup_cost)\n+\t\tif (is_greater_cost(path1->startup_cost, path2->startup_cost))\n \t\t\treturn +1;\n \t}\n \treturn 0;\n@@ -126,9 +126,9 @@ compare_fractional_path_costs(Path *path1, Path *path2,\n \t\tfraction * (path1->total_cost - path1->startup_cost);\n \tcost2 = path2->startup_cost +\n \t\tfraction * (path2->total_cost - path2->startup_cost);\n-\tif (cost1 < cost2)\n+\tif (is_lower_cost(cost1, cost2))\n \t\treturn -1;\n-\tif (cost1 > cost2)\n+\tif (is_greater_cost(cost1, cost2))\n \t\treturn +1;\n \treturn 0;\n }\n@@ -172,11 +172,11 @@ compare_path_costs_fuzzily(Path *path1, Path *path2,\ndouble fuzz_factor)\n \t * Check total cost first since it's more likely to be different; many\n \t * paths have zero startup cost.\n \t */\n-\tif (path1->total_cost > path2->total_cost * fuzz_factor)\n+\tif (is_greater_cost(path1->total_cost, path2->total_cost * fuzz_factor))\n \t{\n \t\t/* path1 fuzzily worse on total cost */\n \t\tif (CONSIDER_PATH_STARTUP_COST(path1) &&\n-\t\t\tpath2->startup_cost > path1->startup_cost * fuzz_factor)\n+\t\t\tis_greater_cost(path2->startup_cost, path1->startup_cost * fuzz_factor))\n \t\t{\n \t\t\t/* ... but path2 fuzzily worse on startup, so DIFFERENT */\n \t\t\treturn COSTS_DIFFERENT;\n@@ -184,11 +184,11 @@ compare_path_costs_fuzzily(Path *path1, Path *path2,\ndouble fuzz_factor)\n \t\t/* else path2 dominates */\n \t\treturn COSTS_BETTER2;\n \t}\n-\tif (path2->total_cost > path1->total_cost * fuzz_factor)\n+\tif (is_greater_cost(path2->total_cost, path1->total_cost * fuzz_factor))\n \t{\n \t\t/* path2 fuzzily worse on total cost */\n \t\tif (CONSIDER_PATH_STARTUP_COST(path2) &&\n-\t\t\tpath1->startup_cost > path2->startup_cost * fuzz_factor)\n+\t\t\tis_greater_cost(path1->startup_cost, path2->startup_cost * fuzz_factor))\n \t\t{\n \t\t\t/* ... but path1 fuzzily worse on startup, so DIFFERENT */\n \t\t\treturn COSTS_DIFFERENT;\n@@ -197,12 +197,12 @@ compare_path_costs_fuzzily(Path *path1, Path *path2,\ndouble fuzz_factor)\n \t\treturn COSTS_BETTER1;\n \t}\n \t/* fuzzily the same on total cost ... */\n-\tif (path1->startup_cost > path2->startup_cost * fuzz_factor)\n+\tif (is_greater_cost(path1->startup_cost, path2->startup_cost *\nfuzz_factor))\n \t{\n \t\t/* ... but path1 fuzzily worse on startup, so path2 wins */\n \t\treturn COSTS_BETTER2;\n \t}\n-\tif (path2->startup_cost > path1->startup_cost * fuzz_factor)\n+\tif (is_greater_cost(path2->startup_cost, path1->startup_cost *\nfuzz_factor))\n \t{\n \t\t/* ... but path2 fuzzily worse on startup, so path1 wins */\n \t\treturn COSTS_BETTER1;\n@@ -605,7 +605,7 @@ add_path(RelOptInfo *parent_rel, Path *new_path)\n \t\telse\n \t\t{\n \t\t\t/* new belongs after this old path if it has cost >= old's */\n-\t\t\tif (new_path->total_cost >= old_path->total_cost)\n+\t\t\tif (is_geq_cost(new_path->total_cost, old_path->total_cost))\n \t\t\t\tinsert_after = p1;\n \t\t\t/* p1_prev advances */\n \t\t\tp1_prev = p1;\n@@ -681,7 +681,7 @@ add_path_precheck(RelOptInfo *parent_rel,\n \t\t *\n \t\t * Cost comparisons here should match compare_path_costs_fuzzily.\n \t\t */\n-\t\tif (total_cost > old_path->total_cost * STD_FUZZ_FACTOR)\n+\t\tif (is_greater_cost(total_cost, old_path->total_cost * STD_FUZZ_FACTOR))\n \t\t{\n \t\t\t/* new path can win on startup cost only if consider_startup */\n \t\t\tif (startup_cost > old_path->startup_cost * STD_FUZZ_FACTOR ||\n@@ -796,14 +796,14 @@ add_partial_path(RelOptInfo *parent_rel, Path\n*new_path)\n \t\t/* Unless pathkeys are incompable, keep just one of the two paths. */\n \t\tif (keyscmp != PATHKEYS_DIFFERENT)\n \t\t{\n-\t\t\tif (new_path->total_cost > old_path->total_cost * STD_FUZZ_FACTOR)\n+\t\t\tif (is_greater_cost(new_path->total_cost, old_path->total_cost *\nSTD_FUZZ_FACTOR))\n \t\t\t{\n \t\t\t\t/* New path costs more; keep it only if pathkeys are better. */\n \t\t\t\tif (keyscmp != PATHKEYS_BETTER1)\n \t\t\t\t\taccept_new = false;\n \t\t\t}\n-\t\t\telse if (old_path->total_cost > new_path->total_cost\n-\t\t\t\t\t * STD_FUZZ_FACTOR)\n+\t\t\telse if (is_greater_cost(old_path->total_cost, new_path->total_cost\n+\t\t\t\t\t\t\t\t\t * STD_FUZZ_FACTOR))\n \t\t\t{\n \t\t\t\t/* Old path costs more; keep it only if pathkeys are better. */\n \t\t\t\tif (keyscmp != PATHKEYS_BETTER2)\n@@ -819,7 +819,7 @@ add_partial_path(RelOptInfo *parent_rel, Path *new_path)\n \t\t\t\t/* Costs are about the same, old path has better pathkeys. */\n \t\t\t\taccept_new = false;\n \t\t\t}\n-\t\t\telse if (old_path->total_cost > new_path->total_cost * 1.0000000001)\n+\t\t\telse if (is_greater_cost(old_path->total_cost, new_path->total_cost *\n1.0000000001))\n \t\t\t{\n \t\t\t\t/* Pathkeys are the same, and the old path costs more. */\n \t\t\t\tremove_old = true;\n@@ -847,7 +847,7 @@ add_partial_path(RelOptInfo *parent_rel, Path *new_path)\n \t\telse\n \t\t{\n \t\t\t/* new belongs after this old path if it has cost >= old's */\n-\t\t\tif (new_path->total_cost >= old_path->total_cost)\n+\t\t\tif (is_geq_cost(new_path->total_cost, old_path->total_cost))\n \t\t\t\tinsert_after = p1;\n \t\t\t/* p1_prev advances */\n \t\t\tp1_prev = p1;\n@@ -913,10 +913,10 @@ add_partial_path_precheck(RelOptInfo *parent_rel, Cost\ntotal_cost,\n \t\tkeyscmp = compare_pathkeys(pathkeys, old_path->pathkeys);\n \t\tif (keyscmp != PATHKEYS_DIFFERENT)\n \t\t{\n-\t\t\tif (total_cost > old_path->total_cost * STD_FUZZ_FACTOR &&\n+\t\t\tif (is_greater_cost(total_cost, old_path->total_cost * STD_FUZZ_FACTOR)\n&&\n \t\t\t\tkeyscmp != PATHKEYS_BETTER1)\n \t\t\t\treturn false;\n-\t\t\tif (old_path->total_cost > total_cost * STD_FUZZ_FACTOR &&\n+\t\t\tif (is_greater_cost(old_path->total_cost, total_cost * STD_FUZZ_FACTOR)\n&&\n \t\t\t\tkeyscmp != PATHKEYS_BETTER2)\n \t\t\t\treturn true;\n \t\t}\n@@ -1697,7 +1697,7 @@ create_unique_path(PlannerInfo *root, RelOptInfo *rel,\nPath *subpath,\n \n \tif (sjinfo->semi_can_btree && sjinfo->semi_can_hash)\n \t{\n-\t\tif (agg_path.total_cost < sort_path.total_cost)\n+\t\tif (is_lower_cost(agg_path.total_cost, sort_path.total_cost))\n \t\t\tpathnode->umethod = UNIQUE_PATH_HASH;\n \t\telse\n \t\t\tpathnode->umethod = UNIQUE_PATH_SORT;\ndiff --git a/src/backend/utils/cache/relcache.c\nb/src/backend/utils/cache/relcache.c\nindex 78f3b99a76..c261a9d790 100644\n--- a/src/backend/utils/cache/relcache.c\n+++ b/src/backend/utils/cache/relcache.c\n@@ -5076,8 +5076,8 @@ IsProjectionFunctionalIndex(Relation index)\n \t\t * when values differ because the expression is recalculated when\n \t\t * inserting a new index entry for the changed value.\n \t\t */\n-\t\tif ((index_expr_cost.startup + index_expr_cost.per_tuple) >\n-\t\t\tHEURISTIC_MAX_HOT_RECHECK_EXPR_COST)\n+\t\tif (is_greater_cost((index_expr_cost.startup +\nindex_expr_cost.per_tuple),\n+\t\t\t\t\t\t\tHEURISTIC_MAX_HOT_RECHECK_EXPR_COST))\n \t\t\tis_projection = false;\n \n \t\ttuple = SearchSysCache1(RELOID,\nObjectIdGetDatum(RelationGetRelid(index)));\ndiff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h\nindex 9159f2bab1..c01d08eae5 100644\n--- a/src/include/optimizer/cost.h\n+++ b/src/include/optimizer/cost.h\n@@ -251,6 +251,12 @@ extern PathTarget\n*set_pathtarget_cost_width(PlannerInfo *root, PathTarget *targ\n extern double compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel,\n \t\t\t\t\t Path *bitmapqual, int loop_count, Cost *cost, double *tuple);\n \n+extern Cost add_cost(Cost cost, Cost delta_cost);\n+extern bool is_lower_cost(Cost cost1, Cost cost2);\n+extern bool is_greater_cost(Cost cost1, Cost cost2);\n+extern bool is_geq_cost(Cost cost1, Cost cost2);\n+\n+\n /*\n * prototypes for clausesel.c\n *\t routines to compute clause selectivities\n\n\n\n\n\n-----\nJim Finnerty, AWS, Amazon Aurora PostgreSQL\n--\nSent from: https://www.postgresql-archive.org/PostgreSQL-hackers-f1928748.html\n\n\n","msg_date":"Tue, 10 Dec 2019 15:50:29 -0700 (MST)","msg_from":"Jim Finnerty <jfinnert@amazon.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, 2019-12-10 at 15:50 -0700, Jim Finnerty wrote:\n> As a proof of concept, I hacked around a bit today to re-purpose one of the\n> bits of the Cost structure to mean \"is_disabled\" so that we can distinguish\n> 'disabled' from 'non-disabled' paths without making the Cost structure any\n> bigger. In fact, it's still a valid double. The obvious choice would have\n> been to re-purpose the sign bit, but I've had occasion to exploit negative\n> costs before so for this POC I used the high-order bit of the fractional\n> bits of the double. (see Wikipedia for double precision floating point for\n> the layout).\n> \n> The idea is to set a special bit when disable_cost is added to a cost. \n> Dedicating multiple bits instead of just 1 would be easily done, but as it\n> is we can accumulate many disable_costs without overflowing, so just\n> comparing the cost suffices.\n\nDoesn't that rely on a specific implementation of double precision (IEEE)?\nI thought that we don't want to limit ourselves to platforms with IEEE floats.\n\nYours,\nLaurenz Albe\n\n\n\n","msg_date":"Wed, 11 Dec 2019 07:23:51 +0100","msg_from":"Laurenz Albe <laurenz.albe@cybertec.at>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, 11 Dec 2019 at 01:24, Laurenz Albe <laurenz.albe@cybertec.at> wrote:\n>\n> On Tue, 2019-12-10 at 15:50 -0700, Jim Finnerty wrote:\n> > As a proof of concept, I hacked around a bit today to re-purpose one of the\n> > bits of the Cost structure to mean \"is_disabled\" so that we can distinguish\n> > 'disabled' from 'non-disabled' paths without making the Cost structure any\n> > bigger. In fact, it's still a valid double. The obvious choice would have\n> > been to re-purpose the sign bit, but I've had occasion to exploit negative\n> > costs before so for this POC I used the high-order bit of the fractional\n> > bits of the double. (see Wikipedia for double precision floating point for\n> > the layout).\n> >\n> > The idea is to set a special bit when disable_cost is added to a cost.\n> > Dedicating multiple bits instead of just 1 would be easily done, but as it\n> > is we can accumulate many disable_costs without overflowing, so just\n> > comparing the cost suffices.\n>\n> Doesn't that rely on a specific implementation of double precision (IEEE)?\n> I thought that we don't want to limit ourselves to platforms with IEEE floats.\n\nWe could always implement it again in another format....\n\nHowever, I wouldn't have expected to be bit twiddling. I would have\nexpected to use standard functions like ldexp to do this. In fact I\nthink if you use the high bit of the exponent you could do it entirely\nusing ldexp and regular double comparisons (with fabs).\n\nIe, to set the bit you set cost = ldexp(cost, __DBL_MAX_EXP__/2). And\nto check for the bit being set you compare ilogb(cost,\n__DBL_MAX_EXP__/2). Hm. that doesn't handle if the cost is already < 1\nin which case I guess you would have to set it to 1 first. Or reserve\nthe two high bits of the cost so you can represent disabled values\nthat had negative exponents before being disabled.\n\nI wonder if it wouldn't be a lot cleaner and more flexible to just go\nwith a plain float for Cost and use the other 32 bits for counters and\nbitmasks and still be ahead of the game. A double can store 2^1024 but\na float 2^128 which still feels like it should be more than enough to\nstore the kinds of costs plans have without the disabled costs. 2^128\nmilliseconds is still 10^28 years which is an awfully expensive\nquery....\n\n-- \ngreg\n\n\n","msg_date":"Thu, 12 Dec 2019 15:42:37 -0500","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, Dec 11, 2019 at 7:24 PM Laurenz Albe <laurenz.albe@cybertec.at> wrote:\n> Doesn't that rely on a specific implementation of double precision (IEEE)?\n> I thought that we don't want to limit ourselves to platforms with IEEE floats.\n\nJust by the way, you might want to read the second last paragraph of\nthe commit message for 02ddd499. The dream is over, we're never going\nto run on Vax.\n\n\n","msg_date":"Fri, 13 Dec 2019 15:59:32 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Thomas Munro <thomas.munro@gmail.com> writes:\n> On Wed, Dec 11, 2019 at 7:24 PM Laurenz Albe <laurenz.albe@cybertec.at> wrote:\n>> Doesn't that rely on a specific implementation of double precision (IEEE)?\n>> I thought that we don't want to limit ourselves to platforms with IEEE floats.\n\n> Just by the way, you might want to read the second last paragraph of\n> the commit message for 02ddd499. The dream is over, we're never going\n> to run on Vax.\n\nStill, the proposed hack is doubling down on IEEE dependency in a way\nthat I quite dislike, in that (a) it doesn't just read float values\nbut generates new ones (and assumes that the hardware/libc will react in\na predictable way to them), (b) in a part of the code that has no damn\nbusiness having close dependencies on float format, and (c) for a gain\nfar smaller than what we got from the Ryu code.\n\nWe have had prior discussions about whether 02ddd499 justifies adding\nmore IEEE dependencies elsewhere. I don't think it does. IEEE 754\nis not the last word that will ever be said on floating-point arithmetic,\nany more than x86_64 is the last CPU architecture that anyone will ever\ncare about. We should keep our dependencies on it well circumscribed.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 15 Dec 2019 14:54:03 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"I think this would be ready to abstract away behind a few functions that\ncould always be replaced by something else later...\n\n\nHowever on further thought I really think just using a 32-bit float and 32\nbits of other bitmaps or counters would be a better approach.\n\n\nOn Sun., Dec. 15, 2019, 14:54 Tom Lane, <tgl@sss.pgh.pa.us> wrote:\n\n> Thomas Munro <thomas.munro@gmail.com> writes:\n> > On Wed, Dec 11, 2019 at 7:24 PM Laurenz Albe <laurenz.albe@cybertec.at>\n> wrote:\n> >> Doesn't that rely on a specific implementation of double precision\n> (IEEE)?\n> >> I thought that we don't want to limit ourselves to platforms with IEEE\n> floats.\n>\n> > Just by the way, you might want to read the second last paragraph of\n> > the commit message for 02ddd499. The dream is over, we're never going\n> > to run on Vax.\n>\n> Still, the proposed hack is doubling down on IEEE dependency in a way\n> that I quite dislike, in that (a) it doesn't just read float values\n> but generates new ones (and assumes that the hardware/libc will react in\n> a predictable way to them), (b) in a part of the code that has no damn\n> business having close dependencies on float format, and (c) for a gain\n> far smaller than what we got from the Ryu code.\n>\n> We have had prior discussions about whether 02ddd499 justifies adding\n> more IEEE dependencies elsewhere. I don't think it does. IEEE 754\n> is not the last word that will ever be said on floating-point arithmetic,\n> any more than x86_64 is the last CPU architecture that anyone will ever\n> care about. We should keep our dependencies on it well circumscribed.\n>\n> regards, tom lane\n>\n>\n>\n\nI think this would be ready to abstract away behind a few functions that could always be replaced by something else later...However on further thought I really think just using a 32-bit float and 32 bits of other bitmaps or counters would be a better approach. On Sun., Dec. 15, 2019, 14:54 Tom Lane, <tgl@sss.pgh.pa.us> wrote:Thomas Munro <thomas.munro@gmail.com> writes:\n> On Wed, Dec 11, 2019 at 7:24 PM Laurenz Albe <laurenz.albe@cybertec.at> wrote:\n>> Doesn't that rely on a specific implementation of double precision (IEEE)?\n>> I thought that we don't want to limit ourselves to platforms with IEEE floats.\n\n> Just by the way, you might want to read the second last paragraph of\n> the commit message for 02ddd499. The dream is over, we're never going\n> to run on Vax.\n\nStill, the proposed hack is doubling down on IEEE dependency in a way\nthat I quite dislike, in that (a) it doesn't just read float values\nbut generates new ones (and assumes that the hardware/libc will react in\na predictable way to them), (b) in a part of the code that has no damn\nbusiness having close dependencies on float format, and (c) for a gain\nfar smaller than what we got from the Ryu code.\n\nWe have had prior discussions about whether 02ddd499 justifies adding\nmore IEEE dependencies elsewhere. I don't think it does. IEEE 754\nis not the last word that will ever be said on floating-point arithmetic,\nany more than x86_64 is the last CPU architecture that anyone will ever\ncare about. We should keep our dependencies on it well circumscribed.\n\n regards, tom lane","msg_date":"Sun, 15 Dec 2019 15:45:31 -0500","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Hi hackers,\r\n\r\nI have write an initial patch to retire the disable_cost GUC, which labeled a flag on the Path struct instead of adding up a big cost which is hard to estimate. Though it involved in tons of plan changes in regression tests, I have tested on some simple test cases such as eagerly generate a two-stage agg plans and it worked. Could someone help to review?\r\n\r\n\r\nregards,\r\n\r\nJian\r\n________________________________\r\nFrom: Euler Taveira <euler@timbira.com.br>\r\nSent: Friday, November 1, 2019 22:48\r\nTo: Zhenghua Lyu <zlyu@vmware.com>\r\nCc: PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>\r\nSubject: Re: On disable_cost\r\n\r\n!! External Email\r\n\r\nEm sex, 1 de nov de 2019 às 03:42, Zhenghua Lyu <zlv@pivotal.io> escreveu:\r\n>\r\n> My issue: I did some spikes and tests on TPCDS 1TB Bytes data. For query 104, it generates\r\n> nestloop join even with enable_nestloop set off. And the final plan's total cost is very huge (about 1e24). But If I enlarge the disable_cost to 1e30, then, planner will generate hash join.\r\n>\r\n> So I guess that disable_cost is not large enough for huge amount of data.\r\n>\r\n> It is tricky to set disable_cost a huge number. Can we come up with better solution?\r\n>\r\nIsn't it a case for a GUC disable_cost? As Thomas suggested, DBL_MAX\r\nupper limit should be sufficient.\r\n\r\n> The following thoughts are from Heikki:\r\n>>\r\n>> Aside from not having a large enough disable cost, there's also the fact that the high cost might affect the rest of the plan, if we have to use a plan type that's disabled. For example, if a table doesn't have any indexes, but enable_seqscan is off, we might put the unavoidable Seq Scan on different side of a join than we we would with enable_seqscan=on, because of the high cost estimate.\r\n>\r\n>\r\n>>\r\n>> I think a more robust way to disable forbidden plan types would be to handle the disabling in add_path(). Instead of having a high disable cost on the Path itself, the comparison add_path() would always consider disabled paths as more expensive than others, regardless of the cost.\r\n>\r\nI'm afraid it is not as cheap as using diable_cost as a node cost. Are\r\nyou proposing to add a new boolean variable in Path struct to handle\r\nthose cases in compare_path_costs_fuzzily?\r\n\r\n\r\n--\r\n Euler Taveira Timbira -\r\nhttps://nam04.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.timbira.com.br%2F&data=05%7C01%7Cgjian%40vmware.com%7C12a30b2852dd4651667608db9401d056%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C0%7C0%7C638266507757076648%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=v54JhsW8FX4mSmjgt2yP59t7xtv1mZvC%2BBhtKrfp%2FBY%3D&reserved=0<http://www.timbira.com.br/>\r\n PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento\r\n\r\n\r\n\r\n\r\n\r\n!! External Email: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the sender.","msg_date":"Thu, 3 Aug 2023 09:21:39 +0000","msg_from":"Jian Guo <gjian@vmware.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Thu, Aug 3, 2023 at 5:22 AM Jian Guo <gjian@vmware.com> wrote:\n> I have write an initial patch to retire the disable_cost GUC, which labeled a flag on the Path struct instead of adding up a big cost which is hard to estimate. Though it involved in tons of plan changes in regression tests, I have tested on some simple test cases such as eagerly generate a two-stage agg plans and it worked. Could someone help to review?\n\nI took a look at this patch today. I believe that overall this may\nwell be an approach worth pursuing. However, more work is going to be\nneeded. Here are some comments:\n\n1. You stated that it changes lots of plans in the regression tests,\nbut you haven't provided any sort of analysis of why those plans\nchanged. I'm kind of surprised that there would be \"tons\" of plan\nchanges. You (or someone) should look into why that's happening.\n\n2. The change to compare_path_costs_fuzzily() seems incorrect to me.\nWhen path1->is_disabled && path2->is_disabled, costs should be\ncompared just as they are when neither path is disabled. Instead, the\npatch treats any two such paths as having equal cost. That seems\ncatastrophically bad. Maybe it accounts for some of those plan\nchanges, although that would only be true if those plans were created\nwhile using some disabling GUC.\n\n3. Instead of adding is_disabled at the end of the Path structure, I\nsuggest adding it between param_info and parallel_aware. I think if\nyou do that, the space used by the new field will use up padding bytes\nthat are currently included in the struct, instead of making it\nlonger.\n\n4. A critical issue for any patch of this type is performance. This\nconcern was raised earlier on this thread, but your path doesn't\naddress it. There's no performance analysis or benchmarking included\nin your email. One idea that I have is to write the cost-comparison\ntest like this:\n\nif (unlikely(path1->is_disabled || path2->is_disabled))\n{\n if (!path1->is_disabled)\n return COSTS_BETTER1;\n if (!path2->is_disabled)\n return COSTS_BETTER2;\n /* if both disabled, fall through */\n}\n\nI'm not sure that would be enough to prevent the patch from adding\nnoticeably to the cost of path comparison, but maybe it would help.\n\n5. The patch changes only compare_path_costs_fuzzily() but I wonder\nwhether compare_path_costs() and compare_fractional_path_costs() need\nsimilar surgery. Whether they do or don't, there should likely be some\ncomments explaining the situation.\n\n6. In fact, the patch changes no comments at all, anywhere. I'm not\nsure how many comment changes a patch like this needs to make, but the\nanswer definitely isn't \"none\".\n\n7. The patch doesn't actually remove disable_cost. I guess it should.\n\n8. When you submit a patch, it's a good idea to also add it on\ncommitfest.postgresql.org. It doesn't look like that was done in this\ncase.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 12 Mar 2024 10:27:48 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Thu, Aug 3, 2023 at 5:22 AM Jian Guo <gjian@vmware.com> wrote:\n>> I have write an initial patch to retire the disable_cost GUC, which labeled a flag on the Path struct instead of adding up a big cost which is hard to estimate. Though it involved in tons of plan changes in regression tests, I have tested on some simple test cases such as eagerly generate a two-stage agg plans and it worked. Could someone help to review?\n\n> I took a look at this patch today. I believe that overall this may\n> well be an approach worth pursuing. However, more work is going to be\n> needed. Here are some comments:\n\n> 1. You stated that it changes lots of plans in the regression tests,\n> but you haven't provided any sort of analysis of why those plans\n> changed. I'm kind of surprised that there would be \"tons\" of plan\n> changes. You (or someone) should look into why that's happening.\n\nI've not read the patch, but given this description I would expect\nthere to be *zero* regression changes --- I don't think we have any\ntest cases that depend on disable_cost being finite. If there's more\nthan zero changes, that very likely indicates a bug in the patch.\nEven if there are places where the output legitimately changes, you\nneed to justify each one and make sure that you didn't invalidate the\nintent of that test case.\n\nBTW, having written that paragraph, I wonder if we couldn't get\nthe same end result with a nearly one-line change that consists of\nmaking disable_cost be IEEE infinity. Years ago we didn't want\nto rely on IEEE float semantics in this area, but nowadays I don't\nsee why we shouldn't.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 12 Mar 2024 13:32:58 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, Mar 12, 2024 at 1:32 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> BTW, having written that paragraph, I wonder if we couldn't get\n> the same end result with a nearly one-line change that consists of\n> making disable_cost be IEEE infinity. Years ago we didn't want\n> to rely on IEEE float semantics in this area, but nowadays I don't\n> see why we shouldn't.\n\nI don't think so, because I think that what will happen in that case\nis that we'll pick a completely random plan if we can't pick a plan\nthat avoids incurring disable_cost. Every plan that contains one\ndisabled node anywhere in the plan tree will look like it has exactly\nthe same cost as any other such plan.\n\nIMHO, this is actually one of the problems with disable_cost as it\nworks today. I think the semantics that we want are: if it's possible\nto pick a plan where nothing is disabled, then pick the cheapest such\nplan; if not, pick the cheapest plan overall. But treating\ndisable_cost doesn't really do that. It does the first part -- picking\nthe cheapest plan where nothing is disabled -- but it doesn't do the\nsecond part, because once you add disable_cost into the cost of some\nparticular plan node, it screws up the rest of the planning, because\nthe cost estimates for the disabled nodes have no bearing in reality.\nFast-start plans, for example, will look insanely good compared to\nwhat would be the case in normal planning (and we lean too much toward\nfast-start plans even normally).\n\n(I don't think we should care how MANY disabled nodes appear in a\nplan, particularly. This is a more arguable point. Is a plan with 1\ndisabled node and 10% more cost better or worse than a plan with 2\ndisabled nodes and 10% less cost? I'd argue that counting the number\nof disabled nodes isn't particularly meaningful.)\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 12 Mar 2024 14:01:36 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Tue, Mar 12, 2024 at 1:32 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> BTW, having written that paragraph, I wonder if we couldn't get\n>> the same end result with a nearly one-line change that consists of\n>> making disable_cost be IEEE infinity.\n\n> I don't think so, because I think that what will happen in that case\n> is that we'll pick a completely random plan if we can't pick a plan\n> that avoids incurring disable_cost. Every plan that contains one\n> disabled node anywhere in the plan tree will look like it has exactly\n> the same cost as any other such plan.\n\nGood point.\n\n> IMHO, this is actually one of the problems with disable_cost as it\n> works today.\n\nYeah. I keep thinking that the right solution is to not generate\ndisabled paths in the first place if there are any other ways to\nproduce the same relation. That has obvious order-of-operations\nproblems though, and I've not been able to make it work.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 12 Mar 2024 15:36:38 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, Mar 12, 2024 at 3:36 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Yeah. I keep thinking that the right solution is to not generate\n> disabled paths in the first place if there are any other ways to\n> produce the same relation. That has obvious order-of-operations\n> problems though, and I've not been able to make it work.\n\nI've expressed the same view in the past. It would be nice not to\nwaste planner effort on paths that we're just going to throw away, but\nI'm not entirely sure what you mean by \"obvious order-of-operations\nproblems.\"\n\nTo me, it seems like what we'd need is to be able to restart the whole\nplanner process if we run out of steam before we get done. For\nexample, suppose we're planning a 2-way join where index and\nindex-only scans are disabled, sorts are disabled, and nested loops\nand hash joins are disabled. There's no problem generating just the\nnon-disabled scan types at the baserel level, but when we reach the\njoin, we're going to find that the only non-disabled join type is a\nmerge join, and we're also going to find that we have no paths that\nprovide pre-sorted input, so we need to sort, which we're also not\nallowed to do. If we could give up at that point and restart planning,\ndisabling all of the plan-choice constraints and now creating all\npaths for each RelOptInfo, then everything would, I believe, be just\nfine. We'd end up needing neither disable_cost nor the mechanism\nproposed by this patch.\n\nBut in the absence of that, we need some way to privilege the\nnon-disabled paths over the disabled ones -- and I'd prefer to have\nsomething more principled than disable_cost, if we can work out the\ndetails.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 12 Mar 2024 15:54:47 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, 13 Mar 2024 at 08:55, Robert Haas <robertmhaas@gmail.com> wrote:\n> But in the absence of that, we need some way to privilege the\n> non-disabled paths over the disabled ones -- and I'd prefer to have\n> something more principled than disable_cost, if we can work out the\n> details.\n\nThe primary place I see issues with disabled_cost is caused by\nSTD_FUZZ_FACTOR. When you add 1.0e10 to a couple of modestly costly\npaths, it makes them appear fuzzily the same in cases where one could\nbe significantly cheaper than the other. If we were to bump up the\ndisable_cost it would make this problem worse.\n\nI think we do still need some way to pick the cheapest disabled path\nwhen there are no other options.\n\nOne way would be to set fuzz_factor to 1.0 when either of the paths\ncosts in compare_path_costs_fuzzily() is >= disable_cost.\nclamp_row_est() does cap row estimates at MAXIMUM_ROWCOUNT (1e100), so\nI think there is some value of disable_cost that could almost\ncertainly ensure we don't reach it because the path is truly expensive\nrather than disabled.\n\nSo maybe the fix could be to set disable_cost to something like\n1.0e110 and adjust compare_path_costs_fuzzily to not apply the\nfuzz_factor for paths >= disable_cost. However, I wonder if that\nrisks the costs going infinite after a couple of cartesian joins.\n\nDavid\n\n\n","msg_date":"Wed, 13 Mar 2024 09:55:22 +1300","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"David Rowley <dgrowleyml@gmail.com> writes:\n> So maybe the fix could be to set disable_cost to something like\n> 1.0e110 and adjust compare_path_costs_fuzzily to not apply the\n> fuzz_factor for paths >= disable_cost. However, I wonder if that\n> risks the costs going infinite after a couple of cartesian joins.\n\nPerhaps. It still does nothing for Robert's point that once we're\nforced into using a \"disabled\" plan type, it'd be better if the\ndisabled-ness didn't skew subsequent planning choices.\n\nOn the whole I agree that getting rid of disable_cost entirely\nwould be the way to go, if we can replace that with a separate\nboolean without driving up the cost of add_path too much.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 12 Mar 2024 17:18:09 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, Mar 12, 2024 at 4:55 PM David Rowley <dgrowleyml@gmail.com> wrote:\n> The primary place I see issues with disabled_cost is caused by\n> STD_FUZZ_FACTOR. When you add 1.0e10 to a couple of modestly costly\n> paths, it makes them appear fuzzily the same in cases where one could\n> be significantly cheaper than the other. If we were to bump up the\n> disable_cost it would make this problem worse.\n\nHmm, good point.\n\n> So maybe the fix could be to set disable_cost to something like\n> 1.0e110 and adjust compare_path_costs_fuzzily to not apply the\n> fuzz_factor for paths >= disable_cost. However, I wonder if that\n> risks the costs going infinite after a couple of cartesian joins.\n\nYeah, I think the disabled flag is a better answer if we can make it\nwork. No matter what value we pick for disable_cost, it's bound to\nskew the planning sometimes.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 13 Mar 2024 09:05:00 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, Mar 12, 2024 at 1:32 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > 1. You stated that it changes lots of plans in the regression tests,\n> > but you haven't provided any sort of analysis of why those plans\n> > changed. I'm kind of surprised that there would be \"tons\" of plan\n> > changes. You (or someone) should look into why that's happening.\n>\n> I've not read the patch, but given this description I would expect\n> there to be *zero* regression changes --- I don't think we have any\n> test cases that depend on disable_cost being finite. If there's more\n> than zero changes, that very likely indicates a bug in the patch.\n> Even if there are places where the output legitimately changes, you\n> need to justify each one and make sure that you didn't invalidate the\n> intent of that test case.\n\nI spent some more time poking at this patch. It's missing a ton of\nimportant stuff and is wrong in a whole bunch of really serious ways,\nand I'm not going to try to mention all of them in this email. But I\ndo want to talk about some of the more interesting realizations that\ncame to me as I was working my way through this.\n\nOne of the things I realized relatively early is that the patch does\nnothing to propagate disable_cost upward through the plan tree. That\nmeans that if you have a choice between, say,\nSort-over-Append-over-SeqScan and MergeAppend-over-IndexScan, as we do\nin the regression tests, disabling IndexScan doesn't change the plan\nwith the patch applied, as it does in master. That's because only the\nIndexScan node ends up flagged as disabled. Once we start stacking\nother plan nodes on top of disabled plan nodes, the resultant plans\nare at no disadvantage compared to plans containing no disabled nodes.\nThe IndexScan plan survives initially, despite being disabled, because\nit's got a sort order. That lets us use it to build a MergeAppend\npath, and that MergeAppend path is not disabled, and compares\nfavorably on cost.\n\nAfter straining my brain over various plan changes for a long time,\nand hacking on the code somewhat, I realized that just propagating the\nBoolean upward is insufficient to set things right. That's basically\nbecause I was being dumb when I said this:\n\n> I don't think we should care how MANY disabled nodes appear in a\n> plan, particularly.\n\nSuppose we try to plan a Nested Loop with SeqScan disabled, but\nthere's no alternative to a SeqScan for the outer side of the join. If\nwe suppose an upward-propagating Boolean, every path for the join is\ndisabled because every path for the outer side is disabled. That means\nthat we have no reason to avoid paths where the inner side also uses a\ndisabled path. When we loop over the inner rel's pathlist looking for\nways to build a path for the join, we may find some disabled paths\nthere, and the join paths we build from those paths are disabled, but\nso are the join paths where we use a non-disabled path on the inner\nside. So those paths are just competing with each other on cost, and\nthe path type that is supposedly disabled on the outer side of the\njoin ends up not really being very disabled at all. More precisely, if\ndisabling a plan type causes paths to be discarded completely before\nthe join paths are constructed, then they actually do get removed from\nconsideration. But if those paths make it into inner rel's path list,\neven way out towards the end, then paths derived from them can jump to\nthe front of the joinrel's path list.\n\nThe same kind of problem happens with Append or MergeAppend nodes. The\nregression tests expect that we'll avoid disabled plan types whenever\npossible even if we can't avoid them completely; for instance, the\nmatest0 table intentionally omits an index on one child table.\nDisabling sequential scans is expected to disable them for all of the\nother child tables even though for that particular child table there\nis no other option. But that behavior is hard to achieve if every path\nfor the parent rel is \"equally disabled\". You either want the path\nthat uses only the one required SeqScan to be not-disabled even though\none of its children is disabled ... or you want the disabled flag to\nbe more than a Boolean. And while there's probably more than one way\nto make it work, the easiest thing seems to be to just have a\ndisabled-counter in every node that gets initialized to the total\ndisabled-counter values of all of its children, and then you add 1 if\nthat node is itself doing something that is disabled, i.e. the exact\nopposite of what I said in the quote above.\n\nI haven't done enough work to know whether that would match the\ncurrent behavior, let alone whether it would have acceptable\nperformance, and I'm not at all convinced that's the right direction\nanyway. Actually, at the moment, I don't have a very good idea at all\nwhat the right direction is. I do have a feeling that this patch is\nprobably not going in the right direction, but I might be wrong about\nthat, too.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 1 Apr 2024 16:40:44 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> One of the things I realized relatively early is that the patch does\n> nothing to propagate disable_cost upward through the plan tree.\n> ...\n> After straining my brain over various plan changes for a long time,\n> and hacking on the code somewhat, I realized that just propagating the\n> Boolean upward is insufficient to set things right. That's basically\n> because I was being dumb when I said this:\n>> I don't think we should care how MANY disabled nodes appear in a\n>> plan, particularly.\n\nVery interesting, thanks for the summary. So the fact that\ndisable_cost is additive across plan nodes is actually a pretty\nimportant property of the current setup. I think this is closely\nrelated to one argument you made against my upthread idea of using\nIEEE Infinity for disable_cost: that'd mask whether more than one\nof the sub-plans had been disabled.\n\n> ... And while there's probably more than one way\n> to make it work, the easiest thing seems to be to just have a\n> disabled-counter in every node that gets initialized to the total\n> disabled-counter values of all of its children, and then you add 1 if\n> that node is itself doing something that is disabled, i.e. the exact\n> opposite of what I said in the quote above.\n\nYeah, that seems like the next thing to try if anyone plans to pursue\nthis further. That'd essentially do what we're doing now except that\ndisable_cost is its own \"order of infinity\", entirely separate from\nnormal costs.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 01 Apr 2024 17:00:11 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Mon, Apr 1, 2024 at 5:00 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Very interesting, thanks for the summary. So the fact that\n> disable_cost is additive across plan nodes is actually a pretty\n> important property of the current setup. I think this is closely\n> related to one argument you made against my upthread idea of using\n> IEEE Infinity for disable_cost: that'd mask whether more than one\n> of the sub-plans had been disabled.\n\nYes, exactly. I just hadn't quite put the pieces together.\n\n> Yeah, that seems like the next thing to try if anyone plans to pursue\n> this further. That'd essentially do what we're doing now except that\n> disable_cost is its own \"order of infinity\", entirely separate from\n> normal costs.\n\nRight. I think that's actually what I had in mind in the last\nparagraph of http://postgr.es/m/CA+TgmoY+Ltw7B=1FSFSN4yHcu2roWrz-ijBovj-99LZU=9h1dA@mail.gmail.com\nbut that was a while ago and I'd lost track of why it actually\nmattered. But I also have questions about whether that's really the\nright approach.\n\nI think the approach of just not generating paths we don't want in the\nfirst place merits more consideration. We do that in some cases\nalready, but not in others, and I'm not clear why. Like, if\nindex-scans, index-only scans, sorts, nested loops, and hash joins are\ndisabled, something is going to have to give, because the only\nremaining join type is a merge join yet we've ruled out every possible\nway of getting the day into some order, but I'm not sure whether\nthere's some reason that we need exactly the behavior that we have\nright now rather than anything else. Maybe it would be OK to just\ninsist on at least one unparameterized, non-partial path at the\nbaserel level, and then if that ends up forcing us to ignore the\njoin-type restrictions higher up, so be it. Or maybe that's not OK and\nafter I try that out I'll end up writing another email about how I was\na bit clueless about all of this. I don't know. But I feel like it\nmerits more investigation, because I'm having trouble shaking the\ntheory that what we've got right now is pretty arbitrary.\n\nAnd also ... looking at the regression tests, and also thinking about\nthe kinds of problems that I think people run into in real\ndeployments, I can't help feeling like we've somehow got this whole\nthing backwards. enable_wunk imagines that you want to plan as normal\nexcept with one particular plan type excluded from consideration. And\nmaybe that makes sense if the point of the enable_wunk GUC is that the\nplanner feature might be buggy and you might therefore want to turn it\noff to protect yourself, or if the planner feature might be expensive\nand you might want to turn it off to save cycles. But surely that's\nnot the case with something like enable_seqscan or enable_indexscan.\nWhat I think we're mostly doing in the regression tests is shutting\noff every relevant type of plan except one. I theorize that what we\nactually want to do is tell the planner what we do want to happen,\nrather than what we don't want to happen, but we've got this weird set\nof GUCs that do the opposite of that and we're super-attached to them\nbecause they've existed forever. I don't really have a concrete\nproposal here, but I wonder if what we're actually talking about here\nis spending time and energy polishing a mechanism that nobody likes in\nthe first place.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 1 Apr 2024 19:53:58 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Mon, Apr 1, 2024 at 7:54 PM Robert Haas <robertmhaas@gmail.com> wrote:\n\n> What I think we're mostly doing in the regression tests is shutting\n> off every relevant type of plan except one. I theorize that what we\n> actually want to do is tell the planner what we do want to happen,\n> rather than what we don't want to happen, but we've got this weird set\n> of GUCs that do the opposite of that and we're super-attached to them\n> because they've existed forever.\n\n\nSo rather than listing all the things we don't want to happen, we need a\nway to force (nay, highly encourage) a particular solution. As our costing\nis a based on positive numbers, what if we did something like this in\ncostsize.c?\n\n Cost disable_cost = 1.0e10;\n Cost promotion_cost = 1.0e10; // or higher or lower, depending on\nhow strongly we want to \"beat\" disable_costs effects.\n...\n\n if (!enable_seqscan)\n startup_cost += disable_cost;\n else if (promote_seqscan)\n startup_cost -= promotion_cost; // or replace \"promote\" with\n\"encourage\"?\n\n\nCheers,\nGreg\n\nOn Mon, Apr 1, 2024 at 7:54 PM Robert Haas <robertmhaas@gmail.com> wrote:\nWhat I think we're mostly doing in the regression tests is shutting\noff every relevant type of plan except one. I theorize that what we\nactually want to do is tell the planner what we do want to happen,\nrather than what we don't want to happen, but we've got this weird set\nof GUCs that do the opposite of that and we're super-attached to them\nbecause they've existed forever.So rather than listing all the things we don't want to happen, we need a way to force (nay, highly encourage) a particular solution. As our costing is a based on positive numbers, what if we did something like this in costsize.c? Cost disable_cost = 1.0e10; Cost promotion_cost = 1.0e10; // or higher or lower, depending on how strongly we want to \"beat\" disable_costs effects.... if (!enable_seqscan) startup_cost += disable_cost; else if (promote_seqscan) startup_cost -= promotion_cost; // or replace \"promote\" with \"encourage\"?Cheers,Greg","msg_date":"Tue, 2 Apr 2024 10:03:57 -0400","msg_from":"Greg Sabino Mullane <htamfids@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, Apr 2, 2024 at 10:04 AM Greg Sabino Mullane <htamfids@gmail.com> wrote:\n> So rather than listing all the things we don't want to happen, we need a way to force (nay, highly encourage) a particular solution. As our costing is a based on positive numbers, what if we did something like this in costsize.c?\n>\n> Cost disable_cost = 1.0e10;\n> Cost promotion_cost = 1.0e10; // or higher or lower, depending on how strongly we want to \"beat\" disable_costs effects.\n> ...\n>\n> if (!enable_seqscan)\n> startup_cost += disable_cost;\n> else if (promote_seqscan)\n> startup_cost -= promotion_cost; // or replace \"promote\" with \"encourage\"?\n\nI'm pretty sure negative costs are going to create a variety of\nunpleasant planning artifacts. The large positive costs do, too, which\nis where this whole discussion started. If I disable (or promote) some\nparticular plan, I want the rest of the plan tree to come out looking\nas much as possible like what would have happened if the same\nalternative had won organically on cost. I think the only reason we're\ndriving this off of costing today is that making add_path() more\ncomplicated is unappealing, mostly on performance grounds, and if you\nadd disabled-ness (or promoted-ness) as a separate axis of value then\nadd_path() has to know about that on top of everything else. I think\nthe goal here is to come up with a more principled alternative that\nisn't just based on whacking large numbers into the cost and hoping\nsomething good happens ... but it is a whole lot easier to be unhappy\nwith the status quo than it is to come up with something that's\nactually better.\n\nI am planning to spend some more time thinking about it, though.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 2 Apr 2024 11:01:20 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Tue, Apr 2, 2024 at 10:04 AM Greg Sabino Mullane <htamfids@gmail.com> wrote:\n>> if (!enable_seqscan)\n>> startup_cost += disable_cost;\n>> else if (promote_seqscan)\n>> startup_cost -= promotion_cost; // or replace \"promote\" with \"encourage\"?\n\n> I'm pretty sure negative costs are going to create a variety of\n> unpleasant planning artifacts.\n\nIndeed. It might be okay to have negative values for disabled-ness\nif we treat disabled-ness as a \"separate order of infinity\", but\nI suspect that it'd behave poorly when there are both disabled and\npromoted sub-paths in a tree, for pretty much the same reasons you\nexplained just upthread.\n\n> I think the only reason we're\n> driving this off of costing today is that making add_path() more\n> complicated is unappealing, mostly on performance grounds, and if you\n> add disabled-ness (or promoted-ness) as a separate axis of value then\n> add_path() has to know about that on top of everything else.\n\nIt doesn't seem to me that it's a separate axis of value, just a\nhigher-order component of the cost metric. Nonetheless, adding even\na few instructions to add_path comparisons sounds expensive. Maybe\nit'd be fine, but we'd need to do some performance testing.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 02 Apr 2024 11:54:24 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, Apr 2, 2024 at 11:54 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > I'm pretty sure negative costs are going to create a variety of\n> > unpleasant planning artifacts.\n>\n> Indeed. It might be okay to have negative values for disabled-ness\n> if we treat disabled-ness as a \"separate order of infinity\", but\n> I suspect that it'd behave poorly when there are both disabled and\n> promoted sub-paths in a tree, for pretty much the same reasons you\n> explained just upthread.\n\nHmm, can you explain further? I think essentially you'd be maximizing\n#(promoted notes)-#(disabled nodes), but I have no real idea whether\nthat behavior will be exactly what people want or extremely\nunintuitive or something in the middle. It seems like it should be\nfine if there's only promoting or only disabling or if we can respect\nboth the promoting and the disabling, assuming we even want to have\nboth, but I'm suspicious that it will be weird somehow in other cases.\nI can't say exactly in what way, though. Do you have more insight?\n\n> > I think the only reason we're\n> > driving this off of costing today is that making add_path() more\n> > complicated is unappealing, mostly on performance grounds, and if you\n> > add disabled-ness (or promoted-ness) as a separate axis of value then\n> > add_path() has to know about that on top of everything else.\n>\n> It doesn't seem to me that it's a separate axis of value, just a\n> higher-order component of the cost metric. Nonetheless, adding even\n> a few instructions to add_path comparisons sounds expensive. Maybe\n> it'd be fine, but we'd need to do some performance testing.\n\nHmm, yeah. I'm not sure how much difference there is between these\nthings in practice. I didn't run down everything that was happening,\nbut I think what I did was equivalent to making it a higher-order\ncomponent of the cost metric, and it seemed like an awful lot of paths\nwere surviving anyway, e.g. index scans survived\nenable_indexscan=false because they had a sort order, and I think\nsequential scans were surviving enable_seqscan=false too, perhaps\nbecause they had no startup cost. At any rate there's no question that\nadd_path() is hot.\n\n--\nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 2 Apr 2024 12:26:27 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Tue, Apr 2, 2024 at 11:54 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> I suspect that it'd behave poorly when there are both disabled and\n>> promoted sub-paths in a tree, for pretty much the same reasons you\n>> explained just upthread.\n\n> Hmm, can you explain further? I think essentially you'd be maximizing\n> #(promoted notes)-#(disabled nodes), but I have no real idea whether\n> that behavior will be exactly what people want or extremely\n> unintuitive or something in the middle. It seems like it should be\n> fine if there's only promoting or only disabling or if we can respect\n> both the promoting and the disabling, assuming we even want to have\n> both, but I'm suspicious that it will be weird somehow in other cases.\n> I can't say exactly in what way, though. Do you have more insight?\n\nNot really. But if you had, say, a join of a promoted path to a\ndisabled path, that would be treated as on-par with a join of two\nregular paths, which seems like it'd lead to odd choices. Maybe\nit'd be fine, but my gut says it'd likely not act nicely. As you\nsay, it's a lot easier to believe that only-promoted or only-disabled\nsituations would behave sanely.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 02 Apr 2024 12:58:18 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, Apr 2, 2024 at 12:58 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Not really. But if you had, say, a join of a promoted path to a\n> disabled path, that would be treated as on-par with a join of two\n> regular paths, which seems like it'd lead to odd choices. Maybe\n> it'd be fine, but my gut says it'd likely not act nicely. As you\n> say, it's a lot easier to believe that only-promoted or only-disabled\n> situations would behave sanely.\n\nMakes sense.\n\nI wanted to further explore the idea of just not generating plans of\ntypes that are currently disabled. I looked into doing this for\nenable_indexscan and enable_indexonlyscan. As a first step, I\ninvestigated how those settings work now, and was horrified. I don't\nknow whether I just wasn't paying attention back when the original\nindex-only scan work was done -- I remember discussing\nenable_indexonlyscan with you at the time -- or whether it got changed\nsubsequently. Anyway, the current behavior is:\n\n[A] enable_indexscan=false adds disable_cost to the cost of every\nIndex Scan path *and also* every Index-Only Scan path. So disabling\nindex-scans also in effect discourages the use of index-only scans,\nwhich would make sense if we didn't have a separate setting called\nenable_indexonlyscan, but we do. Given that, I think this is\ncompletely and utterly wrong.\n\n[b] enable_indexonlyscan=false causes index-only scan paths not to be\ngenerated at all, but instead, we generate index-scan paths to do the\nsame thing that we would not have generated otherwise. This is weird\nbecause it means that disabling one plan type causes us to consider\nadditional plans of another type, which seems like a thing that a user\nmight not expect. It's more defensible than [A], though, because you\ncould argue that we only omit the index scan path as an optimization,\non the theory that it will always lose to the index-only scan path,\nand thus if the index-only scan path is not generated, there's a point\nto generating the index scan path after all, so we should. However, it\nseems unlikely to me that someone reading the one line of\ndocumentation that we have about this parameter would be able to guess\nthat it works this way.\n\nHere's an example of how the current system behaves:\n\nrobert.haas=# explain select count(*) from pgbench_accounts;\n QUERY PLAN\n-----------------------------------------------------------------------------------------------------------------\n Aggregate (cost=2854.29..2854.30 rows=1 width=8)\n -> Index Only Scan using pgbench_accounts_pkey on pgbench_accounts\n (cost=0.29..2604.29 rows=100000 width=0)\n(2 rows)\n\nrobert.haas=# set enable_indexscan=false;\nSET\nrobert.haas=# explain select count(*) from pgbench_accounts;\n QUERY PLAN\n------------------------------------------------------------------------------\n Aggregate (cost=2890.00..2890.01 rows=1 width=8)\n -> Seq Scan on pgbench_accounts (cost=0.00..2640.00 rows=100000 width=0)\n(2 rows)\n\nrobert.haas=# set enable_seqscan=false;\nSET\nrobert.haas=# set enable_bitmapscan=false;\nSET\nrobert.haas=# explain select count(*) from pgbench_accounts;\n QUERY PLAN\n----------------------------------------------------------------------------------------------------------------------------------\n Aggregate (cost=10000002854.29..10000002854.30 rows=1 width=8)\n -> Index Only Scan using pgbench_accounts_pkey on pgbench_accounts\n (cost=10000000000.29..10000002604.29 rows=100000 width=0)\n(2 rows)\n\nrobert.haas=# set enable_indexonlyscan=false;\nSET\nrobert.haas=# explain select count(*) from pgbench_accounts;\n QUERY PLAN\n-----------------------------------------------------------------------------------------------\n Aggregate (cost=10000002890.00..10000002890.01 rows=1 width=8)\n -> Seq Scan on pgbench_accounts\n(cost=10000000000.00..10000002640.00 rows=100000 width=0)\n(2 rows)\n\nThe first time we run the query, it picks an index-only scan because\nit's the cheapest. When index scans are disabled, the query now picks\na sequential scan, even though it wasn't use an index-only scan,\nbecause the index scan that it was using is perceived to have become\nvery expensive. When we then shut off sequential scans and bitmap-only\nscans, it switches back to an index-only scan, because setting\nenable_indexscan=false didn't completely disable index-only scans, but\njust made them expensive. But now everything looks expensive, so we go\nback to the same plan we had initially, except with the cost increased\nby a bazillion. Finally, when we disable index-only scans, that\nremoves that plan from the pool, so now we pick the second-cheapest\nplan overall, which in this case is a sequential scan.\n\nSo just to see what would happen, I wrote a patch to make\nenable_indexscan and enable_indexonlyscan do exactly what they say on\nthe tin: when you set one of them to false, paths of that type are not\ngenerated, and nothing else changes. I found that there are a\nsurprisingly large number of regression tests that rely on the current\nbehavior, so I took a crack at fixing them to achieve their goals (or\nwhat I believed their goals to be) in other ways. The resulting patch\nis attached for your (or anyone's) possible edification.\n\nJust to be clear, I have no immediate plans to press forward with\ntrying to get something committed here. It seems pretty clear to me\nthat we should fix [A] in some way, but maybe not in the way I did it\nhere. It's also pretty clear to me that the fact that enable_indexscan\nand enable_indexonlyscan work completely differently from each other\nis surprising at best, wrong at worst, but here again, what this patch\ndoes about that is not above reproach. I think it may make sense to\ndig through the behavior of some of the remaining enable_* GUCs before\nsettling on a final strategy here, but I thought that the finds above\nwere interesting enough and bizarre enough that it made sense to drop\nan email now and see what people think of all this before going\nfurther.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com","msg_date":"Wed, 3 Apr 2024 15:21:24 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, Apr 3, 2024 at 3:21 PM Robert Haas <robertmhaas@gmail.com> wrote:\n\n> It's also pretty clear to me that the fact that enable_indexscan\n> and enable_indexonlyscan work completely differently from each other\n> is surprising at best, wrong at worst, but here again, what this patch\n> does about that is not above reproach.\n\n\nYes, that is wrong, surely there is a reason we have two vars. Thanks for\ndigging into this: if nothing else, the code will be better for this\ndiscussion, even if we do nothing for now with disable_cost.\n\nCheers,\nGreg\n\nOn Wed, Apr 3, 2024 at 3:21 PM Robert Haas <robertmhaas@gmail.com> wrote: It's also pretty clear to me that the fact that enable_indexscan\nand enable_indexonlyscan work completely differently from each other\nis surprising at best, wrong at worst, but here again, what this patch\ndoes about that is not above reproach.Yes, that is wrong, surely there is a reason we have two vars. Thanks for digging into this: if nothing else, the code will be better for this discussion, even if we do nothing for now with disable_cost.Cheers,Greg","msg_date":"Wed, 3 Apr 2024 16:03:44 -0400","msg_from":"Greg Sabino Mullane <htamfids@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Thu, 4 Apr 2024 at 08:21, Robert Haas <robertmhaas@gmail.com> wrote:\n> I wanted to further explore the idea of just not generating plans of\n> types that are currently disabled. I looked into doing this for\n> enable_indexscan and enable_indexonlyscan. As a first step, I\n> investigated how those settings work now, and was horrified. I don't\n> know whether I just wasn't paying attention back when the original\n> index-only scan work was done -- I remember discussing\n> enable_indexonlyscan with you at the time -- or whether it got changed\n> subsequently. Anyway, the current behavior is:\n>\n> [A] enable_indexscan=false adds disable_cost to the cost of every\n> Index Scan path *and also* every Index-Only Scan path. So disabling\n> index-scans also in effect discourages the use of index-only scans,\n> which would make sense if we didn't have a separate setting called\n> enable_indexonlyscan, but we do. Given that, I think this is\n> completely and utterly wrong.\n>\n> [b] enable_indexonlyscan=false causes index-only scan paths not to be\n> generated at all, but instead, we generate index-scan paths to do the\n> same thing that we would not have generated otherwise.\n\nFWIW, I think changing this is a bad idea and I don't think the\nbehaviour that's in your patch is useful. With your patch, if I SET\nenable_indexonlyscan=false, any index that *can* support an IOS for my\nquery will now not be considered at all!\n\nWith your patch applied, I see:\n\n-- default enable_* GUC values.\npostgres=# explain select oid from pg_class order by oid;\n QUERY PLAN\n-------------------------------------------------------------------------------------------\n Index Only Scan using pg_class_oid_index on pg_class\n(cost=0.27..22.50 rows=415 width=4)\n(1 row)\n\n\npostgres=# set enable_indexonlyscan=0; -- no index scan?\nSET\npostgres=# explain select oid from pg_class order by oid;\n QUERY PLAN\n-----------------------------------------------------------------\n Sort (cost=36.20..37.23 rows=415 width=4)\n Sort Key: oid\n -> Seq Scan on pg_class (cost=0.00..18.15 rows=415 width=4)\n(3 rows)\n\npostgres=# set enable_seqscan=0; -- still no index scan!\nSET\npostgres=# explain select oid from pg_class order by oid;\n QUERY PLAN\n------------------------------------------------------------------------------------\n Sort (cost=10000000036.20..10000000037.23 rows=415 width=4)\n Sort Key: oid\n -> Seq Scan on pg_class (cost=10000000000.00..10000000018.15\nrows=415 width=4)\n(3 rows)\n\npostgres=# explain select oid from pg_class order by oid,relname; --\nnow an index scan?!\n QUERY PLAN\n---------------------------------------------------------------------------------------------\n Incremental Sort (cost=0.43..79.50 rows=415 width=68)\n Sort Key: oid, relname\n Presorted Key: oid\n -> Index Scan using pg_class_oid_index on pg_class\n(cost=0.27..60.82 rows=415 width=68)\n(4 rows)\n\nI don't think this is good as pg_class_oid_index effectively won't be\nused as long as the particular query could use that index with an\nindex *only* scan. You can see above that as soon as I adjust the\nquery slightly so that IOS isn't possible, the index can be used\nagain. I think an Index Scan would have been a much better option for\nthe 2nd query than the seq scan and sort.\n\nI think if I do SET enable_indexonlyscan=0; the index should still be\nused with an Index Scan and it definitely shouldn't result in Index\nScan also being disabled if that index happens to contain all the\ncolumns required to support an IOS.\n\nFWIW, I'm fine with the current behaviour. It looks like we've\nassumed that, when possible, IOS are always superior to Index Scan, so\nthere's no point in generating an Index Scan path when we can generate\nan IOS path. I think this makes sense. For that not to be true,\nchecking the all visible flag would have to be more costly than\nvisiting the heap. Perhaps that could be true if the visibility map\npage had to come from disk and the heap page was cached and the disk\nwas slow, but I don't think that scenario is worthy of considering\nboth Index scan and IOS path types when IOS is possible. We've no way\nto accurately cost that anyway.\n\nThis all seems similar to enable_sort vs enable_incremental_sort. For\na while, we did consider both an incremental sort and a sort when an\nincremental sort was possible, but it seemed to me that an incremental\nsort would always be better when it was possible, so I changed that in\n4a29eabd1. I've not seen anyone complain. I made it so that when an\nincremental sort is possible but is disabled, we do a sort instead.\nThat seems fairly similar to how master handles\nenable_indexonlyscan=false.\n\nIn short, I don't find it strange that disabling one node type results\nin considering another type that we'd otherwise not consider in cases\nwhere we assume that the disabled node type is always superior and\nshould always be used when it is possible.\n\nI do agree that adding disable_cost to IOS when enable_indexscan=0 is\na bit weird. We don't penalise incremental sorts when sorts are\ndisabled, so aligning those might make sense.\n\nDavid\n\n\n","msg_date":"Thu, 4 Apr 2024 10:15:59 +1300","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Thu, 4 Apr 2024 at 10:15, David Rowley <dgrowleyml@gmail.com> wrote:\n> In short, I don't find it strange that disabling one node type results\n> in considering another type that we'd otherwise not consider in cases\n> where we assume that the disabled node type is always superior and\n> should always be used when it is possible.\n\nIn addition to what I said earlier, I think the current\nenable_indexonlyscan is implemented in a way that has the planner do\nwhat it did before IOS was added. I think that goal makes sense with\nany patch that make the planner try something new. We want to have\nsome method to get the previous behaviour for the cases where the\nplanner makes a dumb choice or to avoid some bug in the new feature.\n\nI think using that logic, the current scenario with enable_indexscan\nand enable_indexonlyscan makes complete sense. I mean, including\nenable_indexscan=0 adding disable_cost to IOS Paths.\n\nDavid\n\n\n","msg_date":"Thu, 4 Apr 2024 16:09:03 +1300","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, Apr 3, 2024 at 11:09 PM David Rowley <dgrowleyml@gmail.com> wrote:\n> On Thu, 4 Apr 2024 at 10:15, David Rowley <dgrowleyml@gmail.com> wrote:\n> > In short, I don't find it strange that disabling one node type results\n> > in considering another type that we'd otherwise not consider in cases\n> > where we assume that the disabled node type is always superior and\n> > should always be used when it is possible.\n>\n> In addition to what I said earlier, I think the current\n> enable_indexonlyscan is implemented in a way that has the planner do\n> what it did before IOS was added. I think that goal makes sense with\n> any patch that make the planner try something new. We want to have\n> some method to get the previous behaviour for the cases where the\n> planner makes a dumb choice or to avoid some bug in the new feature.\n\nI see the logic of this, and I agree that the resulting behavior might\nbe more intuitive than what I posted before. I'll do some experiments.\n\n> I think using that logic, the current scenario with enable_indexscan\n> and enable_indexonlyscan makes complete sense. I mean, including\n> enable_indexscan=0 adding disable_cost to IOS Paths.\n\nThis, for me, is a bridge too far. I don't think there's a real\nargument that \"what the planner did before IOS was added\" was add\ndisable_cost to the cost of index-only scan paths. There was no such\npath type. Independently of that argument, I also think the behavior\nof a setting needs to be something that a user can understand. Right\nnow, the documentation says:\n\nEnables or disables the query planner's use of index-scan plan types.\nThe default is on.\nEnables or disables the query planner's use of index-only-scan plan\ntypes (see Section 11.9). The default is on.\n\nI do not think that a user can be expected to guess from these\ndescriptions that the first one also affects index-only scans, or that\nthe two GUCs disable their respective plan types in completely\ndifferent ways. Granted, the latter inconsistency affects a whole\nbunch of these settings, not just this one, but still.\n\n--\nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 4 Apr 2024 14:02:42 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Sat, Nov 2, 2019 at 10:57 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> The idea that I've been thinking about is to not generate disabled\n> Paths in the first place, thus not only fixing the problem but saving\n> some cycles. While this seems easy enough for \"optional\" paths,\n> we have to reserve the ability to generate certain path types regardless,\n> if there's no other way to implement the query. This is a bit of a\n> stumbling block :-(. At the base relation level, we could do something\n> like generating seqscan last, and only if no other path has been\n> successfully generated.\n\nContinuing my investigation into this rather old thread, I did a\nrather primitive implementation of this idea, for baserels only, and\ndiscovered that it caused a small number of planner failures running\nthe regression tests. Here is a slightly simplified example:\n\nCREATE TABLE strtest (n text, t text);\nCREATE INDEX strtest_n_idx ON strtest (n);\nSET enable_seqscan=false;\nEXPLAIN SELECT * FROM strtest s1 INNER JOIN strtest s2 ON s1.n >= s2.n;\n\nWith the patch, I get:\n\nERROR: could not devise a query plan for the given query\n\nThe problem here is that it's perfectly possible to generate a valid\npath for s1 -- and likewise for s2, since it's the same underlying\nrelation -- while respecting the enable_seqscan=false constraint.\nHowever, all such paths are parameterized by the other of the two\nrelations, which means that if we do that, we can't plan the join,\nbecause we need an unparameterized path for at least one of the two\nsides in order to build a nested loop join, which is the only way to\nsatisfy the parameterization on the other side.\n\nNow, you could try to fix this by deciding that planning for a baserel\nhasn't really succeeded unless we got at least one *unparameterized*\npath for that baserel. I haven't tried that, but I presume that if you\ndo, it fixes the above example, because now there will be a last-ditch\nsequential scan on both sides and so this example will behave as\nexpected. But if you do that, then in other cases, that sequential\nscan is going to get picked even when it isn't strictly necessary to\ndo so, just because some plan that uses it looks better on cost.\nPresumably that problem can in turn be fixed by deciding that we also\nneed to keep disable_cost around (or the separate disable-counter idea\nthat we were discussing recently in another branch of this thread),\nbut that's arguably missing the point of this exercise.\n\nAnother idea is to remove the ERROR mentioned above from\nset_cheapest() and just allow planning to continue even if some\nrelations end up with no paths. (This would necessitate finding and\nfixing any code that could be confused by a pathless relation.) Then,\nif you get to the top of the plan tree and you have no paths there,\nredo the join search discarding the constraints (or maybe just some of\nthe constraints, e.g. allow sequential scans and nested loops, or\nsomething). Conceptually, I like this idea a lot, but I think there\nare a few problems. One is that I'm not quite sure how to find all the\ncode that would need to be adjusted to make it work, though the header\ncomment for standard_join_search() seems like it's got some helpful\ntips. A second is that it's another version of the disable_cost =\ninfinity problem: once you find that you can't generate a path while\nenforcing all of the restrictions, you just disregard the restrictions\ncompletely, instead of discarding them only to the extent necessary. I\nhave a feeling that's not going to be very appealing.\n\nNow, I suppose it might be that even if we can't remove disable_cost,\nsomething along these lines is still worth doing, just to save CPU\ncycles. You could for example try planning with only non-disabled\nstuff and then do it over again with everything if that doesn't work\nout, still keeping disable_cost around so that you avoid disabled\nnodes where you can. But I'm kind of hoping that I'm missing something\nand there's some approach that could both kill disable_cost and save\nsome cycles at the same time. If (any of) you have an idea, I'd love\nto hear it!\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 3 May 2024 16:33:33 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Sat, 4 May 2024 at 08:34, Robert Haas <robertmhaas@gmail.com> wrote:\n> Another idea is to remove the ERROR mentioned above from\n> set_cheapest() and just allow planning to continue even if some\n> relations end up with no paths. (This would necessitate finding and\n> fixing any code that could be confused by a pathless relation.) Then,\n> if you get to the top of the plan tree and you have no paths there,\n> redo the join search discarding the constraints (or maybe just some of\n> the constraints, e.g. allow sequential scans and nested loops, or\n> something).\n\nI don't think you'd need to wait longer than where we do set_cheapest\nand find no paths to find out that there's going to be a problem.\n\nI don't think redoing planning is going to be easy or even useful. I\nmean what do you change when you replan? You can't just do\nenable_seqscan and enable_nestloop as if there's no index to provide\nsorted input and the plan requires some sort, then you still can't\nproduce a plan. Adding enable_sort to the list does not give me much\nconfidence we'll never fail to produce a plan either. It just seems\nimpossible to know which of the disabled ones caused the RelOptInfo to\nhave no paths. Also, you might end up enabling one that caused the\nplanner to do something different than it would do today. For\nexample, a Path that today incurs 2x disable_cost vs a Path that only\nreceives 1x disable_cost might do something different if you just went\nand enabled a bunch of enable* GUCs before replanning.\n\n> Now, I suppose it might be that even if we can't remove disable_cost,\n> something along these lines is still worth doing, just to save CPU\n> cycles. You could for example try planning with only non-disabled\n> stuff and then do it over again with everything if that doesn't work\n> out, still keeping disable_cost around so that you avoid disabled\n> nodes where you can. But I'm kind of hoping that I'm missing something\n> and there's some approach that could both kill disable_cost and save\n> some cycles at the same time. If (any of) you have an idea, I'd love\n> to hear it!\n\nI think the int Path.disabledness idea is worth coding up to try it.\nI imagine that a Path will incur the maximum of its subpath's\ndisabledness's then add_path() just needs to prefer lower-valued\ndisabledness Paths.\n\nThat doesn't get you the benefits of fewer CPU cycles, but where did\nthat come from as a motive to change this? There's no shortage of\nother ways to make the planner faster if that's an issue.\n\nDavid\n\n\n","msg_date":"Sun, 5 May 2024 01:16:19 +1200","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"David Rowley <dgrowleyml@gmail.com> writes:\n> I don't think you'd need to wait longer than where we do set_cheapest\n> and find no paths to find out that there's going to be a problem.\n\nAt a base relation, yes, but that doesn't work for joins: it may be\nthat a particular join cannot be formed, yet other join sequences\nwill work. We have that all the time from outer-join ordering\nrestrictions, never mind enable_xxxjoin flags. So I'm not sure\nthat we can usefully declare early failure for joins.\n\n> I think the int Path.disabledness idea is worth coding up to try it.\n> I imagine that a Path will incur the maximum of its subpath's\n> disabledness's then add_path() just needs to prefer lower-valued\n> disabledness Paths.\n\nI would think sum not maximum, but that's a detail.\n\n> That doesn't get you the benefits of fewer CPU cycles, but where did\n> that come from as a motive to change this? There's no shortage of\n> other ways to make the planner faster if that's an issue.\n\nThe concern was to not *add* CPU cycles in order to make this area\nbetter. But I do tend to agree that we've exhausted all the other\noptions.\n\nBTW, I looked through costsize.c just now to see exactly what we are\nusing disable_cost for, and it seemed like a majority of the cases are\njust wrong. Where possible, we should implement a plan-type-disable\nflag by not generating the associated Path in the first place, not by\napplying disable_cost to it. But it looks like a lot of people have\nerroneously copied the wrong logic. I would say that only these plan\ntypes should use the disable_cost method:\n\n\tseqscan\n\tnestloop join\n\tsort\n\nas those are the only ones where we risk not being able to make a\nplan at all for lack of other alternatives.\n\nThere is also some weirdness around needing to force use of tidscan\nif we have WHERE CURRENT OF. But perhaps a different hack could be\nused for that.\n\nWe also have this for hashjoin:\n\n\t * If the bucket holding the inner MCV would exceed hash_mem, we don't\n\t * want to hash unless there is really no other alternative, so apply\n\t * disable_cost.\n\nI'm content to leave that be, if we can't remove disable_cost\nentirely.\n\nWhat I'm wondering at this point is whether we need to trouble with\nimplementing the separate-disabledness-count method, if we trim back\nthe number of places using disable_cost to the absolute minimum.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 04 May 2024 12:57:32 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Sun, 5 May 2024 at 04:57, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> David Rowley <dgrowleyml@gmail.com> writes:\n> > That doesn't get you the benefits of fewer CPU cycles, but where did\n> > that come from as a motive to change this? There's no shortage of\n> > other ways to make the planner faster if that's an issue.\n>\n> The concern was to not *add* CPU cycles in order to make this area\n> better. But I do tend to agree that we've exhausted all the other\n> options.\n\nIt really looks to me that Robert was talking about not generating\npaths for disabled path types. He did write \"just to save CPU cycles\"\nin the paragraph I quoted.\n\nI think we should concern ourselves with adding overhead to add_path()\n*only* when we actually see a patch which slows it down in a way that\nwe can measure. I find it hard to imagine that adding a single\ncomparison for every Path is measurable. Each of these paths has been\npalloced and costed, both of which are significantly more expensive\nthan adding another comparison to compare_path_costs_fuzzily(). I'm\nonly willing for benchmarks on an actual patch to prove me wrong on\nthat. Nothing else. add_path() has become a rat's nest of conditions\nover the years and those seem to have made it without concerns about\nperformance.\n\n> BTW, I looked through costsize.c just now to see exactly what we are\n> using disable_cost for, and it seemed like a majority of the cases are\n> just wrong. Where possible, we should implement a plan-type-disable\n> flag by not generating the associated Path in the first place, not by\n> applying disable_cost to it. But it looks like a lot of people have\n> erroneously copied the wrong logic. I would say that only these plan\n> types should use the disable_cost method:\n>\n> seqscan\n> nestloop join\n> sort\n\nI think this oversimplifies the situation. I only spent 30 seconds\nlooking and I saw cases where this would cause issues. If\nenable_hashagg is false, we could fail to produce some plans where the\ntype is sortable but not hashable. There's also an issue with nested\nloops being unable to FULL OUTER JOIN. However, I do agree that there\nare some in there that are adding disable_cost that should be done by\njust not creating the Path. enable_gathermerge is one.\nenable_bitmapscan is probably another.\n\nI understand you only talked about the cases adding disable_cost in\ncostsize.c. But just as a reminder, there are other things we need to\nbe careful not to break. For example, enable_indexonlyscan=false\nshould defer to still making an index scan. Nobody who disables\nenable_indexonlyscan without disabling enable_indexscan wants queries\nthat are eligible to use IOS to use seq scan instead. They'd still\nwant Index Scan to be considered, otherwise they'd have disabled\nenable_indexscan.\n\nDavid\n\n\n","msg_date":"Sun, 5 May 2024 12:07:30 +1200","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Sat, May 4, 2024 at 9:16 AM David Rowley <dgrowleyml@gmail.com> wrote:\n> I don't think you'd need to wait longer than where we do set_cheapest\n> and find no paths to find out that there's going to be a problem.\n\nI'm confused by this response, because I thought that the main point\nof my previous email was explaining why that's not true. I showed an\nexample where you do find paths at set_cheapest() time and yet are\nunable to complete planning.\n\n> I don't think redoing planning is going to be easy or even useful. I\n> mean what do you change when you replan? You can't just do\n> enable_seqscan and enable_nestloop as if there's no index to provide\n> sorted input and the plan requires some sort, then you still can't\n> produce a plan. Adding enable_sort to the list does not give me much\n> confidence we'll never fail to produce a plan either. It just seems\n> impossible to know which of the disabled ones caused the RelOptInfo to\n> have no paths. Also, you might end up enabling one that caused the\n> planner to do something different than it would do today. For\n> example, a Path that today incurs 2x disable_cost vs a Path that only\n> receives 1x disable_cost might do something different if you just went\n> and enabled a bunch of enable* GUCs before replanning.\n\nI agree that there are problems here, both in terms of implementation\ncomplexity and also in terms of what behavior you actually get, but I\ndo not think that a proposal which changes some current behavior\nshould be considered dead on arrival. Whatever new behavior we might\nwant to implement needs to make sense, and there need to be good\nreasons for making whatever changes are contemplated, but I don't\nthink we should take the position that it has to be identical to\ncurrent.\n\n> I think the int Path.disabledness idea is worth coding up to try it.\n> I imagine that a Path will incur the maximum of its subpath's\n> disabledness's then add_path() just needs to prefer lower-valued\n> disabledness Paths.\n\nIt definitely needs to be sum, not max. Otherwise you can't get the\nmatest example from the regression tests right, where one child lacks\nthe ability to comply with the GUC setting.\n\n> That doesn't get you the benefits of fewer CPU cycles, but where did\n> that come from as a motive to change this? There's no shortage of\n> other ways to make the planner faster if that's an issue.\n\nWell, I don't agree with that at all. If there are lots of ways to\nmake the planner faster, we should definitely do a bunch of that\nstuff, because \"will slow down the planner too much\" has been a\nleading cause of proposed planner patches being rejected for as long\nas I've been involved with the project. My belief was that we were\nrather short of good ideas in that area, actually. But even if it's\ntrue that we have lots of other ways to speed up the planner, that\ndoesn't mean that it wouldn't be good to do it here, too.\n\nStepping back a bit, my current view of this area is: disable_cost is\nhighly imperfect both as an idea and as implemented in PostgreSQL.\nAlthough I'm discovering that the current implementation gets more\nthings right than I had realized, it also sometimes gets things wrong.\nThe original poster gave an example of that, and there are others.\nFurthermore, the current implementation has some weird\ninconsistencies. Therefore, I would like something better. Better, to\nme, could mean any combination of (a) superior behavior, (b) superior\nperformance, and (c) simpler, more elegant code. In a perfect world,\nwe'd be able to come up with something that wins in all three of those\nareas, but I'm not seeing a way to achieve that, so I'm trying to\nfigure out what is achievable. And because we need to reach consensus\non whatever is to be done, I'm sharing raw research results rather\nthan just dropping a completed patch. I don't think it's at all easy\nto understand what the realistic possibilities are in this area;\ncertainly it isn't for me. At some point I'm hoping that there will be\na patch (or a bunch of patches) that we can all agree are an\nimprovement over now and the best we can reasonably do, but I don't\nyet know what the shape of those will be, because I'm still trying to\nunderstand (and document on-list) what all the problems are.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 6 May 2024 08:27:38 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Sat, May 4, 2024 at 12:57 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> There is also some weirdness around needing to force use of tidscan\n> if we have WHERE CURRENT OF. But perhaps a different hack could be\n> used for that.\n\nYeah, figuring out what to do about this was the trickiest part of the\nexperimental patch that I wrote last week. The idea of the current\ncode is that cost_qual_eval_walker charges disable_cost for\nCurrentOfExpr, but cost_tidscan then subtracts disable_cost if\ntidquals contains a CurrentOfExpr, so that we effectively disable\neverything except TID scan paths and, I think, also any TID scan paths\nthat don't use the CurrentOfExpr as a qual. I'm not entirely sure\nwhether the last can happen, but I imagine that it might be possible\nif the cursor refers to a query that itself contains some other kind\nof TID qual.\n\nIt's not very clear that this mechanism is actually 100% reliable,\nbecause we know it's possible in general for the costs of two paths to\nbe different by more than disable_cost. Maybe that's not possible in\nthis specific context, though: I'm not sure.\n\nThe approach I took for my experimental patch was pretty grotty, and\nprobably not quite complete, but basically I defined the case where we\ncurrently subtract out disable_cost as a \"forced TID-scan\". I passed\naround a Boolean called forcedTidScan which gets set to true if we\ndiscover that some plan is a forced TID-scan path, and then we discard\nany other paths and then only add other forced TID-scan paths after\nthat point. There can be more than one, because of parameterization.\n\nBut I think that the right thing to do is probably to pull some of the\nlogic up out of create_tidscan_paths() and decide ONCE whether we're\nin a forced TID-scan situation or not. If we are, then\nset_plain_rel_pathlist() should arrange to create only forced TID-scan\npaths; otherwise, it should proceed as it does now.\n\nMaybe if I try to do that I'll find problems, but the current approach\nseems backwards to me, like going to a restaurant and ordering one of\neverything on the menu, then cancelling all of the orders except the\nstuff you actually want.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 6 May 2024 09:39:38 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Mon, May 6, 2024 at 9:39 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> It's not very clear that this mechanism is actually 100% reliable,\n\nIt isn't. Here's a test case. As a non-superuser, do this:\n\ncreate table foo (a int, b text, primary key (a));\ninsert into foo values (1, 'Apple');\nalter table foo enable row level security;\nalter table foo force row level security;\ncreate policy p1 on foo as permissive using (ctid in ('(0,1)', '(0,2)'));\nbegin;\ndeclare c cursor for select * from foo;\nfetch from c;\nexplain update foo set b = 'Manzana' where current of c;\nupdate foo set b = 'Manzana' where current of c;\n\nThe explain produces this output:\n\n Update on foo (cost=10000000000.00..10000000008.02 rows=0 width=0)\n -> Tid Scan on foo (cost=10000000000.00..10000000008.02 rows=1 width=38)\n TID Cond: (ctid = ANY ('{\"(0,1)\",\"(0,2)\"}'::tid[]))\n Filter: CURRENT OF c\n\nUnless I'm quite confused, the point of the code is to force\nCurrentOfExpr to be a TID Cond, and it normally succeeds in doing so,\nbecause WHERE CURRENT OF cursor_name has to be the one and only WHERE\ncondition for a normal UPDATE. I tried various cases involving views\nand CTEs and got nowhere. But then I wrote a patch to make the\nregression tests fail if a baserel's restrictinfo list contains a\nCurrentOfExpr and also some other qual, and a couple of row-level\nsecurity tests failed (and nothing else). Which then allowed me to\nconstruct the example above, where there are two possible TID quals\nand the logic in tidpath.c latches onto the wrong one. The actual\nUPDATE fails like this:\n\nERROR: WHERE CURRENT OF is not supported for this table type\n\n...because ExecEvalCurrentOfExpr supposes that the only way it can be\nreached is for an FDW without the necessary support, but actually in\nthis case it's planner error that gets us here.\n\nFortunately, there's no real reason for anyone to ever do something\nlike this, or at least I can't see one, so the fact that it doesn't\nwork probably doesn't really matter that much. And you can argue that\nthe only problem here is that the costing hack just didn't get updated\nfor RLS and now needs to be a bit more clever. But I think it'd be\nbetter to find a way of making it less hacky. With the way the code is\nstructured right now, the chances of anyone understanding that RLS\nmight have an impact on its correctness were just about nil, IMHO.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 6 May 2024 13:51:30 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Mon, May 6, 2024 at 9:39 AM Robert Haas <robertmhaas@gmail.com> wrote:\n>> It's not very clear that this mechanism is actually 100% reliable,\n\n> It isn't. Here's a test case.\n\nVery interesting.\n\n> ... Which then allowed me to\n> construct the example above, where there are two possible TID quals\n> and the logic in tidpath.c latches onto the wrong one.\n\nHmm. Without having traced through it, I'm betting that the\nCurrentOfExpr qual is rejected as a tidqual because it's not\nconsidered leakproof. It's not obvious to me why we couldn't consider\nit as leakproof, though. If we don't want to do that in general,\nthen we need some kind of hack in TidQualFromRestrictInfo to accept\nCurrentOfExpr quals anyway.\n\nIn general I think you're right that something less rickety than\nthe disable_cost hack would be a good idea to ensure the desired\nTidPath gets chosen, but this problem is not the fault of that.\nWe're not making the TidPath with the correct contents in the first\nplace.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 06 May 2024 14:44:40 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"I wrote:\n> Robert Haas <robertmhaas@gmail.com> writes:\n>> ... Which then allowed me to\n>> construct the example above, where there are two possible TID quals\n>> and the logic in tidpath.c latches onto the wrong one.\n\n> Hmm. Without having traced through it, I'm betting that the\n> CurrentOfExpr qual is rejected as a tidqual because it's not\n> considered leakproof.\n\nNah, I'm wrong: we do treat it as leakproof, and the comment about\nthat in contain_leaked_vars_walker shows that the interaction with\nRLS quals *was* thought about. What wasn't thought about was the\npossibility of RLS quals that themselves could be usable as tidquals,\nwhich breaks this assumption in TidQualFromRestrictInfoList:\n\n * Stop as soon as we find any usable CTID condition. In theory we\n * could get CTID equality conditions from different AND'ed clauses,\n * in which case we could try to pick the most efficient one. In\n * practice, such usage seems very unlikely, so we don't bother; we\n * just exit as soon as we find the first candidate.\n\nThe executor doesn't seem to be prepared to cope with multiple AND'ed\nTID clauses (only OR'ed ones). So we need to fix this at least to the\nextent of looking for a CurrentOfExpr qual, and preferring that over\nanything else.\n\nI'm also now wondering about this assumption in the executor:\n\n /* CurrentOfExpr could never appear OR'd with something else */\n Assert(list_length(tidstate->tss_tidexprs) == 1 ||\n !tidstate->tss_isCurrentOf);\n\nIt still seems OK, because anything that might come in from RLS quals\nwould be AND'ed not OR'ed with the CurrentOfExpr.\n\n> In general I think you're right that something less rickety than\n> the disable_cost hack would be a good idea to ensure the desired\n> TidPath gets chosen, but this problem is not the fault of that.\n> We're not making the TidPath with the correct contents in the first\n> place.\n\nStill true.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 06 May 2024 15:26:52 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Mon, May 6, 2024 at 3:26 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Nah, I'm wrong: we do treat it as leakproof, and the comment about\n> that in contain_leaked_vars_walker shows that the interaction with\n> RLS quals *was* thought about. What wasn't thought about was the\n> possibility of RLS quals that themselves could be usable as tidquals,\n> which breaks this assumption in TidQualFromRestrictInfoList:\n>\n> * Stop as soon as we find any usable CTID condition. In theory we\n> * could get CTID equality conditions from different AND'ed clauses,\n> * in which case we could try to pick the most efficient one. In\n> * practice, such usage seems very unlikely, so we don't bother; we\n> * just exit as soon as we find the first candidate.\n\nRight. I had noticed this but didn't spell it out.\n\n> The executor doesn't seem to be prepared to cope with multiple AND'ed\n> TID clauses (only OR'ed ones). So we need to fix this at least to the\n> extent of looking for a CurrentOfExpr qual, and preferring that over\n> anything else.\n>\n> I'm also now wondering about this assumption in the executor:\n>\n> /* CurrentOfExpr could never appear OR'd with something else */\n> Assert(list_length(tidstate->tss_tidexprs) == 1 ||\n> !tidstate->tss_isCurrentOf);\n>\n> It still seems OK, because anything that might come in from RLS quals\n> would be AND'ed not OR'ed with the CurrentOfExpr.\n\nThis stuff I had not noticed.\n\n> > In general I think you're right that something less rickety than\n> > the disable_cost hack would be a good idea to ensure the desired\n> > TidPath gets chosen, but this problem is not the fault of that.\n> > We're not making the TidPath with the correct contents in the first\n> > place.\n>\n> Still true.\n\nI'll look into this, unless you want to do it.\n\nIncidentally, another thing I just noticed is that\nIsCurrentOfClause()'s test for (node->cvarno == rel->relid) is\npossibly dead code. At least, there are no examples in our test suite\nwhere it fails to hold. Which seems like it makes sense, because if it\ndidn't, then how did the clause end up in baserestrictinfo? Maybe this\nis worth keeping as defensive coding, or maybe it should be changed to\nan Assert or something.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 6 May 2024 15:58:47 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> I'll look into this, unless you want to do it.\n\nI have a draft patch already. Need to add a test case.\n\n> Incidentally, another thing I just noticed is that\n> IsCurrentOfClause()'s test for (node->cvarno == rel->relid) is\n> possibly dead code. At least, there are no examples in our test suite\n> where it fails to hold. Which seems like it makes sense, because if it\n> didn't, then how did the clause end up in baserestrictinfo? Maybe this\n> is worth keeping as defensive coding, or maybe it should be changed to\n> an Assert or something.\n\nI wouldn't remove it, but maybe an Assert is good enough. The tests\non Vars' varno should be equally pointless no?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 06 May 2024 16:10:44 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Mon, May 6, 2024 at 8:27 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> Stepping back a bit, my current view of this area is: disable_cost is\n> highly imperfect both as an idea and as implemented in PostgreSQL.\n> Although I'm discovering that the current implementation gets more\n> things right than I had realized, it also sometimes gets things wrong.\n> The original poster gave an example of that, and there are others.\n> Furthermore, the current implementation has some weird\n> inconsistencies. Therefore, I would like something better.\n\nFWIW I always found those weird inconsistencies to be annoying at\nbest, and confusing at worst. I speak as somebody that uses\ndisable_cost a lot.\n\nI certainly wouldn't ask anybody to make it a priority for that reason\nalone -- it's not *that* bad. I've given my opinion on this because\nit's already under discussion.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Mon, 6 May 2024 16:30:20 -0400","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Mon, May 6, 2024 at 4:30 PM Peter Geoghegan <pg@bowt.ie> wrote:\n> FWIW I always found those weird inconsistencies to be annoying at\n> best, and confusing at worst. I speak as somebody that uses\n> disable_cost a lot.\n>\n> I certainly wouldn't ask anybody to make it a priority for that reason\n> alone -- it's not *that* bad. I've given my opinion on this because\n> it's already under discussion.\n\nThanks, it's good to have other perspectives.\n\nHere are some patches for discussion.\n\n0001 gets rid of disable_cost as a mechanism for forcing a TID scan\nplan to be chosen when CurrentOfExpr is present. Instead, it arranges\nto generate only the valid path when that case occurs, and skip\neverything else. I think this is a good cleanup, and it doesn't seem\ntotally impossible that it actually prevents a failure in some extreme\ncase.\n\n0002 cleans up the behavior of enable_indexscan and\nenable_indexonlyscan. Currently, setting enable_indexscan=false adds\ndisable_cost to both the cost of index scans and the cost of\nindex-only scans. I think that's indefensible and, in fact, a bug,\nalthough I believe David Rowley disagrees. With this patch, we simply\ndon't generate index scans if enable_indexscan=false, and we don't\ngenerate index-only scans if enable_indexonlyscan=false, which seems a\nlot more consistent to me. However, I did revise one major thing from\nthe patch I posted before, per feedback from David Rowley and also per\nmy own observations: in this version, if enable_indexscan=true and\nenable_indexonlyscan=false, we'll generate index-scan paths for any\ncases where, with both set to true, we would have only generated\nindex-only scan paths. That change makes the behavior of this patch a\nlot more comprehensible and intuitive: the only regression test\nchanges are places where somebody expected that they could disable\nboth index scans and index-only scans by setting\nenable_indexscan=false.\n\n0003 and 0004 extend the approach of \"just don't generate the disabled\npath\" to bitmap scans and gather merge, respectively. I think these\nare more debatable, mostly because it's not clear how far we can\nreally take this approach. Neither breaks any test cases, and 0003 is\nclosely related to the work done in 0002, which seems like a point in\nits favor. 0004 was simply the only other case where it was obvious to\nme that this kind of approach made sense. In my view, it makes most\nsense to use this kind of approach for planner behaviors that seem\nlike they're sort of optional: like if you don't use gather merge, you\ncan still use gather, and if you don't use index scans, you can still\nuse sequential scans. With all these patches applied, the remaining\ncases where we rely on disable_cost are:\n\nsequential scans\nsorts\nhash aggregation\nall 3 join types\nhash joins where a bucket holding the inner MCV would exceed hash_mem\n\nSequential scans are clearly a last-ditch method. I find it a bit hard\nto decide whether hashing or sorting is the default, especially giving\nthe asymmetry between enable_sort - presumptively anywhere - and\nenable_hashagg - specific to aggregation. As for the join types, it's\ntempting to consider nested-loop the default type -- it's the only way\nto satisfy parameterizations, for instance -- but the fact that it's\nthe only method that can't do a full join undermines that position in\nmy book. But, I don't want to pretend like I have all the answers\nhere, either; I'm just sharing some thoughts.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com","msg_date":"Tue, 7 May 2024 16:19:22 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, May 7, 2024 at 4:19 PM Robert Haas <robertmhaas@gmail.com> wrote:\n> Here are some patches for discussion.\n\nWell, that didn't generate much discussion, but here I am trying\nagain. Here I've got patches 0001 and 0002 from my previous posting;\nI've dropped 0003 and 0004 from the previous set for now so as not to\ndistract from the main event, but they may still be a good idea.\nInstead I've got an 0003 and an 0004 that implement the \"count of\ndisabled nodes\" approach that we have discussed previously. This seems\nto work fine, unlike the approaches I tried earlier. I think this is\nthe right direction to go, but I'd like to know what concerns people\nmight have.\n\nThis doesn't completely remove disable_cost, because hash joins still\nadd it to the cost when it's impossible to fit the MCV value into\nwork_mem. I'm not sure what to do with that. Continuing to use\ndisable_cost in that one scenario seems OK to me. We could\nalternatively make that scenario bump disabled_nodes, but I don't\nreally want to confuse the planner not wanting to do something with\nthe user telling the planner not to do something, so I don't think\nthat's a good idea. Or we could rejigger things so that in that case\nwe don't generate the plan at all. I'm not sure why we don't do that\nalready, actually.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com","msg_date":"Wed, 12 Jun 2024 11:35:48 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Hi,\n\nOn 2024-06-12 11:35:48 -0400, Robert Haas wrote:\n> Subject: [PATCH v2 3/4] Treat the # of disabled nodes in a path as a separate\n> cost metric.\n> \n> Previously, when a path type was disabled by e.g. enable_seqscan=false,\n> we either avoided generating that path type in the first place, or\n> more commonly, we added a large constant, called disable_cost, to the\n> estimated startup cost of that path. This latter approach can distort\n> planning. For instance, an extremely expensive non-disabled path\n> could seem to be worse than a disabled path, especially if the full\n> cost of that path node need not be paid (e.g. due to a Limit).\n> Or, as in the regression test whose expected output changes with this\n> commit, the addition of disable_cost can make two paths that would\n> normally be distinguishible cost seem to have fuzzily the same cost.\n> \n> To fix that, we now count the number of disabled path nodes and\n> consider that a high-order component of both the cost. Hence, the\n> path list is now sorted by disabled_nodes and then by total_cost,\n> instead of just by the latter, and likewise for the partial path list.\n> It is important that this number is a count and not simply a Boolean;\n> else, as soon as we're unable to respect disabled path types in all\n> portions of the path, we stop trying to avoid them where we can.\n\n\n> \tif (criterion == STARTUP_COST)\n> \t{\n> \t\tif (path1->startup_cost < path2->startup_cost)\n> @@ -118,6 +127,15 @@ compare_fractional_path_costs(Path *path1, Path *path2,\n> \tCost\t\tcost1,\n> \t\t\t\tcost2;\n> \n> +\t/* Number of disabled nodes, if different, trumps all else. */\n> +\tif (unlikely(path1->disabled_nodes != path2->disabled_nodes))\n> +\t{\n> +\t\tif (path1->disabled_nodes < path2->disabled_nodes)\n> +\t\t\treturn -1;\n> +\t\telse\n> +\t\t\treturn +1;\n> +\t}\n\nI suspect it's going to be ok, because the branch is going to be very\npredictable in normal workloads, but I still worry a bit about making\ncompare_path_costs_fuzzily() more expensive. For more join-heavy queries it\ncan really show up and there's plenty ORM generated join-heavy query\nworkloads.\n\nIf costs were 32 bit integers, I'd have suggested just stashing the disabled\ncounts in the upper 32 bits of a 64bit integer. But ...\n\n<can't resist trying if I see overhead>\n\n\nIn an extreme case i can see a tiny bit of overhead, but not enough to be\nworth worrying about. Mostly because we're so profligate in doing\nbms_overlap() that cost comparisons don't end up mattering as much - I seem to\nrecall that being different in the not distant past though.\n\n\nAside: I'm somewhat confused by add_paths_to_joinrel()'s handling of\nmergejoins_allowed. If mergejoins are disabled we end up reaching\nmatch_unsorted_outer() in more cases than with mergejoins enabled. E.g. we\nonly set mergejoin_enabled for right joins inside select_mergejoin_clauses(),\nbut we don't call select_mergejoin_clauses() if !enable_mergejoin and jointype\n!= FULL. I, what?\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 12 Jun 2024 11:11:23 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, Jun 12, 2024 at 2:11 PM Andres Freund <andres@anarazel.de> wrote:\n> <can't resist trying if I see overhead>\n>\n> In an extreme case i can see a tiny bit of overhead, but not enough to be\n> worth worrying about. Mostly because we're so profligate in doing\n> bms_overlap() that cost comparisons don't end up mattering as much - I seem to\n> recall that being different in the not distant past though.\n\nThere are very few things I love more than when you can't resist\ntrying to break my patches and yet fail to find a problem. Granted the\nlatter part only happens once a century or so, but I'll take it.\n\n> Aside: I'm somewhat confused by add_paths_to_joinrel()'s handling of\n> mergejoins_allowed. If mergejoins are disabled we end up reaching\n> match_unsorted_outer() in more cases than with mergejoins enabled. E.g. we\n> only set mergejoin_enabled for right joins inside select_mergejoin_clauses(),\n> but we don't call select_mergejoin_clauses() if !enable_mergejoin and jointype\n> != FULL. I, what?\n\nI agree this logic is extremely confusing, but \"we only set\nmergejoin_enabled for right joins inside select_mergejoin_clauses()\"\ndoesn't seem to be true. It starts out true, and always stays true\nexcept for right, right-anti, and full joins, where\nselect_mergejoin_clauses() can set it to false. Since the call to\nmatch_unsorted_outer() is gated by mergejoin_enabled, you might think\nthat we'd skip considering nested loops on the strength of not being\nable to do a merge join, but comment \"2.\" in add_paths_to_joinrel\nexplains that the join types for which mergejoin_enabled can end up\nfalse aren't supported by nested loops anyway. Still, this logic is\nreally tortured.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 12 Jun 2024 14:33:31 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Hi,\n\nOn 2024-06-12 14:33:31 -0400, Robert Haas wrote:\n> On Wed, Jun 12, 2024 at 2:11 PM Andres Freund <andres@anarazel.de> wrote:\n> > <can't resist trying if I see overhead>\n> >\n> > In an extreme case i can see a tiny bit of overhead, but not enough to be\n> > worth worrying about. Mostly because we're so profligate in doing\n> > bms_overlap() that cost comparisons don't end up mattering as much - I seem to\n> > recall that being different in the not distant past though.\n> \n> There are very few things I love more than when you can't resist\n> trying to break my patches and yet fail to find a problem. Granted the\n> latter part only happens once a century or so, but I'll take it.\n\n:)\n\n\nToo high cost in path cost comparison is what made me look at the PG code for\nthe first time, IIRC :)\n\n\n\n> > Aside: I'm somewhat confused by add_paths_to_joinrel()'s handling of\n> > mergejoins_allowed. If mergejoins are disabled we end up reaching\n> > match_unsorted_outer() in more cases than with mergejoins enabled. E.g. we\n> > only set mergejoin_enabled for right joins inside select_mergejoin_clauses(),\n> > but we don't call select_mergejoin_clauses() if !enable_mergejoin and jointype\n> > != FULL. I, what?\n> \n> I agree this logic is extremely confusing, but \"we only set\n> mergejoin_enabled for right joins inside select_mergejoin_clauses()\"\n> doesn't seem to be true.\n\nSorry, should have been more precise. With \"set\" I didn't mean set to true,\nbut that that it's only modified within select_mergejoin_clauses().\n\n\n> It starts out true, and always stays true except for right, right-anti, and\n> full joins, where select_mergejoin_clauses() can set it to false. Since the\n> call to match_unsorted_outer() is gated by mergejoin_enabled, you might\n> think that we'd skip considering nested loops on the strength of not being\n> able to do a merge join, but comment \"2.\" in add_paths_to_joinrel explains\n> that the join types for which mergejoin_enabled can end up false aren't\n> supported by nested loops anyway. Still, this logic is really tortured.\n\nAgree that that's the logic - but doesn't that mean we'll consider nestloops\nfor e.g. right joins iff enable_mergejoin=false?\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 12 Jun 2024 11:48:35 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, Jun 12, 2024 at 2:48 PM Andres Freund <andres@anarazel.de> wrote:\n> Sorry, should have been more precise. With \"set\" I didn't mean set to true,\n> but that that it's only modified within select_mergejoin_clauses().\n\nOh. \"set\" has more than one relevant meaning here.\n\n> > It starts out true, and always stays true except for right, right-anti, and\n> > full joins, where select_mergejoin_clauses() can set it to false. Since the\n> > call to match_unsorted_outer() is gated by mergejoin_enabled, you might\n> > think that we'd skip considering nested loops on the strength of not being\n> > able to do a merge join, but comment \"2.\" in add_paths_to_joinrel explains\n> > that the join types for which mergejoin_enabled can end up false aren't\n> > supported by nested loops anyway. Still, this logic is really tortured.\n>\n> Agree that that's the logic - but doesn't that mean we'll consider nestloops\n> for e.g. right joins iff enable_mergejoin=false?\n\nNo, because that function has its own internal guards. See nestjoinOK.\n\nBut don't misunderstand me: I'm not defending the status quo. The\nwhole thing seems like a Rube Goldberg machine to me.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 12 Jun 2024 15:11:53 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, Jun 12, 2024 at 11:35 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> Well, that didn't generate much discussion, but here I am trying\n> again. Here I've got patches 0001 and 0002 from my previous posting;\n> I've dropped 0003 and 0004 from the previous set for now so as not to\n> distract from the main event, but they may still be a good idea.\n> Instead I've got an 0003 and an 0004 that implement the \"count of\n> disabled nodes\" approach that we have discussed previously. This seems\n> to work fine, unlike the approaches I tried earlier. I think this is\n> the right direction to go, but I'd like to know what concerns people\n> might have.\n\nHere is a rebased patch set, where I also fixed pgindent damage and a\ncouple of small oversights in 0004.\n\nI am hoping to get these committed some time in July. So if somebody\nthinks that's too soon or thinks it shouldn't happen at all, please\ndon't wait too long to let me know about that.\n\nThanks,\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com","msg_date":"Fri, 28 Jun 2024 11:46:24 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On 28/06/2024 18:46, Robert Haas wrote:\n> On Wed, Jun 12, 2024 at 11:35 AM Robert Haas <robertmhaas@gmail.com> wrote:\n>> Well, that didn't generate much discussion, but here I am trying\n>> again. Here I've got patches 0001 and 0002 from my previous posting;\n>> I've dropped 0003 and 0004 from the previous set for now so as not to\n>> distract from the main event, but they may still be a good idea.\n>> Instead I've got an 0003 and an 0004 that implement the \"count of\n>> disabled nodes\" approach that we have discussed previously. This seems\n>> to work fine, unlike the approaches I tried earlier. I think this is\n>> the right direction to go, but I'd like to know what concerns people\n>> might have.\n> \n> Here is a rebased patch set, where I also fixed pgindent damage and a\n> couple of small oversights in 0004.\n> \n> I am hoping to get these committed some time in July. So if somebody\n> thinks that's too soon or thinks it shouldn't happen at all, please\n> don't wait too long to let me know about that.\n\nv3-0001-Remove-grotty-use-of-disable_cost-for-TID-scan-pl.patch:\n\n+1, this seems ready for commit\n\nv3-0002-Rationalize-behavior-of-enable_indexscan-and-enab.patch:\n\nI fear this will break people's applications, if they are currently \nforcing a sequential scan with \"set enable_indexscan=off\". Now they will \nneed to do \"set enable_indexscan=off; set enable_indexonlyscan=off\" for \nthe same effect. Maybe it's acceptable, disabling sequential scans to \nforce an index scan is much more common than the other way round.\n\nv3-0003-Treat-number-of-disabled-nodes-in-a-path-as-a-sep.patch:\n\n> @@ -1318,6 +1342,12 @@ cost_tidscan(Path *path, PlannerInfo *root,\n> \tstartup_cost += path->pathtarget->cost.startup;\n> \trun_cost += path->pathtarget->cost.per_tuple * path->rows;\n> \n> +\t/*\n> +\t * There are assertions above verifying that we only reach this function\n> +\t * either when enable_tidscan=true or when the TID scan is the only legal\n> +\t * path, so it's safe to set disabled_nodes to zero here.\n> +\t */\n> +\tpath->disabled_nodes = 0;\n> \tpath->startup_cost = startup_cost;\n> \tpath->total_cost = startup_cost + run_cost;\n> }\n\nSo if you have enable_tidscan=off, and have a query with \"WHERE CURRENT \nOF foo\" that is planned with a TID scan, we set disable_nodes = 0? That \nsounds wrong, shouldn't disable_nodes be 1 in that case? It probably \ncannot affect the rest of the plan, given that \"WHERE CURRENT OF\" is \nonly valid in an UPDATE or DELETE, but still. At least it deserves a \nbetter explanation in the comment.\n\n> diff --git a/src/backend/optimizer/plan/createplan.c b/src/backend/optimizer/plan/createplan.c\n> index 6b64c4a362..20236e8c4d 100644\n> --- a/src/backend/optimizer/plan/createplan.c\n> +++ b/src/backend/optimizer/plan/createplan.c\n> @@ -25,6 +25,7 @@\n> #include \"nodes/extensible.h\"\n> #include \"nodes/makefuncs.h\"\n> #include \"nodes/nodeFuncs.h\"\n> +#include \"nodes/print.h\"\n> #include \"optimizer/clauses.h\"\n> #include \"optimizer/cost.h\"\n> #include \"optimizer/optimizer.h\"\n\nleft over from debugging?\n\n> @@ -68,6 +68,15 @@ static bool pathlist_is_reparameterizable_by_child(List *pathlist,\n> int\n> compare_path_costs(Path *path1, Path *path2, CostSelector criterion)\n> {\n> +\t/* Number of disabled nodes, if different, trumps all else. */\n> +\tif (unlikely(path1->disabled_nodes != path2->disabled_nodes))\n> +\t{\n> +\t\tif (path1->disabled_nodes < path2->disabled_nodes)\n> +\t\t\treturn -1;\n> +\t\telse\n> +\t\t\treturn +1;\n> +\t}\n> +\n> \tif (criterion == STARTUP_COST)\n> \t{\n> \t\tif (path1->startup_cost < path2->startup_cost)\n\nIs \"unlikely()\" really appropriate here (and elsewhere in the patch)? If \nyou run with enable_seqscan=off but have no indexes, you could take that \npath pretty often.\n\nIf this function needs optimizing, I'd suggest splitting it into two \nfunctions, one for comparing the startup cost and another for the total \ncost. Almost all callers pass a constant for that argument, so they \nmight as well call the correct function directly and avoid the branch \nfor that.\n\n> @@ -658,6 +704,20 @@ add_path_precheck(RelOptInfo *parent_rel,\n> \t\tPath\t *old_path = (Path *) lfirst(p1);\n> \t\tPathKeysComparison keyscmp;\n> \n> +\t\t/*\n> +\t\t * Since the pathlist is sorted by disabled_nodes and then by\n> +\t\t * total_cost, we can stop looking once we reach a path with more\n> +\t\t * disabled nodes, or the same number of disabled nodes plus a\n> +\t\t * total_cost larger than the new path's.\n> +\t\t */\n> +\t\tif (unlikely(old_path->disabled_nodes != disabled_nodes))\n> +\t\t{\n> +\t\t\tif (disabled_nodes < old_path->disabled_nodes)\n> +\t\t\t\tbreak;\n> +\t\t}\n> +\t\telse if (total_cost <= old_path->total_cost * STD_FUZZ_FACTOR)\n> +\t\t\tbreak;\n> +\n> \t\t/*\n> \t\t * We are looking for an old_path with the same parameterization (and\n> \t\t * by assumption the same rowcount) that dominates the new path on\n> @@ -666,39 +726,27 @@ add_path_precheck(RelOptInfo *parent_rel,\n> \t\t *\n> \t\t * Cost comparisons here should match compare_path_costs_fuzzily.\n> \t\t */\n> -\t\tif (total_cost > old_path->total_cost * STD_FUZZ_FACTOR)\n> +\t\t/* new path can win on startup cost only if consider_startup */\n> +\t\tif (startup_cost > old_path->startup_cost * STD_FUZZ_FACTOR ||\n> +\t\t\t!consider_startup)\n> \t\t{\n\nThe \"Cost comparisons here should match compare_path_costs_fuzzily\" \ncomment also applies to the check on total_cost that you moved up. Maybe \nmove up the comment to the beginning of the loop.\n\nv3-0004-Show-number-of-disabled-nodes-in-EXPLAIN-ANALYZE-.patch:\n\nIt's surprising that the \"Disable Nodes\" is printed even with the COSTS \nOFF option. It's handy for our regression tests, it's good to print them \nthere, but it feels wrong.\n\nCould we cram it into the \"cost=... rows=...\" part? And perhaps a marker \nthat a node was disabled would be more user friendly than showing the \ncumulative count? Something like:\n\npostgres=# set enable_material=off;\nSET\npostgres=# set enable_seqscan=off;\nSET\npostgres=# set enable_bitmapscan=off;\nSET\npostgres=# explain select * from foo, bar;\n QUERY PLAN \n\n------------------------------------------------------------------------------------\n Nested Loop (cost=0.15..155632.40 rows=6502500 width=8)\n -> Index Only Scan using foo_i_idx on foo (cost=0.15..82.41 \nrows=2550 width=4)\n -> Seq Scan on bar (cost=0.00..35.50 (disabled) rows=2550 width=4)\n(5 rows)\n\n-- \nHeikki Linnakangas\nNeon (https://neon.tech)\n\n\n\n","msg_date":"Tue, 2 Jul 2024 17:57:27 +0300","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Thanks for the review!\n\nOn Tue, Jul 2, 2024 at 10:57 AM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\n> v3-0001-Remove-grotty-use-of-disable_cost-for-TID-scan-pl.patch:\n>\n> +1, this seems ready for commit\n\nCool.\n\n> v3-0002-Rationalize-behavior-of-enable_indexscan-and-enab.patch:\n>\n> I fear this will break people's applications, if they are currently\n> forcing a sequential scan with \"set enable_indexscan=off\". Now they will\n> need to do \"set enable_indexscan=off; set enable_indexonlyscan=off\" for\n> the same effect. Maybe it's acceptable, disabling sequential scans to\n> force an index scan is much more common than the other way round.\n\nWell, I think it's pretty important that the GUC does what the name\nand documentation say it does. One could of course argue that we ought\nnot to have two different GUCs -- or perhaps even that we ought not to\nhave two different plan nodes -- and I think those arguments might be\nquite defensible. One could also argue for another interface, like a\nGUC enable_indexscan and a value that is a comma-separated list\nconsisting of plain, bitmap, and index-only, or\nnone/0/false/any/1/true -- and that might also be quite defensible.\nBut I don't think one can have a GUC called enable_indexscan and\nanother GUC called enable_indexonlyscan and argue that it's OK for the\nformer one to affect both kinds of scans. That's extremely confusing\nand, well, just plain wrong. I think this is a bug, and I'm not going\nto back-patch the fix precisely because of the considerations you\nnote, but I really don't think we can leave it like this. The current\nbehavior is so nonsensical that the code is essentially unmaintable,\nor at least I think it is.\n\n> v3-0003-Treat-number-of-disabled-nodes-in-a-path-as-a-sep.patch:\n>\n> > @@ -1318,6 +1342,12 @@ cost_tidscan(Path *path, PlannerInfo *root,\n> > startup_cost += path->pathtarget->cost.startup;\n> > run_cost += path->pathtarget->cost.per_tuple * path->rows;\n> >\n> > + /*\n> > + * There are assertions above verifying that we only reach this function\n> > + * either when enable_tidscan=true or when the TID scan is the only legal\n> > + * path, so it's safe to set disabled_nodes to zero here.\n> > + */\n> > + path->disabled_nodes = 0;\n> > path->startup_cost = startup_cost;\n> > path->total_cost = startup_cost + run_cost;\n> > }\n>\n> So if you have enable_tidscan=off, and have a query with \"WHERE CURRENT\n> OF foo\" that is planned with a TID scan, we set disable_nodes = 0? That\n> sounds wrong, shouldn't disable_nodes be 1 in that case? It probably\n> cannot affect the rest of the plan, given that \"WHERE CURRENT OF\" is\n> only valid in an UPDATE or DELETE, but still. At least it deserves a\n> better explanation in the comment.\n\nSo, right now, when the planner disregards enable_WHATEVER because it\nthinks it's the only way to implement something, it doesn't add\ndisable_cost. So, I made the patch not increment disabled_nodes in\nthat case. Maybe we want to rethink that choice at some point, but it\ndoesn't seem like a good idea to do it right now. I've found while\nworking on this stuff that it's super-easy to have seemingly innocuous\nchanges disturb regression test results, and I don't really want to\nhave a bunch of extra regression test changes that are due to\nrethinking things other than disable_cost -> disabled_nodes. So for\nnow I'd like to increment disabled_nodes in just the cases where we\ncurrently add disable_cost.\n\n> left over from debugging?\n\nYeah, will fix.\n\n> Is \"unlikely()\" really appropriate here (and elsewhere in the patch)? If\n> you run with enable_seqscan=off but have no indexes, you could take that\n> path pretty often.\n\nThat's true, but I think it's right to assume that's the uncommon\ncase. If we speed up planning for people who disabled sequential scans\nand slow it down for people running with a normal planner\nconfiguration, no one will thank us.\n\n> If this function needs optimizing, I'd suggest splitting it into two\n> functions, one for comparing the startup cost and another for the total\n> cost. Almost all callers pass a constant for that argument, so they\n> might as well call the correct function directly and avoid the branch\n> for that.\n\nThat's not a bad idea but seems like a separate patch.\n\n> The \"Cost comparisons here should match compare_path_costs_fuzzily\"\n> comment also applies to the check on total_cost that you moved up. Maybe\n> move up the comment to the beginning of the loop.\n\nWill have a look.\n\n> v3-0004-Show-number-of-disabled-nodes-in-EXPLAIN-ANALYZE-.patch:\n>\n> It's surprising that the \"Disable Nodes\" is printed even with the COSTS\n> OFF option. It's handy for our regression tests, it's good to print them\n> there, but it feels wrong.\n\nI'm open to doing what people think is best here. Although we're\nregarding them as part of the cost for purposes of how to compare\npaths, they're not unpredictable in the way that costs are, so I think\nthe current handling is defensible and, as you say, it's useful for\nthe regression tests. However, I'm not going to fight tooth and nail\nif people really want it the other way.\n\n> Could we cram it into the \"cost=... rows=...\" part? And perhaps a marker\n> that a node was disabled would be more user friendly than showing the\n> cumulative count? Something like:\n\nThe problem is that we'd have to derive that. What we actually know is\nthe disable count; to figure out whether the node itself was disabled,\nwe'd have to subtract the value for the underlying nodes back out.\nThat seems like it might be buggy or confusing.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 2 Jul 2024 13:24:44 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Tue, Jul 2, 2024 at 10:57 AM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\n>> I fear this will break people's applications, if they are currently\n>> forcing a sequential scan with \"set enable_indexscan=off\". Now they will\n>> need to do \"set enable_indexscan=off; set enable_indexonlyscan=off\" for\n>> the same effect. Maybe it's acceptable, disabling sequential scans to\n>> force an index scan is much more common than the other way round.\n\n> But I don't think one can have a GUC called enable_indexscan and\n> another GUC called enable_indexonlyscan and argue that it's OK for the\n> former one to affect both kinds of scans. That's extremely confusing\n> and, well, just plain wrong.\n\nFWIW, I disagree completely. I think it's entirely natural to\nconsider bitmap index scans to be a subset of index scans, so that\nenable_indexscan should affect both. I admit that the current set\nof GUCs doesn't let you force a bitmap scan over a plain one, but\nI can't recall many people complaining about that. I don't follow\nthe argument that this definition is somehow unmaintainable, either.\n\n>> Could we cram it into the \"cost=... rows=...\" part? And perhaps a marker\n>> that a node was disabled would be more user friendly than showing the\n>> cumulative count? Something like:\n\n> The problem is that we'd have to derive that.\n\nThe other problem is it'd break an awful lot of client code that knows\nthe format of those lines. (Sure, by now all such code should have\nbeen rewritten to look at JSON or other more machine-friendly output\nformats ... but since we haven't even done that in our own regression\ntests, we should know better than to assume other people have done it.)\n\nI'm not really convinced that we need to show anything about this.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 02 Jul 2024 13:40:38 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, Jul 2, 2024 at 1:40 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> FWIW, I disagree completely. I think it's entirely natural to\n> consider bitmap index scans to be a subset of index scans, so that\n> enable_indexscan should affect both. I admit that the current set\n> of GUCs doesn't let you force a bitmap scan over a plain one, but\n> I can't recall many people complaining about that. I don't follow\n> the argument that this definition is somehow unmaintainable, either.\n\nWell... but that's not what the GUC does either. Not now, and not with\nthe patch.\n\nWhat happens right now is:\n\n- If you set enable_indexscan=false, then disable_cost is added to the\ncost of index scan paths and the cost of index-only scan paths.\n\n- If you set enable_indexonlyscan=false, then index-only scan paths\nare not generated at all.\n\nBitmap scans are controlled by enable_bitmapscan.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 2 Jul 2024 13:54:02 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> What happens right now is:\n\n> - If you set enable_indexscan=false, then disable_cost is added to the\n> cost of index scan paths and the cost of index-only scan paths.\n\n> - If you set enable_indexonlyscan=false, then index-only scan paths\n> are not generated at all.\n\nHm. The first part of that seems pretty weird to me --- why don't\nwe simply not generate the paths at all? There is no case AFAIR\nwhere that would prevent us from generating a valid plan.\n\n(I do seem to recall that index-only paths are built on top of regular\nindex paths, so that there might be implementation issues with trying\nto build the former and not the latter. But you've probably looked\nat that far more recently than I.)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 02 Jul 2024 14:37:25 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, Jul 2, 2024 at 2:37 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Robert Haas <robertmhaas@gmail.com> writes:\n> > What happens right now is:\n>\n> > - If you set enable_indexscan=false, then disable_cost is added to the\n> > cost of index scan paths and the cost of index-only scan paths.\n>\n> > - If you set enable_indexonlyscan=false, then index-only scan paths\n> > are not generated at all.\n>\n> Hm. The first part of that seems pretty weird to me --- why don't\n> we simply not generate the paths at all? There is no case AFAIR\n> where that would prevent us from generating a valid plan.\n\nWell, yeah.\n\nWhat the patch does is: if you set either enable_indexscan=false or\nenable_indexonlyscan=false, then the corresponding path type is not\ngenerated, and the other is unaffected. To me, that seems like the\nlogical way to clean this up.\n\nOne could argue for other things, of course. And maybe those other\nthings are fine, if they're properly justified and documented.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 2 Jul 2024 15:28:38 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> What the patch does is: if you set either enable_indexscan=false or\n> enable_indexonlyscan=false, then the corresponding path type is not\n> generated, and the other is unaffected. To me, that seems like the\n> logical way to clean this up.\n\n> One could argue for other things, of course. And maybe those other\n> things are fine, if they're properly justified and documented.\n\n[ shrug... ] This isn't a hill that I'm prepared to die on.\nBut I see no good reason to change the very long-standing\nbehaviors of these GUCs.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 02 Jul 2024 15:36:56 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, Jul 2, 2024 at 3:36 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > One could argue for other things, of course. And maybe those other\n> > things are fine, if they're properly justified and documented.\n>\n> [ shrug... ] This isn't a hill that I'm prepared to die on.\n> But I see no good reason to change the very long-standing\n> behaviors of these GUCs.\n\nWell, I don't really know where to go from here. I mean, I think that\nthree committers (David, Heikki, yourself) have expressed some\nconcerns about changing the behavior. So maybe we shouldn't. But I\ndon't understand how it's reasonable to have two very similarly named\nGUCs behave (1) inconsistently with each other and (2) in a way that\ncannot be guessed from the documentation.\n\nI feel like we're just clinging to legacy behavior on the theory that\nsomebody, somewhere might be relying on it in some way, which they\ncertainly might be. But that doesn't seem like a great reason, either.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 2 Jul 2024 15:54:17 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> Well, I don't really know where to go from here. I mean, I think that\n> three committers (David, Heikki, yourself) have expressed some\n> concerns about changing the behavior. So maybe we shouldn't. But I\n> don't understand how it's reasonable to have two very similarly named\n> GUCs behave (1) inconsistently with each other and (2) in a way that\n> cannot be guessed from the documentation.\n\nIf the documentation isn't adequate, that's certainly an improvable\nsituation. It doesn't seem hard:\n\n- Enables or disables the query planner's use of index-scan plan\n- types. The default is <literal>on</literal>.\n+ Enables or disables the query planner's use of index-scan plan\n+ types (including index-only scans).\n+ The default is <literal>on</literal>.\n\nMore to the point, if we do change the longstanding meaning of this\nGUC, that will *also* require documentation work IMO.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 02 Jul 2024 16:43:41 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On 02/07/2024 22:54, Robert Haas wrote:\n> On Tue, Jul 2, 2024 at 3:36 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>>> One could argue for other things, of course. And maybe those other\n>>> things are fine, if they're properly justified and documented.\n>>\n>> [ shrug... ] This isn't a hill that I'm prepared to die on.\n>> But I see no good reason to change the very long-standing\n>> behaviors of these GUCs.\n> \n> Well, I don't really know where to go from here. I mean, I think that\n> three committers (David, Heikki, yourself) have expressed some\n> concerns about changing the behavior. So maybe we shouldn't. But I\n> don't understand how it's reasonable to have two very similarly named\n> GUCs behave (1) inconsistently with each other and (2) in a way that\n> cannot be guessed from the documentation.\n> \n> I feel like we're just clinging to legacy behavior on the theory that\n> somebody, somewhere might be relying on it in some way, which they\n> certainly might be. But that doesn't seem like a great reason, either.\n\nI agree the status quo is weird too. I'd be OK to break \nbackwards-compatibility if we can make it better.\n\nTom mentioned enable_bitmapscan, and it reminded me that the current \nbehavior with that is actually a bit annoying. I go through this pattern \nvery often when I'm investigating query plans:\n\n1. Hmm, let's see what this query plan looks like:\n\npostgres=# explain analyze select * from foo where i=10;\n QUERY PLAN \n\n----------------------------------------------------------------------------------------------------------------\n Index Scan using foo_i_idx on foo (cost=0.29..8.31 rows=1 width=36) \n(actual time=0.079..0.090 rows=2 loops=1)\n Index Cond: (i = 10)\n Planning Time: 2.220 ms\n Execution Time: 0.337 ms\n(4 rows)\n\n2. Ok, and how long would it take with a seq scan? Let's see:\n\npostgres=# set enable_indexscan=off;\nSET\npostgres=# explain analyze select * from foo where i=10;\n QUERY PLAN \n\n------------------------------------------------------------------------------------------------------------------\n Bitmap Heap Scan on foo (cost=4.30..8.31 rows=1 width=36) (actual \ntime=0.102..0.113 rows=2 loops=1)\n Recheck Cond: (i = 10)\n Heap Blocks: exact=2\n -> Bitmap Index Scan on foo_i_idx (cost=0.00..4.30 rows=1 width=0) \n(actual time=0.067..0.068 rows=2 loops=1)\n Index Cond: (i = 10)\n Planning Time: 0.211 ms\n Execution Time: 0.215 ms\n(7 rows)\n\n3. Oh right, bitmap scan, I forgot about that one. Let's disable that too:\n\npostgres=# set enable_bitmapscan=off;\nSET\npostgres=# explain analyze select * from foo where i=10;\n QUERY PLAN \n\n--------------------------------------------------------------------------------------------------\n Seq Scan on foo (cost=0.00..1862.00 rows=1 width=36) (actual \ntime=0.042..39.226 rows=2 loops=1)\n Filter: (i = 10)\n Rows Removed by Filter: 109998\n Planning Time: 0.118 ms\n Execution Time: 39.272 ms\n(5 rows)\n\nI would be somewhat annoyed if we add another step to that, to also \ndisable index-only scans separately. It would be nice if \nenable_indexscan=off would also disable bitmap scans, that would \neliminate one step from the above. Almost always when I want to disable \nindex scans, I really want to disable the use of the index altogether. \nThe problem then of course is, how do you force a bitmap scan without \nallowing other index scans, when you want to test them both?\n\nIt almost feels like we should have yet another GUC to disable index \nscans, index-only scans and bitmap index scans. \"enable_indexes=off\" or \nsomething.\n\n-- \nHeikki Linnakangas\nNeon (https://neon.tech)\n\n\n\n","msg_date":"Wed, 3 Jul 2024 00:39:43 +0300","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Heikki Linnakangas <hlinnaka@iki.fi> writes:\n> 3. Oh right, bitmap scan, I forgot about that one. Let's disable that too:\n\nYeah, I've hit that too, although more often (for me) it's the first\nchoice of plan. In any case, it usually takes more than one change\nto get to a seqscan.\n\n> It almost feels like we should have yet another GUC to disable index \n> scans, index-only scans and bitmap index scans. \"enable_indexes=off\" or \n> something.\n\nThere's something to be said for that idea. Breaking compatibility is\na little easier to stomach if there's a clear convenience win, and\nthis'd offer that.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 02 Jul 2024 17:49:09 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Tue, Jul 2, 2024 at 5:39 PM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\n> I would be somewhat annoyed if we add another step to that, to also\n> disable index-only scans separately. It would be nice if\n> enable_indexscan=off would also disable bitmap scans, that would\n> eliminate one step from the above. Almost always when I want to disable\n> index scans, I really want to disable the use of the index altogether.\n> The problem then of course is, how do you force a bitmap scan without\n> allowing other index scans, when you want to test them both?\n>\n> It almost feels like we should have yet another GUC to disable index\n> scans, index-only scans and bitmap index scans. \"enable_indexes=off\" or\n> something.\n\nThis is an interesting idea, and it seems like it could be convenient.\nHowever, the fact that it's so non-orthogonal is definitely not great.\nOne problem I've had with going through regression tests that rely on\nthe enable_* GUCs is that it's often not quite clear what values all\nof those GUCs have at a certain point in the test file, because the\nstatements that set them may be quite a bit higher up in the file and\nsome changes may also have been rolled back. I've found recently that\nthe addition of EXPLAIN (SETTINGS) helps with this quite a bit,\nbecause you can adjust the .sql file to use that option and then see\nwhat shows up in the output file. Still, it's annoying, and the same\nissue could occur in any other situation where you're using these\nGUCs. It's just more confusing when there are multiple ways of turning\nsomething off.\n\nWould we consider merging enable_indexscan, enable_indexonlyscan, and\nenable_bitmapscan into something like:\n\nenable_indexes = on | off | { plain | indexonly | bitmap } [, ...]\n\nI feel like that would solve the usability concern that you raise here\nwhile also (1) preserving orthogonality and (2) reducing the number of\nGUCs rather than first increasing it. When I first joined the project\nthere were a decent number of enable_* GUCs, but there's way more now.\nSome of them are a little random (which is a conversation for another\nday) but just cutting down on the number seems like it might not be\nsuch a bad idea.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 3 Jul 2024 09:01:16 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, 3 Jul 2024 at 09:49, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Heikki Linnakangas <hlinnaka@iki.fi> writes:\n> > 3. Oh right, bitmap scan, I forgot about that one. Let's disable that too:\n>\n> Yeah, I've hit that too, although more often (for me) it's the first\n> choice of plan. In any case, it usually takes more than one change\n> to get to a seqscan.\n\nI commonly hit this too.\n\nI think the current behaviour is born out of the fact that we don't\nproduce both an Index Scan and an Index Only Scan for the same index.\nWe'll just make the IndexPath an index only scan, if possible based\non:\n\nindex_only_scan = (scantype != ST_BITMAPSCAN &&\n check_index_only(rel, index));\n\nThe same isn't true for Bitmap Index Scans. We'll create both\nIndexPaths and BitmapHeapPaths and let them battle it out in\nadd_path().\n\nI suspect this is why it's been coded that enable_indexscan also\ndisables Index Only Scans. Now, of course, it could work another way,\nbut I also still think that doing so is changing well-established\nbehaviour that I don't recall anyone ever complaining about besides\nRobert. Robert's complaint seems to have originated from something he\nnoticed while hacking on code rather than actually using the database\nfor something. I think the argument for changing it should have less\nweight due to that.\n\nI understand that we do have inconsistencies around this stuff. For\nexample, enable_sort has no influence on Incremental Sorts like\nenable_indexscan has over Index Only Scan. That might come from the\nfact that we used to, up until a couple of releases ago, produce both\nsort path types and let them compete in add_path(). That's no longer\nthe case, we now just do incremental sort when we can, just like we do\nIndex Only Scans when we can. Despite those inconsistencies, I\nwouldn't vote for changing either of them to align with the other. It\njust feels too long-established behaviour to be messing with.\n\nI feel it might be best to move this patch to the back of the series\nor just drop it for now as it seems to be holding up the other stuff\nfrom moving forward, and that stuff looks useful and worth changing.\n\nDavid\n\n\n","msg_date":"Thu, 4 Jul 2024 13:29:39 +1200","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"OK, here's a new patch version. I earlier committed the refactoring to\navoid using disable_cost to force WHERE CURRENT OF to be implemented\nby a TID scan. In this version, I've dropped everything related to\nreworking enable_indexscan or any other enable_* GUC. Hence, this\nversion of the patch set just focuses on adding the count of disabled\nnodes and removing the use of disable_cost. In addition to dropping\nthe controversial patches, I've also found and squashed a few bugs in\nthis version.\n\nBehavior: With the patch, whenever an enable_* GUC would cause\ndisable_cost to be added, disabled_nodes is incremented instead. There\nis one remaining use of disable_cost which is not triggered by an\nenable_* GUC but by the desire to avoid plans that we think will\noverflow work_mem. I welcome thoughts on what to do about that case;\nfor now, I do nothing. As before, 0001 adds the disabled_nodes field\nto paths and 0002 adds it to plans. I think we could plausibly commit\nonly 0001, both patches separately, or both patches squashed.\n\nNotes:\n\n- I favor committing both patches. Tom stated that he didn't think\nthat we needed to show anything related to disabled nodes, and that\ncould be true. However, today, you can tell which nodes are disabled\nas long as you print out the costs; if we don't propagate disabled\nnodes into the plan and print them out, that will no longer be\npossible. I found working on the patches that it was really hard to\ndebug the patch set without this, so my guess is that we'll find not\nhaving it pretty annoying, but we can also just commit 0001 for\nstarters and see how long it takes for the lack of 0002 to become\nannoying. If the answer is \"infinite time,\" that's cool; if it isn't,\nwe can reconsider committing 0002.\n\n- If we do commit 0002, I think it's a good idea to have the number of\ndisabled nodes displayed even with COSTS OFF, because it's stable, and\nit's pretty useful to be able to see this in the regression output. I\nhave found while working on this that I often need to adjust the .sql\nfiles to say EXPLAIN (COSTS ON) instead of EXPLAIN (COSTS OFF) in\norder to understand what's happening. Right now, there's no real\nalternative because costs aren't stable, but disabled-node counts\nshould be stable, so I feel this would be a step forward. Apart from\nthat, I also think it's good for features to have regression test\ncoverage, and since we use COSTS OFF everywhere or at least nearly\neverywhere in the regression test, if we don't print out the disabled\nnode counts when COSTS OFF is used, then we don't cover that case in\nour tests. Bummer.\n\nRegression test changes in 0001:\n\n- btree_index.sql executes a query \"select proname from pg_proc where\nproname ilike 'ri%foo' order by 1\" with everything but bitmap scans\ndisabled. Currently, that produces an index-only scan; with the patch,\nit produces a sort over a sequential scan. That's a little odd,\nbecause the test seems to be aimed at demonstrating that we can use a\nbitmap scan, and it doesn't, because we apparently can't. But, why\ndoes the patch change the plan?\nAt least on my machine, the index-only scan is significantly more\ncostly than the sequential scan. I think what's happening here is that\nwhen you add disable_cost to the cost of both paths, they compare\nfuzzily the same; without that, the cheaper one wins.\n\n- select_parallel.out executes a query with sequential scans disabled\nbut tenk2 must nevertheless be sequential-scanned. With the patch,\nthat changes to a parallel sequential scan. I think the explanation\nhere is the same as in the preceding case.\n\n- horizons.spec currently sets enable_seqscan=false,\nenable_indexscan=false, and enable_bitmapscan=false. I suspect that\nAndres thought that this would force the use of an index-only scan,\nsince nothing sets enable_indexonlyscan=false. But as discussed\nupthread, that is not true. Instead everything is disabled. For the\nsame reasons as in the previous two examples, this caused an\nassortment of plan changes which in turn caused the test to fail to\ntest what it was intended to test. So I removed enable_indexscan=false\nfrom the spec file, and now it gets index-only scans everywhere again,\nas desired.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com","msg_date":"Wed, 31 Jul 2024 12:23:22 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Thu, 1 Aug 2024 at 04:23, Robert Haas <robertmhaas@gmail.com> wrote:\n> OK, here's a new patch version.\n\nI think we're going down the right path here.\n\nI've reviewed both patches, here's what I noted down during my review:\n\n0. I've not seen any mention so far about postgres_fdw's\nuse_remote_estimate. Maybe changing the costs is fixing an issue that\nexisted before. I'm just not 100% sure on that.\n\nConsider:\n\nCREATE EXTENSION postgres_fdw;\n\nDO $d$\n BEGIN\n EXECUTE $$CREATE SERVER loopback FOREIGN DATA WRAPPER postgres_fdw\n OPTIONS (use_remote_estimate 'true',\n dbname '$$||current_database()||$$',\n port '$$||current_setting('port')||$$'\n )$$;\n END;\n$d$;\nCREATE USER MAPPING FOR CURRENT_USER SERVER loopback;\n\ncreate table t (a int);\ncreate foreign table ft (a int) server loopback OPTIONS (table_name 't');\n\nalter system set enable_seqscan=0;\nselect pg_Reload_conf();\nset enable_seqscan=1;\nexplain select * from ft;\n\npatched:\n Foreign Scan on ft (cost=100.00..671.00 rows=2550 width=4)\n\nmaster:\n Foreign Scan on ft (cost=10000000100.00..10000000671.00 rows=2550 width=4)\n\nI kinda think that might be fixing an issue that I don't recall being\nreported before. I think we shouldn't really care that much about what\nnodes are disabled on the remote server and not having disabled_cost\napplied to that gives us that.\n\n1. The final sentence of the function header comment needs to be\nupdated in estimate_path_cost_size().\n\n2. Does cost_tidscan() need to update the header comment to say\ntidquals must not be empty?\n\n3. final_cost_nestloop() seems to initially use the disabled_nodes\nfrom initial_cost_nestloop() but then it goes off and calculates it\nagain itself. One of these seems redundant. The \"We could include\ndisable_cost in the preliminary estimate\" comment explains why it was\noriginally left to final_cost_nestloop(), so maybe worth sticking to\nthat? I don't quite know the full implications, but it does not seem\nworth risking a behaviour change here.\n\n4. I wonder if it's worth doing a quick refactor of the code in\ninitial_cost_mergejoin() to get rid of the duplicate code in the \"if\n(outersortkeys)\" and \"if (innersortkeys)\" branches. It seems ok to do\nouter_path = &sort_path. Likewise for inner_path.\n\n5. final_cost_hashjoin() does the same thing as #3\n\n6. createplan.c adds #include \"nodes/print.h\" but doesn't seem to add\nany code that might use anything in there.\n\n7. create_lockrows_path() needs to propagate disabled_nodes.\n\ncreate table a (a int);\nset enable_seqscan=0;\n\nexplain select * from a for update limit 1;\n\n Limit (cost=0.00..0.02 rows=1 width=10)\n -> LockRows (cost=0.00..61.00 rows=2550 width=10)\n -> Seq Scan on a (cost=0.00..35.50 rows=2550 width=10)\n Disabled Nodes: 1\n(4 rows)\n\n\nexplain select * from a limit 1;\n\n Limit (cost=0.00..0.01 rows=1 width=4)\n Disabled Nodes: 1\n -> Seq Scan on a (cost=0.00..35.50 rows=2550 width=4)\n Disabled Nodes: 1\n(4 rows)\n\n8. There's something weird with CTEs too.\n\ncreate table b(a int);\nset enable_sort=0;\n\nPatched:\n\nexplain with cte as materialized (select * from b order by a) select *\nfrom cte order by a desc;\n\n Sort (cost=381.44..387.82 rows=2550 width=4)\n Disabled Nodes: 1\n Sort Key: cte.a DESC\n CTE cte\n -> Sort (cost=179.78..186.16 rows=2550 width=4)\n Disabled Nodes: 1\n Sort Key: b.a\n -> Seq Scan on b (cost=0.00..35.50 rows=2550 width=4)\n -> CTE Scan on cte (cost=0.00..51.00 rows=2550 width=4)\n(9 rows)\n\nmaster:\n\nexplain with cte as materialized (select * from a order by a) select *\nfrom cte order by a desc;\n\n Sort (cost=20000000381.44..20000000387.82 rows=2550 width=4)\n Sort Key: cte.a DESC\n CTE cte\n -> Sort (cost=10000000179.78..10000000186.16 rows=2550 width=4)\n Sort Key: a.a\n -> Seq Scan on a (cost=0.00..35.50 rows=2550 width=4)\n -> CTE Scan on cte (cost=0.00..51.00 rows=2550 width=4)\n(7 rows)\n\nI'd expect the final sort to have disabled_nodes == 2 since\ndisabled_cost has been added twice in master.\n\n9. create_set_projection_path() needs to propagate disabled_nodes too:\n\nexplain select b from (select a,generate_series(1,2) as b from b) a limit 1;\n\n Limit (cost=0.00..0.03 rows=1 width=4)\n -> Subquery Scan on a (cost=0.00..131.12 rows=5100 width=4)\n -> ProjectSet (cost=0.00..80.12 rows=5100 width=8)\n -> Seq Scan on b (cost=0.00..35.50 rows=2550 width=0)\n Disabled Nodes: 1\n\n10. create_setop_path() needs to propagate disabled_nodes.\n\nexplain select * from b except select * from b limit 1;\n\n Limit (cost=0.00..0.80 rows=1 width=8)\n -> HashSetOp Except (cost=0.00..160.25 rows=200 width=8)\n -> Append (cost=0.00..147.50 rows=5100 width=8)\n Disabled Nodes: 2\n -> Subquery Scan on \"*SELECT* 1\" (cost=0.00..61.00\nrows=2550 width=8)\n Disabled Nodes: 1\n -> Seq Scan on b (cost=0.00..35.50 rows=2550 width=4)\n Disabled Nodes: 1\n -> Subquery Scan on \"*SELECT* 2\" (cost=0.00..61.00\nrows=2550 width=8)\n Disabled Nodes: 1\n -> Seq Scan on b b_1 (cost=0.00..35.50 rows=2550 width=4)\n Disabled Nodes: 1\n(12 rows)\n\n11. create_modifytable_path() needs to propagate disabled_nodes.\n\nexplain with cte as (update b set a = a+1 returning *) select * from\ncte limit 1;\n\n Limit (cost=41.88..41.90 rows=1 width=4)\n CTE cte\n -> Update on b (cost=0.00..41.88 rows=2550 width=10)\n -> Seq Scan on b (cost=0.00..41.88 rows=2550 width=10)\n Disabled Nodes: 1\n -> CTE Scan on cte (cost=0.00..51.00 rows=2550 width=4)\n(6 rows)\n\n12. For the 0002 patch, I do agree that having this visible in EXPLAIN\nis a must. I'd much rather see: Disabled: true/false. And just\ndisplay this when the disabled_nodes is greater than the sum of the\nsubpaths. That might be much more complex to implement, but it's\ngoing to make it much easier to track down the disabled nodes in very\nlarge plans.\n\nDavid\n\n\n","msg_date":"Thu, 1 Aug 2024 14:01:07 +1200","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, Jul 31, 2024 at 10:01 PM David Rowley <dgrowleyml@gmail.com> wrote:\n> I've reviewed both patches, here's what I noted down during my review:\n\nThanks.\n\n> 0. I've not seen any mention so far about postgres_fdw's\n> use_remote_estimate. Maybe changing the costs is fixing an issue that\n> existed before. I'm just not 100% sure on that.\n>\n> patched:\n> Foreign Scan on ft (cost=100.00..671.00 rows=2550 width=4)\n>\n> master:\n> Foreign Scan on ft (cost=10000000100.00..10000000671.00 rows=2550 width=4)\n>\n> I kinda think that might be fixing an issue that I don't recall being\n> reported before. I think we shouldn't really care that much about what\n> nodes are disabled on the remote server and not having disabled_cost\n> applied to that gives us that.\n\nHmm, I think it's subjective which behavior is better. If somebody\nthought the new behavior was worse, they might want the remote side's\ncount of disabled nodes to be propagated to the local side, but I'm\ndisinclined to go there. My guess is that it doesn't matter much\neither way what we do here, so I'd rather not add more code.\n\n> 1. The final sentence of the function header comment needs to be\n> updated in estimate_path_cost_size().\n\nFixed.\n\n> 2. Does cost_tidscan() need to update the header comment to say\n> tidquals must not be empty?\n\nIMHO, no. The assertions I added to that function were intended as\ndocumentation of what that function was already assuming about the\nbehavior of its caller. I had to trace through the logic in tidpath.c\nfor quite a while to understand why cost_tidscan() was not completely\nbroken. To spare the next person the trouble of working that out, I\nadded assertions. Now we could additionally add commentary in English\nthat restates what the assertions already say, but I feel like having\nthe assertions is good enough. If somebody ever whacks around\ntidpath.c such that these assertions start failing, I think it will be\nfairly clear to them that they either need to revert their changes in\ntidpath.c or upgrade the logic in this function to cope.\n\n> 3. final_cost_nestloop() seems to initially use the disabled_nodes\n> from initial_cost_nestloop() but then it goes off and calculates it\n> again itself. One of these seems redundant.\n\nOops. Fixed.\n\n> The \"We could include\n> disable_cost in the preliminary estimate\" comment explains why it was\n> originally left to final_cost_nestloop(), so maybe worth sticking to\n> that? I don't quite know the full implications, but it does not seem\n> worth risking a behaviour change here.\n\nI don't really see how there could be a behavior change here, unless\nthere's a bug. Dealing with the enable_* flags in initial_cost_XXX\nrather than final_cost_XXX could be better or worse from a performance\nstandpoint and it could make for cleaner or less clean code, but the\nuser-facing behavior should be identical unless there are bugs.\n\nThe reason why I changed this is because of the logic in\nadd_path_precheck(): it exits early as soon as it sees a path whose\ntotal cost is greater than the cost of the proposed new path. Since\nthe patch's aim is to treat disabled_node as a high-order component of\nthe cost, we need to make the same decision by comparing the count of\ndisabled_nodes first and then if that is equal, we need to compare the\ntotal_cost. We can't do that if we don't have the count of\ndisabled_nodes for the proposed new path.\n\nI think this may be a bit hard to understand, so let me give a\nconcrete example. Suppose we're planning some join where one side can\nonly be planned with a sequential scan and sequential scans are\ndisabled. We have ten paths in the path list and they have costs of\n1e10+100, 1e10+200, ..., 1e10+1000. Now add_path_precheck() is asked\nto consider a new path where there is a disabled node on BOTH sides of\nthe join -- the one side has the disabled sequential scan, but now the\nother side also has something disabled, so the cost is let's say\n2e10+79. add_path_precheck() can see at once that this path is a\nloser: it can't possibly dominate any path that already exists,\nbecause it costs more than any of them. But when you take disable_cost\nout, things look quite different. Now you have a proposed path with a\ntotal_cost of 79 and a path list with costs of 100, ..., 1000. If\nyou're not allowed to know anything about disabled_nodes, the new path\nlooks like it might be valuable. You might decide to construct it and\ntry inserting into the pathlist, which will end up being useless, and\neven if you don't, you're going to compare its pathkeys and\nparameterization to each of the 10 existing paths before giving up.\nBummer.\n\nSo, to avoid getting much stupider than it is currently,\nadd_path_precheck() needs a preliminary estimate of the number of\ndisabled nodes just like it needs a preliminary estimate of the total\ncost. And to avoid regressions, that estimate needs to be pretty good.\nA naive estimate would be to just add up the number of disabled_nodes\non the inner and outer paths, but that would be a regression in the\nmerge-join case, because initial_cost_mergejoin() calls cost_sort()\nfor the inner and outer sides and that will add disable_cost if sorts\nare disabled. If you didn't take the effect of cost_sort() into\naccount, you might think that your number of disabled_nodes was going\nto be substantially lower than it really would be, leading to wasted\nwork as described in the last paragraph. Plus, since\ninitial_cost_mergejoin() is incurring the overhead of calling\ncost_sort() anyway to get the total cost numbers anyway, it would be\nsilly not to save the count of disabled nodes: if we did, we'd have to\nredo the cost_sort() call in final_cost_mergejoin(), which would be\nexpensive.\n\nIf we wanted to make our estimate of the # of disabled nodes exactly\ncomparable to what we now do with disable_cost, we would postpone if\n(!enable_WHATEVERjoin) ++disabled_nodes to the final_cost_XXX\nfunctions and do all of the other accounting related to disabled nodes\nat the initial_cost_XXX phase. But I do not like that approach.\nPostponing one trivial portion of the disabled_nodes calculation to a\nlater time won't save any significant number of CPU cycles, but it\nmight confuse people reading the code. You then have to know that the\ndisabled_nodes count that gets passed to final_cost_XXX is not yet the\nfinal count, but that you may still need to add 1 for the join itself\n(but not for the implicit sorts that the join requires, which have\nalready been accounted for). That's the kind of odd definition that\nbreeds bugs. Besides, it's not as if moving that tiny bit of logic to\nthe initial_cost_XXX functions has no upside: it could allow\nadd_path_precheck() to exit earlier, thus saving cycles.\n\n(For the record, the explanation above took about 3 hours to write, so\nI hope it's managed to be both correct and convincing. This stuff is\nreally complicated.)\n\n> 4. I wonder if it's worth doing a quick refactor of the code in\n> initial_cost_mergejoin() to get rid of the duplicate code in the \"if\n> (outersortkeys)\" and \"if (innersortkeys)\" branches. It seems ok to do\n> outer_path = &sort_path. Likewise for inner_path.\n\nI don't think that's better.\n\n> 5. final_cost_hashjoin() does the same thing as #3\n\nArgh. Fixed.\n\n> 6. createplan.c adds #include \"nodes/print.h\" but doesn't seem to add\n> any code that might use anything in there.\n\nFixed.\n\n> 8. There's something weird with CTEs too.\n>\n> I'd expect the final sort to have disabled_nodes == 2 since\n> disabled_cost has been added twice in master.\n\nRight now, disabled node counts don't propagate through SubPlans (see\nSS_process_ctes). Maybe that needs to be changed, but aside from\nlooking weird, does it do any harm?\n\n> 7. create_lockrows_path() needs to propagate disabled_nodes.\n> 9. create_set_projection_path() needs to propagate disabled_nodes too:\n> 10. create_setop_path() needs to propagate disabled_nodes.\n> 11. create_modifytable_path() needs to propagate disabled_nodes.\n\nI changed all of these, but I think these examples only establish that\nthose nodes DO NOT propagate disabled_nodes, not that they need to. If\nwe're past the point of making any choices based on costs, then\nmaintaining disabled_nodes or not doing so won't affect correctness.\nThat's not to say these aren't good to tidy up, and some of them may\nwell be bugs, but I don't think your test cases prove that. What\nprimarily matters is whether the enable_BLAH GUCs get respected; the\nexact contents of the EXPLAIN output are somewhat more arguable.\n\n> 12. For the 0002 patch, I do agree that having this visible in EXPLAIN\n> is a must. I'd much rather see: Disabled: true/false. And just\n> display this when the disabled_nodes is greater than the sum of the\n> subpaths. That might be much more complex to implement, but it's\n> going to make it much easier to track down the disabled nodes in very\n> large plans.\n\nI think it's going to be very unpleasant if we have the planner add\nthings up and then try to have EXPLAIN subtract them back out again.\nOne problem with that is that all of the test cases where you just\nshowed disabled_nodes not propagating upward wouldn't actually show\nanything any more, because disabled_nodes would not have been greater\nin the parent than in the child. So those are oversights in the code\nthat are easy to spot now but would become hard to spot with this\nimplementation. Another problem is that the EXPLAIN code itself could\ncontain bugs, or slightly more broadly, get out of sync with the logic\nthat decides what to add up. It won't be obvious what's happening:\nsome node that is actually disabled just won't appear to be, or the\nother way around, and it will be hard to understand what happened,\nbecause you won't be able to see the raw counts of disabled nodes that\nwould allow you to deduce where the error actually is.\n\nOne idea that occurs to me is to store TWO counts in each path node\nand each plan node: the count of self-exclusive disabled nodes, and\nthe count of self-include disabled nodes. Then explain can just test\nif they are different. If the answer is 1, the node is disabled; if 0,\nit's enabled; if anything else, there's a bug (and it could print the\ndelta, or each value separately, to help localize such bugs). The\nproblem with that is that it eats up more space in\nperformance-critical data structures, but perhaps that's OK: I don't\nknow.\n\nAnother thought is that right now you just see the disable_cost values\nadded up with the rest of the cost. So maybe propagating upward is not\nreally such a bad behavior; it's what we have now.\n\nThis point probably needs more thought and discussion, but I'm out of\ntime to work on this for today, and out of mental energy too. So for\nnow here's v5 as I have it.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com","msg_date":"Thu, 1 Aug 2024 14:03:18 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, 2 Aug 2024 at 06:03, Robert Haas <robertmhaas@gmail.com> wrote:\n> I think this may be a bit hard to understand, so let me give a\n> concrete example. Suppose we're planning some join where one side can\n> only be planned with a sequential scan and sequential scans are\n> disabled. We have ten paths in the path list and they have costs of\n> 1e10+100, 1e10+200, ..., 1e10+1000. Now add_path_precheck() is asked\n> to consider a new path where there is a disabled node on BOTH sides of\n> the join -- the one side has the disabled sequential scan, but now the\n> other side also has something disabled, so the cost is let's say\n> 2e10+79. add_path_precheck() can see at once that this path is a\n> loser: it can't possibly dominate any path that already exists,\n> because it costs more than any of them. But when you take disable_cost\n> out, things look quite different. Now you have a proposed path with a\n> total_cost of 79 and a path list with costs of 100, ..., 1000. If\n> you're not allowed to know anything about disabled_nodes, the new path\n> looks like it might be valuable. You might decide to construct it and\n> try inserting into the pathlist, which will end up being useless, and\n> even if you don't, you're going to compare its pathkeys and\n> parameterization to each of the 10 existing paths before giving up.\n> Bummer.\n\nOK, so it sounds like you'd like to optimise this code so that the\nplanner does a little less work when node types are disabled. The\nexisting comment does mention explicitly that we don't want to do\nthat:\n\n/*\n* We could include disable_cost in the preliminary estimate, but that\n* would amount to optimizing for the case where the join method is\n* disabled, which doesn't seem like the way to bet.\n*/\n\nAs far as I understand it from reading the comments in that file, I\nsee no offer of guarantees that the initial cost will be cheaper than\nthe final cost. So what you're proposing could end up rejecting paths\nbased on initial cost where the final cost might end up being the\ncheapest path. Imagine you're considering a Nested Loop and a Hash\nJoin, both of which are disabled. Merge Join is unavailable as the\njoin column types are not sortable. If the hash join costs 99 and the\ninitial nested loop costs 110, but the final nested loop ends up\ncosting 90, then the nested loop could be rejected before we even get\nto perform the final cost for it. The current code will run\nfinal_cost_nestloop() and find that 90 is cheaper than 99, whereas\nwhat you want to do is stop bothering with nested loop when we see the\ninitial cost come out at 110.\n\nPerhaps it's actually fine if the initial costs are always less than\nthe final costs as, if that's the case, we won't ever reject any paths\nbased on the initial cost that we wouldn't anyway based on the final\ncost. However, since there does not seem to be any comments mentioning\nthis guarantee and if you're just doing this to squeeze more\nperformance out of the planner, it seems risky to do for that reason\nalone.\n\nI'd say if you want to do this, you should be justifying it on its own\nmerit with some performance numbers and some evidence that we don't\nproduce inferior plans as a result. But per what I quoted above,\nyou're not doing that, you're doing this as a performance\noptimisation.\n\nI'm not planning on pushing this any further. I've just tried to\nhighlight that there's the possibility of a behavioural change. You're\nclaiming there isn't one. I claim there is.\n\nDavid\n\n\n","msg_date":"Fri, 2 Aug 2024 15:34:16 +1200","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Thu, Aug 1, 2024 at 11:34 PM David Rowley <dgrowleyml@gmail.com> wrote:\n> I'm not planning on pushing this any further. I've just tried to\n> highlight that there's the possibility of a behavioural change. You're\n> claiming there isn't one. I claim there is.\n\nI don't know what to tell you. The original version of the patch\ndidn't change this stuff, and the result did not work. So I looked\ninto the problem and fixed it. I may have done that wrongly, or there\nmay be debatable points, but it seems like your argument is\nessentially that I shouldn't have done any of this and I should just\ntake it all back out, and I know that doesn't work because it's the\nfirst thing I tried.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 2 Aug 2024 08:17:16 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Sat, 3 Aug 2024 at 00:17, Robert Haas <robertmhaas@gmail.com> wrote:\n>\n> On Thu, Aug 1, 2024 at 11:34 PM David Rowley <dgrowleyml@gmail.com> wrote:\n> > I'm not planning on pushing this any further. I've just tried to\n> > highlight that there's the possibility of a behavioural change. You're\n> > claiming there isn't one. I claim there is.\n>\n> I don't know what to tell you. The original version of the patch\n> didn't change this stuff, and the result did not work. So I looked\n> into the problem and fixed it. I may have done that wrongly, or there\n> may be debatable points, but it seems like your argument is\n> essentially that I shouldn't have done any of this and I should just\n> take it all back out, and I know that doesn't work because it's the\n> first thing I tried.\n\nI've just read what you wrote again and I now realise something I didn't before.\n\nI now think neither of us got it right. I now think what you'd need to\ndo to be aligned to the current behaviour is have\ninitial_cost_nestloop() add the disabled_nodes for the join's subnodes\n*only* and have final_cost_nestloop() add the additional\ndisabled_nodes if enable_nestloop = off. That way you maintain the\nexisting behaviour of not optimising for disabled node types and don't\nrisk plan changes if the final cost comes out cheaper than the initial\ncost.\n\nDavid\n\n\n","msg_date":"Sat, 3 Aug 2024 01:13:36 +1200","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Aug 2, 2024 at 9:13 AM David Rowley <dgrowleyml@gmail.com> wrote:\n> I now think neither of us got it right. I now think what you'd need to\n> do to be aligned to the current behaviour is have\n> initial_cost_nestloop() add the disabled_nodes for the join's subnodes\n> *only* and have final_cost_nestloop() add the additional\n> disabled_nodes if enable_nestloop = off. That way you maintain the\n> existing behaviour of not optimising for disabled node types and don't\n> risk plan changes if the final cost comes out cheaper than the initial\n> cost.\n\nAll three initial_cost_XXX functions have a comment that says \"This\nmust quickly produce lower-bound estimates of the path's startup and\ntotal costs,\" i.e. the final cost should never be cheaper. I'm pretty\nsure that it was the design intention here that no path ever gets\nrejected at the initial cost stage that would have been accepted at\nthe final cost stage.\n\n(You can also see, as a matter of implementation, that they extract\nthe startup_cost and run_cost from the workspace and then add to those\nvalues.)\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 2 Aug 2024 12:04:24 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Fri, Aug 2, 2024 at 9:13 AM David Rowley <dgrowleyml@gmail.com> wrote:\n>> ... That way you maintain the\n>> existing behaviour of not optimising for disabled node types and don't\n>> risk plan changes if the final cost comes out cheaper than the initial\n>> cost.\n\n> All three initial_cost_XXX functions have a comment that says \"This\n> must quickly produce lower-bound estimates of the path's startup and\n> total costs,\" i.e. the final cost should never be cheaper. I'm pretty\n> sure that it was the design intention here that no path ever gets\n> rejected at the initial cost stage that would have been accepted at\n> the final cost stage.\n\nThat absolutely is the expectation, and we'd better be careful not\nto break it.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 02 Aug 2024 12:51:12 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Aug 2, 2024 at 12:51 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> That absolutely is the expectation, and we'd better be careful not\n> to break it.\n\nI have every intention of not breaking it. :-)\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 2 Aug 2024 12:53:16 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Aug 2, 2024 at 12:53 PM Robert Haas <robertmhaas@gmail.com> wrote:\n> On Fri, Aug 2, 2024 at 12:51 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > That absolutely is the expectation, and we'd better be careful not\n> > to break it.\n>\n> I have every intention of not breaking it. :-)\n\nI went ahead and committed these patches. I know there's some debate\nover whether we want to show the # of disabled nodes and if so whether\nit should be controlled by COSTS, and I suspect I haven't completely\nallayed David's concerns about the initial_cost_XXX functions although\nI think that I did the right thing. But, I don't have the impression\nthat anyone is desperately opposed to the basic concept, so I think it\nmakes sense to put these into the tree and see what happens. We have\nquite a bit of time left in this release cycle to uncover bugs, hear\nfrom users or other developers, etc. about what problems there may be\nwith this. If we end up deciding to reverse course or need to fix a\nbunch of stuff, so be it, but let's see what the feedback is.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 21 Aug 2024 10:29:23 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, 31 Jul 2024 at 18:23, Robert Haas <robertmhaas@gmail.com> wrote:\n> - If we do commit 0002, I think it's a good idea to have the number of\n> disabled nodes displayed even with COSTS OFF, because it's stable, and\n> it's pretty useful to be able to see this in the regression output. I\n> have found while working on this that I often need to adjust the .sql\n> files to say EXPLAIN (COSTS ON) instead of EXPLAIN (COSTS OFF) in\n> order to understand what's happening. Right now, there's no real\n> alternative because costs aren't stable, but disabled-node counts\n> should be stable, so I feel this would be a step forward. Apart from\n> that, I also think it's good for features to have regression test\n> coverage, and since we use COSTS OFF everywhere or at least nearly\n> everywhere in the regression test, if we don't print out the disabled\n> node counts when COSTS OFF is used, then we don't cover that case in\n> our tests. Bummer.\n\nAre the disabled node counts still expected to be stable even with\nGEQO? If not, maybe we should have a way to turn them off after all.\nAlthough I agree that always disabling them when COSTS OFF is set is\nprobably also undesirable. How about a new option, e.g. EXPLAIN\n(DISABLED OFF)\n\n\n","msg_date":"Thu, 22 Aug 2024 14:07:32 +0200","msg_from":"Jelte Fennema-Nio <postgres@jeltef.nl>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Thu, Aug 22, 2024 at 8:07 AM Jelte Fennema-Nio <postgres@jeltef.nl> wrote:\n> Are the disabled node counts still expected to be stable even with\n> GEQO? If not, maybe we should have a way to turn them off after all.\n> Although I agree that always disabling them when COSTS OFF is set is\n> probably also undesirable. How about a new option, e.g. EXPLAIN\n> (DISABLED OFF)\n\nHmm, I hadn't thought about that. There are no GEQO-specific changes\nin this patch, which AFAIK is OK, because I think GEQO just relies on\nthe core planning machinery to decide everything about the cost of\npaths, and is really only experimenting with different join orders. So\nI think if it picks the same join order, it should get the same count\nof disabled nodes everywhere. If it doesn't pick the same order,\nyou'll get a different plan entirely.\n\nI don't think I quite want to jump into inventing a new EXPLAIN option\nright this minute. I'm not against the idea, but I don't want to jump\ninto engineering solutions before I understand what the problems are,\nso I think we should give this a little time. I'll be a bit surprised\nif this doesn't elicit a few strong reactions, but I want to see what\npeople are actually sad (or, potentially, happy) about.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 22 Aug 2024 08:43:18 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On 8/21/24 10:29 AM, Robert Haas wrote:\r\n\r\n> I went ahead and committed these patches. I know there's some debate\r\n> over whether we want to show the # of disabled nodes and if so whether\r\n> it should be controlled by COSTS, and I suspect I haven't completely\r\n> allayed David's concerns about the initial_cost_XXX functions although\r\n> I think that I did the right thing. But, I don't have the impression\r\n> that anyone is desperately opposed to the basic concept, so I think it\r\n> makes sense to put these into the tree and see what happens. We have\r\n> quite a bit of time left in this release cycle to uncover bugs, hear\r\n> from users or other developers, etc. about what problems there may be\r\n> with this. If we end up deciding to reverse course or need to fix a\r\n> bunch of stuff, so be it, but let's see what the feedback is.\r\n\r\nWe hit an issue with pgvector[0] where a regular `SELECT count(*) FROM \r\ntable`[1] is attempting to scan the index on the vector column when \r\n`enable_seqscan` is disabled. Credit to Andrew Kane (CC'd) for flagging it.\r\n\r\nI was able to trace this back to e2225346. Here is a reproducer:\r\n\r\nSetup\r\n=====\r\n\r\nCREATE EXTENSION vector;\r\n\r\nCREATE OR REPLACE FUNCTION public.generate_random_normalized_vector(dim \r\ninteger)\r\nRETURNS vector\r\nLANGUAGE SQL\r\nAS $$\r\n SELECT public.l2_normalize(array_agg(random()::real)::vector)\r\n FROM generate_series(1, $1);\r\n$$;\r\n\r\nCREATE TABLE test (id int, embedding vector(128));\r\nINSERT INTO test\r\n SELECT n, public.generate_random_normalized_vector(128)\r\n FROM generate_series(1,5) n;\r\n\r\nCREATE INDEX ON test USING hnsw (embedding vector_cosine_ops);\r\n\r\nTest\r\n====\r\n\r\nSET enable_seqscan TO off;\r\nEXPLAIN ANALYZE\r\nSELECT count(*) FROM test;\r\n\r\nBefore e2225346:\r\n----------------\r\n\r\nAggregate (cost=10000041965.00..10000041965.01 rows=1 width=8) (actual \r\ntime=189.864..189.864 rows\r\n=1 loops=1)\r\n -> Seq Scan on test (cost=10000000000.00..10000040715.00 rows=5 \r\nwidth=0) (actual time=0.01\r\n8..168.294 rows=5 loops=1)\r\n(4 rows)\r\n\r\nWith e2225346:\r\n-------------\r\nERROR: cannot scan hnsw index without order\r\n\r\nSome things to note with the ivfflat/hnsw index AMs[3] in pgvector are \r\nthat they're used for \"ORDER BY\" scans exclusively. They currently don't \r\nsupport index only scans (noting as I tried reproducing the issue with \r\nGIST and couldn't do so because of that), but we wouldn't want to do a \r\nfull table \"count(*)\" on a IVFFlat/HNSW index anyway as it'd be more \r\nexpensive than just a full table scan.\r\n\r\nThanks,\r\n\r\nJonathan\r\n\r\n[0] https://github.com/pgvector/pgvector\r\n[1] https://github.com/pgvector/pgvector/actions/runs/10519052945\r\n[2] https://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=e2225346\r\n[3] https://github.com/pgvector/pgvector/blob/master/src/hnsw.c#L192","msg_date":"Fri, 23 Aug 2024 11:16:57 -0400","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Aug 23, 2024 at 11:17 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n> We hit an issue with pgvector[0] where a regular `SELECT count(*) FROM\n> table`[1] is attempting to scan the index on the vector column when\n> `enable_seqscan` is disabled. Credit to Andrew Kane (CC'd) for flagging it.\n>\n> I was able to trace this back to e2225346. Here is a reproducer:\n\nIf I change EXPLAIN ANALYZE in this test to just EXPLAIN, I get this:\n\n Aggregate (cost=179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00..179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00\nrows=1 width=8)\n -> Index Only Scan using test_embedding_idx on test\n(cost=179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00..179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00\nrows=5 width=0)\n\nIt took me a moment to wrap my head around this: the cost estimate is\n312 decimal digits long. Apparently hnswcostestimate() just returns\nDBL_MAX when there are no scan keys because it really, really doesn't\nwant to do that. Before e2225346, that kept this plan from being\ngenerated because it was (much) larger than disable_cost. But now it\ndoesn't, because 1 disabled node makes a path more expensive than any\npossible non-disabled path. Since that was the whole point of the\npatch, I don't feel too bad about it.\n\nI find it a little weird that hnsw thinks itself able to return all\nthe tuples in an order the user chooses, but unable to return all of\nthe tuples in an arbitrary order. In core, we have precedent for index\ntypes that can't return individual tuples at all (gin, brin) but not\none that is able to return tuples in concept but has a panic attack if\nyou don't know how you want them sorted. I don't quite see why you\ncouldn't just treat that case the same as ORDER BY\nthe_first_column_of_the_index, or any other arbitrary rule that you\nwant to make up. Sure, it might be more expensive than a sequential\nscan, but the user said they didn't want a sequential scan. I'm not\nquite sure why pgvector thinks it gets to decide that it knows better\nthan the user, or the rest of the optimizer. I don't even think I\nreally believe it would always be worse: I've seen cases where a table\nwas badly bloated and mostly empty but its indexes were not bloated,\nand in that case an index scan can be a HUGE winner even though it\nwould normally be a lot worse than a sequential scan.\n\nIf you don't want to fix hnsw to work the way the core optimizer\nthinks it should, or if there's some reason it can't be done,\nalternatives might include (1) having the cost estimate function hack\nthe count of disabled nodes and (2) adding some kind of core support\nfor an index cost estimator refusing a path entirely. I haven't tested\n(1) so I don't know for sure that there are no issues, but I think we\nhave to do all of our cost estimating before we can think about adding\nthe path so I feel like there's a decent chance it would do what you\nwant.\n\nAlso, while I did take the initiative to download pgvector and compile\nit and hook up a debugger and figure out what was going on here, I'm\nnot really too sure that's my job. I do think I have a responsibility\nto help maintainers of out-of-core extensions who have problems as a\nresult of my commits, but I also think it's fair to hope that those\nmaintainers will try to minimize the amount of time that I need to\nspend trying to read code that I did not write and do not maintain.\nFortunately, this wasn't hard to figure out, but in a way that's kind\nof the point. That DBL_MAX hack was put there by somebody who must've\nunderstood that they were trying to use a very large cost to disable a\ncertain path shape completely, and it seems to me that if that person\nhad studied this case and the commit message for e2225346, they would\nhave likely understood what had happened pretty quickly. Do you think\nthat's an unfair feeling on my part?\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 23 Aug 2024 13:11:45 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Fri, Aug 23, 2024 at 11:17 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>> We hit an issue with pgvector[0] where a regular `SELECT count(*) FROM\n>> table`[1] is attempting to scan the index on the vector column when\n>> `enable_seqscan` is disabled. Credit to Andrew Kane (CC'd) for flagging it.\n\n> It took me a moment to wrap my head around this: the cost estimate is\n> 312 decimal digits long. Apparently hnswcostestimate() just returns\n> DBL_MAX when there are no scan keys because it really, really doesn't\n> want to do that. Before e2225346, that kept this plan from being\n> generated because it was (much) larger than disable_cost. But now it\n> doesn't, because 1 disabled node makes a path more expensive than any\n> possible non-disabled path. Since that was the whole point of the\n> patch, I don't feel too bad about it.\n\nYeah, I don't think it's necessary for v18 to be bug-compatible with\nthis hack.\n\n> If you don't want to fix hnsw to work the way the core optimizer\n> thinks it should, or if there's some reason it can't be done,\n> alternatives might include (1) having the cost estimate function hack\n> the count of disabled nodes and (2) adding some kind of core support\n> for an index cost estimator refusing a path entirely. I haven't tested\n> (1) so I don't know for sure that there are no issues, but I think we\n> have to do all of our cost estimating before we can think about adding\n> the path so I feel like there's a decent chance it would do what you\n> want.\n\nIt looks like amcostestimate could change the path's disabled_nodes\ncount, since that's set up before invoking amcostestimate. I guess\nit could be set to INT_MAX to have a comparable solution to before.\n\nI agree with you that it is not great that hnsw is refusing this case\nrather than finding a way to make it work, so I'm not excited about\nputting in support for refusing it in a less klugy way.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 23 Aug 2024 13:26:26 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Aug 23, 2024 at 1:26 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> It looks like amcostestimate could change the path's disabled_nodes\n> count, since that's set up before invoking amcostestimate. I guess\n> it could be set to INT_MAX to have a comparable solution to before.\n\nIt's probably better to add a more modest value, to avoid overflow.\nYou could add a million or so and be far away from overflow while\npresumably still being more disabled than any other path.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 23 Aug 2024 13:37:53 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Fri, Aug 23, 2024 at 1:26 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> It looks like amcostestimate could change the path's disabled_nodes\n>> count, since that's set up before invoking amcostestimate. I guess\n>> it could be set to INT_MAX to have a comparable solution to before.\n\n> It's probably better to add a more modest value, to avoid overflow.\n> You could add a million or so and be far away from overflow while\n> presumably still being more disabled than any other path.\n\nBut that'd only matter if the path survived its first add_path\ntournament, which it shouldn't. If it does then you're at risk\nof the same run-time failure reported here.\n\n(Having said that, you're likely right that \"a million or so\"\nwould be a safer choice, since it doesn't require the assumption\nthat the path fails instantly.)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 23 Aug 2024 13:42:53 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On 23/08/2024 20:11, Robert Haas wrote:\n> I find it a little weird that hnsw thinks itself able to return all\n> the tuples in an order the user chooses, but unable to return all of\n> the tuples in an arbitrary order.\n\nHNSW is weird in many ways:\n\n- There is no inherent sort order. It cannot do \"ORDER BY column\", only \nkNN-sort like \"ORDER BY column <-> value\".\n\n- It's approximate. It's not guaranteed to return the same set of rows \nas a sequential scan + sort.\n\n- The number of results it returns is limited by the hnsw.ef_search GUC, \ndefault 100.\n\n- It collects all the results (up to hnsw.ef_search) in memory, and only \nthen returns them. So if you tried to use it with a large number of \nresults, it can simply run out of memory.\n\nArguably all of those are bugs in HNSW, but it is what it is. The \nalgorithm is inherently approximate. Despite that, it's useful in practice.\n\n> In core, we have precedent for index\n> types that can't return individual tuples at all (gin, brin) but not\n> one that is able to return tuples in concept but has a panic attack if\n> you don't know how you want them sorted.\n\nWell, we do also have gin_fuzzy_search_limit. Two wrongs doesn't make it \nright, though; I'd love to get rid of that hack too somehow.\n\n> I don't quite see why you\n> couldn't just treat that case the same as ORDER BY\n> the_first_column_of_the_index, or any other arbitrary rule that you\n> want to make up. Sure, it might be more expensive than a sequential\n> scan, but the user said they didn't want a sequential scan. I'm not\n> quite sure why pgvector thinks it gets to decide that it knows better\n> than the user, or the rest of the optimizer. I don't even think I\n> really believe it would always be worse: I've seen cases where a table\n> was badly bloated and mostly empty but its indexes were not bloated,\n> and in that case an index scan can be a HUGE winner even though it\n> would normally be a lot worse than a sequential scan.\n\nSure, you could make it work. It could construct a vector out of thin \nair to compare with, when there's no scan key, or implement a completely \ndifferent codepath that traverses the full graph in no particular order.\n\n> If you don't want to fix hnsw to work the way the core optimizer\n> thinks it should, or if there's some reason it can't be done,\n> alternatives might include (1) having the cost estimate function hack\n> the count of disabled nodes and (2) adding some kind of core support\n> for an index cost estimator refusing a path entirely. I haven't tested\n> (1) so I don't know for sure that there are no issues, but I think we\n> have to do all of our cost estimating before we can think about adding\n> the path so I feel like there's a decent chance it would do what you\n> want.\n\nIt would seem useful for an index AM to be able to say \"nope, I can't do \nthis\". I don't remember how exactly this stuff works, but I'm surprised \nit doesn't already exist.\n\n-- \nHeikki Linnakangas\nNeon (https://neon.tech)\n\n\n\n","msg_date":"Fri, 23 Aug 2024 21:18:32 +0300","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On 8/23/24 1:11 PM, Robert Haas wrote:\r\n> On Fri, Aug 23, 2024 at 11:17 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\r\n>> We hit an issue with pgvector[0] where a regular `SELECT count(*) FROM\r\n>> table`[1] is attempting to scan the index on the vector column when\r\n>> `enable_seqscan` is disabled. Credit to Andrew Kane (CC'd) for flagging it.\r\n>>\r\n>> I was able to trace this back to e2225346. Here is a reproducer:\r\n> \r\n> If I change EXPLAIN ANALYZE in this test to just EXPLAIN, I get this:\r\n> \r\n> Aggregate (cost=179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00..179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00\r\n> rows=1 width=8)\r\n> -> Index Only Scan using test_embedding_idx on test\r\n> (cost=179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00..179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00\r\n> rows=5 width=0)\r\n> \r\n> It took me a moment to wrap my head around this: the cost estimate is\r\n> 312 decimal digits long. Apparently hnswcostestimate() just returns\r\n> DBL_MAX when there are no scan keys because it really, really doesn't\r\n> want to do that. Before e2225346, that kept this plan from being\r\n> generated because it was (much) larger than disable_cost. But now it\r\n> doesn't, because 1 disabled node makes a path more expensive than any\r\n> possible non-disabled path. Since that was the whole point of the\r\n> patch, I don't feel too bad about it.\r\n> \r\n> I find it a little weird that hnsw thinks itself able to return all\r\n> the tuples in an order the user chooses, but unable to return all of\r\n> the tuples in an arbitrary order. \r\n\r\nFor HNSW, \"order\" is approximated - even when it's returning \"in the \r\norder the user chooses,\" the scan is making the best guess at what the \r\ncorrect order is based on the index structure. At the traditional \"leaf\" \r\nlevel of an index, you're actually traversing a graph-based neighborhood \r\nof values. And maybe we could say \"Hey, if you get the equivalent of a \r\ncount(*), just do the count at the bottom layer (Layer 0)\" but I think \r\nthis would be very expensive.\r\n\r\n> In core, we have precedent for index\r\n> types that can't return individual tuples at all (gin, brin) but not\r\n> one that is able to return tuples in concept but has a panic attack if\r\n> you don't know how you want them sorted. I don't quite see why you\r\n> couldn't just treat that case the same as ORDER BY\r\n> the_first_column_of_the_index, or any other arbitrary rule that you\r\n> want to make up. Sure, it might be more expensive than a sequential\r\n> scan, but the user said they didn't want a sequential scan. I'm not\r\n> quite sure why pgvector thinks it gets to decide that it knows better\r\n> than the user, or the rest of the optimizer. I don't even think I\r\n> really believe it would always be worse: I've seen cases where a table\r\n> was badly bloated and mostly empty but its indexes were not bloated,\r\n> and in that case an index scan can be a HUGE winner even though it\r\n> would normally be a lot worse than a sequential scan.\r\n\r\nThe challenge here is that HNSW is used specifically for approximating \r\nordering; it's not used to directly filter results in the traditional \r\nsense (e.g. via. a WHERE clause). It's a bit different than the others \r\nmentioned in that regard. However, maybe there are other options to \r\nconsider here based on this work.\r\n\r\n> If you don't want to fix hnsw to work the way the core optimizer\r\n> thinks it should, or if there's some reason it can't be done,\r\n> alternatives might include (1) having the cost estimate function hack\r\n> the count of disabled nodes and (2) adding some kind of core support\r\n> for an index cost estimator refusing a path entirely. I haven't tested\r\n> (1) so I don't know for sure that there are no issues, but I think we\r\n> have to do all of our cost estimating before we can think about adding\r\n> the path so I feel like there's a decent chance it would do what you\r\n> want.\r\n\r\nThanks for the options.\r\n\r\n> Also, while I did take the initiative to download pgvector and compile\r\n> it and hook up a debugger and figure out what was going on here, I'm\r\n> not really too sure that's my job. I do think I have a responsibility\r\n> to help maintainers of out-of-core extensions who have problems as a\r\n> result of my commits, but I also think it's fair to hope that those\r\n> maintainers will try to minimize the amount of time that I need to\r\n> spend trying to read code that I did not write and do not maintain.\r\n> Fortunately, this wasn't hard to figure out, but in a way that's kind\r\n> of the point. That DBL_MAX hack was put there by somebody who must've\r\n> understood that they were trying to use a very large cost to disable a\r\n> certain path shape completely, and it seems to me that if that person\r\n> had studied this case and the commit message for e2225346, they would\r\n> have likely understood what had happened pretty quickly. Do you think\r\n> that's an unfair feeling on my part?\r\n\r\nI don't think extension maintainers necessarily have the same level of \r\nPostgreSQL internals as you or many of the other people who frequent \r\n-hackers, so I think it's fair for them to ask questions or raise issues \r\nwith patches they don't understand. I was able to glean from the commit \r\nmessage that this was the commit that likely changed the behavior in \r\npgvector, but I can't immediately glean looking through the code as to \r\nwhy. (And using your logic, should an extension maintainer understand \r\nthe optimizer code when PostgreSQL is providing an interface to the \r\nextension maintainer to encapsulate its interactions)?\r\n\r\nYou can always push back and say \"Well, maybe try this, or try that\" - \r\nwhich would be a mentoring approach that could push it back on the \r\nextension maintainer, which is valid, but I don't see why an extension \r\nmaintainer can't raise an issue or ask a question here.\r\n\r\nThanks,\r\n\r\nJonathan","msg_date":"Fri, 23 Aug 2024 14:19:56 -0400","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Aug 23, 2024 at 2:20 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n> I don't think extension maintainers necessarily have the same level of\n> PostgreSQL internals as you or many of the other people who frequent\n> -hackers, so I think it's fair for them to ask questions or raise issues\n> with patches they don't understand. I was able to glean from the commit\n> message that this was the commit that likely changed the behavior in\n> pgvector, but I can't immediately glean looking through the code as to\n> why. (And using your logic, should an extension maintainer understand\n> the optimizer code when PostgreSQL is providing an interface to the\n> extension maintainer to encapsulate its interactions)?\n>\n> You can always push back and say \"Well, maybe try this, or try that\" -\n> which would be a mentoring approach that could push it back on the\n> extension maintainer, which is valid, but I don't see why an extension\n> maintainer can't raise an issue or ask a question here.\n\nI'm certainly not saying that extension maintainers can't raise issues\nor ask questions here. I just feel that the problem could have been\nanalyzed a bit more before posting.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 23 Aug 2024 14:29:16 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Aug 23, 2024 at 2:18 PM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\n> It would seem useful for an index AM to be able to say \"nope, I can't do\n> this\". I don't remember how exactly this stuff works, but I'm surprised\n> it doesn't already exist.\n\nYeah, I think so, too. While this particular problem is due to a\nproblem with an out-of-core AM that may be doing some slightly\nquestionable things, there's not really any reason why we couldn't\nhave similar problems in core for some other reason. For example, we\ncould change amcostestimate's signature so that an extension can\nreturn true or false, with false meaning that the path can't be\nsupported. We could then change cost_index so that it can also return\ntrue or false, and then change create_index_path so it has the option\nto return NULL. Callers of create_index_path could then be adjusted\nnot to call add_path when NULL is returned.\n\nThere might be a more elegant way to do it with more refactoring, but\nthe above seems good enough.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 23 Aug 2024 14:36:09 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Fri, Aug 23, 2024 at 2:18 PM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\n>> It would seem useful for an index AM to be able to say \"nope, I can't do\n>> this\". I don't remember how exactly this stuff works, but I'm surprised\n>> it doesn't already exist.\n\n> Yeah, I think so, too. While this particular problem is due to a\n> problem with an out-of-core AM that may be doing some slightly\n> questionable things, there's not really any reason why we couldn't\n> have similar problems in core for some other reason. For example, we\n> could change amcostestimate's signature so that an extension can\n> return true or false, with false meaning that the path can't be\n> supported. We could then change cost_index so that it can also return\n> true or false, and then change create_index_path so it has the option\n> to return NULL. Callers of create_index_path could then be adjusted\n> not to call add_path when NULL is returned.\n\nIf we're going to do this, I'd prefer a solution that doesn't force\nAPI changes onto the vast majority of index AMs that don't have a\nproblem here.\n\nOne way could be to formalize the hack we were just discussing:\n\"To refuse a proposed path, amcostestimate can set the path's\ndisabled_nodes value to anything larger than 1\". I suspect that\nthat would actually be sufficient, since the path would then lose\nto the seqscan path in add_path even if that were disabled; but\nwe could put in a hack to prevent it from getting add_path'd at all.\n\nAnother way could be to bless what hnsw is already doing:\n\"To refuse a proposed path, amcostestimate can return an\nindexTotalCost of DBL_MAX\" (or maybe insisting on +Inf would\nbe better). That would still require changes comparable to\nwhat you specify above, but only in the core-code call path\nnot in every AM.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 23 Aug 2024 14:48:35 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Aug 23, 2024 at 2:48 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> If we're going to do this, I'd prefer a solution that doesn't force\n> API changes onto the vast majority of index AMs that don't have a\n> problem here.\n\nThat's a fair concern.\n\n> One way could be to formalize the hack we were just discussing:\n> \"To refuse a proposed path, amcostestimate can set the path's\n> disabled_nodes value to anything larger than 1\". I suspect that\n> that would actually be sufficient, since the path would then lose\n> to the seqscan path in add_path even if that were disabled; but\n> we could put in a hack to prevent it from getting add_path'd at all.\n>\n> Another way could be to bless what hnsw is already doing:\n> \"To refuse a proposed path, amcostestimate can return an\n> indexTotalCost of DBL_MAX\" (or maybe insisting on +Inf would\n> be better). That would still require changes comparable to\n> what you specify above, but only in the core-code call path\n> not in every AM.\n\nIf just setting disabled_nodes to a value larger than one works, I'd\nbe inclined to not do anything here at all, except possibly document\nthat you can do that. Otherwise, we should probably change the code\nsomehow.\n\nI find both of your proposed solutions above to be pretty inelegant,\nand I think if this problem occurred with a core AM, I'd push for an\nAPI break rather than accept the ugliness. \"This path is not valid\nbecause the AM cannot support it\", \"this path is crazy expensive\", and\n\"the user told us not to do it this way\" are three different things,\nand signalling two or more of them in the same way muddies the water\nin a way that I don't like. API breaks aren't free, though, so I\ncertainly understand why you're not very keen to introduce one where\nit can reasonably be avoided.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 23 Aug 2024 15:05:34 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On 23/08/2024 22:05, Robert Haas wrote:\n> On Fri, Aug 23, 2024 at 2:48 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> If we're going to do this, I'd prefer a solution that doesn't force\n>> API changes onto the vast majority of index AMs that don't have a\n>> problem here.\n> \n> That's a fair concern.\n\nYeah, although I don't think it's too bad. There are not that many \nout-of-tree index AM implementations to begin with, and we do change \nthings often enough that any interesting AM implementation will likely \nneed a few #ifdef PG_VERSION blocks for each PostgreSQL major version \nanyway. pgvector certainly does.\n\n>> One way could be to formalize the hack we were just discussing:\n>> \"To refuse a proposed path, amcostestimate can set the path's\n>> disabled_nodes value to anything larger than 1\". I suspect that\n>> that would actually be sufficient, since the path would then lose\n>> to the seqscan path in add_path even if that were disabled; but\n>> we could put in a hack to prevent it from getting add_path'd at all.\n>>\n>> Another way could be to bless what hnsw is already doing:\n>> \"To refuse a proposed path, amcostestimate can return an\n>> indexTotalCost of DBL_MAX\" (or maybe insisting on +Inf would\n>> be better). That would still require changes comparable to\n>> what you specify above, but only in the core-code call path\n>> not in every AM.\n> \n> If just setting disabled_nodes to a value larger than one works, I'd\n> be inclined to not do anything here at all, except possibly document\n> that you can do that. Otherwise, we should probably change the code\n> somehow.\n\nModifying the passed-in Path feels hacky. amcostestimate currently \nreturns all the estimates in *output parameters, it doesn't modify the \nPath at all.\n\n> I find both of your proposed solutions above to be pretty inelegant,\n> and I think if this problem occurred with a core AM, I'd push for an\n> API break rather than accept the ugliness. \"This path is not valid\n> because the AM cannot support it\", \"this path is crazy expensive\", and\n> \"the user told us not to do it this way\" are three different things,\n> and signalling two or more of them in the same way muddies the water\n> in a way that I don't like. API breaks aren't free, though, so I\n> certainly understand why you're not very keen to introduce one where\n> it can reasonably be avoided.\n\nThe +Inf approach seems fine to me. Or perhaps NaN. Your proposal would \ncertainly be the cleanest interface if we don't mind incurring churn to \nAM implementations.\n\n-- \nHeikki Linnakangas\nNeon (https://neon.tech)\n\n\n\n","msg_date":"Fri, 23 Aug 2024 22:12:55 +0300","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> I find both of your proposed solutions above to be pretty inelegant,\n\nThey are that. If we were working in a green field I'd not propose\nsuch things ... but we aren't. I believe there are now a fair number\nof out-of-core index AMs, so I'd rather not break all of them if we\ndon't have to.\n\n> and I think if this problem occurred with a core AM, I'd push for an\n> API break rather than accept the ugliness. \"This path is not valid\n> because the AM cannot support it\", \"this path is crazy expensive\", and\n> \"the user told us not to do it this way\" are three different things,\n> and signalling two or more of them in the same way muddies the water\n> in a way that I don't like.\n\nI think it's not that bad, because we can limit the knowledge of this\nhack to the amcostestimate interface, which doesn't really deal in\n\"the user told us not to do it this way\" at all. That argues against\nmy first proposal though (having amcostestimate touch disabled_nodes\ndirectly). I now think that a reasonable compromise is to say that\nsetting indexTotalCost to +Inf signals that \"the AM cannot support\nit\". That's not conflated too much with the other case, since even a\ncrazy-expensive cost estimate surely ought to be finite. We can have\ncost_index untangle that case into a separate failure return so that\nthe within-the-core-optimizer APIs remain clean.\n\nWhile that would require hnsw to make a small code change (return\n+Inf not DBL_MAX), that coding should work in back branches too,\nso they don't even need a version check.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 23 Aug 2024 15:32:11 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On 8/23/24 2:29 PM, Robert Haas wrote:\r\n> On Fri, Aug 23, 2024 at 2:20 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\r\n>> I don't think extension maintainers necessarily have the same level of\r\n>> PostgreSQL internals as you or many of the other people who frequent\r\n>> -hackers, so I think it's fair for them to ask questions or raise issues\r\n>> with patches they don't understand. I was able to glean from the commit\r\n>> message that this was the commit that likely changed the behavior in\r\n>> pgvector, but I can't immediately glean looking through the code as to\r\n>> why. (And using your logic, should an extension maintainer understand\r\n>> the optimizer code when PostgreSQL is providing an interface to the\r\n>> extension maintainer to encapsulate its interactions)?\r\n>>\r\n>> You can always push back and say \"Well, maybe try this, or try that\" -\r\n>> which would be a mentoring approach that could push it back on the\r\n>> extension maintainer, which is valid, but I don't see why an extension\r\n>> maintainer can't raise an issue or ask a question here.\r\n> \r\n> I'm certainly not saying that extension maintainers can't raise issues\r\n> or ask questions here. I just feel that the problem could have been\r\n> analyzed a bit more before posting.\r\n\r\nThis assumes that the person posting the problem has the requisite \r\nexpertise to determine what the issue is. Frankly, I was happy I was \r\nable to at least trace the issue down to the particular commit and \r\nbrought what appeared to be a reliable reproducer, in absence of knowing \r\nif 1/ this was actually an issue with PG or pgvector, 2/ does it \r\nactually require a fix, or 3/ what the problem could actually be, given \r\na lack of understanding of the full inner working of the optimizer.\r\n\r\nBased on the above, I'm not sure what bar this needed to clear to begin \r\na discussion on the mailing list (which further downthread, seems to be \r\nraising some interesting points).\r\n\r\nJonathan","msg_date":"Fri, 23 Aug 2024 17:29:10 -0400","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On 8/23/24 3:32 PM, Tom Lane wrote:\r\n> Robert Haas <robertmhaas@gmail.com> writes:\r\n>> I find both of your proposed solutions above to be pretty inelegant,\r\n> \r\n> They are that. If we were working in a green field I'd not propose\r\n> such things ... but we aren't. I believe there are now a fair number\r\n> of out-of-core index AMs, so I'd rather not break all of them if we\r\n> don't have to.\r\n\r\nFor distribution of index AMs in the wild, it's certainly > 1 now, and \r\nincreasing. They're not the easiest extension types to build out, so \r\nit's not as widely distributed as some of the other APIs, but there are \r\na bunch out there, as well as language-specific libs (e.g. pgrx for \r\nRust) that offer wrappers around them.\r\n\r\n>> and I think if this problem occurred with a core AM, I'd push for an\r\n>> API break rather than accept the ugliness. \"This path is not valid\r\n>> because the AM cannot support it\", \"this path is crazy expensive\", and\r\n>> \"the user told us not to do it this way\" are three different things,\r\n>> and signalling two or more of them in the same way muddies the water\r\n>> in a way that I don't like.\r\n> \r\n> I think it's not that bad, because we can limit the knowledge of this\r\n> hack to the amcostestimate interface, which doesn't really deal in\r\n> \"the user told us not to do it this way\" at all. That argues against\r\n> my first proposal though (having amcostestimate touch disabled_nodes\r\n> directly). I now think that a reasonable compromise is to say that\r\n> setting indexTotalCost to +Inf signals that \"the AM cannot support\r\n> it\". That's not conflated too much with the other case, since even a\r\n> crazy-expensive cost estimate surely ought to be finite. We can have\r\n> cost_index untangle that case into a separate failure return so that\r\n> the within-the-core-optimizer APIs remain clean.\r\n> \r\n> While that would require hnsw to make a small code change (return\r\n> +Inf not DBL_MAX), that coding should work in back branches too,\r\n> so they don't even need a version check.\r\n\r\n+1 for this approach (I'll do a quick test in my pgvector workspace just \r\nto ensure it gets the same results in the older version).\r\n\r\nJonathan","msg_date":"Fri, 23 Aug 2024 17:33:12 -0400","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On 8/23/24 5:33 PM, Jonathan S. Katz wrote:\r\n> On 8/23/24 3:32 PM, Tom Lane wrote:\r\n>> Robert Haas <robertmhaas@gmail.com> writes:\r\n\r\n>> I think it's not that bad, because we can limit the knowledge of this\r\n>> hack to the amcostestimate interface, which doesn't really deal in\r\n>> \"the user told us not to do it this way\" at all. That argues against\r\n>> my first proposal though (having amcostestimate touch disabled_nodes\r\n>> directly). I now think that a reasonable compromise is to say that\r\n>> setting indexTotalCost to +Inf signals that \"the AM cannot support\r\n>> it\". That's not conflated too much with the other case, since even a\r\n>> crazy-expensive cost estimate surely ought to be finite. We can have\r\n>> cost_index untangle that case into a separate failure return so that\r\n>> the within-the-core-optimizer APIs remain clean.\r\n>>\r\n>> While that would require hnsw to make a small code change (return\r\n>> +Inf not DBL_MAX), that coding should work in back branches too,\r\n>> so they don't even need a version check.\r\n> \r\n> +1 for this approach (I'll do a quick test in my pgvector workspace just \r\n> to ensure it gets the same results in the older version).\r\n\r\n...and I confirmed the +inf approach on PG16 +pgvector does still give \r\nthe same expected result.\r\n\r\nThanks,\r\n\r\nJonathan","msg_date":"Fri, 23 Aug 2024 17:44:37 -0400","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Hello Robert,\n\n21.08.2024 17:29, Robert Haas wrote:\n> I went ahead and committed these patches. ...\n\nPlease take a look at the following code:\nstatic void\nlabel_sort_with_costsize(PlannerInfo *root, Sort *plan, double limit_tuples)\n{\n...\n cost_sort(&sort_path, root, NIL,\n lefttree->total_cost,\n plan->plan.disabled_nodes,\n lefttree->plan_rows,\n lefttree->plan_width,\n 0.0,\n work_mem,\n limit_tuples);\n\nGiven the cost_sort() declaration:\nvoid\ncost_sort(Path *path, PlannerInfo *root,\n List *pathkeys, int input_disabled_nodes,\n Cost input_cost, double tuples, int width,\n Cost comparison_cost, int sort_mem,\n double limit_tuples)\n\nAren't the input_disabled_nodes and input_cost arguments swapped in the\nabove call?\n\n(I've discovered this with UBSan, which complained\ncreateplan.c:5457:6: runtime error: 4.40465e+09 is outside the range of representable values of type 'int'\nwhile executing a query with a large estimated cost.)\n\nBest regards,\nAlexander\n\n\n","msg_date":"Fri, 6 Sep 2024 12:00:00 +0300","msg_from":"Alexander Lakhin <exclusion@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Sep 6, 2024 at 5:00 PM Alexander Lakhin <exclusion@gmail.com> wrote:\n> static void\n> label_sort_with_costsize(PlannerInfo *root, Sort *plan, double limit_tuples)\n> {\n> ...\n> cost_sort(&sort_path, root, NIL,\n> lefttree->total_cost,\n> plan->plan.disabled_nodes,\n> lefttree->plan_rows,\n> lefttree->plan_width,\n> 0.0,\n> work_mem,\n> limit_tuples);\n>\n> Given the cost_sort() declaration:\n> void\n> cost_sort(Path *path, PlannerInfo *root,\n> List *pathkeys, int input_disabled_nodes,\n> Cost input_cost, double tuples, int width,\n> Cost comparison_cost, int sort_mem,\n> double limit_tuples)\n>\n> Aren't the input_disabled_nodes and input_cost arguments swapped in the\n> above call?\n\nNice catch! I checked other callers to cost_sort, and they are all\ngood.\n\n(I'm a little surprised that this does not cause any plan diffs in the\nregression tests.)\n\nThanks\nRichard\n\n\n","msg_date":"Fri, 6 Sep 2024 17:27:22 +0800","msg_from":"Richard Guo <guofenglinux@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Sep 6, 2024 at 5:27 PM Richard Guo <guofenglinux@gmail.com> wrote:\n> On Fri, Sep 6, 2024 at 5:00 PM Alexander Lakhin <exclusion@gmail.com> wrote:\n> > static void\n> > label_sort_with_costsize(PlannerInfo *root, Sort *plan, double limit_tuples)\n\n> (I'm a little surprised that this does not cause any plan diffs in the\n> regression tests.)\n\nAh I see. label_sort_with_costsize is only used to label the Sort\nnode nicely for EXPLAIN, and usually we do not display the cost\nnumbers in regression tests.\n\nThanks\nRichard\n\n\n","msg_date":"Fri, 6 Sep 2024 17:51:05 +0800","msg_from":"Richard Guo <guofenglinux@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"Hello Richard,\n\n06.09.2024 12:51, Richard Guo wrote:\n> Ah I see. label_sort_with_costsize is only used to label the Sort\n> node nicely for EXPLAIN, and usually we do not display the cost\n> numbers in regression tests.\n\nIn fact, I see the error with the following (EXPLAIN-less) query:\ncreate table t (x int);\n\nselect * from t natural inner join\n(select * from (values(1)) v(x)\n union all\n select 1 from t t1 full join t t2 using (x),\n t t3 full join t t4 using (x)\n);\n\n2024-09-06 10:01:48.034 UTC [696535:5] psql LOG: statement: select * from t natural inner join\n (select * from (values(1)) v(x)\n union all\n select 1 from t t1 full join t t2 using (x),\n t t3 full join t t4 using (x)\n );\ncreateplan.c:5457:6: runtime error: 4.99254e+09 is outside the range of representable values of type 'int'\nSUMMARY: UndefinedBehaviorSanitizer: undefined-behavior createplan.c:5457:6 in\n\n(An UBSan-enabled build --with-blocksize=32 is required for this query to\ntrigger the failure.)\n\nBest regards,\nAlexander\n\n\n","msg_date":"Fri, 6 Sep 2024 13:10:10 +0300","msg_from":"Alexander Lakhin <exclusion@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, Sep 6, 2024 at 5:27 PM Richard Guo <guofenglinux@gmail.com> wrote:\n> On Fri, Sep 6, 2024 at 5:00 PM Alexander Lakhin <exclusion@gmail.com> wrote:\n> > static void\n> > label_sort_with_costsize(PlannerInfo *root, Sort *plan, double limit_tuples)\n> > {\n> > ...\n> > cost_sort(&sort_path, root, NIL,\n> > lefttree->total_cost,\n> > plan->plan.disabled_nodes,\n> > lefttree->plan_rows,\n> > lefttree->plan_width,\n> > 0.0,\n> > work_mem,\n> > limit_tuples);\n> >\n> > Given the cost_sort() declaration:\n> > void\n> > cost_sort(Path *path, PlannerInfo *root,\n> > List *pathkeys, int input_disabled_nodes,\n> > Cost input_cost, double tuples, int width,\n> > Cost comparison_cost, int sort_mem,\n> > double limit_tuples)\n> >\n> > Aren't the input_disabled_nodes and input_cost arguments swapped in the\n> > above call?\n>\n> Nice catch! I checked other callers to cost_sort, and they are all\n> good.\n\nFixed.\n\nThanks\nRichard\n\n\n","msg_date":"Mon, 9 Sep 2024 12:09:26 +0800","msg_from":"Richard Guo <guofenglinux@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Mon, Sep 9, 2024 at 12:09 AM Richard Guo <guofenglinux@gmail.com> wrote:\n> Fixed.\n\nThanks to Alexander for the very good catch and to Richard for pushing the fix.\n\n(I started to respond to this last week but didn't quite get to it\nbefore I ran out of time/energy.)\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 9 Sep 2024 11:28:44 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Wed, 2024-08-21 at 10:29 -0400, Robert Haas wrote:\n> I went ahead and committed these patches. I know there's some debate\n> over whether we want to show the # of disabled nodes and if so whether\n> it should be controlled by COSTS, and I suspect I haven't completely\n> allayed David's concerns about the initial_cost_XXX functions although\n> I think that I did the right thing. But, I don't have the impression\n> that anyone is desperately opposed to the basic concept, so I think it\n> makes sense to put these into the tree and see what happens. We have\n> quite a bit of time left in this release cycle to uncover bugs, hear\n> from users or other developers, etc. about what problems there may be\n> with this. If we end up deciding to reverse course or need to fix a\n> bunch of stuff, so be it, but let's see what the feedback is.\n\nI am somewhat unhappy about the \"Disabled Nodes\" in EXPLAIN.\n\nFirst, the commit message confused me: it claims that the information\nis displayed with EXPLAIN ANALYZE, but it's shown with every EXPLAIN.\n\nBut that's not important. My complaints are:\n\n1. The \"disabled nodes\" are always displayed.\n I'd be happier if it were only shown for COSTS ON, but I think it\n would be best if they were only shown with VERBOSE ON.\n\n After all, the messages are pretty verbose...\n\n2. The \"disabled nodes\" are not only shown at the nodes where nodes\n were actually disabled, but also at every nodes above these nodes.\n\n This would be fine:\n\n Sort\n -> Nested Loop Join\n -> Hash Join\n -> Index Scan\n Disabled Nodes: 1\n -> Hash\n -> Index Scan\n Disabled Nodes: 1\n -> Index Scan\n Disabled Nodes: 1\n\n This is annoying:\n\n Sort\n Disabled Nodes: 3\n -> Nested Loop Join\n Disabled Nodes: 3\n -> Hash Join\n Disabled Nodes: 2\n -> Index Scan\n Disabled Nodes: 1\n -> Hash\n -> Index Scan\n Disabled Nodes: 1\n -> Index Scan\n Disabled Nodes: 1\n\nI have no idea how #2 could be implemented, but it would be nice to have.\nPlease, please, can we show the \"disabled nodes\" only with VERBOSE?\n\nYours,\nLaurenz Albe\n\n\n","msg_date":"Fri, 27 Sep 2024 10:42:32 +0200","msg_from":"Laurenz Albe <laurenz.albe@cybertec.at>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Fri, 27 Sept 2024 at 20:42, Laurenz Albe <laurenz.albe@cybertec.at> wrote:\n> 2. The \"disabled nodes\" are not only shown at the nodes where nodes\n> were actually disabled, but also at every nodes above these nodes.\n\nI'm also not a fan either and I'd like to see this output improved.\n\nIt seems like it's easy enough to implement some logic to detect when\na given node is disabled just by checking if the disable_nodes count\nis higher than the sum of the disabled_node field of the node's\nchildren. If there are no children (a scan node) and disabed_nodes >\n0 then it must be disabled. There's even a nice fast path where we\ndon't need to check the children if disabled_nodes == 0.\n\nHere's a POC grade patch of how I'd rather see it looking.\n\nI opted to have a boolean field as I didn't see any need for an\ninteger count. I also changed things around so we always display the\nboolean property in non-text EXPLAIN. Normally, we don't mind being\nmore verbose there.\n\nI also fixed a bug in make_sort() where disabled_nodes isn't being set\nproperly. I'll do an independent patch for that if this goes nowhere.\n\nDavid","msg_date":"Sat, 28 Sep 2024 00:04:20 +1200","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: On disable_cost"},{"msg_contents":"On Sat, 2024-09-28 at 00:04 +1200, David Rowley wrote:\n> On Fri, 27 Sept 2024 at 20:42, Laurenz Albe <laurenz.albe@cybertec.at> wrote:\n> > 2. The \"disabled nodes\" are not only shown at the nodes where nodes\n> > were actually disabled, but also at every nodes above these nodes.\n> \n> I'm also not a fan either and I'd like to see this output improved.\n> \n> It seems like it's easy enough to implement some logic to detect when\n> a given node is disabled just by checking if the disable_nodes count\n> is higher than the sum of the disabled_node field of the node's\n> children. If there are no children (a scan node) and disabed_nodes >\n> 0 then it must be disabled. There's even a nice fast path where we\n> don't need to check the children if disabled_nodes == 0.\n> \n> Here's a POC grade patch of how I'd rather see it looking.\n> \n> I opted to have a boolean field as I didn't see any need for an\n> integer count. I also changed things around so we always display the\n> boolean property in non-text EXPLAIN. Normally, we don't mind being\n> more verbose there.\n> \n> I also fixed a bug in make_sort() where disabled_nodes isn't being set\n> properly. I'll do an independent patch for that if this goes nowhere.\n\nThanks, and the patch looks good.\n\nWhy did you change \"Disabled\" from an integer to a boolean?\nIf you see a join where two plans were disabled, that's useful information.\n\nI would still prefer to see the disabled nodes only in VERBOSE explain,\nbut I'm satisfied if the disabled nodes don't show up all over the place.\n\nYours,\nLaurenz Albe\n\n\n","msg_date":"Mon, 30 Sep 2024 19:17:38 +0200","msg_from":"Laurenz Albe <laurenz.albe@cybertec.at>","msg_from_op":false,"msg_subject":"Re: On disable_cost"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\n Postgres has a global variable `disable_cost`. It is set the value\\n1.0e10.\\n\\n This value will be added to the cost of path if related GUC is set off.\\nFor example,\\n if enable_nestloop is set off, when planner trys to add nestloop join\\npath, it continues\\n to add such path but with a huge cost `disable_cost`.\\n\\n But 1.0e10 may not be large enough. I encounter this issue in\\nGreenplum(based on postgres).\\nHeikki tolds me that someone also encountered the same issue on Postgres.\\nSo I send it here to\\nhave a discussion.\\n\\n My issue: I did some spikes and tests on TPCDS 1TB Bytes data. For\\nquery 104, it generates\\n nestloop join even with enable_nestloop set off. And the final plan's\\ntotal cost is very huge (about 1e24). But If I enlarge the disable_cost to\\n1e30, then, planner will generate hash join.\\n\\n So I guess that disable_cost is not large enough for huge amount of\\ndata.\\n\\n It is tricky to set disable_cost a huge number. Can we come up with\\nbetter solution?\\n\\n The following thoughts are from Heikki:\\n\\n> Aside from not having a large enough disable cost, there's also the\\n> fact that the high cost might affect the rest of the plan, if we have to\\n> use a plan type that's disabled. For example, if a table doesn't have any\\n> indexes, but enable_seqscan is off, we might put the unavoidable Seq Scan\\n> on different side of a join than we we would with enable_seqscan=on,\\n> because of the high cost estimate.\\n\\n\\n\\n> I think a more robust way to disable forbidden plan types would be to\\n> handle the disabling in add_path(). Instead of having a high disable cost\\n> on the Path itself, the comparison add_path() would always consider\\n> disabled paths as more expensive than others, regardless of the cost.\\n\\n\\n Any thoughts or ideas on the problem? Thanks!\\n\\nBest Regards,\\nZhenghua Lyu\\n\\nHi, Postgres has a global variable `disable_cost`. It is set the value 1.0e10. This value will be added to the cost of path if related GUC is set off. For example, if enable_nestloop is set off, when planner trys to add nestloop join path, it continues to add such path but with a huge cost `disable_cost`. But 1.0e10 may not be large enough. I encounter this issue in Greenplum(based on postgres).Heikki tolds me that someone also encountered the same issue on Postgres. So I send it here tohave a discussion. My issue: I did some spikes and tests on TPCDS 1TB Bytes data. For query 104, it generates nestloop join even with enable_nestloop set off. And the final plan's total cost is very huge (about 1e24). But If I enlarge the disable_cost to 1e30, then, planner will generate hash join. So I guess that disable_cost is not large enough for huge amount of data. It is tricky to set disable_cost a huge number. Can we come up with better solution? The following thoughts are from Heikki: Aside from not having a large enough disable cost, there's also the fact that the high cost might affect the rest of the plan, if we have to use a plan type that's disabled. For example, if a table doesn't have any indexes, but enable_seqscan is off, we might put the unavoidable Seq Scan on different side of a join than we we would with enable_seqscan=on, because of the high cost estimate. I think a more robust way to disable forbidden plan types would be to handle the disabling in add_path(). Instead of having a high disable cost on the Path itself, the comparison add_path() would always consider disabled paths as more expensive than others, regardless of the cost. Any thoughts or ideas on the problem? Thanks!Best Regards,Zhenghua Lyu\",\n \"msg_date\": \"Fri, 1 Nov 2019 14:42:25 +0800\",\n \"msg_from\": \"Zhenghua Lyu <zlv@pivotal.io>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 7:42 PM Zhenghua Lyu <zlv@pivotal.io> wrote:\\n> It is tricky to set disable_cost a huge number. Can we come up with better solution?\\n\\nWhat happens if you use DBL_MAX?\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 19:58:04 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Em sex, 1 de nov de 2019 às 03:42, Zhenghua Lyu <zlv@pivotal.io> escreveu:\\n>\\n> My issue: I did some spikes and tests on TPCDS 1TB Bytes data. For query 104, it generates\\n> nestloop join even with enable_nestloop set off. And the final plan's total cost is very huge (about 1e24). But If I enlarge the disable_cost to 1e30, then, planner will generate hash join.\\n>\\n> So I guess that disable_cost is not large enough for huge amount of data.\\n>\\n> It is tricky to set disable_cost a huge number. Can we come up with better solution?\\n>\\nIsn't it a case for a GUC disable_cost? As Thomas suggested, DBL_MAX\\nupper limit should be sufficient.\\n\\n> The following thoughts are from Heikki:\\n>>\\n>> Aside from not having a large enough disable cost, there's also the fact that the high cost might affect the rest of the plan, if we have to use a plan type that's disabled. For example, if a table doesn't have any indexes, but enable_seqscan is off, we might put the unavoidable Seq Scan on different side of a join than we we would with enable_seqscan=on, because of the high cost estimate.\\n>\\n>\\n>>\\n>> I think a more robust way to disable forbidden plan types would be to handle the disabling in add_path(). Instead of having a high disable cost on the Path itself, the comparison add_path() would always consider disabled paths as more expensive than others, regardless of the cost.\\n>\\nI'm afraid it is not as cheap as using diable_cost as a node cost. Are\\nyou proposing to add a new boolean variable in Path struct to handle\\nthose cases in compare_path_costs_fuzzily?\\n\\n\\n-- \\n Euler Taveira Timbira -\\nhttp://www.timbira.com.br/\\n PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 11:48:29 -0300\",\n \"msg_from\": \"Euler Taveira <euler@timbira.com.br>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-11-01 19:58:04 +1300, Thomas Munro wrote:\\n> On Fri, Nov 1, 2019 at 7:42 PM Zhenghua Lyu <zlv@pivotal.io> wrote:\\n> > It is tricky to set disable_cost a huge number. Can we come up with better solution?\\n> \\n> What happens if you use DBL_MAX?\\n\\nThat seems like a bad idea - we add the cost multiple times. And we\\nstill want to compare plans that potentially involve that cost, if\\nthere's no other way to plan the query.\\n\\n- Andres\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 09:00:38 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 12:00 PM Andres Freund <andres@anarazel.de> wrote:\\n> That seems like a bad idea - we add the cost multiple times. And we\\n> still want to compare plans that potentially involve that cost, if\\n> there's no other way to plan the query.\\n\\nYeah. I kind of wonder if we shouldn't instead (a) skip adding paths\\nthat use methods which are disabled and then (b) if we don't end up\\nwith any paths for that reloptinfo, try again, ignoring disabling\\nGUCs.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 12:22:06 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"re: coping with adding disable_cost more than once\\n\\nAnother option would be to have a 2-part Cost structure. If disable_cost is\\never added to the Cost, then you set a flag recording this. If any plans\\nexist that have no disable_costs added to them, then the planner chooses the\\nminimum cost among those, otherwise you choose the minimum cost path.\\n\\n\\n\\n-----\\nJim Finnerty, AWS, Amazon Aurora PostgreSQL\\n--\\nSent from: https://www.postgresql-archive.org/PostgreSQL-hackers-f1928748.html\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 09:30:52 -0700 (MST)\",\n \"msg_from\": \"Jim Finnerty <jfinnert@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-11-01 12:22:06 -0400, Robert Haas wrote:\\n> On Fri, Nov 1, 2019 at 12:00 PM Andres Freund <andres@anarazel.de> wrote:\\n> > That seems like a bad idea - we add the cost multiple times. And we\\n> > still want to compare plans that potentially involve that cost, if\\n> > there's no other way to plan the query.\\n> \\n> Yeah. I kind of wonder if we shouldn't instead (a) skip adding paths\\n> that use methods which are disabled and then (b) if we don't end up\\n> with any paths for that reloptinfo, try again, ignoring disabling\\n> GUCs.\\n\\nHm. That seems complicated. Is it clear that we'd always notice that we\\nhave no plan early enough to know which paths to reconsider? I think\\nthere's cases where that'd only happen a few levels up.\\n\\nAs a first step I'd be inclined to \\\"just\\\" adjust disable_cost up to\\nsomething like 1.0e12. Unfortunately much higher and and we're getting\\ninto the area where the loss of precision starts to be significant\\nenough that I'm not sure that we're always careful enough to perform\\nmath in the right order (e.g. 1.0e16 + 1 being 1.0e16, and 1e+20 + 1000\\nbeing 1e+20). I've seen queries with costs above 1e10 where that costing\\nwasn't insane.\\n\\nAnd then, in a larger patch, go for something like Heikki's proposal\\nquoted by Zhenghua Lyu upthread, where we treat 'forbidden' as a\\nseparate factor in comparisons of path costs, rather than fudging the\\ncost upwards. But there's some care to be taken to make sure we don't\\nregress performance too much due to the additional logic in\\ncompare_path_cost et al.\\n\\nI'd also be curious to see if there's some other problem with cost\\ncalculation here - some of the quoted final costs seem high enough to be\\nsuspicious. I'd be curious to see a plan...\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 09:43:14 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 1, 2019 at 12:43 PM Andres Freund <andres@anarazel.de> wrote:\\n> Hm. That seems complicated. Is it clear that we'd always notice that we\\n> have no plan early enough to know which paths to reconsider? I think\\n> there's cases where that'd only happen a few levels up.\\n\\nYeah, there could be problems of that kind. I think if a baserel has\\nno paths, then we know right away that we've got a problem, but for\\njoinrels it might be more complicated.\\n\\n> As a first step I'd be inclined to \\\"just\\\" adjust disable_cost up to\\n> something like 1.0e12. Unfortunately much higher and and we're getting\\n> into the area where the loss of precision starts to be significant\\n> enough that I'm not sure that we're always careful enough to perform\\n> math in the right order (e.g. 1.0e16 + 1 being 1.0e16, and 1e+20 + 1000\\n> being 1e+20). I've seen queries with costs above 1e10 where that costing\\n> wasn't insane.\\n\\nWe've done that before and we can do it again. But we're going to need\\nto have something better eventually, I think, not just keep kicking\\nthe can down the road.\\n\\nAnother point to consider here is that in some cases we could really\\njust skip generating certain paths altogether. We already do this for\\nhash joins: if we're planning a join and enable_hashjoin is disabled,\\nwe just don't generate hash joins paths at all, except for full joins,\\nwhere there might be no other legal method. As this example shows,\\nthis cannot be applied in all cases, but maybe we could do it more\\nwidely than we do today. I'm not sure how beneficial that technique\\nwould be, though, because it doesn't seem like it's quite enough to\\nsolve this problem by itself.\\n\\nYet another approach would be to divide the cost into two parts, a\\n\\\"cost\\\" component and a \\\"violations\\\" component. If two paths are\\ncompared, the one with fewer violations always wins; if it's a tie,\\nthey compare on cost. A path's violation count is the total of its\\nchildren, plus one for itself if it does something that's disabled.\\nThis would be more principled than the current approach, but maybe\\nit's too costly.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 12:56:30 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 01, 2019 at 09:30:52AM -0700, Jim Finnerty wrote:\\n>re: coping with adding disable_cost more than once\\n>\\n>Another option would be to have a 2-part Cost structure. If disable_cost is\\n>ever added to the Cost, then you set a flag recording this. If any plans\\n>exist that have no disable_costs added to them, then the planner chooses the\\n>minimum cost among those, otherwise you choose the minimum cost path.\\n>\\n\\nYeah, I agree having is_disabled flag, and treat all paths with 'true'\\nas more expensive than paths with 'false' (and when both paths have the\\nsame value then actually compare the cost) is probably the way forward.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 18:04:06 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On 2019-11-01 12:56:30 -0400, Robert Haas wrote:\\n> On Fri, Nov 1, 2019 at 12:43 PM Andres Freund <andres@anarazel.de> wrote:\\n> > As a first step I'd be inclined to \\\"just\\\" adjust disable_cost up to\\n> > something like 1.0e12. Unfortunately much higher and and we're getting\\n> > into the area where the loss of precision starts to be significant\\n> > enough that I'm not sure that we're always careful enough to perform\\n> > math in the right order (e.g. 1.0e16 + 1 being 1.0e16, and 1e+20 + 1000\\n> > being 1e+20). I've seen queries with costs above 1e10 where that costing\\n> > wasn't insane.\\n> \\n> We've done that before and we can do it again. But we're going to need\\n> to have something better eventually, I think, not just keep kicking\\n> the can down the road.\\n\\nYea, that's why I continued on to describe what we should do afterwards\\n;)\\n\\n\\n> Yet another approach would be to divide the cost into two parts, a\\n> \\\"cost\\\" component and a \\\"violations\\\" component. If two paths are\\n> compared, the one with fewer violations always wins; if it's a tie,\\n> they compare on cost. A path's violation count is the total of its\\n> children, plus one for itself if it does something that's disabled.\\n> This would be more principled than the current approach, but maybe\\n> it's too costly.\\n\\nNamely go for something like this. I think we probably get away with the\\nadditional comparison, especially if we were to store the violations as\\nan integer and did it like if (unlikely(path1->nviolations !=\\npath2->nviolations)) or such - that ought to be very well predicted in\\nnearly all cases.\\n\\nI wonder how much we'd need to reformulate\\ncompare_path_costs/compare_path_costs_fuzzily to allow the compiler to\\nauto-vectorize. Might not be worth caring...\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 10:34:10 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Zhenghua Lyu <zlv@pivotal.io> writes:\\n>> I think a more robust way to disable forbidden plan types would be to\\n>> handle the disabling in add_path(). Instead of having a high disable cost\\n>> on the Path itself, the comparison add_path() would always consider\\n>> disabled paths as more expensive than others, regardless of the cost.\\n\\nGetting rid of disable_cost would be a nice thing to do, but I would\\nrather not do it by adding still more complexity to add_path(), not\\nto mention having to bloat Paths with a separate \\\"disabled\\\" marker.\\n\\nThe idea that I've been thinking about is to not generate disabled\\nPaths in the first place, thus not only fixing the problem but saving\\nsome cycles. While this seems easy enough for \\\"optional\\\" paths,\\nwe have to reserve the ability to generate certain path types regardless,\\nif there's no other way to implement the query. This is a bit of a\\nstumbling block :-(. At the base relation level, we could do something\\nlike generating seqscan last, and only if no other path has been\\nsuccessfully generated. But I'm not sure how to scale that up to\\njoins. In particular, imagine that we consider joining A to B, and\\nfind that the only way is a nestloop, so we generate a nestloop join\\ndespite that being nominally disabled. The next join level would\\nthen see that as an available path, and it might decide that\\n((A nestjoin B) join C) is the cheapest choice, even though there\\nmight have been a way to do, say, ((A join C) join B) with no use of\\nnestloops. Users would find this surprising.\\n\\nMaybe the only way to do this is a separate number-of-uses-of-\\ndisabled-plan-types cost figure in Paths, but I still don't want\\nto go there. The number of cases where disable_cost's shortcomings\\nreally matter is too small to justify that, IMHO.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 02 Nov 2019 10:57:33 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com> writes:\\n> On Fri, Nov 01, 2019 at 09:30:52AM -0700, Jim Finnerty wrote:\\n>> re: coping with adding disable_cost more than once\\n>> \\n>> Another option would be to have a 2-part Cost structure. If disable_cost is\\n>> ever added to the Cost, then you set a flag recording this. If any plans\\n>> exist that have no disable_costs added to them, then the planner chooses the\\n>> minimum cost among those, otherwise you choose the minimum cost path.\\n\\n> Yeah, I agree having is_disabled flag, and treat all paths with 'true'\\n> as more expensive than paths with 'false' (and when both paths have the\\n> same value then actually compare the cost) is probably the way forward.\\n\\nIt would have to be a count, not a boolean --- for example, you want to\\nprefer a path that uses one disabled SeqScan over a path that uses two.\\n\\nI'm with Andres in being pretty worried about the extra burden imposed\\non add_path comparisons.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 02 Nov 2019 11:04:58 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"As a proof of concept, I hacked around a bit today to re-purpose one of the\\nbits of the Cost structure to mean \\\"is_disabled\\\" so that we can distinguish\\n'disabled' from 'non-disabled' paths without making the Cost structure any\\nbigger. In fact, it's still a valid double. The obvious choice would have\\nbeen to re-purpose the sign bit, but I've had occasion to exploit negative\\ncosts before so for this POC I used the high-order bit of the fractional\\nbits of the double. (see Wikipedia for double precision floating point for\\nthe layout).\\n\\nThe idea is to set a special bit when disable_cost is added to a cost. \\nDedicating multiple bits instead of just 1 would be easily done, but as it\\nis we can accumulate many disable_costs without overflowing, so just\\ncomparing the cost suffices.\\n\\nThe patch is not fully debugged and fails on a couple of tests in the serial\\ntest suite. It seems to fail on Cartesian products, and maybe in one other\\nnon-CP case. I wasn't able to debug it before the day came to an end.\\n\\nIn one place the core code subtracts off the disable_cost. I left the\\n\\\"disabled\\\" bit set in this case, which might be wrong.\\n\\nI don't see an option to attach the patch as an attachment, so here is the\\npatch inline (it is based on PG11). The more interesting part is in a small\\nnumber of lines in costsize.c. Other changes just add functions that assign\\na disable_cost and set the bit, or that compare costs such that a\\nnon-disabled cost always compares less than a disabled cost.\\n\\n------------------\\n\\ndiff --git a/src/backend/optimizer/path/costsize.c\\nb/src/backend/optimizer/path/costsize.c\\nindex 4e86458672..3718639330 100644\\n--- a/src/backend/optimizer/path/costsize.c\\n+++ b/src/backend/optimizer/path/costsize.c\\n@@ -123,6 +123,8 @@ double\\t\\tparallel_setup_cost =\\nDEFAULT_PARALLEL_SETUP_COST;\\n int\\t\\t\\teffective_cache_size = DEFAULT_EFFECTIVE_CACHE_SIZE;\\n \\n Cost\\t\\tdisable_cost = 1.0e10;\\n+uint64 disabled_mask = 0x8000000000000;\\n+#define IS_DISABLED(cost) (((uint64) cost) & disabled_mask)\\n \\n int\\t\\t\\tmax_parallel_workers_per_gather = 2;\\n \\n@@ -205,6 +207,53 @@ clamp_row_est(double nrows)\\n \\treturn nrows;\\n }\\n \\n+Cost\\n+add_cost(Cost cost, Cost delta_cost)\\n+{\\n+\\tuint64 mask = (delta_cost == disable_cost) ? disabled_mask : 0;\\n+\\tCost max_cost = disabled_mask - disable_cost;\\n+\\t\\n+\\tif (cost + delta_cost < max_cost)\\n+\\t\\treturn ((Cost) ((uint64)(cost + delta_cost) | mask));\\n+\\telse\\n+\\t\\treturn ((Cost) ((uint64)(max_cost) | mask));\\n+}\\n+\\n+bool\\n+is_lower_cost(Cost cost1, Cost cost2)\\n+{\\n+\\tif ((uint64)cost1 & disabled_mask && !((uint64)cost2 & disabled_mask))\\n+\\t\\treturn false;\\n+\\t\\n+\\tif (!((uint64)cost1 & disabled_mask) && (uint64)cost2 & disabled_mask)\\n+\\t\\treturn true;\\n+\\t\\n+\\treturn (cost1 < cost2);\\n+}\\n+\\n+bool\\n+is_greater_cost(Cost cost1, Cost cost2)\\n+{\\n+\\tif ((uint64)cost1 & disabled_mask && !((uint64)cost2 & disabled_mask))\\n+\\t\\treturn true;\\n+\\t\\n+\\tif (!((uint64)cost1 & disabled_mask) && (uint64)cost2 & disabled_mask)\\n+\\t\\treturn false;\\n+\\t\\n+\\treturn (cost1 > cost2);\\n+}\\n+\\n+bool\\n+is_geq_cost(Cost cost1, Cost cost2)\\n+{\\n+\\tif ((uint64)cost1 & disabled_mask && !((uint64)cost2 & disabled_mask))\\n+\\t\\treturn true;\\n+\\t\\n+\\tif (!((uint64)cost1 & disabled_mask) && (uint64)cost2 & disabled_mask)\\n+\\t\\treturn false;\\n+\\t\\n+\\treturn (cost1 >= cost2);\\n+}\\n \\n /*\\n * cost_seqscan\\n@@ -235,7 +284,7 @@ cost_seqscan(Path *path, PlannerInfo *root,\\n \\t\\tpath->rows = baserel->rows;\\n \\n \\tif (!enable_seqscan)\\n-\\t\\tstartup_cost += disable_cost;\\n+\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n \\n \\t/* fetch estimated page cost for tablespace containing table */\\n \\tget_tablespace_page_costs(baserel->reltablespace,\\n@@ -424,7 +473,7 @@ cost_gather_merge(GatherMergePath *path, PlannerInfo\\n*root,\\n \\t\\tpath->path.rows = rel->rows;\\n \\n \\tif (!enable_gathermerge)\\n-\\t\\tstartup_cost += disable_cost;\\n+\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n \\n \\t/*\\n \\t * Add one to the number of workers to account for the leader. This might\\n@@ -538,7 +587,7 @@ cost_index(IndexPath *path, PlannerInfo *root, double\\nloop_count,\\n \\t}\\n \\n \\tif (!enable_indexscan)\\n-\\t\\tstartup_cost += disable_cost;\\n+\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n \\t/* we don't need to check enable_indexonlyscan; indxpath.c does that */\\n \\n \\t/*\\n@@ -976,7 +1025,7 @@ cost_bitmap_heap_scan(Path *path, PlannerInfo *root,\\nRelOptInfo *baserel,\\n \\t\\tpath->rows = baserel->rows;\\n \\n \\tif (!enable_bitmapscan)\\n-\\t\\tstartup_cost += disable_cost;\\n+\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n \\n \\tpages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,\\n \\t\\t\\t\\t\\t\\t\\t\\t\\t\\t loop_count, &indexTotalCost,\\n@@ -1242,10 +1291,10 @@ cost_tidscan(Path *path, PlannerInfo *root,\\n \\tif (isCurrentOf)\\n \\t{\\n \\t\\tAssert(baserel->baserestrictcost.startup >= disable_cost);\\n-\\t\\tstartup_cost -= disable_cost;\\n+\\t\\tstartup_cost -= disable_cost; /* but do not un-set the disabled mark */\\n \\t}\\n \\telse if (!enable_tidscan)\\n-\\t\\tstartup_cost += disable_cost;\\n+\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n \\n \\t/*\\n \\t * The TID qual expressions will be computed once, any other baserestrict\\n@@ -1676,7 +1725,7 @@ cost_sort(Path *path, PlannerInfo *root,\\n \\tlong\\t\\tsort_mem_bytes = sort_mem * 1024L;\\n \\n \\tif (!enable_sort)\\n-\\t\\tstartup_cost += disable_cost;\\n+\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n \\n \\tpath->rows = tuples;\\n \\n@@ -2121,8 +2170,8 @@ cost_agg(Path *path, PlannerInfo *root,\\n \\t\\ttotal_cost = input_total_cost;\\n \\t\\tif (aggstrategy == AGG_MIXED && !enable_hashagg)\\n \\t\\t{\\n-\\t\\t\\tstartup_cost += disable_cost;\\n-\\t\\t\\ttotal_cost += disable_cost;\\n+\\t\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n+\\t\\t\\ttotal_cost = add_cost(total_cost, disable_cost);\\n \\t\\t}\\n \\t\\t/* calcs phrased this way to match HASHED case, see note above */\\n \\t\\ttotal_cost += aggcosts->transCost.startup;\\n@@ -2137,7 +2186,7 @@ cost_agg(Path *path, PlannerInfo *root,\\n \\t\\t/* must be AGG_HASHED */\\n \\t\\tstartup_cost = input_total_cost;\\n \\t\\tif (!enable_hashagg)\\n-\\t\\t\\tstartup_cost += disable_cost;\\n+\\t\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n \\t\\tstartup_cost += aggcosts->transCost.startup;\\n \\t\\tstartup_cost += aggcosts->transCost.per_tuple * input_tuples;\\n \\t\\tstartup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;\\n@@ -2436,7 +2485,7 @@ final_cost_nestloop(PlannerInfo *root, NestPath *path,\\n \\t * disabled, which doesn't seem like the way to bet.\\n \\t */\\n \\tif (!enable_nestloop)\\n-\\t\\tstartup_cost += disable_cost;\\n+\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n \\n \\t/* cost of inner-relation source data (we already dealt with outer rel) */\\n \\n@@ -2882,7 +2931,7 @@ final_cost_mergejoin(PlannerInfo *root, MergePath\\n*path,\\n \\t * disabled, which doesn't seem like the way to bet.\\n \\t */\\n \\tif (!enable_mergejoin)\\n-\\t\\tstartup_cost += disable_cost;\\n+\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n \\n \\t/*\\n \\t * Compute cost of the mergequals and qpquals (other restriction clauses)\\n@@ -3312,7 +3361,7 @@ final_cost_hashjoin(PlannerInfo *root, HashPath *path,\\n \\t * disabled, which doesn't seem like the way to bet.\\n \\t */\\n \\tif (!enable_hashjoin)\\n-\\t\\tstartup_cost += disable_cost;\\n+\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n \\n \\t/* mark the path with estimated # of batches */\\n \\tpath->num_batches = numbatches;\\n@@ -3410,7 +3459,7 @@ final_cost_hashjoin(PlannerInfo *root, HashPath *path,\\n \\tif (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),\\n \\t\\t\\t\\t\\t\\t inner_path->pathtarget->width) >\\n \\t\\t(work_mem * 1024L))\\n-\\t\\tstartup_cost += disable_cost;\\n+\\t\\tstartup_cost = add_cost(startup_cost, disable_cost);\\n \\n \\t/*\\n \\t * Compute cost of the hashquals and qpquals (other restriction clauses)\\n@@ -3930,7 +3979,7 @@ cost_qual_eval_walker(Node *node,\\ncost_qual_eval_context *context)\\n \\telse if (IsA(node, CurrentOfExpr))\\n \\t{\\n \\t\\t/* Report high cost to prevent selection of anything but TID scan */\\n-\\t\\tcontext->total.startup += disable_cost;\\n+\\t\\tcontext->total.startup = add_cost(context->total.startup, disable_cost);\\n \\t}\\n \\telse if (IsA(node, SubLink))\\n \\t{\\ndiff --git a/src/backend/optimizer/util/pathnode.c\\nb/src/backend/optimizer/util/pathnode.c\\nindex 4736d84a83..fd746a06bc 100644\\n--- a/src/backend/optimizer/util/pathnode.c\\n+++ b/src/backend/optimizer/util/pathnode.c\\n@@ -72,33 +72,33 @@ compare_path_costs(Path *path1, Path *path2,\\nCostSelector criterion)\\n {\\n \\tif (criterion == STARTUP_COST)\\n \\t{\\n-\\t\\tif (path1->startup_cost < path2->startup_cost)\\n+\\t\\tif (is_lower_cost(path1->startup_cost, path2->startup_cost))\\n \\t\\t\\treturn -1;\\n-\\t\\tif (path1->startup_cost > path2->startup_cost)\\n+\\t\\tif (is_greater_cost(path1->startup_cost, path2->startup_cost))\\n \\t\\t\\treturn +1;\\n \\n \\t\\t/*\\n \\t\\t * If paths have the same startup cost (not at all unlikely), order\\n \\t\\t * them by total cost.\\n \\t\\t */\\n-\\t\\tif (path1->total_cost < path2->total_cost)\\n+\\t\\tif (is_lower_cost(path1->total_cost, path2->total_cost))\\n \\t\\t\\treturn -1;\\n-\\t\\tif (path1->total_cost > path2->total_cost)\\n+\\t\\tif (is_greater_cost(path1->total_cost, path2->total_cost))\\n \\t\\t\\treturn +1;\\n \\t}\\n \\telse\\n \\t{\\n-\\t\\tif (path1->total_cost < path2->total_cost)\\n+\\t\\tif (is_lower_cost(path1->total_cost, path2->total_cost))\\n \\t\\t\\treturn -1;\\n-\\t\\tif (path1->total_cost > path2->total_cost)\\n+\\t\\tif (is_greater_cost(path1->total_cost, path2->total_cost))\\n \\t\\t\\treturn +1;\\n \\n \\t\\t/*\\n \\t\\t * If paths have the same total cost, order them by startup cost.\\n \\t\\t */\\n-\\t\\tif (path1->startup_cost < path2->startup_cost)\\n+\\t\\tif (is_lower_cost(path1->startup_cost, path2->startup_cost))\\n \\t\\t\\treturn -1;\\n-\\t\\tif (path1->startup_cost > path2->startup_cost)\\n+\\t\\tif (is_greater_cost(path1->startup_cost, path2->startup_cost))\\n \\t\\t\\treturn +1;\\n \\t}\\n \\treturn 0;\\n@@ -126,9 +126,9 @@ compare_fractional_path_costs(Path *path1, Path *path2,\\n \\t\\tfraction * (path1->total_cost - path1->startup_cost);\\n \\tcost2 = path2->startup_cost +\\n \\t\\tfraction * (path2->total_cost - path2->startup_cost);\\n-\\tif (cost1 < cost2)\\n+\\tif (is_lower_cost(cost1, cost2))\\n \\t\\treturn -1;\\n-\\tif (cost1 > cost2)\\n+\\tif (is_greater_cost(cost1, cost2))\\n \\t\\treturn +1;\\n \\treturn 0;\\n }\\n@@ -172,11 +172,11 @@ compare_path_costs_fuzzily(Path *path1, Path *path2,\\ndouble fuzz_factor)\\n \\t * Check total cost first since it's more likely to be different; many\\n \\t * paths have zero startup cost.\\n \\t */\\n-\\tif (path1->total_cost > path2->total_cost * fuzz_factor)\\n+\\tif (is_greater_cost(path1->total_cost, path2->total_cost * fuzz_factor))\\n \\t{\\n \\t\\t/* path1 fuzzily worse on total cost */\\n \\t\\tif (CONSIDER_PATH_STARTUP_COST(path1) &&\\n-\\t\\t\\tpath2->startup_cost > path1->startup_cost * fuzz_factor)\\n+\\t\\t\\tis_greater_cost(path2->startup_cost, path1->startup_cost * fuzz_factor))\\n \\t\\t{\\n \\t\\t\\t/* ... but path2 fuzzily worse on startup, so DIFFERENT */\\n \\t\\t\\treturn COSTS_DIFFERENT;\\n@@ -184,11 +184,11 @@ compare_path_costs_fuzzily(Path *path1, Path *path2,\\ndouble fuzz_factor)\\n \\t\\t/* else path2 dominates */\\n \\t\\treturn COSTS_BETTER2;\\n \\t}\\n-\\tif (path2->total_cost > path1->total_cost * fuzz_factor)\\n+\\tif (is_greater_cost(path2->total_cost, path1->total_cost * fuzz_factor))\\n \\t{\\n \\t\\t/* path2 fuzzily worse on total cost */\\n \\t\\tif (CONSIDER_PATH_STARTUP_COST(path2) &&\\n-\\t\\t\\tpath1->startup_cost > path2->startup_cost * fuzz_factor)\\n+\\t\\t\\tis_greater_cost(path1->startup_cost, path2->startup_cost * fuzz_factor))\\n \\t\\t{\\n \\t\\t\\t/* ... but path1 fuzzily worse on startup, so DIFFERENT */\\n \\t\\t\\treturn COSTS_DIFFERENT;\\n@@ -197,12 +197,12 @@ compare_path_costs_fuzzily(Path *path1, Path *path2,\\ndouble fuzz_factor)\\n \\t\\treturn COSTS_BETTER1;\\n \\t}\\n \\t/* fuzzily the same on total cost ... */\\n-\\tif (path1->startup_cost > path2->startup_cost * fuzz_factor)\\n+\\tif (is_greater_cost(path1->startup_cost, path2->startup_cost *\\nfuzz_factor))\\n \\t{\\n \\t\\t/* ... but path1 fuzzily worse on startup, so path2 wins */\\n \\t\\treturn COSTS_BETTER2;\\n \\t}\\n-\\tif (path2->startup_cost > path1->startup_cost * fuzz_factor)\\n+\\tif (is_greater_cost(path2->startup_cost, path1->startup_cost *\\nfuzz_factor))\\n \\t{\\n \\t\\t/* ... but path2 fuzzily worse on startup, so path1 wins */\\n \\t\\treturn COSTS_BETTER1;\\n@@ -605,7 +605,7 @@ add_path(RelOptInfo *parent_rel, Path *new_path)\\n \\t\\telse\\n \\t\\t{\\n \\t\\t\\t/* new belongs after this old path if it has cost >= old's */\\n-\\t\\t\\tif (new_path->total_cost >= old_path->total_cost)\\n+\\t\\t\\tif (is_geq_cost(new_path->total_cost, old_path->total_cost))\\n \\t\\t\\t\\tinsert_after = p1;\\n \\t\\t\\t/* p1_prev advances */\\n \\t\\t\\tp1_prev = p1;\\n@@ -681,7 +681,7 @@ add_path_precheck(RelOptInfo *parent_rel,\\n \\t\\t *\\n \\t\\t * Cost comparisons here should match compare_path_costs_fuzzily.\\n \\t\\t */\\n-\\t\\tif (total_cost > old_path->total_cost * STD_FUZZ_FACTOR)\\n+\\t\\tif (is_greater_cost(total_cost, old_path->total_cost * STD_FUZZ_FACTOR))\\n \\t\\t{\\n \\t\\t\\t/* new path can win on startup cost only if consider_startup */\\n \\t\\t\\tif (startup_cost > old_path->startup_cost * STD_FUZZ_FACTOR ||\\n@@ -796,14 +796,14 @@ add_partial_path(RelOptInfo *parent_rel, Path\\n*new_path)\\n \\t\\t/* Unless pathkeys are incompable, keep just one of the two paths. */\\n \\t\\tif (keyscmp != PATHKEYS_DIFFERENT)\\n \\t\\t{\\n-\\t\\t\\tif (new_path->total_cost > old_path->total_cost * STD_FUZZ_FACTOR)\\n+\\t\\t\\tif (is_greater_cost(new_path->total_cost, old_path->total_cost *\\nSTD_FUZZ_FACTOR))\\n \\t\\t\\t{\\n \\t\\t\\t\\t/* New path costs more; keep it only if pathkeys are better. */\\n \\t\\t\\t\\tif (keyscmp != PATHKEYS_BETTER1)\\n \\t\\t\\t\\t\\taccept_new = false;\\n \\t\\t\\t}\\n-\\t\\t\\telse if (old_path->total_cost > new_path->total_cost\\n-\\t\\t\\t\\t\\t * STD_FUZZ_FACTOR)\\n+\\t\\t\\telse if (is_greater_cost(old_path->total_cost, new_path->total_cost\\n+\\t\\t\\t\\t\\t\\t\\t\\t\\t * STD_FUZZ_FACTOR))\\n \\t\\t\\t{\\n \\t\\t\\t\\t/* Old path costs more; keep it only if pathkeys are better. */\\n \\t\\t\\t\\tif (keyscmp != PATHKEYS_BETTER2)\\n@@ -819,7 +819,7 @@ add_partial_path(RelOptInfo *parent_rel, Path *new_path)\\n \\t\\t\\t\\t/* Costs are about the same, old path has better pathkeys. */\\n \\t\\t\\t\\taccept_new = false;\\n \\t\\t\\t}\\n-\\t\\t\\telse if (old_path->total_cost > new_path->total_cost * 1.0000000001)\\n+\\t\\t\\telse if (is_greater_cost(old_path->total_cost, new_path->total_cost *\\n1.0000000001))\\n \\t\\t\\t{\\n \\t\\t\\t\\t/* Pathkeys are the same, and the old path costs more. */\\n \\t\\t\\t\\tremove_old = true;\\n@@ -847,7 +847,7 @@ add_partial_path(RelOptInfo *parent_rel, Path *new_path)\\n \\t\\telse\\n \\t\\t{\\n \\t\\t\\t/* new belongs after this old path if it has cost >= old's */\\n-\\t\\t\\tif (new_path->total_cost >= old_path->total_cost)\\n+\\t\\t\\tif (is_geq_cost(new_path->total_cost, old_path->total_cost))\\n \\t\\t\\t\\tinsert_after = p1;\\n \\t\\t\\t/* p1_prev advances */\\n \\t\\t\\tp1_prev = p1;\\n@@ -913,10 +913,10 @@ add_partial_path_precheck(RelOptInfo *parent_rel, Cost\\ntotal_cost,\\n \\t\\tkeyscmp = compare_pathkeys(pathkeys, old_path->pathkeys);\\n \\t\\tif (keyscmp != PATHKEYS_DIFFERENT)\\n \\t\\t{\\n-\\t\\t\\tif (total_cost > old_path->total_cost * STD_FUZZ_FACTOR &&\\n+\\t\\t\\tif (is_greater_cost(total_cost, old_path->total_cost * STD_FUZZ_FACTOR)\\n&&\\n \\t\\t\\t\\tkeyscmp != PATHKEYS_BETTER1)\\n \\t\\t\\t\\treturn false;\\n-\\t\\t\\tif (old_path->total_cost > total_cost * STD_FUZZ_FACTOR &&\\n+\\t\\t\\tif (is_greater_cost(old_path->total_cost, total_cost * STD_FUZZ_FACTOR)\\n&&\\n \\t\\t\\t\\tkeyscmp != PATHKEYS_BETTER2)\\n \\t\\t\\t\\treturn true;\\n \\t\\t}\\n@@ -1697,7 +1697,7 @@ create_unique_path(PlannerInfo *root, RelOptInfo *rel,\\nPath *subpath,\\n \\n \\tif (sjinfo->semi_can_btree && sjinfo->semi_can_hash)\\n \\t{\\n-\\t\\tif (agg_path.total_cost < sort_path.total_cost)\\n+\\t\\tif (is_lower_cost(agg_path.total_cost, sort_path.total_cost))\\n \\t\\t\\tpathnode->umethod = UNIQUE_PATH_HASH;\\n \\t\\telse\\n \\t\\t\\tpathnode->umethod = UNIQUE_PATH_SORT;\\ndiff --git a/src/backend/utils/cache/relcache.c\\nb/src/backend/utils/cache/relcache.c\\nindex 78f3b99a76..c261a9d790 100644\\n--- a/src/backend/utils/cache/relcache.c\\n+++ b/src/backend/utils/cache/relcache.c\\n@@ -5076,8 +5076,8 @@ IsProjectionFunctionalIndex(Relation index)\\n \\t\\t * when values differ because the expression is recalculated when\\n \\t\\t * inserting a new index entry for the changed value.\\n \\t\\t */\\n-\\t\\tif ((index_expr_cost.startup + index_expr_cost.per_tuple) >\\n-\\t\\t\\tHEURISTIC_MAX_HOT_RECHECK_EXPR_COST)\\n+\\t\\tif (is_greater_cost((index_expr_cost.startup +\\nindex_expr_cost.per_tuple),\\n+\\t\\t\\t\\t\\t\\t\\tHEURISTIC_MAX_HOT_RECHECK_EXPR_COST))\\n \\t\\t\\tis_projection = false;\\n \\n \\t\\ttuple = SearchSysCache1(RELOID,\\nObjectIdGetDatum(RelationGetRelid(index)));\\ndiff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h\\nindex 9159f2bab1..c01d08eae5 100644\\n--- a/src/include/optimizer/cost.h\\n+++ b/src/include/optimizer/cost.h\\n@@ -251,6 +251,12 @@ extern PathTarget\\n*set_pathtarget_cost_width(PlannerInfo *root, PathTarget *targ\\n extern double compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel,\\n \\t\\t\\t\\t\\t Path *bitmapqual, int loop_count, Cost *cost, double *tuple);\\n \\n+extern Cost add_cost(Cost cost, Cost delta_cost);\\n+extern bool is_lower_cost(Cost cost1, Cost cost2);\\n+extern bool is_greater_cost(Cost cost1, Cost cost2);\\n+extern bool is_geq_cost(Cost cost1, Cost cost2);\\n+\\n+\\n /*\\n * prototypes for clausesel.c\\n *\\t routines to compute clause selectivities\\n\\n\\n\\n\\n\\n-----\\nJim Finnerty, AWS, Amazon Aurora PostgreSQL\\n--\\nSent from: https://www.postgresql-archive.org/PostgreSQL-hackers-f1928748.html\\n\\n\\n\",\n \"msg_date\": \"Tue, 10 Dec 2019 15:50:29 -0700 (MST)\",\n \"msg_from\": \"Jim Finnerty <jfinnert@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, 2019-12-10 at 15:50 -0700, Jim Finnerty wrote:\\n> As a proof of concept, I hacked around a bit today to re-purpose one of the\\n> bits of the Cost structure to mean \\\"is_disabled\\\" so that we can distinguish\\n> 'disabled' from 'non-disabled' paths without making the Cost structure any\\n> bigger. In fact, it's still a valid double. The obvious choice would have\\n> been to re-purpose the sign bit, but I've had occasion to exploit negative\\n> costs before so for this POC I used the high-order bit of the fractional\\n> bits of the double. (see Wikipedia for double precision floating point for\\n> the layout).\\n> \\n> The idea is to set a special bit when disable_cost is added to a cost. \\n> Dedicating multiple bits instead of just 1 would be easily done, but as it\\n> is we can accumulate many disable_costs without overflowing, so just\\n> comparing the cost suffices.\\n\\nDoesn't that rely on a specific implementation of double precision (IEEE)?\\nI thought that we don't want to limit ourselves to platforms with IEEE floats.\\n\\nYours,\\nLaurenz Albe\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 11 Dec 2019 07:23:51 +0100\",\n \"msg_from\": \"Laurenz Albe <laurenz.albe@cybertec.at>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, 11 Dec 2019 at 01:24, Laurenz Albe <laurenz.albe@cybertec.at> wrote:\\n>\\n> On Tue, 2019-12-10 at 15:50 -0700, Jim Finnerty wrote:\\n> > As a proof of concept, I hacked around a bit today to re-purpose one of the\\n> > bits of the Cost structure to mean \\\"is_disabled\\\" so that we can distinguish\\n> > 'disabled' from 'non-disabled' paths without making the Cost structure any\\n> > bigger. In fact, it's still a valid double. The obvious choice would have\\n> > been to re-purpose the sign bit, but I've had occasion to exploit negative\\n> > costs before so for this POC I used the high-order bit of the fractional\\n> > bits of the double. (see Wikipedia for double precision floating point for\\n> > the layout).\\n> >\\n> > The idea is to set a special bit when disable_cost is added to a cost.\\n> > Dedicating multiple bits instead of just 1 would be easily done, but as it\\n> > is we can accumulate many disable_costs without overflowing, so just\\n> > comparing the cost suffices.\\n>\\n> Doesn't that rely on a specific implementation of double precision (IEEE)?\\n> I thought that we don't want to limit ourselves to platforms with IEEE floats.\\n\\nWe could always implement it again in another format....\\n\\nHowever, I wouldn't have expected to be bit twiddling. I would have\\nexpected to use standard functions like ldexp to do this. In fact I\\nthink if you use the high bit of the exponent you could do it entirely\\nusing ldexp and regular double comparisons (with fabs).\\n\\nIe, to set the bit you set cost = ldexp(cost, __DBL_MAX_EXP__/2). And\\nto check for the bit being set you compare ilogb(cost,\\n__DBL_MAX_EXP__/2). Hm. that doesn't handle if the cost is already < 1\\nin which case I guess you would have to set it to 1 first. Or reserve\\nthe two high bits of the cost so you can represent disabled values\\nthat had negative exponents before being disabled.\\n\\nI wonder if it wouldn't be a lot cleaner and more flexible to just go\\nwith a plain float for Cost and use the other 32 bits for counters and\\nbitmasks and still be ahead of the game. A double can store 2^1024 but\\na float 2^128 which still feels like it should be more than enough to\\nstore the kinds of costs plans have without the disabled costs. 2^128\\nmilliseconds is still 10^28 years which is an awfully expensive\\nquery....\\n\\n-- \\ngreg\\n\\n\\n\",\n \"msg_date\": \"Thu, 12 Dec 2019 15:42:37 -0500\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 11, 2019 at 7:24 PM Laurenz Albe <laurenz.albe@cybertec.at> wrote:\\n> Doesn't that rely on a specific implementation of double precision (IEEE)?\\n> I thought that we don't want to limit ourselves to platforms with IEEE floats.\\n\\nJust by the way, you might want to read the second last paragraph of\\nthe commit message for 02ddd499. The dream is over, we're never going\\nto run on Vax.\\n\\n\\n\",\n \"msg_date\": \"Fri, 13 Dec 2019 15:59:32 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Thomas Munro <thomas.munro@gmail.com> writes:\\n> On Wed, Dec 11, 2019 at 7:24 PM Laurenz Albe <laurenz.albe@cybertec.at> wrote:\\n>> Doesn't that rely on a specific implementation of double precision (IEEE)?\\n>> I thought that we don't want to limit ourselves to platforms with IEEE floats.\\n\\n> Just by the way, you might want to read the second last paragraph of\\n> the commit message for 02ddd499. The dream is over, we're never going\\n> to run on Vax.\\n\\nStill, the proposed hack is doubling down on IEEE dependency in a way\\nthat I quite dislike, in that (a) it doesn't just read float values\\nbut generates new ones (and assumes that the hardware/libc will react in\\na predictable way to them), (b) in a part of the code that has no damn\\nbusiness having close dependencies on float format, and (c) for a gain\\nfar smaller than what we got from the Ryu code.\\n\\nWe have had prior discussions about whether 02ddd499 justifies adding\\nmore IEEE dependencies elsewhere. I don't think it does. IEEE 754\\nis not the last word that will ever be said on floating-point arithmetic,\\nany more than x86_64 is the last CPU architecture that anyone will ever\\ncare about. We should keep our dependencies on it well circumscribed.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 15 Dec 2019 14:54:03 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"I think this would be ready to abstract away behind a few functions that\\ncould always be replaced by something else later...\\n\\n\\nHowever on further thought I really think just using a 32-bit float and 32\\nbits of other bitmaps or counters would be a better approach.\\n\\n\\nOn Sun., Dec. 15, 2019, 14:54 Tom Lane, <tgl@sss.pgh.pa.us> wrote:\\n\\n> Thomas Munro <thomas.munro@gmail.com> writes:\\n> > On Wed, Dec 11, 2019 at 7:24 PM Laurenz Albe <laurenz.albe@cybertec.at>\\n> wrote:\\n> >> Doesn't that rely on a specific implementation of double precision\\n> (IEEE)?\\n> >> I thought that we don't want to limit ourselves to platforms with IEEE\\n> floats.\\n>\\n> > Just by the way, you might want to read the second last paragraph of\\n> > the commit message for 02ddd499. The dream is over, we're never going\\n> > to run on Vax.\\n>\\n> Still, the proposed hack is doubling down on IEEE dependency in a way\\n> that I quite dislike, in that (a) it doesn't just read float values\\n> but generates new ones (and assumes that the hardware/libc will react in\\n> a predictable way to them), (b) in a part of the code that has no damn\\n> business having close dependencies on float format, and (c) for a gain\\n> far smaller than what we got from the Ryu code.\\n>\\n> We have had prior discussions about whether 02ddd499 justifies adding\\n> more IEEE dependencies elsewhere. I don't think it does. IEEE 754\\n> is not the last word that will ever be said on floating-point arithmetic,\\n> any more than x86_64 is the last CPU architecture that anyone will ever\\n> care about. We should keep our dependencies on it well circumscribed.\\n>\\n> regards, tom lane\\n>\\n>\\n>\\n\\nI think this would be ready to abstract away behind a few functions that could always be replaced by something else later...However on further thought I really think just using a 32-bit float and 32 bits of other bitmaps or counters would be a better approach. On Sun., Dec. 15, 2019, 14:54 Tom Lane, <tgl@sss.pgh.pa.us> wrote:Thomas Munro <thomas.munro@gmail.com> writes:\\n> On Wed, Dec 11, 2019 at 7:24 PM Laurenz Albe <laurenz.albe@cybertec.at> wrote:\\n>> Doesn't that rely on a specific implementation of double precision (IEEE)?\\n>> I thought that we don't want to limit ourselves to platforms with IEEE floats.\\n\\n> Just by the way, you might want to read the second last paragraph of\\n> the commit message for 02ddd499. The dream is over, we're never going\\n> to run on Vax.\\n\\nStill, the proposed hack is doubling down on IEEE dependency in a way\\nthat I quite dislike, in that (a) it doesn't just read float values\\nbut generates new ones (and assumes that the hardware/libc will react in\\na predictable way to them), (b) in a part of the code that has no damn\\nbusiness having close dependencies on float format, and (c) for a gain\\nfar smaller than what we got from the Ryu code.\\n\\nWe have had prior discussions about whether 02ddd499 justifies adding\\nmore IEEE dependencies elsewhere. I don't think it does. IEEE 754\\nis not the last word that will ever be said on floating-point arithmetic,\\nany more than x86_64 is the last CPU architecture that anyone will ever\\ncare about. We should keep our dependencies on it well circumscribed.\\n\\n regards, tom lane\",\n \"msg_date\": \"Sun, 15 Dec 2019 15:45:31 -0500\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Hi hackers,\\r\\n\\r\\nI have write an initial patch to retire the disable_cost GUC, which labeled a flag on the Path struct instead of adding up a big cost which is hard to estimate. Though it involved in tons of plan changes in regression tests, I have tested on some simple test cases such as eagerly generate a two-stage agg plans and it worked. Could someone help to review?\\r\\n\\r\\n\\r\\nregards,\\r\\n\\r\\nJian\\r\\n________________________________\\r\\nFrom: Euler Taveira <euler@timbira.com.br>\\r\\nSent: Friday, November 1, 2019 22:48\\r\\nTo: Zhenghua Lyu <zlyu@vmware.com>\\r\\nCc: PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>\\r\\nSubject: Re: On disable_cost\\r\\n\\r\\n!! External Email\\r\\n\\r\\nEm sex, 1 de nov de 2019 às 03:42, Zhenghua Lyu <zlv@pivotal.io> escreveu:\\r\\n>\\r\\n> My issue: I did some spikes and tests on TPCDS 1TB Bytes data. For query 104, it generates\\r\\n> nestloop join even with enable_nestloop set off. And the final plan's total cost is very huge (about 1e24). But If I enlarge the disable_cost to 1e30, then, planner will generate hash join.\\r\\n>\\r\\n> So I guess that disable_cost is not large enough for huge amount of data.\\r\\n>\\r\\n> It is tricky to set disable_cost a huge number. Can we come up with better solution?\\r\\n>\\r\\nIsn't it a case for a GUC disable_cost? As Thomas suggested, DBL_MAX\\r\\nupper limit should be sufficient.\\r\\n\\r\\n> The following thoughts are from Heikki:\\r\\n>>\\r\\n>> Aside from not having a large enough disable cost, there's also the fact that the high cost might affect the rest of the plan, if we have to use a plan type that's disabled. For example, if a table doesn't have any indexes, but enable_seqscan is off, we might put the unavoidable Seq Scan on different side of a join than we we would with enable_seqscan=on, because of the high cost estimate.\\r\\n>\\r\\n>\\r\\n>>\\r\\n>> I think a more robust way to disable forbidden plan types would be to handle the disabling in add_path(). Instead of having a high disable cost on the Path itself, the comparison add_path() would always consider disabled paths as more expensive than others, regardless of the cost.\\r\\n>\\r\\nI'm afraid it is not as cheap as using diable_cost as a node cost. Are\\r\\nyou proposing to add a new boolean variable in Path struct to handle\\r\\nthose cases in compare_path_costs_fuzzily?\\r\\n\\r\\n\\r\\n--\\r\\n Euler Taveira Timbira -\\r\\nhttps://nam04.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.timbira.com.br%2F&data=05%7C01%7Cgjian%40vmware.com%7C12a30b2852dd4651667608db9401d056%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C0%7C0%7C638266507757076648%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=v54JhsW8FX4mSmjgt2yP59t7xtv1mZvC%2BBhtKrfp%2FBY%3D&reserved=0<http://www.timbira.com.br/>\\r\\n PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento\\r\\n\\r\\n\\r\\n\\r\\n\\r\\n\\r\\n!! External Email: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the sender.\",\n \"msg_date\": \"Thu, 3 Aug 2023 09:21:39 +0000\",\n \"msg_from\": \"Jian Guo <gjian@vmware.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 3, 2023 at 5:22 AM Jian Guo <gjian@vmware.com> wrote:\\n> I have write an initial patch to retire the disable_cost GUC, which labeled a flag on the Path struct instead of adding up a big cost which is hard to estimate. Though it involved in tons of plan changes in regression tests, I have tested on some simple test cases such as eagerly generate a two-stage agg plans and it worked. Could someone help to review?\\n\\nI took a look at this patch today. I believe that overall this may\\nwell be an approach worth pursuing. However, more work is going to be\\nneeded. Here are some comments:\\n\\n1. You stated that it changes lots of plans in the regression tests,\\nbut you haven't provided any sort of analysis of why those plans\\nchanged. I'm kind of surprised that there would be \\\"tons\\\" of plan\\nchanges. You (or someone) should look into why that's happening.\\n\\n2. The change to compare_path_costs_fuzzily() seems incorrect to me.\\nWhen path1->is_disabled && path2->is_disabled, costs should be\\ncompared just as they are when neither path is disabled. Instead, the\\npatch treats any two such paths as having equal cost. That seems\\ncatastrophically bad. Maybe it accounts for some of those plan\\nchanges, although that would only be true if those plans were created\\nwhile using some disabling GUC.\\n\\n3. Instead of adding is_disabled at the end of the Path structure, I\\nsuggest adding it between param_info and parallel_aware. I think if\\nyou do that, the space used by the new field will use up padding bytes\\nthat are currently included in the struct, instead of making it\\nlonger.\\n\\n4. A critical issue for any patch of this type is performance. This\\nconcern was raised earlier on this thread, but your path doesn't\\naddress it. There's no performance analysis or benchmarking included\\nin your email. One idea that I have is to write the cost-comparison\\ntest like this:\\n\\nif (unlikely(path1->is_disabled || path2->is_disabled))\\n{\\n if (!path1->is_disabled)\\n return COSTS_BETTER1;\\n if (!path2->is_disabled)\\n return COSTS_BETTER2;\\n /* if both disabled, fall through */\\n}\\n\\nI'm not sure that would be enough to prevent the patch from adding\\nnoticeably to the cost of path comparison, but maybe it would help.\\n\\n5. The patch changes only compare_path_costs_fuzzily() but I wonder\\nwhether compare_path_costs() and compare_fractional_path_costs() need\\nsimilar surgery. Whether they do or don't, there should likely be some\\ncomments explaining the situation.\\n\\n6. In fact, the patch changes no comments at all, anywhere. I'm not\\nsure how many comment changes a patch like this needs to make, but the\\nanswer definitely isn't \\\"none\\\".\\n\\n7. The patch doesn't actually remove disable_cost. I guess it should.\\n\\n8. When you submit a patch, it's a good idea to also add it on\\ncommitfest.postgresql.org. It doesn't look like that was done in this\\ncase.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Mar 2024 10:27:48 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Thu, Aug 3, 2023 at 5:22 AM Jian Guo <gjian@vmware.com> wrote:\\n>> I have write an initial patch to retire the disable_cost GUC, which labeled a flag on the Path struct instead of adding up a big cost which is hard to estimate. Though it involved in tons of plan changes in regression tests, I have tested on some simple test cases such as eagerly generate a two-stage agg plans and it worked. Could someone help to review?\\n\\n> I took a look at this patch today. I believe that overall this may\\n> well be an approach worth pursuing. However, more work is going to be\\n> needed. Here are some comments:\\n\\n> 1. You stated that it changes lots of plans in the regression tests,\\n> but you haven't provided any sort of analysis of why those plans\\n> changed. I'm kind of surprised that there would be \\\"tons\\\" of plan\\n> changes. You (or someone) should look into why that's happening.\\n\\nI've not read the patch, but given this description I would expect\\nthere to be *zero* regression changes --- I don't think we have any\\ntest cases that depend on disable_cost being finite. If there's more\\nthan zero changes, that very likely indicates a bug in the patch.\\nEven if there are places where the output legitimately changes, you\\nneed to justify each one and make sure that you didn't invalidate the\\nintent of that test case.\\n\\nBTW, having written that paragraph, I wonder if we couldn't get\\nthe same end result with a nearly one-line change that consists of\\nmaking disable_cost be IEEE infinity. Years ago we didn't want\\nto rely on IEEE float semantics in this area, but nowadays I don't\\nsee why we shouldn't.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Mar 2024 13:32:58 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 12, 2024 at 1:32 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> BTW, having written that paragraph, I wonder if we couldn't get\\n> the same end result with a nearly one-line change that consists of\\n> making disable_cost be IEEE infinity. Years ago we didn't want\\n> to rely on IEEE float semantics in this area, but nowadays I don't\\n> see why we shouldn't.\\n\\nI don't think so, because I think that what will happen in that case\\nis that we'll pick a completely random plan if we can't pick a plan\\nthat avoids incurring disable_cost. Every plan that contains one\\ndisabled node anywhere in the plan tree will look like it has exactly\\nthe same cost as any other such plan.\\n\\nIMHO, this is actually one of the problems with disable_cost as it\\nworks today. I think the semantics that we want are: if it's possible\\nto pick a plan where nothing is disabled, then pick the cheapest such\\nplan; if not, pick the cheapest plan overall. But treating\\ndisable_cost doesn't really do that. It does the first part -- picking\\nthe cheapest plan where nothing is disabled -- but it doesn't do the\\nsecond part, because once you add disable_cost into the cost of some\\nparticular plan node, it screws up the rest of the planning, because\\nthe cost estimates for the disabled nodes have no bearing in reality.\\nFast-start plans, for example, will look insanely good compared to\\nwhat would be the case in normal planning (and we lean too much toward\\nfast-start plans even normally).\\n\\n(I don't think we should care how MANY disabled nodes appear in a\\nplan, particularly. This is a more arguable point. Is a plan with 1\\ndisabled node and 10% more cost better or worse than a plan with 2\\ndisabled nodes and 10% less cost? I'd argue that counting the number\\nof disabled nodes isn't particularly meaningful.)\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Mar 2024 14:01:36 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Tue, Mar 12, 2024 at 1:32 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> BTW, having written that paragraph, I wonder if we couldn't get\\n>> the same end result with a nearly one-line change that consists of\\n>> making disable_cost be IEEE infinity.\\n\\n> I don't think so, because I think that what will happen in that case\\n> is that we'll pick a completely random plan if we can't pick a plan\\n> that avoids incurring disable_cost. Every plan that contains one\\n> disabled node anywhere in the plan tree will look like it has exactly\\n> the same cost as any other such plan.\\n\\nGood point.\\n\\n> IMHO, this is actually one of the problems with disable_cost as it\\n> works today.\\n\\nYeah. I keep thinking that the right solution is to not generate\\ndisabled paths in the first place if there are any other ways to\\nproduce the same relation. That has obvious order-of-operations\\nproblems though, and I've not been able to make it work.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Mar 2024 15:36:38 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 12, 2024 at 3:36 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Yeah. I keep thinking that the right solution is to not generate\\n> disabled paths in the first place if there are any other ways to\\n> produce the same relation. That has obvious order-of-operations\\n> problems though, and I've not been able to make it work.\\n\\nI've expressed the same view in the past. It would be nice not to\\nwaste planner effort on paths that we're just going to throw away, but\\nI'm not entirely sure what you mean by \\\"obvious order-of-operations\\nproblems.\\\"\\n\\nTo me, it seems like what we'd need is to be able to restart the whole\\nplanner process if we run out of steam before we get done. For\\nexample, suppose we're planning a 2-way join where index and\\nindex-only scans are disabled, sorts are disabled, and nested loops\\nand hash joins are disabled. There's no problem generating just the\\nnon-disabled scan types at the baserel level, but when we reach the\\njoin, we're going to find that the only non-disabled join type is a\\nmerge join, and we're also going to find that we have no paths that\\nprovide pre-sorted input, so we need to sort, which we're also not\\nallowed to do. If we could give up at that point and restart planning,\\ndisabling all of the plan-choice constraints and now creating all\\npaths for each RelOptInfo, then everything would, I believe, be just\\nfine. We'd end up needing neither disable_cost nor the mechanism\\nproposed by this patch.\\n\\nBut in the absence of that, we need some way to privilege the\\nnon-disabled paths over the disabled ones -- and I'd prefer to have\\nsomething more principled than disable_cost, if we can work out the\\ndetails.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Mar 2024 15:54:47 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, 13 Mar 2024 at 08:55, Robert Haas <robertmhaas@gmail.com> wrote:\\n> But in the absence of that, we need some way to privilege the\\n> non-disabled paths over the disabled ones -- and I'd prefer to have\\n> something more principled than disable_cost, if we can work out the\\n> details.\\n\\nThe primary place I see issues with disabled_cost is caused by\\nSTD_FUZZ_FACTOR. When you add 1.0e10 to a couple of modestly costly\\npaths, it makes them appear fuzzily the same in cases where one could\\nbe significantly cheaper than the other. If we were to bump up the\\ndisable_cost it would make this problem worse.\\n\\nI think we do still need some way to pick the cheapest disabled path\\nwhen there are no other options.\\n\\nOne way would be to set fuzz_factor to 1.0 when either of the paths\\ncosts in compare_path_costs_fuzzily() is >= disable_cost.\\nclamp_row_est() does cap row estimates at MAXIMUM_ROWCOUNT (1e100), so\\nI think there is some value of disable_cost that could almost\\ncertainly ensure we don't reach it because the path is truly expensive\\nrather than disabled.\\n\\nSo maybe the fix could be to set disable_cost to something like\\n1.0e110 and adjust compare_path_costs_fuzzily to not apply the\\nfuzz_factor for paths >= disable_cost. However, I wonder if that\\nrisks the costs going infinite after a couple of cartesian joins.\\n\\nDavid\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Mar 2024 09:55:22 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"David Rowley <dgrowleyml@gmail.com> writes:\\n> So maybe the fix could be to set disable_cost to something like\\n> 1.0e110 and adjust compare_path_costs_fuzzily to not apply the\\n> fuzz_factor for paths >= disable_cost. However, I wonder if that\\n> risks the costs going infinite after a couple of cartesian joins.\\n\\nPerhaps. It still does nothing for Robert's point that once we're\\nforced into using a \\\"disabled\\\" plan type, it'd be better if the\\ndisabled-ness didn't skew subsequent planning choices.\\n\\nOn the whole I agree that getting rid of disable_cost entirely\\nwould be the way to go, if we can replace that with a separate\\nboolean without driving up the cost of add_path too much.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Mar 2024 17:18:09 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 12, 2024 at 4:55 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n> The primary place I see issues with disabled_cost is caused by\\n> STD_FUZZ_FACTOR. When you add 1.0e10 to a couple of modestly costly\\n> paths, it makes them appear fuzzily the same in cases where one could\\n> be significantly cheaper than the other. If we were to bump up the\\n> disable_cost it would make this problem worse.\\n\\nHmm, good point.\\n\\n> So maybe the fix could be to set disable_cost to something like\\n> 1.0e110 and adjust compare_path_costs_fuzzily to not apply the\\n> fuzz_factor for paths >= disable_cost. However, I wonder if that\\n> risks the costs going infinite after a couple of cartesian joins.\\n\\nYeah, I think the disabled flag is a better answer if we can make it\\nwork. No matter what value we pick for disable_cost, it's bound to\\nskew the planning sometimes.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Mar 2024 09:05:00 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 12, 2024 at 1:32 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > 1. You stated that it changes lots of plans in the regression tests,\\n> > but you haven't provided any sort of analysis of why those plans\\n> > changed. I'm kind of surprised that there would be \\\"tons\\\" of plan\\n> > changes. You (or someone) should look into why that's happening.\\n>\\n> I've not read the patch, but given this description I would expect\\n> there to be *zero* regression changes --- I don't think we have any\\n> test cases that depend on disable_cost being finite. If there's more\\n> than zero changes, that very likely indicates a bug in the patch.\\n> Even if there are places where the output legitimately changes, you\\n> need to justify each one and make sure that you didn't invalidate the\\n> intent of that test case.\\n\\nI spent some more time poking at this patch. It's missing a ton of\\nimportant stuff and is wrong in a whole bunch of really serious ways,\\nand I'm not going to try to mention all of them in this email. But I\\ndo want to talk about some of the more interesting realizations that\\ncame to me as I was working my way through this.\\n\\nOne of the things I realized relatively early is that the patch does\\nnothing to propagate disable_cost upward through the plan tree. That\\nmeans that if you have a choice between, say,\\nSort-over-Append-over-SeqScan and MergeAppend-over-IndexScan, as we do\\nin the regression tests, disabling IndexScan doesn't change the plan\\nwith the patch applied, as it does in master. That's because only the\\nIndexScan node ends up flagged as disabled. Once we start stacking\\nother plan nodes on top of disabled plan nodes, the resultant plans\\nare at no disadvantage compared to plans containing no disabled nodes.\\nThe IndexScan plan survives initially, despite being disabled, because\\nit's got a sort order. That lets us use it to build a MergeAppend\\npath, and that MergeAppend path is not disabled, and compares\\nfavorably on cost.\\n\\nAfter straining my brain over various plan changes for a long time,\\nand hacking on the code somewhat, I realized that just propagating the\\nBoolean upward is insufficient to set things right. That's basically\\nbecause I was being dumb when I said this:\\n\\n> I don't think we should care how MANY disabled nodes appear in a\\n> plan, particularly.\\n\\nSuppose we try to plan a Nested Loop with SeqScan disabled, but\\nthere's no alternative to a SeqScan for the outer side of the join. If\\nwe suppose an upward-propagating Boolean, every path for the join is\\ndisabled because every path for the outer side is disabled. That means\\nthat we have no reason to avoid paths where the inner side also uses a\\ndisabled path. When we loop over the inner rel's pathlist looking for\\nways to build a path for the join, we may find some disabled paths\\nthere, and the join paths we build from those paths are disabled, but\\nso are the join paths where we use a non-disabled path on the inner\\nside. So those paths are just competing with each other on cost, and\\nthe path type that is supposedly disabled on the outer side of the\\njoin ends up not really being very disabled at all. More precisely, if\\ndisabling a plan type causes paths to be discarded completely before\\nthe join paths are constructed, then they actually do get removed from\\nconsideration. But if those paths make it into inner rel's path list,\\neven way out towards the end, then paths derived from them can jump to\\nthe front of the joinrel's path list.\\n\\nThe same kind of problem happens with Append or MergeAppend nodes. The\\nregression tests expect that we'll avoid disabled plan types whenever\\npossible even if we can't avoid them completely; for instance, the\\nmatest0 table intentionally omits an index on one child table.\\nDisabling sequential scans is expected to disable them for all of the\\nother child tables even though for that particular child table there\\nis no other option. But that behavior is hard to achieve if every path\\nfor the parent rel is \\\"equally disabled\\\". You either want the path\\nthat uses only the one required SeqScan to be not-disabled even though\\none of its children is disabled ... or you want the disabled flag to\\nbe more than a Boolean. And while there's probably more than one way\\nto make it work, the easiest thing seems to be to just have a\\ndisabled-counter in every node that gets initialized to the total\\ndisabled-counter values of all of its children, and then you add 1 if\\nthat node is itself doing something that is disabled, i.e. the exact\\nopposite of what I said in the quote above.\\n\\nI haven't done enough work to know whether that would match the\\ncurrent behavior, let alone whether it would have acceptable\\nperformance, and I'm not at all convinced that's the right direction\\nanyway. Actually, at the moment, I don't have a very good idea at all\\nwhat the right direction is. I do have a feeling that this patch is\\nprobably not going in the right direction, but I might be wrong about\\nthat, too.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 1 Apr 2024 16:40:44 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> One of the things I realized relatively early is that the patch does\\n> nothing to propagate disable_cost upward through the plan tree.\\n> ...\\n> After straining my brain over various plan changes for a long time,\\n> and hacking on the code somewhat, I realized that just propagating the\\n> Boolean upward is insufficient to set things right. That's basically\\n> because I was being dumb when I said this:\\n>> I don't think we should care how MANY disabled nodes appear in a\\n>> plan, particularly.\\n\\nVery interesting, thanks for the summary. So the fact that\\ndisable_cost is additive across plan nodes is actually a pretty\\nimportant property of the current setup. I think this is closely\\nrelated to one argument you made against my upthread idea of using\\nIEEE Infinity for disable_cost: that'd mask whether more than one\\nof the sub-plans had been disabled.\\n\\n> ... And while there's probably more than one way\\n> to make it work, the easiest thing seems to be to just have a\\n> disabled-counter in every node that gets initialized to the total\\n> disabled-counter values of all of its children, and then you add 1 if\\n> that node is itself doing something that is disabled, i.e. the exact\\n> opposite of what I said in the quote above.\\n\\nYeah, that seems like the next thing to try if anyone plans to pursue\\nthis further. That'd essentially do what we're doing now except that\\ndisable_cost is its own \\\"order of infinity\\\", entirely separate from\\nnormal costs.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 01 Apr 2024 17:00:11 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Mon, Apr 1, 2024 at 5:00 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Very interesting, thanks for the summary. So the fact that\\n> disable_cost is additive across plan nodes is actually a pretty\\n> important property of the current setup. I think this is closely\\n> related to one argument you made against my upthread idea of using\\n> IEEE Infinity for disable_cost: that'd mask whether more than one\\n> of the sub-plans had been disabled.\\n\\nYes, exactly. I just hadn't quite put the pieces together.\\n\\n> Yeah, that seems like the next thing to try if anyone plans to pursue\\n> this further. That'd essentially do what we're doing now except that\\n> disable_cost is its own \\\"order of infinity\\\", entirely separate from\\n> normal costs.\\n\\nRight. I think that's actually what I had in mind in the last\\nparagraph of http://postgr.es/m/CA+TgmoY+Ltw7B=1FSFSN4yHcu2roWrz-ijBovj-99LZU=9h1dA@mail.gmail.com\\nbut that was a while ago and I'd lost track of why it actually\\nmattered. But I also have questions about whether that's really the\\nright approach.\\n\\nI think the approach of just not generating paths we don't want in the\\nfirst place merits more consideration. We do that in some cases\\nalready, but not in others, and I'm not clear why. Like, if\\nindex-scans, index-only scans, sorts, nested loops, and hash joins are\\ndisabled, something is going to have to give, because the only\\nremaining join type is a merge join yet we've ruled out every possible\\nway of getting the day into some order, but I'm not sure whether\\nthere's some reason that we need exactly the behavior that we have\\nright now rather than anything else. Maybe it would be OK to just\\ninsist on at least one unparameterized, non-partial path at the\\nbaserel level, and then if that ends up forcing us to ignore the\\njoin-type restrictions higher up, so be it. Or maybe that's not OK and\\nafter I try that out I'll end up writing another email about how I was\\na bit clueless about all of this. I don't know. But I feel like it\\nmerits more investigation, because I'm having trouble shaking the\\ntheory that what we've got right now is pretty arbitrary.\\n\\nAnd also ... looking at the regression tests, and also thinking about\\nthe kinds of problems that I think people run into in real\\ndeployments, I can't help feeling like we've somehow got this whole\\nthing backwards. enable_wunk imagines that you want to plan as normal\\nexcept with one particular plan type excluded from consideration. And\\nmaybe that makes sense if the point of the enable_wunk GUC is that the\\nplanner feature might be buggy and you might therefore want to turn it\\noff to protect yourself, or if the planner feature might be expensive\\nand you might want to turn it off to save cycles. But surely that's\\nnot the case with something like enable_seqscan or enable_indexscan.\\nWhat I think we're mostly doing in the regression tests is shutting\\noff every relevant type of plan except one. I theorize that what we\\nactually want to do is tell the planner what we do want to happen,\\nrather than what we don't want to happen, but we've got this weird set\\nof GUCs that do the opposite of that and we're super-attached to them\\nbecause they've existed forever. I don't really have a concrete\\nproposal here, but I wonder if what we're actually talking about here\\nis spending time and energy polishing a mechanism that nobody likes in\\nthe first place.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 1 Apr 2024 19:53:58 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Mon, Apr 1, 2024 at 7:54 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> What I think we're mostly doing in the regression tests is shutting\\n> off every relevant type of plan except one. I theorize that what we\\n> actually want to do is tell the planner what we do want to happen,\\n> rather than what we don't want to happen, but we've got this weird set\\n> of GUCs that do the opposite of that and we're super-attached to them\\n> because they've existed forever.\\n\\n\\nSo rather than listing all the things we don't want to happen, we need a\\nway to force (nay, highly encourage) a particular solution. As our costing\\nis a based on positive numbers, what if we did something like this in\\ncostsize.c?\\n\\n Cost disable_cost = 1.0e10;\\n Cost promotion_cost = 1.0e10; // or higher or lower, depending on\\nhow strongly we want to \\\"beat\\\" disable_costs effects.\\n...\\n\\n if (!enable_seqscan)\\n startup_cost += disable_cost;\\n else if (promote_seqscan)\\n startup_cost -= promotion_cost; // or replace \\\"promote\\\" with\\n\\\"encourage\\\"?\\n\\n\\nCheers,\\nGreg\\n\\nOn Mon, Apr 1, 2024 at 7:54 PM Robert Haas <robertmhaas@gmail.com> wrote:\\nWhat I think we're mostly doing in the regression tests is shutting\\noff every relevant type of plan except one. I theorize that what we\\nactually want to do is tell the planner what we do want to happen,\\nrather than what we don't want to happen, but we've got this weird set\\nof GUCs that do the opposite of that and we're super-attached to them\\nbecause they've existed forever.So rather than listing all the things we don't want to happen, we need a way to force (nay, highly encourage) a particular solution. As our costing is a based on positive numbers, what if we did something like this in costsize.c? Cost disable_cost = 1.0e10; Cost promotion_cost = 1.0e10; // or higher or lower, depending on how strongly we want to \\\"beat\\\" disable_costs effects.... if (!enable_seqscan) startup_cost += disable_cost; else if (promote_seqscan) startup_cost -= promotion_cost; // or replace \\\"promote\\\" with \\\"encourage\\\"?Cheers,Greg\",\n \"msg_date\": \"Tue, 2 Apr 2024 10:03:57 -0400\",\n \"msg_from\": \"Greg Sabino Mullane <htamfids@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 2, 2024 at 10:04 AM Greg Sabino Mullane <htamfids@gmail.com> wrote:\\n> So rather than listing all the things we don't want to happen, we need a way to force (nay, highly encourage) a particular solution. As our costing is a based on positive numbers, what if we did something like this in costsize.c?\\n>\\n> Cost disable_cost = 1.0e10;\\n> Cost promotion_cost = 1.0e10; // or higher or lower, depending on how strongly we want to \\\"beat\\\" disable_costs effects.\\n> ...\\n>\\n> if (!enable_seqscan)\\n> startup_cost += disable_cost;\\n> else if (promote_seqscan)\\n> startup_cost -= promotion_cost; // or replace \\\"promote\\\" with \\\"encourage\\\"?\\n\\nI'm pretty sure negative costs are going to create a variety of\\nunpleasant planning artifacts. The large positive costs do, too, which\\nis where this whole discussion started. If I disable (or promote) some\\nparticular plan, I want the rest of the plan tree to come out looking\\nas much as possible like what would have happened if the same\\nalternative had won organically on cost. I think the only reason we're\\ndriving this off of costing today is that making add_path() more\\ncomplicated is unappealing, mostly on performance grounds, and if you\\nadd disabled-ness (or promoted-ness) as a separate axis of value then\\nadd_path() has to know about that on top of everything else. I think\\nthe goal here is to come up with a more principled alternative that\\nisn't just based on whacking large numbers into the cost and hoping\\nsomething good happens ... but it is a whole lot easier to be unhappy\\nwith the status quo than it is to come up with something that's\\nactually better.\\n\\nI am planning to spend some more time thinking about it, though.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Apr 2024 11:01:20 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Tue, Apr 2, 2024 at 10:04 AM Greg Sabino Mullane <htamfids@gmail.com> wrote:\\n>> if (!enable_seqscan)\\n>> startup_cost += disable_cost;\\n>> else if (promote_seqscan)\\n>> startup_cost -= promotion_cost; // or replace \\\"promote\\\" with \\\"encourage\\\"?\\n\\n> I'm pretty sure negative costs are going to create a variety of\\n> unpleasant planning artifacts.\\n\\nIndeed. It might be okay to have negative values for disabled-ness\\nif we treat disabled-ness as a \\\"separate order of infinity\\\", but\\nI suspect that it'd behave poorly when there are both disabled and\\npromoted sub-paths in a tree, for pretty much the same reasons you\\nexplained just upthread.\\n\\n> I think the only reason we're\\n> driving this off of costing today is that making add_path() more\\n> complicated is unappealing, mostly on performance grounds, and if you\\n> add disabled-ness (or promoted-ness) as a separate axis of value then\\n> add_path() has to know about that on top of everything else.\\n\\nIt doesn't seem to me that it's a separate axis of value, just a\\nhigher-order component of the cost metric. Nonetheless, adding even\\na few instructions to add_path comparisons sounds expensive. Maybe\\nit'd be fine, but we'd need to do some performance testing.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 02 Apr 2024 11:54:24 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 2, 2024 at 11:54 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > I'm pretty sure negative costs are going to create a variety of\\n> > unpleasant planning artifacts.\\n>\\n> Indeed. It might be okay to have negative values for disabled-ness\\n> if we treat disabled-ness as a \\\"separate order of infinity\\\", but\\n> I suspect that it'd behave poorly when there are both disabled and\\n> promoted sub-paths in a tree, for pretty much the same reasons you\\n> explained just upthread.\\n\\nHmm, can you explain further? I think essentially you'd be maximizing\\n#(promoted notes)-#(disabled nodes), but I have no real idea whether\\nthat behavior will be exactly what people want or extremely\\nunintuitive or something in the middle. It seems like it should be\\nfine if there's only promoting or only disabling or if we can respect\\nboth the promoting and the disabling, assuming we even want to have\\nboth, but I'm suspicious that it will be weird somehow in other cases.\\nI can't say exactly in what way, though. Do you have more insight?\\n\\n> > I think the only reason we're\\n> > driving this off of costing today is that making add_path() more\\n> > complicated is unappealing, mostly on performance grounds, and if you\\n> > add disabled-ness (or promoted-ness) as a separate axis of value then\\n> > add_path() has to know about that on top of everything else.\\n>\\n> It doesn't seem to me that it's a separate axis of value, just a\\n> higher-order component of the cost metric. Nonetheless, adding even\\n> a few instructions to add_path comparisons sounds expensive. Maybe\\n> it'd be fine, but we'd need to do some performance testing.\\n\\nHmm, yeah. I'm not sure how much difference there is between these\\nthings in practice. I didn't run down everything that was happening,\\nbut I think what I did was equivalent to making it a higher-order\\ncomponent of the cost metric, and it seemed like an awful lot of paths\\nwere surviving anyway, e.g. index scans survived\\nenable_indexscan=false because they had a sort order, and I think\\nsequential scans were surviving enable_seqscan=false too, perhaps\\nbecause they had no startup cost. At any rate there's no question that\\nadd_path() is hot.\\n\\n--\\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Apr 2024 12:26:27 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Tue, Apr 2, 2024 at 11:54 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> I suspect that it'd behave poorly when there are both disabled and\\n>> promoted sub-paths in a tree, for pretty much the same reasons you\\n>> explained just upthread.\\n\\n> Hmm, can you explain further? I think essentially you'd be maximizing\\n> #(promoted notes)-#(disabled nodes), but I have no real idea whether\\n> that behavior will be exactly what people want or extremely\\n> unintuitive or something in the middle. It seems like it should be\\n> fine if there's only promoting or only disabling or if we can respect\\n> both the promoting and the disabling, assuming we even want to have\\n> both, but I'm suspicious that it will be weird somehow in other cases.\\n> I can't say exactly in what way, though. Do you have more insight?\\n\\nNot really. But if you had, say, a join of a promoted path to a\\ndisabled path, that would be treated as on-par with a join of two\\nregular paths, which seems like it'd lead to odd choices. Maybe\\nit'd be fine, but my gut says it'd likely not act nicely. As you\\nsay, it's a lot easier to believe that only-promoted or only-disabled\\nsituations would behave sanely.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 02 Apr 2024 12:58:18 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 2, 2024 at 12:58 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Not really. But if you had, say, a join of a promoted path to a\\n> disabled path, that would be treated as on-par with a join of two\\n> regular paths, which seems like it'd lead to odd choices. Maybe\\n> it'd be fine, but my gut says it'd likely not act nicely. As you\\n> say, it's a lot easier to believe that only-promoted or only-disabled\\n> situations would behave sanely.\\n\\nMakes sense.\\n\\nI wanted to further explore the idea of just not generating plans of\\ntypes that are currently disabled. I looked into doing this for\\nenable_indexscan and enable_indexonlyscan. As a first step, I\\ninvestigated how those settings work now, and was horrified. I don't\\nknow whether I just wasn't paying attention back when the original\\nindex-only scan work was done -- I remember discussing\\nenable_indexonlyscan with you at the time -- or whether it got changed\\nsubsequently. Anyway, the current behavior is:\\n\\n[A] enable_indexscan=false adds disable_cost to the cost of every\\nIndex Scan path *and also* every Index-Only Scan path. So disabling\\nindex-scans also in effect discourages the use of index-only scans,\\nwhich would make sense if we didn't have a separate setting called\\nenable_indexonlyscan, but we do. Given that, I think this is\\ncompletely and utterly wrong.\\n\\n[b] enable_indexonlyscan=false causes index-only scan paths not to be\\ngenerated at all, but instead, we generate index-scan paths to do the\\nsame thing that we would not have generated otherwise. This is weird\\nbecause it means that disabling one plan type causes us to consider\\nadditional plans of another type, which seems like a thing that a user\\nmight not expect. It's more defensible than [A], though, because you\\ncould argue that we only omit the index scan path as an optimization,\\non the theory that it will always lose to the index-only scan path,\\nand thus if the index-only scan path is not generated, there's a point\\nto generating the index scan path after all, so we should. However, it\\nseems unlikely to me that someone reading the one line of\\ndocumentation that we have about this parameter would be able to guess\\nthat it works this way.\\n\\nHere's an example of how the current system behaves:\\n\\nrobert.haas=# explain select count(*) from pgbench_accounts;\\n QUERY PLAN\\n-----------------------------------------------------------------------------------------------------------------\\n Aggregate (cost=2854.29..2854.30 rows=1 width=8)\\n -> Index Only Scan using pgbench_accounts_pkey on pgbench_accounts\\n (cost=0.29..2604.29 rows=100000 width=0)\\n(2 rows)\\n\\nrobert.haas=# set enable_indexscan=false;\\nSET\\nrobert.haas=# explain select count(*) from pgbench_accounts;\\n QUERY PLAN\\n------------------------------------------------------------------------------\\n Aggregate (cost=2890.00..2890.01 rows=1 width=8)\\n -> Seq Scan on pgbench_accounts (cost=0.00..2640.00 rows=100000 width=0)\\n(2 rows)\\n\\nrobert.haas=# set enable_seqscan=false;\\nSET\\nrobert.haas=# set enable_bitmapscan=false;\\nSET\\nrobert.haas=# explain select count(*) from pgbench_accounts;\\n QUERY PLAN\\n----------------------------------------------------------------------------------------------------------------------------------\\n Aggregate (cost=10000002854.29..10000002854.30 rows=1 width=8)\\n -> Index Only Scan using pgbench_accounts_pkey on pgbench_accounts\\n (cost=10000000000.29..10000002604.29 rows=100000 width=0)\\n(2 rows)\\n\\nrobert.haas=# set enable_indexonlyscan=false;\\nSET\\nrobert.haas=# explain select count(*) from pgbench_accounts;\\n QUERY PLAN\\n-----------------------------------------------------------------------------------------------\\n Aggregate (cost=10000002890.00..10000002890.01 rows=1 width=8)\\n -> Seq Scan on pgbench_accounts\\n(cost=10000000000.00..10000002640.00 rows=100000 width=0)\\n(2 rows)\\n\\nThe first time we run the query, it picks an index-only scan because\\nit's the cheapest. When index scans are disabled, the query now picks\\na sequential scan, even though it wasn't use an index-only scan,\\nbecause the index scan that it was using is perceived to have become\\nvery expensive. When we then shut off sequential scans and bitmap-only\\nscans, it switches back to an index-only scan, because setting\\nenable_indexscan=false didn't completely disable index-only scans, but\\njust made them expensive. But now everything looks expensive, so we go\\nback to the same plan we had initially, except with the cost increased\\nby a bazillion. Finally, when we disable index-only scans, that\\nremoves that plan from the pool, so now we pick the second-cheapest\\nplan overall, which in this case is a sequential scan.\\n\\nSo just to see what would happen, I wrote a patch to make\\nenable_indexscan and enable_indexonlyscan do exactly what they say on\\nthe tin: when you set one of them to false, paths of that type are not\\ngenerated, and nothing else changes. I found that there are a\\nsurprisingly large number of regression tests that rely on the current\\nbehavior, so I took a crack at fixing them to achieve their goals (or\\nwhat I believed their goals to be) in other ways. The resulting patch\\nis attached for your (or anyone's) possible edification.\\n\\nJust to be clear, I have no immediate plans to press forward with\\ntrying to get something committed here. It seems pretty clear to me\\nthat we should fix [A] in some way, but maybe not in the way I did it\\nhere. It's also pretty clear to me that the fact that enable_indexscan\\nand enable_indexonlyscan work completely differently from each other\\nis surprising at best, wrong at worst, but here again, what this patch\\ndoes about that is not above reproach. I think it may make sense to\\ndig through the behavior of some of the remaining enable_* GUCs before\\nsettling on a final strategy here, but I thought that the finds above\\nwere interesting enough and bizarre enough that it made sense to drop\\nan email now and see what people think of all this before going\\nfurther.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Wed, 3 Apr 2024 15:21:24 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 3, 2024 at 3:21 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n\\n> It's also pretty clear to me that the fact that enable_indexscan\\n> and enable_indexonlyscan work completely differently from each other\\n> is surprising at best, wrong at worst, but here again, what this patch\\n> does about that is not above reproach.\\n\\n\\nYes, that is wrong, surely there is a reason we have two vars. Thanks for\\ndigging into this: if nothing else, the code will be better for this\\ndiscussion, even if we do nothing for now with disable_cost.\\n\\nCheers,\\nGreg\\n\\nOn Wed, Apr 3, 2024 at 3:21 PM Robert Haas <robertmhaas@gmail.com> wrote: It's also pretty clear to me that the fact that enable_indexscan\\nand enable_indexonlyscan work completely differently from each other\\nis surprising at best, wrong at worst, but here again, what this patch\\ndoes about that is not above reproach.Yes, that is wrong, surely there is a reason we have two vars. Thanks for digging into this: if nothing else, the code will be better for this discussion, even if we do nothing for now with disable_cost.Cheers,Greg\",\n \"msg_date\": \"Wed, 3 Apr 2024 16:03:44 -0400\",\n \"msg_from\": \"Greg Sabino Mullane <htamfids@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Thu, 4 Apr 2024 at 08:21, Robert Haas <robertmhaas@gmail.com> wrote:\\n> I wanted to further explore the idea of just not generating plans of\\n> types that are currently disabled. I looked into doing this for\\n> enable_indexscan and enable_indexonlyscan. As a first step, I\\n> investigated how those settings work now, and was horrified. I don't\\n> know whether I just wasn't paying attention back when the original\\n> index-only scan work was done -- I remember discussing\\n> enable_indexonlyscan with you at the time -- or whether it got changed\\n> subsequently. Anyway, the current behavior is:\\n>\\n> [A] enable_indexscan=false adds disable_cost to the cost of every\\n> Index Scan path *and also* every Index-Only Scan path. So disabling\\n> index-scans also in effect discourages the use of index-only scans,\\n> which would make sense if we didn't have a separate setting called\\n> enable_indexonlyscan, but we do. Given that, I think this is\\n> completely and utterly wrong.\\n>\\n> [b] enable_indexonlyscan=false causes index-only scan paths not to be\\n> generated at all, but instead, we generate index-scan paths to do the\\n> same thing that we would not have generated otherwise.\\n\\nFWIW, I think changing this is a bad idea and I don't think the\\nbehaviour that's in your patch is useful. With your patch, if I SET\\nenable_indexonlyscan=false, any index that *can* support an IOS for my\\nquery will now not be considered at all!\\n\\nWith your patch applied, I see:\\n\\n-- default enable_* GUC values.\\npostgres=# explain select oid from pg_class order by oid;\\n QUERY PLAN\\n-------------------------------------------------------------------------------------------\\n Index Only Scan using pg_class_oid_index on pg_class\\n(cost=0.27..22.50 rows=415 width=4)\\n(1 row)\\n\\n\\npostgres=# set enable_indexonlyscan=0; -- no index scan?\\nSET\\npostgres=# explain select oid from pg_class order by oid;\\n QUERY PLAN\\n-----------------------------------------------------------------\\n Sort (cost=36.20..37.23 rows=415 width=4)\\n Sort Key: oid\\n -> Seq Scan on pg_class (cost=0.00..18.15 rows=415 width=4)\\n(3 rows)\\n\\npostgres=# set enable_seqscan=0; -- still no index scan!\\nSET\\npostgres=# explain select oid from pg_class order by oid;\\n QUERY PLAN\\n------------------------------------------------------------------------------------\\n Sort (cost=10000000036.20..10000000037.23 rows=415 width=4)\\n Sort Key: oid\\n -> Seq Scan on pg_class (cost=10000000000.00..10000000018.15\\nrows=415 width=4)\\n(3 rows)\\n\\npostgres=# explain select oid from pg_class order by oid,relname; --\\nnow an index scan?!\\n QUERY PLAN\\n---------------------------------------------------------------------------------------------\\n Incremental Sort (cost=0.43..79.50 rows=415 width=68)\\n Sort Key: oid, relname\\n Presorted Key: oid\\n -> Index Scan using pg_class_oid_index on pg_class\\n(cost=0.27..60.82 rows=415 width=68)\\n(4 rows)\\n\\nI don't think this is good as pg_class_oid_index effectively won't be\\nused as long as the particular query could use that index with an\\nindex *only* scan. You can see above that as soon as I adjust the\\nquery slightly so that IOS isn't possible, the index can be used\\nagain. I think an Index Scan would have been a much better option for\\nthe 2nd query than the seq scan and sort.\\n\\nI think if I do SET enable_indexonlyscan=0; the index should still be\\nused with an Index Scan and it definitely shouldn't result in Index\\nScan also being disabled if that index happens to contain all the\\ncolumns required to support an IOS.\\n\\nFWIW, I'm fine with the current behaviour. It looks like we've\\nassumed that, when possible, IOS are always superior to Index Scan, so\\nthere's no point in generating an Index Scan path when we can generate\\nan IOS path. I think this makes sense. For that not to be true,\\nchecking the all visible flag would have to be more costly than\\nvisiting the heap. Perhaps that could be true if the visibility map\\npage had to come from disk and the heap page was cached and the disk\\nwas slow, but I don't think that scenario is worthy of considering\\nboth Index scan and IOS path types when IOS is possible. We've no way\\nto accurately cost that anyway.\\n\\nThis all seems similar to enable_sort vs enable_incremental_sort. For\\na while, we did consider both an incremental sort and a sort when an\\nincremental sort was possible, but it seemed to me that an incremental\\nsort would always be better when it was possible, so I changed that in\\n4a29eabd1. I've not seen anyone complain. I made it so that when an\\nincremental sort is possible but is disabled, we do a sort instead.\\nThat seems fairly similar to how master handles\\nenable_indexonlyscan=false.\\n\\nIn short, I don't find it strange that disabling one node type results\\nin considering another type that we'd otherwise not consider in cases\\nwhere we assume that the disabled node type is always superior and\\nshould always be used when it is possible.\\n\\nI do agree that adding disable_cost to IOS when enable_indexscan=0 is\\na bit weird. We don't penalise incremental sorts when sorts are\\ndisabled, so aligning those might make sense.\\n\\nDavid\\n\\n\\n\",\n \"msg_date\": \"Thu, 4 Apr 2024 10:15:59 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Thu, 4 Apr 2024 at 10:15, David Rowley <dgrowleyml@gmail.com> wrote:\\n> In short, I don't find it strange that disabling one node type results\\n> in considering another type that we'd otherwise not consider in cases\\n> where we assume that the disabled node type is always superior and\\n> should always be used when it is possible.\\n\\nIn addition to what I said earlier, I think the current\\nenable_indexonlyscan is implemented in a way that has the planner do\\nwhat it did before IOS was added. I think that goal makes sense with\\nany patch that make the planner try something new. We want to have\\nsome method to get the previous behaviour for the cases where the\\nplanner makes a dumb choice or to avoid some bug in the new feature.\\n\\nI think using that logic, the current scenario with enable_indexscan\\nand enable_indexonlyscan makes complete sense. I mean, including\\nenable_indexscan=0 adding disable_cost to IOS Paths.\\n\\nDavid\\n\\n\\n\",\n \"msg_date\": \"Thu, 4 Apr 2024 16:09:03 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 3, 2024 at 11:09 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n> On Thu, 4 Apr 2024 at 10:15, David Rowley <dgrowleyml@gmail.com> wrote:\\n> > In short, I don't find it strange that disabling one node type results\\n> > in considering another type that we'd otherwise not consider in cases\\n> > where we assume that the disabled node type is always superior and\\n> > should always be used when it is possible.\\n>\\n> In addition to what I said earlier, I think the current\\n> enable_indexonlyscan is implemented in a way that has the planner do\\n> what it did before IOS was added. I think that goal makes sense with\\n> any patch that make the planner try something new. We want to have\\n> some method to get the previous behaviour for the cases where the\\n> planner makes a dumb choice or to avoid some bug in the new feature.\\n\\nI see the logic of this, and I agree that the resulting behavior might\\nbe more intuitive than what I posted before. I'll do some experiments.\\n\\n> I think using that logic, the current scenario with enable_indexscan\\n> and enable_indexonlyscan makes complete sense. I mean, including\\n> enable_indexscan=0 adding disable_cost to IOS Paths.\\n\\nThis, for me, is a bridge too far. I don't think there's a real\\nargument that \\\"what the planner did before IOS was added\\\" was add\\ndisable_cost to the cost of index-only scan paths. There was no such\\npath type. Independently of that argument, I also think the behavior\\nof a setting needs to be something that a user can understand. Right\\nnow, the documentation says:\\n\\nEnables or disables the query planner's use of index-scan plan types.\\nThe default is on.\\nEnables or disables the query planner's use of index-only-scan plan\\ntypes (see Section 11.9). The default is on.\\n\\nI do not think that a user can be expected to guess from these\\ndescriptions that the first one also affects index-only scans, or that\\nthe two GUCs disable their respective plan types in completely\\ndifferent ways. Granted, the latter inconsistency affects a whole\\nbunch of these settings, not just this one, but still.\\n\\n--\\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 4 Apr 2024 14:02:42 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 2, 2019 at 10:57 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> The idea that I've been thinking about is to not generate disabled\\n> Paths in the first place, thus not only fixing the problem but saving\\n> some cycles. While this seems easy enough for \\\"optional\\\" paths,\\n> we have to reserve the ability to generate certain path types regardless,\\n> if there's no other way to implement the query. This is a bit of a\\n> stumbling block :-(. At the base relation level, we could do something\\n> like generating seqscan last, and only if no other path has been\\n> successfully generated.\\n\\nContinuing my investigation into this rather old thread, I did a\\nrather primitive implementation of this idea, for baserels only, and\\ndiscovered that it caused a small number of planner failures running\\nthe regression tests. Here is a slightly simplified example:\\n\\nCREATE TABLE strtest (n text, t text);\\nCREATE INDEX strtest_n_idx ON strtest (n);\\nSET enable_seqscan=false;\\nEXPLAIN SELECT * FROM strtest s1 INNER JOIN strtest s2 ON s1.n >= s2.n;\\n\\nWith the patch, I get:\\n\\nERROR: could not devise a query plan for the given query\\n\\nThe problem here is that it's perfectly possible to generate a valid\\npath for s1 -- and likewise for s2, since it's the same underlying\\nrelation -- while respecting the enable_seqscan=false constraint.\\nHowever, all such paths are parameterized by the other of the two\\nrelations, which means that if we do that, we can't plan the join,\\nbecause we need an unparameterized path for at least one of the two\\nsides in order to build a nested loop join, which is the only way to\\nsatisfy the parameterization on the other side.\\n\\nNow, you could try to fix this by deciding that planning for a baserel\\nhasn't really succeeded unless we got at least one *unparameterized*\\npath for that baserel. I haven't tried that, but I presume that if you\\ndo, it fixes the above example, because now there will be a last-ditch\\nsequential scan on both sides and so this example will behave as\\nexpected. But if you do that, then in other cases, that sequential\\nscan is going to get picked even when it isn't strictly necessary to\\ndo so, just because some plan that uses it looks better on cost.\\nPresumably that problem can in turn be fixed by deciding that we also\\nneed to keep disable_cost around (or the separate disable-counter idea\\nthat we were discussing recently in another branch of this thread),\\nbut that's arguably missing the point of this exercise.\\n\\nAnother idea is to remove the ERROR mentioned above from\\nset_cheapest() and just allow planning to continue even if some\\nrelations end up with no paths. (This would necessitate finding and\\nfixing any code that could be confused by a pathless relation.) Then,\\nif you get to the top of the plan tree and you have no paths there,\\nredo the join search discarding the constraints (or maybe just some of\\nthe constraints, e.g. allow sequential scans and nested loops, or\\nsomething). Conceptually, I like this idea a lot, but I think there\\nare a few problems. One is that I'm not quite sure how to find all the\\ncode that would need to be adjusted to make it work, though the header\\ncomment for standard_join_search() seems like it's got some helpful\\ntips. A second is that it's another version of the disable_cost =\\ninfinity problem: once you find that you can't generate a path while\\nenforcing all of the restrictions, you just disregard the restrictions\\ncompletely, instead of discarding them only to the extent necessary. I\\nhave a feeling that's not going to be very appealing.\\n\\nNow, I suppose it might be that even if we can't remove disable_cost,\\nsomething along these lines is still worth doing, just to save CPU\\ncycles. You could for example try planning with only non-disabled\\nstuff and then do it over again with everything if that doesn't work\\nout, still keeping disable_cost around so that you avoid disabled\\nnodes where you can. But I'm kind of hoping that I'm missing something\\nand there's some approach that could both kill disable_cost and save\\nsome cycles at the same time. If (any of) you have an idea, I'd love\\nto hear it!\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 3 May 2024 16:33:33 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Sat, 4 May 2024 at 08:34, Robert Haas <robertmhaas@gmail.com> wrote:\\n> Another idea is to remove the ERROR mentioned above from\\n> set_cheapest() and just allow planning to continue even if some\\n> relations end up with no paths. (This would necessitate finding and\\n> fixing any code that could be confused by a pathless relation.) Then,\\n> if you get to the top of the plan tree and you have no paths there,\\n> redo the join search discarding the constraints (or maybe just some of\\n> the constraints, e.g. allow sequential scans and nested loops, or\\n> something).\\n\\nI don't think you'd need to wait longer than where we do set_cheapest\\nand find no paths to find out that there's going to be a problem.\\n\\nI don't think redoing planning is going to be easy or even useful. I\\nmean what do you change when you replan? You can't just do\\nenable_seqscan and enable_nestloop as if there's no index to provide\\nsorted input and the plan requires some sort, then you still can't\\nproduce a plan. Adding enable_sort to the list does not give me much\\nconfidence we'll never fail to produce a plan either. It just seems\\nimpossible to know which of the disabled ones caused the RelOptInfo to\\nhave no paths. Also, you might end up enabling one that caused the\\nplanner to do something different than it would do today. For\\nexample, a Path that today incurs 2x disable_cost vs a Path that only\\nreceives 1x disable_cost might do something different if you just went\\nand enabled a bunch of enable* GUCs before replanning.\\n\\n> Now, I suppose it might be that even if we can't remove disable_cost,\\n> something along these lines is still worth doing, just to save CPU\\n> cycles. You could for example try planning with only non-disabled\\n> stuff and then do it over again with everything if that doesn't work\\n> out, still keeping disable_cost around so that you avoid disabled\\n> nodes where you can. But I'm kind of hoping that I'm missing something\\n> and there's some approach that could both kill disable_cost and save\\n> some cycles at the same time. If (any of) you have an idea, I'd love\\n> to hear it!\\n\\nI think the int Path.disabledness idea is worth coding up to try it.\\nI imagine that a Path will incur the maximum of its subpath's\\ndisabledness's then add_path() just needs to prefer lower-valued\\ndisabledness Paths.\\n\\nThat doesn't get you the benefits of fewer CPU cycles, but where did\\nthat come from as a motive to change this? There's no shortage of\\nother ways to make the planner faster if that's an issue.\\n\\nDavid\\n\\n\\n\",\n \"msg_date\": \"Sun, 5 May 2024 01:16:19 +1200\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"David Rowley <dgrowleyml@gmail.com> writes:\\n> I don't think you'd need to wait longer than where we do set_cheapest\\n> and find no paths to find out that there's going to be a problem.\\n\\nAt a base relation, yes, but that doesn't work for joins: it may be\\nthat a particular join cannot be formed, yet other join sequences\\nwill work. We have that all the time from outer-join ordering\\nrestrictions, never mind enable_xxxjoin flags. So I'm not sure\\nthat we can usefully declare early failure for joins.\\n\\n> I think the int Path.disabledness idea is worth coding up to try it.\\n> I imagine that a Path will incur the maximum of its subpath's\\n> disabledness's then add_path() just needs to prefer lower-valued\\n> disabledness Paths.\\n\\nI would think sum not maximum, but that's a detail.\\n\\n> That doesn't get you the benefits of fewer CPU cycles, but where did\\n> that come from as a motive to change this? There's no shortage of\\n> other ways to make the planner faster if that's an issue.\\n\\nThe concern was to not *add* CPU cycles in order to make this area\\nbetter. But I do tend to agree that we've exhausted all the other\\noptions.\\n\\nBTW, I looked through costsize.c just now to see exactly what we are\\nusing disable_cost for, and it seemed like a majority of the cases are\\njust wrong. Where possible, we should implement a plan-type-disable\\nflag by not generating the associated Path in the first place, not by\\napplying disable_cost to it. But it looks like a lot of people have\\nerroneously copied the wrong logic. I would say that only these plan\\ntypes should use the disable_cost method:\\n\\n\\tseqscan\\n\\tnestloop join\\n\\tsort\\n\\nas those are the only ones where we risk not being able to make a\\nplan at all for lack of other alternatives.\\n\\nThere is also some weirdness around needing to force use of tidscan\\nif we have WHERE CURRENT OF. But perhaps a different hack could be\\nused for that.\\n\\nWe also have this for hashjoin:\\n\\n\\t * If the bucket holding the inner MCV would exceed hash_mem, we don't\\n\\t * want to hash unless there is really no other alternative, so apply\\n\\t * disable_cost.\\n\\nI'm content to leave that be, if we can't remove disable_cost\\nentirely.\\n\\nWhat I'm wondering at this point is whether we need to trouble with\\nimplementing the separate-disabledness-count method, if we trim back\\nthe number of places using disable_cost to the absolute minimum.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 04 May 2024 12:57:32 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Sun, 5 May 2024 at 04:57, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> David Rowley <dgrowleyml@gmail.com> writes:\\n> > That doesn't get you the benefits of fewer CPU cycles, but where did\\n> > that come from as a motive to change this? There's no shortage of\\n> > other ways to make the planner faster if that's an issue.\\n>\\n> The concern was to not *add* CPU cycles in order to make this area\\n> better. But I do tend to agree that we've exhausted all the other\\n> options.\\n\\nIt really looks to me that Robert was talking about not generating\\npaths for disabled path types. He did write \\\"just to save CPU cycles\\\"\\nin the paragraph I quoted.\\n\\nI think we should concern ourselves with adding overhead to add_path()\\n*only* when we actually see a patch which slows it down in a way that\\nwe can measure. I find it hard to imagine that adding a single\\ncomparison for every Path is measurable. Each of these paths has been\\npalloced and costed, both of which are significantly more expensive\\nthan adding another comparison to compare_path_costs_fuzzily(). I'm\\nonly willing for benchmarks on an actual patch to prove me wrong on\\nthat. Nothing else. add_path() has become a rat's nest of conditions\\nover the years and those seem to have made it without concerns about\\nperformance.\\n\\n> BTW, I looked through costsize.c just now to see exactly what we are\\n> using disable_cost for, and it seemed like a majority of the cases are\\n> just wrong. Where possible, we should implement a plan-type-disable\\n> flag by not generating the associated Path in the first place, not by\\n> applying disable_cost to it. But it looks like a lot of people have\\n> erroneously copied the wrong logic. I would say that only these plan\\n> types should use the disable_cost method:\\n>\\n> seqscan\\n> nestloop join\\n> sort\\n\\nI think this oversimplifies the situation. I only spent 30 seconds\\nlooking and I saw cases where this would cause issues. If\\nenable_hashagg is false, we could fail to produce some plans where the\\ntype is sortable but not hashable. There's also an issue with nested\\nloops being unable to FULL OUTER JOIN. However, I do agree that there\\nare some in there that are adding disable_cost that should be done by\\njust not creating the Path. enable_gathermerge is one.\\nenable_bitmapscan is probably another.\\n\\nI understand you only talked about the cases adding disable_cost in\\ncostsize.c. But just as a reminder, there are other things we need to\\nbe careful not to break. For example, enable_indexonlyscan=false\\nshould defer to still making an index scan. Nobody who disables\\nenable_indexonlyscan without disabling enable_indexscan wants queries\\nthat are eligible to use IOS to use seq scan instead. They'd still\\nwant Index Scan to be considered, otherwise they'd have disabled\\nenable_indexscan.\\n\\nDavid\\n\\n\\n\",\n \"msg_date\": \"Sun, 5 May 2024 12:07:30 +1200\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Sat, May 4, 2024 at 9:16 AM David Rowley <dgrowleyml@gmail.com> wrote:\\n> I don't think you'd need to wait longer than where we do set_cheapest\\n> and find no paths to find out that there's going to be a problem.\\n\\nI'm confused by this response, because I thought that the main point\\nof my previous email was explaining why that's not true. I showed an\\nexample where you do find paths at set_cheapest() time and yet are\\nunable to complete planning.\\n\\n> I don't think redoing planning is going to be easy or even useful. I\\n> mean what do you change when you replan? You can't just do\\n> enable_seqscan and enable_nestloop as if there's no index to provide\\n> sorted input and the plan requires some sort, then you still can't\\n> produce a plan. Adding enable_sort to the list does not give me much\\n> confidence we'll never fail to produce a plan either. It just seems\\n> impossible to know which of the disabled ones caused the RelOptInfo to\\n> have no paths. Also, you might end up enabling one that caused the\\n> planner to do something different than it would do today. For\\n> example, a Path that today incurs 2x disable_cost vs a Path that only\\n> receives 1x disable_cost might do something different if you just went\\n> and enabled a bunch of enable* GUCs before replanning.\\n\\nI agree that there are problems here, both in terms of implementation\\ncomplexity and also in terms of what behavior you actually get, but I\\ndo not think that a proposal which changes some current behavior\\nshould be considered dead on arrival. Whatever new behavior we might\\nwant to implement needs to make sense, and there need to be good\\nreasons for making whatever changes are contemplated, but I don't\\nthink we should take the position that it has to be identical to\\ncurrent.\\n\\n> I think the int Path.disabledness idea is worth coding up to try it.\\n> I imagine that a Path will incur the maximum of its subpath's\\n> disabledness's then add_path() just needs to prefer lower-valued\\n> disabledness Paths.\\n\\nIt definitely needs to be sum, not max. Otherwise you can't get the\\nmatest example from the regression tests right, where one child lacks\\nthe ability to comply with the GUC setting.\\n\\n> That doesn't get you the benefits of fewer CPU cycles, but where did\\n> that come from as a motive to change this? There's no shortage of\\n> other ways to make the planner faster if that's an issue.\\n\\nWell, I don't agree with that at all. If there are lots of ways to\\nmake the planner faster, we should definitely do a bunch of that\\nstuff, because \\\"will slow down the planner too much\\\" has been a\\nleading cause of proposed planner patches being rejected for as long\\nas I've been involved with the project. My belief was that we were\\nrather short of good ideas in that area, actually. But even if it's\\ntrue that we have lots of other ways to speed up the planner, that\\ndoesn't mean that it wouldn't be good to do it here, too.\\n\\nStepping back a bit, my current view of this area is: disable_cost is\\nhighly imperfect both as an idea and as implemented in PostgreSQL.\\nAlthough I'm discovering that the current implementation gets more\\nthings right than I had realized, it also sometimes gets things wrong.\\nThe original poster gave an example of that, and there are others.\\nFurthermore, the current implementation has some weird\\ninconsistencies. Therefore, I would like something better. Better, to\\nme, could mean any combination of (a) superior behavior, (b) superior\\nperformance, and (c) simpler, more elegant code. In a perfect world,\\nwe'd be able to come up with something that wins in all three of those\\nareas, but I'm not seeing a way to achieve that, so I'm trying to\\nfigure out what is achievable. And because we need to reach consensus\\non whatever is to be done, I'm sharing raw research results rather\\nthan just dropping a completed patch. I don't think it's at all easy\\nto understand what the realistic possibilities are in this area;\\ncertainly it isn't for me. At some point I'm hoping that there will be\\na patch (or a bunch of patches) that we can all agree are an\\nimprovement over now and the best we can reasonably do, but I don't\\nyet know what the shape of those will be, because I'm still trying to\\nunderstand (and document on-list) what all the problems are.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 May 2024 08:27:38 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Sat, May 4, 2024 at 12:57 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> There is also some weirdness around needing to force use of tidscan\\n> if we have WHERE CURRENT OF. But perhaps a different hack could be\\n> used for that.\\n\\nYeah, figuring out what to do about this was the trickiest part of the\\nexperimental patch that I wrote last week. The idea of the current\\ncode is that cost_qual_eval_walker charges disable_cost for\\nCurrentOfExpr, but cost_tidscan then subtracts disable_cost if\\ntidquals contains a CurrentOfExpr, so that we effectively disable\\neverything except TID scan paths and, I think, also any TID scan paths\\nthat don't use the CurrentOfExpr as a qual. I'm not entirely sure\\nwhether the last can happen, but I imagine that it might be possible\\nif the cursor refers to a query that itself contains some other kind\\nof TID qual.\\n\\nIt's not very clear that this mechanism is actually 100% reliable,\\nbecause we know it's possible in general for the costs of two paths to\\nbe different by more than disable_cost. Maybe that's not possible in\\nthis specific context, though: I'm not sure.\\n\\nThe approach I took for my experimental patch was pretty grotty, and\\nprobably not quite complete, but basically I defined the case where we\\ncurrently subtract out disable_cost as a \\\"forced TID-scan\\\". I passed\\naround a Boolean called forcedTidScan which gets set to true if we\\ndiscover that some plan is a forced TID-scan path, and then we discard\\nany other paths and then only add other forced TID-scan paths after\\nthat point. There can be more than one, because of parameterization.\\n\\nBut I think that the right thing to do is probably to pull some of the\\nlogic up out of create_tidscan_paths() and decide ONCE whether we're\\nin a forced TID-scan situation or not. If we are, then\\nset_plain_rel_pathlist() should arrange to create only forced TID-scan\\npaths; otherwise, it should proceed as it does now.\\n\\nMaybe if I try to do that I'll find problems, but the current approach\\nseems backwards to me, like going to a restaurant and ordering one of\\neverything on the menu, then cancelling all of the orders except the\\nstuff you actually want.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 May 2024 09:39:38 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Mon, May 6, 2024 at 9:39 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> It's not very clear that this mechanism is actually 100% reliable,\\n\\nIt isn't. Here's a test case. As a non-superuser, do this:\\n\\ncreate table foo (a int, b text, primary key (a));\\ninsert into foo values (1, 'Apple');\\nalter table foo enable row level security;\\nalter table foo force row level security;\\ncreate policy p1 on foo as permissive using (ctid in ('(0,1)', '(0,2)'));\\nbegin;\\ndeclare c cursor for select * from foo;\\nfetch from c;\\nexplain update foo set b = 'Manzana' where current of c;\\nupdate foo set b = 'Manzana' where current of c;\\n\\nThe explain produces this output:\\n\\n Update on foo (cost=10000000000.00..10000000008.02 rows=0 width=0)\\n -> Tid Scan on foo (cost=10000000000.00..10000000008.02 rows=1 width=38)\\n TID Cond: (ctid = ANY ('{\\\"(0,1)\\\",\\\"(0,2)\\\"}'::tid[]))\\n Filter: CURRENT OF c\\n\\nUnless I'm quite confused, the point of the code is to force\\nCurrentOfExpr to be a TID Cond, and it normally succeeds in doing so,\\nbecause WHERE CURRENT OF cursor_name has to be the one and only WHERE\\ncondition for a normal UPDATE. I tried various cases involving views\\nand CTEs and got nowhere. But then I wrote a patch to make the\\nregression tests fail if a baserel's restrictinfo list contains a\\nCurrentOfExpr and also some other qual, and a couple of row-level\\nsecurity tests failed (and nothing else). Which then allowed me to\\nconstruct the example above, where there are two possible TID quals\\nand the logic in tidpath.c latches onto the wrong one. The actual\\nUPDATE fails like this:\\n\\nERROR: WHERE CURRENT OF is not supported for this table type\\n\\n...because ExecEvalCurrentOfExpr supposes that the only way it can be\\nreached is for an FDW without the necessary support, but actually in\\nthis case it's planner error that gets us here.\\n\\nFortunately, there's no real reason for anyone to ever do something\\nlike this, or at least I can't see one, so the fact that it doesn't\\nwork probably doesn't really matter that much. And you can argue that\\nthe only problem here is that the costing hack just didn't get updated\\nfor RLS and now needs to be a bit more clever. But I think it'd be\\nbetter to find a way of making it less hacky. With the way the code is\\nstructured right now, the chances of anyone understanding that RLS\\nmight have an impact on its correctness were just about nil, IMHO.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 May 2024 13:51:30 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Mon, May 6, 2024 at 9:39 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n>> It's not very clear that this mechanism is actually 100% reliable,\\n\\n> It isn't. Here's a test case.\\n\\nVery interesting.\\n\\n> ... Which then allowed me to\\n> construct the example above, where there are two possible TID quals\\n> and the logic in tidpath.c latches onto the wrong one.\\n\\nHmm. Without having traced through it, I'm betting that the\\nCurrentOfExpr qual is rejected as a tidqual because it's not\\nconsidered leakproof. It's not obvious to me why we couldn't consider\\nit as leakproof, though. If we don't want to do that in general,\\nthen we need some kind of hack in TidQualFromRestrictInfo to accept\\nCurrentOfExpr quals anyway.\\n\\nIn general I think you're right that something less rickety than\\nthe disable_cost hack would be a good idea to ensure the desired\\nTidPath gets chosen, but this problem is not the fault of that.\\nWe're not making the TidPath with the correct contents in the first\\nplace.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 06 May 2024 14:44:40 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Robert Haas <robertmhaas@gmail.com> writes:\\n>> ... Which then allowed me to\\n>> construct the example above, where there are two possible TID quals\\n>> and the logic in tidpath.c latches onto the wrong one.\\n\\n> Hmm. Without having traced through it, I'm betting that the\\n> CurrentOfExpr qual is rejected as a tidqual because it's not\\n> considered leakproof.\\n\\nNah, I'm wrong: we do treat it as leakproof, and the comment about\\nthat in contain_leaked_vars_walker shows that the interaction with\\nRLS quals *was* thought about. What wasn't thought about was the\\npossibility of RLS quals that themselves could be usable as tidquals,\\nwhich breaks this assumption in TidQualFromRestrictInfoList:\\n\\n * Stop as soon as we find any usable CTID condition. In theory we\\n * could get CTID equality conditions from different AND'ed clauses,\\n * in which case we could try to pick the most efficient one. In\\n * practice, such usage seems very unlikely, so we don't bother; we\\n * just exit as soon as we find the first candidate.\\n\\nThe executor doesn't seem to be prepared to cope with multiple AND'ed\\nTID clauses (only OR'ed ones). So we need to fix this at least to the\\nextent of looking for a CurrentOfExpr qual, and preferring that over\\nanything else.\\n\\nI'm also now wondering about this assumption in the executor:\\n\\n /* CurrentOfExpr could never appear OR'd with something else */\\n Assert(list_length(tidstate->tss_tidexprs) == 1 ||\\n !tidstate->tss_isCurrentOf);\\n\\nIt still seems OK, because anything that might come in from RLS quals\\nwould be AND'ed not OR'ed with the CurrentOfExpr.\\n\\n> In general I think you're right that something less rickety than\\n> the disable_cost hack would be a good idea to ensure the desired\\n> TidPath gets chosen, but this problem is not the fault of that.\\n> We're not making the TidPath with the correct contents in the first\\n> place.\\n\\nStill true.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 06 May 2024 15:26:52 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Mon, May 6, 2024 at 3:26 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Nah, I'm wrong: we do treat it as leakproof, and the comment about\\n> that in contain_leaked_vars_walker shows that the interaction with\\n> RLS quals *was* thought about. What wasn't thought about was the\\n> possibility of RLS quals that themselves could be usable as tidquals,\\n> which breaks this assumption in TidQualFromRestrictInfoList:\\n>\\n> * Stop as soon as we find any usable CTID condition. In theory we\\n> * could get CTID equality conditions from different AND'ed clauses,\\n> * in which case we could try to pick the most efficient one. In\\n> * practice, such usage seems very unlikely, so we don't bother; we\\n> * just exit as soon as we find the first candidate.\\n\\nRight. I had noticed this but didn't spell it out.\\n\\n> The executor doesn't seem to be prepared to cope with multiple AND'ed\\n> TID clauses (only OR'ed ones). So we need to fix this at least to the\\n> extent of looking for a CurrentOfExpr qual, and preferring that over\\n> anything else.\\n>\\n> I'm also now wondering about this assumption in the executor:\\n>\\n> /* CurrentOfExpr could never appear OR'd with something else */\\n> Assert(list_length(tidstate->tss_tidexprs) == 1 ||\\n> !tidstate->tss_isCurrentOf);\\n>\\n> It still seems OK, because anything that might come in from RLS quals\\n> would be AND'ed not OR'ed with the CurrentOfExpr.\\n\\nThis stuff I had not noticed.\\n\\n> > In general I think you're right that something less rickety than\\n> > the disable_cost hack would be a good idea to ensure the desired\\n> > TidPath gets chosen, but this problem is not the fault of that.\\n> > We're not making the TidPath with the correct contents in the first\\n> > place.\\n>\\n> Still true.\\n\\nI'll look into this, unless you want to do it.\\n\\nIncidentally, another thing I just noticed is that\\nIsCurrentOfClause()'s test for (node->cvarno == rel->relid) is\\npossibly dead code. At least, there are no examples in our test suite\\nwhere it fails to hold. Which seems like it makes sense, because if it\\ndidn't, then how did the clause end up in baserestrictinfo? Maybe this\\nis worth keeping as defensive coding, or maybe it should be changed to\\nan Assert or something.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 May 2024 15:58:47 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> I'll look into this, unless you want to do it.\\n\\nI have a draft patch already. Need to add a test case.\\n\\n> Incidentally, another thing I just noticed is that\\n> IsCurrentOfClause()'s test for (node->cvarno == rel->relid) is\\n> possibly dead code. At least, there are no examples in our test suite\\n> where it fails to hold. Which seems like it makes sense, because if it\\n> didn't, then how did the clause end up in baserestrictinfo? Maybe this\\n> is worth keeping as defensive coding, or maybe it should be changed to\\n> an Assert or something.\\n\\nI wouldn't remove it, but maybe an Assert is good enough. The tests\\non Vars' varno should be equally pointless no?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 06 May 2024 16:10:44 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Mon, May 6, 2024 at 8:27 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> Stepping back a bit, my current view of this area is: disable_cost is\\n> highly imperfect both as an idea and as implemented in PostgreSQL.\\n> Although I'm discovering that the current implementation gets more\\n> things right than I had realized, it also sometimes gets things wrong.\\n> The original poster gave an example of that, and there are others.\\n> Furthermore, the current implementation has some weird\\n> inconsistencies. Therefore, I would like something better.\\n\\nFWIW I always found those weird inconsistencies to be annoying at\\nbest, and confusing at worst. I speak as somebody that uses\\ndisable_cost a lot.\\n\\nI certainly wouldn't ask anybody to make it a priority for that reason\\nalone -- it's not *that* bad. I've given my opinion on this because\\nit's already under discussion.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 May 2024 16:30:20 -0400\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Mon, May 6, 2024 at 4:30 PM Peter Geoghegan <pg@bowt.ie> wrote:\\n> FWIW I always found those weird inconsistencies to be annoying at\\n> best, and confusing at worst. I speak as somebody that uses\\n> disable_cost a lot.\\n>\\n> I certainly wouldn't ask anybody to make it a priority for that reason\\n> alone -- it's not *that* bad. I've given my opinion on this because\\n> it's already under discussion.\\n\\nThanks, it's good to have other perspectives.\\n\\nHere are some patches for discussion.\\n\\n0001 gets rid of disable_cost as a mechanism for forcing a TID scan\\nplan to be chosen when CurrentOfExpr is present. Instead, it arranges\\nto generate only the valid path when that case occurs, and skip\\neverything else. I think this is a good cleanup, and it doesn't seem\\ntotally impossible that it actually prevents a failure in some extreme\\ncase.\\n\\n0002 cleans up the behavior of enable_indexscan and\\nenable_indexonlyscan. Currently, setting enable_indexscan=false adds\\ndisable_cost to both the cost of index scans and the cost of\\nindex-only scans. I think that's indefensible and, in fact, a bug,\\nalthough I believe David Rowley disagrees. With this patch, we simply\\ndon't generate index scans if enable_indexscan=false, and we don't\\ngenerate index-only scans if enable_indexonlyscan=false, which seems a\\nlot more consistent to me. However, I did revise one major thing from\\nthe patch I posted before, per feedback from David Rowley and also per\\nmy own observations: in this version, if enable_indexscan=true and\\nenable_indexonlyscan=false, we'll generate index-scan paths for any\\ncases where, with both set to true, we would have only generated\\nindex-only scan paths. That change makes the behavior of this patch a\\nlot more comprehensible and intuitive: the only regression test\\nchanges are places where somebody expected that they could disable\\nboth index scans and index-only scans by setting\\nenable_indexscan=false.\\n\\n0003 and 0004 extend the approach of \\\"just don't generate the disabled\\npath\\\" to bitmap scans and gather merge, respectively. I think these\\nare more debatable, mostly because it's not clear how far we can\\nreally take this approach. Neither breaks any test cases, and 0003 is\\nclosely related to the work done in 0002, which seems like a point in\\nits favor. 0004 was simply the only other case where it was obvious to\\nme that this kind of approach made sense. In my view, it makes most\\nsense to use this kind of approach for planner behaviors that seem\\nlike they're sort of optional: like if you don't use gather merge, you\\ncan still use gather, and if you don't use index scans, you can still\\nuse sequential scans. With all these patches applied, the remaining\\ncases where we rely on disable_cost are:\\n\\nsequential scans\\nsorts\\nhash aggregation\\nall 3 join types\\nhash joins where a bucket holding the inner MCV would exceed hash_mem\\n\\nSequential scans are clearly a last-ditch method. I find it a bit hard\\nto decide whether hashing or sorting is the default, especially giving\\nthe asymmetry between enable_sort - presumptively anywhere - and\\nenable_hashagg - specific to aggregation. As for the join types, it's\\ntempting to consider nested-loop the default type -- it's the only way\\nto satisfy parameterizations, for instance -- but the fact that it's\\nthe only method that can't do a full join undermines that position in\\nmy book. But, I don't want to pretend like I have all the answers\\nhere, either; I'm just sharing some thoughts.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Tue, 7 May 2024 16:19:22 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, May 7, 2024 at 4:19 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n> Here are some patches for discussion.\\n\\nWell, that didn't generate much discussion, but here I am trying\\nagain. Here I've got patches 0001 and 0002 from my previous posting;\\nI've dropped 0003 and 0004 from the previous set for now so as not to\\ndistract from the main event, but they may still be a good idea.\\nInstead I've got an 0003 and an 0004 that implement the \\\"count of\\ndisabled nodes\\\" approach that we have discussed previously. This seems\\nto work fine, unlike the approaches I tried earlier. I think this is\\nthe right direction to go, but I'd like to know what concerns people\\nmight have.\\n\\nThis doesn't completely remove disable_cost, because hash joins still\\nadd it to the cost when it's impossible to fit the MCV value into\\nwork_mem. I'm not sure what to do with that. Continuing to use\\ndisable_cost in that one scenario seems OK to me. We could\\nalternatively make that scenario bump disabled_nodes, but I don't\\nreally want to confuse the planner not wanting to do something with\\nthe user telling the planner not to do something, so I don't think\\nthat's a good idea. Or we could rejigger things so that in that case\\nwe don't generate the plan at all. I'm not sure why we don't do that\\nalready, actually.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Wed, 12 Jun 2024 11:35:48 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2024-06-12 11:35:48 -0400, Robert Haas wrote:\\n> Subject: [PATCH v2 3/4] Treat the # of disabled nodes in a path as a separate\\n> cost metric.\\n> \\n> Previously, when a path type was disabled by e.g. enable_seqscan=false,\\n> we either avoided generating that path type in the first place, or\\n> more commonly, we added a large constant, called disable_cost, to the\\n> estimated startup cost of that path. This latter approach can distort\\n> planning. For instance, an extremely expensive non-disabled path\\n> could seem to be worse than a disabled path, especially if the full\\n> cost of that path node need not be paid (e.g. due to a Limit).\\n> Or, as in the regression test whose expected output changes with this\\n> commit, the addition of disable_cost can make two paths that would\\n> normally be distinguishible cost seem to have fuzzily the same cost.\\n> \\n> To fix that, we now count the number of disabled path nodes and\\n> consider that a high-order component of both the cost. Hence, the\\n> path list is now sorted by disabled_nodes and then by total_cost,\\n> instead of just by the latter, and likewise for the partial path list.\\n> It is important that this number is a count and not simply a Boolean;\\n> else, as soon as we're unable to respect disabled path types in all\\n> portions of the path, we stop trying to avoid them where we can.\\n\\n\\n> \\tif (criterion == STARTUP_COST)\\n> \\t{\\n> \\t\\tif (path1->startup_cost < path2->startup_cost)\\n> @@ -118,6 +127,15 @@ compare_fractional_path_costs(Path *path1, Path *path2,\\n> \\tCost\\t\\tcost1,\\n> \\t\\t\\t\\tcost2;\\n> \\n> +\\t/* Number of disabled nodes, if different, trumps all else. */\\n> +\\tif (unlikely(path1->disabled_nodes != path2->disabled_nodes))\\n> +\\t{\\n> +\\t\\tif (path1->disabled_nodes < path2->disabled_nodes)\\n> +\\t\\t\\treturn -1;\\n> +\\t\\telse\\n> +\\t\\t\\treturn +1;\\n> +\\t}\\n\\nI suspect it's going to be ok, because the branch is going to be very\\npredictable in normal workloads, but I still worry a bit about making\\ncompare_path_costs_fuzzily() more expensive. For more join-heavy queries it\\ncan really show up and there's plenty ORM generated join-heavy query\\nworkloads.\\n\\nIf costs were 32 bit integers, I'd have suggested just stashing the disabled\\ncounts in the upper 32 bits of a 64bit integer. But ...\\n\\n<can't resist trying if I see overhead>\\n\\n\\nIn an extreme case i can see a tiny bit of overhead, but not enough to be\\nworth worrying about. Mostly because we're so profligate in doing\\nbms_overlap() that cost comparisons don't end up mattering as much - I seem to\\nrecall that being different in the not distant past though.\\n\\n\\nAside: I'm somewhat confused by add_paths_to_joinrel()'s handling of\\nmergejoins_allowed. If mergejoins are disabled we end up reaching\\nmatch_unsorted_outer() in more cases than with mergejoins enabled. E.g. we\\nonly set mergejoin_enabled for right joins inside select_mergejoin_clauses(),\\nbut we don't call select_mergejoin_clauses() if !enable_mergejoin and jointype\\n!= FULL. I, what?\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 12 Jun 2024 11:11:23 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 12, 2024 at 2:11 PM Andres Freund <andres@anarazel.de> wrote:\\n> <can't resist trying if I see overhead>\\n>\\n> In an extreme case i can see a tiny bit of overhead, but not enough to be\\n> worth worrying about. Mostly because we're so profligate in doing\\n> bms_overlap() that cost comparisons don't end up mattering as much - I seem to\\n> recall that being different in the not distant past though.\\n\\nThere are very few things I love more than when you can't resist\\ntrying to break my patches and yet fail to find a problem. Granted the\\nlatter part only happens once a century or so, but I'll take it.\\n\\n> Aside: I'm somewhat confused by add_paths_to_joinrel()'s handling of\\n> mergejoins_allowed. If mergejoins are disabled we end up reaching\\n> match_unsorted_outer() in more cases than with mergejoins enabled. E.g. we\\n> only set mergejoin_enabled for right joins inside select_mergejoin_clauses(),\\n> but we don't call select_mergejoin_clauses() if !enable_mergejoin and jointype\\n> != FULL. I, what?\\n\\nI agree this logic is extremely confusing, but \\\"we only set\\nmergejoin_enabled for right joins inside select_mergejoin_clauses()\\\"\\ndoesn't seem to be true. It starts out true, and always stays true\\nexcept for right, right-anti, and full joins, where\\nselect_mergejoin_clauses() can set it to false. Since the call to\\nmatch_unsorted_outer() is gated by mergejoin_enabled, you might think\\nthat we'd skip considering nested loops on the strength of not being\\nable to do a merge join, but comment \\\"2.\\\" in add_paths_to_joinrel\\nexplains that the join types for which mergejoin_enabled can end up\\nfalse aren't supported by nested loops anyway. Still, this logic is\\nreally tortured.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 12 Jun 2024 14:33:31 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2024-06-12 14:33:31 -0400, Robert Haas wrote:\\n> On Wed, Jun 12, 2024 at 2:11 PM Andres Freund <andres@anarazel.de> wrote:\\n> > <can't resist trying if I see overhead>\\n> >\\n> > In an extreme case i can see a tiny bit of overhead, but not enough to be\\n> > worth worrying about. Mostly because we're so profligate in doing\\n> > bms_overlap() that cost comparisons don't end up mattering as much - I seem to\\n> > recall that being different in the not distant past though.\\n> \\n> There are very few things I love more than when you can't resist\\n> trying to break my patches and yet fail to find a problem. Granted the\\n> latter part only happens once a century or so, but I'll take it.\\n\\n:)\\n\\n\\nToo high cost in path cost comparison is what made me look at the PG code for\\nthe first time, IIRC :)\\n\\n\\n\\n> > Aside: I'm somewhat confused by add_paths_to_joinrel()'s handling of\\n> > mergejoins_allowed. If mergejoins are disabled we end up reaching\\n> > match_unsorted_outer() in more cases than with mergejoins enabled. E.g. we\\n> > only set mergejoin_enabled for right joins inside select_mergejoin_clauses(),\\n> > but we don't call select_mergejoin_clauses() if !enable_mergejoin and jointype\\n> > != FULL. I, what?\\n> \\n> I agree this logic is extremely confusing, but \\\"we only set\\n> mergejoin_enabled for right joins inside select_mergejoin_clauses()\\\"\\n> doesn't seem to be true.\\n\\nSorry, should have been more precise. With \\\"set\\\" I didn't mean set to true,\\nbut that that it's only modified within select_mergejoin_clauses().\\n\\n\\n> It starts out true, and always stays true except for right, right-anti, and\\n> full joins, where select_mergejoin_clauses() can set it to false. Since the\\n> call to match_unsorted_outer() is gated by mergejoin_enabled, you might\\n> think that we'd skip considering nested loops on the strength of not being\\n> able to do a merge join, but comment \\\"2.\\\" in add_paths_to_joinrel explains\\n> that the join types for which mergejoin_enabled can end up false aren't\\n> supported by nested loops anyway. Still, this logic is really tortured.\\n\\nAgree that that's the logic - but doesn't that mean we'll consider nestloops\\nfor e.g. right joins iff enable_mergejoin=false?\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 12 Jun 2024 11:48:35 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 12, 2024 at 2:48 PM Andres Freund <andres@anarazel.de> wrote:\\n> Sorry, should have been more precise. With \\\"set\\\" I didn't mean set to true,\\n> but that that it's only modified within select_mergejoin_clauses().\\n\\nOh. \\\"set\\\" has more than one relevant meaning here.\\n\\n> > It starts out true, and always stays true except for right, right-anti, and\\n> > full joins, where select_mergejoin_clauses() can set it to false. Since the\\n> > call to match_unsorted_outer() is gated by mergejoin_enabled, you might\\n> > think that we'd skip considering nested loops on the strength of not being\\n> > able to do a merge join, but comment \\\"2.\\\" in add_paths_to_joinrel explains\\n> > that the join types for which mergejoin_enabled can end up false aren't\\n> > supported by nested loops anyway. Still, this logic is really tortured.\\n>\\n> Agree that that's the logic - but doesn't that mean we'll consider nestloops\\n> for e.g. right joins iff enable_mergejoin=false?\\n\\nNo, because that function has its own internal guards. See nestjoinOK.\\n\\nBut don't misunderstand me: I'm not defending the status quo. The\\nwhole thing seems like a Rube Goldberg machine to me.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 12 Jun 2024 15:11:53 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 12, 2024 at 11:35 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> Well, that didn't generate much discussion, but here I am trying\\n> again. Here I've got patches 0001 and 0002 from my previous posting;\\n> I've dropped 0003 and 0004 from the previous set for now so as not to\\n> distract from the main event, but they may still be a good idea.\\n> Instead I've got an 0003 and an 0004 that implement the \\\"count of\\n> disabled nodes\\\" approach that we have discussed previously. This seems\\n> to work fine, unlike the approaches I tried earlier. I think this is\\n> the right direction to go, but I'd like to know what concerns people\\n> might have.\\n\\nHere is a rebased patch set, where I also fixed pgindent damage and a\\ncouple of small oversights in 0004.\\n\\nI am hoping to get these committed some time in July. So if somebody\\nthinks that's too soon or thinks it shouldn't happen at all, please\\ndon't wait too long to let me know about that.\\n\\nThanks,\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Fri, 28 Jun 2024 11:46:24 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On 28/06/2024 18:46, Robert Haas wrote:\\n> On Wed, Jun 12, 2024 at 11:35 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n>> Well, that didn't generate much discussion, but here I am trying\\n>> again. Here I've got patches 0001 and 0002 from my previous posting;\\n>> I've dropped 0003 and 0004 from the previous set for now so as not to\\n>> distract from the main event, but they may still be a good idea.\\n>> Instead I've got an 0003 and an 0004 that implement the \\\"count of\\n>> disabled nodes\\\" approach that we have discussed previously. This seems\\n>> to work fine, unlike the approaches I tried earlier. I think this is\\n>> the right direction to go, but I'd like to know what concerns people\\n>> might have.\\n> \\n> Here is a rebased patch set, where I also fixed pgindent damage and a\\n> couple of small oversights in 0004.\\n> \\n> I am hoping to get these committed some time in July. So if somebody\\n> thinks that's too soon or thinks it shouldn't happen at all, please\\n> don't wait too long to let me know about that.\\n\\nv3-0001-Remove-grotty-use-of-disable_cost-for-TID-scan-pl.patch:\\n\\n+1, this seems ready for commit\\n\\nv3-0002-Rationalize-behavior-of-enable_indexscan-and-enab.patch:\\n\\nI fear this will break people's applications, if they are currently \\nforcing a sequential scan with \\\"set enable_indexscan=off\\\". Now they will \\nneed to do \\\"set enable_indexscan=off; set enable_indexonlyscan=off\\\" for \\nthe same effect. Maybe it's acceptable, disabling sequential scans to \\nforce an index scan is much more common than the other way round.\\n\\nv3-0003-Treat-number-of-disabled-nodes-in-a-path-as-a-sep.patch:\\n\\n> @@ -1318,6 +1342,12 @@ cost_tidscan(Path *path, PlannerInfo *root,\\n> \\tstartup_cost += path->pathtarget->cost.startup;\\n> \\trun_cost += path->pathtarget->cost.per_tuple * path->rows;\\n> \\n> +\\t/*\\n> +\\t * There are assertions above verifying that we only reach this function\\n> +\\t * either when enable_tidscan=true or when the TID scan is the only legal\\n> +\\t * path, so it's safe to set disabled_nodes to zero here.\\n> +\\t */\\n> +\\tpath->disabled_nodes = 0;\\n> \\tpath->startup_cost = startup_cost;\\n> \\tpath->total_cost = startup_cost + run_cost;\\n> }\\n\\nSo if you have enable_tidscan=off, and have a query with \\\"WHERE CURRENT \\nOF foo\\\" that is planned with a TID scan, we set disable_nodes = 0? That \\nsounds wrong, shouldn't disable_nodes be 1 in that case? It probably \\ncannot affect the rest of the plan, given that \\\"WHERE CURRENT OF\\\" is \\nonly valid in an UPDATE or DELETE, but still. At least it deserves a \\nbetter explanation in the comment.\\n\\n> diff --git a/src/backend/optimizer/plan/createplan.c b/src/backend/optimizer/plan/createplan.c\\n> index 6b64c4a362..20236e8c4d 100644\\n> --- a/src/backend/optimizer/plan/createplan.c\\n> +++ b/src/backend/optimizer/plan/createplan.c\\n> @@ -25,6 +25,7 @@\\n> #include \\\"nodes/extensible.h\\\"\\n> #include \\\"nodes/makefuncs.h\\\"\\n> #include \\\"nodes/nodeFuncs.h\\\"\\n> +#include \\\"nodes/print.h\\\"\\n> #include \\\"optimizer/clauses.h\\\"\\n> #include \\\"optimizer/cost.h\\\"\\n> #include \\\"optimizer/optimizer.h\\\"\\n\\nleft over from debugging?\\n\\n> @@ -68,6 +68,15 @@ static bool pathlist_is_reparameterizable_by_child(List *pathlist,\\n> int\\n> compare_path_costs(Path *path1, Path *path2, CostSelector criterion)\\n> {\\n> +\\t/* Number of disabled nodes, if different, trumps all else. */\\n> +\\tif (unlikely(path1->disabled_nodes != path2->disabled_nodes))\\n> +\\t{\\n> +\\t\\tif (path1->disabled_nodes < path2->disabled_nodes)\\n> +\\t\\t\\treturn -1;\\n> +\\t\\telse\\n> +\\t\\t\\treturn +1;\\n> +\\t}\\n> +\\n> \\tif (criterion == STARTUP_COST)\\n> \\t{\\n> \\t\\tif (path1->startup_cost < path2->startup_cost)\\n\\nIs \\\"unlikely()\\\" really appropriate here (and elsewhere in the patch)? If \\nyou run with enable_seqscan=off but have no indexes, you could take that \\npath pretty often.\\n\\nIf this function needs optimizing, I'd suggest splitting it into two \\nfunctions, one for comparing the startup cost and another for the total \\ncost. Almost all callers pass a constant for that argument, so they \\nmight as well call the correct function directly and avoid the branch \\nfor that.\\n\\n> @@ -658,6 +704,20 @@ add_path_precheck(RelOptInfo *parent_rel,\\n> \\t\\tPath\\t *old_path = (Path *) lfirst(p1);\\n> \\t\\tPathKeysComparison keyscmp;\\n> \\n> +\\t\\t/*\\n> +\\t\\t * Since the pathlist is sorted by disabled_nodes and then by\\n> +\\t\\t * total_cost, we can stop looking once we reach a path with more\\n> +\\t\\t * disabled nodes, or the same number of disabled nodes plus a\\n> +\\t\\t * total_cost larger than the new path's.\\n> +\\t\\t */\\n> +\\t\\tif (unlikely(old_path->disabled_nodes != disabled_nodes))\\n> +\\t\\t{\\n> +\\t\\t\\tif (disabled_nodes < old_path->disabled_nodes)\\n> +\\t\\t\\t\\tbreak;\\n> +\\t\\t}\\n> +\\t\\telse if (total_cost <= old_path->total_cost * STD_FUZZ_FACTOR)\\n> +\\t\\t\\tbreak;\\n> +\\n> \\t\\t/*\\n> \\t\\t * We are looking for an old_path with the same parameterization (and\\n> \\t\\t * by assumption the same rowcount) that dominates the new path on\\n> @@ -666,39 +726,27 @@ add_path_precheck(RelOptInfo *parent_rel,\\n> \\t\\t *\\n> \\t\\t * Cost comparisons here should match compare_path_costs_fuzzily.\\n> \\t\\t */\\n> -\\t\\tif (total_cost > old_path->total_cost * STD_FUZZ_FACTOR)\\n> +\\t\\t/* new path can win on startup cost only if consider_startup */\\n> +\\t\\tif (startup_cost > old_path->startup_cost * STD_FUZZ_FACTOR ||\\n> +\\t\\t\\t!consider_startup)\\n> \\t\\t{\\n\\nThe \\\"Cost comparisons here should match compare_path_costs_fuzzily\\\" \\ncomment also applies to the check on total_cost that you moved up. Maybe \\nmove up the comment to the beginning of the loop.\\n\\nv3-0004-Show-number-of-disabled-nodes-in-EXPLAIN-ANALYZE-.patch:\\n\\nIt's surprising that the \\\"Disable Nodes\\\" is printed even with the COSTS \\nOFF option. It's handy for our regression tests, it's good to print them \\nthere, but it feels wrong.\\n\\nCould we cram it into the \\\"cost=... rows=...\\\" part? And perhaps a marker \\nthat a node was disabled would be more user friendly than showing the \\ncumulative count? Something like:\\n\\npostgres=# set enable_material=off;\\nSET\\npostgres=# set enable_seqscan=off;\\nSET\\npostgres=# set enable_bitmapscan=off;\\nSET\\npostgres=# explain select * from foo, bar;\\n QUERY PLAN \\n\\n------------------------------------------------------------------------------------\\n Nested Loop (cost=0.15..155632.40 rows=6502500 width=8)\\n -> Index Only Scan using foo_i_idx on foo (cost=0.15..82.41 \\nrows=2550 width=4)\\n -> Seq Scan on bar (cost=0.00..35.50 (disabled) rows=2550 width=4)\\n(5 rows)\\n\\n-- \\nHeikki Linnakangas\\nNeon (https://neon.tech)\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Jul 2024 17:57:27 +0300\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Thanks for the review!\\n\\nOn Tue, Jul 2, 2024 at 10:57 AM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n> v3-0001-Remove-grotty-use-of-disable_cost-for-TID-scan-pl.patch:\\n>\\n> +1, this seems ready for commit\\n\\nCool.\\n\\n> v3-0002-Rationalize-behavior-of-enable_indexscan-and-enab.patch:\\n>\\n> I fear this will break people's applications, if they are currently\\n> forcing a sequential scan with \\\"set enable_indexscan=off\\\". Now they will\\n> need to do \\\"set enable_indexscan=off; set enable_indexonlyscan=off\\\" for\\n> the same effect. Maybe it's acceptable, disabling sequential scans to\\n> force an index scan is much more common than the other way round.\\n\\nWell, I think it's pretty important that the GUC does what the name\\nand documentation say it does. One could of course argue that we ought\\nnot to have two different GUCs -- or perhaps even that we ought not to\\nhave two different plan nodes -- and I think those arguments might be\\nquite defensible. One could also argue for another interface, like a\\nGUC enable_indexscan and a value that is a comma-separated list\\nconsisting of plain, bitmap, and index-only, or\\nnone/0/false/any/1/true -- and that might also be quite defensible.\\nBut I don't think one can have a GUC called enable_indexscan and\\nanother GUC called enable_indexonlyscan and argue that it's OK for the\\nformer one to affect both kinds of scans. That's extremely confusing\\nand, well, just plain wrong. I think this is a bug, and I'm not going\\nto back-patch the fix precisely because of the considerations you\\nnote, but I really don't think we can leave it like this. The current\\nbehavior is so nonsensical that the code is essentially unmaintable,\\nor at least I think it is.\\n\\n> v3-0003-Treat-number-of-disabled-nodes-in-a-path-as-a-sep.patch:\\n>\\n> > @@ -1318,6 +1342,12 @@ cost_tidscan(Path *path, PlannerInfo *root,\\n> > startup_cost += path->pathtarget->cost.startup;\\n> > run_cost += path->pathtarget->cost.per_tuple * path->rows;\\n> >\\n> > + /*\\n> > + * There are assertions above verifying that we only reach this function\\n> > + * either when enable_tidscan=true or when the TID scan is the only legal\\n> > + * path, so it's safe to set disabled_nodes to zero here.\\n> > + */\\n> > + path->disabled_nodes = 0;\\n> > path->startup_cost = startup_cost;\\n> > path->total_cost = startup_cost + run_cost;\\n> > }\\n>\\n> So if you have enable_tidscan=off, and have a query with \\\"WHERE CURRENT\\n> OF foo\\\" that is planned with a TID scan, we set disable_nodes = 0? That\\n> sounds wrong, shouldn't disable_nodes be 1 in that case? It probably\\n> cannot affect the rest of the plan, given that \\\"WHERE CURRENT OF\\\" is\\n> only valid in an UPDATE or DELETE, but still. At least it deserves a\\n> better explanation in the comment.\\n\\nSo, right now, when the planner disregards enable_WHATEVER because it\\nthinks it's the only way to implement something, it doesn't add\\ndisable_cost. So, I made the patch not increment disabled_nodes in\\nthat case. Maybe we want to rethink that choice at some point, but it\\ndoesn't seem like a good idea to do it right now. I've found while\\nworking on this stuff that it's super-easy to have seemingly innocuous\\nchanges disturb regression test results, and I don't really want to\\nhave a bunch of extra regression test changes that are due to\\nrethinking things other than disable_cost -> disabled_nodes. So for\\nnow I'd like to increment disabled_nodes in just the cases where we\\ncurrently add disable_cost.\\n\\n> left over from debugging?\\n\\nYeah, will fix.\\n\\n> Is \\\"unlikely()\\\" really appropriate here (and elsewhere in the patch)? If\\n> you run with enable_seqscan=off but have no indexes, you could take that\\n> path pretty often.\\n\\nThat's true, but I think it's right to assume that's the uncommon\\ncase. If we speed up planning for people who disabled sequential scans\\nand slow it down for people running with a normal planner\\nconfiguration, no one will thank us.\\n\\n> If this function needs optimizing, I'd suggest splitting it into two\\n> functions, one for comparing the startup cost and another for the total\\n> cost. Almost all callers pass a constant for that argument, so they\\n> might as well call the correct function directly and avoid the branch\\n> for that.\\n\\nThat's not a bad idea but seems like a separate patch.\\n\\n> The \\\"Cost comparisons here should match compare_path_costs_fuzzily\\\"\\n> comment also applies to the check on total_cost that you moved up. Maybe\\n> move up the comment to the beginning of the loop.\\n\\nWill have a look.\\n\\n> v3-0004-Show-number-of-disabled-nodes-in-EXPLAIN-ANALYZE-.patch:\\n>\\n> It's surprising that the \\\"Disable Nodes\\\" is printed even with the COSTS\\n> OFF option. It's handy for our regression tests, it's good to print them\\n> there, but it feels wrong.\\n\\nI'm open to doing what people think is best here. Although we're\\nregarding them as part of the cost for purposes of how to compare\\npaths, they're not unpredictable in the way that costs are, so I think\\nthe current handling is defensible and, as you say, it's useful for\\nthe regression tests. However, I'm not going to fight tooth and nail\\nif people really want it the other way.\\n\\n> Could we cram it into the \\\"cost=... rows=...\\\" part? And perhaps a marker\\n> that a node was disabled would be more user friendly than showing the\\n> cumulative count? Something like:\\n\\nThe problem is that we'd have to derive that. What we actually know is\\nthe disable count; to figure out whether the node itself was disabled,\\nwe'd have to subtract the value for the underlying nodes back out.\\nThat seems like it might be buggy or confusing.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Jul 2024 13:24:44 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Tue, Jul 2, 2024 at 10:57 AM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n>> I fear this will break people's applications, if they are currently\\n>> forcing a sequential scan with \\\"set enable_indexscan=off\\\". Now they will\\n>> need to do \\\"set enable_indexscan=off; set enable_indexonlyscan=off\\\" for\\n>> the same effect. Maybe it's acceptable, disabling sequential scans to\\n>> force an index scan is much more common than the other way round.\\n\\n> But I don't think one can have a GUC called enable_indexscan and\\n> another GUC called enable_indexonlyscan and argue that it's OK for the\\n> former one to affect both kinds of scans. That's extremely confusing\\n> and, well, just plain wrong.\\n\\nFWIW, I disagree completely. I think it's entirely natural to\\nconsider bitmap index scans to be a subset of index scans, so that\\nenable_indexscan should affect both. I admit that the current set\\nof GUCs doesn't let you force a bitmap scan over a plain one, but\\nI can't recall many people complaining about that. I don't follow\\nthe argument that this definition is somehow unmaintainable, either.\\n\\n>> Could we cram it into the \\\"cost=... rows=...\\\" part? And perhaps a marker\\n>> that a node was disabled would be more user friendly than showing the\\n>> cumulative count? Something like:\\n\\n> The problem is that we'd have to derive that.\\n\\nThe other problem is it'd break an awful lot of client code that knows\\nthe format of those lines. (Sure, by now all such code should have\\nbeen rewritten to look at JSON or other more machine-friendly output\\nformats ... but since we haven't even done that in our own regression\\ntests, we should know better than to assume other people have done it.)\\n\\nI'm not really convinced that we need to show anything about this.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 02 Jul 2024 13:40:38 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 2, 2024 at 1:40 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> FWIW, I disagree completely. I think it's entirely natural to\\n> consider bitmap index scans to be a subset of index scans, so that\\n> enable_indexscan should affect both. I admit that the current set\\n> of GUCs doesn't let you force a bitmap scan over a plain one, but\\n> I can't recall many people complaining about that. I don't follow\\n> the argument that this definition is somehow unmaintainable, either.\\n\\nWell... but that's not what the GUC does either. Not now, and not with\\nthe patch.\\n\\nWhat happens right now is:\\n\\n- If you set enable_indexscan=false, then disable_cost is added to the\\ncost of index scan paths and the cost of index-only scan paths.\\n\\n- If you set enable_indexonlyscan=false, then index-only scan paths\\nare not generated at all.\\n\\nBitmap scans are controlled by enable_bitmapscan.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Jul 2024 13:54:02 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> What happens right now is:\\n\\n> - If you set enable_indexscan=false, then disable_cost is added to the\\n> cost of index scan paths and the cost of index-only scan paths.\\n\\n> - If you set enable_indexonlyscan=false, then index-only scan paths\\n> are not generated at all.\\n\\nHm. The first part of that seems pretty weird to me --- why don't\\nwe simply not generate the paths at all? There is no case AFAIR\\nwhere that would prevent us from generating a valid plan.\\n\\n(I do seem to recall that index-only paths are built on top of regular\\nindex paths, so that there might be implementation issues with trying\\nto build the former and not the latter. But you've probably looked\\nat that far more recently than I.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 02 Jul 2024 14:37:25 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 2, 2024 at 2:37 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Robert Haas <robertmhaas@gmail.com> writes:\\n> > What happens right now is:\\n>\\n> > - If you set enable_indexscan=false, then disable_cost is added to the\\n> > cost of index scan paths and the cost of index-only scan paths.\\n>\\n> > - If you set enable_indexonlyscan=false, then index-only scan paths\\n> > are not generated at all.\\n>\\n> Hm. The first part of that seems pretty weird to me --- why don't\\n> we simply not generate the paths at all? There is no case AFAIR\\n> where that would prevent us from generating a valid plan.\\n\\nWell, yeah.\\n\\nWhat the patch does is: if you set either enable_indexscan=false or\\nenable_indexonlyscan=false, then the corresponding path type is not\\ngenerated, and the other is unaffected. To me, that seems like the\\nlogical way to clean this up.\\n\\nOne could argue for other things, of course. And maybe those other\\nthings are fine, if they're properly justified and documented.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Jul 2024 15:28:38 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> What the patch does is: if you set either enable_indexscan=false or\\n> enable_indexonlyscan=false, then the corresponding path type is not\\n> generated, and the other is unaffected. To me, that seems like the\\n> logical way to clean this up.\\n\\n> One could argue for other things, of course. And maybe those other\\n> things are fine, if they're properly justified and documented.\\n\\n[ shrug... ] This isn't a hill that I'm prepared to die on.\\nBut I see no good reason to change the very long-standing\\nbehaviors of these GUCs.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 02 Jul 2024 15:36:56 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 2, 2024 at 3:36 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > One could argue for other things, of course. And maybe those other\\n> > things are fine, if they're properly justified and documented.\\n>\\n> [ shrug... ] This isn't a hill that I'm prepared to die on.\\n> But I see no good reason to change the very long-standing\\n> behaviors of these GUCs.\\n\\nWell, I don't really know where to go from here. I mean, I think that\\nthree committers (David, Heikki, yourself) have expressed some\\nconcerns about changing the behavior. So maybe we shouldn't. But I\\ndon't understand how it's reasonable to have two very similarly named\\nGUCs behave (1) inconsistently with each other and (2) in a way that\\ncannot be guessed from the documentation.\\n\\nI feel like we're just clinging to legacy behavior on the theory that\\nsomebody, somewhere might be relying on it in some way, which they\\ncertainly might be. But that doesn't seem like a great reason, either.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Jul 2024 15:54:17 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> Well, I don't really know where to go from here. I mean, I think that\\n> three committers (David, Heikki, yourself) have expressed some\\n> concerns about changing the behavior. So maybe we shouldn't. But I\\n> don't understand how it's reasonable to have two very similarly named\\n> GUCs behave (1) inconsistently with each other and (2) in a way that\\n> cannot be guessed from the documentation.\\n\\nIf the documentation isn't adequate, that's certainly an improvable\\nsituation. It doesn't seem hard:\\n\\n- Enables or disables the query planner's use of index-scan plan\\n- types. The default is <literal>on</literal>.\\n+ Enables or disables the query planner's use of index-scan plan\\n+ types (including index-only scans).\\n+ The default is <literal>on</literal>.\\n\\nMore to the point, if we do change the longstanding meaning of this\\nGUC, that will *also* require documentation work IMO.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 02 Jul 2024 16:43:41 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On 02/07/2024 22:54, Robert Haas wrote:\\n> On Tue, Jul 2, 2024 at 3:36 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>>> One could argue for other things, of course. And maybe those other\\n>>> things are fine, if they're properly justified and documented.\\n>>\\n>> [ shrug... ] This isn't a hill that I'm prepared to die on.\\n>> But I see no good reason to change the very long-standing\\n>> behaviors of these GUCs.\\n> \\n> Well, I don't really know where to go from here. I mean, I think that\\n> three committers (David, Heikki, yourself) have expressed some\\n> concerns about changing the behavior. So maybe we shouldn't. But I\\n> don't understand how it's reasonable to have two very similarly named\\n> GUCs behave (1) inconsistently with each other and (2) in a way that\\n> cannot be guessed from the documentation.\\n> \\n> I feel like we're just clinging to legacy behavior on the theory that\\n> somebody, somewhere might be relying on it in some way, which they\\n> certainly might be. But that doesn't seem like a great reason, either.\\n\\nI agree the status quo is weird too. I'd be OK to break \\nbackwards-compatibility if we can make it better.\\n\\nTom mentioned enable_bitmapscan, and it reminded me that the current \\nbehavior with that is actually a bit annoying. I go through this pattern \\nvery often when I'm investigating query plans:\\n\\n1. Hmm, let's see what this query plan looks like:\\n\\npostgres=# explain analyze select * from foo where i=10;\\n QUERY PLAN \\n\\n----------------------------------------------------------------------------------------------------------------\\n Index Scan using foo_i_idx on foo (cost=0.29..8.31 rows=1 width=36) \\n(actual time=0.079..0.090 rows=2 loops=1)\\n Index Cond: (i = 10)\\n Planning Time: 2.220 ms\\n Execution Time: 0.337 ms\\n(4 rows)\\n\\n2. Ok, and how long would it take with a seq scan? Let's see:\\n\\npostgres=# set enable_indexscan=off;\\nSET\\npostgres=# explain analyze select * from foo where i=10;\\n QUERY PLAN \\n\\n------------------------------------------------------------------------------------------------------------------\\n Bitmap Heap Scan on foo (cost=4.30..8.31 rows=1 width=36) (actual \\ntime=0.102..0.113 rows=2 loops=1)\\n Recheck Cond: (i = 10)\\n Heap Blocks: exact=2\\n -> Bitmap Index Scan on foo_i_idx (cost=0.00..4.30 rows=1 width=0) \\n(actual time=0.067..0.068 rows=2 loops=1)\\n Index Cond: (i = 10)\\n Planning Time: 0.211 ms\\n Execution Time: 0.215 ms\\n(7 rows)\\n\\n3. Oh right, bitmap scan, I forgot about that one. Let's disable that too:\\n\\npostgres=# set enable_bitmapscan=off;\\nSET\\npostgres=# explain analyze select * from foo where i=10;\\n QUERY PLAN \\n\\n--------------------------------------------------------------------------------------------------\\n Seq Scan on foo (cost=0.00..1862.00 rows=1 width=36) (actual \\ntime=0.042..39.226 rows=2 loops=1)\\n Filter: (i = 10)\\n Rows Removed by Filter: 109998\\n Planning Time: 0.118 ms\\n Execution Time: 39.272 ms\\n(5 rows)\\n\\nI would be somewhat annoyed if we add another step to that, to also \\ndisable index-only scans separately. It would be nice if \\nenable_indexscan=off would also disable bitmap scans, that would \\neliminate one step from the above. Almost always when I want to disable \\nindex scans, I really want to disable the use of the index altogether. \\nThe problem then of course is, how do you force a bitmap scan without \\nallowing other index scans, when you want to test them both?\\n\\nIt almost feels like we should have yet another GUC to disable index \\nscans, index-only scans and bitmap index scans. \\\"enable_indexes=off\\\" or \\nsomething.\\n\\n-- \\nHeikki Linnakangas\\nNeon (https://neon.tech)\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 3 Jul 2024 00:39:43 +0300\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Heikki Linnakangas <hlinnaka@iki.fi> writes:\\n> 3. Oh right, bitmap scan, I forgot about that one. Let's disable that too:\\n\\nYeah, I've hit that too, although more often (for me) it's the first\\nchoice of plan. In any case, it usually takes more than one change\\nto get to a seqscan.\\n\\n> It almost feels like we should have yet another GUC to disable index \\n> scans, index-only scans and bitmap index scans. \\\"enable_indexes=off\\\" or \\n> something.\\n\\nThere's something to be said for that idea. Breaking compatibility is\\na little easier to stomach if there's a clear convenience win, and\\nthis'd offer that.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 02 Jul 2024 17:49:09 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 2, 2024 at 5:39 PM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n> I would be somewhat annoyed if we add another step to that, to also\\n> disable index-only scans separately. It would be nice if\\n> enable_indexscan=off would also disable bitmap scans, that would\\n> eliminate one step from the above. Almost always when I want to disable\\n> index scans, I really want to disable the use of the index altogether.\\n> The problem then of course is, how do you force a bitmap scan without\\n> allowing other index scans, when you want to test them both?\\n>\\n> It almost feels like we should have yet another GUC to disable index\\n> scans, index-only scans and bitmap index scans. \\\"enable_indexes=off\\\" or\\n> something.\\n\\nThis is an interesting idea, and it seems like it could be convenient.\\nHowever, the fact that it's so non-orthogonal is definitely not great.\\nOne problem I've had with going through regression tests that rely on\\nthe enable_* GUCs is that it's often not quite clear what values all\\nof those GUCs have at a certain point in the test file, because the\\nstatements that set them may be quite a bit higher up in the file and\\nsome changes may also have been rolled back. I've found recently that\\nthe addition of EXPLAIN (SETTINGS) helps with this quite a bit,\\nbecause you can adjust the .sql file to use that option and then see\\nwhat shows up in the output file. Still, it's annoying, and the same\\nissue could occur in any other situation where you're using these\\nGUCs. It's just more confusing when there are multiple ways of turning\\nsomething off.\\n\\nWould we consider merging enable_indexscan, enable_indexonlyscan, and\\nenable_bitmapscan into something like:\\n\\nenable_indexes = on | off | { plain | indexonly | bitmap } [, ...]\\n\\nI feel like that would solve the usability concern that you raise here\\nwhile also (1) preserving orthogonality and (2) reducing the number of\\nGUCs rather than first increasing it. When I first joined the project\\nthere were a decent number of enable_* GUCs, but there's way more now.\\nSome of them are a little random (which is a conversation for another\\nday) but just cutting down on the number seems like it might not be\\nsuch a bad idea.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 3 Jul 2024 09:01:16 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, 3 Jul 2024 at 09:49, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Heikki Linnakangas <hlinnaka@iki.fi> writes:\\n> > 3. Oh right, bitmap scan, I forgot about that one. Let's disable that too:\\n>\\n> Yeah, I've hit that too, although more often (for me) it's the first\\n> choice of plan. In any case, it usually takes more than one change\\n> to get to a seqscan.\\n\\nI commonly hit this too.\\n\\nI think the current behaviour is born out of the fact that we don't\\nproduce both an Index Scan and an Index Only Scan for the same index.\\nWe'll just make the IndexPath an index only scan, if possible based\\non:\\n\\nindex_only_scan = (scantype != ST_BITMAPSCAN &&\\n check_index_only(rel, index));\\n\\nThe same isn't true for Bitmap Index Scans. We'll create both\\nIndexPaths and BitmapHeapPaths and let them battle it out in\\nadd_path().\\n\\nI suspect this is why it's been coded that enable_indexscan also\\ndisables Index Only Scans. Now, of course, it could work another way,\\nbut I also still think that doing so is changing well-established\\nbehaviour that I don't recall anyone ever complaining about besides\\nRobert. Robert's complaint seems to have originated from something he\\nnoticed while hacking on code rather than actually using the database\\nfor something. I think the argument for changing it should have less\\nweight due to that.\\n\\nI understand that we do have inconsistencies around this stuff. For\\nexample, enable_sort has no influence on Incremental Sorts like\\nenable_indexscan has over Index Only Scan. That might come from the\\nfact that we used to, up until a couple of releases ago, produce both\\nsort path types and let them compete in add_path(). That's no longer\\nthe case, we now just do incremental sort when we can, just like we do\\nIndex Only Scans when we can. Despite those inconsistencies, I\\nwouldn't vote for changing either of them to align with the other. It\\njust feels too long-established behaviour to be messing with.\\n\\nI feel it might be best to move this patch to the back of the series\\nor just drop it for now as it seems to be holding up the other stuff\\nfrom moving forward, and that stuff looks useful and worth changing.\\n\\nDavid\\n\\n\\n\",\n \"msg_date\": \"Thu, 4 Jul 2024 13:29:39 +1200\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"OK, here's a new patch version. I earlier committed the refactoring to\\navoid using disable_cost to force WHERE CURRENT OF to be implemented\\nby a TID scan. In this version, I've dropped everything related to\\nreworking enable_indexscan or any other enable_* GUC. Hence, this\\nversion of the patch set just focuses on adding the count of disabled\\nnodes and removing the use of disable_cost. In addition to dropping\\nthe controversial patches, I've also found and squashed a few bugs in\\nthis version.\\n\\nBehavior: With the patch, whenever an enable_* GUC would cause\\ndisable_cost to be added, disabled_nodes is incremented instead. There\\nis one remaining use of disable_cost which is not triggered by an\\nenable_* GUC but by the desire to avoid plans that we think will\\noverflow work_mem. I welcome thoughts on what to do about that case;\\nfor now, I do nothing. As before, 0001 adds the disabled_nodes field\\nto paths and 0002 adds it to plans. I think we could plausibly commit\\nonly 0001, both patches separately, or both patches squashed.\\n\\nNotes:\\n\\n- I favor committing both patches. Tom stated that he didn't think\\nthat we needed to show anything related to disabled nodes, and that\\ncould be true. However, today, you can tell which nodes are disabled\\nas long as you print out the costs; if we don't propagate disabled\\nnodes into the plan and print them out, that will no longer be\\npossible. I found working on the patches that it was really hard to\\ndebug the patch set without this, so my guess is that we'll find not\\nhaving it pretty annoying, but we can also just commit 0001 for\\nstarters and see how long it takes for the lack of 0002 to become\\nannoying. If the answer is \\\"infinite time,\\\" that's cool; if it isn't,\\nwe can reconsider committing 0002.\\n\\n- If we do commit 0002, I think it's a good idea to have the number of\\ndisabled nodes displayed even with COSTS OFF, because it's stable, and\\nit's pretty useful to be able to see this in the regression output. I\\nhave found while working on this that I often need to adjust the .sql\\nfiles to say EXPLAIN (COSTS ON) instead of EXPLAIN (COSTS OFF) in\\norder to understand what's happening. Right now, there's no real\\nalternative because costs aren't stable, but disabled-node counts\\nshould be stable, so I feel this would be a step forward. Apart from\\nthat, I also think it's good for features to have regression test\\ncoverage, and since we use COSTS OFF everywhere or at least nearly\\neverywhere in the regression test, if we don't print out the disabled\\nnode counts when COSTS OFF is used, then we don't cover that case in\\nour tests. Bummer.\\n\\nRegression test changes in 0001:\\n\\n- btree_index.sql executes a query \\\"select proname from pg_proc where\\nproname ilike 'ri%foo' order by 1\\\" with everything but bitmap scans\\ndisabled. Currently, that produces an index-only scan; with the patch,\\nit produces a sort over a sequential scan. That's a little odd,\\nbecause the test seems to be aimed at demonstrating that we can use a\\nbitmap scan, and it doesn't, because we apparently can't. But, why\\ndoes the patch change the plan?\\nAt least on my machine, the index-only scan is significantly more\\ncostly than the sequential scan. I think what's happening here is that\\nwhen you add disable_cost to the cost of both paths, they compare\\nfuzzily the same; without that, the cheaper one wins.\\n\\n- select_parallel.out executes a query with sequential scans disabled\\nbut tenk2 must nevertheless be sequential-scanned. With the patch,\\nthat changes to a parallel sequential scan. I think the explanation\\nhere is the same as in the preceding case.\\n\\n- horizons.spec currently sets enable_seqscan=false,\\nenable_indexscan=false, and enable_bitmapscan=false. I suspect that\\nAndres thought that this would force the use of an index-only scan,\\nsince nothing sets enable_indexonlyscan=false. But as discussed\\nupthread, that is not true. Instead everything is disabled. For the\\nsame reasons as in the previous two examples, this caused an\\nassortment of plan changes which in turn caused the test to fail to\\ntest what it was intended to test. So I removed enable_indexscan=false\\nfrom the spec file, and now it gets index-only scans everywhere again,\\nas desired.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Wed, 31 Jul 2024 12:23:22 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Thu, 1 Aug 2024 at 04:23, Robert Haas <robertmhaas@gmail.com> wrote:\\n> OK, here's a new patch version.\\n\\nI think we're going down the right path here.\\n\\nI've reviewed both patches, here's what I noted down during my review:\\n\\n0. I've not seen any mention so far about postgres_fdw's\\nuse_remote_estimate. Maybe changing the costs is fixing an issue that\\nexisted before. I'm just not 100% sure on that.\\n\\nConsider:\\n\\nCREATE EXTENSION postgres_fdw;\\n\\nDO $d$\\n BEGIN\\n EXECUTE $$CREATE SERVER loopback FOREIGN DATA WRAPPER postgres_fdw\\n OPTIONS (use_remote_estimate 'true',\\n dbname '$$||current_database()||$$',\\n port '$$||current_setting('port')||$$'\\n )$$;\\n END;\\n$d$;\\nCREATE USER MAPPING FOR CURRENT_USER SERVER loopback;\\n\\ncreate table t (a int);\\ncreate foreign table ft (a int) server loopback OPTIONS (table_name 't');\\n\\nalter system set enable_seqscan=0;\\nselect pg_Reload_conf();\\nset enable_seqscan=1;\\nexplain select * from ft;\\n\\npatched:\\n Foreign Scan on ft (cost=100.00..671.00 rows=2550 width=4)\\n\\nmaster:\\n Foreign Scan on ft (cost=10000000100.00..10000000671.00 rows=2550 width=4)\\n\\nI kinda think that might be fixing an issue that I don't recall being\\nreported before. I think we shouldn't really care that much about what\\nnodes are disabled on the remote server and not having disabled_cost\\napplied to that gives us that.\\n\\n1. The final sentence of the function header comment needs to be\\nupdated in estimate_path_cost_size().\\n\\n2. Does cost_tidscan() need to update the header comment to say\\ntidquals must not be empty?\\n\\n3. final_cost_nestloop() seems to initially use the disabled_nodes\\nfrom initial_cost_nestloop() but then it goes off and calculates it\\nagain itself. One of these seems redundant. The \\\"We could include\\ndisable_cost in the preliminary estimate\\\" comment explains why it was\\noriginally left to final_cost_nestloop(), so maybe worth sticking to\\nthat? I don't quite know the full implications, but it does not seem\\nworth risking a behaviour change here.\\n\\n4. I wonder if it's worth doing a quick refactor of the code in\\ninitial_cost_mergejoin() to get rid of the duplicate code in the \\\"if\\n(outersortkeys)\\\" and \\\"if (innersortkeys)\\\" branches. It seems ok to do\\nouter_path = &sort_path. Likewise for inner_path.\\n\\n5. final_cost_hashjoin() does the same thing as #3\\n\\n6. createplan.c adds #include \\\"nodes/print.h\\\" but doesn't seem to add\\nany code that might use anything in there.\\n\\n7. create_lockrows_path() needs to propagate disabled_nodes.\\n\\ncreate table a (a int);\\nset enable_seqscan=0;\\n\\nexplain select * from a for update limit 1;\\n\\n Limit (cost=0.00..0.02 rows=1 width=10)\\n -> LockRows (cost=0.00..61.00 rows=2550 width=10)\\n -> Seq Scan on a (cost=0.00..35.50 rows=2550 width=10)\\n Disabled Nodes: 1\\n(4 rows)\\n\\n\\nexplain select * from a limit 1;\\n\\n Limit (cost=0.00..0.01 rows=1 width=4)\\n Disabled Nodes: 1\\n -> Seq Scan on a (cost=0.00..35.50 rows=2550 width=4)\\n Disabled Nodes: 1\\n(4 rows)\\n\\n8. There's something weird with CTEs too.\\n\\ncreate table b(a int);\\nset enable_sort=0;\\n\\nPatched:\\n\\nexplain with cte as materialized (select * from b order by a) select *\\nfrom cte order by a desc;\\n\\n Sort (cost=381.44..387.82 rows=2550 width=4)\\n Disabled Nodes: 1\\n Sort Key: cte.a DESC\\n CTE cte\\n -> Sort (cost=179.78..186.16 rows=2550 width=4)\\n Disabled Nodes: 1\\n Sort Key: b.a\\n -> Seq Scan on b (cost=0.00..35.50 rows=2550 width=4)\\n -> CTE Scan on cte (cost=0.00..51.00 rows=2550 width=4)\\n(9 rows)\\n\\nmaster:\\n\\nexplain with cte as materialized (select * from a order by a) select *\\nfrom cte order by a desc;\\n\\n Sort (cost=20000000381.44..20000000387.82 rows=2550 width=4)\\n Sort Key: cte.a DESC\\n CTE cte\\n -> Sort (cost=10000000179.78..10000000186.16 rows=2550 width=4)\\n Sort Key: a.a\\n -> Seq Scan on a (cost=0.00..35.50 rows=2550 width=4)\\n -> CTE Scan on cte (cost=0.00..51.00 rows=2550 width=4)\\n(7 rows)\\n\\nI'd expect the final sort to have disabled_nodes == 2 since\\ndisabled_cost has been added twice in master.\\n\\n9. create_set_projection_path() needs to propagate disabled_nodes too:\\n\\nexplain select b from (select a,generate_series(1,2) as b from b) a limit 1;\\n\\n Limit (cost=0.00..0.03 rows=1 width=4)\\n -> Subquery Scan on a (cost=0.00..131.12 rows=5100 width=4)\\n -> ProjectSet (cost=0.00..80.12 rows=5100 width=8)\\n -> Seq Scan on b (cost=0.00..35.50 rows=2550 width=0)\\n Disabled Nodes: 1\\n\\n10. create_setop_path() needs to propagate disabled_nodes.\\n\\nexplain select * from b except select * from b limit 1;\\n\\n Limit (cost=0.00..0.80 rows=1 width=8)\\n -> HashSetOp Except (cost=0.00..160.25 rows=200 width=8)\\n -> Append (cost=0.00..147.50 rows=5100 width=8)\\n Disabled Nodes: 2\\n -> Subquery Scan on \\\"*SELECT* 1\\\" (cost=0.00..61.00\\nrows=2550 width=8)\\n Disabled Nodes: 1\\n -> Seq Scan on b (cost=0.00..35.50 rows=2550 width=4)\\n Disabled Nodes: 1\\n -> Subquery Scan on \\\"*SELECT* 2\\\" (cost=0.00..61.00\\nrows=2550 width=8)\\n Disabled Nodes: 1\\n -> Seq Scan on b b_1 (cost=0.00..35.50 rows=2550 width=4)\\n Disabled Nodes: 1\\n(12 rows)\\n\\n11. create_modifytable_path() needs to propagate disabled_nodes.\\n\\nexplain with cte as (update b set a = a+1 returning *) select * from\\ncte limit 1;\\n\\n Limit (cost=41.88..41.90 rows=1 width=4)\\n CTE cte\\n -> Update on b (cost=0.00..41.88 rows=2550 width=10)\\n -> Seq Scan on b (cost=0.00..41.88 rows=2550 width=10)\\n Disabled Nodes: 1\\n -> CTE Scan on cte (cost=0.00..51.00 rows=2550 width=4)\\n(6 rows)\\n\\n12. For the 0002 patch, I do agree that having this visible in EXPLAIN\\nis a must. I'd much rather see: Disabled: true/false. And just\\ndisplay this when the disabled_nodes is greater than the sum of the\\nsubpaths. That might be much more complex to implement, but it's\\ngoing to make it much easier to track down the disabled nodes in very\\nlarge plans.\\n\\nDavid\\n\\n\\n\",\n \"msg_date\": \"Thu, 1 Aug 2024 14:01:07 +1200\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, Jul 31, 2024 at 10:01 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n> I've reviewed both patches, here's what I noted down during my review:\\n\\nThanks.\\n\\n> 0. I've not seen any mention so far about postgres_fdw's\\n> use_remote_estimate. Maybe changing the costs is fixing an issue that\\n> existed before. I'm just not 100% sure on that.\\n>\\n> patched:\\n> Foreign Scan on ft (cost=100.00..671.00 rows=2550 width=4)\\n>\\n> master:\\n> Foreign Scan on ft (cost=10000000100.00..10000000671.00 rows=2550 width=4)\\n>\\n> I kinda think that might be fixing an issue that I don't recall being\\n> reported before. I think we shouldn't really care that much about what\\n> nodes are disabled on the remote server and not having disabled_cost\\n> applied to that gives us that.\\n\\nHmm, I think it's subjective which behavior is better. If somebody\\nthought the new behavior was worse, they might want the remote side's\\ncount of disabled nodes to be propagated to the local side, but I'm\\ndisinclined to go there. My guess is that it doesn't matter much\\neither way what we do here, so I'd rather not add more code.\\n\\n> 1. The final sentence of the function header comment needs to be\\n> updated in estimate_path_cost_size().\\n\\nFixed.\\n\\n> 2. Does cost_tidscan() need to update the header comment to say\\n> tidquals must not be empty?\\n\\nIMHO, no. The assertions I added to that function were intended as\\ndocumentation of what that function was already assuming about the\\nbehavior of its caller. I had to trace through the logic in tidpath.c\\nfor quite a while to understand why cost_tidscan() was not completely\\nbroken. To spare the next person the trouble of working that out, I\\nadded assertions. Now we could additionally add commentary in English\\nthat restates what the assertions already say, but I feel like having\\nthe assertions is good enough. If somebody ever whacks around\\ntidpath.c such that these assertions start failing, I think it will be\\nfairly clear to them that they either need to revert their changes in\\ntidpath.c or upgrade the logic in this function to cope.\\n\\n> 3. final_cost_nestloop() seems to initially use the disabled_nodes\\n> from initial_cost_nestloop() but then it goes off and calculates it\\n> again itself. One of these seems redundant.\\n\\nOops. Fixed.\\n\\n> The \\\"We could include\\n> disable_cost in the preliminary estimate\\\" comment explains why it was\\n> originally left to final_cost_nestloop(), so maybe worth sticking to\\n> that? I don't quite know the full implications, but it does not seem\\n> worth risking a behaviour change here.\\n\\nI don't really see how there could be a behavior change here, unless\\nthere's a bug. Dealing with the enable_* flags in initial_cost_XXX\\nrather than final_cost_XXX could be better or worse from a performance\\nstandpoint and it could make for cleaner or less clean code, but the\\nuser-facing behavior should be identical unless there are bugs.\\n\\nThe reason why I changed this is because of the logic in\\nadd_path_precheck(): it exits early as soon as it sees a path whose\\ntotal cost is greater than the cost of the proposed new path. Since\\nthe patch's aim is to treat disabled_node as a high-order component of\\nthe cost, we need to make the same decision by comparing the count of\\ndisabled_nodes first and then if that is equal, we need to compare the\\ntotal_cost. We can't do that if we don't have the count of\\ndisabled_nodes for the proposed new path.\\n\\nI think this may be a bit hard to understand, so let me give a\\nconcrete example. Suppose we're planning some join where one side can\\nonly be planned with a sequential scan and sequential scans are\\ndisabled. We have ten paths in the path list and they have costs of\\n1e10+100, 1e10+200, ..., 1e10+1000. Now add_path_precheck() is asked\\nto consider a new path where there is a disabled node on BOTH sides of\\nthe join -- the one side has the disabled sequential scan, but now the\\nother side also has something disabled, so the cost is let's say\\n2e10+79. add_path_precheck() can see at once that this path is a\\nloser: it can't possibly dominate any path that already exists,\\nbecause it costs more than any of them. But when you take disable_cost\\nout, things look quite different. Now you have a proposed path with a\\ntotal_cost of 79 and a path list with costs of 100, ..., 1000. If\\nyou're not allowed to know anything about disabled_nodes, the new path\\nlooks like it might be valuable. You might decide to construct it and\\ntry inserting into the pathlist, which will end up being useless, and\\neven if you don't, you're going to compare its pathkeys and\\nparameterization to each of the 10 existing paths before giving up.\\nBummer.\\n\\nSo, to avoid getting much stupider than it is currently,\\nadd_path_precheck() needs a preliminary estimate of the number of\\ndisabled nodes just like it needs a preliminary estimate of the total\\ncost. And to avoid regressions, that estimate needs to be pretty good.\\nA naive estimate would be to just add up the number of disabled_nodes\\non the inner and outer paths, but that would be a regression in the\\nmerge-join case, because initial_cost_mergejoin() calls cost_sort()\\nfor the inner and outer sides and that will add disable_cost if sorts\\nare disabled. If you didn't take the effect of cost_sort() into\\naccount, you might think that your number of disabled_nodes was going\\nto be substantially lower than it really would be, leading to wasted\\nwork as described in the last paragraph. Plus, since\\ninitial_cost_mergejoin() is incurring the overhead of calling\\ncost_sort() anyway to get the total cost numbers anyway, it would be\\nsilly not to save the count of disabled nodes: if we did, we'd have to\\nredo the cost_sort() call in final_cost_mergejoin(), which would be\\nexpensive.\\n\\nIf we wanted to make our estimate of the # of disabled nodes exactly\\ncomparable to what we now do with disable_cost, we would postpone if\\n(!enable_WHATEVERjoin) ++disabled_nodes to the final_cost_XXX\\nfunctions and do all of the other accounting related to disabled nodes\\nat the initial_cost_XXX phase. But I do not like that approach.\\nPostponing one trivial portion of the disabled_nodes calculation to a\\nlater time won't save any significant number of CPU cycles, but it\\nmight confuse people reading the code. You then have to know that the\\ndisabled_nodes count that gets passed to final_cost_XXX is not yet the\\nfinal count, but that you may still need to add 1 for the join itself\\n(but not for the implicit sorts that the join requires, which have\\nalready been accounted for). That's the kind of odd definition that\\nbreeds bugs. Besides, it's not as if moving that tiny bit of logic to\\nthe initial_cost_XXX functions has no upside: it could allow\\nadd_path_precheck() to exit earlier, thus saving cycles.\\n\\n(For the record, the explanation above took about 3 hours to write, so\\nI hope it's managed to be both correct and convincing. This stuff is\\nreally complicated.)\\n\\n> 4. I wonder if it's worth doing a quick refactor of the code in\\n> initial_cost_mergejoin() to get rid of the duplicate code in the \\\"if\\n> (outersortkeys)\\\" and \\\"if (innersortkeys)\\\" branches. It seems ok to do\\n> outer_path = &sort_path. Likewise for inner_path.\\n\\nI don't think that's better.\\n\\n> 5. final_cost_hashjoin() does the same thing as #3\\n\\nArgh. Fixed.\\n\\n> 6. createplan.c adds #include \\\"nodes/print.h\\\" but doesn't seem to add\\n> any code that might use anything in there.\\n\\nFixed.\\n\\n> 8. There's something weird with CTEs too.\\n>\\n> I'd expect the final sort to have disabled_nodes == 2 since\\n> disabled_cost has been added twice in master.\\n\\nRight now, disabled node counts don't propagate through SubPlans (see\\nSS_process_ctes). Maybe that needs to be changed, but aside from\\nlooking weird, does it do any harm?\\n\\n> 7. create_lockrows_path() needs to propagate disabled_nodes.\\n> 9. create_set_projection_path() needs to propagate disabled_nodes too:\\n> 10. create_setop_path() needs to propagate disabled_nodes.\\n> 11. create_modifytable_path() needs to propagate disabled_nodes.\\n\\nI changed all of these, but I think these examples only establish that\\nthose nodes DO NOT propagate disabled_nodes, not that they need to. If\\nwe're past the point of making any choices based on costs, then\\nmaintaining disabled_nodes or not doing so won't affect correctness.\\nThat's not to say these aren't good to tidy up, and some of them may\\nwell be bugs, but I don't think your test cases prove that. What\\nprimarily matters is whether the enable_BLAH GUCs get respected; the\\nexact contents of the EXPLAIN output are somewhat more arguable.\\n\\n> 12. For the 0002 patch, I do agree that having this visible in EXPLAIN\\n> is a must. I'd much rather see: Disabled: true/false. And just\\n> display this when the disabled_nodes is greater than the sum of the\\n> subpaths. That might be much more complex to implement, but it's\\n> going to make it much easier to track down the disabled nodes in very\\n> large plans.\\n\\nI think it's going to be very unpleasant if we have the planner add\\nthings up and then try to have EXPLAIN subtract them back out again.\\nOne problem with that is that all of the test cases where you just\\nshowed disabled_nodes not propagating upward wouldn't actually show\\nanything any more, because disabled_nodes would not have been greater\\nin the parent than in the child. So those are oversights in the code\\nthat are easy to spot now but would become hard to spot with this\\nimplementation. Another problem is that the EXPLAIN code itself could\\ncontain bugs, or slightly more broadly, get out of sync with the logic\\nthat decides what to add up. It won't be obvious what's happening:\\nsome node that is actually disabled just won't appear to be, or the\\nother way around, and it will be hard to understand what happened,\\nbecause you won't be able to see the raw counts of disabled nodes that\\nwould allow you to deduce where the error actually is.\\n\\nOne idea that occurs to me is to store TWO counts in each path node\\nand each plan node: the count of self-exclusive disabled nodes, and\\nthe count of self-include disabled nodes. Then explain can just test\\nif they are different. If the answer is 1, the node is disabled; if 0,\\nit's enabled; if anything else, there's a bug (and it could print the\\ndelta, or each value separately, to help localize such bugs). The\\nproblem with that is that it eats up more space in\\nperformance-critical data structures, but perhaps that's OK: I don't\\nknow.\\n\\nAnother thought is that right now you just see the disable_cost values\\nadded up with the rest of the cost. So maybe propagating upward is not\\nreally such a bad behavior; it's what we have now.\\n\\nThis point probably needs more thought and discussion, but I'm out of\\ntime to work on this for today, and out of mental energy too. So for\\nnow here's v5 as I have it.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Thu, 1 Aug 2024 14:03:18 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, 2 Aug 2024 at 06:03, Robert Haas <robertmhaas@gmail.com> wrote:\\n> I think this may be a bit hard to understand, so let me give a\\n> concrete example. Suppose we're planning some join where one side can\\n> only be planned with a sequential scan and sequential scans are\\n> disabled. We have ten paths in the path list and they have costs of\\n> 1e10+100, 1e10+200, ..., 1e10+1000. Now add_path_precheck() is asked\\n> to consider a new path where there is a disabled node on BOTH sides of\\n> the join -- the one side has the disabled sequential scan, but now the\\n> other side also has something disabled, so the cost is let's say\\n> 2e10+79. add_path_precheck() can see at once that this path is a\\n> loser: it can't possibly dominate any path that already exists,\\n> because it costs more than any of them. But when you take disable_cost\\n> out, things look quite different. Now you have a proposed path with a\\n> total_cost of 79 and a path list with costs of 100, ..., 1000. If\\n> you're not allowed to know anything about disabled_nodes, the new path\\n> looks like it might be valuable. You might decide to construct it and\\n> try inserting into the pathlist, which will end up being useless, and\\n> even if you don't, you're going to compare its pathkeys and\\n> parameterization to each of the 10 existing paths before giving up.\\n> Bummer.\\n\\nOK, so it sounds like you'd like to optimise this code so that the\\nplanner does a little less work when node types are disabled. The\\nexisting comment does mention explicitly that we don't want to do\\nthat:\\n\\n/*\\n* We could include disable_cost in the preliminary estimate, but that\\n* would amount to optimizing for the case where the join method is\\n* disabled, which doesn't seem like the way to bet.\\n*/\\n\\nAs far as I understand it from reading the comments in that file, I\\nsee no offer of guarantees that the initial cost will be cheaper than\\nthe final cost. So what you're proposing could end up rejecting paths\\nbased on initial cost where the final cost might end up being the\\ncheapest path. Imagine you're considering a Nested Loop and a Hash\\nJoin, both of which are disabled. Merge Join is unavailable as the\\njoin column types are not sortable. If the hash join costs 99 and the\\ninitial nested loop costs 110, but the final nested loop ends up\\ncosting 90, then the nested loop could be rejected before we even get\\nto perform the final cost for it. The current code will run\\nfinal_cost_nestloop() and find that 90 is cheaper than 99, whereas\\nwhat you want to do is stop bothering with nested loop when we see the\\ninitial cost come out at 110.\\n\\nPerhaps it's actually fine if the initial costs are always less than\\nthe final costs as, if that's the case, we won't ever reject any paths\\nbased on the initial cost that we wouldn't anyway based on the final\\ncost. However, since there does not seem to be any comments mentioning\\nthis guarantee and if you're just doing this to squeeze more\\nperformance out of the planner, it seems risky to do for that reason\\nalone.\\n\\nI'd say if you want to do this, you should be justifying it on its own\\nmerit with some performance numbers and some evidence that we don't\\nproduce inferior plans as a result. But per what I quoted above,\\nyou're not doing that, you're doing this as a performance\\noptimisation.\\n\\nI'm not planning on pushing this any further. I've just tried to\\nhighlight that there's the possibility of a behavioural change. You're\\nclaiming there isn't one. I claim there is.\\n\\nDavid\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Aug 2024 15:34:16 +1200\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 1, 2024 at 11:34 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n> I'm not planning on pushing this any further. I've just tried to\\n> highlight that there's the possibility of a behavioural change. You're\\n> claiming there isn't one. I claim there is.\\n\\nI don't know what to tell you. The original version of the patch\\ndidn't change this stuff, and the result did not work. So I looked\\ninto the problem and fixed it. I may have done that wrongly, or there\\nmay be debatable points, but it seems like your argument is\\nessentially that I shouldn't have done any of this and I should just\\ntake it all back out, and I know that doesn't work because it's the\\nfirst thing I tried.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Aug 2024 08:17:16 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Sat, 3 Aug 2024 at 00:17, Robert Haas <robertmhaas@gmail.com> wrote:\\n>\\n> On Thu, Aug 1, 2024 at 11:34 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n> > I'm not planning on pushing this any further. I've just tried to\\n> > highlight that there's the possibility of a behavioural change. You're\\n> > claiming there isn't one. I claim there is.\\n>\\n> I don't know what to tell you. The original version of the patch\\n> didn't change this stuff, and the result did not work. So I looked\\n> into the problem and fixed it. I may have done that wrongly, or there\\n> may be debatable points, but it seems like your argument is\\n> essentially that I shouldn't have done any of this and I should just\\n> take it all back out, and I know that doesn't work because it's the\\n> first thing I tried.\\n\\nI've just read what you wrote again and I now realise something I didn't before.\\n\\nI now think neither of us got it right. I now think what you'd need to\\ndo to be aligned to the current behaviour is have\\ninitial_cost_nestloop() add the disabled_nodes for the join's subnodes\\n*only* and have final_cost_nestloop() add the additional\\ndisabled_nodes if enable_nestloop = off. That way you maintain the\\nexisting behaviour of not optimising for disabled node types and don't\\nrisk plan changes if the final cost comes out cheaper than the initial\\ncost.\\n\\nDavid\\n\\n\\n\",\n \"msg_date\": \"Sat, 3 Aug 2024 01:13:36 +1200\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 2, 2024 at 9:13 AM David Rowley <dgrowleyml@gmail.com> wrote:\\n> I now think neither of us got it right. I now think what you'd need to\\n> do to be aligned to the current behaviour is have\\n> initial_cost_nestloop() add the disabled_nodes for the join's subnodes\\n> *only* and have final_cost_nestloop() add the additional\\n> disabled_nodes if enable_nestloop = off. That way you maintain the\\n> existing behaviour of not optimising for disabled node types and don't\\n> risk plan changes if the final cost comes out cheaper than the initial\\n> cost.\\n\\nAll three initial_cost_XXX functions have a comment that says \\\"This\\nmust quickly produce lower-bound estimates of the path's startup and\\ntotal costs,\\\" i.e. the final cost should never be cheaper. I'm pretty\\nsure that it was the design intention here that no path ever gets\\nrejected at the initial cost stage that would have been accepted at\\nthe final cost stage.\\n\\n(You can also see, as a matter of implementation, that they extract\\nthe startup_cost and run_cost from the workspace and then add to those\\nvalues.)\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Aug 2024 12:04:24 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Fri, Aug 2, 2024 at 9:13 AM David Rowley <dgrowleyml@gmail.com> wrote:\\n>> ... That way you maintain the\\n>> existing behaviour of not optimising for disabled node types and don't\\n>> risk plan changes if the final cost comes out cheaper than the initial\\n>> cost.\\n\\n> All three initial_cost_XXX functions have a comment that says \\\"This\\n> must quickly produce lower-bound estimates of the path's startup and\\n> total costs,\\\" i.e. the final cost should never be cheaper. I'm pretty\\n> sure that it was the design intention here that no path ever gets\\n> rejected at the initial cost stage that would have been accepted at\\n> the final cost stage.\\n\\nThat absolutely is the expectation, and we'd better be careful not\\nto break it.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 02 Aug 2024 12:51:12 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 2, 2024 at 12:51 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> That absolutely is the expectation, and we'd better be careful not\\n> to break it.\\n\\nI have every intention of not breaking it. :-)\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Aug 2024 12:53:16 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 2, 2024 at 12:53 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Fri, Aug 2, 2024 at 12:51 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > That absolutely is the expectation, and we'd better be careful not\\n> > to break it.\\n>\\n> I have every intention of not breaking it. :-)\\n\\nI went ahead and committed these patches. I know there's some debate\\nover whether we want to show the # of disabled nodes and if so whether\\nit should be controlled by COSTS, and I suspect I haven't completely\\nallayed David's concerns about the initial_cost_XXX functions although\\nI think that I did the right thing. But, I don't have the impression\\nthat anyone is desperately opposed to the basic concept, so I think it\\nmakes sense to put these into the tree and see what happens. We have\\nquite a bit of time left in this release cycle to uncover bugs, hear\\nfrom users or other developers, etc. about what problems there may be\\nwith this. If we end up deciding to reverse course or need to fix a\\nbunch of stuff, so be it, but let's see what the feedback is.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 21 Aug 2024 10:29:23 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, 31 Jul 2024 at 18:23, Robert Haas <robertmhaas@gmail.com> wrote:\\n> - If we do commit 0002, I think it's a good idea to have the number of\\n> disabled nodes displayed even with COSTS OFF, because it's stable, and\\n> it's pretty useful to be able to see this in the regression output. I\\n> have found while working on this that I often need to adjust the .sql\\n> files to say EXPLAIN (COSTS ON) instead of EXPLAIN (COSTS OFF) in\\n> order to understand what's happening. Right now, there's no real\\n> alternative because costs aren't stable, but disabled-node counts\\n> should be stable, so I feel this would be a step forward. Apart from\\n> that, I also think it's good for features to have regression test\\n> coverage, and since we use COSTS OFF everywhere or at least nearly\\n> everywhere in the regression test, if we don't print out the disabled\\n> node counts when COSTS OFF is used, then we don't cover that case in\\n> our tests. Bummer.\\n\\nAre the disabled node counts still expected to be stable even with\\nGEQO? If not, maybe we should have a way to turn them off after all.\\nAlthough I agree that always disabling them when COSTS OFF is set is\\nprobably also undesirable. How about a new option, e.g. EXPLAIN\\n(DISABLED OFF)\\n\\n\\n\",\n \"msg_date\": \"Thu, 22 Aug 2024 14:07:32 +0200\",\n \"msg_from\": \"Jelte Fennema-Nio <postgres@jeltef.nl>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 22, 2024 at 8:07 AM Jelte Fennema-Nio <postgres@jeltef.nl> wrote:\\n> Are the disabled node counts still expected to be stable even with\\n> GEQO? If not, maybe we should have a way to turn them off after all.\\n> Although I agree that always disabling them when COSTS OFF is set is\\n> probably also undesirable. How about a new option, e.g. EXPLAIN\\n> (DISABLED OFF)\\n\\nHmm, I hadn't thought about that. There are no GEQO-specific changes\\nin this patch, which AFAIK is OK, because I think GEQO just relies on\\nthe core planning machinery to decide everything about the cost of\\npaths, and is really only experimenting with different join orders. So\\nI think if it picks the same join order, it should get the same count\\nof disabled nodes everywhere. If it doesn't pick the same order,\\nyou'll get a different plan entirely.\\n\\nI don't think I quite want to jump into inventing a new EXPLAIN option\\nright this minute. I'm not against the idea, but I don't want to jump\\ninto engineering solutions before I understand what the problems are,\\nso I think we should give this a little time. I'll be a bit surprised\\nif this doesn't elicit a few strong reactions, but I want to see what\\npeople are actually sad (or, potentially, happy) about.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 22 Aug 2024 08:43:18 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On 8/21/24 10:29 AM, Robert Haas wrote:\\r\\n\\r\\n> I went ahead and committed these patches. I know there's some debate\\r\\n> over whether we want to show the # of disabled nodes and if so whether\\r\\n> it should be controlled by COSTS, and I suspect I haven't completely\\r\\n> allayed David's concerns about the initial_cost_XXX functions although\\r\\n> I think that I did the right thing. But, I don't have the impression\\r\\n> that anyone is desperately opposed to the basic concept, so I think it\\r\\n> makes sense to put these into the tree and see what happens. We have\\r\\n> quite a bit of time left in this release cycle to uncover bugs, hear\\r\\n> from users or other developers, etc. about what problems there may be\\r\\n> with this. If we end up deciding to reverse course or need to fix a\\r\\n> bunch of stuff, so be it, but let's see what the feedback is.\\r\\n\\r\\nWe hit an issue with pgvector[0] where a regular `SELECT count(*) FROM \\r\\ntable`[1] is attempting to scan the index on the vector column when \\r\\n`enable_seqscan` is disabled. Credit to Andrew Kane (CC'd) for flagging it.\\r\\n\\r\\nI was able to trace this back to e2225346. Here is a reproducer:\\r\\n\\r\\nSetup\\r\\n=====\\r\\n\\r\\nCREATE EXTENSION vector;\\r\\n\\r\\nCREATE OR REPLACE FUNCTION public.generate_random_normalized_vector(dim \\r\\ninteger)\\r\\nRETURNS vector\\r\\nLANGUAGE SQL\\r\\nAS $$\\r\\n SELECT public.l2_normalize(array_agg(random()::real)::vector)\\r\\n FROM generate_series(1, $1);\\r\\n$$;\\r\\n\\r\\nCREATE TABLE test (id int, embedding vector(128));\\r\\nINSERT INTO test\\r\\n SELECT n, public.generate_random_normalized_vector(128)\\r\\n FROM generate_series(1,5) n;\\r\\n\\r\\nCREATE INDEX ON test USING hnsw (embedding vector_cosine_ops);\\r\\n\\r\\nTest\\r\\n====\\r\\n\\r\\nSET enable_seqscan TO off;\\r\\nEXPLAIN ANALYZE\\r\\nSELECT count(*) FROM test;\\r\\n\\r\\nBefore e2225346:\\r\\n----------------\\r\\n\\r\\nAggregate (cost=10000041965.00..10000041965.01 rows=1 width=8) (actual \\r\\ntime=189.864..189.864 rows\\r\\n=1 loops=1)\\r\\n -> Seq Scan on test (cost=10000000000.00..10000040715.00 rows=5 \\r\\nwidth=0) (actual time=0.01\\r\\n8..168.294 rows=5 loops=1)\\r\\n(4 rows)\\r\\n\\r\\nWith e2225346:\\r\\n-------------\\r\\nERROR: cannot scan hnsw index without order\\r\\n\\r\\nSome things to note with the ivfflat/hnsw index AMs[3] in pgvector are \\r\\nthat they're used for \\\"ORDER BY\\\" scans exclusively. They currently don't \\r\\nsupport index only scans (noting as I tried reproducing the issue with \\r\\nGIST and couldn't do so because of that), but we wouldn't want to do a \\r\\nfull table \\\"count(*)\\\" on a IVFFlat/HNSW index anyway as it'd be more \\r\\nexpensive than just a full table scan.\\r\\n\\r\\nThanks,\\r\\n\\r\\nJonathan\\r\\n\\r\\n[0] https://github.com/pgvector/pgvector\\r\\n[1] https://github.com/pgvector/pgvector/actions/runs/10519052945\\r\\n[2] https://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=e2225346\\r\\n[3] https://github.com/pgvector/pgvector/blob/master/src/hnsw.c#L192\",\n \"msg_date\": \"Fri, 23 Aug 2024 11:16:57 -0400\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 23, 2024 at 11:17 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n> We hit an issue with pgvector[0] where a regular `SELECT count(*) FROM\\n> table`[1] is attempting to scan the index on the vector column when\\n> `enable_seqscan` is disabled. Credit to Andrew Kane (CC'd) for flagging it.\\n>\\n> I was able to trace this back to e2225346. Here is a reproducer:\\n\\nIf I change EXPLAIN ANALYZE in this test to just EXPLAIN, I get this:\\n\\n Aggregate (cost=179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00..179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00\\nrows=1 width=8)\\n -> Index Only Scan using test_embedding_idx on test\\n(cost=179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00..179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00\\nrows=5 width=0)\\n\\nIt took me a moment to wrap my head around this: the cost estimate is\\n312 decimal digits long. Apparently hnswcostestimate() just returns\\nDBL_MAX when there are no scan keys because it really, really doesn't\\nwant to do that. Before e2225346, that kept this plan from being\\ngenerated because it was (much) larger than disable_cost. But now it\\ndoesn't, because 1 disabled node makes a path more expensive than any\\npossible non-disabled path. Since that was the whole point of the\\npatch, I don't feel too bad about it.\\n\\nI find it a little weird that hnsw thinks itself able to return all\\nthe tuples in an order the user chooses, but unable to return all of\\nthe tuples in an arbitrary order. In core, we have precedent for index\\ntypes that can't return individual tuples at all (gin, brin) but not\\none that is able to return tuples in concept but has a panic attack if\\nyou don't know how you want them sorted. I don't quite see why you\\ncouldn't just treat that case the same as ORDER BY\\nthe_first_column_of_the_index, or any other arbitrary rule that you\\nwant to make up. Sure, it might be more expensive than a sequential\\nscan, but the user said they didn't want a sequential scan. I'm not\\nquite sure why pgvector thinks it gets to decide that it knows better\\nthan the user, or the rest of the optimizer. I don't even think I\\nreally believe it would always be worse: I've seen cases where a table\\nwas badly bloated and mostly empty but its indexes were not bloated,\\nand in that case an index scan can be a HUGE winner even though it\\nwould normally be a lot worse than a sequential scan.\\n\\nIf you don't want to fix hnsw to work the way the core optimizer\\nthinks it should, or if there's some reason it can't be done,\\nalternatives might include (1) having the cost estimate function hack\\nthe count of disabled nodes and (2) adding some kind of core support\\nfor an index cost estimator refusing a path entirely. I haven't tested\\n(1) so I don't know for sure that there are no issues, but I think we\\nhave to do all of our cost estimating before we can think about adding\\nthe path so I feel like there's a decent chance it would do what you\\nwant.\\n\\nAlso, while I did take the initiative to download pgvector and compile\\nit and hook up a debugger and figure out what was going on here, I'm\\nnot really too sure that's my job. I do think I have a responsibility\\nto help maintainers of out-of-core extensions who have problems as a\\nresult of my commits, but I also think it's fair to hope that those\\nmaintainers will try to minimize the amount of time that I need to\\nspend trying to read code that I did not write and do not maintain.\\nFortunately, this wasn't hard to figure out, but in a way that's kind\\nof the point. That DBL_MAX hack was put there by somebody who must've\\nunderstood that they were trying to use a very large cost to disable a\\ncertain path shape completely, and it seems to me that if that person\\nhad studied this case and the commit message for e2225346, they would\\nhave likely understood what had happened pretty quickly. Do you think\\nthat's an unfair feeling on my part?\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Aug 2024 13:11:45 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Fri, Aug 23, 2024 at 11:17 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>> We hit an issue with pgvector[0] where a regular `SELECT count(*) FROM\\n>> table`[1] is attempting to scan the index on the vector column when\\n>> `enable_seqscan` is disabled. Credit to Andrew Kane (CC'd) for flagging it.\\n\\n> It took me a moment to wrap my head around this: the cost estimate is\\n> 312 decimal digits long. Apparently hnswcostestimate() just returns\\n> DBL_MAX when there are no scan keys because it really, really doesn't\\n> want to do that. Before e2225346, that kept this plan from being\\n> generated because it was (much) larger than disable_cost. But now it\\n> doesn't, because 1 disabled node makes a path more expensive than any\\n> possible non-disabled path. Since that was the whole point of the\\n> patch, I don't feel too bad about it.\\n\\nYeah, I don't think it's necessary for v18 to be bug-compatible with\\nthis hack.\\n\\n> If you don't want to fix hnsw to work the way the core optimizer\\n> thinks it should, or if there's some reason it can't be done,\\n> alternatives might include (1) having the cost estimate function hack\\n> the count of disabled nodes and (2) adding some kind of core support\\n> for an index cost estimator refusing a path entirely. I haven't tested\\n> (1) so I don't know for sure that there are no issues, but I think we\\n> have to do all of our cost estimating before we can think about adding\\n> the path so I feel like there's a decent chance it would do what you\\n> want.\\n\\nIt looks like amcostestimate could change the path's disabled_nodes\\ncount, since that's set up before invoking amcostestimate. I guess\\nit could be set to INT_MAX to have a comparable solution to before.\\n\\nI agree with you that it is not great that hnsw is refusing this case\\nrather than finding a way to make it work, so I'm not excited about\\nputting in support for refusing it in a less klugy way.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Aug 2024 13:26:26 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 23, 2024 at 1:26 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> It looks like amcostestimate could change the path's disabled_nodes\\n> count, since that's set up before invoking amcostestimate. I guess\\n> it could be set to INT_MAX to have a comparable solution to before.\\n\\nIt's probably better to add a more modest value, to avoid overflow.\\nYou could add a million or so and be far away from overflow while\\npresumably still being more disabled than any other path.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Aug 2024 13:37:53 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Fri, Aug 23, 2024 at 1:26 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> It looks like amcostestimate could change the path's disabled_nodes\\n>> count, since that's set up before invoking amcostestimate. I guess\\n>> it could be set to INT_MAX to have a comparable solution to before.\\n\\n> It's probably better to add a more modest value, to avoid overflow.\\n> You could add a million or so and be far away from overflow while\\n> presumably still being more disabled than any other path.\\n\\nBut that'd only matter if the path survived its first add_path\\ntournament, which it shouldn't. If it does then you're at risk\\nof the same run-time failure reported here.\\n\\n(Having said that, you're likely right that \\\"a million or so\\\"\\nwould be a safer choice, since it doesn't require the assumption\\nthat the path fails instantly.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Aug 2024 13:42:53 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On 23/08/2024 20:11, Robert Haas wrote:\\n> I find it a little weird that hnsw thinks itself able to return all\\n> the tuples in an order the user chooses, but unable to return all of\\n> the tuples in an arbitrary order.\\n\\nHNSW is weird in many ways:\\n\\n- There is no inherent sort order. It cannot do \\\"ORDER BY column\\\", only \\nkNN-sort like \\\"ORDER BY column <-> value\\\".\\n\\n- It's approximate. It's not guaranteed to return the same set of rows \\nas a sequential scan + sort.\\n\\n- The number of results it returns is limited by the hnsw.ef_search GUC, \\ndefault 100.\\n\\n- It collects all the results (up to hnsw.ef_search) in memory, and only \\nthen returns them. So if you tried to use it with a large number of \\nresults, it can simply run out of memory.\\n\\nArguably all of those are bugs in HNSW, but it is what it is. The \\nalgorithm is inherently approximate. Despite that, it's useful in practice.\\n\\n> In core, we have precedent for index\\n> types that can't return individual tuples at all (gin, brin) but not\\n> one that is able to return tuples in concept but has a panic attack if\\n> you don't know how you want them sorted.\\n\\nWell, we do also have gin_fuzzy_search_limit. Two wrongs doesn't make it \\nright, though; I'd love to get rid of that hack too somehow.\\n\\n> I don't quite see why you\\n> couldn't just treat that case the same as ORDER BY\\n> the_first_column_of_the_index, or any other arbitrary rule that you\\n> want to make up. Sure, it might be more expensive than a sequential\\n> scan, but the user said they didn't want a sequential scan. I'm not\\n> quite sure why pgvector thinks it gets to decide that it knows better\\n> than the user, or the rest of the optimizer. I don't even think I\\n> really believe it would always be worse: I've seen cases where a table\\n> was badly bloated and mostly empty but its indexes were not bloated,\\n> and in that case an index scan can be a HUGE winner even though it\\n> would normally be a lot worse than a sequential scan.\\n\\nSure, you could make it work. It could construct a vector out of thin \\nair to compare with, when there's no scan key, or implement a completely \\ndifferent codepath that traverses the full graph in no particular order.\\n\\n> If you don't want to fix hnsw to work the way the core optimizer\\n> thinks it should, or if there's some reason it can't be done,\\n> alternatives might include (1) having the cost estimate function hack\\n> the count of disabled nodes and (2) adding some kind of core support\\n> for an index cost estimator refusing a path entirely. I haven't tested\\n> (1) so I don't know for sure that there are no issues, but I think we\\n> have to do all of our cost estimating before we can think about adding\\n> the path so I feel like there's a decent chance it would do what you\\n> want.\\n\\nIt would seem useful for an index AM to be able to say \\\"nope, I can't do \\nthis\\\". I don't remember how exactly this stuff works, but I'm surprised \\nit doesn't already exist.\\n\\n-- \\nHeikki Linnakangas\\nNeon (https://neon.tech)\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Aug 2024 21:18:32 +0300\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On 8/23/24 1:11 PM, Robert Haas wrote:\\r\\n> On Fri, Aug 23, 2024 at 11:17 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\r\\n>> We hit an issue with pgvector[0] where a regular `SELECT count(*) FROM\\r\\n>> table`[1] is attempting to scan the index on the vector column when\\r\\n>> `enable_seqscan` is disabled. Credit to Andrew Kane (CC'd) for flagging it.\\r\\n>>\\r\\n>> I was able to trace this back to e2225346. Here is a reproducer:\\r\\n> \\r\\n> If I change EXPLAIN ANALYZE in this test to just EXPLAIN, I get this:\\r\\n> \\r\\n> Aggregate (cost=179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00..179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00\\r\\n> rows=1 width=8)\\r\\n> -> Index Only Scan using test_embedding_idx on test\\r\\n> (cost=179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00..179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368.00\\r\\n> rows=5 width=0)\\r\\n> \\r\\n> It took me a moment to wrap my head around this: the cost estimate is\\r\\n> 312 decimal digits long. Apparently hnswcostestimate() just returns\\r\\n> DBL_MAX when there are no scan keys because it really, really doesn't\\r\\n> want to do that. Before e2225346, that kept this plan from being\\r\\n> generated because it was (much) larger than disable_cost. But now it\\r\\n> doesn't, because 1 disabled node makes a path more expensive than any\\r\\n> possible non-disabled path. Since that was the whole point of the\\r\\n> patch, I don't feel too bad about it.\\r\\n> \\r\\n> I find it a little weird that hnsw thinks itself able to return all\\r\\n> the tuples in an order the user chooses, but unable to return all of\\r\\n> the tuples in an arbitrary order. \\r\\n\\r\\nFor HNSW, \\\"order\\\" is approximated - even when it's returning \\\"in the \\r\\norder the user chooses,\\\" the scan is making the best guess at what the \\r\\ncorrect order is based on the index structure. At the traditional \\\"leaf\\\" \\r\\nlevel of an index, you're actually traversing a graph-based neighborhood \\r\\nof values. And maybe we could say \\\"Hey, if you get the equivalent of a \\r\\ncount(*), just do the count at the bottom layer (Layer 0)\\\" but I think \\r\\nthis would be very expensive.\\r\\n\\r\\n> In core, we have precedent for index\\r\\n> types that can't return individual tuples at all (gin, brin) but not\\r\\n> one that is able to return tuples in concept but has a panic attack if\\r\\n> you don't know how you want them sorted. I don't quite see why you\\r\\n> couldn't just treat that case the same as ORDER BY\\r\\n> the_first_column_of_the_index, or any other arbitrary rule that you\\r\\n> want to make up. Sure, it might be more expensive than a sequential\\r\\n> scan, but the user said they didn't want a sequential scan. I'm not\\r\\n> quite sure why pgvector thinks it gets to decide that it knows better\\r\\n> than the user, or the rest of the optimizer. I don't even think I\\r\\n> really believe it would always be worse: I've seen cases where a table\\r\\n> was badly bloated and mostly empty but its indexes were not bloated,\\r\\n> and in that case an index scan can be a HUGE winner even though it\\r\\n> would normally be a lot worse than a sequential scan.\\r\\n\\r\\nThe challenge here is that HNSW is used specifically for approximating \\r\\nordering; it's not used to directly filter results in the traditional \\r\\nsense (e.g. via. a WHERE clause). It's a bit different than the others \\r\\nmentioned in that regard. However, maybe there are other options to \\r\\nconsider here based on this work.\\r\\n\\r\\n> If you don't want to fix hnsw to work the way the core optimizer\\r\\n> thinks it should, or if there's some reason it can't be done,\\r\\n> alternatives might include (1) having the cost estimate function hack\\r\\n> the count of disabled nodes and (2) adding some kind of core support\\r\\n> for an index cost estimator refusing a path entirely. I haven't tested\\r\\n> (1) so I don't know for sure that there are no issues, but I think we\\r\\n> have to do all of our cost estimating before we can think about adding\\r\\n> the path so I feel like there's a decent chance it would do what you\\r\\n> want.\\r\\n\\r\\nThanks for the options.\\r\\n\\r\\n> Also, while I did take the initiative to download pgvector and compile\\r\\n> it and hook up a debugger and figure out what was going on here, I'm\\r\\n> not really too sure that's my job. I do think I have a responsibility\\r\\n> to help maintainers of out-of-core extensions who have problems as a\\r\\n> result of my commits, but I also think it's fair to hope that those\\r\\n> maintainers will try to minimize the amount of time that I need to\\r\\n> spend trying to read code that I did not write and do not maintain.\\r\\n> Fortunately, this wasn't hard to figure out, but in a way that's kind\\r\\n> of the point. That DBL_MAX hack was put there by somebody who must've\\r\\n> understood that they were trying to use a very large cost to disable a\\r\\n> certain path shape completely, and it seems to me that if that person\\r\\n> had studied this case and the commit message for e2225346, they would\\r\\n> have likely understood what had happened pretty quickly. Do you think\\r\\n> that's an unfair feeling on my part?\\r\\n\\r\\nI don't think extension maintainers necessarily have the same level of \\r\\nPostgreSQL internals as you or many of the other people who frequent \\r\\n-hackers, so I think it's fair for them to ask questions or raise issues \\r\\nwith patches they don't understand. I was able to glean from the commit \\r\\nmessage that this was the commit that likely changed the behavior in \\r\\npgvector, but I can't immediately glean looking through the code as to \\r\\nwhy. (And using your logic, should an extension maintainer understand \\r\\nthe optimizer code when PostgreSQL is providing an interface to the \\r\\nextension maintainer to encapsulate its interactions)?\\r\\n\\r\\nYou can always push back and say \\\"Well, maybe try this, or try that\\\" - \\r\\nwhich would be a mentoring approach that could push it back on the \\r\\nextension maintainer, which is valid, but I don't see why an extension \\r\\nmaintainer can't raise an issue or ask a question here.\\r\\n\\r\\nThanks,\\r\\n\\r\\nJonathan\",\n \"msg_date\": \"Fri, 23 Aug 2024 14:19:56 -0400\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 23, 2024 at 2:20 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n> I don't think extension maintainers necessarily have the same level of\\n> PostgreSQL internals as you or many of the other people who frequent\\n> -hackers, so I think it's fair for them to ask questions or raise issues\\n> with patches they don't understand. I was able to glean from the commit\\n> message that this was the commit that likely changed the behavior in\\n> pgvector, but I can't immediately glean looking through the code as to\\n> why. (And using your logic, should an extension maintainer understand\\n> the optimizer code when PostgreSQL is providing an interface to the\\n> extension maintainer to encapsulate its interactions)?\\n>\\n> You can always push back and say \\\"Well, maybe try this, or try that\\\" -\\n> which would be a mentoring approach that could push it back on the\\n> extension maintainer, which is valid, but I don't see why an extension\\n> maintainer can't raise an issue or ask a question here.\\n\\nI'm certainly not saying that extension maintainers can't raise issues\\nor ask questions here. I just feel that the problem could have been\\nanalyzed a bit more before posting.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Aug 2024 14:29:16 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 23, 2024 at 2:18 PM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n> It would seem useful for an index AM to be able to say \\\"nope, I can't do\\n> this\\\". I don't remember how exactly this stuff works, but I'm surprised\\n> it doesn't already exist.\\n\\nYeah, I think so, too. While this particular problem is due to a\\nproblem with an out-of-core AM that may be doing some slightly\\nquestionable things, there's not really any reason why we couldn't\\nhave similar problems in core for some other reason. For example, we\\ncould change amcostestimate's signature so that an extension can\\nreturn true or false, with false meaning that the path can't be\\nsupported. We could then change cost_index so that it can also return\\ntrue or false, and then change create_index_path so it has the option\\nto return NULL. Callers of create_index_path could then be adjusted\\nnot to call add_path when NULL is returned.\\n\\nThere might be a more elegant way to do it with more refactoring, but\\nthe above seems good enough.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Aug 2024 14:36:09 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Fri, Aug 23, 2024 at 2:18 PM Heikki Linnakangas <hlinnaka@iki.fi> wrote:\\n>> It would seem useful for an index AM to be able to say \\\"nope, I can't do\\n>> this\\\". I don't remember how exactly this stuff works, but I'm surprised\\n>> it doesn't already exist.\\n\\n> Yeah, I think so, too. While this particular problem is due to a\\n> problem with an out-of-core AM that may be doing some slightly\\n> questionable things, there's not really any reason why we couldn't\\n> have similar problems in core for some other reason. For example, we\\n> could change amcostestimate's signature so that an extension can\\n> return true or false, with false meaning that the path can't be\\n> supported. We could then change cost_index so that it can also return\\n> true or false, and then change create_index_path so it has the option\\n> to return NULL. Callers of create_index_path could then be adjusted\\n> not to call add_path when NULL is returned.\\n\\nIf we're going to do this, I'd prefer a solution that doesn't force\\nAPI changes onto the vast majority of index AMs that don't have a\\nproblem here.\\n\\nOne way could be to formalize the hack we were just discussing:\\n\\\"To refuse a proposed path, amcostestimate can set the path's\\ndisabled_nodes value to anything larger than 1\\\". I suspect that\\nthat would actually be sufficient, since the path would then lose\\nto the seqscan path in add_path even if that were disabled; but\\nwe could put in a hack to prevent it from getting add_path'd at all.\\n\\nAnother way could be to bless what hnsw is already doing:\\n\\\"To refuse a proposed path, amcostestimate can return an\\nindexTotalCost of DBL_MAX\\\" (or maybe insisting on +Inf would\\nbe better). That would still require changes comparable to\\nwhat you specify above, but only in the core-code call path\\nnot in every AM.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Aug 2024 14:48:35 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 23, 2024 at 2:48 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> If we're going to do this, I'd prefer a solution that doesn't force\\n> API changes onto the vast majority of index AMs that don't have a\\n> problem here.\\n\\nThat's a fair concern.\\n\\n> One way could be to formalize the hack we were just discussing:\\n> \\\"To refuse a proposed path, amcostestimate can set the path's\\n> disabled_nodes value to anything larger than 1\\\". I suspect that\\n> that would actually be sufficient, since the path would then lose\\n> to the seqscan path in add_path even if that were disabled; but\\n> we could put in a hack to prevent it from getting add_path'd at all.\\n>\\n> Another way could be to bless what hnsw is already doing:\\n> \\\"To refuse a proposed path, amcostestimate can return an\\n> indexTotalCost of DBL_MAX\\\" (or maybe insisting on +Inf would\\n> be better). That would still require changes comparable to\\n> what you specify above, but only in the core-code call path\\n> not in every AM.\\n\\nIf just setting disabled_nodes to a value larger than one works, I'd\\nbe inclined to not do anything here at all, except possibly document\\nthat you can do that. Otherwise, we should probably change the code\\nsomehow.\\n\\nI find both of your proposed solutions above to be pretty inelegant,\\nand I think if this problem occurred with a core AM, I'd push for an\\nAPI break rather than accept the ugliness. \\\"This path is not valid\\nbecause the AM cannot support it\\\", \\\"this path is crazy expensive\\\", and\\n\\\"the user told us not to do it this way\\\" are three different things,\\nand signalling two or more of them in the same way muddies the water\\nin a way that I don't like. API breaks aren't free, though, so I\\ncertainly understand why you're not very keen to introduce one where\\nit can reasonably be avoided.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Aug 2024 15:05:34 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On 23/08/2024 22:05, Robert Haas wrote:\\n> On Fri, Aug 23, 2024 at 2:48 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> If we're going to do this, I'd prefer a solution that doesn't force\\n>> API changes onto the vast majority of index AMs that don't have a\\n>> problem here.\\n> \\n> That's a fair concern.\\n\\nYeah, although I don't think it's too bad. There are not that many \\nout-of-tree index AM implementations to begin with, and we do change \\nthings often enough that any interesting AM implementation will likely \\nneed a few #ifdef PG_VERSION blocks for each PostgreSQL major version \\nanyway. pgvector certainly does.\\n\\n>> One way could be to formalize the hack we were just discussing:\\n>> \\\"To refuse a proposed path, amcostestimate can set the path's\\n>> disabled_nodes value to anything larger than 1\\\". I suspect that\\n>> that would actually be sufficient, since the path would then lose\\n>> to the seqscan path in add_path even if that were disabled; but\\n>> we could put in a hack to prevent it from getting add_path'd at all.\\n>>\\n>> Another way could be to bless what hnsw is already doing:\\n>> \\\"To refuse a proposed path, amcostestimate can return an\\n>> indexTotalCost of DBL_MAX\\\" (or maybe insisting on +Inf would\\n>> be better). That would still require changes comparable to\\n>> what you specify above, but only in the core-code call path\\n>> not in every AM.\\n> \\n> If just setting disabled_nodes to a value larger than one works, I'd\\n> be inclined to not do anything here at all, except possibly document\\n> that you can do that. Otherwise, we should probably change the code\\n> somehow.\\n\\nModifying the passed-in Path feels hacky. amcostestimate currently \\nreturns all the estimates in *output parameters, it doesn't modify the \\nPath at all.\\n\\n> I find both of your proposed solutions above to be pretty inelegant,\\n> and I think if this problem occurred with a core AM, I'd push for an\\n> API break rather than accept the ugliness. \\\"This path is not valid\\n> because the AM cannot support it\\\", \\\"this path is crazy expensive\\\", and\\n> \\\"the user told us not to do it this way\\\" are three different things,\\n> and signalling two or more of them in the same way muddies the water\\n> in a way that I don't like. API breaks aren't free, though, so I\\n> certainly understand why you're not very keen to introduce one where\\n> it can reasonably be avoided.\\n\\nThe +Inf approach seems fine to me. Or perhaps NaN. Your proposal would \\ncertainly be the cleanest interface if we don't mind incurring churn to \\nAM implementations.\\n\\n-- \\nHeikki Linnakangas\\nNeon (https://neon.tech)\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Aug 2024 22:12:55 +0300\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> I find both of your proposed solutions above to be pretty inelegant,\\n\\nThey are that. If we were working in a green field I'd not propose\\nsuch things ... but we aren't. I believe there are now a fair number\\nof out-of-core index AMs, so I'd rather not break all of them if we\\ndon't have to.\\n\\n> and I think if this problem occurred with a core AM, I'd push for an\\n> API break rather than accept the ugliness. \\\"This path is not valid\\n> because the AM cannot support it\\\", \\\"this path is crazy expensive\\\", and\\n> \\\"the user told us not to do it this way\\\" are three different things,\\n> and signalling two or more of them in the same way muddies the water\\n> in a way that I don't like.\\n\\nI think it's not that bad, because we can limit the knowledge of this\\nhack to the amcostestimate interface, which doesn't really deal in\\n\\\"the user told us not to do it this way\\\" at all. That argues against\\nmy first proposal though (having amcostestimate touch disabled_nodes\\ndirectly). I now think that a reasonable compromise is to say that\\nsetting indexTotalCost to +Inf signals that \\\"the AM cannot support\\nit\\\". That's not conflated too much with the other case, since even a\\ncrazy-expensive cost estimate surely ought to be finite. We can have\\ncost_index untangle that case into a separate failure return so that\\nthe within-the-core-optimizer APIs remain clean.\\n\\nWhile that would require hnsw to make a small code change (return\\n+Inf not DBL_MAX), that coding should work in back branches too,\\nso they don't even need a version check.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 23 Aug 2024 15:32:11 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On 8/23/24 2:29 PM, Robert Haas wrote:\\r\\n> On Fri, Aug 23, 2024 at 2:20 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\r\\n>> I don't think extension maintainers necessarily have the same level of\\r\\n>> PostgreSQL internals as you or many of the other people who frequent\\r\\n>> -hackers, so I think it's fair for them to ask questions or raise issues\\r\\n>> with patches they don't understand. I was able to glean from the commit\\r\\n>> message that this was the commit that likely changed the behavior in\\r\\n>> pgvector, but I can't immediately glean looking through the code as to\\r\\n>> why. (And using your logic, should an extension maintainer understand\\r\\n>> the optimizer code when PostgreSQL is providing an interface to the\\r\\n>> extension maintainer to encapsulate its interactions)?\\r\\n>>\\r\\n>> You can always push back and say \\\"Well, maybe try this, or try that\\\" -\\r\\n>> which would be a mentoring approach that could push it back on the\\r\\n>> extension maintainer, which is valid, but I don't see why an extension\\r\\n>> maintainer can't raise an issue or ask a question here.\\r\\n> \\r\\n> I'm certainly not saying that extension maintainers can't raise issues\\r\\n> or ask questions here. I just feel that the problem could have been\\r\\n> analyzed a bit more before posting.\\r\\n\\r\\nThis assumes that the person posting the problem has the requisite \\r\\nexpertise to determine what the issue is. Frankly, I was happy I was \\r\\nable to at least trace the issue down to the particular commit and \\r\\nbrought what appeared to be a reliable reproducer, in absence of knowing \\r\\nif 1/ this was actually an issue with PG or pgvector, 2/ does it \\r\\nactually require a fix, or 3/ what the problem could actually be, given \\r\\na lack of understanding of the full inner working of the optimizer.\\r\\n\\r\\nBased on the above, I'm not sure what bar this needed to clear to begin \\r\\na discussion on the mailing list (which further downthread, seems to be \\r\\nraising some interesting points).\\r\\n\\r\\nJonathan\",\n \"msg_date\": \"Fri, 23 Aug 2024 17:29:10 -0400\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On 8/23/24 3:32 PM, Tom Lane wrote:\\r\\n> Robert Haas <robertmhaas@gmail.com> writes:\\r\\n>> I find both of your proposed solutions above to be pretty inelegant,\\r\\n> \\r\\n> They are that. If we were working in a green field I'd not propose\\r\\n> such things ... but we aren't. I believe there are now a fair number\\r\\n> of out-of-core index AMs, so I'd rather not break all of them if we\\r\\n> don't have to.\\r\\n\\r\\nFor distribution of index AMs in the wild, it's certainly > 1 now, and \\r\\nincreasing. They're not the easiest extension types to build out, so \\r\\nit's not as widely distributed as some of the other APIs, but there are \\r\\na bunch out there, as well as language-specific libs (e.g. pgrx for \\r\\nRust) that offer wrappers around them.\\r\\n\\r\\n>> and I think if this problem occurred with a core AM, I'd push for an\\r\\n>> API break rather than accept the ugliness. \\\"This path is not valid\\r\\n>> because the AM cannot support it\\\", \\\"this path is crazy expensive\\\", and\\r\\n>> \\\"the user told us not to do it this way\\\" are three different things,\\r\\n>> and signalling two or more of them in the same way muddies the water\\r\\n>> in a way that I don't like.\\r\\n> \\r\\n> I think it's not that bad, because we can limit the knowledge of this\\r\\n> hack to the amcostestimate interface, which doesn't really deal in\\r\\n> \\\"the user told us not to do it this way\\\" at all. That argues against\\r\\n> my first proposal though (having amcostestimate touch disabled_nodes\\r\\n> directly). I now think that a reasonable compromise is to say that\\r\\n> setting indexTotalCost to +Inf signals that \\\"the AM cannot support\\r\\n> it\\\". That's not conflated too much with the other case, since even a\\r\\n> crazy-expensive cost estimate surely ought to be finite. We can have\\r\\n> cost_index untangle that case into a separate failure return so that\\r\\n> the within-the-core-optimizer APIs remain clean.\\r\\n> \\r\\n> While that would require hnsw to make a small code change (return\\r\\n> +Inf not DBL_MAX), that coding should work in back branches too,\\r\\n> so they don't even need a version check.\\r\\n\\r\\n+1 for this approach (I'll do a quick test in my pgvector workspace just \\r\\nto ensure it gets the same results in the older version).\\r\\n\\r\\nJonathan\",\n \"msg_date\": \"Fri, 23 Aug 2024 17:33:12 -0400\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On 8/23/24 5:33 PM, Jonathan S. Katz wrote:\\r\\n> On 8/23/24 3:32 PM, Tom Lane wrote:\\r\\n>> Robert Haas <robertmhaas@gmail.com> writes:\\r\\n\\r\\n>> I think it's not that bad, because we can limit the knowledge of this\\r\\n>> hack to the amcostestimate interface, which doesn't really deal in\\r\\n>> \\\"the user told us not to do it this way\\\" at all. That argues against\\r\\n>> my first proposal though (having amcostestimate touch disabled_nodes\\r\\n>> directly). I now think that a reasonable compromise is to say that\\r\\n>> setting indexTotalCost to +Inf signals that \\\"the AM cannot support\\r\\n>> it\\\". That's not conflated too much with the other case, since even a\\r\\n>> crazy-expensive cost estimate surely ought to be finite. We can have\\r\\n>> cost_index untangle that case into a separate failure return so that\\r\\n>> the within-the-core-optimizer APIs remain clean.\\r\\n>>\\r\\n>> While that would require hnsw to make a small code change (return\\r\\n>> +Inf not DBL_MAX), that coding should work in back branches too,\\r\\n>> so they don't even need a version check.\\r\\n> \\r\\n> +1 for this approach (I'll do a quick test in my pgvector workspace just \\r\\n> to ensure it gets the same results in the older version).\\r\\n\\r\\n...and I confirmed the +inf approach on PG16 +pgvector does still give \\r\\nthe same expected result.\\r\\n\\r\\nThanks,\\r\\n\\r\\nJonathan\",\n \"msg_date\": \"Fri, 23 Aug 2024 17:44:37 -0400\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Hello Robert,\\n\\n21.08.2024 17:29, Robert Haas wrote:\\n> I went ahead and committed these patches. ...\\n\\nPlease take a look at the following code:\\nstatic void\\nlabel_sort_with_costsize(PlannerInfo *root, Sort *plan, double limit_tuples)\\n{\\n...\\n cost_sort(&sort_path, root, NIL,\\n lefttree->total_cost,\\n plan->plan.disabled_nodes,\\n lefttree->plan_rows,\\n lefttree->plan_width,\\n 0.0,\\n work_mem,\\n limit_tuples);\\n\\nGiven the cost_sort() declaration:\\nvoid\\ncost_sort(Path *path, PlannerInfo *root,\\n List *pathkeys, int input_disabled_nodes,\\n Cost input_cost, double tuples, int width,\\n Cost comparison_cost, int sort_mem,\\n double limit_tuples)\\n\\nAren't the input_disabled_nodes and input_cost arguments swapped in the\\nabove call?\\n\\n(I've discovered this with UBSan, which complained\\ncreateplan.c:5457:6: runtime error: 4.40465e+09 is outside the range of representable values of type 'int'\\nwhile executing a query with a large estimated cost.)\\n\\nBest regards,\\nAlexander\\n\\n\\n\",\n \"msg_date\": \"Fri, 6 Sep 2024 12:00:00 +0300\",\n \"msg_from\": \"Alexander Lakhin <exclusion@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Sep 6, 2024 at 5:00 PM Alexander Lakhin <exclusion@gmail.com> wrote:\\n> static void\\n> label_sort_with_costsize(PlannerInfo *root, Sort *plan, double limit_tuples)\\n> {\\n> ...\\n> cost_sort(&sort_path, root, NIL,\\n> lefttree->total_cost,\\n> plan->plan.disabled_nodes,\\n> lefttree->plan_rows,\\n> lefttree->plan_width,\\n> 0.0,\\n> work_mem,\\n> limit_tuples);\\n>\\n> Given the cost_sort() declaration:\\n> void\\n> cost_sort(Path *path, PlannerInfo *root,\\n> List *pathkeys, int input_disabled_nodes,\\n> Cost input_cost, double tuples, int width,\\n> Cost comparison_cost, int sort_mem,\\n> double limit_tuples)\\n>\\n> Aren't the input_disabled_nodes and input_cost arguments swapped in the\\n> above call?\\n\\nNice catch! I checked other callers to cost_sort, and they are all\\ngood.\\n\\n(I'm a little surprised that this does not cause any plan diffs in the\\nregression tests.)\\n\\nThanks\\nRichard\\n\\n\\n\",\n \"msg_date\": \"Fri, 6 Sep 2024 17:27:22 +0800\",\n \"msg_from\": \"Richard Guo <guofenglinux@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Sep 6, 2024 at 5:27 PM Richard Guo <guofenglinux@gmail.com> wrote:\\n> On Fri, Sep 6, 2024 at 5:00 PM Alexander Lakhin <exclusion@gmail.com> wrote:\\n> > static void\\n> > label_sort_with_costsize(PlannerInfo *root, Sort *plan, double limit_tuples)\\n\\n> (I'm a little surprised that this does not cause any plan diffs in the\\n> regression tests.)\\n\\nAh I see. label_sort_with_costsize is only used to label the Sort\\nnode nicely for EXPLAIN, and usually we do not display the cost\\nnumbers in regression tests.\\n\\nThanks\\nRichard\\n\\n\\n\",\n \"msg_date\": \"Fri, 6 Sep 2024 17:51:05 +0800\",\n \"msg_from\": \"Richard Guo <guofenglinux@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"Hello Richard,\\n\\n06.09.2024 12:51, Richard Guo wrote:\\n> Ah I see. label_sort_with_costsize is only used to label the Sort\\n> node nicely for EXPLAIN, and usually we do not display the cost\\n> numbers in regression tests.\\n\\nIn fact, I see the error with the following (EXPLAIN-less) query:\\ncreate table t (x int);\\n\\nselect * from t natural inner join\\n(select * from (values(1)) v(x)\\n union all\\n select 1 from t t1 full join t t2 using (x),\\n t t3 full join t t4 using (x)\\n);\\n\\n2024-09-06 10:01:48.034 UTC [696535:5] psql LOG: statement: select * from t natural inner join\\n (select * from (values(1)) v(x)\\n union all\\n select 1 from t t1 full join t t2 using (x),\\n t t3 full join t t4 using (x)\\n );\\ncreateplan.c:5457:6: runtime error: 4.99254e+09 is outside the range of representable values of type 'int'\\nSUMMARY: UndefinedBehaviorSanitizer: undefined-behavior createplan.c:5457:6 in\\n\\n(An UBSan-enabled build --with-blocksize=32 is required for this query to\\ntrigger the failure.)\\n\\nBest regards,\\nAlexander\\n\\n\\n\",\n \"msg_date\": \"Fri, 6 Sep 2024 13:10:10 +0300\",\n \"msg_from\": \"Alexander Lakhin <exclusion@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, Sep 6, 2024 at 5:27 PM Richard Guo <guofenglinux@gmail.com> wrote:\\n> On Fri, Sep 6, 2024 at 5:00 PM Alexander Lakhin <exclusion@gmail.com> wrote:\\n> > static void\\n> > label_sort_with_costsize(PlannerInfo *root, Sort *plan, double limit_tuples)\\n> > {\\n> > ...\\n> > cost_sort(&sort_path, root, NIL,\\n> > lefttree->total_cost,\\n> > plan->plan.disabled_nodes,\\n> > lefttree->plan_rows,\\n> > lefttree->plan_width,\\n> > 0.0,\\n> > work_mem,\\n> > limit_tuples);\\n> >\\n> > Given the cost_sort() declaration:\\n> > void\\n> > cost_sort(Path *path, PlannerInfo *root,\\n> > List *pathkeys, int input_disabled_nodes,\\n> > Cost input_cost, double tuples, int width,\\n> > Cost comparison_cost, int sort_mem,\\n> > double limit_tuples)\\n> >\\n> > Aren't the input_disabled_nodes and input_cost arguments swapped in the\\n> > above call?\\n>\\n> Nice catch! I checked other callers to cost_sort, and they are all\\n> good.\\n\\nFixed.\\n\\nThanks\\nRichard\\n\\n\\n\",\n \"msg_date\": \"Mon, 9 Sep 2024 12:09:26 +0800\",\n \"msg_from\": \"Richard Guo <guofenglinux@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 9, 2024 at 12:09 AM Richard Guo <guofenglinux@gmail.com> wrote:\\n> Fixed.\\n\\nThanks to Alexander for the very good catch and to Richard for pushing the fix.\\n\\n(I started to respond to this last week but didn't quite get to it\\nbefore I ran out of time/energy.)\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 9 Sep 2024 11:28:44 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Wed, 2024-08-21 at 10:29 -0400, Robert Haas wrote:\\n> I went ahead and committed these patches. I know there's some debate\\n> over whether we want to show the # of disabled nodes and if so whether\\n> it should be controlled by COSTS, and I suspect I haven't completely\\n> allayed David's concerns about the initial_cost_XXX functions although\\n> I think that I did the right thing. But, I don't have the impression\\n> that anyone is desperately opposed to the basic concept, so I think it\\n> makes sense to put these into the tree and see what happens. We have\\n> quite a bit of time left in this release cycle to uncover bugs, hear\\n> from users or other developers, etc. about what problems there may be\\n> with this. If we end up deciding to reverse course or need to fix a\\n> bunch of stuff, so be it, but let's see what the feedback is.\\n\\nI am somewhat unhappy about the \\\"Disabled Nodes\\\" in EXPLAIN.\\n\\nFirst, the commit message confused me: it claims that the information\\nis displayed with EXPLAIN ANALYZE, but it's shown with every EXPLAIN.\\n\\nBut that's not important. My complaints are:\\n\\n1. The \\\"disabled nodes\\\" are always displayed.\\n I'd be happier if it were only shown for COSTS ON, but I think it\\n would be best if they were only shown with VERBOSE ON.\\n\\n After all, the messages are pretty verbose...\\n\\n2. The \\\"disabled nodes\\\" are not only shown at the nodes where nodes\\n were actually disabled, but also at every nodes above these nodes.\\n\\n This would be fine:\\n\\n Sort\\n -> Nested Loop Join\\n -> Hash Join\\n -> Index Scan\\n Disabled Nodes: 1\\n -> Hash\\n -> Index Scan\\n Disabled Nodes: 1\\n -> Index Scan\\n Disabled Nodes: 1\\n\\n This is annoying:\\n\\n Sort\\n Disabled Nodes: 3\\n -> Nested Loop Join\\n Disabled Nodes: 3\\n -> Hash Join\\n Disabled Nodes: 2\\n -> Index Scan\\n Disabled Nodes: 1\\n -> Hash\\n -> Index Scan\\n Disabled Nodes: 1\\n -> Index Scan\\n Disabled Nodes: 1\\n\\nI have no idea how #2 could be implemented, but it would be nice to have.\\nPlease, please, can we show the \\\"disabled nodes\\\" only with VERBOSE?\\n\\nYours,\\nLaurenz Albe\\n\\n\\n\",\n \"msg_date\": \"Fri, 27 Sep 2024 10:42:32 +0200\",\n \"msg_from\": \"Laurenz Albe <laurenz.albe@cybertec.at>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Fri, 27 Sept 2024 at 20:42, Laurenz Albe <laurenz.albe@cybertec.at> wrote:\\n> 2. The \\\"disabled nodes\\\" are not only shown at the nodes where nodes\\n> were actually disabled, but also at every nodes above these nodes.\\n\\nI'm also not a fan either and I'd like to see this output improved.\\n\\nIt seems like it's easy enough to implement some logic to detect when\\na given node is disabled just by checking if the disable_nodes count\\nis higher than the sum of the disabled_node field of the node's\\nchildren. If there are no children (a scan node) and disabed_nodes >\\n0 then it must be disabled. There's even a nice fast path where we\\ndon't need to check the children if disabled_nodes == 0.\\n\\nHere's a POC grade patch of how I'd rather see it looking.\\n\\nI opted to have a boolean field as I didn't see any need for an\\ninteger count. I also changed things around so we always display the\\nboolean property in non-text EXPLAIN. Normally, we don't mind being\\nmore verbose there.\\n\\nI also fixed a bug in make_sort() where disabled_nodes isn't being set\\nproperly. I'll do an independent patch for that if this goes nowhere.\\n\\nDavid\",\n \"msg_date\": \"Sat, 28 Sep 2024 00:04:20 +1200\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n },\n {\n \"msg_contents\": \"On Sat, 2024-09-28 at 00:04 +1200, David Rowley wrote:\\n> On Fri, 27 Sept 2024 at 20:42, Laurenz Albe <laurenz.albe@cybertec.at> wrote:\\n> > 2. The \\\"disabled nodes\\\" are not only shown at the nodes where nodes\\n> > were actually disabled, but also at every nodes above these nodes.\\n> \\n> I'm also not a fan either and I'd like to see this output improved.\\n> \\n> It seems like it's easy enough to implement some logic to detect when\\n> a given node is disabled just by checking if the disable_nodes count\\n> is higher than the sum of the disabled_node field of the node's\\n> children. If there are no children (a scan node) and disabed_nodes >\\n> 0 then it must be disabled. There's even a nice fast path where we\\n> don't need to check the children if disabled_nodes == 0.\\n> \\n> Here's a POC grade patch of how I'd rather see it looking.\\n> \\n> I opted to have a boolean field as I didn't see any need for an\\n> integer count. I also changed things around so we always display the\\n> boolean property in non-text EXPLAIN. Normally, we don't mind being\\n> more verbose there.\\n> \\n> I also fixed a bug in make_sort() where disabled_nodes isn't being set\\n> properly. I'll do an independent patch for that if this goes nowhere.\\n\\nThanks, and the patch looks good.\\n\\nWhy did you change \\\"Disabled\\\" from an integer to a boolean?\\nIf you see a join where two plans were disabled, that's useful information.\\n\\nI would still prefer to see the disabled nodes only in VERBOSE explain,\\nbut I'm satisfied if the disabled nodes don't show up all over the place.\\n\\nYours,\\nLaurenz Albe\\n\\n\\n\",\n \"msg_date\": \"Mon, 30 Sep 2024 19:17:38 +0200\",\n \"msg_from\": \"Laurenz Albe <laurenz.albe@cybertec.at>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: On disable_cost\"\n }\n]","isConversational":false}}},{"rowIdx":49390,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"This patch moves the parse analysis component of ExecuteQuery() and\nEvaluateParams() into a new transformExecuteStmt() that is called from\ntransformStmt(). This makes EXECUTE behave more like other utility\ncommands.\n\nEffects are that error messages can have position information (see \nregression test case), and it allows using external parameters in the \narguments of the EXECUTE command.\n\nI had previously inquired about this in [0] and some vague concerns were \nraised. I haven't dug very deep on this, but I figure with an actual \npatch it might be easier to review and figure out if there are any problems.\n\n\n[0]: \nhttps://www.postgresql.org/message-id/flat/ed2767e5-c506-048d-8ddf-280ecbc9e1b7%402ndquadrant.com\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 1 Nov 2019 08:07:16 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Refactor parse analysis of EXECUTE command"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> This patch moves the parse analysis component of ExecuteQuery() and\n> EvaluateParams() into a new transformExecuteStmt() that is called from\n> transformStmt().\n\nUhmm ... no actual patch attached?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 02 Nov 2019 11:00:30 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"On 2019-11-02 16:00, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n>> This patch moves the parse analysis component of ExecuteQuery() and\n>> EvaluateParams() into a new transformExecuteStmt() that is called from\n>> transformStmt().\n> \n> Uhmm ... no actual patch attached?\n\nOops, here it is.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Mon, 4 Nov 2019 08:53:18 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"Hello.\n\nAt Mon, 4 Nov 2019 08:53:18 +0100, Peter Eisentraut <peter.eisentraut@2ndquadrant.com> wrote in \n> On 2019-11-02 16:00, Tom Lane wrote:\n> > Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> >> This patch moves the parse analysis component of ExecuteQuery() and\n> >> EvaluateParams() into a new transformExecuteStmt() that is called from\n> >> transformStmt().\n> > Uhmm ... no actual patch attached?\n> \n> Oops, here it is.\n\nThe patch just moves the first half of EvaluateParams that is\nirrelevant to executor state to before portal parameters are set. I\nlooked with a suspect that extended protocol or SPI are affected but\nAFAICS it doesn't seem to.\n\nI dug into repository and found that transformExecuteStmt existed at\nthe time of implementing PREPARE-EXECUTE statements(28e82066a1) and\nremoved by the commit b9527e9840 which is related to\nplan-invalidation.\n\ngit show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\n> In service of this, rearrange utility-statement processing so that parse\n> analysis does not assume table schemas can't change before execution for\n> utility statements (necessary because we don't attempt to re-acquire locks\n> for utility statements when reusing a stored plan). This requires some\n\nIsn't this related to the current structure?\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Tue, 05 Nov 2019 19:27:48 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"út 5. 11. 2019 v 11:28 odesílatel Kyotaro Horiguchi <horikyota.ntt@gmail.com>\nnapsal:\n\n> Hello.\n>\n> At Mon, 4 Nov 2019 08:53:18 +0100, Peter Eisentraut <\n> peter.eisentraut@2ndquadrant.com> wrote in\n> > On 2019-11-02 16:00, Tom Lane wrote:\n> > > Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> > >> This patch moves the parse analysis component of ExecuteQuery() and\n> > >> EvaluateParams() into a new transformExecuteStmt() that is called from\n> > >> transformStmt().\n> > > Uhmm ... no actual patch attached?\n> >\n> > Oops, here it is.\n>\n> The patch just moves the first half of EvaluateParams that is\n> irrelevant to executor state to before portal parameters are set. I\n> looked with a suspect that extended protocol or SPI are affected but\n> AFAICS it doesn't seem to.\n>\n> I dug into repository and found that transformExecuteStmt existed at\n> the time of implementing PREPARE-EXECUTE statements(28e82066a1) and\n> removed by the commit b9527e9840 which is related to\n> plan-invalidation.\n>\n> git show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\n> > In service of this, rearrange utility-statement processing so that parse\n> > analysis does not assume table schemas can't change before execution for\n> > utility statements (necessary because we don't attempt to re-acquire\n> locks\n> > for utility statements when reusing a stored plan). This requires some\n>\n> Isn't this related to the current structure?\n>\n\nI think so it should be ok, because the transformation is still in same\nstatement - if I understand well.\n\nSo visibility of system catalogue or access to plan cache should not be\nchanged.\n\nRegards\n\nPavel\n\n\n> regards.\n>\n> --\n> Kyotaro Horiguchi\n> NTT Open Source Software Center\n>\n>\n>\n\nút 5. 11. 2019 v 11:28 odesílatel Kyotaro Horiguchi <horikyota.ntt@gmail.com> napsal:Hello.\n\nAt Mon, 4 Nov 2019 08:53:18 +0100, Peter Eisentraut <peter.eisentraut@2ndquadrant.com> wrote in \n> On 2019-11-02 16:00, Tom Lane wrote:\n> > Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> >> This patch moves the parse analysis component of ExecuteQuery() and\n> >> EvaluateParams() into a new transformExecuteStmt() that is called from\n> >> transformStmt().\n> > Uhmm ... no actual patch attached?\n> \n> Oops, here it is.\n\nThe patch just moves the first half of EvaluateParams that is\nirrelevant to executor state to before portal parameters are set. I\nlooked with a suspect that extended protocol or SPI are affected but\nAFAICS it doesn't seem to.\n\nI dug into repository and found that transformExecuteStmt existed at\nthe time of implementing PREPARE-EXECUTE statements(28e82066a1) and\nremoved by the commit b9527e9840 which is related to\nplan-invalidation.\n\ngit show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\n> In service of this, rearrange utility-statement processing so that parse\n> analysis does not assume table schemas can't change before execution for\n> utility statements (necessary because we don't attempt to re-acquire locks\n> for utility statements when reusing a stored plan). This requires some\n\nIsn't this related to the current structure?I think so it should be ok, because the transformation is still in same statement - if I understand well. So visibility of system catalogue or access to plan cache should not be changed.RegardsPavel\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center","msg_date":"Fri, 8 Nov 2019 08:13:55 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"po 4. 11. 2019 v 8:53 odesílatel Peter Eisentraut <\npeter.eisentraut@2ndquadrant.com> napsal:\n\n> On 2019-11-02 16:00, Tom Lane wrote:\n> > Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> >> This patch moves the parse analysis component of ExecuteQuery() and\n> >> EvaluateParams() into a new transformExecuteStmt() that is called from\n> >> transformStmt().\n> >\n> > Uhmm ... no actual patch attached?\n>\n> Oops, here it is.\n>\n\nI checked this patch, and I think so it's correct and wanted. It introduce\ntransform stage for EXECUTE command, and move there the argument\ntransformation.\n\nThis has sensible change - the code is much more correct now.\n\nThe patching, compilation was without any problems, make check-world too.\n\nI was little bit confused about regress tests - the patch did some code\nrefactoring and I expect so main target is same behave before and after\npatching. But the regress tests shows new feature that is just side effect\n(nice) of patch. More, the example is little bit strange - nobody will use\nprepared statements and execution in SQL function. It should be better\ncommented.\n\nI'll mark this patch as ready for commiters.\n\nRegards\n\nPavel\n\n\n> --\n> Peter Eisentraut http://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\npo 4. 11. 2019 v 8:53 odesílatel Peter Eisentraut <peter.eisentraut@2ndquadrant.com> napsal:On 2019-11-02 16:00, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n>> This patch moves the parse analysis component of ExecuteQuery() and\n>> EvaluateParams() into a new transformExecuteStmt() that is called from\n>> transformStmt().\n> \n> Uhmm ... no actual patch attached?\n\nOops, here it is.I checked this patch, and I think so it's correct and wanted. It introduce transform stage for EXECUTE command, and move there the argument transformation.This has sensible change - the code is much more correct now. The patching, compilation was without any problems, make check-world too. I was little bit confused about regress tests - the patch did some code refactoring and I expect so main target is same behave before and after patching. But the regress tests shows new feature that is just side effect (nice) of patch. More, the example is little bit strange - nobody will use prepared statements and execution in SQL function. It should be better commented.I'll mark this patch as ready for commiters.RegardsPavel\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Fri, 8 Nov 2019 08:38:57 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"On 2019-11-08 08:13, Pavel Stehule wrote:\n> I dug into repository and found that transformExecuteStmt existed at\n> the time of implementing PREPARE-EXECUTE statements(28e82066a1) and\n> removed by the commit b9527e9840 which is related to\n> plan-invalidation.\n> \n> git show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\n> > In service of this, rearrange utility-statement processing so\n> that parse\n> > analysis does not assume table schemas can't change before\n> execution for\n> > utility statements (necessary because we don't attempt to\n> re-acquire locks\n> > for utility statements when reusing a stored plan). This\n> requires some\n> \n> Isn't this related to the current structure?\n> \n> I think so it should be ok, because the transformation is still in same \n> statement - if I understand well.\n> \n> So visibility of system catalogue or access to plan cache should not be \n> changed.\n\nI think what that patch was addressing is, if you use a protocol-level \nprepare+execute with commands like CREATE INDEX, CREATE VIEW, or COPY \nand you change the table schema between the prepare and execute, things \nwould break, for the reasons explained in the commit message. So any \nparse analysis in utility statements that accesses table schemas needs \nto be done in the execute phase, not in the prepare phase, as one might \nthink.\n\nParse analysis of EXECUTE does not access any tables, so if I understood \nthis correctly, this concern doesn't apply here.\n\nInterestingly, the above commit also removed the prepare-time \ntransformation of ExplainStmt, but it was later put back and now has the \ncomment \"We used to postpone that until execution, but it's really \nnecessary to do it during the normal parse analysis phase to ensure that \nside effects of parser hooks happen at the expected time.\" So there \nappears to be a generally uneasy situation still about how to do this \ncorrectly.\n\nPerhaps something could be done about the issue \"because we don't \nattempt to re-acquire locks for utility statements when reusing a stored \nplan\"?\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 8 Nov 2019 08:54:11 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"pá 8. 11. 2019 v 8:54 odesílatel Peter Eisentraut <\npeter.eisentraut@2ndquadrant.com> napsal:\n\n> On 2019-11-08 08:13, Pavel Stehule wrote:\n> > I dug into repository and found that transformExecuteStmt existed at\n> > the time of implementing PREPARE-EXECUTE statements(28e82066a1) and\n> > removed by the commit b9527e9840 which is related to\n> > plan-invalidation.\n> >\n> > git show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\n> > > In service of this, rearrange utility-statement processing so\n> > that parse\n> > > analysis does not assume table schemas can't change before\n> > execution for\n> > > utility statements (necessary because we don't attempt to\n> > re-acquire locks\n> > > for utility statements when reusing a stored plan). This\n> > requires some\n> >\n> > Isn't this related to the current structure?\n> >\n> > I think so it should be ok, because the transformation is still in same\n> > statement - if I understand well.\n> >\n> > So visibility of system catalogue or access to plan cache should not be\n> > changed.\n>\n> I think what that patch was addressing is, if you use a protocol-level\n> prepare+execute with commands like CREATE INDEX, CREATE VIEW, or COPY\n> and you change the table schema between the prepare and execute, things\n> would break, for the reasons explained in the commit message. So any\n> parse analysis in utility statements that accesses table schemas needs\n> to be done in the execute phase, not in the prepare phase, as one might\n> think.\n>\n> Parse analysis of EXECUTE does not access any tables, so if I understood\n> this correctly, this concern doesn't apply here.\n>\n\nit should not be true - the subquery can be a expression.\n\nMinimally on SQL level is not possible do prepare on execute. So execute\nshould be evaluate as one step.\n\n\n\n> Interestingly, the above commit also removed the prepare-time\n> transformation of ExplainStmt, but it was later put back and now has the\n> comment \"We used to postpone that until execution, but it's really\n> necessary to do it during the normal parse analysis phase to ensure that\n> side effects of parser hooks happen at the expected time.\" So there\n> appears to be a generally uneasy situation still about how to do this\n> correctly.\n>\n> Perhaps something could be done about the issue \"because we don't\n> attempt to re-acquire locks for utility statements when reusing a stored\n> plan\"?\n>\n> --\n> Peter Eisentraut http://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\npá 8. 11. 2019 v 8:54 odesílatel Peter Eisentraut <peter.eisentraut@2ndquadrant.com> napsal:On 2019-11-08 08:13, Pavel Stehule wrote:\n> I dug into repository and found that transformExecuteStmt existed at\n> the time of implementing PREPARE-EXECUTE statements(28e82066a1) and\n> removed by the commit b9527e9840 which is related to\n> plan-invalidation.\n> \n> git show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\n> > In service of this, rearrange utility-statement processing so\n> that parse\n> > analysis does not assume table schemas can't change before\n> execution for\n> > utility statements (necessary because we don't attempt to\n> re-acquire locks\n> > for utility statements when reusing a stored plan). This\n> requires some\n> \n> Isn't this related to the current structure?\n> \n> I think so it should be ok, because the transformation is still in same \n> statement - if I understand well.\n> \n> So visibility of system catalogue or access to plan cache should not be \n> changed.\n\nI think what that patch was addressing is, if you use a protocol-level \nprepare+execute with commands like CREATE INDEX, CREATE VIEW, or COPY \nand you change the table schema between the prepare and execute, things \nwould break, for the reasons explained in the commit message. So any \nparse analysis in utility statements that accesses table schemas needs \nto be done in the execute phase, not in the prepare phase, as one might \nthink.\n\nParse analysis of EXECUTE does not access any tables, so if I understood \nthis correctly, this concern doesn't apply here.it should not be true - the subquery can be a expression.Minimally on SQL level is not possible do prepare on execute. So execute should be evaluate as one step.\n\nInterestingly, the above commit also removed the prepare-time \ntransformation of ExplainStmt, but it was later put back and now has the \ncomment \"We used to postpone that until execution, but it's really \nnecessary to do it during the normal parse analysis phase to ensure that \nside effects of parser hooks happen at the expected time.\" So there \nappears to be a generally uneasy situation still about how to do this \ncorrectly.\n\nPerhaps something could be done about the issue \"because we don't \nattempt to re-acquire locks for utility statements when reusing a stored \nplan\"?\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Fri, 8 Nov 2019 09:03:18 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"On 2019-11-08 09:03, Pavel Stehule wrote:\n> Parse analysis of EXECUTE does not access any tables, so if I\n> understood\n> this correctly, this concern doesn't apply here.\n> \n> \n> it should not be true - the subquery can be a expression.\n\nArguments of EXECUTE cannot be subqueries.\n\n> Minimally on SQL level is not possible do prepare on execute. So execute \n> should be evaluate as one step.\n\nWell, that's kind of the question that is being discussed in this thread.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Fri, 8 Nov 2019 13:34:40 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"pá 8. 11. 2019 v 13:34 odesílatel Peter Eisentraut <\npeter.eisentraut@2ndquadrant.com> napsal:\n\n> On 2019-11-08 09:03, Pavel Stehule wrote:\n> > Parse analysis of EXECUTE does not access any tables, so if I\n> > understood\n> > this correctly, this concern doesn't apply here.\n> >\n> >\n> > it should not be true - the subquery can be a expression.\n>\n> Arguments of EXECUTE cannot be subqueries.\n>\nok\n\n>\n> > Minimally on SQL level is not possible do prepare on execute. So execute\n> > should be evaluate as one step.\n>\n> Well, that's kind of the question that is being discussed in this thread.\n>\n\nI say it not cleanly - I think so this change should be safe, because\nparsing, transforming, and execution must be in one statement.\n\nRegards\n\nPavel\n\n>\n> --\n> Peter Eisentraut http://www.2ndQuadrant.com/\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n>\n\npá 8. 11. 2019 v 13:34 odesílatel Peter Eisentraut <peter.eisentraut@2ndquadrant.com> napsal:On 2019-11-08 09:03, Pavel Stehule wrote:\n> Parse analysis of EXECUTE does not access any tables, so if I\n> understood\n> this correctly, this concern doesn't apply here.\n> \n> \n> it should not be true - the subquery can be a expression.\n\nArguments of EXECUTE cannot be subqueries.ok \n\n> Minimally on SQL level is not possible do prepare on execute. So execute \n> should be evaluate as one step.\n\nWell, that's kind of the question that is being discussed in this thread.I say it not cleanly - I think so this change should be safe, because parsing, transforming, and execution must be in one statement.RegardsPavel\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Fri, 8 Nov 2019 16:20:46 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\n> On 2019-11-08 09:03, Pavel Stehule wrote:\n>> Minimally on SQL level is not possible do prepare on execute. So execute \n>> should be evaluate as one step.\n\n> Well, that's kind of the question that is being discussed in this thread.\n\nYeah. Having now taken a quick look at this patch, it makes me pretty\nqueasy. In particular, it doesn't appear to add any support for\ninvalidation of cached EXECUTE commands when their parameter expressions\nchange. You dismissed that as irrelevant because no table schemas would\nbe involved, but there's also the possibility of replacements of user\ndefined functions. I'm not sure how easy it is to create a situation\nwhere an EXECUTE statement is in plancache, but it's probably possible\n(maybe using some other PL than plpgsql). In that case, we really would\nneed the EXECUTE's transformed expressions to get invalidated if the\nuser drops or replaces a function they use.\n\nIn view of the ALTER TABLE bugs I'm struggling with over in [1], I feel\nlike this patch is probably going in the wrong direction. We should\ngenerally be striving to do all transformation of utility commands as\nlate as possible. As long as a plancached utility statement contains\nnothing beyond raw-parser output, it never needs invalidation.\n\nYou pointed to an old comment of mine about EXPLAIN that seems to argue\nin the other direction, but digging in the commit log, I see that it\ncame from commit 08f8d478, whose log entry is perhaps more informative\nthan the comment:\n\n Do parse analysis of an EXPLAIN's contained statement during the normal\n parse analysis phase, rather than at execution time. This makes parameter\n handling work the same as it does in ordinary plannable queries, and in\n particular fixes the incompatibility that Pavel pointed out with plpgsql's\n new handling of variable references. plancache.c gets a little bit\n grottier, but the alternatives seem worse.\n\nSo what this really is all about is still the same old issue of how we\nhandle external parameter references in utility statements. Maybe we\nought to focus on a redesign addressing that specific problem, rather\nthan nibbling around the edges. It seems like the core of the issue\nis that we have mechanisms for PLs to capture parameter references\nduring parse analysis, and those hooks aren't managed in a way that\nlets them be invoked if we do parse analysis during utility statement\nexecution. But we *need* to be able to do that. ALTER TABLE already\ndoes do that, yet we need to postpone its analysis to even later than\nit's doing it now.\n\nAnother issue in all this is that for many utility statements, you\ndon't actually want injections of PL parameter references, for instance\nit'd make little sense to allow \"alter table ... add check (f1 > p1)\"\nif p1 is a local variable in the function doing the ALTER. It's\nprobably time to have some explicit recognition and management of such\ncases, rather than just dodging them by not invoking the hooks.\n\ntl;dr: I think that we need to embrace parse analysis during utility\nstatement execution as a fully supported thing, not a stepchild.\nTrying to make it go away is the wrong approach.\n\n\t\t\tregards, tom lane\n\n[1] https://www.postgresql.org/message-id/flat/10365.1558909428@sss.pgh.pa.us\n\n\n","msg_date":"Fri, 08 Nov 2019 11:21:58 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"After digesting the discussion, let's reshuffle this a bit.\n\nI have committed the change that adds the error location in one place. \nThat worked independently.\n\nAttached is a new patch that refactors things a bit to pass the \nParseState into functions such as PrepareQuery() and ExecuteQuery() \ninstead of passing the query string and query environment as a separate \narguments. We had already done that for most utility commands; there \nwere just some left that happened to be involved in the current thread's \ndiscussion anyway.\n\nThat's a nice cosmetic improvement in any case, but I think that it \nwould also help with the issue of passing parameters into some utility \ncommands later on. I will look into that some other time.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services","msg_date":"Fri, 29 Nov 2019 11:39:31 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"The following review has been posted through the commitfest application:\nmake installcheck-world: tested, passed\nImplements feature: not tested\nSpec compliant: not tested\nDocumentation: not tested\n\nThis patch replaced query string by parse state on few places. It increase code consistency.\n\nThe new status of this patch is: Ready for Committer\n","msg_date":"Thu, 02 Jan 2020 13:26:04 +0000","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"},{"msg_contents":"On 2020-01-02 14:26, Pavel Stehule wrote:\n> The following review has been posted through the commitfest application:\n> make installcheck-world: tested, passed\n> Implements feature: not tested\n> Spec compliant: not tested\n> Documentation: not tested\n> \n> This patch replaced query string by parse state on few places. It increase code consistency.\n> \n> The new status of this patch is: Ready for Committer\n\ncommitted, thanks\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n","msg_date":"Sat, 4 Jan 2020 13:43:52 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: Refactor parse analysis of EXECUTE command"}],"string":"[\n {\n \"msg_contents\": \"This patch moves the parse analysis component of ExecuteQuery() and\\nEvaluateParams() into a new transformExecuteStmt() that is called from\\ntransformStmt(). This makes EXECUTE behave more like other utility\\ncommands.\\n\\nEffects are that error messages can have position information (see \\nregression test case), and it allows using external parameters in the \\narguments of the EXECUTE command.\\n\\nI had previously inquired about this in [0] and some vague concerns were \\nraised. I haven't dug very deep on this, but I figure with an actual \\npatch it might be easier to review and figure out if there are any problems.\\n\\n\\n[0]: \\nhttps://www.postgresql.org/message-id/flat/ed2767e5-c506-048d-8ddf-280ecbc9e1b7%402ndquadrant.com\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 08:07:16 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> This patch moves the parse analysis component of ExecuteQuery() and\\n> EvaluateParams() into a new transformExecuteStmt() that is called from\\n> transformStmt().\\n\\nUhmm ... no actual patch attached?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 02 Nov 2019 11:00:30 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"On 2019-11-02 16:00, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n>> This patch moves the parse analysis component of ExecuteQuery() and\\n>> EvaluateParams() into a new transformExecuteStmt() that is called from\\n>> transformStmt().\\n> \\n> Uhmm ... no actual patch attached?\\n\\nOops, here it is.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Mon, 4 Nov 2019 08:53:18 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"Hello.\\n\\nAt Mon, 4 Nov 2019 08:53:18 +0100, Peter Eisentraut <peter.eisentraut@2ndquadrant.com> wrote in \\n> On 2019-11-02 16:00, Tom Lane wrote:\\n> > Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> >> This patch moves the parse analysis component of ExecuteQuery() and\\n> >> EvaluateParams() into a new transformExecuteStmt() that is called from\\n> >> transformStmt().\\n> > Uhmm ... no actual patch attached?\\n> \\n> Oops, here it is.\\n\\nThe patch just moves the first half of EvaluateParams that is\\nirrelevant to executor state to before portal parameters are set. I\\nlooked with a suspect that extended protocol or SPI are affected but\\nAFAICS it doesn't seem to.\\n\\nI dug into repository and found that transformExecuteStmt existed at\\nthe time of implementing PREPARE-EXECUTE statements(28e82066a1) and\\nremoved by the commit b9527e9840 which is related to\\nplan-invalidation.\\n\\ngit show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\\n> In service of this, rearrange utility-statement processing so that parse\\n> analysis does not assume table schemas can't change before execution for\\n> utility statements (necessary because we don't attempt to re-acquire locks\\n> for utility statements when reusing a stored plan). This requires some\\n\\nIsn't this related to the current structure?\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Tue, 05 Nov 2019 19:27:48 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"út 5. 11. 2019 v 11:28 odesílatel Kyotaro Horiguchi <horikyota.ntt@gmail.com>\\nnapsal:\\n\\n> Hello.\\n>\\n> At Mon, 4 Nov 2019 08:53:18 +0100, Peter Eisentraut <\\n> peter.eisentraut@2ndquadrant.com> wrote in\\n> > On 2019-11-02 16:00, Tom Lane wrote:\\n> > > Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> > >> This patch moves the parse analysis component of ExecuteQuery() and\\n> > >> EvaluateParams() into a new transformExecuteStmt() that is called from\\n> > >> transformStmt().\\n> > > Uhmm ... no actual patch attached?\\n> >\\n> > Oops, here it is.\\n>\\n> The patch just moves the first half of EvaluateParams that is\\n> irrelevant to executor state to before portal parameters are set. I\\n> looked with a suspect that extended protocol or SPI are affected but\\n> AFAICS it doesn't seem to.\\n>\\n> I dug into repository and found that transformExecuteStmt existed at\\n> the time of implementing PREPARE-EXECUTE statements(28e82066a1) and\\n> removed by the commit b9527e9840 which is related to\\n> plan-invalidation.\\n>\\n> git show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\\n> > In service of this, rearrange utility-statement processing so that parse\\n> > analysis does not assume table schemas can't change before execution for\\n> > utility statements (necessary because we don't attempt to re-acquire\\n> locks\\n> > for utility statements when reusing a stored plan). This requires some\\n>\\n> Isn't this related to the current structure?\\n>\\n\\nI think so it should be ok, because the transformation is still in same\\nstatement - if I understand well.\\n\\nSo visibility of system catalogue or access to plan cache should not be\\nchanged.\\n\\nRegards\\n\\nPavel\\n\\n\\n> regards.\\n>\\n> --\\n> Kyotaro Horiguchi\\n> NTT Open Source Software Center\\n>\\n>\\n>\\n\\nút 5. 11. 2019 v 11:28 odesílatel Kyotaro Horiguchi <horikyota.ntt@gmail.com> napsal:Hello.\\n\\nAt Mon, 4 Nov 2019 08:53:18 +0100, Peter Eisentraut <peter.eisentraut@2ndquadrant.com> wrote in \\n> On 2019-11-02 16:00, Tom Lane wrote:\\n> > Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> >> This patch moves the parse analysis component of ExecuteQuery() and\\n> >> EvaluateParams() into a new transformExecuteStmt() that is called from\\n> >> transformStmt().\\n> > Uhmm ... no actual patch attached?\\n> \\n> Oops, here it is.\\n\\nThe patch just moves the first half of EvaluateParams that is\\nirrelevant to executor state to before portal parameters are set. I\\nlooked with a suspect that extended protocol or SPI are affected but\\nAFAICS it doesn't seem to.\\n\\nI dug into repository and found that transformExecuteStmt existed at\\nthe time of implementing PREPARE-EXECUTE statements(28e82066a1) and\\nremoved by the commit b9527e9840 which is related to\\nplan-invalidation.\\n\\ngit show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\\n> In service of this, rearrange utility-statement processing so that parse\\n> analysis does not assume table schemas can't change before execution for\\n> utility statements (necessary because we don't attempt to re-acquire locks\\n> for utility statements when reusing a stored plan). This requires some\\n\\nIsn't this related to the current structure?I think so it should be ok, because the transformation is still in same statement - if I understand well. So visibility of system catalogue or access to plan cache should not be changed.RegardsPavel\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\",\n \"msg_date\": \"Fri, 8 Nov 2019 08:13:55 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"po 4. 11. 2019 v 8:53 odesílatel Peter Eisentraut <\\npeter.eisentraut@2ndquadrant.com> napsal:\\n\\n> On 2019-11-02 16:00, Tom Lane wrote:\\n> > Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> >> This patch moves the parse analysis component of ExecuteQuery() and\\n> >> EvaluateParams() into a new transformExecuteStmt() that is called from\\n> >> transformStmt().\\n> >\\n> > Uhmm ... no actual patch attached?\\n>\\n> Oops, here it is.\\n>\\n\\nI checked this patch, and I think so it's correct and wanted. It introduce\\ntransform stage for EXECUTE command, and move there the argument\\ntransformation.\\n\\nThis has sensible change - the code is much more correct now.\\n\\nThe patching, compilation was without any problems, make check-world too.\\n\\nI was little bit confused about regress tests - the patch did some code\\nrefactoring and I expect so main target is same behave before and after\\npatching. But the regress tests shows new feature that is just side effect\\n(nice) of patch. More, the example is little bit strange - nobody will use\\nprepared statements and execution in SQL function. It should be better\\ncommented.\\n\\nI'll mark this patch as ready for commiters.\\n\\nRegards\\n\\nPavel\\n\\n\\n> --\\n> Peter Eisentraut http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\npo 4. 11. 2019 v 8:53 odesílatel Peter Eisentraut <peter.eisentraut@2ndquadrant.com> napsal:On 2019-11-02 16:00, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n>> This patch moves the parse analysis component of ExecuteQuery() and\\n>> EvaluateParams() into a new transformExecuteStmt() that is called from\\n>> transformStmt().\\n> \\n> Uhmm ... no actual patch attached?\\n\\nOops, here it is.I checked this patch, and I think so it's correct and wanted. It introduce transform stage for EXECUTE command, and move there the argument transformation.This has sensible change - the code is much more correct now. The patching, compilation was without any problems, make check-world too. I was little bit confused about regress tests - the patch did some code refactoring and I expect so main target is same behave before and after patching. But the regress tests shows new feature that is just side effect (nice) of patch. More, the example is little bit strange - nobody will use prepared statements and execution in SQL function. It should be better commented.I'll mark this patch as ready for commiters.RegardsPavel\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 8 Nov 2019 08:38:57 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"On 2019-11-08 08:13, Pavel Stehule wrote:\\n> I dug into repository and found that transformExecuteStmt existed at\\n> the time of implementing PREPARE-EXECUTE statements(28e82066a1) and\\n> removed by the commit b9527e9840 which is related to\\n> plan-invalidation.\\n> \\n> git show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\\n> > In service of this, rearrange utility-statement processing so\\n> that parse\\n> > analysis does not assume table schemas can't change before\\n> execution for\\n> > utility statements (necessary because we don't attempt to\\n> re-acquire locks\\n> > for utility statements when reusing a stored plan). This\\n> requires some\\n> \\n> Isn't this related to the current structure?\\n> \\n> I think so it should be ok, because the transformation is still in same \\n> statement - if I understand well.\\n> \\n> So visibility of system catalogue or access to plan cache should not be \\n> changed.\\n\\nI think what that patch was addressing is, if you use a protocol-level \\nprepare+execute with commands like CREATE INDEX, CREATE VIEW, or COPY \\nand you change the table schema between the prepare and execute, things \\nwould break, for the reasons explained in the commit message. So any \\nparse analysis in utility statements that accesses table schemas needs \\nto be done in the execute phase, not in the prepare phase, as one might \\nthink.\\n\\nParse analysis of EXECUTE does not access any tables, so if I understood \\nthis correctly, this concern doesn't apply here.\\n\\nInterestingly, the above commit also removed the prepare-time \\ntransformation of ExplainStmt, but it was later put back and now has the \\ncomment \\\"We used to postpone that until execution, but it's really \\nnecessary to do it during the normal parse analysis phase to ensure that \\nside effects of parser hooks happen at the expected time.\\\" So there \\nappears to be a generally uneasy situation still about how to do this \\ncorrectly.\\n\\nPerhaps something could be done about the issue \\\"because we don't \\nattempt to re-acquire locks for utility statements when reusing a stored \\nplan\\\"?\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 08:54:11 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"pá 8. 11. 2019 v 8:54 odesílatel Peter Eisentraut <\\npeter.eisentraut@2ndquadrant.com> napsal:\\n\\n> On 2019-11-08 08:13, Pavel Stehule wrote:\\n> > I dug into repository and found that transformExecuteStmt existed at\\n> > the time of implementing PREPARE-EXECUTE statements(28e82066a1) and\\n> > removed by the commit b9527e9840 which is related to\\n> > plan-invalidation.\\n> >\\n> > git show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\\n> > > In service of this, rearrange utility-statement processing so\\n> > that parse\\n> > > analysis does not assume table schemas can't change before\\n> > execution for\\n> > > utility statements (necessary because we don't attempt to\\n> > re-acquire locks\\n> > > for utility statements when reusing a stored plan). This\\n> > requires some\\n> >\\n> > Isn't this related to the current structure?\\n> >\\n> > I think so it should be ok, because the transformation is still in same\\n> > statement - if I understand well.\\n> >\\n> > So visibility of system catalogue or access to plan cache should not be\\n> > changed.\\n>\\n> I think what that patch was addressing is, if you use a protocol-level\\n> prepare+execute with commands like CREATE INDEX, CREATE VIEW, or COPY\\n> and you change the table schema between the prepare and execute, things\\n> would break, for the reasons explained in the commit message. So any\\n> parse analysis in utility statements that accesses table schemas needs\\n> to be done in the execute phase, not in the prepare phase, as one might\\n> think.\\n>\\n> Parse analysis of EXECUTE does not access any tables, so if I understood\\n> this correctly, this concern doesn't apply here.\\n>\\n\\nit should not be true - the subquery can be a expression.\\n\\nMinimally on SQL level is not possible do prepare on execute. So execute\\nshould be evaluate as one step.\\n\\n\\n\\n> Interestingly, the above commit also removed the prepare-time\\n> transformation of ExplainStmt, but it was later put back and now has the\\n> comment \\\"We used to postpone that until execution, but it's really\\n> necessary to do it during the normal parse analysis phase to ensure that\\n> side effects of parser hooks happen at the expected time.\\\" So there\\n> appears to be a generally uneasy situation still about how to do this\\n> correctly.\\n>\\n> Perhaps something could be done about the issue \\\"because we don't\\n> attempt to re-acquire locks for utility statements when reusing a stored\\n> plan\\\"?\\n>\\n> --\\n> Peter Eisentraut http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\npá 8. 11. 2019 v 8:54 odesílatel Peter Eisentraut <peter.eisentraut@2ndquadrant.com> napsal:On 2019-11-08 08:13, Pavel Stehule wrote:\\n> I dug into repository and found that transformExecuteStmt existed at\\n> the time of implementing PREPARE-EXECUTE statements(28e82066a1) and\\n> removed by the commit b9527e9840 which is related to\\n> plan-invalidation.\\n> \\n> git show -s --format=%B b9527e984092e838790b543b014c0c2720ea4f11\\n> > In service of this, rearrange utility-statement processing so\\n> that parse\\n> > analysis does not assume table schemas can't change before\\n> execution for\\n> > utility statements (necessary because we don't attempt to\\n> re-acquire locks\\n> > for utility statements when reusing a stored plan). This\\n> requires some\\n> \\n> Isn't this related to the current structure?\\n> \\n> I think so it should be ok, because the transformation is still in same \\n> statement - if I understand well.\\n> \\n> So visibility of system catalogue or access to plan cache should not be \\n> changed.\\n\\nI think what that patch was addressing is, if you use a protocol-level \\nprepare+execute with commands like CREATE INDEX, CREATE VIEW, or COPY \\nand you change the table schema between the prepare and execute, things \\nwould break, for the reasons explained in the commit message. So any \\nparse analysis in utility statements that accesses table schemas needs \\nto be done in the execute phase, not in the prepare phase, as one might \\nthink.\\n\\nParse analysis of EXECUTE does not access any tables, so if I understood \\nthis correctly, this concern doesn't apply here.it should not be true - the subquery can be a expression.Minimally on SQL level is not possible do prepare on execute. So execute should be evaluate as one step.\\n\\nInterestingly, the above commit also removed the prepare-time \\ntransformation of ExplainStmt, but it was later put back and now has the \\ncomment \\\"We used to postpone that until execution, but it's really \\nnecessary to do it during the normal parse analysis phase to ensure that \\nside effects of parser hooks happen at the expected time.\\\" So there \\nappears to be a generally uneasy situation still about how to do this \\ncorrectly.\\n\\nPerhaps something could be done about the issue \\\"because we don't \\nattempt to re-acquire locks for utility statements when reusing a stored \\nplan\\\"?\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 8 Nov 2019 09:03:18 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"On 2019-11-08 09:03, Pavel Stehule wrote:\\n> Parse analysis of EXECUTE does not access any tables, so if I\\n> understood\\n> this correctly, this concern doesn't apply here.\\n> \\n> \\n> it should not be true - the subquery can be a expression.\\n\\nArguments of EXECUTE cannot be subqueries.\\n\\n> Minimally on SQL level is not possible do prepare on execute. So execute \\n> should be evaluate as one step.\\n\\nWell, that's kind of the question that is being discussed in this thread.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 13:34:40 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"pá 8. 11. 2019 v 13:34 odesílatel Peter Eisentraut <\\npeter.eisentraut@2ndquadrant.com> napsal:\\n\\n> On 2019-11-08 09:03, Pavel Stehule wrote:\\n> > Parse analysis of EXECUTE does not access any tables, so if I\\n> > understood\\n> > this correctly, this concern doesn't apply here.\\n> >\\n> >\\n> > it should not be true - the subquery can be a expression.\\n>\\n> Arguments of EXECUTE cannot be subqueries.\\n>\\nok\\n\\n>\\n> > Minimally on SQL level is not possible do prepare on execute. So execute\\n> > should be evaluate as one step.\\n>\\n> Well, that's kind of the question that is being discussed in this thread.\\n>\\n\\nI say it not cleanly - I think so this change should be safe, because\\nparsing, transforming, and execution must be in one statement.\\n\\nRegards\\n\\nPavel\\n\\n>\\n> --\\n> Peter Eisentraut http://www.2ndQuadrant.com/\\n> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n>\\n\\npá 8. 11. 2019 v 13:34 odesílatel Peter Eisentraut <peter.eisentraut@2ndquadrant.com> napsal:On 2019-11-08 09:03, Pavel Stehule wrote:\\n> Parse analysis of EXECUTE does not access any tables, so if I\\n> understood\\n> this correctly, this concern doesn't apply here.\\n> \\n> \\n> it should not be true - the subquery can be a expression.\\n\\nArguments of EXECUTE cannot be subqueries.ok \\n\\n> Minimally on SQL level is not possible do prepare on execute. So execute \\n> should be evaluate as one step.\\n\\nWell, that's kind of the question that is being discussed in this thread.I say it not cleanly - I think so this change should be safe, because parsing, transforming, and execution must be in one statement.RegardsPavel\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 8 Nov 2019 16:20:46 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:\\n> On 2019-11-08 09:03, Pavel Stehule wrote:\\n>> Minimally on SQL level is not possible do prepare on execute. So execute \\n>> should be evaluate as one step.\\n\\n> Well, that's kind of the question that is being discussed in this thread.\\n\\nYeah. Having now taken a quick look at this patch, it makes me pretty\\nqueasy. In particular, it doesn't appear to add any support for\\ninvalidation of cached EXECUTE commands when their parameter expressions\\nchange. You dismissed that as irrelevant because no table schemas would\\nbe involved, but there's also the possibility of replacements of user\\ndefined functions. I'm not sure how easy it is to create a situation\\nwhere an EXECUTE statement is in plancache, but it's probably possible\\n(maybe using some other PL than plpgsql). In that case, we really would\\nneed the EXECUTE's transformed expressions to get invalidated if the\\nuser drops or replaces a function they use.\\n\\nIn view of the ALTER TABLE bugs I'm struggling with over in [1], I feel\\nlike this patch is probably going in the wrong direction. We should\\ngenerally be striving to do all transformation of utility commands as\\nlate as possible. As long as a plancached utility statement contains\\nnothing beyond raw-parser output, it never needs invalidation.\\n\\nYou pointed to an old comment of mine about EXPLAIN that seems to argue\\nin the other direction, but digging in the commit log, I see that it\\ncame from commit 08f8d478, whose log entry is perhaps more informative\\nthan the comment:\\n\\n Do parse analysis of an EXPLAIN's contained statement during the normal\\n parse analysis phase, rather than at execution time. This makes parameter\\n handling work the same as it does in ordinary plannable queries, and in\\n particular fixes the incompatibility that Pavel pointed out with plpgsql's\\n new handling of variable references. plancache.c gets a little bit\\n grottier, but the alternatives seem worse.\\n\\nSo what this really is all about is still the same old issue of how we\\nhandle external parameter references in utility statements. Maybe we\\nought to focus on a redesign addressing that specific problem, rather\\nthan nibbling around the edges. It seems like the core of the issue\\nis that we have mechanisms for PLs to capture parameter references\\nduring parse analysis, and those hooks aren't managed in a way that\\nlets them be invoked if we do parse analysis during utility statement\\nexecution. But we *need* to be able to do that. ALTER TABLE already\\ndoes do that, yet we need to postpone its analysis to even later than\\nit's doing it now.\\n\\nAnother issue in all this is that for many utility statements, you\\ndon't actually want injections of PL parameter references, for instance\\nit'd make little sense to allow \\\"alter table ... add check (f1 > p1)\\\"\\nif p1 is a local variable in the function doing the ALTER. It's\\nprobably time to have some explicit recognition and management of such\\ncases, rather than just dodging them by not invoking the hooks.\\n\\ntl;dr: I think that we need to embrace parse analysis during utility\\nstatement execution as a fully supported thing, not a stepchild.\\nTrying to make it go away is the wrong approach.\\n\\n\\t\\t\\tregards, tom lane\\n\\n[1] https://www.postgresql.org/message-id/flat/10365.1558909428@sss.pgh.pa.us\\n\\n\\n\",\n \"msg_date\": \"Fri, 08 Nov 2019 11:21:58 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"After digesting the discussion, let's reshuffle this a bit.\\n\\nI have committed the change that adds the error location in one place. \\nThat worked independently.\\n\\nAttached is a new patch that refactors things a bit to pass the \\nParseState into functions such as PrepareQuery() and ExecuteQuery() \\ninstead of passing the query string and query environment as a separate \\narguments. We had already done that for most utility commands; there \\nwere just some left that happened to be involved in the current thread's \\ndiscussion anyway.\\n\\nThat's a nice cosmetic improvement in any case, but I think that it \\nwould also help with the issue of passing parameters into some utility \\ncommands later on. I will look into that some other time.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\",\n \"msg_date\": \"Fri, 29 Nov 2019 11:39:31 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"The following review has been posted through the commitfest application:\\nmake installcheck-world: tested, passed\\nImplements feature: not tested\\nSpec compliant: not tested\\nDocumentation: not tested\\n\\nThis patch replaced query string by parse state on few places. It increase code consistency.\\n\\nThe new status of this patch is: Ready for Committer\\n\",\n \"msg_date\": \"Thu, 02 Jan 2020 13:26:04 +0000\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n },\n {\n \"msg_contents\": \"On 2020-01-02 14:26, Pavel Stehule wrote:\\n> The following review has been posted through the commitfest application:\\n> make installcheck-world: tested, passed\\n> Implements feature: not tested\\n> Spec compliant: not tested\\n> Documentation: not tested\\n> \\n> This patch replaced query string by parse state on few places. It increase code consistency.\\n> \\n> The new status of this patch is: Ready for Committer\\n\\ncommitted, thanks\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\",\n \"msg_date\": \"Sat, 4 Jan 2020 13:43:52 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Refactor parse analysis of EXECUTE command\"\n }\n]","isConversational":false}}},{"rowIdx":49391,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi.sorry for my English.I want to once again open the topic of 64 bit transaction id. I did not manage to find in the archive of the option that I want to discuss, so I write. If I searched poorly, then please forgive me.The idea is not very original and probably has already been considered, again I repeat - I did not find it. Therefore, please do not scold me severely.In discussions of 64-bit transaction id, I did not find mention of an algorithm for storing them, as it was done, for example, in MS SQL Server.What if instead of 2 fields (xmin and xmax) with a total length of 64 bits - use 1 field (let's call it xid) with a length of 64 bits in tuple header? In this field store the xid of the transaction that created the version. In this case, the new transaction in order to understand whether the read version is suitable for it or not, will have to read the next version as well. Those. The downside of such decision is of course an increase in I / O. Transactions will have to read the +1 version. On the plus side, the title of the tuple remains the same length. regards, Eremin Pavel.","msg_date":"Fri, 01 Nov 2019 12:05:12 +0300","msg_from":"=?utf-8?B?0J/QsNCy0LXQuyDQldGA0ZHQvNC40L0=?= <shnoor111gmail@yandex.ru>","msg_from_op":true,"msg_subject":"64 bit transaction id"},{"msg_contents":"Hi\n\npá 1. 11. 2019 v 10:11 odesílatel Павел Ерёмин <shnoor111gmail@yandex.ru>\nnapsal:\n\n> Hi.\n> sorry for my English.\n> I want to once again open the topic of 64 bit transaction id. I did not\n> manage to find in the archive of the option that I want to discuss, so I\n> write. If I searched poorly, then please forgive me.\n> The idea is not very original and probably has already been considered,\n> again I repeat - I did not find it. Therefore, please do not scold me\n> severely.\n> In discussions of 64-bit transaction id, I did not find mention of an\n> algorithm for storing them, as it was done, for example, in MS SQL Server.\n> What if instead of 2 fields (xmin and xmax) with a total length of 64 bits\n> - use 1 field (let's call it xid) with a length of 64 bits in tuple header?\n> In this field store the xid of the transaction that created the version. In\n> this case, the new transaction in order to understand whether the read\n> version is suitable for it or not, will have to read the next version as\n> well. Those. The downside of such decision is of course an increase in I /\n> O. Transactions will have to read the +1 version. On the plus side, the\n> title of the tuple remains the same length.\n>\n\nis 32 bit tid really problem? Why you need to know state of last 2^31\ntransactions? Is not problem in too low usage (or maybe too high overhead)\nof VACUUM FREEZE.\n\nI am not sure if increasing this range can has much more fatal problems\n(maybe later)\n\nPavel\n\n\n\n>\n> regards, Eremin Pavel.\n>\n\nHipá 1. 11. 2019 v 10:11 odesílatel Павел Ерёмин <shnoor111gmail@yandex.ru> napsal:Hi.sorry for my English.I want to once again open the topic of 64 bit transaction id. I did not manage to find in the archive of the option that I want to discuss, so I write. If I searched poorly, then please forgive me.The idea is not very original and probably has already been considered, again I repeat - I did not find it. Therefore, please do not scold me severely.In discussions of 64-bit transaction id, I did not find mention of an algorithm for storing them, as it was done, for example, in MS SQL Server.What if instead of 2 fields (xmin and xmax) with a total length of 64 bits - use 1 field (let's call it xid) with a length of 64 bits in tuple header? In this field store the xid of the transaction that created the version. In this case, the new transaction in order to understand whether the read version is suitable for it or not, will have to read the next version as well. Those. The downside of such decision is of course an increase in I / O. Transactions will have to read the +1 version. On the plus side, the title of the tuple remains the same length.is 32 bit tid really problem? Why you need to know state of last 2^31 transactions? Is not problem in too low usage (or maybe too high overhead) of VACUUM FREEZE.I am not sure if increasing this range can has much more fatal problems (maybe later)Pavel regards, Eremin Pavel.","msg_date":"Fri, 1 Nov 2019 10:25:17 +0100","msg_from":"Pavel Stehule <pavel.stehule@gmail.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Fri, Nov 01, 2019 at 10:25:17AM +0100, Pavel Stehule wrote:\n>Hi\n>\n>pá 1. 11. 2019 v 10:11 odesílatel Павел Ерёмин <shnoor111gmail@yandex.ru>\n>napsal:\n>\n>> Hi.\n>> sorry for my English.\n>> I want to once again open the topic of 64 bit transaction id. I did not\n>> manage to find in the archive of the option that I want to discuss, so I\n>> write. If I searched poorly, then please forgive me.\n>> The idea is not very original and probably has already been considered,\n>> again I repeat - I did not find it. Therefore, please do not scold me\n>> severely.\n>> In discussions of 64-bit transaction id, I did not find mention of an\n>> algorithm for storing them, as it was done, for example, in MS SQL Server.\n>> What if instead of 2 fields (xmin and xmax) with a total length of 64 bits\n>> - use 1 field (let's call it xid) with a length of 64 bits in tuple header?\n>> In this field store the xid of the transaction that created the version. In\n>> this case, the new transaction in order to understand whether the read\n>> version is suitable for it or not, will have to read the next version as\n>> well. Those. The downside of such decision is of course an increase in I /\n>> O. Transactions will have to read the +1 version. On the plus side, the\n>> title of the tuple remains the same length.\n>>\n>\n>is 32 bit tid really problem? Why you need to know state of last 2^31\n>transactions? Is not problem in too low usage (or maybe too high overhead)\n>of VACUUM FREEZE.\n>\n\nIt certainly can be an issue for large and busy systems, that may need\nanti-wraparoud vacuum every couple of days. If that requires rewriting a\ncouple of TB of data, it's not particularly nice. That's why 64-bit XIDs\nwere discussed repeatedly in the past, and it's likely to get even more\npressing as the systems get larger.\n\n>I am not sure if increasing this range can has much more fatal problems\n>(maybe later)\n>\n\nWell, not fatal, but naive approaches can increase per-tuple overhead.\nAnd we already have plenty of that, hence there were proposals to use\npage epochs and so on.\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Fri, 1 Nov 2019 18:10:38 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Fri, Nov 01, 2019 at 12:05:12PM +0300, Павел Ерёмин wrote:\n> Hi.\n> sorry for my English.\n> I want to once again open the topic of 64 bit transaction id. I did not\n> manage to find in the archive of the option that I want to discuss, so I\n> write. If I searched poorly, then please forgive me.\n> The idea is not very original and probably has already been considered,\n> again I repeat - I did not find it. Therefore, please do not scold me\n> severely.\n> In discussions of 64-bit transaction id, I did not find mention of an\n> algorithm for storing them, as it was done, for example, in MS SQL Server.\n> What if instead of 2 fields (xmin and xmax) with a total length of 64 bits\n> - use 1 field (let's call it xid) with a length of 64 bits in tuple\n> header? In this field store the xid of the transaction that created the\n> version. In this case, the new transaction in order to understand whether\n> the read version is suitable for it or not, will have to read the next\n> version as well. Those. The downside of such decision is of course an\n> increase in I / O. Transactions will have to read the +1 version. On the\n> plus side, the title of the tuple remains the same length.\n> \n\nI think that assumes we can easily identify the next version of a tuple,\nand I don't think we can do that. We may be able to do that for for HOT\nchains, but that only works when the next version fits onto the same\npage (and does not update indexed columns). But when we store the new\nversion on a separate page, we don't have any link between those tuples.\nAnd adding it may easily mean more overhead than the 8B we'd save by\nonly storing a single XID.\n\nIMO the most promising solution to this is the \"page epoch\" approach\ndiscussed some time ago (1-2 years?).\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Fri, 1 Nov 2019 18:14:50 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":" The proposed option does not need to change the length of either the page header or tuple header. Therefore, you will not need to physically change the data. regards01.11.2019, 20:10, \"Tomas Vondra\" <tomas.vondra@2ndquadrant.com>:On Fri, Nov 01, 2019 at 10:25:17AM +0100, Pavel Stehule wrote:Hipá 1. 11. 2019 v 10:11 odesílatel Павел Ерёмин <shnoor111gmail@yandex.ru>napsal: Hi. sorry for my English. I want to once again open the topic of 64 bit transaction id. I did not manage to find in the archive of the option that I want to discuss, so I write. If I searched poorly, then please forgive me. The idea is not very original and probably has already been considered, again I repeat - I did not find it. Therefore, please do not scold me severely. In discussions of 64-bit transaction id, I did not find mention of an algorithm for storing them, as it was done, for example, in MS SQL Server. What if instead of 2 fields (xmin and xmax) with a total length of 64 bits - use 1 field (let's call it xid) with a length of 64 bits in tuple header? In this field store the xid of the transaction that created the version. In this case, the new transaction in order to understand whether the read version is suitable for it or not, will have to read the next version as well. Those. The downside of such decision is of course an increase in I / O. Transactions will have to read the +1 version. On the plus side, the title of the tuple remains the same length.is 32 bit tid really problem? Why you need to know state of last 2^31transactions? Is not problem in too low usage (or maybe too high overhead)of VACUUM FREEZE.It certainly can be an issue for large and busy systems, that may needanti-wraparoud vacuum every couple of days. If that requires rewriting acouple of TB of data, it's not particularly nice. That's why 64-bit XIDswere discussed repeatedly in the past, and it's likely to get even morepressing as the systems get larger.I am not sure if increasing this range can has much more fatal problems(maybe later)Well, not fatal, but naive approaches can increase per-tuple overhead.And we already have plenty of that, hence there were proposals to usepage epochs and so on.regards-- Tomas Vondra http://www.2ndQuadrant.comPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services ","msg_date":"Sat, 02 Nov 2019 19:07:17 +0300","msg_from":"=?utf-8?B?0J/QsNCy0LXQuyDQldGA0ZHQvNC40L0=?= <shnoor111gmail@yandex.ru>","msg_from_op":true,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Sat, Nov 02, 2019 at 07:07:17PM +0300, Павел Ерёмин wrote:\n> \n> The proposed option does not need to change the length of either the page\n> header or tuple header. Therefore, you will not need to physically change\n> the data.\n> \n\nSo how do you link the tuple versions together? Clearly, that has to be\nstored somewhere ...\n\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Sat, 2 Nov 2019 19:15:34 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"The proposed option is not much different from what it is now.We are not trying to save some space - we will reuse the existing one. We just work in 64 bit transaction counters. Correct me if I'm wrong - the address of the next version of the line is stored in the 6 byte field t_cid in the tuple header - which is not attached to the current page in any way - and can be stored anywhere in the table. Nothing changes.I often explain things very poorly, but I will try.for exampleEach transaction is assigned a unique 64-bit xid.In the tuple header, we replace 32-bit xmin and xmax - with one 64-bit field - let's call it xid.SupposeTransaction 1 does INSERTThe first version is created (Tuple1).Tuple1. Tuple_header.xid = Transacrion1.xid and Tuple1. Tuple_header. t_cid = 0;Transaction 3 (started after transaction 1) does UPDATEThe second version is created (Tuple2).Tuple1. Tuple_header. t_cid = (address) Tuple2;Tuple2. Tuple_header.xid = Transacrion3.xid and Tuple2. Tuple_header. t_cid = 0;Transaction 2 (started between transaction1 and transaction2) makes SELECTReads Tuple1. Transaction 2 sees that Tuple1.Tuple_header.xid <Transacrion2.xid, sees that the address Tuple1 is filled. Tuple_header. t_cid, follow it and read the version of Tuple2.02.11.2019, 21:15, \"Tomas Vondra\" <tomas.vondra@2ndquadrant.com>:On Sat, Nov 02, 2019 at 07:07:17PM +0300, Павел Ерёмин wrote: The proposed option does not need to change the length of either the page header or tuple header. Therefore, you will not need to physically change the data. So how do you link the tuple versions together? Clearly, that has to bestored somewhere ...-- Tomas Vondra http://www.2ndQuadrant.comPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services ","msg_date":"Sat, 02 Nov 2019 23:35:09 +0300","msg_from":"=?utf-8?B?0J/QsNCy0LXQuyDQldGA0ZHQvNC40L0=?= <shnoor111gmail@yandex.ru>","msg_from_op":true,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Sat, Nov 02, 2019 at 11:35:09PM +0300, Павел Ерёмин wrote:\n> The proposed option is not much different from what it is now.\n> We are not trying to save some space - we will reuse the existing one. We\n> just work in 64 bit transaction counters. Correct me if I'm wrong - the\n> address of the next version of the line is stored in the 6 byte field\n> t_cid in the tuple header - which is not attached to the current page in\n> any way - and can be stored anywhere in the table. Nothing changes.\n\nI think you mean t_ctid, not t_cid (which is a 4-byte CommandId, not any\nsort of item pointer).\n\nI think this comment from htup_details.h explains the issue:\n\n * ... Beware however that VACUUM might\n * erase the pointed-to (newer) tuple before erasing the pointing (older)\n * tuple. Hence, when following a t_ctid link, it is necessary to check\n * to see if the referenced slot is empty or contains an unrelated tuple.\n * Check that the referenced tuple has XMIN equal to the referencing tuple's\n * XMAX to verify that it is actually the descendant version and not an\n * unrelated tuple stored into a slot recently freed by VACUUM. If either\n * check fails, one may assume that there is no live descendant version.\n\nNow, imagine you have a tuple that gets updated repeatedly (say, 3x) and\neach version gets to a different page. Say, pages #1, #2, #3. And then\nVACUUM happens on some of the \"middle\" page (this may happen when trying\nto fit new row onto a page to allow HOT, but it might happen even during\nregular VACUUM).\n\nSo we started with 3 tuples on pages #1, #2, #3, but now we have this\n\n #1 - tuple exists, points to tuple on page #2\n #2 - tuple no longer exists, cleaned up by vacuum\n #3 - tuple exists\n\nThe scheme you proposed requires existence of all the tuples in the\nchain to determine visibility. When tuple #2 no longer exists, it's\nimpossible to decide whether tuple on page #1 is visible or not.\n\nThis also significantly increases the amount of random I/O, pretty much\nby factor of 2, because whenever you look at a row, you also have to\nlook at the \"next version\" which may be on another page. That's pretty\nbad, bot for I/O and cache hit ratio. I don't think that's a reasonable\ntrade-off (at least compared to simply making the XIDs 64bit).\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Sun, 3 Nov 2019 00:20:43 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"I completely agree with all of the above. Therefore, the proposed mechanism may entail larger improvements (and not only VACUUM).I can offer the following solution.For VACUUM, create a hash table.VACUUM scanning the table sees that the version (tuple1) has t_ctid filled and refers to the address tuple2, it creates a structure into which it writes the address tuple1, tuple1.xid, length tuple1 (well, and other information that is needed), puts this structure in the hash table by key tuple2 addresses.VACUUM reaches tuple2, checks the address of tuple2 in the hash table - if it finds it, it evaluates the connection between them and makes a decision on cleaning. regards03.11.2019, 02:20, \"Tomas Vondra\" <tomas.vondra@2ndquadrant.com>:On Sat, Nov 02, 2019 at 11:35:09PM +0300, Павел Ерёмин wrote: The proposed option is not much different from what it is now. We are not trying to save some space - we will reuse the existing one. We just work in 64 bit transaction counters. Correct me if I'm wrong - the address of the next version of the line is stored in the 6 byte field t_cid in the tuple header - which is not attached to the current page in any way - and can be stored anywhere in the table. Nothing changes.I think you mean t_ctid, not t_cid (which is a 4-byte CommandId, not anysort of item pointer).I think this comment from htup_details.h explains the issue: * ... Beware however that VACUUM might * erase the pointed-to (newer) tuple before erasing the pointing (older) * tuple. Hence, when following a t_ctid link, it is necessary to check * to see if the referenced slot is empty or contains an unrelated tuple. * Check that the referenced tuple has XMIN equal to the referencing tuple's * XMAX to verify that it is actually the descendant version and not an * unrelated tuple stored into a slot recently freed by VACUUM. If either * check fails, one may assume that there is no live descendant version.Now, imagine you have a tuple that gets updated repeatedly (say, 3x) andeach version gets to a different page. Say, pages #1, #2, #3. And thenVACUUM happens on some of the \"middle\" page (this may happen when tryingto fit new row onto a page to allow HOT, but it might happen even duringregular VACUUM).So we started with 3 tuples on pages #1, #2, #3, but now we have this #1 - tuple exists, points to tuple on page #2 #2 - tuple no longer exists, cleaned up by vacuum #3 - tuple existsThe scheme you proposed requires existence of all the tuples in thechain to determine visibility. When tuple #2 no longer exists, it'simpossible to decide whether tuple on page #1 is visible or not.This also significantly increases the amount of random I/O, pretty muchby factor of 2, because whenever you look at a row, you also have tolook at the \"next version\" which may be on another page. That's prettybad, bot for I/O and cache hit ratio. I don't think that's a reasonabletrade-off (at least compared to simply making the XIDs 64bit).regards-- Tomas Vondra http://www.2ndQuadrant.comPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services ","msg_date":"Sun, 03 Nov 2019 14:17:15 +0300","msg_from":"=?utf-8?B?0J/QsNCy0LXQuyDQldGA0ZHQvNC40L0=?= <shnoor111gmail@yandex.ru>","msg_from_op":true,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Sun, Nov 03, 2019 at 02:17:15PM +0300, Павел Ерёмин wrote:\n> I completely agree with all of the above. Therefore, the proposed\n> mechanism may entail larger improvements (and not only VACUUM).\n\nI think the best think you can do is try implementing this ...\n\nI'm afraid the \"improvements\" essentially mean making various imporant\nparts of the system much more complicated and expensive. There's a\ntrade-off between saving 8B per row and additional overhead (during\nvacuum etc.), and it does not seem like a winning strategy. What started\nas \"we can simply look at the next row version\" is clearly way more\ncomplicated and expensive.\n\nThe trouble here is that it adds dependency between pages in the data\nfile. That for example means that during cleanup of a page it may be \nnecessary to modify the other page, when originally that would be \nread-only in that checkpoint interval. That's essentially write \namplification, and may significantly increase the amount of WAL due to \ngenerating FPW for the other page.\n\n> I can offer the following solution.\n> For VACUUM, create a hash table.\n> VACUUM scanning the table sees that the version (tuple1) has t_ctid filled\n> and refers to the address tuple2, it creates a structure into which it\n> writes the address tuple1, tuple1.xid, length tuple1 (well, and other\n> information that is needed), puts this structure in the hash table by key\n> tuple2 addresses.\n> VACUUM reaches tuple2, checks the address of tuple2 in the hash table - if\n> it finds it, it evaluates the connection between them and makes a decision\n> on cleaning.\n> \n\nWe know VACUUM is already pretty expensive, so making it even more\nexpensive seems pretty awful. And the proposed solution seems damn\nexpensive. We already do something similar for indexes - we track\npointers for removed rows, so that we can remove them from indexes. And\nit's damn expensive because we don't know where in the index the tuples\nare - so we have to scan the whole indexes.\n\nThis would mean we have to do the same thing for table, because we don't\nknow where in the table are the older versions of those rows, because we\ndon't know where the other rows are. That seems mighty expensive.\n\nNot to mention that this does nothing for page-level vacuum, which we\ndo when trying to fit another row on a page (e.g. for HOT). This has to\nbe absolutely cheap, we certainly are not going to do lookups of other\npages or looking for older versions of the row, and so on.\n\nBeing able to do visibility decisions based on the tuple alone (or\npossibly page-level + tuple information) has a lot of value, and I don't\nthink we want to make this more complicated.\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Sun, 3 Nov 2019 20:15:22 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Sat, Nov 2, 2019 at 6:15 AM Tomas Vondra\n<tomas.vondra@2ndquadrant.com> wrote:\n> On Fri, Nov 01, 2019 at 12:05:12PM +0300, Павел Ерёмин wrote:\n> > Hi.\n> > sorry for my English.\n> > I want to once again open the topic of 64 bit transaction id. I did not\n> > manage to find in the archive of the option that I want to discuss, so I\n> > write. If I searched poorly, then please forgive me.\n> > The idea is not very original and probably has already been considered,\n> > again I repeat - I did not find it. Therefore, please do not scold me\n> > severely.\n> > In discussions of 64-bit transaction id, I did not find mention of an\n> > algorithm for storing them, as it was done, for example, in MS SQL Server.\n> > What if instead of 2 fields (xmin and xmax) with a total length of 64 bits\n> > - use 1 field (let's call it xid) with a length of 64 bits in tuple\n> > header? In this field store the xid of the transaction that created the\n> > version. In this case, the new transaction in order to understand whether\n> > the read version is suitable for it or not, will have to read the next\n> > version as well. Those. The downside of such decision is of course an\n> > increase in I / O. Transactions will have to read the +1 version. On the\n> > plus side, the title of the tuple remains the same length.\n> >\n>\n> I think that assumes we can easily identify the next version of a tuple,\n> and I don't think we can do that. We may be able to do that for for HOT\n> chains, but that only works when the next version fits onto the same\n> page (and does not update indexed columns). But when we store the new\n> version on a separate page, we don't have any link between those tuples.\n> And adding it may easily mean more overhead than the 8B we'd save by\n> only storing a single XID.\n>\n> IMO the most promising solution to this is the \"page epoch\" approach\n> discussed some time ago (1-2 years?).\n\nThere have been so many discussions of this topic that it's hard to search for.\n\nSince we have in fact begun to take some baby steps towards using 64\nbit transaction IDs in a few places, I decided to create a new wiki\npage to try to keep track of the various discussions. If you know\nwhere to find the 'best' discussions (for example the one where, if I\nrecall correctly, it was Heikki who proposed a 'reference'\nFullTranasctionId on the page header) and any proposals that came with\npatches, then I'd be grateful if you could add links to it to this\nwiki page!\n\nhttps://wiki.postgresql.org/wiki/FullTransactionId\n\n\n","msg_date":"Mon, 4 Nov 2019 10:41:54 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"And yet, if I try to implement a similar mechanism, if successful, will my revision be considered? regards03.11.2019, 22:15, \"Tomas Vondra\" <tomas.vondra@2ndquadrant.com>:On Sun, Nov 03, 2019 at 02:17:15PM +0300, Павел Ерёмин wrote: I completely agree with all of the above. Therefore, the proposed mechanism may entail larger improvements (and not only VACUUM).I think the best think you can do is try implementing this ...I'm afraid the \"improvements\" essentially mean making various imporantparts of the system much more complicated and expensive. There's atrade-off between saving 8B per row and additional overhead (duringvacuum etc.), and it does not seem like a winning strategy. What startedas \"we can simply look at the next row version\" is clearly way morecomplicated and expensive.The trouble here is that it adds dependency between pages in the datafile. That for example means that during cleanup of a page it may be necessary to modify the other page, when originally that would be read-only in that checkpoint interval. That's essentially write amplification, and may significantly increase the amount of WAL due to generating FPW for the other page. I can offer the following solution. For VACUUM, create a hash table. VACUUM scanning the table sees that the version (tuple1) has t_ctid filled and refers to the address tuple2, it creates a structure into which it writes the address tuple1, tuple1.xid, length tuple1 (well, and other information that is needed), puts this structure in the hash table by key tuple2 addresses. VACUUM reaches tuple2, checks the address of tuple2 in the hash table - if it finds it, it evaluates the connection between them and makes a decision on cleaning.We know VACUUM is already pretty expensive, so making it even moreexpensive seems pretty awful. And the proposed solution seems damnexpensive. We already do something similar for indexes - we trackpointers for removed rows, so that we can remove them from indexes. Andit's damn expensive because we don't know where in the index the tuplesare - so we have to scan the whole indexes.This would mean we have to do the same thing for table, because we don'tknow where in the table are the older versions of those rows, because wedon't know where the other rows are. That seems mighty expensive.Not to mention that this does nothing for page-level vacuum, which wedo when trying to fit another row on a page (e.g. for HOT). This has tobe absolutely cheap, we certainly are not going to do lookups of otherpages or looking for older versions of the row, and so on.Being able to do visibility decisions based on the tuple alone (orpossibly page-level + tuple information) has a lot of value, and I don'tthink we want to make this more complicated.regards-- Tomas Vondra http://www.2ndQuadrant.comPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services ","msg_date":"Mon, 04 Nov 2019 16:39:44 +0300","msg_from":"=?utf-8?B?0J/QsNCy0LXQuyDQldGA0ZHQvNC40L0=?= <shnoor111gmail@yandex.ru>","msg_from_op":true,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Mon, Nov 04, 2019 at 04:39:44PM +0300, Павел Ерёмин wrote:\n> And yet, if I try to implement a similar mechanism, if successful, will my\n> revision be considered?\n> \n\nWhy wouldn't it be considered? If you submit a patch that demonstrably\nimproves the behavior (in this case reduces per-tuple overhead without\ncausing significant issues elsewhere), we'd be crazy not to consider it.\n\nThe bar is pretty high, though, because this touches one of the core\npieces of the database.\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Mon, 4 Nov 2019 16:07:23 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"Hi,\n\n(I've not read the rest of this thread yet)\n\nOn 2019-11-04 16:07:23 +0100, Tomas Vondra wrote:\n> On Mon, Nov 04, 2019 at 04:39:44PM +0300, Павел Ерёмин wrote:\n> > And yet, if I try to implement a similar mechanism, if successful, will my\n> > revision be considered?\n> > \n> \n> Why wouldn't it be considered? If you submit a patch that demonstrably\n> improves the behavior (in this case reduces per-tuple overhead without\n> causing significant issues elsewhere), we'd be crazy not to consider it.\n\nAnd \"without causing significant issues elsewhere\" unfortunately\nincludes continuing to allow pg_upgrade to work.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 4 Nov 2019 10:04:09 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Mon, Nov 04, 2019 at 10:04:09AM -0800, Andres Freund wrote:\n>Hi,\n>\n>(I've not read the rest of this thread yet)\n>\n>On 2019-11-04 16:07:23 +0100, Tomas Vondra wrote:\n>> On Mon, Nov 04, 2019 at 04:39:44PM +0300, Павел Ерёмин wrote:\n>> > And yet, if I try to implement a similar mechanism, if successful, will my\n>> > revision be considered?\n>> > \n>>\n>> Why wouldn't it be considered? If you submit a patch that demonstrably\n>> improves the behavior (in this case reduces per-tuple overhead without\n>> causing significant issues elsewhere), we'd be crazy not to consider it.\n>\n>And \"without causing significant issues elsewhere\" unfortunately\n>includes continuing to allow pg_upgrade to work.\n>\n\nYeah. I suppose we could have a different AM implementing this, but\nmaybe that's not possible ...\n\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Mon, 4 Nov 2019 19:39:18 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"Hi,\n\nOn 2019-11-04 19:39:18 +0100, Tomas Vondra wrote:\n> On Mon, Nov 04, 2019 at 10:04:09AM -0800, Andres Freund wrote:\n> > And \"without causing significant issues elsewhere\" unfortunately\n> > includes continuing to allow pg_upgrade to work.\n\n> Yeah. I suppose we could have a different AM implementing this, but\n> maybe that's not possible ...\n\nEntirely possible. But the amount of code duplication / unnecessary\nbranching and the user confusion from two very similar AMs, would have\nto be weighed against the benefits.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 4 Nov 2019 10:44:53 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Mon, Nov 04, 2019 at 10:44:53AM -0800, Andres Freund wrote:\n>Hi,\n>\n>On 2019-11-04 19:39:18 +0100, Tomas Vondra wrote:\n>> On Mon, Nov 04, 2019 at 10:04:09AM -0800, Andres Freund wrote:\n>> > And \"without causing significant issues elsewhere\" unfortunately\n>> > includes continuing to allow pg_upgrade to work.\n>\n>> Yeah. I suppose we could have a different AM implementing this, but\n>> maybe that's not possible ...\n>\n>Entirely possible. But the amount of code duplication / unnecessary\n>branching and the user confusion from two very similar AMs, would have\n>to be weighed against the benefits.\n>\n\nAgreed. I think code complexity is part of the trade-off. IMO it's fine\nto hack existing heap AM initially, and only explore the separate AM if\nthat turns out to be promising.\n\nregards\n\n-- \nTomas Vondra http://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \n\n\n","msg_date":"Mon, 4 Nov 2019 20:44:55 +0100","msg_from":"Tomas Vondra <tomas.vondra@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Tue, Nov 5, 2019 at 8:45 AM Tomas Vondra\n<tomas.vondra@2ndquadrant.com> wrote:\n> On Mon, Nov 04, 2019 at 10:44:53AM -0800, Andres Freund wrote:\n> >On 2019-11-04 19:39:18 +0100, Tomas Vondra wrote:\n> >> On Mon, Nov 04, 2019 at 10:04:09AM -0800, Andres Freund wrote:\n> >> > And \"without causing significant issues elsewhere\" unfortunately\n> >> > includes continuing to allow pg_upgrade to work.\n> >\n> >> Yeah. I suppose we could have a different AM implementing this, but\n> >> maybe that's not possible ...\n> >\n> >Entirely possible. But the amount of code duplication / unnecessary\n> >branching and the user confusion from two very similar AMs, would have\n> >to be weighed against the benefits.\n> >\n>\n> Agreed. I think code complexity is part of the trade-off. IMO it's fine\n> to hack existing heap AM initially, and only explore the separate AM if\n> that turns out to be promising.\n\nI thought a bit about how to make a minimally-diffferent-from-heap\nnon-freezing table AM using 64 bit xids, as a thought experiment when\ntrying to understand or explain to others what zheap is about.\nCommitted transactions are easy (you don't have to freeze fxid\nreferences from the ancient past because they don't wrap around so\nthey always look old), but how do you deal with *aborted* transactions\nwhen truncating the CLOG (given that our current rule is \"if it's\nbefore the CLOG begins, it must be committed\")? I see three\npossibilities: (1) don't truncate the CLOG anymore (use 64 bit\naddressing and let it leak disk forever, like we did before commit\n2589735d and later work), (2) freeze aborted transactions only, using\na wraparound vacuum (and now you have failed, if the goal was to avoid\nhaving to scan all tuples periodically to freeze stuff, though\nadmittedly it will require less IO to freeze only the aborted\ntransactions), (3) go and remove aborted fxid references eagerly, when\nyou roll back (this could be done using the undo technology that we\nhave been developing to support zheap). Another way to explain (3) is\nthat this hypothetical table AM, let's call it \"yheap\", takes the\nminimum parts of the zheap technology stack required to get rid of\nvacuum-for-wraparound, without doing in-place updates or any of that\nhard stuff. To make this really work you'd also have to deal with\nmultixacts, which also require freezing. If that all sounds too\ncomplicated, you're back to (2) which seems a bit weak to me. Or\nperhaps I'm missing something?\n\n\n","msg_date":"Tue, 5 Nov 2019 09:34:52 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Tue, Nov 5, 2019 at 09:34:52AM +1300, Thomas Munro wrote:\n> On Tue, Nov 5, 2019 at 8:45 AM Tomas Vondra\n> > Agreed. I think code complexity is part of the trade-off. IMO it's fine\n> > to hack existing heap AM initially, and only explore the separate AM if\n> > that turns out to be promising.\n> \n> I thought a bit about how to make a minimally-diffferent-from-heap\n> non-freezing table AM using 64 bit xids, as a thought experiment when\n> trying to understand or explain to others what zheap is about.\n> Committed transactions are easy (you don't have to freeze fxid\n> references from the ancient past because they don't wrap around so\n> they always look old), but how do you deal with *aborted* transactions\n> when truncating the CLOG (given that our current rule is \"if it's\n> before the CLOG begins, it must be committed\")? I see three\n> possibilities: (1) don't truncate the CLOG anymore (use 64 bit\n> addressing and let it leak disk forever, like we did before commit\n> 2589735d and later work), (2) freeze aborted transactions only, using\n> a wraparound vacuum (and now you have failed, if the goal was to avoid\n> having to scan all tuples periodically to freeze stuff, though\n> admittedly it will require less IO to freeze only the aborted\n> transactions), (3) go and remove aborted fxid references eagerly, when\n> you roll back (this could be done using the undo technology that we\n> have been developing to support zheap). Another way to explain (3) is\n> that this hypothetical table AM, let's call it \"yheap\", takes the\n> minimum parts of the zheap technology stack required to get rid of\n> vacuum-for-wraparound, without doing in-place updates or any of that\n> hard stuff. To make this really work you'd also have to deal with\n> multixacts, which also require freezing. If that all sounds too\n> complicated, you're back to (2) which seems a bit weak to me. Or\n> perhaps I'm missing something?\n\nThe above is a very good summary of the constraints that have led to our\ncurrent handling of XID wraparound. If we are concerned about excessive\nvacuum freeze overhead, why is the default autovacuum_freeze_max_age =\n200000000 so low? That causes feezing when the pg_xact directory holds\n200 million xids or 50 megabytes of xid status?\n\nAs far as I understand it, we cause the database to stop writes when the\nxid counter gets within 2 billion xids of the current transaction\ncounter, so 200 million is only 1/10th to that limit, and even then, I\nam not sure why we couldn't make it stop writes at 3 billion or\nsomething. My point is that increasing the default\nautovacuum_freeze_max_age value seems like an easy way to reduce vacuum\nfreeze. (While, the visibility map helps avoid vacuum freeze from\nreading all heap pages, and we still need to read all index pages.)\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n+ As you are, so once was I. As I am, so you will be. +\n+ Ancient Roman grave inscription +\n\n\n","msg_date":"Thu, 7 Nov 2019 10:28:31 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"},{"msg_contents":"On Thu, Nov 7, 2019 at 10:28 AM Bruce Momjian <bruce@momjian.us> wrote:\n> The above is a very good summary of the constraints that have led to our\n> current handling of XID wraparound. If we are concerned about excessive\n> vacuum freeze overhead, why is the default autovacuum_freeze_max_age =\n> 200000000 so low? That causes feezing when the pg_xact directory holds\n> 200 million xids or 50 megabytes of xid status?\n>\n> As far as I understand it, we cause the database to stop writes when the\n> xid counter gets within 2 billion xids of the current transaction\n> counter, so 200 million is only 1/10th to that limit, and even then, I\n> am not sure why we couldn't make it stop writes at 3 billion or\n> something. My point is that increasing the default\n> autovacuum_freeze_max_age value seems like an easy way to reduce vacuum\n> freeze. (While, the visibility map helps avoid vacuum freeze from\n> reading all heap pages, and we still need to read all index pages.)\n\nYeah, I've also wondered why this isn't higher by default, but it's a\nsomewhat tricky topic.\n\nThree billion won't work, because it's deeply baked into PostgreSQL's\narchitecture that at most two billion XIDs are used at one time. For\ncomparison purposes, the four billion XIDs form a ring, so that from\nthe perspective of any individual XID, half of the XIDs are in the\nfuture and the other half are in the past. If three billion XIDs were\nin use simultaneously, say starting with XID 5 and ending with XID\n3,000,000,004, then XID 5 would see XID 3,000,000,004 as being the\npast rather than the future, while XID 1,500,000,000 would (correctly)\nsee XID 5 as in the past and XID 3,000,000,004 as in the future. So\nXID comparison would not be transitive, which would break a lot of\ncode. Allowing at most two billion XIDs to be in use at one time fixes\nthis problem.\n\nThat doesn't mean we couldn't raise the setting. It just means that\nthe hard limit is two billion, not four billion. But, how high should\nwe raise it? The highest safe value depends on how many XIDs you'll\nburn while the freezing vacuums are running, which depends on both the\nsize of the database and the rate of XID consumption, and those values\ncan be very different on different systems. I think most people could\nget by with a significantly higher value, but even with the current\nvalue I think there are probably some people who run out of XIDs, at\nwhich point they can no longer write to the database. The higher we\nmake the default, the more people are going to have that problem. It's\ntrue that a lot of people only hit the limit because something has\ngone wrong, like they've forgotten about a prepared transaction or an\nunused replication slot, but still, on high-velocity systems you can't\nafford to cut it too close because you're still going to be burning\nthrough XIDs while vacuum is running.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 6 Dec 2019 08:30:28 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: 64 bit transaction id"}],"string":"[\n {\n \"msg_contents\": \"Hi.sorry for my English.I want to once again open the topic of 64 bit transaction id. I did not manage to find in the archive of the option that I want to discuss, so I write. If I searched poorly, then please forgive me.The idea is not very original and probably has already been considered, again I repeat - I did not find it. Therefore, please do not scold me severely.In discussions of 64-bit transaction id, I did not find mention of an algorithm for storing them, as it was done, for example, in MS SQL Server.What if instead of 2 fields (xmin and xmax) with a total length of 64 bits - use 1 field (let's call it xid) with a length of 64 bits in tuple header? In this field store the xid of the transaction that created the version. In this case, the new transaction in order to understand whether the read version is suitable for it or not, will have to read the next version as well. Those. The downside of such decision is of course an increase in I / O. Transactions will have to read the +1 version. On the plus side, the title of the tuple remains the same length. regards, Eremin Pavel.\",\n \"msg_date\": \"Fri, 01 Nov 2019 12:05:12 +0300\",\n \"msg_from\": \"=?utf-8?B?0J/QsNCy0LXQuyDQldGA0ZHQvNC40L0=?= <shnoor111gmail@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"64 bit transaction id\"\n },\n {\n \"msg_contents\": \"Hi\\n\\npá 1. 11. 2019 v 10:11 odesílatel Павел Ерёмин <shnoor111gmail@yandex.ru>\\nnapsal:\\n\\n> Hi.\\n> sorry for my English.\\n> I want to once again open the topic of 64 bit transaction id. I did not\\n> manage to find in the archive of the option that I want to discuss, so I\\n> write. If I searched poorly, then please forgive me.\\n> The idea is not very original and probably has already been considered,\\n> again I repeat - I did not find it. Therefore, please do not scold me\\n> severely.\\n> In discussions of 64-bit transaction id, I did not find mention of an\\n> algorithm for storing them, as it was done, for example, in MS SQL Server.\\n> What if instead of 2 fields (xmin and xmax) with a total length of 64 bits\\n> - use 1 field (let's call it xid) with a length of 64 bits in tuple header?\\n> In this field store the xid of the transaction that created the version. In\\n> this case, the new transaction in order to understand whether the read\\n> version is suitable for it or not, will have to read the next version as\\n> well. Those. The downside of such decision is of course an increase in I /\\n> O. Transactions will have to read the +1 version. On the plus side, the\\n> title of the tuple remains the same length.\\n>\\n\\nis 32 bit tid really problem? Why you need to know state of last 2^31\\ntransactions? Is not problem in too low usage (or maybe too high overhead)\\nof VACUUM FREEZE.\\n\\nI am not sure if increasing this range can has much more fatal problems\\n(maybe later)\\n\\nPavel\\n\\n\\n\\n>\\n> regards, Eremin Pavel.\\n>\\n\\nHipá 1. 11. 2019 v 10:11 odesílatel Павел Ерёмин <shnoor111gmail@yandex.ru> napsal:Hi.sorry for my English.I want to once again open the topic of 64 bit transaction id. I did not manage to find in the archive of the option that I want to discuss, so I write. If I searched poorly, then please forgive me.The idea is not very original and probably has already been considered, again I repeat - I did not find it. Therefore, please do not scold me severely.In discussions of 64-bit transaction id, I did not find mention of an algorithm for storing them, as it was done, for example, in MS SQL Server.What if instead of 2 fields (xmin and xmax) with a total length of 64 bits - use 1 field (let's call it xid) with a length of 64 bits in tuple header? In this field store the xid of the transaction that created the version. In this case, the new transaction in order to understand whether the read version is suitable for it or not, will have to read the next version as well. Those. The downside of such decision is of course an increase in I / O. Transactions will have to read the +1 version. On the plus side, the title of the tuple remains the same length.is 32 bit tid really problem? Why you need to know state of last 2^31 transactions? Is not problem in too low usage (or maybe too high overhead) of VACUUM FREEZE.I am not sure if increasing this range can has much more fatal problems (maybe later)Pavel regards, Eremin Pavel.\",\n \"msg_date\": \"Fri, 1 Nov 2019 10:25:17 +0100\",\n \"msg_from\": \"Pavel Stehule <pavel.stehule@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 01, 2019 at 10:25:17AM +0100, Pavel Stehule wrote:\\n>Hi\\n>\\n>pá 1. 11. 2019 v 10:11 odesílatel Павел Ерёмин <shnoor111gmail@yandex.ru>\\n>napsal:\\n>\\n>> Hi.\\n>> sorry for my English.\\n>> I want to once again open the topic of 64 bit transaction id. I did not\\n>> manage to find in the archive of the option that I want to discuss, so I\\n>> write. If I searched poorly, then please forgive me.\\n>> The idea is not very original and probably has already been considered,\\n>> again I repeat - I did not find it. Therefore, please do not scold me\\n>> severely.\\n>> In discussions of 64-bit transaction id, I did not find mention of an\\n>> algorithm for storing them, as it was done, for example, in MS SQL Server.\\n>> What if instead of 2 fields (xmin and xmax) with a total length of 64 bits\\n>> - use 1 field (let's call it xid) with a length of 64 bits in tuple header?\\n>> In this field store the xid of the transaction that created the version. In\\n>> this case, the new transaction in order to understand whether the read\\n>> version is suitable for it or not, will have to read the next version as\\n>> well. Those. The downside of such decision is of course an increase in I /\\n>> O. Transactions will have to read the +1 version. On the plus side, the\\n>> title of the tuple remains the same length.\\n>>\\n>\\n>is 32 bit tid really problem? Why you need to know state of last 2^31\\n>transactions? Is not problem in too low usage (or maybe too high overhead)\\n>of VACUUM FREEZE.\\n>\\n\\nIt certainly can be an issue for large and busy systems, that may need\\nanti-wraparoud vacuum every couple of days. If that requires rewriting a\\ncouple of TB of data, it's not particularly nice. That's why 64-bit XIDs\\nwere discussed repeatedly in the past, and it's likely to get even more\\npressing as the systems get larger.\\n\\n>I am not sure if increasing this range can has much more fatal problems\\n>(maybe later)\\n>\\n\\nWell, not fatal, but naive approaches can increase per-tuple overhead.\\nAnd we already have plenty of that, hence there were proposals to use\\npage epochs and so on.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 18:10:38 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 01, 2019 at 12:05:12PM +0300, Павел Ерёмин wrote:\\n> Hi.\\n> sorry for my English.\\n> I want to once again open the topic of 64 bit transaction id. I did not\\n> manage to find in the archive of the option that I want to discuss, so I\\n> write. If I searched poorly, then please forgive me.\\n> The idea is not very original and probably has already been considered,\\n> again I repeat - I did not find it. Therefore, please do not scold me\\n> severely.\\n> In discussions of 64-bit transaction id, I did not find mention of an\\n> algorithm for storing them, as it was done, for example, in MS SQL Server.\\n> What if instead of 2 fields (xmin and xmax) with a total length of 64 bits\\n> - use 1 field (let's call it xid) with a length of 64 bits in tuple\\n> header? In this field store the xid of the transaction that created the\\n> version. In this case, the new transaction in order to understand whether\\n> the read version is suitable for it or not, will have to read the next\\n> version as well. Those. The downside of such decision is of course an\\n> increase in I / O. Transactions will have to read the +1 version. On the\\n> plus side, the title of the tuple remains the same length.\\n> \\n\\nI think that assumes we can easily identify the next version of a tuple,\\nand I don't think we can do that. We may be able to do that for for HOT\\nchains, but that only works when the next version fits onto the same\\npage (and does not update indexed columns). But when we store the new\\nversion on a separate page, we don't have any link between those tuples.\\nAnd adding it may easily mean more overhead than the 8B we'd save by\\nonly storing a single XID.\\n\\nIMO the most promising solution to this is the \\\"page epoch\\\" approach\\ndiscussed some time ago (1-2 years?).\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 18:14:50 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \" The proposed option does not need to change the length of either the page header or tuple header. Therefore, you will not need to physically change the data. regards01.11.2019, 20:10, \\\"Tomas Vondra\\\" <tomas.vondra@2ndquadrant.com>:On Fri, Nov 01, 2019 at 10:25:17AM +0100, Pavel Stehule wrote:Hipá 1. 11. 2019 v 10:11 odesílatel Павел Ерёмин <shnoor111gmail@yandex.ru>napsal: Hi. sorry for my English. I want to once again open the topic of 64 bit transaction id. I did not manage to find in the archive of the option that I want to discuss, so I write. If I searched poorly, then please forgive me. The idea is not very original and probably has already been considered, again I repeat - I did not find it. Therefore, please do not scold me severely. In discussions of 64-bit transaction id, I did not find mention of an algorithm for storing them, as it was done, for example, in MS SQL Server. What if instead of 2 fields (xmin and xmax) with a total length of 64 bits - use 1 field (let's call it xid) with a length of 64 bits in tuple header? In this field store the xid of the transaction that created the version. In this case, the new transaction in order to understand whether the read version is suitable for it or not, will have to read the next version as well. Those. The downside of such decision is of course an increase in I / O. Transactions will have to read the +1 version. On the plus side, the title of the tuple remains the same length.is 32 bit tid really problem? Why you need to know state of last 2^31transactions? Is not problem in too low usage (or maybe too high overhead)of VACUUM FREEZE.It certainly can be an issue for large and busy systems, that may needanti-wraparoud vacuum every couple of days. If that requires rewriting acouple of TB of data, it's not particularly nice. That's why 64-bit XIDswere discussed repeatedly in the past, and it's likely to get even morepressing as the systems get larger.I am not sure if increasing this range can has much more fatal problems(maybe later)Well, not fatal, but naive approaches can increase per-tuple overhead.And we already have plenty of that, hence there were proposals to usepage epochs and so on.regards-- Tomas Vondra http://www.2ndQuadrant.comPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \",\n \"msg_date\": \"Sat, 02 Nov 2019 19:07:17 +0300\",\n \"msg_from\": \"=?utf-8?B?0J/QsNCy0LXQuyDQldGA0ZHQvNC40L0=?= <shnoor111gmail@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 02, 2019 at 07:07:17PM +0300, Павел Ерёмин wrote:\\n> \\n> The proposed option does not need to change the length of either the page\\n> header or tuple header. Therefore, you will not need to physically change\\n> the data.\\n> \\n\\nSo how do you link the tuple versions together? Clearly, that has to be\\nstored somewhere ...\\n\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Sat, 2 Nov 2019 19:15:34 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"The proposed option is not much different from what it is now.We are not trying to save some space - we will reuse the existing one. We just work in 64 bit transaction counters. Correct me if I'm wrong - the address of the next version of the line is stored in the 6 byte field t_cid in the tuple header - which is not attached to the current page in any way - and can be stored anywhere in the table. Nothing changes.I often explain things very poorly, but I will try.for exampleEach transaction is assigned a unique 64-bit xid.In the tuple header, we replace 32-bit xmin and xmax - with one 64-bit field - let's call it xid.SupposeTransaction 1 does INSERTThe first version is created (Tuple1).Tuple1. Tuple_header.xid = Transacrion1.xid and Tuple1. Tuple_header. t_cid = 0;Transaction 3 (started after transaction 1) does UPDATEThe second version is created (Tuple2).Tuple1. Tuple_header. t_cid = (address) Tuple2;Tuple2. Tuple_header.xid = Transacrion3.xid and Tuple2. Tuple_header. t_cid = 0;Transaction 2 (started between transaction1 and transaction2) makes SELECTReads Tuple1. Transaction 2 sees that Tuple1.Tuple_header.xid <Transacrion2.xid, sees that the address Tuple1 is filled. Tuple_header. t_cid, follow it and read the version of Tuple2.02.11.2019, 21:15, \\\"Tomas Vondra\\\" <tomas.vondra@2ndquadrant.com>:On Sat, Nov 02, 2019 at 07:07:17PM +0300, Павел Ерёмин wrote: The proposed option does not need to change the length of either the page header or tuple header. Therefore, you will not need to physically change the data. So how do you link the tuple versions together? Clearly, that has to bestored somewhere ...-- Tomas Vondra http://www.2ndQuadrant.comPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \",\n \"msg_date\": \"Sat, 02 Nov 2019 23:35:09 +0300\",\n \"msg_from\": \"=?utf-8?B?0J/QsNCy0LXQuyDQldGA0ZHQvNC40L0=?= <shnoor111gmail@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 02, 2019 at 11:35:09PM +0300, Павел Ерёмин wrote:\\n> The proposed option is not much different from what it is now.\\n> We are not trying to save some space - we will reuse the existing one. We\\n> just work in 64 bit transaction counters. Correct me if I'm wrong - the\\n> address of the next version of the line is stored in the 6 byte field\\n> t_cid in the tuple header - which is not attached to the current page in\\n> any way - and can be stored anywhere in the table. Nothing changes.\\n\\nI think you mean t_ctid, not t_cid (which is a 4-byte CommandId, not any\\nsort of item pointer).\\n\\nI think this comment from htup_details.h explains the issue:\\n\\n * ... Beware however that VACUUM might\\n * erase the pointed-to (newer) tuple before erasing the pointing (older)\\n * tuple. Hence, when following a t_ctid link, it is necessary to check\\n * to see if the referenced slot is empty or contains an unrelated tuple.\\n * Check that the referenced tuple has XMIN equal to the referencing tuple's\\n * XMAX to verify that it is actually the descendant version and not an\\n * unrelated tuple stored into a slot recently freed by VACUUM. If either\\n * check fails, one may assume that there is no live descendant version.\\n\\nNow, imagine you have a tuple that gets updated repeatedly (say, 3x) and\\neach version gets to a different page. Say, pages #1, #2, #3. And then\\nVACUUM happens on some of the \\\"middle\\\" page (this may happen when trying\\nto fit new row onto a page to allow HOT, but it might happen even during\\nregular VACUUM).\\n\\nSo we started with 3 tuples on pages #1, #2, #3, but now we have this\\n\\n #1 - tuple exists, points to tuple on page #2\\n #2 - tuple no longer exists, cleaned up by vacuum\\n #3 - tuple exists\\n\\nThe scheme you proposed requires existence of all the tuples in the\\nchain to determine visibility. When tuple #2 no longer exists, it's\\nimpossible to decide whether tuple on page #1 is visible or not.\\n\\nThis also significantly increases the amount of random I/O, pretty much\\nby factor of 2, because whenever you look at a row, you also have to\\nlook at the \\\"next version\\\" which may be on another page. That's pretty\\nbad, bot for I/O and cache hit ratio. I don't think that's a reasonable\\ntrade-off (at least compared to simply making the XIDs 64bit).\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Sun, 3 Nov 2019 00:20:43 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"I completely agree with all of the above. Therefore, the proposed mechanism may entail larger improvements (and not only VACUUM).I can offer the following solution.For VACUUM, create a hash table.VACUUM scanning the table sees that the version (tuple1) has t_ctid filled and refers to the address tuple2, it creates a structure into which it writes the address tuple1, tuple1.xid, length tuple1 (well, and other information that is needed), puts this structure in the hash table by key tuple2 addresses.VACUUM reaches tuple2, checks the address of tuple2 in the hash table - if it finds it, it evaluates the connection between them and makes a decision on cleaning. regards03.11.2019, 02:20, \\\"Tomas Vondra\\\" <tomas.vondra@2ndquadrant.com>:On Sat, Nov 02, 2019 at 11:35:09PM +0300, Павел Ерёмин wrote: The proposed option is not much different from what it is now. We are not trying to save some space - we will reuse the existing one. We just work in 64 bit transaction counters. Correct me if I'm wrong - the address of the next version of the line is stored in the 6 byte field t_cid in the tuple header - which is not attached to the current page in any way - and can be stored anywhere in the table. Nothing changes.I think you mean t_ctid, not t_cid (which is a 4-byte CommandId, not anysort of item pointer).I think this comment from htup_details.h explains the issue: * ... Beware however that VACUUM might * erase the pointed-to (newer) tuple before erasing the pointing (older) * tuple. Hence, when following a t_ctid link, it is necessary to check * to see if the referenced slot is empty or contains an unrelated tuple. * Check that the referenced tuple has XMIN equal to the referencing tuple's * XMAX to verify that it is actually the descendant version and not an * unrelated tuple stored into a slot recently freed by VACUUM. If either * check fails, one may assume that there is no live descendant version.Now, imagine you have a tuple that gets updated repeatedly (say, 3x) andeach version gets to a different page. Say, pages #1, #2, #3. And thenVACUUM happens on some of the \\\"middle\\\" page (this may happen when tryingto fit new row onto a page to allow HOT, but it might happen even duringregular VACUUM).So we started with 3 tuples on pages #1, #2, #3, but now we have this #1 - tuple exists, points to tuple on page #2 #2 - tuple no longer exists, cleaned up by vacuum #3 - tuple existsThe scheme you proposed requires existence of all the tuples in thechain to determine visibility. When tuple #2 no longer exists, it'simpossible to decide whether tuple on page #1 is visible or not.This also significantly increases the amount of random I/O, pretty muchby factor of 2, because whenever you look at a row, you also have tolook at the \\\"next version\\\" which may be on another page. That's prettybad, bot for I/O and cache hit ratio. I don't think that's a reasonabletrade-off (at least compared to simply making the XIDs 64bit).regards-- Tomas Vondra http://www.2ndQuadrant.comPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \",\n \"msg_date\": \"Sun, 03 Nov 2019 14:17:15 +0300\",\n \"msg_from\": \"=?utf-8?B?0J/QsNCy0LXQuyDQldGA0ZHQvNC40L0=?= <shnoor111gmail@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 03, 2019 at 02:17:15PM +0300, Павел Ерёмин wrote:\\n> I completely agree with all of the above. Therefore, the proposed\\n> mechanism may entail larger improvements (and not only VACUUM).\\n\\nI think the best think you can do is try implementing this ...\\n\\nI'm afraid the \\\"improvements\\\" essentially mean making various imporant\\nparts of the system much more complicated and expensive. There's a\\ntrade-off between saving 8B per row and additional overhead (during\\nvacuum etc.), and it does not seem like a winning strategy. What started\\nas \\\"we can simply look at the next row version\\\" is clearly way more\\ncomplicated and expensive.\\n\\nThe trouble here is that it adds dependency between pages in the data\\nfile. That for example means that during cleanup of a page it may be \\nnecessary to modify the other page, when originally that would be \\nread-only in that checkpoint interval. That's essentially write \\namplification, and may significantly increase the amount of WAL due to \\ngenerating FPW for the other page.\\n\\n> I can offer the following solution.\\n> For VACUUM, create a hash table.\\n> VACUUM scanning the table sees that the version (tuple1) has t_ctid filled\\n> and refers to the address tuple2, it creates a structure into which it\\n> writes the address tuple1, tuple1.xid, length tuple1 (well, and other\\n> information that is needed), puts this structure in the hash table by key\\n> tuple2 addresses.\\n> VACUUM reaches tuple2, checks the address of tuple2 in the hash table - if\\n> it finds it, it evaluates the connection between them and makes a decision\\n> on cleaning.\\n> \\n\\nWe know VACUUM is already pretty expensive, so making it even more\\nexpensive seems pretty awful. And the proposed solution seems damn\\nexpensive. We already do something similar for indexes - we track\\npointers for removed rows, so that we can remove them from indexes. And\\nit's damn expensive because we don't know where in the index the tuples\\nare - so we have to scan the whole indexes.\\n\\nThis would mean we have to do the same thing for table, because we don't\\nknow where in the table are the older versions of those rows, because we\\ndon't know where the other rows are. That seems mighty expensive.\\n\\nNot to mention that this does nothing for page-level vacuum, which we\\ndo when trying to fit another row on a page (e.g. for HOT). This has to\\nbe absolutely cheap, we certainly are not going to do lookups of other\\npages or looking for older versions of the row, and so on.\\n\\nBeing able to do visibility decisions based on the tuple alone (or\\npossibly page-level + tuple information) has a lot of value, and I don't\\nthink we want to make this more complicated.\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Sun, 3 Nov 2019 20:15:22 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 2, 2019 at 6:15 AM Tomas Vondra\\n<tomas.vondra@2ndquadrant.com> wrote:\\n> On Fri, Nov 01, 2019 at 12:05:12PM +0300, Павел Ерёмин wrote:\\n> > Hi.\\n> > sorry for my English.\\n> > I want to once again open the topic of 64 bit transaction id. I did not\\n> > manage to find in the archive of the option that I want to discuss, so I\\n> > write. If I searched poorly, then please forgive me.\\n> > The idea is not very original and probably has already been considered,\\n> > again I repeat - I did not find it. Therefore, please do not scold me\\n> > severely.\\n> > In discussions of 64-bit transaction id, I did not find mention of an\\n> > algorithm for storing them, as it was done, for example, in MS SQL Server.\\n> > What if instead of 2 fields (xmin and xmax) with a total length of 64 bits\\n> > - use 1 field (let's call it xid) with a length of 64 bits in tuple\\n> > header? In this field store the xid of the transaction that created the\\n> > version. In this case, the new transaction in order to understand whether\\n> > the read version is suitable for it or not, will have to read the next\\n> > version as well. Those. The downside of such decision is of course an\\n> > increase in I / O. Transactions will have to read the +1 version. On the\\n> > plus side, the title of the tuple remains the same length.\\n> >\\n>\\n> I think that assumes we can easily identify the next version of a tuple,\\n> and I don't think we can do that. We may be able to do that for for HOT\\n> chains, but that only works when the next version fits onto the same\\n> page (and does not update indexed columns). But when we store the new\\n> version on a separate page, we don't have any link between those tuples.\\n> And adding it may easily mean more overhead than the 8B we'd save by\\n> only storing a single XID.\\n>\\n> IMO the most promising solution to this is the \\\"page epoch\\\" approach\\n> discussed some time ago (1-2 years?).\\n\\nThere have been so many discussions of this topic that it's hard to search for.\\n\\nSince we have in fact begun to take some baby steps towards using 64\\nbit transaction IDs in a few places, I decided to create a new wiki\\npage to try to keep track of the various discussions. If you know\\nwhere to find the 'best' discussions (for example the one where, if I\\nrecall correctly, it was Heikki who proposed a 'reference'\\nFullTranasctionId on the page header) and any proposals that came with\\npatches, then I'd be grateful if you could add links to it to this\\nwiki page!\\n\\nhttps://wiki.postgresql.org/wiki/FullTransactionId\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Nov 2019 10:41:54 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"And yet, if I try to implement a similar mechanism, if successful, will my revision be considered? regards03.11.2019, 22:15, \\\"Tomas Vondra\\\" <tomas.vondra@2ndquadrant.com>:On Sun, Nov 03, 2019 at 02:17:15PM +0300, Павел Ерёмин wrote: I completely agree with all of the above. Therefore, the proposed mechanism may entail larger improvements (and not only VACUUM).I think the best think you can do is try implementing this ...I'm afraid the \\\"improvements\\\" essentially mean making various imporantparts of the system much more complicated and expensive. There's atrade-off between saving 8B per row and additional overhead (duringvacuum etc.), and it does not seem like a winning strategy. What startedas \\\"we can simply look at the next row version\\\" is clearly way morecomplicated and expensive.The trouble here is that it adds dependency between pages in the datafile. That for example means that during cleanup of a page it may be necessary to modify the other page, when originally that would be read-only in that checkpoint interval. That's essentially write amplification, and may significantly increase the amount of WAL due to generating FPW for the other page. I can offer the following solution. For VACUUM, create a hash table. VACUUM scanning the table sees that the version (tuple1) has t_ctid filled and refers to the address tuple2, it creates a structure into which it writes the address tuple1, tuple1.xid, length tuple1 (well, and other information that is needed), puts this structure in the hash table by key tuple2 addresses. VACUUM reaches tuple2, checks the address of tuple2 in the hash table - if it finds it, it evaluates the connection between them and makes a decision on cleaning.We know VACUUM is already pretty expensive, so making it even moreexpensive seems pretty awful. And the proposed solution seems damnexpensive. We already do something similar for indexes - we trackpointers for removed rows, so that we can remove them from indexes. Andit's damn expensive because we don't know where in the index the tuplesare - so we have to scan the whole indexes.This would mean we have to do the same thing for table, because we don'tknow where in the table are the older versions of those rows, because wedon't know where the other rows are. That seems mighty expensive.Not to mention that this does nothing for page-level vacuum, which wedo when trying to fit another row on a page (e.g. for HOT). This has tobe absolutely cheap, we certainly are not going to do lookups of otherpages or looking for older versions of the row, and so on.Being able to do visibility decisions based on the tuple alone (orpossibly page-level + tuple information) has a lot of value, and I don'tthink we want to make this more complicated.regards-- Tomas Vondra http://www.2ndQuadrant.comPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \",\n \"msg_date\": \"Mon, 04 Nov 2019 16:39:44 +0300\",\n \"msg_from\": \"=?utf-8?B?0J/QsNCy0LXQuyDQldGA0ZHQvNC40L0=?= <shnoor111gmail@yandex.ru>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 04, 2019 at 04:39:44PM +0300, Павел Ерёмин wrote:\\n> And yet, if I try to implement a similar mechanism, if successful, will my\\n> revision be considered?\\n> \\n\\nWhy wouldn't it be considered? If you submit a patch that demonstrably\\nimproves the behavior (in this case reduces per-tuple overhead without\\ncausing significant issues elsewhere), we'd be crazy not to consider it.\\n\\nThe bar is pretty high, though, because this touches one of the core\\npieces of the database.\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Nov 2019 16:07:23 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\n(I've not read the rest of this thread yet)\\n\\nOn 2019-11-04 16:07:23 +0100, Tomas Vondra wrote:\\n> On Mon, Nov 04, 2019 at 04:39:44PM +0300, Павел Ерёмин wrote:\\n> > And yet, if I try to implement a similar mechanism, if successful, will my\\n> > revision be considered?\\n> > \\n> \\n> Why wouldn't it be considered? If you submit a patch that demonstrably\\n> improves the behavior (in this case reduces per-tuple overhead without\\n> causing significant issues elsewhere), we'd be crazy not to consider it.\\n\\nAnd \\\"without causing significant issues elsewhere\\\" unfortunately\\nincludes continuing to allow pg_upgrade to work.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Nov 2019 10:04:09 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 04, 2019 at 10:04:09AM -0800, Andres Freund wrote:\\n>Hi,\\n>\\n>(I've not read the rest of this thread yet)\\n>\\n>On 2019-11-04 16:07:23 +0100, Tomas Vondra wrote:\\n>> On Mon, Nov 04, 2019 at 04:39:44PM +0300, Павел Ерёмин wrote:\\n>> > And yet, if I try to implement a similar mechanism, if successful, will my\\n>> > revision be considered?\\n>> > \\n>>\\n>> Why wouldn't it be considered? If you submit a patch that demonstrably\\n>> improves the behavior (in this case reduces per-tuple overhead without\\n>> causing significant issues elsewhere), we'd be crazy not to consider it.\\n>\\n>And \\\"without causing significant issues elsewhere\\\" unfortunately\\n>includes continuing to allow pg_upgrade to work.\\n>\\n\\nYeah. I suppose we could have a different AM implementing this, but\\nmaybe that's not possible ...\\n\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Nov 2019 19:39:18 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-11-04 19:39:18 +0100, Tomas Vondra wrote:\\n> On Mon, Nov 04, 2019 at 10:04:09AM -0800, Andres Freund wrote:\\n> > And \\\"without causing significant issues elsewhere\\\" unfortunately\\n> > includes continuing to allow pg_upgrade to work.\\n\\n> Yeah. I suppose we could have a different AM implementing this, but\\n> maybe that's not possible ...\\n\\nEntirely possible. But the amount of code duplication / unnecessary\\nbranching and the user confusion from two very similar AMs, would have\\nto be weighed against the benefits.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Nov 2019 10:44:53 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 04, 2019 at 10:44:53AM -0800, Andres Freund wrote:\\n>Hi,\\n>\\n>On 2019-11-04 19:39:18 +0100, Tomas Vondra wrote:\\n>> On Mon, Nov 04, 2019 at 10:04:09AM -0800, Andres Freund wrote:\\n>> > And \\\"without causing significant issues elsewhere\\\" unfortunately\\n>> > includes continuing to allow pg_upgrade to work.\\n>\\n>> Yeah. I suppose we could have a different AM implementing this, but\\n>> maybe that's not possible ...\\n>\\n>Entirely possible. But the amount of code duplication / unnecessary\\n>branching and the user confusion from two very similar AMs, would have\\n>to be weighed against the benefits.\\n>\\n\\nAgreed. I think code complexity is part of the trade-off. IMO it's fine\\nto hack existing heap AM initially, and only explore the separate AM if\\nthat turns out to be promising.\\n\\nregards\\n\\n-- \\nTomas Vondra http://www.2ndQuadrant.com\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services \\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Nov 2019 20:44:55 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 5, 2019 at 8:45 AM Tomas Vondra\\n<tomas.vondra@2ndquadrant.com> wrote:\\n> On Mon, Nov 04, 2019 at 10:44:53AM -0800, Andres Freund wrote:\\n> >On 2019-11-04 19:39:18 +0100, Tomas Vondra wrote:\\n> >> On Mon, Nov 04, 2019 at 10:04:09AM -0800, Andres Freund wrote:\\n> >> > And \\\"without causing significant issues elsewhere\\\" unfortunately\\n> >> > includes continuing to allow pg_upgrade to work.\\n> >\\n> >> Yeah. I suppose we could have a different AM implementing this, but\\n> >> maybe that's not possible ...\\n> >\\n> >Entirely possible. But the amount of code duplication / unnecessary\\n> >branching and the user confusion from two very similar AMs, would have\\n> >to be weighed against the benefits.\\n> >\\n>\\n> Agreed. I think code complexity is part of the trade-off. IMO it's fine\\n> to hack existing heap AM initially, and only explore the separate AM if\\n> that turns out to be promising.\\n\\nI thought a bit about how to make a minimally-diffferent-from-heap\\nnon-freezing table AM using 64 bit xids, as a thought experiment when\\ntrying to understand or explain to others what zheap is about.\\nCommitted transactions are easy (you don't have to freeze fxid\\nreferences from the ancient past because they don't wrap around so\\nthey always look old), but how do you deal with *aborted* transactions\\nwhen truncating the CLOG (given that our current rule is \\\"if it's\\nbefore the CLOG begins, it must be committed\\\")? I see three\\npossibilities: (1) don't truncate the CLOG anymore (use 64 bit\\naddressing and let it leak disk forever, like we did before commit\\n2589735d and later work), (2) freeze aborted transactions only, using\\na wraparound vacuum (and now you have failed, if the goal was to avoid\\nhaving to scan all tuples periodically to freeze stuff, though\\nadmittedly it will require less IO to freeze only the aborted\\ntransactions), (3) go and remove aborted fxid references eagerly, when\\nyou roll back (this could be done using the undo technology that we\\nhave been developing to support zheap). Another way to explain (3) is\\nthat this hypothetical table AM, let's call it \\\"yheap\\\", takes the\\nminimum parts of the zheap technology stack required to get rid of\\nvacuum-for-wraparound, without doing in-place updates or any of that\\nhard stuff. To make this really work you'd also have to deal with\\nmultixacts, which also require freezing. If that all sounds too\\ncomplicated, you're back to (2) which seems a bit weak to me. Or\\nperhaps I'm missing something?\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Nov 2019 09:34:52 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 5, 2019 at 09:34:52AM +1300, Thomas Munro wrote:\\n> On Tue, Nov 5, 2019 at 8:45 AM Tomas Vondra\\n> > Agreed. I think code complexity is part of the trade-off. IMO it's fine\\n> > to hack existing heap AM initially, and only explore the separate AM if\\n> > that turns out to be promising.\\n> \\n> I thought a bit about how to make a minimally-diffferent-from-heap\\n> non-freezing table AM using 64 bit xids, as a thought experiment when\\n> trying to understand or explain to others what zheap is about.\\n> Committed transactions are easy (you don't have to freeze fxid\\n> references from the ancient past because they don't wrap around so\\n> they always look old), but how do you deal with *aborted* transactions\\n> when truncating the CLOG (given that our current rule is \\\"if it's\\n> before the CLOG begins, it must be committed\\\")? I see three\\n> possibilities: (1) don't truncate the CLOG anymore (use 64 bit\\n> addressing and let it leak disk forever, like we did before commit\\n> 2589735d and later work), (2) freeze aborted transactions only, using\\n> a wraparound vacuum (and now you have failed, if the goal was to avoid\\n> having to scan all tuples periodically to freeze stuff, though\\n> admittedly it will require less IO to freeze only the aborted\\n> transactions), (3) go and remove aborted fxid references eagerly, when\\n> you roll back (this could be done using the undo technology that we\\n> have been developing to support zheap). Another way to explain (3) is\\n> that this hypothetical table AM, let's call it \\\"yheap\\\", takes the\\n> minimum parts of the zheap technology stack required to get rid of\\n> vacuum-for-wraparound, without doing in-place updates or any of that\\n> hard stuff. To make this really work you'd also have to deal with\\n> multixacts, which also require freezing. If that all sounds too\\n> complicated, you're back to (2) which seems a bit weak to me. Or\\n> perhaps I'm missing something?\\n\\nThe above is a very good summary of the constraints that have led to our\\ncurrent handling of XID wraparound. If we are concerned about excessive\\nvacuum freeze overhead, why is the default autovacuum_freeze_max_age =\\n200000000 so low? That causes feezing when the pg_xact directory holds\\n200 million xids or 50 megabytes of xid status?\\n\\nAs far as I understand it, we cause the database to stop writes when the\\nxid counter gets within 2 billion xids of the current transaction\\ncounter, so 200 million is only 1/10th to that limit, and even then, I\\nam not sure why we couldn't make it stop writes at 3 billion or\\nsomething. My point is that increasing the default\\nautovacuum_freeze_max_age value seems like an easy way to reduce vacuum\\nfreeze. (While, the visibility map helps avoid vacuum freeze from\\nreading all heap pages, and we still need to read all index pages.)\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n+ As you are, so once was I. As I am, so you will be. +\\n+ Ancient Roman grave inscription +\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Nov 2019 10:28:31 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 7, 2019 at 10:28 AM Bruce Momjian <bruce@momjian.us> wrote:\\n> The above is a very good summary of the constraints that have led to our\\n> current handling of XID wraparound. If we are concerned about excessive\\n> vacuum freeze overhead, why is the default autovacuum_freeze_max_age =\\n> 200000000 so low? That causes feezing when the pg_xact directory holds\\n> 200 million xids or 50 megabytes of xid status?\\n>\\n> As far as I understand it, we cause the database to stop writes when the\\n> xid counter gets within 2 billion xids of the current transaction\\n> counter, so 200 million is only 1/10th to that limit, and even then, I\\n> am not sure why we couldn't make it stop writes at 3 billion or\\n> something. My point is that increasing the default\\n> autovacuum_freeze_max_age value seems like an easy way to reduce vacuum\\n> freeze. (While, the visibility map helps avoid vacuum freeze from\\n> reading all heap pages, and we still need to read all index pages.)\\n\\nYeah, I've also wondered why this isn't higher by default, but it's a\\nsomewhat tricky topic.\\n\\nThree billion won't work, because it's deeply baked into PostgreSQL's\\narchitecture that at most two billion XIDs are used at one time. For\\ncomparison purposes, the four billion XIDs form a ring, so that from\\nthe perspective of any individual XID, half of the XIDs are in the\\nfuture and the other half are in the past. If three billion XIDs were\\nin use simultaneously, say starting with XID 5 and ending with XID\\n3,000,000,004, then XID 5 would see XID 3,000,000,004 as being the\\npast rather than the future, while XID 1,500,000,000 would (correctly)\\nsee XID 5 as in the past and XID 3,000,000,004 as in the future. So\\nXID comparison would not be transitive, which would break a lot of\\ncode. Allowing at most two billion XIDs to be in use at one time fixes\\nthis problem.\\n\\nThat doesn't mean we couldn't raise the setting. It just means that\\nthe hard limit is two billion, not four billion. But, how high should\\nwe raise it? The highest safe value depends on how many XIDs you'll\\nburn while the freezing vacuums are running, which depends on both the\\nsize of the database and the rate of XID consumption, and those values\\ncan be very different on different systems. I think most people could\\nget by with a significantly higher value, but even with the current\\nvalue I think there are probably some people who run out of XIDs, at\\nwhich point they can no longer write to the database. The higher we\\nmake the default, the more people are going to have that problem. It's\\ntrue that a lot of people only hit the limit because something has\\ngone wrong, like they've forgotten about a prepared transaction or an\\nunused replication slot, but still, on high-velocity systems you can't\\nafford to cut it too close because you're still going to be burning\\nthrough XIDs while vacuum is running.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 6 Dec 2019 08:30:28 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: 64 bit transaction id\"\n }\n]","isConversational":false}}},{"rowIdx":49392,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi, hackers\n\nAs the $Subject, does anyone have one? I'd like to refer to it, and\nwrite an example for people who is also looking for the document.\n\nThanks.\n\n-- \nAdam Lee\n\n\n","msg_date":"Fri, 1 Nov 2019 20:18:02 +0800","msg_from":"Adam Lee <ali@pivotal.io>","msg_from_op":true,"msg_subject":"Looking for a demo of extensible nodes"},{"msg_contents":"Hi,\n\nI've a basic experimental extension where I use extensible nodes. This is\nthe commit which adds the extensible node to the project:\nhttps://github.com/onderkalaci/pgcolor/commit/10cba5d02a828dbee4bc140f5e88d6c44b40e5c2\n\nHope that gives you some pointers.\n\n\nOn Fri, Nov 1, 2019 at 1:18 PM Adam Lee <ali@pivotal.io> wrote:\n\n> Hi, hackers\n>\n> As the $Subject, does anyone have one? I'd like to refer to it, and\n> write an example for people who is also looking for the document.\n>\n> Thanks.\n>\n> --\n> Adam Lee\n>\n>\n>\n\nHi,I've a basic experimental extension where I use extensible nodes. This is the commit which adds the extensible node to the project: https://github.com/onderkalaci/pgcolor/commit/10cba5d02a828dbee4bc140f5e88d6c44b40e5c2Hope that gives you some pointers.On Fri, Nov 1, 2019 at 1:18 PM Adam Lee <ali@pivotal.io> wrote:Hi, hackers\n\nAs the $Subject, does anyone have one? I'd like to refer to it, and\nwrite an example for people who is also looking for the document.\n\nThanks.\n\n-- \nAdam Lee","msg_date":"Fri, 1 Nov 2019 17:43:44 +0100","msg_from":"Onder Kalaci <onder@citusdata.com>","msg_from_op":false,"msg_subject":"Re: Looking for a demo of extensible nodes"}],"string":"[\n {\n \"msg_contents\": \"Hi, hackers\\n\\nAs the $Subject, does anyone have one? I'd like to refer to it, and\\nwrite an example for people who is also looking for the document.\\n\\nThanks.\\n\\n-- \\nAdam Lee\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 20:18:02 +0800\",\n \"msg_from\": \"Adam Lee <ali@pivotal.io>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Looking for a demo of extensible nodes\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nI've a basic experimental extension where I use extensible nodes. This is\\nthe commit which adds the extensible node to the project:\\nhttps://github.com/onderkalaci/pgcolor/commit/10cba5d02a828dbee4bc140f5e88d6c44b40e5c2\\n\\nHope that gives you some pointers.\\n\\n\\nOn Fri, Nov 1, 2019 at 1:18 PM Adam Lee <ali@pivotal.io> wrote:\\n\\n> Hi, hackers\\n>\\n> As the $Subject, does anyone have one? I'd like to refer to it, and\\n> write an example for people who is also looking for the document.\\n>\\n> Thanks.\\n>\\n> --\\n> Adam Lee\\n>\\n>\\n>\\n\\nHi,I've a basic experimental extension where I use extensible nodes. This is the commit which adds the extensible node to the project: https://github.com/onderkalaci/pgcolor/commit/10cba5d02a828dbee4bc140f5e88d6c44b40e5c2Hope that gives you some pointers.On Fri, Nov 1, 2019 at 1:18 PM Adam Lee <ali@pivotal.io> wrote:Hi, hackers\\n\\nAs the $Subject, does anyone have one? I'd like to refer to it, and\\nwrite an example for people who is also looking for the document.\\n\\nThanks.\\n\\n-- \\nAdam Lee\",\n \"msg_date\": \"Fri, 1 Nov 2019 17:43:44 +0100\",\n \"msg_from\": \"Onder Kalaci <onder@citusdata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Looking for a demo of extensible nodes\"\n }\n]","isConversational":false}}},{"rowIdx":49393,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\n\nAs per the following code, t1 is a remote table through postgres_fdw:\n\n\n test=# BEGIN;\n BEGIN\n test=# SELECT * FROM t1;\n ...\n\n test=# PREPARE TRANSACTION 'gxid1';\n ERROR: cannot prepare a transaction that modified remote tables\n\n\nI have attached a patch to the documentation that adds remote tables to\nthe list of objects where any operation prevent using a prepared\ntransaction, currently it is just notified \"operations involving\ntemporary tables or the session's temporary namespace\".\n\n\nThe second patch modify the message returned by postgres_fdw as per the\nSELECT statement above the message should be more comprehensible with:\n\n ERROR: cannot PREPARE a transaction that has operated on remote tables\n\nlike for temporary objects:\n\n ERROR: cannot PREPARE a transaction that has operated on temporary\nobjects\n\n\nBest regards,\n\n--\n\nGilles\n\n-- \nGilles Darold\nhttp://www.darold.net/","msg_date":"Fri, 1 Nov 2019 17:29:23 +0100","msg_from":"Gilles Darold <gilles@darold.net>","msg_from_op":true,"msg_subject":"[PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"On Fri, Nov 01, 2019 at 05:29:23PM +0100, Gilles Darold wrote:\n> I have attached a patch to the documentation that adds remote tables to\n> the list of objects where any operation prevent using a prepared\n> transaction, currently it is just notified \"operations involving\n> temporary tables or the session's temporary namespace\".\n\nPerhaps we had better use foreign tables for the error message and the\ndocs?\n--\nMichael","msg_date":"Sat, 2 Nov 2019 16:31:29 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Le 02/11/2019 � 08:31, Michael Paquier a �crit�:\n> On Fri, Nov 01, 2019 at 05:29:23PM +0100, Gilles Darold wrote:\n>> I have attached a patch to the documentation that adds remote tables to\n>> the list of objects where any operation prevent using a prepared\n>> transaction, currently it is just notified \"operations involving\n>> temporary tables or the session's temporary namespace\".\n> Perhaps we had better use foreign tables for the error message and the\n> docs?\n> --\n> Michael\n\n\nAgree, attached is a new version of the patches that replace word remote\nby foreign.\n\n--\n\nGilles","msg_date":"Sun, 3 Nov 2019 09:12:38 +0100","msg_from":"Gilles Darold <gillesdarold@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hi Gilles,\n\nOn Sat, Nov 2, 2019 at 1:29 AM Gilles Darold <gilles@darold.net> wrote:\n> As per the following code, t1 is a remote table through postgres_fdw:\n\n> test=# BEGIN;\n> BEGIN\n> test=# SELECT * FROM t1;\n> ...\n>\n> test=# PREPARE TRANSACTION 'gxid1';\n> ERROR: cannot prepare a transaction that modified remote tables\n\n> I have attached a patch to the documentation that adds remote tables to the list of objects where any operation prevent using a prepared transaction, currently it is just notified \"operations involving temporary tables or the session's temporary namespace\".\n\nI'm not sure that's a good idea because file_fdw works well for\nPREPARE TRANSACTION! How about adding a note about that to the\nsection of Transaction Management in the postgres_fdw documentation\nlike the attached?\n\n> The second patch modify the message returned by postgres_fdw as per the SELECT statement above the message should be more comprehensible with:\n>\n> ERROR: cannot PREPARE a transaction that has operated on remote tables\n>\n> like for temporary objects:\n>\n> ERROR: cannot PREPARE a transaction that has operated on temporary objects\n\n+1 (I too think it would be better to use \"foreign tables\" rather\nthan \"remote tables\" as pointed by Michael-san, but I think it might\nbe much better to use \"postgres_fdw foreign tables\", not just \"foreign\ntables\".)\n\nBest regards,\nEtsuro Fujita","msg_date":"Tue, 5 Nov 2019 18:35:54 +0900","msg_from":"Etsuro Fujita <etsuro.fujita@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hi Esturo,\n\nLe 05/11/2019 à 10:35, Etsuro Fujita a écrit :\n> Hi Gilles,\n>\n> On Sat, Nov 2, 2019 at 1:29 AM Gilles Darold <gilles@darold.net> wrote:\n>> As per the following code, t1 is a remote table through postgres_fdw:\n>> test=# BEGIN;\n>> BEGIN\n>> test=# SELECT * FROM t1;\n>> ...\n>>\n>> test=# PREPARE TRANSACTION 'gxid1';\n>> ERROR: cannot prepare a transaction that modified remote tables\n>> I have attached a patch to the documentation that adds remote tables to the list of objects where any operation prevent using a prepared transaction, currently it is just notified \"operations involving temporary tables or the session's temporary namespace\".\n> I'm not sure that's a good idea because file_fdw works well for\n> PREPARE TRANSACTION! How about adding a note about that to the\n> section of Transaction Management in the postgres_fdw documentation\n> like the attached?\n\n\nYou are right, read only FDW can be used. A second point in favor of\nyour remark is that this is the responsibility of the FDW implementation\nto throw an error when used with a prepared transaction and I see that\nthis is not the case for all FDW.\n\n\nI have attached a single patch that include Etsuro Fujita's patch on\npostgres_fdw documentation and mine on postgres_fdw error message with\nthe precision that it comes from postgres_fdw. The modification about\nprepared transaction in PostgreSQL documentation has been removed.\n\n\n-- \nGilles Darold","msg_date":"Tue, 5 Nov 2019 12:41:41 +0100","msg_from":"Gilles Darold <gilles@darold.net>","msg_from_op":true,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hi Gilles,\n\nOn Tue, Nov 5, 2019 at 8:41 PM Gilles Darold <gilles@darold.net> wrote:\n> I have attached a single patch that include Etsuro Fujita's patch on\n> postgres_fdw documentation and mine on postgres_fdw error message with\n> the precision that it comes from postgres_fdw. The modification about\n> prepared transaction in PostgreSQL documentation has been removed.\n\nThanks for the patch! I added the commit message. Does that make\nsense? If there are no objections, I'll apply the patch to all\nsupported branches.\n\nBest regards,\nEtsuro Fujita","msg_date":"Wed, 6 Nov 2019 12:57:10 +0900","msg_from":"Etsuro Fujita <etsuro.fujita@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"On Wed, Nov 06, 2019 at 12:57:10PM +0900, Etsuro Fujita wrote:\n> Thanks for the patch! I added the commit message. Does that make\n> sense? If there are no objections, I'll apply the patch to all\n> supported branches.\n\n\"postgres_fdw foreign tables\" sounds weird to me. Could \"foreign\ntables using postgres_fdw\" be a better wording? I am wondering as\nwell if we should not split this information into two parts: one for\nthe actual error message which only mentions foreign tables, and a\nsecond one with a hint to mention that postgres_fdw has been used.\n\nWe could have more test cases with 2PC in this module, not necessarily\nthe responsibility of this patch, but while we're on it..\n--\nMichael","msg_date":"Wed, 6 Nov 2019 13:13:12 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hi Michael-san,\n\nOn Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\n> \"postgres_fdw foreign tables\" sounds weird to me. Could \"foreign\n> tables using postgres_fdw\" be a better wording? I am wondering as\n> well if we should not split this information into two parts: one for\n> the actual error message which only mentions foreign tables, and a\n> second one with a hint to mention that postgres_fdw has been used.\n\nWe use \"postgres_fdw foreign tables\" or \"postgres_fdw tables\" in\nrelease notes, so I thought it was OK to use that in error messages as\nwell. But actually, these wordings are not suitable for error\nmessages?\n\n> We could have more test cases with 2PC in this module, not necessarily\n> the responsibility of this patch, but while we're on it..\n\nAgreed. Will do.\n\nBest regards,\nEtsuro Fujita\n\n\n","msg_date":"Wed, 6 Nov 2019 15:12:04 +0900","msg_from":"Etsuro Fujita <etsuro.fujita@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"On Wed, Nov 06, 2019 at 03:12:04PM +0900, Etsuro Fujita wrote:\n> On Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\n>> \"postgres_fdw foreign tables\" sounds weird to me. Could \"foreign\n>> tables using postgres_fdw\" be a better wording? I am wondering as\n>> well if we should not split this information into two parts: one for\n>> the actual error message which only mentions foreign tables, and a\n>> second one with a hint to mention that postgres_fdw has been used.\n> \n> We use \"postgres_fdw foreign tables\" or \"postgres_fdw tables\" in\n> release notes, so I thought it was OK to use that in error messages as\n> well. But actually, these wordings are not suitable for error\n> messages?\n\nIt is true that the docs of postgres_fdw use that and that it is used\nin some comments. Still, I found this wording a bit weird.. If you\nthink that what you have is better, I am also fine to let you have the \nfinal word, so please feel to ignore me :)\n--\nMichael","msg_date":"Wed, 6 Nov 2019 16:35:10 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hi Michael-san,\n\nOn Wed, Nov 6, 2019 at 4:35 PM Michael Paquier <michael@paquier.xyz> wrote:\n> On Wed, Nov 06, 2019 at 03:12:04PM +0900, Etsuro Fujita wrote:\n> > On Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\n> >> \"postgres_fdw foreign tables\" sounds weird to me. Could \"foreign\n> >> tables using postgres_fdw\" be a better wording? I am wondering as\n> >> well if we should not split this information into two parts: one for\n> >> the actual error message which only mentions foreign tables, and a\n> >> second one with a hint to mention that postgres_fdw has been used.\n> >\n> > We use \"postgres_fdw foreign tables\" or \"postgres_fdw tables\" in\n> > release notes, so I thought it was OK to use that in error messages as\n> > well. But actually, these wordings are not suitable for error\n> > messages?\n>\n> It is true that the docs of postgres_fdw use that and that it is used\n> in some comments. Still, I found this wording a bit weird.. If you\n> think that what you have is better, I am also fine to let you have the\n> final word, so please feel to ignore me :)\n\nI'd like to hear the opinions of others.\n\nBest regards,\nEtsuro Fujita\n\n\n","msg_date":"Wed, 6 Nov 2019 20:13:10 +0900","msg_from":"Etsuro Fujita <etsuro.fujita@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hello.\n\nAt Wed, 6 Nov 2019 20:13:10 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in \n> Hi Michael-san,\n> \n> On Wed, Nov 6, 2019 at 4:35 PM Michael Paquier <michael@paquier.xyz> wrote:\n> > On Wed, Nov 06, 2019 at 03:12:04PM +0900, Etsuro Fujita wrote:\n> > > On Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\n> > >> \"postgres_fdw foreign tables\" sounds weird to me. Could \"foreign\n> > >> tables using postgres_fdw\" be a better wording? I am wondering as\n> > >> well if we should not split this information into two parts: one for\n> > >> the actual error message which only mentions foreign tables, and a\n> > >> second one with a hint to mention that postgres_fdw has been used.\n> > >\n> > > We use \"postgres_fdw foreign tables\" or \"postgres_fdw tables\" in\n> > > release notes, so I thought it was OK to use that in error messages as\n> > > well. But actually, these wordings are not suitable for error\n> > > messages?\n> >\n> > It is true that the docs of postgres_fdw use that and that it is used\n> > in some comments. Still, I found this wording a bit weird.. If you\n> > think that what you have is better, I am also fine to let you have the\n> > final word, so please feel to ignore me :)\n> \n> I'd like to hear the opinions of others.\n\nFWIW, I see it a bit weird, too. And perhaps \"prepare\" should be in\nupper case letters. Plus, any operation including a SELECT on a\ntemporary table inhibits PREAPRE TRANSACTION, but the same on a\npostgres_fdw foreign tables is not. So the error message is rather\nwrong.\n\nA verbose alternative can be:\n\n\"cannot PREPARE a transaction that has modified data on foreign tables using postgres_fdw\"\n\nOr I think different style is OK here since the message is not of core\nbut of an extension.\n\n\"postgres_fdw doesn't support PREPARE of a transaction that has modified data on foreign tables\"\n\nThat could be shorter or simpler, of course.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Thu, 07 Nov 2019 16:10:47 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hi Kyotaro,\n\nLe 07/11/2019 à 08:10, Kyotaro Horiguchi a écrit :\n> Hello.\n>\n> At Wed, 6 Nov 2019 20:13:10 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in \n>> Hi Michael-san,\n>>\n>> On Wed, Nov 6, 2019 at 4:35 PM Michael Paquier <michael@paquier.xyz> wrote:\n>>> On Wed, Nov 06, 2019 at 03:12:04PM +0900, Etsuro Fujita wrote:\n>>>> On Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\n>>>>> \"postgres_fdw foreign tables\" sounds weird to me. Could \"foreign\n>>>>> tables using postgres_fdw\" be a better wording? I am wondering as\n>>>>> well if we should not split this information into two parts: one for\n>>>>> the actual error message which only mentions foreign tables, and a\n>>>>> second one with a hint to mention that postgres_fdw has been used.\n>>>> We use \"postgres_fdw foreign tables\" or \"postgres_fdw tables\" in\n>>>> release notes, so I thought it was OK to use that in error messages as\n>>>> well. But actually, these wordings are not suitable for error\n>>>> messages?\n>>> It is true that the docs of postgres_fdw use that and that it is used\n>>> in some comments. Still, I found this wording a bit weird.. If you\n>>> think that what you have is better, I am also fine to let you have the\n>>> final word, so please feel to ignore me :)\n>> I'd like to hear the opinions of others.\n> FWIW, I see it a bit weird, too. And perhaps \"prepare\" should be in\n> upper case letters. Plus, any operation including a SELECT on a\n> temporary table inhibits PREAPRE TRANSACTION, but the same on a\n> postgres_fdw foreign tables is not. So the error message is rather\n> wrong.\n\n\nThis is not what I've experienced, see the first message of the thread.\nA SELECT on foreign table prevent to use PREPARE TRANSACTION like with\ntemporary table. Perhaps postgres_fdw should not throw an error with\nreadonly queries on foreign tables but I guess that it is pretty hard to\nknow especially on a later PREPARE event. But maybe I'm wrong, it is not\neasy every day :-) Can you share the SQL code you have executed to allow\nPREPARE transaction after a SELECT on a postgres_fdw foreign table?\n\n\n-- \nGilles Darold\n\n\n\n\n","msg_date":"Thu, 7 Nov 2019 09:05:55 +0100","msg_from":"Gilles Darold <gilles@darold.net>","msg_from_op":true,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Horiguchi-san,\n\nOn Thu, Nov 7, 2019 at 4:11 PM Kyotaro Horiguchi\n<horikyota.ntt@gmail.com> wrote:\n> At Wed, 6 Nov 2019 20:13:10 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in\n> > On Wed, Nov 6, 2019 at 4:35 PM Michael Paquier <michael@paquier.xyz> wrote:\n> > > On Wed, Nov 06, 2019 at 03:12:04PM +0900, Etsuro Fujita wrote:\n> > > > On Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\n> > > >> \"postgres_fdw foreign tables\" sounds weird to me. Could \"foreign\n> > > >> tables using postgres_fdw\" be a better wording? I am wondering as\n> > > >> well if we should not split this information into two parts: one for\n> > > >> the actual error message which only mentions foreign tables, and a\n> > > >> second one with a hint to mention that postgres_fdw has been used.\n> > > >\n> > > > We use \"postgres_fdw foreign tables\" or \"postgres_fdw tables\" in\n> > > > release notes, so I thought it was OK to use that in error messages as\n> > > > well. But actually, these wordings are not suitable for error\n> > > > messages?\n> > >\n> > > It is true that the docs of postgres_fdw use that and that it is used\n> > > in some comments. Still, I found this wording a bit weird.. If you\n> > > think that what you have is better, I am also fine to let you have the\n> > > final word, so please feel to ignore me :)\n> >\n> > I'd like to hear the opinions of others.\n>\n> FWIW, I see it a bit weird, too.\n\nOnly two people complaining about the wording? Considering as well\nthat we use that wording in the docs and there were no complains about\nthat IIRC, I don't feel a need to change that, TBH.\n\n> And perhaps \"prepare\" should be in\n> upper case letters.\n\nSeems like a good idea.\n\n> Plus, any operation including a SELECT on a\n> temporary table inhibits PREAPRE TRANSACTION, but the same on a\n> postgres_fdw foreign tables is not. So the error message is rather\n> wrong.\n>\n> A verbose alternative can be:\n>\n> \"cannot PREPARE a transaction that has modified data on foreign tables using postgres_fdw\"\n\nI don't think that's better, because that doesn't address the original\nissue reported in this thread, as Gilles pointed out just before in\nhis email. See the commit message in the patch I posted.\n\nThanks for the comments!\n\nBest regards,\nEtsuro Fujita\n\n\n","msg_date":"Thu, 7 Nov 2019 17:20:07 +0900","msg_from":"Etsuro Fujita <etsuro.fujita@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hello Gilles. I made a silly mistake.\n\nAt Thu, 7 Nov 2019 09:05:55 +0100, Gilles Darold <gilles@darold.net> wrote in \n> > FWIW, I see it a bit weird, too. And perhaps \"prepare\" should be in\n> > upper case letters. Plus, any operation including a SELECT on a\n> > temporary table inhibits PREAPRE TRANSACTION, but the same on a\n> > postgres_fdw foreign tables is not. So the error message is rather\n> > wrong.\n> \n> \n> This is not what I've experienced, see the first message of the thread.\n> A SELECT on foreign table prevent to use PREPARE TRANSACTION like with\n> temporary table. Perhaps postgres_fdw should not throw an error with\n> readonly queries on foreign tables but I guess that it is pretty hard to\n> know especially on a later PREPARE event. But maybe I'm wrong, it is not\n> easy every day :-) Can you share the SQL code you have executed to allow\n> PREPARE transaction after a SELECT on a postgres_fdw foreign table?\n\nOooops!\n\nAfter reading this, I came to be afraid that I did something wrong,\nthen I rechecked actual behavior. Finally I found that SELECT * FROM\nforegn_tbl prohibits PREPARE TRANSACTION. I might have used a local\ntable instead of foreign tabel at the previous trial.\n\nSorry for the mistake and thank you for pointing it.\n\nSo my fixed proposals are:\n\n\"cannot PREPARE a transaction that has operated on foreign tables using postgres_fdw\"\n\nOr\n\n\"postgres_fdw doesn't support PREPARE of a transaction that has accessed foreign tables\"\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Thu, 07 Nov 2019 17:22:18 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hello, Fujita-san.\n\nAt Thu, 7 Nov 2019 17:20:07 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in \n> Only two people complaining about the wording? Considering as well\n> that we use that wording in the docs and there were no complains about\n> that IIRC, I don't feel a need to change that, TBH.\n> \n> > And perhaps \"prepare\" should be in\n> > upper case letters.\n> \n> Seems like a good idea.\n> \n> > Plus, any operation including a SELECT on a\n> > temporary table inhibits PREAPRE TRANSACTION, but the same on a\n> > postgres_fdw foreign tables is not. So the error message is rather\n> > wrong.\n> >\n> > A verbose alternative can be:\n> >\n> > \"cannot PREPARE a transaction that has modified data on foreign tables using postgres_fdw\"\n> \n> I don't think that's better, because that doesn't address the original\n> issue reported in this thread, as Gilles pointed out just before in\n> his email. See the commit message in the patch I posted.\n\n\"modified\" is my mistake as in the just posted mail. But the most\nsignificant point in the previous mail is using \"foreign tables using\npostgres_fdw\" instead of \"postgres_fdw foreign tables\". And the other\npoint is using different message from temporary tables.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Thu, 07 Nov 2019 17:27:47 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"At Thu, 07 Nov 2019 17:27:47 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in \n> \"modified\" is my mistake as in the just posted mail. But the most\n> significant point in the previous mail is using \"foreign tables using\n> postgres_fdw\" instead of \"postgres_fdw foreign tables\". And the other\n> point is using different message from temporary tables.\n\nI forgot to mention that the comment in XACT_EVENT_PRE_PREPARE\ncontains the same mistake and needs more or less the same fix.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Thu, 07 Nov 2019 17:31:29 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Horiguchi-san,\n\nOn Thu, Nov 7, 2019 at 5:28 PM Kyotaro Horiguchi\n<horikyota.ntt@gmail.com> wrote:\n> At Thu, 7 Nov 2019 17:20:07 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in\n> > Only two people complaining about the wording? Considering as well\n> > that we use that wording in the docs and there were no complains about\n> > that IIRC, I don't feel a need to change that, TBH.\n\n> But the most\n> significant point in the previous mail is using \"foreign tables using\n> postgres_fdw\" instead of \"postgres_fdw foreign tables\".\n\nOK, but as I said above, I don't feel the need to change that. How\nabout leaving it for another patch to improve the wording in that\nmessage and/or the documentation if we really need to do it.\n\n> And the other\n> point is using different message from temporary tables.\n\nYou mean we should do s/prepare/PREPARE/?\n\nBest regards,\nEtsuro Fujita\n\n\n","msg_date":"Thu, 7 Nov 2019 18:40:36 +0900","msg_from":"Etsuro Fujita <etsuro.fujita@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Horiguchi-san,\n\nOn Thu, Nov 7, 2019 at 5:31 PM Kyotaro Horiguchi\n<horikyota.ntt@gmail.com> wrote:\n> I forgot to mention that the comment in XACT_EVENT_PRE_PREPARE\n> contains the same mistake and needs more or less the same fix.\n\nGood catch! How about rewriting \"We disallow remote transactions that\nmodified anything\" in the comment simply to \"We disallow any remote\ntransactions\" or something like that? Attached is an updated patch.\nIn the patch, I did s/prepare/PREPARE/ to the error message as well,\nas you proposed.\n\nThanks again for reviewing!\n\nBest regards,\nEtsuro Fujita","msg_date":"Thu, 7 Nov 2019 19:52:42 +0900","msg_from":"Etsuro Fujita <etsuro.fujita@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"On Thu, Nov 07, 2019 at 06:40:36PM +0900, Etsuro Fujita wrote:\n> On Thu, Nov 7, 2019 at 5:28 PM Kyotaro Horiguchi\n> <horikyota.ntt@gmail.com> wrote:\n>> At Thu, 7 Nov 2019 17:20:07 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in\n>>> Only two people complaining about the wording? Considering as well\n\nThat's like.. Half the folks participating to this thread ;)\n\n>>> that we use that wording in the docs and there were no complains about\n>>> that IIRC, I don't feel a need to change that, TBH.\n>> But the most\n>> significant point in the previous mail is using \"foreign tables using\n>> postgres_fdw\" instead of \"postgres_fdw foreign tables\".\n> \n> OK, but as I said above, I don't feel the need to change that. How\n> about leaving it for another patch to improve the wording in that\n> message and/or the documentation if we really need to do it.\n\nFine by me. If I were to put a number on that, I would be around +-0,\nso I don't have an actual objection with your point of view either.\n--\nMichael","msg_date":"Fri, 8 Nov 2019 09:10:04 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Le 07/11/2019 à 11:52, Etsuro Fujita a écrit :\n> Horiguchi-san,\n>\n> On Thu, Nov 7, 2019 at 5:31 PM Kyotaro Horiguchi\n> <horikyota.ntt@gmail.com> wrote:\n>> I forgot to mention that the comment in XACT_EVENT_PRE_PREPARE\n>> contains the same mistake and needs more or less the same fix.\n> Good catch! How about rewriting \"We disallow remote transactions that\n> modified anything\" in the comment simply to \"We disallow any remote\n> transactions\" or something like that? Attached is an updated patch.\n> In the patch, I did s/prepare/PREPARE/ to the error message as well,\n> as you proposed.\n>\n> Thanks again for reviewing!\n>\n> Best regards,\n> Etsuro Fujita\n\n\nLooks good for me,\n\n-- \nGilles Darold\n\n\n\n","msg_date":"Fri, 8 Nov 2019 08:55:08 +0100","msg_from":"Gilles Darold <gilles@darold.net>","msg_from_op":true,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hi Michael-san,\n\nOn Fri, Nov 8, 2019 at 9:10 AM Michael Paquier <michael@paquier.xyz> wrote:\n> On Thu, Nov 07, 2019 at 06:40:36PM +0900, Etsuro Fujita wrote:\n> > On Thu, Nov 7, 2019 at 5:28 PM Kyotaro Horiguchi\n> > <horikyota.ntt@gmail.com> wrote:\n> >> At Thu, 7 Nov 2019 17:20:07 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in\n> >>> Only two people complaining about the wording? Considering as well\n>\n> That's like.. Half the folks participating to this thread ;)\n\nRight...\n\n> >>> that we use that wording in the docs and there were no complains about\n> >>> that IIRC, I don't feel a need to change that, TBH.\n> >> But the most\n> >> significant point in the previous mail is using \"foreign tables using\n> >> postgres_fdw\" instead of \"postgres_fdw foreign tables\".\n> >\n> > OK, but as I said above, I don't feel the need to change that. How\n> > about leaving it for another patch to improve the wording in that\n> > message and/or the documentation if we really need to do it.\n>\n> Fine by me. If I were to put a number on that, I would be around +-0,\n> so I don't have an actual objection with your point of view either.\n\nOK, pushed as-is. Thanks for reviewing!\n\nBest regards,\nEtsuro Fujita\n\n\n","msg_date":"Fri, 8 Nov 2019 17:19:34 +0900","msg_from":"Etsuro Fujita <etsuro.fujita@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hi Gilles,\n\nOn Fri, Nov 8, 2019 at 4:55 PM Gilles Darold <gilles@darold.net> wrote:\n> Le 07/11/2019 à 11:52, Etsuro Fujita a écrit :\n> > On Thu, Nov 7, 2019 at 5:31 PM Kyotaro Horiguchi\n> > <horikyota.ntt@gmail.com> wrote:\n> >> I forgot to mention that the comment in XACT_EVENT_PRE_PREPARE\n> >> contains the same mistake and needs more or less the same fix.\n> > Good catch! How about rewriting \"We disallow remote transactions that\n> > modified anything\" in the comment simply to \"We disallow any remote\n> > transactions\" or something like that? Attached is an updated patch.\n> > In the patch, I did s/prepare/PREPARE/ to the error message as well,\n> > as you proposed.\n\n> Looks good for me,\n>\n> --\n> Gilles Darold\n>\n\n\n","msg_date":"Fri, 8 Nov 2019 17:20:33 +0900","msg_from":"Etsuro Fujita <etsuro.fujita@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hi Gilles,\n\nSorry, I have sent an unfinished email.\n\nOn Fri, Nov 8, 2019 at 5:20 PM Etsuro Fujita <etsuro.fujita@gmail.com> wrote:\n> On Fri, Nov 8, 2019 at 4:55 PM Gilles Darold <gilles@darold.net> wrote:\n> > Le 07/11/2019 à 11:52, Etsuro Fujita a écrit :\n> > > On Thu, Nov 7, 2019 at 5:31 PM Kyotaro Horiguchi\n> > > <horikyota.ntt@gmail.com> wrote:\n> > >> I forgot to mention that the comment in XACT_EVENT_PRE_PREPARE\n> > >> contains the same mistake and needs more or less the same fix.\n> > > Good catch! How about rewriting \"We disallow remote transactions that\n> > > modified anything\" in the comment simply to \"We disallow any remote\n> > > transactions\" or something like that? Attached is an updated patch.\n> > > In the patch, I did s/prepare/PREPARE/ to the error message as well,\n> > > as you proposed.\n>\n> > Looks good for me,\n\nPushed after modifying the commit message a bit. Thanks!\n\nBest regards,\nEtsuro Fujita\n\n\n","msg_date":"Fri, 8 Nov 2019 17:25:52 +0900","msg_from":"Etsuro Fujita <etsuro.fujita@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"On Fri, Nov 08, 2019 at 05:25:52PM +0900, Etsuro Fujita wrote:\n> Pushed after modifying the commit message a bit. Thanks!\n\nShould we have more tests for 2PC then?\n--\nMichael","msg_date":"Fri, 8 Nov 2019 18:05:07 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Hi Michael,\n\n\nLe 08/11/2019 à 10:05, Michael Paquier a écrit :\n> On Fri, Nov 08, 2019 at 05:25:52PM +0900, Etsuro Fujita wrote:\n>> Pushed after modifying the commit message a bit. Thanks!\n> Should we have more tests for 2PC then?\n> --\n> Michael\n\n\nI don't think so. The support or not of 2PC is on foreign data wrapper\nside. In postgres_fdw contrib the error for use with 2PC is not part of\nthe test but it will be thrown anyway. I guess that a test will be\nvaluable only if there is support for readonly query.\n\n\n-- \nGilles Darold\n\n\n\n\n","msg_date":"Fri, 8 Nov 2019 10:19:01 +0100","msg_from":"Gilles Darold <gilles@darold.net>","msg_from_op":true,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"On Fri, Nov 08, 2019 at 10:19:01AM +0100, Gilles Darold wrote:\n> I don't think so. The support or not of 2PC is on foreign data wrapper\n> side. In postgres_fdw contrib the error for use with 2PC is not part of\n> the test but it will be thrown anyway. I guess that a test will be\n> valuable only if there is support for readonly query.\n\nThat's what I call a case for negative testing. We don't allow 2PC to\nbe used so there is a point in having a test to make sure of that.\nThis way, if the code in this area is refactored or changed, we still\nmake sure that 2PC is correctly prevented. My suggestion is to close\nthis gap. One point here is that this requires an alternate output\nfile because of max_prepared_transactions and there is no point in\ncreating one with all the tests of postgres_fdw in a single file as we\nhave now as it would create 8k lines of expected file bloat, so it\nwould be better to split the tests first. My 2c.\n--\nMichael","msg_date":"Sat, 9 Nov 2019 10:22:51 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"Le 09/11/2019 à 02:22, Michael Paquier a écrit :\n> On Fri, Nov 08, 2019 at 10:19:01AM +0100, Gilles Darold wrote:\n>> I don't think so. The support or not of 2PC is on foreign data wrapper\n>> side. In postgres_fdw contrib the error for use with 2PC is not part of\n>> the test but it will be thrown anyway. I guess that a test will be\n>> valuable only if there is support for readonly query.\n> That's what I call a case for negative testing. We don't allow 2PC to\n> be used so there is a point in having a test to make sure of that.\n> This way, if the code in this area is refactored or changed, we still\n> make sure that 2PC is correctly prevented. My suggestion is to close\n> this gap. One point here is that this requires an alternate output\n> file because of max_prepared_transactions and there is no point in\n> creating one with all the tests of postgres_fdw in a single file as we\n> have now as it would create 8k lines of expected file bloat, so it\n> would be better to split the tests first. My 2c.\n> --\n> Michael\n\n\nHi Michael, it looks that a separate test is not required at least for\nthis test. Here is a patch that enable the test in\ncontrib/postgres_fdw/, expected output:\n\n\n -- Make sure that 2PC is correctly prevented\n BEGIN;\n SELECT count(*) FROM ft1;\n count\n -------\n 822\n (1 row)\n\n -- Must throw an error\n PREPARE TRANSACTION 'fdw_tpc';\n ERROR: cannot PREPARE a transaction that has operated on\n postgres_fdw foreign tables\n ROLLBACK;\n WARNING: there is no transaction in progress\n\n\n\n-- \nGilles Darold\nhttp://www.darold.net/","msg_date":"Mon, 11 Nov 2019 16:43:18 +0100","msg_from":"Gilles Darold <gilles@darold.net>","msg_from_op":true,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"On Mon, Nov 11, 2019 at 04:43:18PM +0100, Gilles Darold wrote:\n> Hi Michael, it looks that a separate test is not required at least for\n> this test. Here is a patch that enable the test in\n> contrib/postgres_fdw/, expected output:\n\nIndeed, thanks for looking. I thought that the callback was called\nafter checking for max_prepared_transaction, but that's not the case.\nSo let's add at least a test case. Any objections?\n--\nMichael","msg_date":"Tue, 12 Nov 2019 09:35:03 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."},{"msg_contents":"On Tue, Nov 12, 2019 at 09:35:03AM +0900, Michael Paquier wrote:\n> Indeed, thanks for looking. I thought that the callback was called\n> after checking for max_prepared_transaction, but that's not the case.\n> So let's add at least a test case. Any objections?\n\nOkay, done.\n--\nMichael","msg_date":"Wed, 13 Nov 2019 13:53:24 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\n message."}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\n\\nAs per the following code, t1 is a remote table through postgres_fdw:\\n\\n\\n test=# BEGIN;\\n BEGIN\\n test=# SELECT * FROM t1;\\n ...\\n\\n test=# PREPARE TRANSACTION 'gxid1';\\n ERROR: cannot prepare a transaction that modified remote tables\\n\\n\\nI have attached a patch to the documentation that adds remote tables to\\nthe list of objects where any operation prevent using a prepared\\ntransaction, currently it is just notified \\\"operations involving\\ntemporary tables or the session's temporary namespace\\\".\\n\\n\\nThe second patch modify the message returned by postgres_fdw as per the\\nSELECT statement above the message should be more comprehensible with:\\n\\n ERROR: cannot PREPARE a transaction that has operated on remote tables\\n\\nlike for temporary objects:\\n\\n ERROR: cannot PREPARE a transaction that has operated on temporary\\nobjects\\n\\n\\nBest regards,\\n\\n--\\n\\nGilles\\n\\n-- \\nGilles Darold\\nhttp://www.darold.net/\",\n \"msg_date\": \"Fri, 1 Nov 2019 17:29:23 +0100\",\n \"msg_from\": \"Gilles Darold <gilles@darold.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 01, 2019 at 05:29:23PM +0100, Gilles Darold wrote:\\n> I have attached a patch to the documentation that adds remote tables to\\n> the list of objects where any operation prevent using a prepared\\n> transaction, currently it is just notified \\\"operations involving\\n> temporary tables or the session's temporary namespace\\\".\\n\\nPerhaps we had better use foreign tables for the error message and the\\ndocs?\\n--\\nMichael\",\n \"msg_date\": \"Sat, 2 Nov 2019 16:31:29 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Le 02/11/2019 � 08:31, Michael Paquier a �crit�:\\n> On Fri, Nov 01, 2019 at 05:29:23PM +0100, Gilles Darold wrote:\\n>> I have attached a patch to the documentation that adds remote tables to\\n>> the list of objects where any operation prevent using a prepared\\n>> transaction, currently it is just notified \\\"operations involving\\n>> temporary tables or the session's temporary namespace\\\".\\n> Perhaps we had better use foreign tables for the error message and the\\n> docs?\\n> --\\n> Michael\\n\\n\\nAgree, attached is a new version of the patches that replace word remote\\nby foreign.\\n\\n--\\n\\nGilles\",\n \"msg_date\": \"Sun, 3 Nov 2019 09:12:38 +0100\",\n \"msg_from\": \"Gilles Darold <gillesdarold@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hi Gilles,\\n\\nOn Sat, Nov 2, 2019 at 1:29 AM Gilles Darold <gilles@darold.net> wrote:\\n> As per the following code, t1 is a remote table through postgres_fdw:\\n\\n> test=# BEGIN;\\n> BEGIN\\n> test=# SELECT * FROM t1;\\n> ...\\n>\\n> test=# PREPARE TRANSACTION 'gxid1';\\n> ERROR: cannot prepare a transaction that modified remote tables\\n\\n> I have attached a patch to the documentation that adds remote tables to the list of objects where any operation prevent using a prepared transaction, currently it is just notified \\\"operations involving temporary tables or the session's temporary namespace\\\".\\n\\nI'm not sure that's a good idea because file_fdw works well for\\nPREPARE TRANSACTION! How about adding a note about that to the\\nsection of Transaction Management in the postgres_fdw documentation\\nlike the attached?\\n\\n> The second patch modify the message returned by postgres_fdw as per the SELECT statement above the message should be more comprehensible with:\\n>\\n> ERROR: cannot PREPARE a transaction that has operated on remote tables\\n>\\n> like for temporary objects:\\n>\\n> ERROR: cannot PREPARE a transaction that has operated on temporary objects\\n\\n+1 (I too think it would be better to use \\\"foreign tables\\\" rather\\nthan \\\"remote tables\\\" as pointed by Michael-san, but I think it might\\nbe much better to use \\\"postgres_fdw foreign tables\\\", not just \\\"foreign\\ntables\\\".)\\n\\nBest regards,\\nEtsuro Fujita\",\n \"msg_date\": \"Tue, 5 Nov 2019 18:35:54 +0900\",\n \"msg_from\": \"Etsuro Fujita <etsuro.fujita@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hi Esturo,\\n\\nLe 05/11/2019 à 10:35, Etsuro Fujita a écrit :\\n> Hi Gilles,\\n>\\n> On Sat, Nov 2, 2019 at 1:29 AM Gilles Darold <gilles@darold.net> wrote:\\n>> As per the following code, t1 is a remote table through postgres_fdw:\\n>> test=# BEGIN;\\n>> BEGIN\\n>> test=# SELECT * FROM t1;\\n>> ...\\n>>\\n>> test=# PREPARE TRANSACTION 'gxid1';\\n>> ERROR: cannot prepare a transaction that modified remote tables\\n>> I have attached a patch to the documentation that adds remote tables to the list of objects where any operation prevent using a prepared transaction, currently it is just notified \\\"operations involving temporary tables or the session's temporary namespace\\\".\\n> I'm not sure that's a good idea because file_fdw works well for\\n> PREPARE TRANSACTION! How about adding a note about that to the\\n> section of Transaction Management in the postgres_fdw documentation\\n> like the attached?\\n\\n\\nYou are right, read only FDW can be used. A second point in favor of\\nyour remark is that this is the responsibility of the FDW implementation\\nto throw an error when used with a prepared transaction and I see that\\nthis is not the case for all FDW.\\n\\n\\nI have attached a single patch that include Etsuro Fujita's patch on\\npostgres_fdw documentation and mine on postgres_fdw error message with\\nthe precision that it comes from postgres_fdw. The modification about\\nprepared transaction in PostgreSQL documentation has been removed.\\n\\n\\n-- \\nGilles Darold\",\n \"msg_date\": \"Tue, 5 Nov 2019 12:41:41 +0100\",\n \"msg_from\": \"Gilles Darold <gilles@darold.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hi Gilles,\\n\\nOn Tue, Nov 5, 2019 at 8:41 PM Gilles Darold <gilles@darold.net> wrote:\\n> I have attached a single patch that include Etsuro Fujita's patch on\\n> postgres_fdw documentation and mine on postgres_fdw error message with\\n> the precision that it comes from postgres_fdw. The modification about\\n> prepared transaction in PostgreSQL documentation has been removed.\\n\\nThanks for the patch! I added the commit message. Does that make\\nsense? If there are no objections, I'll apply the patch to all\\nsupported branches.\\n\\nBest regards,\\nEtsuro Fujita\",\n \"msg_date\": \"Wed, 6 Nov 2019 12:57:10 +0900\",\n \"msg_from\": \"Etsuro Fujita <etsuro.fujita@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 06, 2019 at 12:57:10PM +0900, Etsuro Fujita wrote:\\n> Thanks for the patch! I added the commit message. Does that make\\n> sense? If there are no objections, I'll apply the patch to all\\n> supported branches.\\n\\n\\\"postgres_fdw foreign tables\\\" sounds weird to me. Could \\\"foreign\\ntables using postgres_fdw\\\" be a better wording? I am wondering as\\nwell if we should not split this information into two parts: one for\\nthe actual error message which only mentions foreign tables, and a\\nsecond one with a hint to mention that postgres_fdw has been used.\\n\\nWe could have more test cases with 2PC in this module, not necessarily\\nthe responsibility of this patch, but while we're on it..\\n--\\nMichael\",\n \"msg_date\": \"Wed, 6 Nov 2019 13:13:12 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hi Michael-san,\\n\\nOn Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> \\\"postgres_fdw foreign tables\\\" sounds weird to me. Could \\\"foreign\\n> tables using postgres_fdw\\\" be a better wording? I am wondering as\\n> well if we should not split this information into two parts: one for\\n> the actual error message which only mentions foreign tables, and a\\n> second one with a hint to mention that postgres_fdw has been used.\\n\\nWe use \\\"postgres_fdw foreign tables\\\" or \\\"postgres_fdw tables\\\" in\\nrelease notes, so I thought it was OK to use that in error messages as\\nwell. But actually, these wordings are not suitable for error\\nmessages?\\n\\n> We could have more test cases with 2PC in this module, not necessarily\\n> the responsibility of this patch, but while we're on it..\\n\\nAgreed. Will do.\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Nov 2019 15:12:04 +0900\",\n \"msg_from\": \"Etsuro Fujita <etsuro.fujita@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 06, 2019 at 03:12:04PM +0900, Etsuro Fujita wrote:\\n> On Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\\n>> \\\"postgres_fdw foreign tables\\\" sounds weird to me. Could \\\"foreign\\n>> tables using postgres_fdw\\\" be a better wording? I am wondering as\\n>> well if we should not split this information into two parts: one for\\n>> the actual error message which only mentions foreign tables, and a\\n>> second one with a hint to mention that postgres_fdw has been used.\\n> \\n> We use \\\"postgres_fdw foreign tables\\\" or \\\"postgres_fdw tables\\\" in\\n> release notes, so I thought it was OK to use that in error messages as\\n> well. But actually, these wordings are not suitable for error\\n> messages?\\n\\nIt is true that the docs of postgres_fdw use that and that it is used\\nin some comments. Still, I found this wording a bit weird.. If you\\nthink that what you have is better, I am also fine to let you have the \\nfinal word, so please feel to ignore me :)\\n--\\nMichael\",\n \"msg_date\": \"Wed, 6 Nov 2019 16:35:10 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hi Michael-san,\\n\\nOn Wed, Nov 6, 2019 at 4:35 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> On Wed, Nov 06, 2019 at 03:12:04PM +0900, Etsuro Fujita wrote:\\n> > On Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> >> \\\"postgres_fdw foreign tables\\\" sounds weird to me. Could \\\"foreign\\n> >> tables using postgres_fdw\\\" be a better wording? I am wondering as\\n> >> well if we should not split this information into two parts: one for\\n> >> the actual error message which only mentions foreign tables, and a\\n> >> second one with a hint to mention that postgres_fdw has been used.\\n> >\\n> > We use \\\"postgres_fdw foreign tables\\\" or \\\"postgres_fdw tables\\\" in\\n> > release notes, so I thought it was OK to use that in error messages as\\n> > well. But actually, these wordings are not suitable for error\\n> > messages?\\n>\\n> It is true that the docs of postgres_fdw use that and that it is used\\n> in some comments. Still, I found this wording a bit weird.. If you\\n> think that what you have is better, I am also fine to let you have the\\n> final word, so please feel to ignore me :)\\n\\nI'd like to hear the opinions of others.\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Nov 2019 20:13:10 +0900\",\n \"msg_from\": \"Etsuro Fujita <etsuro.fujita@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hello.\\n\\nAt Wed, 6 Nov 2019 20:13:10 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in \\n> Hi Michael-san,\\n> \\n> On Wed, Nov 6, 2019 at 4:35 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> > On Wed, Nov 06, 2019 at 03:12:04PM +0900, Etsuro Fujita wrote:\\n> > > On Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> > >> \\\"postgres_fdw foreign tables\\\" sounds weird to me. Could \\\"foreign\\n> > >> tables using postgres_fdw\\\" be a better wording? I am wondering as\\n> > >> well if we should not split this information into two parts: one for\\n> > >> the actual error message which only mentions foreign tables, and a\\n> > >> second one with a hint to mention that postgres_fdw has been used.\\n> > >\\n> > > We use \\\"postgres_fdw foreign tables\\\" or \\\"postgres_fdw tables\\\" in\\n> > > release notes, so I thought it was OK to use that in error messages as\\n> > > well. But actually, these wordings are not suitable for error\\n> > > messages?\\n> >\\n> > It is true that the docs of postgres_fdw use that and that it is used\\n> > in some comments. Still, I found this wording a bit weird.. If you\\n> > think that what you have is better, I am also fine to let you have the\\n> > final word, so please feel to ignore me :)\\n> \\n> I'd like to hear the opinions of others.\\n\\nFWIW, I see it a bit weird, too. And perhaps \\\"prepare\\\" should be in\\nupper case letters. Plus, any operation including a SELECT on a\\ntemporary table inhibits PREAPRE TRANSACTION, but the same on a\\npostgres_fdw foreign tables is not. So the error message is rather\\nwrong.\\n\\nA verbose alternative can be:\\n\\n\\\"cannot PREPARE a transaction that has modified data on foreign tables using postgres_fdw\\\"\\n\\nOr I think different style is OK here since the message is not of core\\nbut of an extension.\\n\\n\\\"postgres_fdw doesn't support PREPARE of a transaction that has modified data on foreign tables\\\"\\n\\nThat could be shorter or simpler, of course.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Nov 2019 16:10:47 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hi Kyotaro,\\n\\nLe 07/11/2019 à 08:10, Kyotaro Horiguchi a écrit :\\n> Hello.\\n>\\n> At Wed, 6 Nov 2019 20:13:10 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in \\n>> Hi Michael-san,\\n>>\\n>> On Wed, Nov 6, 2019 at 4:35 PM Michael Paquier <michael@paquier.xyz> wrote:\\n>>> On Wed, Nov 06, 2019 at 03:12:04PM +0900, Etsuro Fujita wrote:\\n>>>> On Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\\n>>>>> \\\"postgres_fdw foreign tables\\\" sounds weird to me. Could \\\"foreign\\n>>>>> tables using postgres_fdw\\\" be a better wording? I am wondering as\\n>>>>> well if we should not split this information into two parts: one for\\n>>>>> the actual error message which only mentions foreign tables, and a\\n>>>>> second one with a hint to mention that postgres_fdw has been used.\\n>>>> We use \\\"postgres_fdw foreign tables\\\" or \\\"postgres_fdw tables\\\" in\\n>>>> release notes, so I thought it was OK to use that in error messages as\\n>>>> well. But actually, these wordings are not suitable for error\\n>>>> messages?\\n>>> It is true that the docs of postgres_fdw use that and that it is used\\n>>> in some comments. Still, I found this wording a bit weird.. If you\\n>>> think that what you have is better, I am also fine to let you have the\\n>>> final word, so please feel to ignore me :)\\n>> I'd like to hear the opinions of others.\\n> FWIW, I see it a bit weird, too. And perhaps \\\"prepare\\\" should be in\\n> upper case letters. Plus, any operation including a SELECT on a\\n> temporary table inhibits PREAPRE TRANSACTION, but the same on a\\n> postgres_fdw foreign tables is not. So the error message is rather\\n> wrong.\\n\\n\\nThis is not what I've experienced, see the first message of the thread.\\nA SELECT on foreign table prevent to use PREPARE TRANSACTION like with\\ntemporary table. Perhaps postgres_fdw should not throw an error with\\nreadonly queries on foreign tables but I guess that it is pretty hard to\\nknow especially on a later PREPARE event. But maybe I'm wrong, it is not\\neasy every day :-) Can you share the SQL code you have executed to allow\\nPREPARE transaction after a SELECT on a postgres_fdw foreign table?\\n\\n\\n-- \\nGilles Darold\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Nov 2019 09:05:55 +0100\",\n \"msg_from\": \"Gilles Darold <gilles@darold.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Horiguchi-san,\\n\\nOn Thu, Nov 7, 2019 at 4:11 PM Kyotaro Horiguchi\\n<horikyota.ntt@gmail.com> wrote:\\n> At Wed, 6 Nov 2019 20:13:10 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in\\n> > On Wed, Nov 6, 2019 at 4:35 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> > > On Wed, Nov 06, 2019 at 03:12:04PM +0900, Etsuro Fujita wrote:\\n> > > > On Wed, Nov 6, 2019 at 1:13 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> > > >> \\\"postgres_fdw foreign tables\\\" sounds weird to me. Could \\\"foreign\\n> > > >> tables using postgres_fdw\\\" be a better wording? I am wondering as\\n> > > >> well if we should not split this information into two parts: one for\\n> > > >> the actual error message which only mentions foreign tables, and a\\n> > > >> second one with a hint to mention that postgres_fdw has been used.\\n> > > >\\n> > > > We use \\\"postgres_fdw foreign tables\\\" or \\\"postgres_fdw tables\\\" in\\n> > > > release notes, so I thought it was OK to use that in error messages as\\n> > > > well. But actually, these wordings are not suitable for error\\n> > > > messages?\\n> > >\\n> > > It is true that the docs of postgres_fdw use that and that it is used\\n> > > in some comments. Still, I found this wording a bit weird.. If you\\n> > > think that what you have is better, I am also fine to let you have the\\n> > > final word, so please feel to ignore me :)\\n> >\\n> > I'd like to hear the opinions of others.\\n>\\n> FWIW, I see it a bit weird, too.\\n\\nOnly two people complaining about the wording? Considering as well\\nthat we use that wording in the docs and there were no complains about\\nthat IIRC, I don't feel a need to change that, TBH.\\n\\n> And perhaps \\\"prepare\\\" should be in\\n> upper case letters.\\n\\nSeems like a good idea.\\n\\n> Plus, any operation including a SELECT on a\\n> temporary table inhibits PREAPRE TRANSACTION, but the same on a\\n> postgres_fdw foreign tables is not. So the error message is rather\\n> wrong.\\n>\\n> A verbose alternative can be:\\n>\\n> \\\"cannot PREPARE a transaction that has modified data on foreign tables using postgres_fdw\\\"\\n\\nI don't think that's better, because that doesn't address the original\\nissue reported in this thread, as Gilles pointed out just before in\\nhis email. See the commit message in the patch I posted.\\n\\nThanks for the comments!\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Nov 2019 17:20:07 +0900\",\n \"msg_from\": \"Etsuro Fujita <etsuro.fujita@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hello Gilles. I made a silly mistake.\\n\\nAt Thu, 7 Nov 2019 09:05:55 +0100, Gilles Darold <gilles@darold.net> wrote in \\n> > FWIW, I see it a bit weird, too. And perhaps \\\"prepare\\\" should be in\\n> > upper case letters. Plus, any operation including a SELECT on a\\n> > temporary table inhibits PREAPRE TRANSACTION, but the same on a\\n> > postgres_fdw foreign tables is not. So the error message is rather\\n> > wrong.\\n> \\n> \\n> This is not what I've experienced, see the first message of the thread.\\n> A SELECT on foreign table prevent to use PREPARE TRANSACTION like with\\n> temporary table. Perhaps postgres_fdw should not throw an error with\\n> readonly queries on foreign tables but I guess that it is pretty hard to\\n> know especially on a later PREPARE event. But maybe I'm wrong, it is not\\n> easy every day :-) Can you share the SQL code you have executed to allow\\n> PREPARE transaction after a SELECT on a postgres_fdw foreign table?\\n\\nOooops!\\n\\nAfter reading this, I came to be afraid that I did something wrong,\\nthen I rechecked actual behavior. Finally I found that SELECT * FROM\\nforegn_tbl prohibits PREPARE TRANSACTION. I might have used a local\\ntable instead of foreign tabel at the previous trial.\\n\\nSorry for the mistake and thank you for pointing it.\\n\\nSo my fixed proposals are:\\n\\n\\\"cannot PREPARE a transaction that has operated on foreign tables using postgres_fdw\\\"\\n\\nOr\\n\\n\\\"postgres_fdw doesn't support PREPARE of a transaction that has accessed foreign tables\\\"\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Nov 2019 17:22:18 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hello, Fujita-san.\\n\\nAt Thu, 7 Nov 2019 17:20:07 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in \\n> Only two people complaining about the wording? Considering as well\\n> that we use that wording in the docs and there were no complains about\\n> that IIRC, I don't feel a need to change that, TBH.\\n> \\n> > And perhaps \\\"prepare\\\" should be in\\n> > upper case letters.\\n> \\n> Seems like a good idea.\\n> \\n> > Plus, any operation including a SELECT on a\\n> > temporary table inhibits PREAPRE TRANSACTION, but the same on a\\n> > postgres_fdw foreign tables is not. So the error message is rather\\n> > wrong.\\n> >\\n> > A verbose alternative can be:\\n> >\\n> > \\\"cannot PREPARE a transaction that has modified data on foreign tables using postgres_fdw\\\"\\n> \\n> I don't think that's better, because that doesn't address the original\\n> issue reported in this thread, as Gilles pointed out just before in\\n> his email. See the commit message in the patch I posted.\\n\\n\\\"modified\\\" is my mistake as in the just posted mail. But the most\\nsignificant point in the previous mail is using \\\"foreign tables using\\npostgres_fdw\\\" instead of \\\"postgres_fdw foreign tables\\\". And the other\\npoint is using different message from temporary tables.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Nov 2019 17:27:47 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"At Thu, 07 Nov 2019 17:27:47 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in \\n> \\\"modified\\\" is my mistake as in the just posted mail. But the most\\n> significant point in the previous mail is using \\\"foreign tables using\\n> postgres_fdw\\\" instead of \\\"postgres_fdw foreign tables\\\". And the other\\n> point is using different message from temporary tables.\\n\\nI forgot to mention that the comment in XACT_EVENT_PRE_PREPARE\\ncontains the same mistake and needs more or less the same fix.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Nov 2019 17:31:29 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Horiguchi-san,\\n\\nOn Thu, Nov 7, 2019 at 5:28 PM Kyotaro Horiguchi\\n<horikyota.ntt@gmail.com> wrote:\\n> At Thu, 7 Nov 2019 17:20:07 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in\\n> > Only two people complaining about the wording? Considering as well\\n> > that we use that wording in the docs and there were no complains about\\n> > that IIRC, I don't feel a need to change that, TBH.\\n\\n> But the most\\n> significant point in the previous mail is using \\\"foreign tables using\\n> postgres_fdw\\\" instead of \\\"postgres_fdw foreign tables\\\".\\n\\nOK, but as I said above, I don't feel the need to change that. How\\nabout leaving it for another patch to improve the wording in that\\nmessage and/or the documentation if we really need to do it.\\n\\n> And the other\\n> point is using different message from temporary tables.\\n\\nYou mean we should do s/prepare/PREPARE/?\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Nov 2019 18:40:36 +0900\",\n \"msg_from\": \"Etsuro Fujita <etsuro.fujita@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Horiguchi-san,\\n\\nOn Thu, Nov 7, 2019 at 5:31 PM Kyotaro Horiguchi\\n<horikyota.ntt@gmail.com> wrote:\\n> I forgot to mention that the comment in XACT_EVENT_PRE_PREPARE\\n> contains the same mistake and needs more or less the same fix.\\n\\nGood catch! How about rewriting \\\"We disallow remote transactions that\\nmodified anything\\\" in the comment simply to \\\"We disallow any remote\\ntransactions\\\" or something like that? Attached is an updated patch.\\nIn the patch, I did s/prepare/PREPARE/ to the error message as well,\\nas you proposed.\\n\\nThanks again for reviewing!\\n\\nBest regards,\\nEtsuro Fujita\",\n \"msg_date\": \"Thu, 7 Nov 2019 19:52:42 +0900\",\n \"msg_from\": \"Etsuro Fujita <etsuro.fujita@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 07, 2019 at 06:40:36PM +0900, Etsuro Fujita wrote:\\n> On Thu, Nov 7, 2019 at 5:28 PM Kyotaro Horiguchi\\n> <horikyota.ntt@gmail.com> wrote:\\n>> At Thu, 7 Nov 2019 17:20:07 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in\\n>>> Only two people complaining about the wording? Considering as well\\n\\nThat's like.. Half the folks participating to this thread ;)\\n\\n>>> that we use that wording in the docs and there were no complains about\\n>>> that IIRC, I don't feel a need to change that, TBH.\\n>> But the most\\n>> significant point in the previous mail is using \\\"foreign tables using\\n>> postgres_fdw\\\" instead of \\\"postgres_fdw foreign tables\\\".\\n> \\n> OK, but as I said above, I don't feel the need to change that. How\\n> about leaving it for another patch to improve the wording in that\\n> message and/or the documentation if we really need to do it.\\n\\nFine by me. If I were to put a number on that, I would be around +-0,\\nso I don't have an actual objection with your point of view either.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 8 Nov 2019 09:10:04 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Le 07/11/2019 à 11:52, Etsuro Fujita a écrit :\\n> Horiguchi-san,\\n>\\n> On Thu, Nov 7, 2019 at 5:31 PM Kyotaro Horiguchi\\n> <horikyota.ntt@gmail.com> wrote:\\n>> I forgot to mention that the comment in XACT_EVENT_PRE_PREPARE\\n>> contains the same mistake and needs more or less the same fix.\\n> Good catch! How about rewriting \\\"We disallow remote transactions that\\n> modified anything\\\" in the comment simply to \\\"We disallow any remote\\n> transactions\\\" or something like that? Attached is an updated patch.\\n> In the patch, I did s/prepare/PREPARE/ to the error message as well,\\n> as you proposed.\\n>\\n> Thanks again for reviewing!\\n>\\n> Best regards,\\n> Etsuro Fujita\\n\\n\\nLooks good for me,\\n\\n-- \\nGilles Darold\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 08:55:08 +0100\",\n \"msg_from\": \"Gilles Darold <gilles@darold.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hi Michael-san,\\n\\nOn Fri, Nov 8, 2019 at 9:10 AM Michael Paquier <michael@paquier.xyz> wrote:\\n> On Thu, Nov 07, 2019 at 06:40:36PM +0900, Etsuro Fujita wrote:\\n> > On Thu, Nov 7, 2019 at 5:28 PM Kyotaro Horiguchi\\n> > <horikyota.ntt@gmail.com> wrote:\\n> >> At Thu, 7 Nov 2019 17:20:07 +0900, Etsuro Fujita <etsuro.fujita@gmail.com> wrote in\\n> >>> Only two people complaining about the wording? Considering as well\\n>\\n> That's like.. Half the folks participating to this thread ;)\\n\\nRight...\\n\\n> >>> that we use that wording in the docs and there were no complains about\\n> >>> that IIRC, I don't feel a need to change that, TBH.\\n> >> But the most\\n> >> significant point in the previous mail is using \\\"foreign tables using\\n> >> postgres_fdw\\\" instead of \\\"postgres_fdw foreign tables\\\".\\n> >\\n> > OK, but as I said above, I don't feel the need to change that. How\\n> > about leaving it for another patch to improve the wording in that\\n> > message and/or the documentation if we really need to do it.\\n>\\n> Fine by me. If I were to put a number on that, I would be around +-0,\\n> so I don't have an actual objection with your point of view either.\\n\\nOK, pushed as-is. Thanks for reviewing!\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 17:19:34 +0900\",\n \"msg_from\": \"Etsuro Fujita <etsuro.fujita@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hi Gilles,\\n\\nOn Fri, Nov 8, 2019 at 4:55 PM Gilles Darold <gilles@darold.net> wrote:\\n> Le 07/11/2019 à 11:52, Etsuro Fujita a écrit :\\n> > On Thu, Nov 7, 2019 at 5:31 PM Kyotaro Horiguchi\\n> > <horikyota.ntt@gmail.com> wrote:\\n> >> I forgot to mention that the comment in XACT_EVENT_PRE_PREPARE\\n> >> contains the same mistake and needs more or less the same fix.\\n> > Good catch! How about rewriting \\\"We disallow remote transactions that\\n> > modified anything\\\" in the comment simply to \\\"We disallow any remote\\n> > transactions\\\" or something like that? Attached is an updated patch.\\n> > In the patch, I did s/prepare/PREPARE/ to the error message as well,\\n> > as you proposed.\\n\\n> Looks good for me,\\n>\\n> --\\n> Gilles Darold\\n>\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 17:20:33 +0900\",\n \"msg_from\": \"Etsuro Fujita <etsuro.fujita@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hi Gilles,\\n\\nSorry, I have sent an unfinished email.\\n\\nOn Fri, Nov 8, 2019 at 5:20 PM Etsuro Fujita <etsuro.fujita@gmail.com> wrote:\\n> On Fri, Nov 8, 2019 at 4:55 PM Gilles Darold <gilles@darold.net> wrote:\\n> > Le 07/11/2019 à 11:52, Etsuro Fujita a écrit :\\n> > > On Thu, Nov 7, 2019 at 5:31 PM Kyotaro Horiguchi\\n> > > <horikyota.ntt@gmail.com> wrote:\\n> > >> I forgot to mention that the comment in XACT_EVENT_PRE_PREPARE\\n> > >> contains the same mistake and needs more or less the same fix.\\n> > > Good catch! How about rewriting \\\"We disallow remote transactions that\\n> > > modified anything\\\" in the comment simply to \\\"We disallow any remote\\n> > > transactions\\\" or something like that? Attached is an updated patch.\\n> > > In the patch, I did s/prepare/PREPARE/ to the error message as well,\\n> > > as you proposed.\\n>\\n> > Looks good for me,\\n\\nPushed after modifying the commit message a bit. Thanks!\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 17:25:52 +0900\",\n \"msg_from\": \"Etsuro Fujita <etsuro.fujita@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 08, 2019 at 05:25:52PM +0900, Etsuro Fujita wrote:\\n> Pushed after modifying the commit message a bit. Thanks!\\n\\nShould we have more tests for 2PC then?\\n--\\nMichael\",\n \"msg_date\": \"Fri, 8 Nov 2019 18:05:07 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Hi Michael,\\n\\n\\nLe 08/11/2019 à 10:05, Michael Paquier a écrit :\\n> On Fri, Nov 08, 2019 at 05:25:52PM +0900, Etsuro Fujita wrote:\\n>> Pushed after modifying the commit message a bit. Thanks!\\n> Should we have more tests for 2PC then?\\n> --\\n> Michael\\n\\n\\nI don't think so. The support or not of 2PC is on foreign data wrapper\\nside. In postgres_fdw contrib the error for use with 2PC is not part of\\nthe test but it will be thrown anyway. I guess that a test will be\\nvaluable only if there is support for readonly query.\\n\\n\\n-- \\nGilles Darold\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 10:19:01 +0100\",\n \"msg_from\": \"Gilles Darold <gilles@darold.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 08, 2019 at 10:19:01AM +0100, Gilles Darold wrote:\\n> I don't think so. The support or not of 2PC is on foreign data wrapper\\n> side. In postgres_fdw contrib the error for use with 2PC is not part of\\n> the test but it will be thrown anyway. I guess that a test will be\\n> valuable only if there is support for readonly query.\\n\\nThat's what I call a case for negative testing. We don't allow 2PC to\\nbe used so there is a point in having a test to make sure of that.\\nThis way, if the code in this area is refactored or changed, we still\\nmake sure that 2PC is correctly prevented. My suggestion is to close\\nthis gap. One point here is that this requires an alternate output\\nfile because of max_prepared_transactions and there is no point in\\ncreating one with all the tests of postgres_fdw in a single file as we\\nhave now as it would create 8k lines of expected file bloat, so it\\nwould be better to split the tests first. My 2c.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 9 Nov 2019 10:22:51 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"Le 09/11/2019 à 02:22, Michael Paquier a écrit :\\n> On Fri, Nov 08, 2019 at 10:19:01AM +0100, Gilles Darold wrote:\\n>> I don't think so. The support or not of 2PC is on foreign data wrapper\\n>> side. In postgres_fdw contrib the error for use with 2PC is not part of\\n>> the test but it will be thrown anyway. I guess that a test will be\\n>> valuable only if there is support for readonly query.\\n> That's what I call a case for negative testing. We don't allow 2PC to\\n> be used so there is a point in having a test to make sure of that.\\n> This way, if the code in this area is refactored or changed, we still\\n> make sure that 2PC is correctly prevented. My suggestion is to close\\n> this gap. One point here is that this requires an alternate output\\n> file because of max_prepared_transactions and there is no point in\\n> creating one with all the tests of postgres_fdw in a single file as we\\n> have now as it would create 8k lines of expected file bloat, so it\\n> would be better to split the tests first. My 2c.\\n> --\\n> Michael\\n\\n\\nHi Michael, it looks that a separate test is not required at least for\\nthis test. Here is a patch that enable the test in\\ncontrib/postgres_fdw/, expected output:\\n\\n\\n -- Make sure that 2PC is correctly prevented\\n BEGIN;\\n SELECT count(*) FROM ft1;\\n count\\n -------\\n 822\\n (1 row)\\n\\n -- Must throw an error\\n PREPARE TRANSACTION 'fdw_tpc';\\n ERROR: cannot PREPARE a transaction that has operated on\\n postgres_fdw foreign tables\\n ROLLBACK;\\n WARNING: there is no transaction in progress\\n\\n\\n\\n-- \\nGilles Darold\\nhttp://www.darold.net/\",\n \"msg_date\": \"Mon, 11 Nov 2019 16:43:18 +0100\",\n \"msg_from\": \"Gilles Darold <gilles@darold.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 11, 2019 at 04:43:18PM +0100, Gilles Darold wrote:\\n> Hi Michael, it looks that a separate test is not required at least for\\n> this test. Here is a patch that enable the test in\\n> contrib/postgres_fdw/, expected output:\\n\\nIndeed, thanks for looking. I thought that the callback was called\\nafter checking for max_prepared_transaction, but that's not the case.\\nSo let's add at least a test case. Any objections?\\n--\\nMichael\",\n \"msg_date\": \"Tue, 12 Nov 2019 09:35:03 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 12, 2019 at 09:35:03AM +0900, Michael Paquier wrote:\\n> Indeed, thanks for looking. I thought that the callback was called\\n> after checking for max_prepared_transaction, but that's not the case.\\n> So let's add at least a test case. Any objections?\\n\\nOkay, done.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 13 Nov 2019 13:53:24 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH][DOC] Fix for PREPARE TRANSACTION doc and postgres_fdw\\n message.\"\n }\n]","isConversational":false}}},{"rowIdx":49394,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"It would be useful to have CREATE INDEX CONCURRENTLY be ignored by\nvacuuming's OldestXmin. Frequently in OLTP scenarios, CIC transactions\nare severely disruptive because they are the only long-running\ntransactions in the system, and VACUUM has to keep rows only for their\nsake, pointlessly. The motivation for this change seems well justified\nto me (but feel free to argue if you think otherwise).\n\nSo the question is how to implement this. Here's a very small patch for\nit:\n\ndiff --git a/src/backend/commands/indexcmds.c b/src/backend/commands/indexcmds.c\nindex 374e2d0efe..9081dfe384 100644\n--- a/src/backend/commands/indexcmds.c\n+++ b/src/backend/commands/indexcmds.c\n@@ -532,6 +532,12 @@ DefineIndex(Oid relationId,\n \t\t\t\t errmsg(\"cannot use more than %d columns in an index\",\n \t\t\t\t\t\tINDEX_MAX_KEYS)));\n \n+\tif (stmt->concurrent && !IsTransactionBlock())\n+\t{\n+\t\tAssert(GetCurrentTransactionIdIfAny() == InvalidTransactionId);\n+\t\tMyPgXact->vacuumFlags |= PROC_IN_VACUUM;\n+\t}\n+\n \t/*\n \t * Only SELECT ... FOR UPDATE/SHARE are allowed while doing a standard\n \t * index build; but for concurrent builds we allow INSERT/UPDATE/DELETE\n\nThere's an obvious flaw, which is that this doesn't consider expressions\nin partial indexes and column definitions. That's moderately easy to\nfix. But there are less obvious flaws, such as: is it possible that\nCIC's xmin is required for other reasons? (such as access to catalogs,\nwhich get cleaned by concurrent vacuum) If it is, can we fix that\nproblem by keeping track of a separate Xmin for catalog vacuuming\npurposes? (We already have catalog_xmin for replication slots, so this\nis not completely ridiculous I think ...)\n\n-- \n�lvaro Herrera\n\n\n","msg_date":"Fri, 1 Nov 2019 17:33:10 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":true,"msg_subject":"excluding CREATE INDEX CONCURRENTLY from OldestXmin"}],"string":"[\n {\n \"msg_contents\": \"It would be useful to have CREATE INDEX CONCURRENTLY be ignored by\\nvacuuming's OldestXmin. Frequently in OLTP scenarios, CIC transactions\\nare severely disruptive because they are the only long-running\\ntransactions in the system, and VACUUM has to keep rows only for their\\nsake, pointlessly. The motivation for this change seems well justified\\nto me (but feel free to argue if you think otherwise).\\n\\nSo the question is how to implement this. Here's a very small patch for\\nit:\\n\\ndiff --git a/src/backend/commands/indexcmds.c b/src/backend/commands/indexcmds.c\\nindex 374e2d0efe..9081dfe384 100644\\n--- a/src/backend/commands/indexcmds.c\\n+++ b/src/backend/commands/indexcmds.c\\n@@ -532,6 +532,12 @@ DefineIndex(Oid relationId,\\n \\t\\t\\t\\t errmsg(\\\"cannot use more than %d columns in an index\\\",\\n \\t\\t\\t\\t\\t\\tINDEX_MAX_KEYS)));\\n \\n+\\tif (stmt->concurrent && !IsTransactionBlock())\\n+\\t{\\n+\\t\\tAssert(GetCurrentTransactionIdIfAny() == InvalidTransactionId);\\n+\\t\\tMyPgXact->vacuumFlags |= PROC_IN_VACUUM;\\n+\\t}\\n+\\n \\t/*\\n \\t * Only SELECT ... FOR UPDATE/SHARE are allowed while doing a standard\\n \\t * index build; but for concurrent builds we allow INSERT/UPDATE/DELETE\\n\\nThere's an obvious flaw, which is that this doesn't consider expressions\\nin partial indexes and column definitions. That's moderately easy to\\nfix. But there are less obvious flaws, such as: is it possible that\\nCIC's xmin is required for other reasons? (such as access to catalogs,\\nwhich get cleaned by concurrent vacuum) If it is, can we fix that\\nproblem by keeping track of a separate Xmin for catalog vacuuming\\npurposes? (We already have catalog_xmin for replication slots, so this\\nis not completely ridiculous I think ...)\\n\\n-- \\n�lvaro Herrera\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Nov 2019 17:33:10 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"excluding CREATE INDEX CONCURRENTLY from OldestXmin\"\n }\n]","isConversational":false}}},{"rowIdx":49395,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello,\n\nI came across a surprising behavior when upgrading our PostgreSQL 10 DBs that\nalso serve as a destination for the logical replication of some reference tables.\n\npg_upgrade turns off all subscriptions on the cluster and doesn't turn them on.\nSpecifically, it creates them with connect = false, as discussed at the thread\nstarting at\nhttps://www.postgresql.org/message-id/e4fbfad5-c6ac-fd50-6777-18c84b34eb2f@2ndquadrant.com\n\nUnfortunately, I can't find any mention of this in the docs of pg_upgrade, so\nI am at leas willing to add those if we can't resolve this in a more automated\nway (read below).\n\nSince we can determine those subscriptions that were active on the old cluster\nimmediately before the upgrade, we could collect those and emit a script at\nthe end of the pg_upgrade to turn them on, similar to the one pg_upgrade\nproduces for the analyze.\n\nThere are two options when re-enabling the subscriptions: either continue from\nthe current position (possibly discarding all changes that happened while the\ndatabase was offline), or truncate the destination tables and copy the data again.\nThe first one corresponds to setting copy_data=false when doing refresh publication.\nThe second one is a combination of a prior truncate + refresh publication with\ncopy_data=true and doesn't seem like an action that is performed in a\nsingle transaction. Would it make sense to add a copy_truncate flag, false\nby default, that would instruct the logical replication worker to truncate the\ndestination table immediately before resyncing it from the origin?\n\nRegards,\nOleksii\n\n\n\n","msg_date":"Sat, 2 Nov 2019 17:55:25 +0100","msg_from":"Oleksii Kliukin <alexk@hintbits.com>","msg_from_op":true,"msg_subject":"pg_upgrade and subscriptions"},{"msg_contents":"Em sáb., 2 de nov. de 2019 às 13:55, Oleksii Kliukin\n<alexk@hintbits.com> escreveu:\n>\n> I came across a surprising behavior when upgrading our PostgreSQL 10 DBs that\n> also serve as a destination for the logical replication of some reference tables.\n>\n> pg_upgrade turns off all subscriptions on the cluster and doesn't turn them on.\n> Specifically, it creates them with connect = false, as discussed at the thread\n> starting at\n> https://www.postgresql.org/message-id/e4fbfad5-c6ac-fd50-6777-18c84b34eb2f@2ndquadrant.com\n>\n> Unfortunately, I can't find any mention of this in the docs of pg_upgrade, so\n> I am at leas willing to add those if we can't resolve this in a more automated\n> way (read below).\n>\nIt is documented in step 13 \"Post-upgrade processing\". In this case,\nwe need to provide a new script for subscriptions.\n\n> Since we can determine those subscriptions that were active on the old cluster\n> immediately before the upgrade, we could collect those and emit a script at\n> the end of the pg_upgrade to turn them on, similar to the one pg_upgrade\n> produces for the analyze.\n>\n+1. It seems to be an oversight (missing feature) that nobody bothers to fix it.\n\n> There are two options when re-enabling the subscriptions: either continue from\n> the current position (possibly discarding all changes that happened while the\n> database was offline), or truncate the destination tables and copy the data again.\n> The first one corresponds to setting copy_data=false when doing refresh publication.\n> The second one is a combination of a prior truncate + refresh publication with\n> copy_data=true and doesn't seem like an action that is performed in a\n> single transaction. Would it make sense to add a copy_truncate flag, false\n> by default, that would instruct the logical replication worker to truncate the\n> destination table immediately before resyncing it from the origin?\n>\nIt seems the common case is the former. However, I don't think people\nwant to \"discard all the changes that happen while database was\noffline\" because the slot will remain in the publisher and we are\nupgrading the subscriber. Since we are providing a script for the\ncommon case, you are free to ignore it and create a new script that\nfulfill your requirements.\n\n\n-- \n Euler Taveira Timbira -\nhttp://www.timbira.com.br/\n PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento\n\n\n","msg_date":"Fri, 8 Nov 2019 11:38:53 -0300","msg_from":"Euler Taveira <euler@timbira.com.br>","msg_from_op":false,"msg_subject":"Re: pg_upgrade and subscriptions"}],"string":"[\n {\n \"msg_contents\": \"Hello,\\n\\nI came across a surprising behavior when upgrading our PostgreSQL 10 DBs that\\nalso serve as a destination for the logical replication of some reference tables.\\n\\npg_upgrade turns off all subscriptions on the cluster and doesn't turn them on.\\nSpecifically, it creates them with connect = false, as discussed at the thread\\nstarting at\\nhttps://www.postgresql.org/message-id/e4fbfad5-c6ac-fd50-6777-18c84b34eb2f@2ndquadrant.com\\n\\nUnfortunately, I can't find any mention of this in the docs of pg_upgrade, so\\nI am at leas willing to add those if we can't resolve this in a more automated\\nway (read below).\\n\\nSince we can determine those subscriptions that were active on the old cluster\\nimmediately before the upgrade, we could collect those and emit a script at\\nthe end of the pg_upgrade to turn them on, similar to the one pg_upgrade\\nproduces for the analyze.\\n\\nThere are two options when re-enabling the subscriptions: either continue from\\nthe current position (possibly discarding all changes that happened while the\\ndatabase was offline), or truncate the destination tables and copy the data again.\\nThe first one corresponds to setting copy_data=false when doing refresh publication.\\nThe second one is a combination of a prior truncate + refresh publication with\\ncopy_data=true and doesn't seem like an action that is performed in a\\nsingle transaction. Would it make sense to add a copy_truncate flag, false\\nby default, that would instruct the logical replication worker to truncate the\\ndestination table immediately before resyncing it from the origin?\\n\\nRegards,\\nOleksii\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 2 Nov 2019 17:55:25 +0100\",\n \"msg_from\": \"Oleksii Kliukin <alexk@hintbits.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pg_upgrade and subscriptions\"\n },\n {\n \"msg_contents\": \"Em sáb., 2 de nov. de 2019 às 13:55, Oleksii Kliukin\\n<alexk@hintbits.com> escreveu:\\n>\\n> I came across a surprising behavior when upgrading our PostgreSQL 10 DBs that\\n> also serve as a destination for the logical replication of some reference tables.\\n>\\n> pg_upgrade turns off all subscriptions on the cluster and doesn't turn them on.\\n> Specifically, it creates them with connect = false, as discussed at the thread\\n> starting at\\n> https://www.postgresql.org/message-id/e4fbfad5-c6ac-fd50-6777-18c84b34eb2f@2ndquadrant.com\\n>\\n> Unfortunately, I can't find any mention of this in the docs of pg_upgrade, so\\n> I am at leas willing to add those if we can't resolve this in a more automated\\n> way (read below).\\n>\\nIt is documented in step 13 \\\"Post-upgrade processing\\\". In this case,\\nwe need to provide a new script for subscriptions.\\n\\n> Since we can determine those subscriptions that were active on the old cluster\\n> immediately before the upgrade, we could collect those and emit a script at\\n> the end of the pg_upgrade to turn them on, similar to the one pg_upgrade\\n> produces for the analyze.\\n>\\n+1. It seems to be an oversight (missing feature) that nobody bothers to fix it.\\n\\n> There are two options when re-enabling the subscriptions: either continue from\\n> the current position (possibly discarding all changes that happened while the\\n> database was offline), or truncate the destination tables and copy the data again.\\n> The first one corresponds to setting copy_data=false when doing refresh publication.\\n> The second one is a combination of a prior truncate + refresh publication with\\n> copy_data=true and doesn't seem like an action that is performed in a\\n> single transaction. Would it make sense to add a copy_truncate flag, false\\n> by default, that would instruct the logical replication worker to truncate the\\n> destination table immediately before resyncing it from the origin?\\n>\\nIt seems the common case is the former. However, I don't think people\\nwant to \\\"discard all the changes that happen while database was\\noffline\\\" because the slot will remain in the publisher and we are\\nupgrading the subscriber. Since we are providing a script for the\\ncommon case, you are free to ignore it and create a new script that\\nfulfill your requirements.\\n\\n\\n-- \\n Euler Taveira Timbira -\\nhttp://www.timbira.com.br/\\n PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Nov 2019 11:38:53 -0300\",\n \"msg_from\": \"Euler Taveira <euler@timbira.com.br>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_upgrade and subscriptions\"\n }\n]","isConversational":false}}},{"rowIdx":49396,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi hackers,\n\nI just noticed that when contrib/seg was converted to V1 calling\nconvention (commit 389bb2818f4), the PG_GETARG_SEG_P() macro got defined\nin terms of PG_GETARG_POINTER(). But it itself calls DatumGetPointer(),\nso shouldn't it be using PG_GETARG_DATUM()?\n\nAttached is a patch that fixes it, and brings it in line with all the\nother PG_GETARG_FOO_P() macros.\n\n- ilmari\n-- \n\"A disappointingly low fraction of the human race is,\n at any given time, on fire.\" - Stig Sandbeck Mathisen","msg_date":"Sat, 02 Nov 2019 23:14:47 +0000","msg_from":"ilmari@ilmari.org (Dagfinn Ilmari =?utf-8?Q?Manns=C3=A5ker?=)","msg_from_op":true,"msg_subject":"[PATCH] contrib/seg: Fix PG_GETARG_SEG_P definition"},{"msg_contents":"ilmari@ilmari.org (Dagfinn Ilmari =?utf-8?Q?Manns=C3=A5ker?=) writes:\n> I just noticed that when contrib/seg was converted to V1 calling\n> convention (commit 389bb2818f4), the PG_GETARG_SEG_P() macro got defined\n> in terms of PG_GETARG_POINTER(). But it itself calls DatumGetPointer(),\n> so shouldn't it be using PG_GETARG_DATUM()?\n\nYup, I agree. Pushed.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 03 Nov 2019 10:58:30 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] contrib/seg: Fix PG_GETARG_SEG_P definition"},{"msg_contents":"Tom Lane <tgl@sss.pgh.pa.us> writes:\n\n> ilmari@ilmari.org (Dagfinn Ilmari =?utf-8?Q?Manns=C3=A5ker?=) writes:\n>> I just noticed that when contrib/seg was converted to V1 calling\n>> convention (commit 389bb2818f4), the PG_GETARG_SEG_P() macro got defined\n>> in terms of PG_GETARG_POINTER(). But it itself calls DatumGetPointer(),\n>> so shouldn't it be using PG_GETARG_DATUM()?\n>\n> Yup, I agree. Pushed.\n\nThanks!\n\n> \t\t\tregards, tom lane\n\n- ilmari\n-- \n\"The surreality of the universe tends towards a maximum\" -- Skud's Law\n\"Never formulate a law or axiom that you're not prepared to live with\n the consequences of.\" -- Skud's Meta-Law\n\n\n","msg_date":"Mon, 04 Nov 2019 11:30:23 +0000","msg_from":"ilmari@ilmari.org (Dagfinn Ilmari =?utf-8?Q?Manns=C3=A5ker?=)","msg_from_op":true,"msg_subject":"Re: [PATCH] contrib/seg: Fix PG_GETARG_SEG_P definition"},{"msg_contents":"Hi,\n\nOn 2019-11-04 11:30:23 +0000, Dagfinn Ilmari Manns�ker wrote:\n> Tom Lane <tgl@sss.pgh.pa.us> writes:\n> \n> > ilmari@ilmari.org (Dagfinn Ilmari =?utf-8?Q?Manns=C3=A5ker?=) writes:\n> >> I just noticed that when contrib/seg was converted to V1 calling\n> >> convention (commit 389bb2818f4), the PG_GETARG_SEG_P() macro got defined\n> >> in terms of PG_GETARG_POINTER(). But it itself calls DatumGetPointer(),\n> >> so shouldn't it be using PG_GETARG_DATUM()?\n> >\n> > Yup, I agree. Pushed.\n> \n> Thanks!\n\nThanks both of you.\n\n- Andres\n\n\n","msg_date":"Mon, 4 Nov 2019 10:01:44 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] contrib/seg: Fix PG_GETARG_SEG_P definition"}],"string":"[\n {\n \"msg_contents\": \"Hi hackers,\\n\\nI just noticed that when contrib/seg was converted to V1 calling\\nconvention (commit 389bb2818f4), the PG_GETARG_SEG_P() macro got defined\\nin terms of PG_GETARG_POINTER(). But it itself calls DatumGetPointer(),\\nso shouldn't it be using PG_GETARG_DATUM()?\\n\\nAttached is a patch that fixes it, and brings it in line with all the\\nother PG_GETARG_FOO_P() macros.\\n\\n- ilmari\\n-- \\n\\\"A disappointingly low fraction of the human race is,\\n at any given time, on fire.\\\" - Stig Sandbeck Mathisen\",\n \"msg_date\": \"Sat, 02 Nov 2019 23:14:47 +0000\",\n \"msg_from\": \"ilmari@ilmari.org (Dagfinn Ilmari =?utf-8?Q?Manns=C3=A5ker?=)\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[PATCH] contrib/seg: Fix PG_GETARG_SEG_P definition\"\n },\n {\n \"msg_contents\": \"ilmari@ilmari.org (Dagfinn Ilmari =?utf-8?Q?Manns=C3=A5ker?=) writes:\\n> I just noticed that when contrib/seg was converted to V1 calling\\n> convention (commit 389bb2818f4), the PG_GETARG_SEG_P() macro got defined\\n> in terms of PG_GETARG_POINTER(). But it itself calls DatumGetPointer(),\\n> so shouldn't it be using PG_GETARG_DATUM()?\\n\\nYup, I agree. Pushed.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 03 Nov 2019 10:58:30 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] contrib/seg: Fix PG_GETARG_SEG_P definition\"\n },\n {\n \"msg_contents\": \"Tom Lane <tgl@sss.pgh.pa.us> writes:\\n\\n> ilmari@ilmari.org (Dagfinn Ilmari =?utf-8?Q?Manns=C3=A5ker?=) writes:\\n>> I just noticed that when contrib/seg was converted to V1 calling\\n>> convention (commit 389bb2818f4), the PG_GETARG_SEG_P() macro got defined\\n>> in terms of PG_GETARG_POINTER(). But it itself calls DatumGetPointer(),\\n>> so shouldn't it be using PG_GETARG_DATUM()?\\n>\\n> Yup, I agree. Pushed.\\n\\nThanks!\\n\\n> \\t\\t\\tregards, tom lane\\n\\n- ilmari\\n-- \\n\\\"The surreality of the universe tends towards a maximum\\\" -- Skud's Law\\n\\\"Never formulate a law or axiom that you're not prepared to live with\\n the consequences of.\\\" -- Skud's Meta-Law\\n\\n\\n\",\n \"msg_date\": \"Mon, 04 Nov 2019 11:30:23 +0000\",\n \"msg_from\": \"ilmari@ilmari.org (Dagfinn Ilmari =?utf-8?Q?Manns=C3=A5ker?=)\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] contrib/seg: Fix PG_GETARG_SEG_P definition\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2019-11-04 11:30:23 +0000, Dagfinn Ilmari Manns�ker wrote:\\n> Tom Lane <tgl@sss.pgh.pa.us> writes:\\n> \\n> > ilmari@ilmari.org (Dagfinn Ilmari =?utf-8?Q?Manns=C3=A5ker?=) writes:\\n> >> I just noticed that when contrib/seg was converted to V1 calling\\n> >> convention (commit 389bb2818f4), the PG_GETARG_SEG_P() macro got defined\\n> >> in terms of PG_GETARG_POINTER(). But it itself calls DatumGetPointer(),\\n> >> so shouldn't it be using PG_GETARG_DATUM()?\\n> >\\n> > Yup, I agree. Pushed.\\n> \\n> Thanks!\\n\\nThanks both of you.\\n\\n- Andres\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Nov 2019 10:01:44 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] contrib/seg: Fix PG_GETARG_SEG_P definition\"\n }\n]","isConversational":false}}},{"rowIdx":49397,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"While monitoring pg_stat_subscription, I noticed that last_msg_send_time\nwas usually NULL, which doesn't make sense. Why would we have a message,\nbut not know when it was sent?\n\nLooking in src/backend/replication/walsender.c, there is this:\n\n /* output previously gathered data in a CopyData packet */\n pq_putmessage_noblock('d', ctx->out->data, ctx->out->len);\n\n /*\n * Fill the send timestamp last, so that it is taken as late as\npossible.\n * This is somewhat ugly, but the protocol is set as it's already used\nfor\n * several releases by streaming physical replication.\n */\n resetStringInfo(&tmpbuf);\n now = GetCurrentTimestamp();\n pq_sendint64(&tmpbuf, now);\n memcpy(&ctx->out->data[1 + sizeof(int64) + sizeof(int64)],\n tmpbuf.data, sizeof(int64));\n\nFilling out the timestamp after the message has already been sent is taking\n\"as late as possible\" a little too far. It results in every message having\na zero timestamp, other than keep-alives which go through a different path.\n\nRe-ordering the code blocks as in the attached seems to fix it. But I have\nto wonder, if this has been broken from the start and no one ever noticed,\nis this even valuable information to relay in the first place? We could\njust take the column out of the view, and not bother calling\nGetCurrentTimestamp() for each message.\n\nCheers,\n\nJeff","msg_date":"Sat, 2 Nov 2019 21:54:54 -0400","msg_from":"Jeff Janes <jeff.janes@gmail.com>","msg_from_op":true,"msg_subject":"Logical replication wal sender timestamp bug"},{"msg_contents":"On Sat, Nov 02, 2019 at 09:54:54PM -0400, Jeff Janes wrote:\n> While monitoring pg_stat_subscription, I noticed that last_msg_send_time\n> was usually NULL, which doesn't make sense. Why would we have a message,\n> but not know when it was sent?\n\nSo... The timestamp is received and stored in LogicalRepApplyLoop()\nwith send_time when a 'w' message is received in the subscription\ncluster. And it gets computed with a two-phase process:\n- WalSndPrepareWrite() reserves the space in the message for the\ntimestamp.\n- WalSndWriteData() computes the timestamp in the reserved space once\nthe write message is computed and ready to go.\n\n> Filling out the timestamp after the message has already been sent is taking\n> \"as late as possible\" a little too far. It results in every message having\n> a zero timestamp, other than keep-alives which go through a different path.\n\nIt seems to me that you are right: the timestamp is computed too\nlate.\n\n> Re-ordering the code blocks as in the attached seems to fix it. But I have\n> to wonder, if this has been broken from the start and no one ever noticed,\n> is this even valuable information to relay in the first place? We could\n> just take the column out of the view, and not bother calling\n> GetCurrentTimestamp() for each message.\n\nI think that there are use cases for such monitoring capabilities,\nsee for example 7fee252. So I would rather keep it.\n--\nMichael","msg_date":"Tue, 5 Nov 2019 13:19:37 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Logical replication wal sender timestamp bug"},{"msg_contents":"On Tue, Nov 05, 2019 at 01:19:37PM +0900, Michael Paquier wrote:\n> On Sat, Nov 02, 2019 at 09:54:54PM -0400, Jeff Janes wrote:\n>> Filling out the timestamp after the message has already been sent is taking\n>> \"as late as possible\" a little too far. It results in every message having\n>> a zero timestamp, other than keep-alives which go through a different path.\n> \n> It seems to me that you are right: the timestamp is computed too\n> late.\n\nIt is easy enough to reproduce the problem by setting for example\nlogical replication between two nodes and pgbench to produce some\nload and then monitor pg_stat_subscription periodically. However, it\nis a problem since logical decoding has been introduced (5a991ef) so\ncommitted your fix down to 9.4.\n--\nMichael","msg_date":"Wed, 6 Nov 2019 16:15:00 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Logical replication wal sender timestamp bug"},{"msg_contents":"On Wed, Nov 6, 2019 at 2:15 AM Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Tue, Nov 05, 2019 at 01:19:37PM +0900, Michael Paquier wrote:\n> > On Sat, Nov 02, 2019 at 09:54:54PM -0400, Jeff Janes wrote:\n> >> Filling out the timestamp after the message has already been sent is\n> taking\n> >> \"as late as possible\" a little too far. It results in every message\n> having\n> >> a zero timestamp, other than keep-alives which go through a different\n> path.\n> >\n> > It seems to me that you are right: the timestamp is computed too\n> > late.\n>\n> It is easy enough to reproduce the problem by setting for example\n> logical replication between two nodes and pgbench to produce some\n> load and then monitor pg_stat_subscription periodically. However, it\n> is a problem since logical decoding has been introduced (5a991ef) so\n> committed your fix down to 9.4.\n>\n\nThanks. This column looks much more reasonable now.\n\nCheers,\n\nJeff\n\nOn Wed, Nov 6, 2019 at 2:15 AM Michael Paquier <michael@paquier.xyz> wrote:On Tue, Nov 05, 2019 at 01:19:37PM +0900, Michael Paquier wrote:\n> On Sat, Nov 02, 2019 at 09:54:54PM -0400, Jeff Janes wrote:\n>> Filling out the timestamp after the message has already been sent is taking\n>> \"as late as possible\" a little too far. It results in every message having\n>> a zero timestamp, other than keep-alives which go through a different path.\n> \n> It seems to me that you are right: the timestamp is computed too\n> late.\n\nIt is easy enough to reproduce the problem by setting for example\nlogical replication between two nodes and pgbench to produce some\nload and then monitor pg_stat_subscription periodically. However, it\nis a problem since logical decoding has been introduced (5a991ef) so\ncommitted your fix down to 9.4.Thanks. This column looks much more reasonable now.Cheers,Jeff","msg_date":"Fri, 8 Nov 2019 15:01:35 -0500","msg_from":"Jeff Janes <jeff.janes@gmail.com>","msg_from_op":true,"msg_subject":"Re: Logical replication wal sender timestamp bug"}],"string":"[\n {\n \"msg_contents\": \"While monitoring pg_stat_subscription, I noticed that last_msg_send_time\\nwas usually NULL, which doesn't make sense. Why would we have a message,\\nbut not know when it was sent?\\n\\nLooking in src/backend/replication/walsender.c, there is this:\\n\\n /* output previously gathered data in a CopyData packet */\\n pq_putmessage_noblock('d', ctx->out->data, ctx->out->len);\\n\\n /*\\n * Fill the send timestamp last, so that it is taken as late as\\npossible.\\n * This is somewhat ugly, but the protocol is set as it's already used\\nfor\\n * several releases by streaming physical replication.\\n */\\n resetStringInfo(&tmpbuf);\\n now = GetCurrentTimestamp();\\n pq_sendint64(&tmpbuf, now);\\n memcpy(&ctx->out->data[1 + sizeof(int64) + sizeof(int64)],\\n tmpbuf.data, sizeof(int64));\\n\\nFilling out the timestamp after the message has already been sent is taking\\n\\\"as late as possible\\\" a little too far. It results in every message having\\na zero timestamp, other than keep-alives which go through a different path.\\n\\nRe-ordering the code blocks as in the attached seems to fix it. But I have\\nto wonder, if this has been broken from the start and no one ever noticed,\\nis this even valuable information to relay in the first place? We could\\njust take the column out of the view, and not bother calling\\nGetCurrentTimestamp() for each message.\\n\\nCheers,\\n\\nJeff\",\n \"msg_date\": \"Sat, 2 Nov 2019 21:54:54 -0400\",\n \"msg_from\": \"Jeff Janes <jeff.janes@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Logical replication wal sender timestamp bug\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 02, 2019 at 09:54:54PM -0400, Jeff Janes wrote:\\n> While monitoring pg_stat_subscription, I noticed that last_msg_send_time\\n> was usually NULL, which doesn't make sense. Why would we have a message,\\n> but not know when it was sent?\\n\\nSo... The timestamp is received and stored in LogicalRepApplyLoop()\\nwith send_time when a 'w' message is received in the subscription\\ncluster. And it gets computed with a two-phase process:\\n- WalSndPrepareWrite() reserves the space in the message for the\\ntimestamp.\\n- WalSndWriteData() computes the timestamp in the reserved space once\\nthe write message is computed and ready to go.\\n\\n> Filling out the timestamp after the message has already been sent is taking\\n> \\\"as late as possible\\\" a little too far. It results in every message having\\n> a zero timestamp, other than keep-alives which go through a different path.\\n\\nIt seems to me that you are right: the timestamp is computed too\\nlate.\\n\\n> Re-ordering the code blocks as in the attached seems to fix it. But I have\\n> to wonder, if this has been broken from the start and no one ever noticed,\\n> is this even valuable information to relay in the first place? We could\\n> just take the column out of the view, and not bother calling\\n> GetCurrentTimestamp() for each message.\\n\\nI think that there are use cases for such monitoring capabilities,\\nsee for example 7fee252. So I would rather keep it.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 5 Nov 2019 13:19:37 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical replication wal sender timestamp bug\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 05, 2019 at 01:19:37PM +0900, Michael Paquier wrote:\\n> On Sat, Nov 02, 2019 at 09:54:54PM -0400, Jeff Janes wrote:\\n>> Filling out the timestamp after the message has already been sent is taking\\n>> \\\"as late as possible\\\" a little too far. It results in every message having\\n>> a zero timestamp, other than keep-alives which go through a different path.\\n> \\n> It seems to me that you are right: the timestamp is computed too\\n> late.\\n\\nIt is easy enough to reproduce the problem by setting for example\\nlogical replication between two nodes and pgbench to produce some\\nload and then monitor pg_stat_subscription periodically. However, it\\nis a problem since logical decoding has been introduced (5a991ef) so\\ncommitted your fix down to 9.4.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 6 Nov 2019 16:15:00 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Logical replication wal sender timestamp bug\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 6, 2019 at 2:15 AM Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Tue, Nov 05, 2019 at 01:19:37PM +0900, Michael Paquier wrote:\\n> > On Sat, Nov 02, 2019 at 09:54:54PM -0400, Jeff Janes wrote:\\n> >> Filling out the timestamp after the message has already been sent is\\n> taking\\n> >> \\\"as late as possible\\\" a little too far. It results in every message\\n> having\\n> >> a zero timestamp, other than keep-alives which go through a different\\n> path.\\n> >\\n> > It seems to me that you are right: the timestamp is computed too\\n> > late.\\n>\\n> It is easy enough to reproduce the problem by setting for example\\n> logical replication between two nodes and pgbench to produce some\\n> load and then monitor pg_stat_subscription periodically. However, it\\n> is a problem since logical decoding has been introduced (5a991ef) so\\n> committed your fix down to 9.4.\\n>\\n\\nThanks. This column looks much more reasonable now.\\n\\nCheers,\\n\\nJeff\\n\\nOn Wed, Nov 6, 2019 at 2:15 AM Michael Paquier <michael@paquier.xyz> wrote:On Tue, Nov 05, 2019 at 01:19:37PM +0900, Michael Paquier wrote:\\n> On Sat, Nov 02, 2019 at 09:54:54PM -0400, Jeff Janes wrote:\\n>> Filling out the timestamp after the message has already been sent is taking\\n>> \\\"as late as possible\\\" a little too far. It results in every message having\\n>> a zero timestamp, other than keep-alives which go through a different path.\\n> \\n> It seems to me that you are right: the timestamp is computed too\\n> late.\\n\\nIt is easy enough to reproduce the problem by setting for example\\nlogical replication between two nodes and pgbench to produce some\\nload and then monitor pg_stat_subscription periodically. However, it\\nis a problem since logical decoding has been introduced (5a991ef) so\\ncommitted your fix down to 9.4.Thanks. This column looks much more reasonable now.Cheers,Jeff\",\n \"msg_date\": \"Fri, 8 Nov 2019 15:01:35 -0500\",\n \"msg_from\": \"Jeff Janes <jeff.janes@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Logical replication wal sender timestamp bug\"\n }\n]","isConversational":false}}},{"rowIdx":49398,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello Folks\n\nI would like to update unaccent.rules file to support Arabic letters. so could someone help me or tell me how could I add such contribution. I attached the file including the modifications, only the last 4 lines.\n\nthank you","msg_date":"Sun, 3 Nov 2019 06:02:19 +0000","msg_from":"kerbrose khaled <kerbrose@hotmail.com>","msg_from_op":true,"msg_subject":"updating unaccent.rules for Arabic letters"},{"msg_contents":"Hello Folks\n\nI would like to update unaccent.rules file to support Arabic letters. so could someone help me or tell me how could I add such contribution. I attached the file including the modifications, only the last 4 lines.\n\nthank you","msg_date":"Sun, 3 Nov 2019 06:05:25 +0000","msg_from":"kerbrose khaled <kerbrose@hotmail.com>","msg_from_op":true,"msg_subject":"updating unaccent.rules for Arabic letters"},{"msg_contents":"kerbrose khaled <kerbrose@hotmail.com> writes:\n> I would like to update unaccent.rules file to support Arabic letters. so could someone help me or tell me how could I add such contribution. I attached the file including the modifications, only the last 4 lines.\n\nHi! I've got no objection to including Arabic in the set of covered\nlanguages, but handing us a new unaccent.rules file isn't the way to\ndo it, because that's a generated file. The adjacent script\ngenerate_unaccent_rules.py generates it from the official Unicode\nsource data (see comments in that script). What we need, ultimately,\nis a patch to that script so it will emit these additional translations.\nPast commits that might be useful sources of inspiration include\n\nhttps://git.postgresql.org/gitweb/?p=postgresql.git&a=commitdiff&h=456e3718e7b72efe4d2639437fcbca2e4ad83099\nhttps://git.postgresql.org/gitweb/?p=postgresql.git&a=commitdiff&h=5e8d670c313531c0dca245943fb84c94a477ddc4\nhttps://git.postgresql.org/gitweb/?p=postgresql.git&a=commitdiff&h=ec0a69e49bf41a37b5c2d6f6be66d8abae00ee05\n\nIf you're not good with Python, maybe you could just explain to us\nhow to recognize these characters from Unicode character properties.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 03 Nov 2019 11:12:15 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: updating unaccent.rules for Arabic letters"},{"msg_contents":"kerbrose khaled wrote:\n\n> I would like to update unaccent.rules file to support Arabic letters. so\n> could someone help me or tell me how could I add such contribution. I\n> attached the file including the modifications, only the last 4 lines.\n\nThe Arabic letters are found in the Unicode block U+0600 to U+06FF \n(https://www.fileformat.info/info/unicode/block/arabic/list.htm)\nThere has been no coverage of this block until now by the unaccent\nmodule. Since Arabic uses several diacritics [1] , it would be best to\nfigure out all the transliterations that should go in and let them in\none go (plus coding that in the Python script).\n\nThe canonical way to unaccent is normally to apply a Unicode\ntransformation: NFC -> NFD and remove the non-spacing marks.\n\nI've tentatively did that with each codepoint in the 0600-06FF block\nin SQL with icu_transform in icu_ext [2], and it produces the\nattached result, with 60 (!) entries, along with Unicode names for\nreadability.\n\nDoes that make sense to people who know Arabic?\n\nFor the record, here's the query:\n\nWITH block(cp) AS (select * FROM generate_series(x'600'::int,x'6ff'::int) AS\ncp),\n dest AS (select cp, icu_transform(chr(cp), 'any-NFD;[:nonspacing mark:]\nany-remove; any-NFC') AS unaccented FROM block)\nSELECT\n chr(cp) as \"src\",\n icu_transform(chr(cp), 'Name') as \"srcName\",\n dest.unaccented as \"dest\",\n icu_transform(dest.unaccented, 'Name') as \"destName\"\nFROM dest\nWHERE chr(cp) <> dest.unaccented;\n\n\n[1] https://en.wikipedia.org/wiki/Arabic_diacritics\n[2] https://github.com/dverite/icu_ext#icu_transform\n\n\nBest regards,\n-- \nDaniel Vérité\nPostgreSQL-powered mailer: http://www.manitou-mail.org\nTwitter: @DanielVerite","msg_date":"Mon, 04 Nov 2019 18:41:59 +0100","msg_from":"\"Daniel Verite\" <daniel@manitou-mail.org>","msg_from_op":false,"msg_subject":"Re: updating unaccent.rules for Arabic letters"}],"string":"[\n {\n \"msg_contents\": \"Hello Folks\\n\\nI would like to update unaccent.rules file to support Arabic letters. so could someone help me or tell me how could I add such contribution. I attached the file including the modifications, only the last 4 lines.\\n\\nthank you\",\n \"msg_date\": \"Sun, 3 Nov 2019 06:02:19 +0000\",\n \"msg_from\": \"kerbrose khaled <kerbrose@hotmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"updating unaccent.rules for Arabic letters\"\n },\n {\n \"msg_contents\": \"Hello Folks\\n\\nI would like to update unaccent.rules file to support Arabic letters. so could someone help me or tell me how could I add such contribution. I attached the file including the modifications, only the last 4 lines.\\n\\nthank you\",\n \"msg_date\": \"Sun, 3 Nov 2019 06:05:25 +0000\",\n \"msg_from\": \"kerbrose khaled <kerbrose@hotmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"updating unaccent.rules for Arabic letters\"\n },\n {\n \"msg_contents\": \"kerbrose khaled <kerbrose@hotmail.com> writes:\\n> I would like to update unaccent.rules file to support Arabic letters. so could someone help me or tell me how could I add such contribution. I attached the file including the modifications, only the last 4 lines.\\n\\nHi! I've got no objection to including Arabic in the set of covered\\nlanguages, but handing us a new unaccent.rules file isn't the way to\\ndo it, because that's a generated file. The adjacent script\\ngenerate_unaccent_rules.py generates it from the official Unicode\\nsource data (see comments in that script). What we need, ultimately,\\nis a patch to that script so it will emit these additional translations.\\nPast commits that might be useful sources of inspiration include\\n\\nhttps://git.postgresql.org/gitweb/?p=postgresql.git&a=commitdiff&h=456e3718e7b72efe4d2639437fcbca2e4ad83099\\nhttps://git.postgresql.org/gitweb/?p=postgresql.git&a=commitdiff&h=5e8d670c313531c0dca245943fb84c94a477ddc4\\nhttps://git.postgresql.org/gitweb/?p=postgresql.git&a=commitdiff&h=ec0a69e49bf41a37b5c2d6f6be66d8abae00ee05\\n\\nIf you're not good with Python, maybe you could just explain to us\\nhow to recognize these characters from Unicode character properties.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 03 Nov 2019 11:12:15 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: updating unaccent.rules for Arabic letters\"\n },\n {\n \"msg_contents\": \"kerbrose khaled wrote:\\n\\n> I would like to update unaccent.rules file to support Arabic letters. so\\n> could someone help me or tell me how could I add such contribution. I\\n> attached the file including the modifications, only the last 4 lines.\\n\\nThe Arabic letters are found in the Unicode block U+0600 to U+06FF \\n(https://www.fileformat.info/info/unicode/block/arabic/list.htm)\\nThere has been no coverage of this block until now by the unaccent\\nmodule. Since Arabic uses several diacritics [1] , it would be best to\\nfigure out all the transliterations that should go in and let them in\\none go (plus coding that in the Python script).\\n\\nThe canonical way to unaccent is normally to apply a Unicode\\ntransformation: NFC -> NFD and remove the non-spacing marks.\\n\\nI've tentatively did that with each codepoint in the 0600-06FF block\\nin SQL with icu_transform in icu_ext [2], and it produces the\\nattached result, with 60 (!) entries, along with Unicode names for\\nreadability.\\n\\nDoes that make sense to people who know Arabic?\\n\\nFor the record, here's the query:\\n\\nWITH block(cp) AS (select * FROM generate_series(x'600'::int,x'6ff'::int) AS\\ncp),\\n dest AS (select cp, icu_transform(chr(cp), 'any-NFD;[:nonspacing mark:]\\nany-remove; any-NFC') AS unaccented FROM block)\\nSELECT\\n chr(cp) as \\\"src\\\",\\n icu_transform(chr(cp), 'Name') as \\\"srcName\\\",\\n dest.unaccented as \\\"dest\\\",\\n icu_transform(dest.unaccented, 'Name') as \\\"destName\\\"\\nFROM dest\\nWHERE chr(cp) <> dest.unaccented;\\n\\n\\n[1] https://en.wikipedia.org/wiki/Arabic_diacritics\\n[2] https://github.com/dverite/icu_ext#icu_transform\\n\\n\\nBest regards,\\n-- \\nDaniel Vérité\\nPostgreSQL-powered mailer: http://www.manitou-mail.org\\nTwitter: @DanielVerite\",\n \"msg_date\": \"Mon, 04 Nov 2019 18:41:59 +0100\",\n \"msg_from\": \"\\\"Daniel Verite\\\" <daniel@manitou-mail.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: updating unaccent.rules for Arabic letters\"\n }\n]","isConversational":false}}},{"rowIdx":49399,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello!\n\nRecently I got few times into situation where I was trying to find out what\nis blocking DELETE queries. Running EXPLAIN (even VERBOSE one) wasn't\nuseful, since the reason was slow trigger (missing index on foreign key\ncolumn). I had to create testing entry and run EXPLAIN ANALYZE DELETE to\nget this information.\n\nIt will be really valuable for me to show triggers in EXPLAIN query since\nit will make clear for me there will be any trigger \"activated\" during\nexecution of DELETE query and that can be the reason for slow DELETE.\n\nI have seen initial discussion at\nhttps://www.postgresql.org/message-id/flat/20693.1111732761%40sss.pgh.pa.us\nto show time spent in triggers in EXPLAIN ANALYZE including quick\ndiscussion to possibly show triggers during EXPLAIN. Anyway since it\ndoesn't show any additional cost and just inform about the possibilities, I\nstill consider this feature useful. This is probably implementation of idea\nmentioned at\nhttps://www.postgresql.org/message-id/21221.1111736869%40sss.pgh.pa.us by\nTom Lane.\n\nAfter initial discussion with Pavel Stěhule and Tomáš Vondra at czech\npostgresql maillist (\nhttps://groups.google.com/forum/#!topic/postgresql-cz/Dq1sT7huVho) I was\nable to prepare initial version of this patch. I have added EXPLAIN option\ncalled TRIGGERS enabled by default.There's already autoexplain property for\nthis. I understand it is not possible to show only triggers which will be\nreally activated unless query is really executed. EXPLAIN ANALYZE remains\nuntouched with this patch.\n\n- patch with examples can be found at\nhttps://github.com/simi/postgres/pull/2\n- DIFF format https://github.com/simi/postgres/pull/2.diff\n- PATCH format (also attached) https://github.com/simi/postgres/pull/2.patch\n\nAll regression tests passed with this change locally on latest git master.\nI would like to cover this patch with more regression tests, but I wasn't\nsure where to place them, since there's no \"EXPLAIN\" related test \"group\".\nIs \"src/test/regress/sql/triggers.sql\" the best place to add tests related\nto this change?\n\nPS: This is my first try to contribute to postgresql codebase. The quality\nof patch is probably not the best, but I will be more than happy to do any\nrequested change if needed.\n\nRegards,\nJosef Šimánek","msg_date":"Sun, 3 Nov 2019 18:25:45 +0100","msg_from":"=?UTF-8?B?Sm9zZWYgxaBpbcOhbmVr?= <josef.simanek@gmail.com>","msg_from_op":true,"msg_subject":"[PATCH] Include triggers in EXPLAIN"},{"msg_contents":"=?UTF-8?B?Sm9zZWYgxaBpbcOhbmVr?= <josef.simanek@gmail.com> writes:\n> Recently I got few times into situation where I was trying to find out what\n> is blocking DELETE queries. Running EXPLAIN (even VERBOSE one) wasn't\n> useful, since the reason was slow trigger (missing index on foreign key\n> column). I had to create testing entry and run EXPLAIN ANALYZE DELETE to\n> get this information.\n\n> It will be really valuable for me to show triggers in EXPLAIN query since\n> it will make clear for me there will be any trigger \"activated\" during\n> execution of DELETE query and that can be the reason for slow DELETE.\n\nI don't really see the point of this patch? You do get the trigger\ntimes during EXPLAIN ANALYZE, and I don't believe that a plain EXPLAIN\nis going to have full information about what triggers might fire or\nnot fire.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 03 Nov 2019 16:49:36 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Include triggers in EXPLAIN"},{"msg_contents":"Hello Tom.\n\nThanks for quick response. As I was testing this feature it shows all\n\"possible\" triggers to be executed running given query. The benefit of\nhaving this information in EXPLAIN as well is you do not need to execute\nthe query (as EXPLAIN ANALYZE does). My usecase is to take a look at query\nbefore it is executed to get some idea about the plan with EXPLAIN.\n\nDo you have idea about some case where actual trigger will be missing in\nEXPLAIN with current implementation, but will be present in EXPLAIN\nANALYZE? I can take a look if there's any way how to handle those cases as\nwell.\n\nne 3. 11. 2019 v 22:49 odesílatel Tom Lane <tgl@sss.pgh.pa.us> napsal:\n\n> =?UTF-8?B?Sm9zZWYgxaBpbcOhbmVr?= <josef.simanek@gmail.com> writes:\n> > Recently I got few times into situation where I was trying to find out\n> what\n> > is blocking DELETE queries. Running EXPLAIN (even VERBOSE one) wasn't\n> > useful, since the reason was slow trigger (missing index on foreign key\n> > column). I had to create testing entry and run EXPLAIN ANALYZE DELETE to\n> > get this information.\n>\n> > It will be really valuable for me to show triggers in EXPLAIN query since\n> > it will make clear for me there will be any trigger \"activated\" during\n> > execution of DELETE query and that can be the reason for slow DELETE.\n>\n> I don't really see the point of this patch? You do get the trigger\n> times during EXPLAIN ANALYZE, and I don't believe that a plain EXPLAIN\n> is going to have full information about what triggers might fire or\n> not fire.\n>\n> regards, tom lane\n>\n\nHello Tom.Thanks for quick response. As I was testing this feature it shows all \"possible\" triggers to be executed running given query. The benefit of having this information in EXPLAIN as well is you do not need to execute the query (as EXPLAIN ANALYZE does). My usecase is to take a look at query before it is executed to get some idea about the plan with EXPLAIN.Do you have idea about some case where actual trigger will be missing in EXPLAIN with current implementation, but will be present in EXPLAIN ANALYZE? I can take a look if there's any way how to handle those cases as well.ne 3. 11. 2019 v 22:49 odesílatel Tom Lane <tgl@sss.pgh.pa.us> napsal:=?UTF-8?B?Sm9zZWYgxaBpbcOhbmVr?= <josef.simanek@gmail.com> writes:\n> Recently I got few times into situation where I was trying to find out what\n> is blocking DELETE queries. Running EXPLAIN (even VERBOSE one) wasn't\n> useful, since the reason was slow trigger (missing index on foreign key\n> column). I had to create testing entry and run EXPLAIN ANALYZE DELETE to\n> get this information.\n\n> It will be really valuable for me to show triggers in EXPLAIN query since\n> it will make clear for me there will be any trigger \"activated\" during\n> execution of DELETE query and that can be the reason for slow DELETE.\n\nI don't really see the point of this patch? You do get the trigger\ntimes during EXPLAIN ANALYZE, and I don't believe that a plain EXPLAIN\nis going to have full information about what triggers might fire or\nnot fire.\n\n regards, tom lane","msg_date":"Mon, 4 Nov 2019 10:35:25 +0100","msg_from":"=?UTF-8?B?Sm9zZWYgxaBpbcOhbmVr?= <josef.simanek@gmail.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Include triggers in EXPLAIN"},{"msg_contents":"Hi,\n\n(minor note - on PG lists the style is to quote in-line and trip)\n\nOn 2019-11-04 10:35:25 +0100, Josef Šimánek wrote:\n> Thanks for quick response. As I was testing this feature it shows all\n> \"possible\" triggers to be executed running given query. The benefit of\n> having this information in EXPLAIN as well is you do not need to execute\n> the query (as EXPLAIN ANALYZE does). My usecase is to take a look at query\n> before it is executed to get some idea about the plan with EXPLAIN.\n\nI can actually see some value in additional information here, but I'd\nprobably want to change the format a bit. When explicitly desired (or\nperhaps just in verbose mode?), I see value in counting the number of\ntriggers we know about that need to be checked, how many were excluded\non the basis of the trigger's WHEN clause etc.\n\n\n> Do you have idea about some case where actual trigger will be missing in\n> EXPLAIN with current implementation, but will be present in EXPLAIN\n> ANALYZE? I can take a look if there's any way how to handle those cases as\n> well.\n\nAny triggers that are fired because of other, listed, triggers causing\nother changes. E.g. a logging trigger that inserts into a log table -\nEXPLAIN, without ANALYZE, doesn't have a way of knowing about that.\n\nAnd before you say that sounds like a niche issue - it's not in my\nexperience. Forgetting the necessary indexes two or three foreign keys\ndown a CASCADE chain seems to be more common than doing so for tables\ndirectly \"linked\" with busy ones.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 4 Nov 2019 10:00:28 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Include triggers in EXPLAIN"}],"string":"[\n {\n \"msg_contents\": \"Hello!\\n\\nRecently I got few times into situation where I was trying to find out what\\nis blocking DELETE queries. Running EXPLAIN (even VERBOSE one) wasn't\\nuseful, since the reason was slow trigger (missing index on foreign key\\ncolumn). I had to create testing entry and run EXPLAIN ANALYZE DELETE to\\nget this information.\\n\\nIt will be really valuable for me to show triggers in EXPLAIN query since\\nit will make clear for me there will be any trigger \\\"activated\\\" during\\nexecution of DELETE query and that can be the reason for slow DELETE.\\n\\nI have seen initial discussion at\\nhttps://www.postgresql.org/message-id/flat/20693.1111732761%40sss.pgh.pa.us\\nto show time spent in triggers in EXPLAIN ANALYZE including quick\\ndiscussion to possibly show triggers during EXPLAIN. Anyway since it\\ndoesn't show any additional cost and just inform about the possibilities, I\\nstill consider this feature useful. This is probably implementation of idea\\nmentioned at\\nhttps://www.postgresql.org/message-id/21221.1111736869%40sss.pgh.pa.us by\\nTom Lane.\\n\\nAfter initial discussion with Pavel Stěhule and Tomáš Vondra at czech\\npostgresql maillist (\\nhttps://groups.google.com/forum/#!topic/postgresql-cz/Dq1sT7huVho) I was\\nable to prepare initial version of this patch. I have added EXPLAIN option\\ncalled TRIGGERS enabled by default.There's already autoexplain property for\\nthis. I understand it is not possible to show only triggers which will be\\nreally activated unless query is really executed. EXPLAIN ANALYZE remains\\nuntouched with this patch.\\n\\n- patch with examples can be found at\\nhttps://github.com/simi/postgres/pull/2\\n- DIFF format https://github.com/simi/postgres/pull/2.diff\\n- PATCH format (also attached) https://github.com/simi/postgres/pull/2.patch\\n\\nAll regression tests passed with this change locally on latest git master.\\nI would like to cover this patch with more regression tests, but I wasn't\\nsure where to place them, since there's no \\\"EXPLAIN\\\" related test \\\"group\\\".\\nIs \\\"src/test/regress/sql/triggers.sql\\\" the best place to add tests related\\nto this change?\\n\\nPS: This is my first try to contribute to postgresql codebase. The quality\\nof patch is probably not the best, but I will be more than happy to do any\\nrequested change if needed.\\n\\nRegards,\\nJosef Šimánek\",\n \"msg_date\": \"Sun, 3 Nov 2019 18:25:45 +0100\",\n \"msg_from\": \"=?UTF-8?B?Sm9zZWYgxaBpbcOhbmVr?= <josef.simanek@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[PATCH] Include triggers in EXPLAIN\"\n },\n {\n \"msg_contents\": \"=?UTF-8?B?Sm9zZWYgxaBpbcOhbmVr?= <josef.simanek@gmail.com> writes:\\n> Recently I got few times into situation where I was trying to find out what\\n> is blocking DELETE queries. Running EXPLAIN (even VERBOSE one) wasn't\\n> useful, since the reason was slow trigger (missing index on foreign key\\n> column). I had to create testing entry and run EXPLAIN ANALYZE DELETE to\\n> get this information.\\n\\n> It will be really valuable for me to show triggers in EXPLAIN query since\\n> it will make clear for me there will be any trigger \\\"activated\\\" during\\n> execution of DELETE query and that can be the reason for slow DELETE.\\n\\nI don't really see the point of this patch? You do get the trigger\\ntimes during EXPLAIN ANALYZE, and I don't believe that a plain EXPLAIN\\nis going to have full information about what triggers might fire or\\nnot fire.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 03 Nov 2019 16:49:36 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Include triggers in EXPLAIN\"\n },\n {\n \"msg_contents\": \"Hello Tom.\\n\\nThanks for quick response. As I was testing this feature it shows all\\n\\\"possible\\\" triggers to be executed running given query. The benefit of\\nhaving this information in EXPLAIN as well is you do not need to execute\\nthe query (as EXPLAIN ANALYZE does). My usecase is to take a look at query\\nbefore it is executed to get some idea about the plan with EXPLAIN.\\n\\nDo you have idea about some case where actual trigger will be missing in\\nEXPLAIN with current implementation, but will be present in EXPLAIN\\nANALYZE? I can take a look if there's any way how to handle those cases as\\nwell.\\n\\nne 3. 11. 2019 v 22:49 odesílatel Tom Lane <tgl@sss.pgh.pa.us> napsal:\\n\\n> =?UTF-8?B?Sm9zZWYgxaBpbcOhbmVr?= <josef.simanek@gmail.com> writes:\\n> > Recently I got few times into situation where I was trying to find out\\n> what\\n> > is blocking DELETE queries. Running EXPLAIN (even VERBOSE one) wasn't\\n> > useful, since the reason was slow trigger (missing index on foreign key\\n> > column). I had to create testing entry and run EXPLAIN ANALYZE DELETE to\\n> > get this information.\\n>\\n> > It will be really valuable for me to show triggers in EXPLAIN query since\\n> > it will make clear for me there will be any trigger \\\"activated\\\" during\\n> > execution of DELETE query and that can be the reason for slow DELETE.\\n>\\n> I don't really see the point of this patch? You do get the trigger\\n> times during EXPLAIN ANALYZE, and I don't believe that a plain EXPLAIN\\n> is going to have full information about what triggers might fire or\\n> not fire.\\n>\\n> regards, tom lane\\n>\\n\\nHello Tom.Thanks for quick response. As I was testing this feature it shows all \\\"possible\\\" triggers to be executed running given query. The benefit of having this information in EXPLAIN as well is you do not need to execute the query (as EXPLAIN ANALYZE does). My usecase is to take a look at query before it is executed to get some idea about the plan with EXPLAIN.Do you have idea about some case where actual trigger will be missing in EXPLAIN with current implementation, but will be present in EXPLAIN ANALYZE? I can take a look if there's any way how to handle those cases as well.ne 3. 11. 2019 v 22:49 odesílatel Tom Lane <tgl@sss.pgh.pa.us> napsal:=?UTF-8?B?Sm9zZWYgxaBpbcOhbmVr?= <josef.simanek@gmail.com> writes:\\n> Recently I got few times into situation where I was trying to find out what\\n> is blocking DELETE queries. Running EXPLAIN (even VERBOSE one) wasn't\\n> useful, since the reason was slow trigger (missing index on foreign key\\n> column). I had to create testing entry and run EXPLAIN ANALYZE DELETE to\\n> get this information.\\n\\n> It will be really valuable for me to show triggers in EXPLAIN query since\\n> it will make clear for me there will be any trigger \\\"activated\\\" during\\n> execution of DELETE query and that can be the reason for slow DELETE.\\n\\nI don't really see the point of this patch? You do get the trigger\\ntimes during EXPLAIN ANALYZE, and I don't believe that a plain EXPLAIN\\nis going to have full information about what triggers might fire or\\nnot fire.\\n\\n regards, tom lane\",\n \"msg_date\": \"Mon, 4 Nov 2019 10:35:25 +0100\",\n \"msg_from\": \"=?UTF-8?B?Sm9zZWYgxaBpbcOhbmVr?= <josef.simanek@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Include triggers in EXPLAIN\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\n(minor note - on PG lists the style is to quote in-line and trip)\\n\\nOn 2019-11-04 10:35:25 +0100, Josef Šimánek wrote:\\n> Thanks for quick response. As I was testing this feature it shows all\\n> \\\"possible\\\" triggers to be executed running given query. The benefit of\\n> having this information in EXPLAIN as well is you do not need to execute\\n> the query (as EXPLAIN ANALYZE does). My usecase is to take a look at query\\n> before it is executed to get some idea about the plan with EXPLAIN.\\n\\nI can actually see some value in additional information here, but I'd\\nprobably want to change the format a bit. When explicitly desired (or\\nperhaps just in verbose mode?), I see value in counting the number of\\ntriggers we know about that need to be checked, how many were excluded\\non the basis of the trigger's WHEN clause etc.\\n\\n\\n> Do you have idea about some case where actual trigger will be missing in\\n> EXPLAIN with current implementation, but will be present in EXPLAIN\\n> ANALYZE? I can take a look if there's any way how to handle those cases as\\n> well.\\n\\nAny triggers that are fired because of other, listed, triggers causing\\nother changes. E.g. a logging trigger that inserts into a log table -\\nEXPLAIN, without ANALYZE, doesn't have a way of knowing about that.\\n\\nAnd before you say that sounds like a niche issue - it's not in my\\nexperience. Forgetting the necessary indexes two or three foreign keys\\ndown a CASCADE chain seems to be more common than doing so for tables\\ndirectly \\\"linked\\\" with busy ones.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Nov 2019 10:00:28 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Include triggers in EXPLAIN\"\n }\n]","isConversational":false}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":493,"numItemsPerPage":100,"numTotalItems":58371,"offset":49300,"length":100}},"jwt":"eyJhbGciOiJFZERTQSJ9.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc4MDkyOTc3Mywic3ViIjoiL2RhdGFzZXRzL3dhbnNoZW5sL3Bnc3FsLWhhY2tlcnMiLCJleHAiOjE3ODA5MzMzNzMsImlzcyI6Imh0dHBzOi8vaHVnZ2luZ2ZhY2UuY28ifQ.E9rEIcLljMntrEEomcYjEa2bOuu0SImFLSrAiOvc-iJ84WJQZnwBXmZeupZFZJ-OIi7rEH8a9scbHriBAMp3Ag","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": "Hi all,\n\n\n\nHow to fetch certain number of tuples from a postgres table.\n\n\n\nSame I am doing in oracle using following lines by setting prefetch attribute.\n\n\n\nFor oracle\n// Prepare query\n if( OCIStmtPrepare( myOciStatement, myOciError, (text *)aSqlStatement,\n// Get statement t... |
[
{
"msg_contents": "I'm not sure if this can be considered a bug or not, but it is perhaps\nunexpected. I found that when using a view that is simply select * from\ntable, then doing INSERT ... ON CONFLICT ON CONSTRAINT constraint_name on\nthat view, it does not find the constraint and errors out. But it does\... |
[
{
"msg_contents": "Hello,\n\nJehan-Guillaume (in Cc) reported me today a problem with logical\nreplication, where in case of network issue the walsender is correctly\nterminating at the given wal_sender_timeout but the logical worker\nkept waiting indefinitely.\n\nThe issue is apparently a simple thinko, the ti... |
[
{
"msg_contents": "Greetings,\n\nlibpq since PostgreSQL-12 has stricter checks for integer values in\nconnection parameters. They were introduced by commit\nhttps://github.com/postgres/postgres/commit/e7a2217978d9cbb2149bfcb4ef1e45716cfcbefb\n.\n\nHowever in case of \"connect_timeout\" such an invalid integer v... |
[
{
"msg_contents": "While reviewing the partitionwise-join patch, I noticed $Subject，ie,\nthis in create_list_bounds():\n\n /*\n * Never put a null into the values array, flag instead for\n * the code further down below where we construct the actual\n ... |
[
{
"msg_contents": "Attached is a patch to implement change badges in our documentation.\n\nWhat's a change badge? It's my term for a visual cue in the documentation\nused to indicate that the nearby section of documentation is new in this\nversion or otherwise changed from the previous version.\n\nOne example o... |
[
{
"msg_contents": "I've been struggling with how we're going to upgrade launchpad.net to\nPostgreSQL 12 or newer (we're currently on 10). We're one of those\napplications that deliberately uses CTEs as optimization fences in a few\ndifficult places. The provision of the MATERIALIZED keyword in 12 is\ngreat, b... |
[
{
"msg_contents": "Hi Hackers,\n\nBy optimising our application I stumbled over the join quals used very often in our application.\nIn general this concerns datasets, which are subdivided into chunks, like years, seasons (here half a year), multiple tenants in OLTP system etc.\nIn these cases many tables are jo... |
[
{
"msg_contents": "Hello,\n\nI am one of the primary maintainers of Pleroma, a federated social\nnetworking application written in Elixir, which uses PostgreSQL in\nways that may be considered outside the typical usage scenarios for\nPostgreSQL.\n\nNamely, we leverage JSONB heavily as a backing store for JSON-L... |
[
{
"msg_contents": "I am not sure if this causes any potential problems or not, but for\nconsistency of code seems we are missing below. All other places in code\nwhere sigsetjmp() exists for top level handling has error_context_stack set\nto NULL.\n\ndiff --git a/src/backend/postmaster/autovacuum.c\nb/src/backe... |
[
{
"msg_contents": "For all ppc compilers, implement compare_exchange and fetch_add with asm.\n\nThis is more like how we handle s_lock.h and arch-x86.h.\n\nReviewed by Tom Lane.\n\nDiscussion: https://postgr.es/m/20191005173400.GA3979129@rfd.leadboat.com\n\nBranch\n------\nmaster\n\nDetails\n-------\nhttps://gi... |
[
{
"msg_contents": "Hello,\n\n\nThe attached trivial patch fixes the initialization of the fake unlogged LSN. Currently, BootstrapXLOG() in initdb sets the initial fake unlogged LSN to FirstNormalUnloggedLSN (=1000), but the recovery and pg_resetwal sets it to 1. The patch modifies the latter two cases to matc... |
[
{
"msg_contents": "Hello hackers,\n\nI've noticed that the createuser utility supports two undocumented\noptions (--adduser, --no-adduser), that became obsolete in 2005.\nI believe that their existence should come to end someday (maybe\ntoday?). The patch to remove them is attached.\n\nBest regards.\nAlexander"... |
[
{
"msg_contents": "Hello,\n\nThis patch propose a new way to sample statement to logs.\n\nAs a reminder, this feature was committed in PG12[1] then reverted[2] after the\nproposition of log_statement_sample_limit[3]\n\nThe first implementation added a new GUC to sample statement logged by\nlog_min_duration_stat... |
[
{
"msg_contents": "Here’s a cut-down version of Umair Shahid’s blog post here:\n\nhttps://www.2ndquadrant.com/en/blog/postgresql-11-server-side-procedures-part-1/ <https://www.2ndquadrant.com/en/blog/postgresql-11-server-side-procedures-part-1/>\n__________\n\ncreate table t(k int primary key, v int not null);\... |
[
{
"msg_contents": "While reviewing Pavel's patch for a new option in Drop Database\ncommand [1], I noticed that the check for CountDBSubscriptions in\ndropdb() is done after we kill the autovac workers and allowed other\nbackends to exit via CountOtherDBBackends. Now, if there are already\nactive subscritions ... |
[
{
"msg_contents": "Hi,\n\nI found that the argument name of XLogFileInit() is wrong in its comment.\nAttached is the patch that fixes that typo.\n\nRegards,\n\n-- \nFujii Masao",
"msg_date": "Mon, 21 Oct 2019 15:57:43 +0900",
"msg_from": "Fujii Masao <masao.fujii@gmail.com>",
"msg_from_op": true,
... |
[
{
"msg_contents": "Hi hackers,\nI found this issue when restart standby node and then try to connect it.\nIt return \"psql: FATAL: the database system is starting up\".\n\n\nThe steps to reproduce this issue.\n1. Create a session to run uncommit_trans.sql\n2. Create the other session to do checkpoint\n3. Re... |
[
{
"msg_contents": "Hi all,\n\nWhile digging into the issues reported lately about REINDEX\nCONCURRENTLY, I have bumped into the following, independent, issue:\n/* Now open the relation of the new index, a lock is also needed on it */\nnewIndexRel = index_open(indexId, ShareUpdateExclusiveLock)\n\nIn this code p... |
[
{
"msg_contents": "\nBowerbird (Visual Studio 2017 / Windows 10 pro) just had a failure on\nthe pg_ctl test :\n<https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2019-10-21%2011%3A50%3A21>\n\n\nThere was a similar failure 17 days ago.\n\n\nI surmise that what's happening here is that the test... |
[
{
"msg_contents": "After setting up logical replication of a slowly changing table using the\nbuilt in pub/sub facility, I noticed way more network traffic than made\nsense. Looking into I see that every transaction in that database on the\nmaster gets sent to the replica. 99.999+% of them are empty transacti... |
[
{
"msg_contents": "Hello,\n\nI ran across an EXPLAIN plan and had some questions about some of its\ndetails. The BUFFERS docs say\n\n>The number of blocks shown for an upper-level node includes those used by\nall its child nodes.\n\nI initially assumed this would be cumulative, but I realized it's probably\nnot... |
[
{
"msg_contents": "Hello,\n\nWhile developing pgbench to allow partitioned tabled, I reproduced the \nstring management style used in the corresponding functions, but was \npretty unhappy with that kind of pattern:\n\n \tsnprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ...)\n\nHowever adding a feature is ... |
[
{
"msg_contents": "Hi,\n\nWhen I run pg_basebackup in v12 against v11, standby server fails to connecto\nprimary with the following error:\n\n2019-10-22 09:28:23.673 UTC [2375] FATAL: could not connect to the primary\nserver: invalid connection option \"gssencmode\"\n\nWhen I remove this from recovery.conf, it... |
[
{
"msg_contents": "Hello,\n\nI originally reported this in pgsql-bugs [0], but there wasn't much\nfeedback there, so moving the discussion here. When using JSON, YAML, or\nXML-format EXPLAIN on a plan that uses a parallelized sort, the Sort nodes\nlist two different entries for \"Workers\", one for the sort-rel... |
[
{
"msg_contents": "Hello\n\nTo benchmark with tpcb model, I tried to create a foreign key in the partitioned history table, but backend process killed by OOM.\nthe number of partitions is 8192. I tried in master(commit: ad4b7aeb84).\n\nI did the same thing in another server which has 200GB memory, but creating ... |
[
{
"msg_contents": "Hi,\n\nWhile working on some slides explaining EXPLAIN, I couldn't resist the\nurge to add the missing $SUBJECT. The attached 0001 patch gives the\nfollowing:\n\nGather ... time=0.146..33.077 rows=1 loops=1)\n Workers Planned: 2\n Workers Launched: 2\n Buffers: shared hit=4425\n -> Par... |
[
{
"msg_contents": "Hi all ,\n\nTemporal table is one of the main new features added in sql standard 2011.\n From that I will like to implement system versioned temporal table which\nallows to keep past and present data so old data can be queried. Am propose\nto implement it like below\n\nCREATE\n\nIn create tab... |
[
{
"msg_contents": "Hi,\n\nTL;DR: Some performance figures at the end. Lots of details before.\n\n\nFor a while I've been on and off (unfortunately more the latter), been\nhacking on improving expression evaluation further.\n\nThis is motivated by mainly two factors:\na) Expression evaluation is still often a ve... |
[
{
"msg_contents": "Hi,\n\nI tried to install PostgreSQL 12 with the \"Norwegian Bokmål, Norway\" locale in hope that it would, among other things, provide proper support for Norwegian characters out-of-the-box.\n\nBut initcluster.vbs appear to fail during post-install because the locale name contains a Norwegia... |
[
{
"msg_contents": "Hello,\n\nI noticed that our existing 2-param json{,b}_object functions take \ntext[] for both keys and values, so they are only able to build \none-layer-deep JSON objects. I'm interested in adding json{,b}_object \nfunctions that take text[] for the keys and json{,b}[] for the values. \nIt ... |
[
{
"msg_contents": "We have primary and hot standby databases running Postgres 11.3 inside Docker, with their data directories bind-mounted to a reflink-enabled XFS filesystem. The VM is running Debian's 4.19.16-1~bpo9+1 kernel inside an AWS EC2 instance.\n\nI've seen TOAST corruption in one of the standby datab... |
[
{
"msg_contents": "Currently I see the vacuum behavior for a table is that, even if a long\nrunning query on a different table is executing in another read committed\ntransaction.\nThat vacuum in the 1st transaction skips the dead rows until the long\nrunning query finishes.\nWhy that is the case, On same table... |
[
{
"msg_contents": "Hi all,\n\nWhile digging into a separate issue, I have found a new bug with\nREINDEX CONCURRENTLY. Once the new index is built and validated,\na couple of things are done at the swap phase, like switching\nconstraints, comments, and dependencies. The current code moves all\nthe dependency e... |
[
{
"msg_contents": "Hi,\n\nI recently discovered two possible bugs about synchronous replication.\n\n1. SyncRepCleanupAtProcExit may delete an element that has been deleted\nSyncRepCleanupAtProcExit first checks whether the queue is detached, if it is not detached, \nacquires the SyncRepLock lock and deletes it.... |
[
{
"msg_contents": "Today, I committed a patch (dddf4cdc) to reorder some of the file\nheader inclusions and buildfarm members prairiedog and locust failed\nas a result of that. The reason turns out to be that we have defined\na bool in pgtypeslib_extern.h and that definition is different from\nwhat we define i... |
[
{
"msg_contents": "Dear all,\n\nWe stumbled upon a few cases in which retrieving information from the\nforeign server may turn pretty useful before creating any foreign\ntable, especially info related to the catalog. E.g: a list of schemas\nor tables the user has access to.\n\nI thought of using dblink for it, ... |
[
{
"msg_contents": "Hi list,\n\nWhen investigating for the bug reported in thread \"logical replication -\nnegative bitmapset member not allowed\", I found a way to seg fault postgresql\nonly when cassert is enabled.\n\nSee the scenario in attachment.\n\nWhen executed against binaries compiled with --enable-cass... |
[
{
"msg_contents": "Under MinGW, when compiling the ecpg test files, you get these warnings:\n\nsqlda.pgc: In function 'dump_sqlda':\nsqlda.pgc:44:11: warning: unknown conversion type character 'l' in format [-Wformat=]\n printf(\"name sqlda descriptor: '%s' value %lld\\n\", sqlda->sqlvar[i].sqlname.data, *(l... |
[
{
"msg_contents": "The following bug has been logged on the website:\n\nBug reference: 16079\nLogged by: Yudhveer Kandukuri\nEmail address: k.yudhveer@gmail.com\nPostgreSQL version: 10.10\nOperating system: UBUNTU\nDescription: \n\nAs your team mentioned that LDAP process is not secu... |
[
{
"msg_contents": "Hi, \n\nI found a missing column value in the pg_stat_progress_cluster view document.\nI read the src/backend/catalog/system_views.sql file, there seems to be a possibility that 'writing new heap' is output in the 'phase' column.\nThe attached patch adds a description of the 'writing new heap... |
[
{
"msg_contents": "It seems to me that using IDENT_USERNAME_MAX for peer authentication is\nsome kind of historical leftover and not really appropriate or useful,\nso I propose the attached cleanup.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote D... |
[
{
"msg_contents": "Hi,\n\nOne of the function apply_typmod in numeric.c file present within #if 0.\nThis is like this for many years.\nI felt this can be removed.\nAttached patch contains the changes to handle removal of apply_typmod\npresent in #if 0.\nThoughts?\n\nRegards,\nVignesh\nEnterpriseDB: http://www.e... |
[
{
"msg_contents": "When the user modifies the REPLICA IDENTIFY field type, the logical \nreplication settings are lost.\n\nFor example:\n\npostgres=# \\d+ t1\n Table \"public.t1\"\n Column | Type | Collation | Nullable | Default | Storage | Stats \ntarget | Description\n-... |
[
{
"msg_contents": "Hi.\n\nI have noticed that it would be cool to use '==' in place of 'IS NOT\nDISTICT FROM'\n\nWhat do you think about this crazy idea?\n\n-- \nBest regards,\nEugen Konkov\n\n\n\n",
"msg_date": "Sat, 26 Oct 2019 18:41:10 +0300",
"msg_from": "Eugen Konkov <kes-kes@yandex.ru>",
"ms... |
[
{
"msg_contents": "At pgconf.eu, someone whose name I've forgotten pointed out to me\nthat this doesn't work:\n\nregression=# select (row(1, 2.0)).f1;\nERROR: could not identify column \"f1\" in record data type\nLINE 1: select (row(1, 2.0)).f1;\n ^\n\nThe fields of an anonymous rowtype are cert... |
[
{
"msg_contents": "commit #898e5e32 (Allow ATTACH PARTITION with only ShareUpdateExclusiveLock)\nupdates ddl.sgml but not alter_table.sgml, which only says:\n\nhttps://www.postgresql.org/docs/12/release-12.html\n|An ACCESS EXCLUSIVE lock is held unless explicitly noted.\n\nFind attached patch, which also improv... |
[
{
"msg_contents": "Hi folks\n\nI was recently surprised to notice that log_line_prefix doesn't support a\ncluster_name placeholder. I suggest adding one. If I don't hear objections\nI'll send a patch.\n\nBefore anyone asks \"but why?!\":\n\n* A constant (short) string in log_line_prefix is immensely useful when... |
[
{
"msg_contents": "Hi,\n\nCurrently, we need to scan the WHOLE shared buffers when VACUUM\ntruncated off any empty pages at end of transaction or when relation\nis TRUNCATEd.\nAs for our customer case, we periodically truncate thousands of tables,\nand it's possible to TRUNCATE single table per transaction. Thi... |
[
{
"msg_contents": "This patch adds const qualifiers to internal range type APIs. It \ndoesn't require any new casts or remove any old ones.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services",
"msg_date": "Mon, 28 Oct 201... |
[
{
"msg_contents": "As mentioned in [0], pg_upgrade currently does not preserve the version \nof collation objects created by initdb. Here is an attempt to fix that.\n\nThe way I deal with this here is by having the binary-upgrade mode in \npg_dump delete all the collations created by initdb and then dump out \... |
[
{
"msg_contents": "Greetings hackers,\n\nBefore PG12, select strpos('test', '') returns 1 (empty substring found at\nfirst position of the string), whereas starting with PG12 it returns 0\n(empty substring not found).\n\nIs this behavior change intentional? If so, it doesn't seem to be\ndocumented in the releas... |
[
{
"msg_contents": "Hi,\n\nPostgreSQL 10 introduced extended statistics, allowing us to consider\ncorrelation between columns to improve estimates, and PostgreSQL 12\nadded support for MCV statistics. But we still had the limitation that\nwe only allowed using a single extended statistics per relation, i.e.\ngiv... |
[
{
"msg_contents": "Hi,\n\nPlease find the attached patch having the fix for the typos and\ninconsistencies present in code.\nThe patch contains the following changes:\n1) attibute -> attribute\n2) efficent -> efficient\n3) becuase -> because\n4) fallthru -> fall through\n5) uncoming -> upcoming\n6) ans -> and\n... |
[
{
"msg_contents": "Hi,\n\nI've groused about this a few times, but to me it seems wrong that\nHashJoin and Hash are separate nodes. They're so tightly bound together\nthat keeping them separate just doesn't architecturally makes sense,\nimo. So I wrote a prototype.\n\nEvidence of being tightly bound together:\n... |
[
{
"msg_contents": "The attached patch teaches psql to redisplay any not-yet-executed\nquery text after editing with \\e. The fact that you don't get to\nsee what you're about to execute has been complained of before,\nmost recently at bug #16034 [1]. In that thread I complained that\nwe needed some probably-n... |
[
{
"msg_contents": "Hello,\nI think this demonstrates a bug, tested in 11.5:\nhttps://gist.github.com/wkalt/a298fe82c564668c803b3537561e67a0\n\nThe same script succeeds if the index on line 11 is either dropped, made to\nbe non-partial on b, or shifted to a different column (the others are used\nin the partition... |
[
{
"msg_contents": "The cache_plan argument to ri_PlanCheck has not been used since\ne8c9fd5fdf768323911f7088e8287f63b513c3c6. I propose to remove it.\n\nThat commit said \"I left it alone in case there is any future need for \nit\" but there hasn't been a need in 7 years, and I find it confusing to \nhave an u... |
[
{
"msg_contents": "Continuing the discussion from [0] and [1], here is a patch that \nautomates the process of updating Unicode derived files. Summary:\n\n- Edit UNICODE_VERSION and/or CLDR_VERSION in src/Makefile.global.in\n- Run make update-unicode\n- Commit\n\nI have added that to the release checklist in R... |
[
{
"msg_contents": "When joining tables with USING, the listed columns are merged and no\nlonger belong to either the left or the right side. That means they can\nno longer be qualified which can often be an inconvenience.\n\n\nSELECT a.x, b.y, z FROM a INNER JOIN b USING (z);\n\n\nThe SQL standard provides a w... |
[
{
"msg_contents": "Hi.\n\nThis is not clear from doc, so I have asked on IRC too.\n\nfrom the DOC: https://www.postgresql.org/docs/current/trigger-definition.html\nIn the case of INSTEAD OF triggers, the possibly-modified row returned by each trigger becomes the input to the next trigger\n\nI modify OLD row, t... |
[
{
"msg_contents": "Hi,\n\none of the most frequent conflicts I see is that two patches add files\nto OBJS (or one of its other spellings), and there are conflicts because\nanother file has been added.\n\nRight now there's two reasons why that's likely to happen:\n1) By listing multiple objects for each line, we... |
[
{
"msg_contents": "Hi,\n\nWhen using -b, --bkp-details pg_waldump outputs an unnecessary newline\nfor blocks that contain an FPW.\n\nIn --bkp-details block references are output on their own lines, like:\n\nrmgr: SPGist len (rec/tot): 4348/ 4348, tx: 980, lsn: 0/01985818, prev 0/01983850, desc: P... |
[
{
"msg_contents": "Hi,\n\nThis patch, in a slightly rougher form, was submitted as part of [1],\nbut it seems worth bringing up separately, rather than just committing\nhearing no objections.\n\n From the commit message:\n\n Make StringInfo available to frontend code.\n\n There's plenty places in frontend... |
[
{
"msg_contents": "Hello,\n\n\nI propose new simple sql query, which shows total block numbers in the \nrelation.\n\nI now reviewing this patch (https://commitfest.postgresql.org/25/2211/) \nand I think,\nit is usefull for knowing how many blocks there are in the relation to \ndetermine whether we use VACUUM RE... |
[
{
"msg_contents": "A global index by very definition is a single index on the parent table\nthat maps to many\nunderlying table partitions. The parent table itself does not have any\nunderlying storage,\nso it must, therefore, retrieve the data satisfying index constraints from\nthe underlying tables.\nIn very ... |
[
{
"msg_contents": "Please consider this scenario (race conditions):\n\n1. FlushBuffer() has written the buffer but hasn't yet managed to clear the\nBM_DIRTY flag (however BM_JUST_DIRTIED could be cleared by now).\n\n2. Another backend modified a hint bit and called MarkBufferDirtyHint().\n\n3. In MarkBufferDirt... |
[
{
"msg_contents": "HAVE_LONG_LONG_INT is now implied by the requirement for C99, so the \nseparate Autoconf check can be removed. The uses are almost all in ecpg \ncode, and AFAICT the check was originally added specifically for ecpg.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQ... |
[
{
"msg_contents": "While fooling with the NetBSD-vs-libpython issue noted in a nearby\nthread, I observed that the core regression tests sometimes hang up\nin the \"stats\" test on this platform (NetBSD 8.1/amd64). Investigation\nfound that the stats collector process was sometimes exiting like\nthis:\n\n2019-... |
[
{
"msg_contents": "Hi,\n\nCurrently CREATE OR REPLACE VIEW command fails if the column names\nare changed. For example,\n\n =# CREATE VIEW test AS SELECT 0 AS a;\n =# CREATE OR REPLACE VIEW test AS SELECT 0 AS x;\n ERROR: cannot change name of view column \"a\" to \"x\"\n\nI'd like to propose the atta... |
[
{
"msg_contents": "Hi all,\n\nI wondered can we have a shortcut somewhat similar to following POC\nin recomputeNamespacePath () when the recomputed path is the same as the\nprevious baseSearchPath/activeSearchPath :\n\n== POC patch ==\ndiff --git a/src/backend/catalog/namespace.c\nb/src/backend/catalog/namespac... |
[
{
"msg_contents": "AFAICT, these build options were only useful to maintain compatibility \nfor version-0 functions, but those are no longer supported, so these \noptions can be removed. There is a fair amount of code all over the \nplace to support these options, so the cleanup is quite significant.\n\nThe cu... |
[
{
"msg_contents": "Hi,\n\nI want to know how postgres stores catalog relations in cache in-depth. Is\nthere any documentation for that?\n\nHi,I want to know how postgres stores catalog relations in cache in-depth. Is there any documentation for that?",
"msg_date": "Thu, 31 Oct 2019 15:19:23 +0530",
"msg... |
[
{
"msg_contents": "Hi,\n\nThe command tag of ALTER MATERIALIZED VIEW is basically\n\"ALTER MATERIALIZED VIEW\". For example,\n\n =# ALTER MATERIALIZED VIEW test ALTER COLUMN j SET STATISTICS 100;\n ALTER MATERIALIZED VIEW\n =# ALTER MATERIALIZED VIEW test OWNER TO CURRENT_USER;\n ALTER MATERIALIZED ... |
[
{
"msg_contents": "Dear hackers,\n\nAs declared last month, I propose again the new ECPG grammar, DECLARE STATEMENT.\nThis had been committed once, but it removed from PG12 because of\nsome problems. \nIn this mail, I want to report some problems that previous implementation has,\nproduce a new solution, and at... |
[
{
"msg_contents": "Hi,\n\nThe attached patch adds an -a / --appname command line switch to \npg_basebackup, pg_receivewal and pg_recvlogical.\n\nThis is useful when f.ex. pg_receivewal needs to connect as a \nsynchronous client (synchronous_standby_names),\n\n pg_receivewal -h myhost -p 5432 -S replica1 -a rep... |
[
{
"msg_contents": "Hi\n\nI almost finished patch optimizing non volatile function calls.\n\nselect f(t.n) from t where f(t.n) > 10 and f(t.n) < 100; needs 3 calls of\nf() for each tuple,\nafter applying patch only 1.\n\nAny pros and cons ?\n\nHiI almost finished patch optimizing non volatile function calls.se... |
[
{
"msg_contents": "This is the first of a number of patches to enhance SSL functionality,\nparticularly w.r.t. passphrases.\n\n\nThis patch provides a hook for a function that can supply an SSL\npassphrase. The hook can be filled in by a shared preloadable module. In\norder for that to be effective, the startup... |
[
{
"msg_contents": "Hi everyone,\n\nI was taking a look at pg_stat_statements module and noticed that it does\nnot collect any percentile metrics. I believe that It would be really handy\nto have those available and I'd love to contribute with this feature.\n\nThe basic idea is to accumulate the the query execut... |
[
{
"msg_contents": "Hi\n\nlong time we are think how to allow add some custom commands in psql. I had\na following idea\n\n1. psql can has special buffer for custom queries. This buffer can be\nfilled by special command \\gdefq. This command will have two parameters -\nname and number of arguments.\n\nsome like\... |
[
{
"msg_contents": "This small patch authored by my colleague Craig Ringer enhances\nTestlib's command_fails_like by allowing the passing of extra keyword\ntype arguments. The keyword initially recognized is 'extra_ipcrun_opts'.\nThe value for this keyword needs to be an array, and is passed through\nto the call... |
[
{
"msg_contents": "Hello Devs,\n\nThis patch moves duplicated query cancellation code code from psql & \nscripts to fe-utils, so that it is shared and may be used by other \ncommands.\n\nThis is because Masao-san suggested to add a query cancellation feature to \npgbench for long queries (server-side data gener... |
[
{
"msg_contents": "Hi,\n\nWe currently align byval types such as int4/8, float4/8, timestamp *,\ndate etc, even though we mostly don't need to. When tuples are deformed,\nall byval types are copied out from the tuple data into the\ncorresponding Datum array, therefore the original alignment in the tuple\ndata d... |
[
{
"msg_contents": "This patch allows the superuser to grant passwordless connection rights\nin postgres_fdw user mappings.\n\n\nThe patch is authored by my colleague Craig Ringer, with slight bitrot\nfixed by me.\n\n\nOne use case for this is with passphrase-protected client certificates,\na patch for which wil... |
[
{
"msg_contents": "\nThis patch provides for an sslpassword parameter for libpq, and a hook\nthat a client can fill in for a callback function to set the password.\n\n\nThis provides similar facilities to those already available in the JDBC\ndriver.\n\n\nThere is also a function to fetch the sslpassword from th... |
[
{
"msg_contents": "Hello hackers,\n\nPlease feel free to edit this new page, which I'd like to use to keep\ntrack of observations, ideas and threads relating to hash joins.\n\nhttps://wiki.postgresql.org/wiki/Hash_Join\n\n\n",
"msg_date": "Fri, 1 Nov 2019 10:16:35 +1100",
"msg_from": "Thomas Munro <thom... |
[
{
"msg_contents": "Hi!\n\nOur customer faced with issue, when index is invisible after creation.\nThe reproducible case is following.\n\n $ psql db2\n # begin;\n # select txid_current();\n$ psql db1\n# select i as id, 0 as v into t from generate_series(1, 100000) i;\n# create unique index idx on t (id)... |
[
{
"msg_contents": "This patch achieves $SUBJECT and also provides some testing of the\nsslpassword setting.\n\n\ncheers\n\n\nandrew\n\n\n\n-- \nAndrew Dunstan https://www.2ndQuadrant.com\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services",
"msg_date": "Thu, 31 Oct 2019 19... |
[
{
"msg_contents": "Hi,\n\nSometimes you want to answer if a difference between two timestamps is\nlesser than x minutes but you are not sure which timestamp is greater\nthan the other one (to obtain a positive result -- it is not always\npossible). However, if you cannot obtain the absolute value of\nsubtractio... |
[
{
"msg_contents": "Hi,\n\n Postgres has a global variable `disable_cost`. It is set the value\n1.0e10.\n\n This value will be added to the cost of path if related GUC is set off.\nFor example,\n if enable_nestloop is set off, when planner trys to add nestloop join\npath, it continues\n to add such p... |
[
{
"msg_contents": "This patch moves the parse analysis component of ExecuteQuery() and\nEvaluateParams() into a new transformExecuteStmt() that is called from\ntransformStmt(). This makes EXECUTE behave more like other utility\ncommands.\n\nEffects are that error messages can have position information (see \nr... |
[
{
"msg_contents": "Hi.sorry for my English.I want to once again open the topic of 64 bit transaction id. I did not manage to find in the archive of the option that I want to discuss, so I write. If I searched poorly, then please forgive me.The idea is not very original and probably has already been considered, ... |
[
{
"msg_contents": "Hi, hackers\n\nAs the $Subject, does anyone have one? I'd like to refer to it, and\nwrite an example for people who is also looking for the document.\n\nThanks.\n\n-- \nAdam Lee\n\n\n",
"msg_date": "Fri, 1 Nov 2019 20:18:02 +0800",
"msg_from": "Adam Lee <ali@pivotal.io>",
"msg_fro... |
[
{
"msg_contents": "Hi,\n\n\nAs per the following code, t1 is a remote table through postgres_fdw:\n\n\n test=# BEGIN;\n BEGIN\n test=# SELECT * FROM t1;\n ...\n\n test=# PREPARE TRANSACTION 'gxid1';\n ERROR: cannot prepare a transaction that modified remote tables\n\n\nI have attached a patch... |
[
{
"msg_contents": "It would be useful to have CREATE INDEX CONCURRENTLY be ignored by\nvacuuming's OldestXmin. Frequently in OLTP scenarios, CIC transactions\nare severely disruptive because they are the only long-running\ntransactions in the system, and VACUUM has to keep rows only for their\nsake, pointlessl... |
[
{
"msg_contents": "Hello,\n\nI came across a surprising behavior when upgrading our PostgreSQL 10 DBs that\nalso serve as a destination for the logical replication of some reference tables.\n\npg_upgrade turns off all subscriptions on the cluster and doesn't turn them on.\nSpecifically, it creates them with con... |
[
{
"msg_contents": "Hi hackers,\n\nI just noticed that when contrib/seg was converted to V1 calling\nconvention (commit 389bb2818f4), the PG_GETARG_SEG_P() macro got defined\nin terms of PG_GETARG_POINTER(). But it itself calls DatumGetPointer(),\nso shouldn't it be using PG_GETARG_DATUM()?\n\nAttached is a pat... |
[
{
"msg_contents": "While monitoring pg_stat_subscription, I noticed that last_msg_send_time\nwas usually NULL, which doesn't make sense. Why would we have a message,\nbut not know when it was sent?\n\nLooking in src/backend/replication/walsender.c, there is this:\n\n /* output previously gathered data in a ... |
[
{
"msg_contents": "Hello Folks\n\nI would like to update unaccent.rules file to support Arabic letters. so could someone help me or tell me how could I add such contribution. I attached the file including the modifications, only the last 4 lines.\n\nthank you",
"msg_date": "Sun, 3 Nov 2019 06:02:19 +0000",
... |
[
{
"msg_contents": "Hello!\n\nRecently I got few times into situation where I was trying to find out what\nis blocking DELETE queries. Running EXPLAIN (even VERBOSE one) wasn't\nuseful, since the reason was slow trigger (missing index on foreign key\ncolumn). I had to create testing entry and run EXPLAIN ANALYZE... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.