writes:\\n> And in the end, if we set values like this from PG --- whether\\n> hard-wired or via a GUC --- the SSL library people will have exactly\\n> the same perspective with regards to *our* values. And not without\\n> reason; we were forcing very obsolete settings up till recently,\\n> because nobody had looked at the issue for a decade. I see no reason\\n> to expect that that history won't repeat itself.\\n\\nI'm not sure what you're saying here, but - I'm not sure how familiar\\nyou are with the OpenSSL API, but it's insecure by default. There is\\n*no other choice* for an application than to explicitly select which\\nprotocols it wants to use (or at least which protocols it wants to\\navoid). And you can't change OpenSSL, because a ton of old crappy\\nsoftware is going to break.\\n\\nDES\\n-- \\nDag-Erling Smørgrav - des@des.no\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 22 Oct 2014 15:16:56 +0200\",\n \"msg_from\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Tom Lane <tgl@sss.pgh.pa.us> writes:\\n> As far as protocol version goes, I think our existing coding basically\\n> says \\\"prefer newest available version, but at least TLS 1.0\\\". I think\\n> that's probably a reasonable approach.\\n\\nThe client side forces TLS 1.0:\\n\\n SSL_context = SSL_CTX_new(TLSv1_method());\\n\\nIn typical OpenSSL fashion, this does *not* mean 1.0 or higher. It\\nmeans 1.0 exactly.\\n\\n> If the patch exposed a GUC that set a \\\"minimum\\\" version, rather than\\n> calling out specific acceptable protocols, it might be less risky.\\n\\nNot necessarily. Someone might find a weakness in TLS 1.1 which is not\\npresent in 1.0 because it involves a specific algorithm or mode that 1.0\\ndoes not support.\\n\\nDES\\n-- \\nDag-Erling Smørgrav - des@des.no\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 22 Oct 2014 15:21:52 +0200\",\n \"msg_from\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Magnus Hagander <magnus@hagander.net> writes:\\n> Yes, it does that. Though it only does it on 9.4,but with the facts we\\n> know now, what 9.4+ does is perfectly safe.\\n\\nAgreed.\\n\\nDES\\n-- \\nDag-Erling Smørgrav - des@des.no\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 22 Oct 2014 15:22:17 +0200\",\n \"msg_from\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 22, 2014 at 3:12 PM, Dag-Erling Smørgrav <des@des.no> wrote:\\n\\n> Tom Lane <tgl@sss.pgh.pa.us> writes:\\n> > This looks to me like re-fighting the last war. Such a GUC has zero\\n> value\\n> > *unless* some situation exactly like the POODLE bug comes up again, and\\n> > the odds of that are not high.\\n>\\n> Many people would have said the exact same thing before POODLE, and\\n> BEAST, and CRIME, and Heartbleed. You never know what sort of bugs or\\n> weaknesses will show up or when; all you know is that there are a lot of\\n> people working very hard to find these things and exploit them, and that\\n> they *will* succeeded, again and again and again. You can gamble that\\n> PostgreSQL will not be vulnerable due to specific details of its\\n> protocol or how it uses TLS, but that's a gamble which you will\\n> eventually lose.\\n>\\nThere are some companies, without naming them, that have increased the\\nresources dedicated to analyze existing security protocols and libraries,\\nso even if the chances are low, IMO the occurence that similar problems\\nshow up are getting to increase wit the time.\\n\\n\\n> > Moreover, the GUC could easily be misused to decrease rather than\\n> increase\\n> > one's security, if it's carelessly set.\\n>\\n> That's the user's responsibility.\\n>\\nI actually just had a user knocking about having a way to control the\\nprotocols used by server... So, changing my opinion on the matter, that\\nwould be nice to have even such a parameter on back-branches, with\\n'default' to let the server decide which one is better.\\n-- \\nMichael\\n\\nOn Wed, Oct 22, 2014 at 3:12 PM, Dag-Erling Smørgrav <des@des.no> wrote:Tom Lane <tgl@sss.pgh.pa.us> writes:\\n> This looks to me like re-fighting the last war. Such a GUC has zero value\\n> *unless* some situation exactly like the POODLE bug comes up again, and\\n> the odds of that are not high.\\n\\nMany people would have said the exact same thing before POODLE, and\\nBEAST, and CRIME, and Heartbleed. You never know what sort of bugs or\\nweaknesses will show up or when; all you know is that there are a lot of\\npeople working very hard to find these things and exploit them, and that\\nthey *will* succeeded, again and again and again. You can gamble that\\nPostgreSQL will not be vulnerable due to specific details of its\\nprotocol or how it uses TLS, but that's a gamble which you will\\neventually lose.There are some companies, without naming them, that have increased the resources dedicated to analyze existing security protocols and libraries, so even if the chances are low, IMO the occurence that similar problems show up are getting to increase wit the time. \\n> Moreover, the GUC could easily be misused to decrease rather than increase\\n> one's security, if it's carelessly set.\\n\\nThat's the user's responsibility.I actually just had a user knocking about having a way to control the protocols used by server... So, changing my opinion on the matter, that would be nice to have even such a parameter on back-branches, with 'default' to let the server decide which one is better.-- Michael\",\n \"msg_date\": \"Wed, 22 Oct 2014 15:34:58 +0200\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 22, 2014 at 03:14:26PM +0200, Dag-Erling Smørgrav wrote:\\n> > In a case like POODLE we probably wouldn't have done it anyway, as it\\n> > doesn't affect our connections (we're not http)\\n> \\n> If I understand correctly, imaps has been shown to be vulnerable as\\n> well, so I wouldn't be so sure.\\n\\nReference? It's an SSL3 specific attack, so most servers are not\\nvulnerable because TLS will negotiate to the highest supported\\nprotocol. So unless one of the client/server doesn't support TLS1.0\\nthere's no issue. The only reason http is vulnerable is because\\nbrowsers do protocol downgrading, something strictly forbidden by the\\nspec.\\n\\nAdditionally, the man-in-the-middle must be able to control the padding\\nin the startup packet, which just isn't possible (no scripting language\\nin the client).\\n\\nSince you can already specify the cipher list, couldn't you just add\\n-SSLv3 to the cipher list and be done?\\n\\nHave a nice day,\\n-- \\nMartijn van Oosterhout <kleptog@svana.org> http://svana.org/kleptog/\\n> He who writes carelessly confesses thereby at the very outset that he does\\n> not attach much importance to his own thoughts.\\n -- Arthur Schopenhauer\",\n \"msg_date\": \"Wed, 22 Oct 2014 20:54:44 +0200\",\n \"msg_from\": \"Martijn van Oosterhout <kleptog@svana.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Martijn van Oosterhout <kleptog@svana.org> writes:\\n> Dag-Erling Smørgrav <des@des.no> writes:\\n> > If I understand correctly, imaps has been shown to be vulnerable as\\n> > well, so I wouldn't be so sure.\\n> Reference?\\n\\nSorry, no reference. I was told that Thunderbird was vulnerable to\\nPOODLE when talking imaps.\\n\\n> Since you can already specify the cipher list, couldn't you just add\\n> -SSLv3 to the cipher list and be done?\\n\\nI didn't want to change the existing behavior; all I wanted was to give\\nusers a way to do so if they wish.\\n\\nDES\\n-- \\nDag-Erling Smørgrav - des@des.no\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 22 Oct 2014 21:36:59 +0200\",\n \"msg_from\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 22, 2014 at 09:36:59PM +0200, Dag-Erling Smørgrav wrote:\\n> Martijn van Oosterhout <kleptog@svana.org> writes:\\n> > Dag-Erling Smørgrav <des@des.no> writes:\\n> > > If I understand correctly, imaps has been shown to be vulnerable as\\n> > > well, so I wouldn't be so sure.\\n> > Reference?\\n> \\n> Sorry, no reference. I was told that Thunderbird was vulnerable to\\n> POODLE when talking imaps.\\n\\nUgh, found it. It does the same connection fallback stuff as firefox.\\n\\nhttps://securityblog.redhat.com/2014/10/20/can-ssl-3-0-be-fixed-an-analysis-of-the-poodle-attack/\\n\\n> > Since you can already specify the cipher list, couldn't you just add\\n> > -SSLv3 to the cipher list and be done?\\n> \\n> I didn't want to change the existing behavior; all I wanted was to give\\n> users a way to do so if they wish.\\n\\nI think we should just disable SSL3.0 altogether. The only way this\\ncould cause problems is if people are using PostgreSQL with an OpenSSL\\nlibrary from last century. As for client libraries, even Windows XP\\nsupports TLS1.0.\\n\\nHave a nice day,\\n-- \\nMartijn van Oosterhout <kleptog@svana.org> http://svana.org/kleptog/\\n> He who writes carelessly confesses thereby at the very outset that he does\\n> not attach much importance to his own thoughts.\\n -- Arthur Schopenhauer\",\n \"msg_date\": \"Thu, 23 Oct 2014 08:30:37 +0200\",\n \"msg_from\": \"Martijn van Oosterhout <kleptog@svana.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Martijn van Oosterhout <kleptog@svana.org> writes:\\n> Dag-Erling Smørgrav <des@des.no> writes:\\n> > Martijn van Oosterhout <kleptog@svana.org> writes:\\n> > > Since you can already specify the cipher list, couldn't you just\\n> > > add -SSLv3 to the cipher list and be done?\\n> > I didn't want to change the existing behavior; all I wanted was to\\n> > give users a way to do so if they wish.\\n> I think we should just disable SSL3.0 altogether. The only way this\\n> could cause problems is if people are using PostgreSQL with an OpenSSL\\n> library from last century. As for client libraries, even Windows XP\\n> supports TLS1.0.\\n\\nAs far as I'm concerned (i.e. as far as FreeBSD and the University of\\nOslo are concerned), I couldn't care less about anything older than\\n0.9.8, which is what FreeBSD 8 and RHEL5 have, but I don't feel\\ncomfortable making that decision for other people. On the gripping\\nhand, no currently supported version of libpq uses anything older than\\nTLS; 9.0 through 9.3 use TLS 1.0 only while 9.4 uses TLS 1.0 or higher.\\n\\nDES\\n-- \\nDag-Erling Smørgrav - des@des.no\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 23 Oct 2014 09:30:24 +0200\",\n \"msg_from\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Dag-Erling Sm�rgrav wrote:\\n> Martijn van Oosterhout <kleptog@svana.org> writes:\\n> > Dag-Erling Sm�rgrav <des@des.no> writes:\\n> > > Martijn van Oosterhout <kleptog@svana.org> writes:\\n> > > > Since you can already specify the cipher list, couldn't you just\\n> > > > add -SSLv3 to the cipher list and be done?\\n> > > I didn't want to change the existing behavior; all I wanted was to\\n> > > give users a way to do so if they wish.\\n> > I think we should just disable SSL3.0 altogether. The only way this\\n> > could cause problems is if people are using PostgreSQL with an OpenSSL\\n> > library from last century. As for client libraries, even Windows XP\\n> > supports TLS1.0.\\n> \\n> As far as I'm concerned (i.e. as far as FreeBSD and the University of\\n> Oslo are concerned), I couldn't care less about anything older than\\n> 0.9.8, which is what FreeBSD 8 and RHEL5 have, but I don't feel\\n> comfortable making that decision for other people. On the gripping\\n> hand, no currently supported version of libpq uses anything older than\\n> TLS; 9.0 through 9.3 use TLS 1.0 only while 9.4 uses TLS 1.0 or higher.\\n\\nOpenSSL just announced a week or two ago that they're abandoning support\\nfor 0.9.8 by the end of next year[1], which means its replacements have\\nbeen around for a really long time. I think it's fine to drop 0.9.7\\nsupport --- we already dropped support for 0.9.6 with the renegotiation\\nrework[2] in the 9.4 timeframe.\\n\\nOpenSSL 0.9.7 has already not gotten fixes for all the latest flurry of\\nsecurity issues, so anyone *is* using SSL but not at least the 0.9.8\\nbranch, they are in trouble.\\n\\n[1] http://openssl.6102.n7.nabble.com/OpenSSL-0-9-8-End-Of-Life-Announcement-td54155.html\\n[2] http://www.postgresql.org/message-id/20130712203252.GH29206@eldon.alvh.no-ip.org\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 23 Oct 2014 06:58:50 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> OpenSSL just announced a week or two ago that they're abandoning support\\n> for 0.9.8 by the end of next year[1], which means its replacements have\\n> been around for a really long time.\\n\\nRHEL5 still has 0.9.8e with backported patches and will be supported\\nuntil 2017-03-31.\\n\\nFreeBSD 8.4, 9.1, 9.2 and 9.3 all have 0.9.8y with backported patches.\\n8.4, 9.1 and 9.2 all expire before OpenSSL 0.9.8, but 9.3 will be\\nsupported until 2016-12-31.\\n\\n0.9.8 and 1.0.1 are not binary compatible, so upgrading is *not* an\\noption. We (as in FreeBSD) will have to make do - either develop our\\nown patches or adapt RedHat's.\\n\\n> OpenSSL 0.9.7 has already not gotten fixes for all the latest flurry of\\n> security issues, so anyone *is* using SSL but not at least the 0.9.8\\n> branch, they are in trouble.\\n\\nThe latest 0.9.8 still only has TLS 1.0, unless they're planning to\\nbackport 1.1 and 1.2 (which I seriously doubt).\\n\\nDES\\n-- \\nDag-Erling Smørgrav - des@des.no\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 23 Oct 2014 18:40:50 +0200\",\n \"msg_from\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no> writes:\\n> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n>> OpenSSL 0.9.7 has already not gotten fixes for all the latest flurry of\\n>> security issues, so anyone *is* using SSL but not at least the 0.9.8\\n>> branch, they are in trouble.\\n\\n> The latest 0.9.8 still only has TLS 1.0, unless they're planning to\\n> backport 1.1 and 1.2 (which I seriously doubt).\\n\\nThe upshot of this conversation still seems to be that we don't need to\\ndo anything. Unless I'm misunderstanding something:\\n\\n(1) No currently supported (or even recently supported) version of either\\nthe backend or libpq will select protocol less than TLS 1.0 unless forced\\nto via (poorly chosen) configuration settings.\\n\\n(2) Anyone who is feeling paranoid about shutting off SSLv3 despite (1)\\ncan do so via the existing ssl_ciphers GUC parameter.\\n\\nSeems to me that's sufficient, not only for now but for the future;\\nexisting OpenSSL practice is that the ciphers string includes categories\\ncorresponding to protocol versions, so you can shut off an old\\nprotocol version there if you need to.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 23 Oct 2014 14:11:09 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Tom Lane <tgl@sss.pgh.pa.us> writes:\\n> Anyone who is feeling paranoid about shutting off SSLv3 despite (1)\\n> can do so via the existing ssl_ciphers GUC parameter [...] the ciphers\\n> string includes categories corresponding to protocol versions, so you\\n> can shut off an old protocol version there if you need to.\\n\\nThe overlap between SSL 3.0 and TLS 1.0 is 100%:\\n\\n% openssl ciphers SSLv2 | md5 \\nfe5ff23432f119364a1126ca0776c5db\\n% openssl ciphers SSLv3 | md5 \\nbde4e4a10b9c3f323c0632ad067e293a\\n% openssl ciphers TLSv1 | md5\\nbde4e4a10b9c3f323c0632ad067e293a\\n% openssl ciphers TLSv1.2 | md5\\n26c375b6bdefb018b9dd7df463658320\\n\\nThus, if you disable all SSL 3.0 ciphers, you also disable TLS 1.0.\\n\\nDES\\n-- \\nDag-Erling Smørgrav - des@des.no\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 23 Oct 2014 20:56:49 +0200\",\n \"msg_from\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 1:08 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> This looks to me like re-fighting the last war. Such a GUC has zero value\\n> *unless* some situation exactly like the POODLE bug comes up again, and\\n> the odds of that are not high.\\n\\nI think it's pretty common for flaws to be discovered in particular\\nprotocols - usually, but not always, the oldest ones. The cost of\\nadding a new GUC seems pretty small tom me compared to the appealing\\npotential for users to secure their installation by changing a\\nconfiguration setting rather than, say, having to wait for new\\npackages to be available to install.\\n\\n> Moreover, the GUC could easily be misused to decrease rather than increase\\n> one's security, if it's carelessly set. Remember that we only recently\\n> fixed bugs that prevented use of the latest TLS version. If we have a\\n> setting like this, I fully anticipate that people will set it to \\\"only use\\n> TLS 1.2\\\" and then forget that they ever did that; years from now they'll\\n> still be using 1.2 when it's deprecated.\\n\\ncheckpoint_segments can easily be misused to decrease rather than\\nincrease performance, and autovacuum_naptime can easily be misused to\\ndestroy rather than improve autovacuum behavior. If we only added\\nGUCs that couldn't be used to make things worse, we'd have no GUCs.\\nThe bottom line for me is that OpenSSL has (a) a seemingly\\nnever-ending supply of serious security flaws and (b) an exposed knob\\nintended to help users of OpenSSL mitigate the effects of those flaws.\\nExposing that knob to our users seems like a good plan; supporting\\nalternate SSL implementations might be an even better one, not that\\nthe two are mutually exclusive.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 23 Oct 2014 19:51:26 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Dag-Erling Sm�rgrav wrote:\\n> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> > OpenSSL just announced a week or two ago that they're abandoning support\\n> > for 0.9.8 by the end of next year[1], which means its replacements have\\n> > been around for a really long time.\\n> \\n> RHEL5 still has 0.9.8e with backported patches and will be supported\\n> until 2017-03-31.\\n> \\n> FreeBSD 8.4, 9.1, 9.2 and 9.3 all have 0.9.8y with backported patches.\\n> 8.4, 9.1 and 9.2 all expire before OpenSSL 0.9.8, but 9.3 will be\\n> supported until 2016-12-31.\\n> \\n> 0.9.8 and 1.0.1 are not binary compatible, so upgrading is *not* an\\n> option. We (as in FreeBSD) will have to make do - either develop our\\n> own patches or adapt RedHat's.\\n\\nI think you misread me. I was saying to desupport 0.9.7, not 0.9.8.\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 23 Oct 2014 20:55:46 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"\\nDag-Erling Smørgrav <des@des.no> writes:\\n>\\n> The attached patches add an ssl_protocols configuration option which\\n> control which versions of SSL or TLS the server will use. The syntax is\\n> similar to Apache's SSLProtocols directive, except that the list is\\n> colon-separated instead of whitespace-separated, although that is easy\\n> to change if it proves unpopular.\\n\\nHello,\\n\\nHere is my review of the patch against master branch:\\n\\n* The patch applies and compiles cleanly, except for sgml docs:\\n\\npostgres.xml:66141: element varlistentry: validity error : Element\\nvarlistentry content does not follow the DTD, expecting (term+ ,\\nlistitem), got (term indexterm listitem)\\n\\n </para></listitem></varlistentry><varlistentry\\n ^\\n\\n The fix is to move <indexterm> under the <term> tag in the material\\n added to config.sgml by the patch.\\n\\n* The patch works as advertised, though the only way to verify that\\n connections made with the protocol disabled by the GUC are indeed\\n rejected is to edit fe-secure-openssl.c to only allow specific TLS\\n versions. Adding configuration on the libpq side as suggested in the\\n original discussion might help here.\\n\\n* The code allows specifying SSLv2 and SSLv3 in the GUC, but removes\\n them forcibly after parsing the complete string (a warning is issued).\\n Should we also add a note about this to the documentation?\\n\\n* In case the list of allowed protocols comes empty a FATAL error\\n message is logged: \\\"could not set the protocol list (no valid\\n protocols available)\\\". I think it's worth changing that to \\\"could not\\n set SSL protocol list...\\\" All other related error/warning logs\\n include \\\"SSL\\\".\\n\\n* The added code follows conventions and looks good to me. I didn't\\n spot any problems in the parser. I've tried a good deal of different\\n values and all seem to be parsed correctly.\\n\\nI would try to apply patches for older branches if there is consensus\\nthat we really need this patch and we want to back-patch it.\\n\\nThanks.\\n--\\nAlex\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 19 Nov 2014 18:34:39 +0300\",\n \"msg_from\": \"Alex Shulgin <ash@commandprompt.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Alex Shulgin <ash@commandprompt.com> writes:\\n> * The patch works as advertised, though the only way to verify that\\n> connections made with the protocol disabled by the GUC are indeed\\n> rejected is to edit fe-secure-openssl.c to only allow specific TLS\\n> versions. Adding configuration on the libpq side as suggested in the\\n> original discussion might help here.\\n\\nI can easily do that, but I won't have time until next week or so.\\n\\nDES\\n-- \\nDag-Erling Smørgrav - des@des.no\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 20 Nov 2014 09:41:52 +0100\",\n \"msg_from\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 19, 2014 at 4:34 PM, Alex Shulgin <ash@commandprompt.com> wrote:\\n>\\n> Dag-Erling Smørgrav <des@des.no> writes:\\n>>\\n>> The attached patches add an ssl_protocols configuration option which\\n>> control which versions of SSL or TLS the server will use. The syntax is\\n>> similar to Apache's SSLProtocols directive, except that the list is\\n>> colon-separated instead of whitespace-separated, although that is easy\\n>> to change if it proves unpopular.\\n>\\n> Hello,\\n>\\n> Here is my review of the patch against master branch:\\n>\\n> * The code allows specifying SSLv2 and SSLv3 in the GUC, but removes\\n> them forcibly after parsing the complete string (a warning is issued).\\n> Should we also add a note about this to the documentation?\\n\\nI see no reason to accept them at all, if we're going to reject them\\nlater anyway.\\n\\nWe can argue (as was done earlier in this thread) if we can drop SSL\\n3.0 completely -- but we can *definitely* drop SSLv2, and we should.\\nBut anything that we're going to reject at a later stage anyway, we\\nshould reject early.\\n\\n-- \\n Magnus Hagander\\n Me: http://www.hagander.net/\\n Work: http://www.redpill-linpro.com/\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 20 Nov 2014 09:49:24 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Magnus Hagander <magnus@hagander.net> writes:\\n> Alex Shulgin <ash@commandprompt.com> writes:\\n> > * The code allows specifying SSLv2 and SSLv3 in the GUC, but removes\\n> > them forcibly after parsing the complete string (a warning is issued).\\n> > Should we also add a note about this to the documentation?\\n> I see no reason to accept them at all, if we're going to reject them\\n> later anyway.\\n>\\n> We can argue (as was done earlier in this thread) if we can drop SSL\\n> 3.0 completely -- but we can *definitely* drop SSLv2, and we should.\\n> But anything that we're going to reject at a later stage anyway, we\\n> should reject early.\\n\\nIt's not really \\\"early or late\\\", but rather \\\"within the loop or at the\\nend of it\\\". From the users' perspective, the difference is that they\\nget (to paraphrase) \\\"SSLv2 is not allowed\\\" instead of \\\"syntax error\\\" and\\nthat they can use constructs such as \\\"ALL:-SSLv2\\\".\\n\\nDES\\n-- \\nDag-Erling Smørgrav - des@des.no\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 20 Nov 2014 10:19:19 +0100\",\n \"msg_from\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 20, 2014 at 10:19 AM, Dag-Erling Smørgrav <des@des.no> wrote:\\n> Magnus Hagander <magnus@hagander.net> writes:\\n>> Alex Shulgin <ash@commandprompt.com> writes:\\n>> > * The code allows specifying SSLv2 and SSLv3 in the GUC, but removes\\n>> > them forcibly after parsing the complete string (a warning is issued).\\n>> > Should we also add a note about this to the documentation?\\n>> I see no reason to accept them at all, if we're going to reject them\\n>> later anyway.\\n>>\\n>> We can argue (as was done earlier in this thread) if we can drop SSL\\n>> 3.0 completely -- but we can *definitely* drop SSLv2, and we should.\\n>> But anything that we're going to reject at a later stage anyway, we\\n>> should reject early.\\n>\\n> It's not really \\\"early or late\\\", but rather \\\"within the loop or at the\\n> end of it\\\". From the users' perspective, the difference is that they\\n> get (to paraphrase) \\\"SSLv2 is not allowed\\\" instead of \\\"syntax error\\\" and\\n> that they can use constructs such as \\\"ALL:-SSLv2\\\".\\n\\nAh, I see now - I hadn't looked at the code, just the review comment.\\nIt's a \\\"fallout\\\" from the reverse logic in openssl. Then it makes a\\nlot more sense.\\n\\n-- \\n Magnus Hagander\\n Me: http://www.hagander.net/\\n Work: http://www.redpill-linpro.com/\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 20 Nov 2014 10:26:37 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"\\nDag-Erling Smørgrav <des@des.no> writes:\\n\\n> Alex Shulgin <ash@commandprompt.com> writes:\\n>> * The patch works as advertised, though the only way to verify that\\n>> connections made with the protocol disabled by the GUC are indeed\\n>> rejected is to edit fe-secure-openssl.c to only allow specific TLS\\n>> versions. Adding configuration on the libpq side as suggested in the\\n>> original discussion might help here.\\n>\\n> I can easily do that, but I won't have time until next week or so.\\n\\nI can do that too, just need a hint where to look at in libpq/psql to\\nadd the option.\\n\\nFor SSL we have sslmode and sslcompression, etc. in conninfo, so adding\\nsslprotocols seems to be an option. As an aside note: should we also\\nexpose a parameter to choose SSL ciphers (would be a separate patch)?\\n\\n--\\nAlex\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 20 Nov 2014 14:57:12 +0300\",\n \"msg_from\": \"Alex Shulgin <ash@commandprompt.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Alex Shulgin <ash@commandprompt.com> writes:\\n> I can do that too, just need a hint where to look at in libpq/psql to\\n> add the option.\\n\\nThe place to *enforce* the option is src/interfaces/libpq/fe-secure.c\\n(look for SSLv23_method() and SSL_CTX_set_options()). I haven't looked\\ninto how to set it.\\n\\nDES\\n-- \\nDag-Erling Smørgrav - des@des.no\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 21 Nov 2014 09:15:36 +0100\",\n \"msg_from\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"\\nDag-Erling Smørgrav <des@des.no> writes:\\n\\n> Alex Shulgin <ash@commandprompt.com> writes:\\n>> I can do that too, just need a hint where to look at in libpq/psql to\\n>> add the option.\\n>\\n> The place to *enforce* the option is src/interfaces/libpq/fe-secure.c\\n> (look for SSLv23_method() and SSL_CTX_set_options()). I haven't looked\\n> into how to set it.\\n\\nYes, I've figured it out. Guess we'd better share the ssl_protocol\\nvalue parser code between libpq and the backend. Any precedent?\\n\\n--\\nAlex\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 21 Nov 2014 12:13:10 +0300\",\n \"msg_from\": \"Alex Shulgin <ash@commandprompt.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Alex Shulgin <ash@commandprompt.com> writes:\\n>>>\\n>>> I can do that too, just need a hint where to look at in libpq/psql to\\n>>> add the option.\\n>>\\n>> The place to *enforce* the option is src/interfaces/libpq/fe-secure.c\\n>> (look for SSLv23_method() and SSL_CTX_set_options()). I haven't looked\\n>> into how to set it.\\n>\\n> Yes, I've figured it out. Guess we'd better share the ssl_protocol\\n> value parser code between libpq and the backend. Any precedent?\\n\\nOK, looks like I've come up with something workable: I've added\\nsslprotocol connection string keyword similar to pre-existing\\nsslcompression, etc.\\n\\nPlease see attached v2 of the original patch. I'm having doubts about\\nthe name of openssl.h header though, libpq-openssl.h?\\n\\n--\\nAlex\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Wed, 26 Nov 2014 21:52:46 +0300\",\n \"msg_from\": \"Alex Shulgin <ash@commandprompt.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Alex Shulgin <ash@commandprompt.com> writes:\\n> OK, looks like I've come up with something workable: I've added\\n> sslprotocol connection string keyword similar to pre-existing\\n> sslcompression, etc. Please see attached v2 of the original patch.\\n> I'm having doubts about the name of openssl.h header though,\\n> libpq-openssl.h?\\n\\nPerhaps ssloptions.[ch], unless you plan to add non-option-related code\\nthere later?\\n\\nBTW, there is no Regent code in your openssl.c, so the copyright\\nstatement is incorrect.\\n\\nDES\\n-- \\nDag-Erling Smørgrav - des@des.no\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 27 Nov 2014 11:47:24 +0100\",\n \"msg_from\": \"=?utf-8?Q?Dag-Erling_Sm=C3=B8rgrav?= <des@des.no>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"\\nDag-Erling Smørgrav <des@des.no> writes:\\n\\n> Alex Shulgin <ash@commandprompt.com> writes:\\n>> OK, looks like I've come up with something workable: I've added\\n>> sslprotocol connection string keyword similar to pre-existing\\n>> sslcompression, etc. Please see attached v2 of the original patch.\\n>> I'm having doubts about the name of openssl.h header though,\\n>> libpq-openssl.h?\\n>\\n> Perhaps ssloptions.[ch], unless you plan to add non-option-related code\\n> there later?\\n\\nI don't think anything else than common options-related code fits in\\nthere, so ssloptions.c makes sense to me.\\n\\n> BTW, there is no Regent code in your openssl.c, so the copyright\\n> statement is incorrect.\\n\\nGood catch, I just blindly copied that from some other file.\\n\\n--\\nAlex\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 27 Nov 2014 14:51:46 +0300\",\n \"msg_from\": \"Alex Shulgin <ash@commandprompt.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 27, 2014 at 8:51 PM, Alex Shulgin <ash@commandprompt.com> wrote:\\n>\\n> Dag-Erling Smørgrav <des@des.no> writes:\\n>\\n>> Alex Shulgin <ash@commandprompt.com> writes:\\n>>> OK, looks like I've come up with something workable: I've added\\n>>> sslprotocol connection string keyword similar to pre-existing\\n>>> sslcompression, etc. Please see attached v2 of the original patch.\\n>>> I'm having doubts about the name of openssl.h header though,\\n>>> libpq-openssl.h?\\n>>\\n>> Perhaps ssloptions.[ch], unless you plan to add non-option-related code\\n>> there later?\\n>\\n> I don't think anything else than common options-related code fits in\\n> there, so ssloptions.c makes sense to me.\\n>\\n>> BTW, there is no Regent code in your openssl.c, so the copyright\\n>> statement is incorrect.\\n>\\n> Good catch, I just blindly copied that from some other file.\\nThere have been arguments in favor and against this patch... But\\nmarking it as returned with feedback because of a lack of activity in\\nthe last couple of weeks. Feel free to update if you think that's not\\ncorrect, and please move this patch to commit fest 2014-12.\\n--\\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 15 Dec 2014 12:02:07 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael.paquier@gmail.com> writes:\\n>>>\\n>>> Perhaps ssloptions.[ch], unless you plan to add non-option-related code\\n>>> there later?\\n>>\\n>> I don't think anything else than common options-related code fits in\\n>> there, so ssloptions.c makes sense to me.\\n>>\\n>>> BTW, there is no Regent code in your openssl.c, so the copyright\\n>>> statement is incorrect.\\n>>\\n>> Good catch, I just blindly copied that from some other file.\\n> There have been arguments in favor and against this patch... But\\n> marking it as returned with feedback because of a lack of activity in\\n> the last couple of weeks. Feel free to update if you think that's not\\n> correct, and please move this patch to commit fest 2014-12.\\n\\nAttached is a new version that addresses the earlier feedback: renamed\\nthe added *.[ch] files and removed incorrect copyright line.\\n\\nI'm moving this to the current CF.\\n\\n--\\nAlex\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Mon, 15 Dec 2014 17:15:56 +0300\",\n \"msg_from\": \"Alex Shulgin <ash@commandprompt.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 15, 2014 at 11:15 PM, Alex Shulgin <ash@commandprompt.com> wrote:\\n> Michael Paquier <michael.paquier@gmail.com> writes:\\n>>>>\\n>>>> Perhaps ssloptions.[ch], unless you plan to add non-option-related code\\n>>>> there later?\\n>>>\\n>>> I don't think anything else than common options-related code fits in\\n>>> there, so ssloptions.c makes sense to me.\\n>>>\\n>>>> BTW, there is no Regent code in your openssl.c, so the copyright\\n>>>> statement is incorrect.\\n>>>\\n>>> Good catch, I just blindly copied that from some other file.\\n>> There have been arguments in favor and against this patch... But\\n>> marking it as returned with feedback because of a lack of activity in\\n>> the last couple of weeks. Feel free to update if you think that's not\\n>> correct, and please move this patch to commit fest 2014-12.\\n>\\n> Attached is a new version that addresses the earlier feedback: renamed\\n> the added *.[ch] files and removed incorrect copyright line.\\n>\\n> I'm moving this to the current CF.\\nThis patch is statuquo since the beginning of this CF, at the\\narguments are standing the same. If there is nothing happening maybe\\nit would be better to mark it as returned with feedback? Thoughts?\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 15 Jan 2015 16:17:02 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 15, 2015 at 4:17 PM, Michael Paquier <michael.paquier@gmail.com>\\nwrote:\\n\\n> On Mon, Dec 15, 2014 at 11:15 PM, Alex Shulgin <ash@commandprompt.com>\\n> wrote:\\n> > Michael Paquier <michael.paquier@gmail.com> writes:\\n> >>>>\\n> >>>> Perhaps ssloptions.[ch], unless you plan to add non-option-related\\n> code\\n> >>>> there later?\\n> >>>\\n> >>> I don't think anything else than common options-related code fits in\\n> >>> there, so ssloptions.c makes sense to me.\\n> >>>\\n> >>>> BTW, there is no Regent code in your openssl.c, so the copyright\\n> >>>> statement is incorrect.\\n> >>>\\n> >>> Good catch, I just blindly copied that from some other file.\\n> >> There have been arguments in favor and against this patch... But\\n> >> marking it as returned with feedback because of a lack of activity in\\n> >> the last couple of weeks. Feel free to update if you think that's not\\n> >> correct, and please move this patch to commit fest 2014-12.\\n> >\\n> > Attached is a new version that addresses the earlier feedback: renamed\\n> > the added *.[ch] files and removed incorrect copyright line.\\n> >\\n> > I'm moving this to the current CF.\\n> This patch is statuquo since the beginning of this CF, at the\\n> arguments are standing the same. If there is nothing happening maybe\\n> it would be better to mark it as returned with feedback? Thoughts?\\n>\\n\\nI am not sure where we are moving on here, and if anybody cares much about\\nthis patch... Hence marked as rejected. Feel free to complain...\\n-- \\nMichael\\n\\nOn Thu, Jan 15, 2015 at 4:17 PM, Michael Paquier <michael.paquier@gmail.com> wrote:On Mon, Dec 15, 2014 at 11:15 PM, Alex Shulgin <ash@commandprompt.com> wrote:\\n> Michael Paquier <michael.paquier@gmail.com> writes:\\n>>>>\\n>>>> Perhaps ssloptions.[ch], unless you plan to add non-option-related code\\n>>>> there later?\\n>>>\\n>>> I don't think anything else than common options-related code fits in\\n>>> there, so ssloptions.c makes sense to me.\\n>>>\\n>>>> BTW, there is no Regent code in your openssl.c, so the copyright\\n>>>> statement is incorrect.\\n>>>\\n>>> Good catch, I just blindly copied that from some other file.\\n>> There have been arguments in favor and against this patch... But\\n>> marking it as returned with feedback because of a lack of activity in\\n>> the last couple of weeks. Feel free to update if you think that's not\\n>> correct, and please move this patch to commit fest 2014-12.\\n>\\n> Attached is a new version that addresses the earlier feedback: renamed\\n> the added *.[ch] files and removed incorrect copyright line.\\n>\\n> I'm moving this to the current CF.\\nThis patch is statuquo since the beginning of this CF, at the\\narguments are standing the same. If there is nothing happening maybe\\nit would be better to mark it as returned with feedback? Thoughts?\\nI am not sure where we are moving on here, and if anybody cares much about this patch... Hence marked as rejected. Feel free to complain...-- Michael\",\n \"msg_date\": \"Fri, 13 Feb 2015 17:13:11 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add ssl_protocols configuration option\"\n }\n]","isConversational":false}}},{"rowIdx":39676,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi everyone,\n\nVitesse DB 9.3.5.S is Postgres 9.3.5 with a LLVM-JIT query executor\ndesigned for compute intensive OLAP workload. We have gotten it to a\nreasonable state and would like to open it up to the pg hackers\ncommunity for testing and suggestions.\n\nVitesse DB offers\n-- JIT Compilation for compute-intensive queries\n-- CSV parsing with SSE instructions\n-- 100% binary compatibility with PG9.3.5.\n\nOur results show CSV imports run up to 2X faster, and TPCH Q1 runs 8X faster.\n\nOur TPCH 1GB benchmark results is also available at\nhttp://vitessedata.com/benchmark/ .\n\nPlease direct any questions by email to cktan@vitessedata.com .\n\nThank you for your help.\n\n--\nCK Tan\nVitesse Data, Inc.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 05:32:13 -0700","msg_from":"CK Tan <cktan@vitessedata.com>","msg_from_op":true,"msg_subject":"Vitesse DB call for testing"},{"msg_contents":"Hi,\n\n\nOn 2014-10-17 05:32:13 -0700, CK Tan wrote:\n> Vitesse DB 9.3.5.S is Postgres 9.3.5 with a LLVM-JIT query executor\n> designed for compute intensive OLAP workload. We have gotten it to a\n> reasonable state and would like to open it up to the pg hackers\n> community for testing and suggestions.\n> \n> Vitesse DB offers\n> -- JIT Compilation for compute-intensive queries\n> -- CSV parsing with SSE instructions\n> -- 100% binary compatibility with PG9.3.5.\n> \n> Our results show CSV imports run up to 2X faster, and TPCH Q1 runs 8X faster.\n\nHow are these modifications licensed?\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 14:37:54 +0200","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"On Fri, Oct 17, 2014 at 7:32 AM, CK Tan <cktan@vitessedata.com> wrote:\n> Hi everyone,\n>\n> Vitesse DB 9.3.5.S is Postgres 9.3.5 with a LLVM-JIT query executor\n> designed for compute intensive OLAP workload. We have gotten it to a\n> reasonable state and would like to open it up to the pg hackers\n> community for testing and suggestions.\n>\n> Vitesse DB offers\n> -- JIT Compilation for compute-intensive queries\n> -- CSV parsing with SSE instructions\n> -- 100% binary compatibility with PG9.3.5.\n>\n> Our results show CSV imports run up to 2X faster, and TPCH Q1 runs 8X faster.\n>\n> Our TPCH 1GB benchmark results is also available at\n> http://vitessedata.com/benchmark/ .\n>\n> Please direct any questions by email to cktan@vitessedata.com .\n\nYou offer a binary with 32k block size...what's the rationale for that?\n\nmerlin\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 08:14:14 -0500","msg_from":"Merlin Moncure <mmoncure@gmail.com>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"On Fri, Oct 17, 2014 at 8:14 AM, Merlin Moncure <mmoncure@gmail.com> wrote:\n> On Fri, Oct 17, 2014 at 7:32 AM, CK Tan <cktan@vitessedata.com> wrote:\n>> Hi everyone,\n>>\n>> Vitesse DB 9.3.5.S is Postgres 9.3.5 with a LLVM-JIT query executor\n>> designed for compute intensive OLAP workload. We have gotten it to a\n>> reasonable state and would like to open it up to the pg hackers\n>> community for testing and suggestions.\n>>\n>> Vitesse DB offers\n>> -- JIT Compilation for compute-intensive queries\n>> -- CSV parsing with SSE instructions\n>> -- 100% binary compatibility with PG9.3.5.\n>>\n>> Our results show CSV imports run up to 2X faster, and TPCH Q1 runs 8X faster.\n>>\n>> Our TPCH 1GB benchmark results is also available at\n>> http://vitessedata.com/benchmark/ .\n>>\n>> Please direct any questions by email to cktan@vitessedata.com .\n>\n> You offer a binary with 32k block size...what's the rationale for that?\n\n(sorry for the double post)\n\nOK, I downloaded the ubuntu binary and ran your benchmarks (after\nmaking some minor .conf tweaks like disabling SSL). I then ran your\nbenchmark (after fixing the typo) of the count/sum/avg test -- *and\nnoticed a 95% reduction in runtime performance* which is really quite\namazing IMNSHO. I also ran a select only test on small scale factor\npgbench and didn't see any regression there -- in fact you beat stock\nby ~ 3% (although this could be measurement noise). So now you've\ngot my attention. So, if you don't mind, quit being coy and explain\nhow the software works and all the neat things it does and doesn't do.\n\nmerlin\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 08:43:30 -0500","msg_from":"Merlin Moncure <mmoncure@gmail.com>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"Merlin, glad you tried it. \n\nWe take the query plan exactly as given by the planner, decide whether to JIT or to punt depending on the cost. If we punt, it goes back to pg executor. If we JIT, and if we could not proceed (usually of some operators we haven't implemented yet), we again punt. Once we were able to generate the code, there is no going back; we call into LLVM to obtain the function entry point, and run it to completion. The 3% improvement you see in OLTP tests is definitely noise.\n\nThe bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\n\nThe speed up is really noticeable when the data type is nonvarlena. In the varlena cases, we still call into pg routines most of the times. Again, try the sum,avg,count query on numeric, and you will see what I mean.\n\nAlso, we don't support UDF at the moment. So all queries involving UDF gets sent to pg executor.\n\nOn your question of 32k page size, the rational is that some of our customers could be interested in a data warehouse on pg. 32k page size is a big win when all you do is seqscan all day long.\n\nWe are looking for bug reports at these stage and some stress tests done without our own prejudices. Some test on real data in non prod setting on queries that are highly CPU bound would be ideal.\n\nThanks,\n-cktan\n\n> On Oct 17, 2014, at 6:43 AM, Merlin Moncure <mmoncure@gmail.com> wrote:\n> \n>> On Fri, Oct 17, 2014 at 8:14 AM, Merlin Moncure <mmoncure@gmail.com> wrote:\n>>> On Fri, Oct 17, 2014 at 7:32 AM, CK Tan <cktan@vitessedata.com> wrote:\n>>> Hi everyone,\n>>> \n>>> Vitesse DB 9.3.5.S is Postgres 9.3.5 with a LLVM-JIT query executor\n>>> designed for compute intensive OLAP workload. We have gotten it to a\n>>> reasonable state and would like to open it up to the pg hackers\n>>> community for testing and suggestions.\n>>> \n>>> Vitesse DB offers\n>>> -- JIT Compilation for compute-intensive queries\n>>> -- CSV parsing with SSE instructions\n>>> -- 100% binary compatibility with PG9.3.5.\n>>> \n>>> Our results show CSV imports run up to 2X faster, and TPCH Q1 runs 8X faster.\n>>> \n>>> Our TPCH 1GB benchmark results is also available at\n>>> http://vitessedata.com/benchmark/ .\n>>> \n>>> Please direct any questions by email to cktan@vitessedata.com .\n>> \n>> You offer a binary with 32k block size...what's the rationale for that?\n> \n> (sorry for the double post)\n> \n> OK, I downloaded the ubuntu binary and ran your benchmarks (after\n> making some minor .conf tweaks like disabling SSL). I then ran your\n> benchmark (after fixing the typo) of the count/sum/avg test -- *and\n> noticed a 95% reduction in runtime performance* which is really quite\n> amazing IMNSHO. I also ran a select only test on small scale factor\n> pgbench and didn't see any regression there -- in fact you beat stock\n> by ~ 3% (although this could be measurement noise). So now you've\n> got my attention. So, if you don't mind, quit being coy and explain\n> how the software works and all the neat things it does and doesn't do.\n> \n> merlin\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 08:47:14 -0700","msg_from":"CK Tan <cktan@vitessedata.com>","msg_from_op":true,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"On Fri, Oct 17, 2014 at 10:47 AM, CK Tan <cktan@vitessedata.com> wrote:\n> Merlin, glad you tried it.\n>\n> We take the query plan exactly as given by the planner, decide whether to JIT or to punt depending on the cost. If we punt, it goes back to pg executor. If we JIT, and if we could not proceed (usually of some operators we haven't implemented yet), we again punt. Once we were able to generate the code, there is no going back; we call into LLVM to obtain the function entry point, and run it to completion. The 3% improvement you see in OLTP tests is definitely noise.\n>\n> The bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\n>\n> The speed up is really noticeable when the data type is nonvarlena. In the varlena cases, we still call into pg routines most of the times. Again, try the sum,avg,count query on numeric, and you will see what I mean.\n>\n> Also, we don't support UDF at the moment. So all queries involving UDF gets sent to pg executor.\n>\n> On your question of 32k page size, the rational is that some of our customers could be interested in a data warehouse on pg. 32k page size is a big win when all you do is seqscan all day long.\n>\n> We are looking for bug reports at these stage and some stress tests done without our own prejudices. Some test on real data in non prod setting on queries that are highly CPU bound would be ideal.\n\nOne thing that I noticed is that when slamming your benchmark query\nvia pgbench, resident memory consumption was really aggressive and\nwould have taken down the server had I not spuriously stopped the\ntest. Memory consumption did return to baseline after that so I\nfigured some type of llvm memory management games were going on. This\nisn't really a problem for most OLAP workloads but it's something to\nbe aware of.\n\nVia 'perf top' on stock postgres, you see the usual suspects: palloc,\nhash_search_etc. On your build though HeapTuplesSatisfiesMVCC zooms\nright to the top of the stack which is pretty interesting...the\nexecutor is you've built is very lean and mean for sure. A drop in\noptimization engine with little no schema/sql changes is pretty neat\n-- your primary competitor here is going to be column organized type\ntable solutions to olap type problems.\n\nmerlin\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 11:00:41 -0500","msg_from":"Merlin Moncure <mmoncure@gmail.com>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"CK Tan <cktan@vitessedata.com> writes:\n> The bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\n\nWell, that's pretty much cheating: it's too hard to disentangle what's\ncoming from JIT vs what's coming from using a different accumulator\ndatatype. If we wanted to depend on having int128 available we could\nget that speedup with a couple hours' work.\n\nBut what exactly are you \"compiling\" here? I trust not the actual data\naccesses; that seems far too complicated to try to inline.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 13:12:27 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"Hi, Tom,\n\nSorry for double post to you.\n\nFeng\n\n---------- Forwarded message ----------\nFrom: Feng Tian <ftian@vitessedata.com>\nDate: Fri, Oct 17, 2014 at 10:29 AM\nSubject: Re: [HACKERS] Vitesse DB call for testing\nTo: Tom Lane <tgl@sss.pgh.pa.us>\n\n\nHi, Tom,\n\nI agree using that using int128 in stock postgres will speed up things\ntoo. On the other hand, that is only one part of the equation. For\nexample, if you look at TPCH Q1, the int128 \"cheating\" does not kick in at\nall, but we are 8x faster.\n\nI am not sure why do you mean by \"actual data access\". Data is still in\nstock postgres format on disk. We indeed jit-ed all data fields access\n(deform tuple). To put things in perspective, I just timed select\ncount(*) and select count(l_orderkey) from tpch1.lineitem; Our code is\nbottlenecked by memory bandwidth and difference is pretty much invisible.\n\nThanks,\nFeng\n\n\nftian=# set vdb_jit = 0;\n\nSET\n\nTime: 0.155 ms\n\nftian=# select count(*) from tpch1.lineitem;\n\n count\n\n---------\n\n 6001215\n\n(1 row)\n\n\nTime: 688.658 ms\n\nftian=# select count(*) from tpch1.lineitem;\n\n count\n\n---------\n\n 6001215\n\n(1 row)\n\n\nTime: 690.753 ms\n\nftian=# select count(l_orderkey) from tpch1.lineitem;\n\n count\n\n---------\n\n 6001215\n\n(1 row)\n\n\nTime: 819.452 ms\n\nftian=# set vdb_jit = 1;\n\nSET\n\nTime: 0.167 ms\n\nftian=# select count(*) from tpch1.lineitem;\n\n count\n\n---------\n\n 6001215\n\n(1 row)\n\n\nTime: 203.543 ms\n\nftian=# select count(l_orderkey) from tpch1.lineitem;\n\n count\n\n---------\n\n 6001215\n\n(1 row)\n\n\nTime: 202.253 ms\n\nftian=#\n\n\n\n\nOn Fri, Oct 17, 2014 at 10:12 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> CK Tan <cktan@vitessedata.com> writes:\n> > The bigint sum,avg,count case in the example you tried has some\n> optimization. We use int128 to accumulate the bigint instead of numeric in\n> pg. Hence the big speed up. Try the same query on int4 for the improvement\n> where both pg and vitessedb are using int4 in the execution.\n>\n> Well, that's pretty much cheating: it's too hard to disentangle what's\n> coming from JIT vs what's coming from using a different accumulator\n> datatype. If we wanted to depend on having int128 available we could\n> get that speedup with a couple hours' work.\n>\n> But what exactly are you \"compiling\" here? I trust not the actual data\n> accesses; that seems far too complicated to try to inline.\n>\n> regards, tom lane\n>\n>\n> --\n> Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\n> To make changes to your subscription:\n> http://www.postgresql.org/mailpref/pgsql-hackers\n>\n\nHi, Tom,Sorry for double post to you.Feng---------- Forwarded message ----------From: Feng Tian <ftian@vitessedata.com>Date: Fri, Oct 17, 2014 at 10:29 AMSubject: Re: [HACKERS] Vitesse DB call for testingTo: Tom Lane <tgl@sss.pgh.pa.us>Hi, Tom,I agree using that using int128 in stock postgres will speed up things too. On the other hand, that is only one part of the equation. For example, if you look at TPCH Q1, the int128 \"cheating\" does not kick in at all, but we are 8x faster.I am not sure why do you mean by \"actual data access\". Data is still in stock postgres format on disk. We indeed jit-ed all data fields access (deform tuple). To put things in perspective, I just timed select count(*) and select count(l_orderkey) from tpch1.lineitem; Our code is bottlenecked by memory bandwidth and difference is pretty much invisible.Thanks,Feng\nftian=# set vdb_jit = 0;\nSET\nTime: 0.155 ms\nftian=# select count(*) from tpch1.lineitem;\n count \n---------\n 6001215\n(1 row)\n\nTime: 688.658 ms\nftian=# select count(*) from tpch1.lineitem;\n count \n---------\n 6001215\n(1 row)\n\nTime: 690.753 ms\nftian=# select count(l_orderkey) from tpch1.lineitem;\n count \n---------\n 6001215\n(1 row)\n\nTime: 819.452 ms\nftian=# set vdb_jit = 1;\nSET\nTime: 0.167 ms\nftian=# select count(*) from tpch1.lineitem;\n count \n---------\n 6001215\n(1 row)\n\nTime: 203.543 ms\nftian=# select count(l_orderkey) from tpch1.lineitem;\n count \n---------\n 6001215\n(1 row)\n\nTime: 202.253 ms\nftian=# On Fri, Oct 17, 2014 at 10:12 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:CK Tan <cktan@vitessedata.com> writes:\n> The bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\n\nWell, that's pretty much cheating: it's too hard to disentangle what's\ncoming from JIT vs what's coming from using a different accumulator\ndatatype. If we wanted to depend on having int128 available we could\nget that speedup with a couple hours' work.\n\nBut what exactly are you \"compiling\" here? I trust not the actual data\naccesses; that seems far too complicated to try to inline.\n\n regards, tom lane\n\n\n--\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Fri, 17 Oct 2014 11:08:48 -0700","msg_from":"Feng Tian <ftian@vitessedata.com>","msg_from_op":false,"msg_subject":"Fwd: Vitesse DB call for testing"},{"msg_contents":"CK,\n\nBefore we go any further on this, how is Vitesse currently licensed?\nlast time we talked it was still proprietary. If it's not being\nopen-sourced, we likely need to take discussion off this list.\n\n-- \nJosh Berkus\nPostgreSQL Experts Inc.\nhttp://pgexperts.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 11:13:00 -0700","msg_from":"Josh Berkus <josh@agliodbs.com>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"On Fri, Oct 17, 2014 at 11:08 AM, Feng Tian <ftian@vitessedata.com> wrote:\n> I agree using that using int128 in stock postgres will speed up things too.\n> On the other hand, that is only one part of the equation. For example, if\n> you look at TPCH Q1, the int128 \"cheating\" does not kick in at all, but we\n> are 8x faster.\n\nI'm curious about how the numbers look when stock Postgres is built\nwith the same page size as your fork. You didn't mention whether or\nnot your Postgres numbers came from a standard build.\n\n-- \nPeter Geoghegan\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 11:21:55 -0700","msg_from":"Peter Geoghegan <pg@heroku.com>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"On 2014-10-17 13:12:27 -0400, Tom Lane wrote:\n> Well, that's pretty much cheating: it's too hard to disentangle what's\n> coming from JIT vs what's coming from using a different accumulator\n> datatype. If we wanted to depend on having int128 available we could\n> get that speedup with a couple hours' work.\n\nI think doing that when configure detects int128 would make a great deal\nof sense. It's not like we'd save a great deal of complicated code by\nremoving the existing accumulator... We'd still have to return a\nnumeric, but that's likely lost in the noise cost wise.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 20:25:00 +0200","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"On Fri, Oct 17, 2014 at 1:21 PM, Peter Geoghegan <pg@heroku.com> wrote:\r\n> On Fri, Oct 17, 2014 at 11:08 AM, Feng Tian <ftian@vitessedata.com> wrote:\r\n>> I agree using that using int128 in stock postgres will speed up things too.\r\n>> On the other hand, that is only one part of the equation. For example, if\r\n>> you look at TPCH Q1, the int128 \"cheating\" does not kick in at all, but we\r\n>> are 8x faster.\r\n>\r\n> I'm curious about how the numbers look when stock Postgres is built\r\n> with the same page size as your fork. You didn't mention whether or\r\n> not your Postgres numbers came from a standard build.\r\n\r\nI downloaded the 8kb varant.\r\n\r\nvitesse (median of 3):\r\npostgres=# select count(*), sum(i*i), avg(i) from t;\r\n count │ sum │ avg\r\n─────────┼────────────────────┼─────────────────────\r\n 1000000 │ 333333833333500000 │ 500000.500000000000\r\n(1 row)\r\n\r\nTime: 39.197 ms\r\n\r\nstock (median of 3):\r\npostgres=# select count(*), sum(i*i), avg(i) from t;\r\n count │ sum │ avg\r\n─────────┼────────────────────┼─────────────────────\r\n 1000000 │ 333333833333500000 │ 500000.500000000000\r\n(1 row)\r\n\r\nTime: 667.362 ms\r\n\r\n(stock int4 ops)\r\npostgres=# select sum(1::int4) from t;\r\n sum\r\n─────────\r\n 1000000\r\n(1 row)\r\n\r\nTime: 75.265 ms\r\n\r\nWhat I'm wondering is how complex the hooks are that tie the\r\ntechnology in. Unless a bsd licensed patch materializes, the\r\nconversation (beyond the intitial wow! factor) should probably focus\r\non a possible integration points and/or implementation of technology\r\ninto core in a general way.\r\n\r\nmerlin\r\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 13:29:09 -0500","msg_from":"Merlin Moncure <mmoncure@gmail.com>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"Andres Freund <andres@2ndquadrant.com> writes:\n> On 2014-10-17 13:12:27 -0400, Tom Lane wrote:\n>> Well, that's pretty much cheating: it's too hard to disentangle what's\n>> coming from JIT vs what's coming from using a different accumulator\n>> datatype. If we wanted to depend on having int128 available we could\n>> get that speedup with a couple hours' work.\n\n> I think doing that when configure detects int128 would make a great deal\n> of sense.\n\nYeah, I was wondering about that myself: use int128 if available,\nelse fall back on existing code path.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 14:35:55 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"Happy to contribute to that decision :-)\n\n\nOn Fri, Oct 17, 2014 at 11:35 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Andres Freund <andres@2ndquadrant.com> writes:\n>> On 2014-10-17 13:12:27 -0400, Tom Lane wrote:\n>>> Well, that's pretty much cheating: it's too hard to disentangle what's\n>>> coming from JIT vs what's coming from using a different accumulator\n>>> datatype. If we wanted to depend on having int128 available we could\n>>> get that speedup with a couple hours' work.\n>\n>> I think doing that when configure detects int128 would make a great deal\n>> of sense.\n>\n> Yeah, I was wondering about that myself: use int128 if available,\n> else fall back on existing code path.\n>\n> regards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 11:52:32 -0700","msg_from":"CK Tan <cktan@vitessedata.com>","msg_from_op":true,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"On Fri, 17 Oct 2014 13:12:27 -0400\nTom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> CK Tan <cktan@vitessedata.com> writes:\n> > The bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\n> \n> Well, that's pretty much cheating: it's too hard to disentangle what's\n> coming from JIT vs what's coming from using a different accumulator\n> datatype. If we wanted to depend on having int128 available we could\n> get that speedup with a couple hours' work.\n> \n> But what exactly are you \"compiling\" here? I trust not the actual data\n> accesses; that seems far too complicated to try to inline.\n> \n> \t\t\tregards, tom lane\n> \n> \n\nI don't have any inside knowledge, but from the presentation given at the\nrecent SFPUG followed by a bit of google-fu I think these papers are\nrelevant:\n\n http://www.vldb.org/pvldb/vol4/p539-neumann.pdf\n http://sites.computer.org/debull/A14mar/p3.pdf\n\n-dg\n\n-- \nDavid Gould 510 282 0869 daveg@sonic.net\nIf simplicity worked, the world would be overrun with insects.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 20:40:14 -0700","msg_from":"David Gould <daveg@sonic.net>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"Indeed! A big part of our implementation is based on the Neumann\npaper. There are also a few other papers that impacted our\nimplemented:\n\nA. Ailamaki, D. DeWitt, M. Hill, D. Wood. DBMSs On A Modern Processor:\nWhere Does Time Go?\n\nPeter Boncz, Marcin Zukowski, Niels Nes. MonetDB/X100:\nHyper-Pipelining Query Execution\n\nM. Zukowski el al. Super-Scalar RAM-CPU Cache Compression\n\nOf course, we need to adapt a lot of the design to Postgres to make\nsomething that could stand up harmoniously with the Postgres system,\nand also to take care that we would be able to merge easily with\nfuture versions of Postgres -- the implementation needs to be as\nnon-invasive as possible.\n\nRegards,\n-cktan\n\nOn Fri, Oct 17, 2014 at 8:40 PM, David Gould <daveg@sonic.net> wrote:\n> On Fri, 17 Oct 2014 13:12:27 -0400\n> Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n>> CK Tan <cktan@vitessedata.com> writes:\n>> > The bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\n>>\n>> Well, that's pretty much cheating: it's too hard to disentangle what's\n>> coming from JIT vs what's coming from using a different accumulator\n>> datatype. If we wanted to depend on having int128 available we could\n>> get that speedup with a couple hours' work.\n>>\n>> But what exactly are you \"compiling\" here? I trust not the actual data\n>> accesses; that seems far too complicated to try to inline.\n>>\n>> regards, tom lane\n>>\n>>\n>\n> I don't have any inside knowledge, but from the presentation given at the\n> recent SFPUG followed by a bit of google-fu I think these papers are\n> relevant:\n>\n> http://www.vldb.org/pvldb/vol4/p539-neumann.pdf\n> http://sites.computer.org/debull/A14mar/p3.pdf\n>\n> -dg\n>\n> --\n> David Gould 510 282 0869 daveg@sonic.net\n> If simplicity worked, the world would be overrun with insects.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 21:46:10 -0700","msg_from":"CK Tan <cktan@vitessedata.com>","msg_from_op":true,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"On 18/10/14 07:13, Josh Berkus wrote:\n> CK,\n>\n> Before we go any further on this, how is Vitesse currently licensed?\n> last time we talked it was still proprietary. If it's not being\n> open-sourced, we likely need to take discussion off this list.\n>\n\n+1\n\nGuys, you need to 'fess up on the licensing!\n\nRegards\n\nMark\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 18:55:06 +1300","msg_from":"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>","msg_from_op":false,"msg_subject":"Re: Vitesse DB call for testing"},{"msg_contents":"Hi Mark,\n\nVitesse DB won't be open-sourced, or it would have been a contrib\nmodule to postgres. We should take further discussions off this list.\nPeople should contact me directly if there is any questions.\n\nThanks,\ncktan@vitessedata.com\n\n\n\nOn Fri, Oct 17, 2014 at 10:55 PM, Mark Kirkwood\n<mark.kirkwood@catalyst.net.nz> wrote:\n> On 18/10/14 07:13, Josh Berkus wrote:\n>>\n>> CK,\n>>\n>> Before we go any further on this, how is Vitesse currently licensed?\n>> last time we talked it was still proprietary. If it's not being\n>> open-sourced, we likely need to take discussion off this list.\n>>\n>\n> +1\n>\n> Guys, you need to 'fess up on the licensing!\n>\n> Regards\n>\n> Mark\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 00:11:51 -0700","msg_from":"CK Tan <cktan@vitessedata.com>","msg_from_op":true,"msg_subject":"Re: Vitesse DB call for testing"}],"string":"[\n {\n \"msg_contents\": \"Hi everyone,\\n\\nVitesse DB 9.3.5.S is Postgres 9.3.5 with a LLVM-JIT query executor\\ndesigned for compute intensive OLAP workload. We have gotten it to a\\nreasonable state and would like to open it up to the pg hackers\\ncommunity for testing and suggestions.\\n\\nVitesse DB offers\\n-- JIT Compilation for compute-intensive queries\\n-- CSV parsing with SSE instructions\\n-- 100% binary compatibility with PG9.3.5.\\n\\nOur results show CSV imports run up to 2X faster, and TPCH Q1 runs 8X faster.\\n\\nOur TPCH 1GB benchmark results is also available at\\nhttp://vitessedata.com/benchmark/ .\\n\\nPlease direct any questions by email to cktan@vitessedata.com .\\n\\nThank you for your help.\\n\\n--\\nCK Tan\\nVitesse Data, Inc.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 05:32:13 -0700\",\n \"msg_from\": \"CK Tan <cktan@vitessedata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\n\\nOn 2014-10-17 05:32:13 -0700, CK Tan wrote:\\n> Vitesse DB 9.3.5.S is Postgres 9.3.5 with a LLVM-JIT query executor\\n> designed for compute intensive OLAP workload. We have gotten it to a\\n> reasonable state and would like to open it up to the pg hackers\\n> community for testing and suggestions.\\n> \\n> Vitesse DB offers\\n> -- JIT Compilation for compute-intensive queries\\n> -- CSV parsing with SSE instructions\\n> -- 100% binary compatibility with PG9.3.5.\\n> \\n> Our results show CSV imports run up to 2X faster, and TPCH Q1 runs 8X faster.\\n\\nHow are these modifications licensed?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 14:37:54 +0200\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 7:32 AM, CK Tan <cktan@vitessedata.com> wrote:\\n> Hi everyone,\\n>\\n> Vitesse DB 9.3.5.S is Postgres 9.3.5 with a LLVM-JIT query executor\\n> designed for compute intensive OLAP workload. We have gotten it to a\\n> reasonable state and would like to open it up to the pg hackers\\n> community for testing and suggestions.\\n>\\n> Vitesse DB offers\\n> -- JIT Compilation for compute-intensive queries\\n> -- CSV parsing with SSE instructions\\n> -- 100% binary compatibility with PG9.3.5.\\n>\\n> Our results show CSV imports run up to 2X faster, and TPCH Q1 runs 8X faster.\\n>\\n> Our TPCH 1GB benchmark results is also available at\\n> http://vitessedata.com/benchmark/ .\\n>\\n> Please direct any questions by email to cktan@vitessedata.com .\\n\\nYou offer a binary with 32k block size...what's the rationale for that?\\n\\nmerlin\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 08:14:14 -0500\",\n \"msg_from\": \"Merlin Moncure <mmoncure@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 8:14 AM, Merlin Moncure <mmoncure@gmail.com> wrote:\\n> On Fri, Oct 17, 2014 at 7:32 AM, CK Tan <cktan@vitessedata.com> wrote:\\n>> Hi everyone,\\n>>\\n>> Vitesse DB 9.3.5.S is Postgres 9.3.5 with a LLVM-JIT query executor\\n>> designed for compute intensive OLAP workload. We have gotten it to a\\n>> reasonable state and would like to open it up to the pg hackers\\n>> community for testing and suggestions.\\n>>\\n>> Vitesse DB offers\\n>> -- JIT Compilation for compute-intensive queries\\n>> -- CSV parsing with SSE instructions\\n>> -- 100% binary compatibility with PG9.3.5.\\n>>\\n>> Our results show CSV imports run up to 2X faster, and TPCH Q1 runs 8X faster.\\n>>\\n>> Our TPCH 1GB benchmark results is also available at\\n>> http://vitessedata.com/benchmark/ .\\n>>\\n>> Please direct any questions by email to cktan@vitessedata.com .\\n>\\n> You offer a binary with 32k block size...what's the rationale for that?\\n\\n(sorry for the double post)\\n\\nOK, I downloaded the ubuntu binary and ran your benchmarks (after\\nmaking some minor .conf tweaks like disabling SSL). I then ran your\\nbenchmark (after fixing the typo) of the count/sum/avg test -- *and\\nnoticed a 95% reduction in runtime performance* which is really quite\\namazing IMNSHO. I also ran a select only test on small scale factor\\npgbench and didn't see any regression there -- in fact you beat stock\\nby ~ 3% (although this could be measurement noise). So now you've\\ngot my attention. So, if you don't mind, quit being coy and explain\\nhow the software works and all the neat things it does and doesn't do.\\n\\nmerlin\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 08:43:30 -0500\",\n \"msg_from\": \"Merlin Moncure <mmoncure@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"Merlin, glad you tried it. \\n\\nWe take the query plan exactly as given by the planner, decide whether to JIT or to punt depending on the cost. If we punt, it goes back to pg executor. If we JIT, and if we could not proceed (usually of some operators we haven't implemented yet), we again punt. Once we were able to generate the code, there is no going back; we call into LLVM to obtain the function entry point, and run it to completion. The 3% improvement you see in OLTP tests is definitely noise.\\n\\nThe bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\\n\\nThe speed up is really noticeable when the data type is nonvarlena. In the varlena cases, we still call into pg routines most of the times. Again, try the sum,avg,count query on numeric, and you will see what I mean.\\n\\nAlso, we don't support UDF at the moment. So all queries involving UDF gets sent to pg executor.\\n\\nOn your question of 32k page size, the rational is that some of our customers could be interested in a data warehouse on pg. 32k page size is a big win when all you do is seqscan all day long.\\n\\nWe are looking for bug reports at these stage and some stress tests done without our own prejudices. Some test on real data in non prod setting on queries that are highly CPU bound would be ideal.\\n\\nThanks,\\n-cktan\\n\\n> On Oct 17, 2014, at 6:43 AM, Merlin Moncure <mmoncure@gmail.com> wrote:\\n> \\n>> On Fri, Oct 17, 2014 at 8:14 AM, Merlin Moncure <mmoncure@gmail.com> wrote:\\n>>> On Fri, Oct 17, 2014 at 7:32 AM, CK Tan <cktan@vitessedata.com> wrote:\\n>>> Hi everyone,\\n>>> \\n>>> Vitesse DB 9.3.5.S is Postgres 9.3.5 with a LLVM-JIT query executor\\n>>> designed for compute intensive OLAP workload. We have gotten it to a\\n>>> reasonable state and would like to open it up to the pg hackers\\n>>> community for testing and suggestions.\\n>>> \\n>>> Vitesse DB offers\\n>>> -- JIT Compilation for compute-intensive queries\\n>>> -- CSV parsing with SSE instructions\\n>>> -- 100% binary compatibility with PG9.3.5.\\n>>> \\n>>> Our results show CSV imports run up to 2X faster, and TPCH Q1 runs 8X faster.\\n>>> \\n>>> Our TPCH 1GB benchmark results is also available at\\n>>> http://vitessedata.com/benchmark/ .\\n>>> \\n>>> Please direct any questions by email to cktan@vitessedata.com .\\n>> \\n>> You offer a binary with 32k block size...what's the rationale for that?\\n> \\n> (sorry for the double post)\\n> \\n> OK, I downloaded the ubuntu binary and ran your benchmarks (after\\n> making some minor .conf tweaks like disabling SSL). I then ran your\\n> benchmark (after fixing the typo) of the count/sum/avg test -- *and\\n> noticed a 95% reduction in runtime performance* which is really quite\\n> amazing IMNSHO. I also ran a select only test on small scale factor\\n> pgbench and didn't see any regression there -- in fact you beat stock\\n> by ~ 3% (although this could be measurement noise). So now you've\\n> got my attention. So, if you don't mind, quit being coy and explain\\n> how the software works and all the neat things it does and doesn't do.\\n> \\n> merlin\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 08:47:14 -0700\",\n \"msg_from\": \"CK Tan <cktan@vitessedata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 10:47 AM, CK Tan <cktan@vitessedata.com> wrote:\\n> Merlin, glad you tried it.\\n>\\n> We take the query plan exactly as given by the planner, decide whether to JIT or to punt depending on the cost. If we punt, it goes back to pg executor. If we JIT, and if we could not proceed (usually of some operators we haven't implemented yet), we again punt. Once we were able to generate the code, there is no going back; we call into LLVM to obtain the function entry point, and run it to completion. The 3% improvement you see in OLTP tests is definitely noise.\\n>\\n> The bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\\n>\\n> The speed up is really noticeable when the data type is nonvarlena. In the varlena cases, we still call into pg routines most of the times. Again, try the sum,avg,count query on numeric, and you will see what I mean.\\n>\\n> Also, we don't support UDF at the moment. So all queries involving UDF gets sent to pg executor.\\n>\\n> On your question of 32k page size, the rational is that some of our customers could be interested in a data warehouse on pg. 32k page size is a big win when all you do is seqscan all day long.\\n>\\n> We are looking for bug reports at these stage and some stress tests done without our own prejudices. Some test on real data in non prod setting on queries that are highly CPU bound would be ideal.\\n\\nOne thing that I noticed is that when slamming your benchmark query\\nvia pgbench, resident memory consumption was really aggressive and\\nwould have taken down the server had I not spuriously stopped the\\ntest. Memory consumption did return to baseline after that so I\\nfigured some type of llvm memory management games were going on. This\\nisn't really a problem for most OLAP workloads but it's something to\\nbe aware of.\\n\\nVia 'perf top' on stock postgres, you see the usual suspects: palloc,\\nhash_search_etc. On your build though HeapTuplesSatisfiesMVCC zooms\\nright to the top of the stack which is pretty interesting...the\\nexecutor is you've built is very lean and mean for sure. A drop in\\noptimization engine with little no schema/sql changes is pretty neat\\n-- your primary competitor here is going to be column organized type\\ntable solutions to olap type problems.\\n\\nmerlin\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 11:00:41 -0500\",\n \"msg_from\": \"Merlin Moncure <mmoncure@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"CK Tan <cktan@vitessedata.com> writes:\\n> The bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\\n\\nWell, that's pretty much cheating: it's too hard to disentangle what's\\ncoming from JIT vs what's coming from using a different accumulator\\ndatatype. If we wanted to depend on having int128 available we could\\nget that speedup with a couple hours' work.\\n\\nBut what exactly are you \\\"compiling\\\" here? I trust not the actual data\\naccesses; that seems far too complicated to try to inline.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 13:12:27 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"Hi, Tom,\\n\\nSorry for double post to you.\\n\\nFeng\\n\\n---------- Forwarded message ----------\\nFrom: Feng Tian <ftian@vitessedata.com>\\nDate: Fri, Oct 17, 2014 at 10:29 AM\\nSubject: Re: [HACKERS] Vitesse DB call for testing\\nTo: Tom Lane <tgl@sss.pgh.pa.us>\\n\\n\\nHi, Tom,\\n\\nI agree using that using int128 in stock postgres will speed up things\\ntoo. On the other hand, that is only one part of the equation. For\\nexample, if you look at TPCH Q1, the int128 \\\"cheating\\\" does not kick in at\\nall, but we are 8x faster.\\n\\nI am not sure why do you mean by \\\"actual data access\\\". Data is still in\\nstock postgres format on disk. We indeed jit-ed all data fields access\\n(deform tuple). To put things in perspective, I just timed select\\ncount(*) and select count(l_orderkey) from tpch1.lineitem; Our code is\\nbottlenecked by memory bandwidth and difference is pretty much invisible.\\n\\nThanks,\\nFeng\\n\\n\\nftian=# set vdb_jit = 0;\\n\\nSET\\n\\nTime: 0.155 ms\\n\\nftian=# select count(*) from tpch1.lineitem;\\n\\n count\\n\\n---------\\n\\n 6001215\\n\\n(1 row)\\n\\n\\nTime: 688.658 ms\\n\\nftian=# select count(*) from tpch1.lineitem;\\n\\n count\\n\\n---------\\n\\n 6001215\\n\\n(1 row)\\n\\n\\nTime: 690.753 ms\\n\\nftian=# select count(l_orderkey) from tpch1.lineitem;\\n\\n count\\n\\n---------\\n\\n 6001215\\n\\n(1 row)\\n\\n\\nTime: 819.452 ms\\n\\nftian=# set vdb_jit = 1;\\n\\nSET\\n\\nTime: 0.167 ms\\n\\nftian=# select count(*) from tpch1.lineitem;\\n\\n count\\n\\n---------\\n\\n 6001215\\n\\n(1 row)\\n\\n\\nTime: 203.543 ms\\n\\nftian=# select count(l_orderkey) from tpch1.lineitem;\\n\\n count\\n\\n---------\\n\\n 6001215\\n\\n(1 row)\\n\\n\\nTime: 202.253 ms\\n\\nftian=#\\n\\n\\n\\n\\nOn Fri, Oct 17, 2014 at 10:12 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> CK Tan <cktan@vitessedata.com> writes:\\n> > The bigint sum,avg,count case in the example you tried has some\\n> optimization. We use int128 to accumulate the bigint instead of numeric in\\n> pg. Hence the big speed up. Try the same query on int4 for the improvement\\n> where both pg and vitessedb are using int4 in the execution.\\n>\\n> Well, that's pretty much cheating: it's too hard to disentangle what's\\n> coming from JIT vs what's coming from using a different accumulator\\n> datatype. If we wanted to depend on having int128 available we could\\n> get that speedup with a couple hours' work.\\n>\\n> But what exactly are you \\\"compiling\\\" here? I trust not the actual data\\n> accesses; that seems far too complicated to try to inline.\\n>\\n> regards, tom lane\\n>\\n>\\n> --\\n> Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\n> To make changes to your subscription:\\n> http://www.postgresql.org/mailpref/pgsql-hackers\\n>\\n\\nHi, Tom,Sorry for double post to you.Feng---------- Forwarded message ----------From: Feng Tian <ftian@vitessedata.com>Date: Fri, Oct 17, 2014 at 10:29 AMSubject: Re: [HACKERS] Vitesse DB call for testingTo: Tom Lane <tgl@sss.pgh.pa.us>Hi, Tom,I agree using that using int128 in stock postgres will speed up things too. On the other hand, that is only one part of the equation. For example, if you look at TPCH Q1, the int128 \\\"cheating\\\" does not kick in at all, but we are 8x faster.I am not sure why do you mean by \\\"actual data access\\\". Data is still in stock postgres format on disk. We indeed jit-ed all data fields access (deform tuple). To put things in perspective, I just timed select count(*) and select count(l_orderkey) from tpch1.lineitem; Our code is bottlenecked by memory bandwidth and difference is pretty much invisible.Thanks,Feng\\nftian=# set vdb_jit = 0;\\nSET\\nTime: 0.155 ms\\nftian=# select count(*) from tpch1.lineitem;\\n count \\n---------\\n 6001215\\n(1 row)\\n\\nTime: 688.658 ms\\nftian=# select count(*) from tpch1.lineitem;\\n count \\n---------\\n 6001215\\n(1 row)\\n\\nTime: 690.753 ms\\nftian=# select count(l_orderkey) from tpch1.lineitem;\\n count \\n---------\\n 6001215\\n(1 row)\\n\\nTime: 819.452 ms\\nftian=# set vdb_jit = 1;\\nSET\\nTime: 0.167 ms\\nftian=# select count(*) from tpch1.lineitem;\\n count \\n---------\\n 6001215\\n(1 row)\\n\\nTime: 203.543 ms\\nftian=# select count(l_orderkey) from tpch1.lineitem;\\n count \\n---------\\n 6001215\\n(1 row)\\n\\nTime: 202.253 ms\\nftian=# On Fri, Oct 17, 2014 at 10:12 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:CK Tan <cktan@vitessedata.com> writes:\\n> The bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\\n\\nWell, that's pretty much cheating: it's too hard to disentangle what's\\ncoming from JIT vs what's coming from using a different accumulator\\ndatatype. If we wanted to depend on having int128 available we could\\nget that speedup with a couple hours' work.\\n\\nBut what exactly are you \\\"compiling\\\" here? I trust not the actual data\\naccesses; that seems far too complicated to try to inline.\\n\\n regards, tom lane\\n\\n\\n--\\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Fri, 17 Oct 2014 11:08:48 -0700\",\n \"msg_from\": \"Feng Tian <ftian@vitessedata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Fwd: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"CK,\\n\\nBefore we go any further on this, how is Vitesse currently licensed?\\nlast time we talked it was still proprietary. If it's not being\\nopen-sourced, we likely need to take discussion off this list.\\n\\n-- \\nJosh Berkus\\nPostgreSQL Experts Inc.\\nhttp://pgexperts.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 11:13:00 -0700\",\n \"msg_from\": \"Josh Berkus <josh@agliodbs.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 11:08 AM, Feng Tian <ftian@vitessedata.com> wrote:\\n> I agree using that using int128 in stock postgres will speed up things too.\\n> On the other hand, that is only one part of the equation. For example, if\\n> you look at TPCH Q1, the int128 \\\"cheating\\\" does not kick in at all, but we\\n> are 8x faster.\\n\\nI'm curious about how the numbers look when stock Postgres is built\\nwith the same page size as your fork. You didn't mention whether or\\nnot your Postgres numbers came from a standard build.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 11:21:55 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@heroku.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"On 2014-10-17 13:12:27 -0400, Tom Lane wrote:\\n> Well, that's pretty much cheating: it's too hard to disentangle what's\\n> coming from JIT vs what's coming from using a different accumulator\\n> datatype. If we wanted to depend on having int128 available we could\\n> get that speedup with a couple hours' work.\\n\\nI think doing that when configure detects int128 would make a great deal\\nof sense. It's not like we'd save a great deal of complicated code by\\nremoving the existing accumulator... We'd still have to return a\\nnumeric, but that's likely lost in the noise cost wise.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 20:25:00 +0200\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 1:21 PM, Peter Geoghegan <pg@heroku.com> wrote:\\r\\n> On Fri, Oct 17, 2014 at 11:08 AM, Feng Tian <ftian@vitessedata.com> wrote:\\r\\n>> I agree using that using int128 in stock postgres will speed up things too.\\r\\n>> On the other hand, that is only one part of the equation. For example, if\\r\\n>> you look at TPCH Q1, the int128 \\\"cheating\\\" does not kick in at all, but we\\r\\n>> are 8x faster.\\r\\n>\\r\\n> I'm curious about how the numbers look when stock Postgres is built\\r\\n> with the same page size as your fork. You didn't mention whether or\\r\\n> not your Postgres numbers came from a standard build.\\r\\n\\r\\nI downloaded the 8kb varant.\\r\\n\\r\\nvitesse (median of 3):\\r\\npostgres=# select count(*), sum(i*i), avg(i) from t;\\r\\n count │ sum │ avg\\r\\n─────────┼────────────────────┼─────────────────────\\r\\n 1000000 │ 333333833333500000 │ 500000.500000000000\\r\\n(1 row)\\r\\n\\r\\nTime: 39.197 ms\\r\\n\\r\\nstock (median of 3):\\r\\npostgres=# select count(*), sum(i*i), avg(i) from t;\\r\\n count │ sum │ avg\\r\\n─────────┼────────────────────┼─────────────────────\\r\\n 1000000 │ 333333833333500000 │ 500000.500000000000\\r\\n(1 row)\\r\\n\\r\\nTime: 667.362 ms\\r\\n\\r\\n(stock int4 ops)\\r\\npostgres=# select sum(1::int4) from t;\\r\\n sum\\r\\n─────────\\r\\n 1000000\\r\\n(1 row)\\r\\n\\r\\nTime: 75.265 ms\\r\\n\\r\\nWhat I'm wondering is how complex the hooks are that tie the\\r\\ntechnology in. Unless a bsd licensed patch materializes, the\\r\\nconversation (beyond the intitial wow! factor) should probably focus\\r\\non a possible integration points and/or implementation of technology\\r\\ninto core in a general way.\\r\\n\\r\\nmerlin\\r\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 13:29:09 -0500\",\n \"msg_from\": \"Merlin Moncure <mmoncure@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@2ndquadrant.com> writes:\\n> On 2014-10-17 13:12:27 -0400, Tom Lane wrote:\\n>> Well, that's pretty much cheating: it's too hard to disentangle what's\\n>> coming from JIT vs what's coming from using a different accumulator\\n>> datatype. If we wanted to depend on having int128 available we could\\n>> get that speedup with a couple hours' work.\\n\\n> I think doing that when configure detects int128 would make a great deal\\n> of sense.\\n\\nYeah, I was wondering about that myself: use int128 if available,\\nelse fall back on existing code path.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 14:35:55 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"Happy to contribute to that decision :-)\\n\\n\\nOn Fri, Oct 17, 2014 at 11:35 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Andres Freund <andres@2ndquadrant.com> writes:\\n>> On 2014-10-17 13:12:27 -0400, Tom Lane wrote:\\n>>> Well, that's pretty much cheating: it's too hard to disentangle what's\\n>>> coming from JIT vs what's coming from using a different accumulator\\n>>> datatype. If we wanted to depend on having int128 available we could\\n>>> get that speedup with a couple hours' work.\\n>\\n>> I think doing that when configure detects int128 would make a great deal\\n>> of sense.\\n>\\n> Yeah, I was wondering about that myself: use int128 if available,\\n> else fall back on existing code path.\\n>\\n> regards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 11:52:32 -0700\",\n \"msg_from\": \"CK Tan <cktan@vitessedata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"On Fri, 17 Oct 2014 13:12:27 -0400\\nTom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> CK Tan <cktan@vitessedata.com> writes:\\n> > The bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\\n> \\n> Well, that's pretty much cheating: it's too hard to disentangle what's\\n> coming from JIT vs what's coming from using a different accumulator\\n> datatype. If we wanted to depend on having int128 available we could\\n> get that speedup with a couple hours' work.\\n> \\n> But what exactly are you \\\"compiling\\\" here? I trust not the actual data\\n> accesses; that seems far too complicated to try to inline.\\n> \\n> \\t\\t\\tregards, tom lane\\n> \\n> \\n\\nI don't have any inside knowledge, but from the presentation given at the\\nrecent SFPUG followed by a bit of google-fu I think these papers are\\nrelevant:\\n\\n http://www.vldb.org/pvldb/vol4/p539-neumann.pdf\\n http://sites.computer.org/debull/A14mar/p3.pdf\\n\\n-dg\\n\\n-- \\nDavid Gould 510 282 0869 daveg@sonic.net\\nIf simplicity worked, the world would be overrun with insects.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 20:40:14 -0700\",\n \"msg_from\": \"David Gould <daveg@sonic.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"Indeed! A big part of our implementation is based on the Neumann\\npaper. There are also a few other papers that impacted our\\nimplemented:\\n\\nA. Ailamaki, D. DeWitt, M. Hill, D. Wood. DBMSs On A Modern Processor:\\nWhere Does Time Go?\\n\\nPeter Boncz, Marcin Zukowski, Niels Nes. MonetDB/X100:\\nHyper-Pipelining Query Execution\\n\\nM. Zukowski el al. Super-Scalar RAM-CPU Cache Compression\\n\\nOf course, we need to adapt a lot of the design to Postgres to make\\nsomething that could stand up harmoniously with the Postgres system,\\nand also to take care that we would be able to merge easily with\\nfuture versions of Postgres -- the implementation needs to be as\\nnon-invasive as possible.\\n\\nRegards,\\n-cktan\\n\\nOn Fri, Oct 17, 2014 at 8:40 PM, David Gould <daveg@sonic.net> wrote:\\n> On Fri, 17 Oct 2014 13:12:27 -0400\\n> Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n>> CK Tan <cktan@vitessedata.com> writes:\\n>> > The bigint sum,avg,count case in the example you tried has some optimization. We use int128 to accumulate the bigint instead of numeric in pg. Hence the big speed up. Try the same query on int4 for the improvement where both pg and vitessedb are using int4 in the execution.\\n>>\\n>> Well, that's pretty much cheating: it's too hard to disentangle what's\\n>> coming from JIT vs what's coming from using a different accumulator\\n>> datatype. If we wanted to depend on having int128 available we could\\n>> get that speedup with a couple hours' work.\\n>>\\n>> But what exactly are you \\\"compiling\\\" here? I trust not the actual data\\n>> accesses; that seems far too complicated to try to inline.\\n>>\\n>> regards, tom lane\\n>>\\n>>\\n>\\n> I don't have any inside knowledge, but from the presentation given at the\\n> recent SFPUG followed by a bit of google-fu I think these papers are\\n> relevant:\\n>\\n> http://www.vldb.org/pvldb/vol4/p539-neumann.pdf\\n> http://sites.computer.org/debull/A14mar/p3.pdf\\n>\\n> -dg\\n>\\n> --\\n> David Gould 510 282 0869 daveg@sonic.net\\n> If simplicity worked, the world would be overrun with insects.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 21:46:10 -0700\",\n \"msg_from\": \"CK Tan <cktan@vitessedata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"On 18/10/14 07:13, Josh Berkus wrote:\\n> CK,\\n>\\n> Before we go any further on this, how is Vitesse currently licensed?\\n> last time we talked it was still proprietary. If it's not being\\n> open-sourced, we likely need to take discussion off this list.\\n>\\n\\n+1\\n\\nGuys, you need to 'fess up on the licensing!\\n\\nRegards\\n\\nMark\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 18:55:06 +1300\",\n \"msg_from\": \"Mark Kirkwood <mark.kirkwood@catalyst.net.nz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n },\n {\n \"msg_contents\": \"Hi Mark,\\n\\nVitesse DB won't be open-sourced, or it would have been a contrib\\nmodule to postgres. We should take further discussions off this list.\\nPeople should contact me directly if there is any questions.\\n\\nThanks,\\ncktan@vitessedata.com\\n\\n\\n\\nOn Fri, Oct 17, 2014 at 10:55 PM, Mark Kirkwood\\n<mark.kirkwood@catalyst.net.nz> wrote:\\n> On 18/10/14 07:13, Josh Berkus wrote:\\n>>\\n>> CK,\\n>>\\n>> Before we go any further on this, how is Vitesse currently licensed?\\n>> last time we talked it was still proprietary. If it's not being\\n>> open-sourced, we likely need to take discussion off this list.\\n>>\\n>\\n> +1\\n>\\n> Guys, you need to 'fess up on the licensing!\\n>\\n> Regards\\n>\\n> Mark\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 00:11:51 -0700\",\n \"msg_from\": \"CK Tan <cktan@vitessedata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Vitesse DB call for testing\"\n }\n]","isConversational":false}}},{"rowIdx":39677,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nConsider the following queries.\n\ncreate table t(i int, j int, k int, t text);\ninsert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from\ngenerate_series(1, 1000000) i;\n\nftian=# explain select count(distinct j) from t group by t, i;\n QUERY PLAN\n------------------------------------------------------------------------\n GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22)\n -> Sort (cost=158029.84..160529.84 rows=1000000 width=22)\n Sort Key: t, i\n -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)\n(4 rows)\n\n\nThe query,\nselect count(distinct j) from t group by t, i;\n\nruns for 35 seconds. However, if I change the query to,\nselect count(distinct j) from t group by i, t; -- note switching the\nordering\nselect count(distinct j) from t group by decode(t, 'escape'), i; -- convert\nt to bytea\n\nRun times are just about 5 and 6.5 seconds. The reason is clear, compare a\nstring with collation is slow, which is well understood by pg hackers.\nHowever, here, the sorting order is forced by the planner, not user.\nPlanner can do the following optimizations,\n\n1. for the sort we generated for sort agg, planner can switch column\nordering, put int before string,\n2. for the sort we generated for sort agg, use bytea compare instead of\nstring compare.\n\nThey will bring big improvement to this common query. Is this something\nreasonable?\n\nThanks,\n\nHi, Consider the following queries.create table t(i int, j int, k int, t text);insert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from generate_series(1, 1000000) i;ftian=# explain select count(distinct j) from t group by t, i; QUERY PLAN ------------------------------------------------------------------------ GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22) -> Sort (cost=158029.84..160529.84 rows=1000000 width=22) Sort Key: t, i -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)(4 rows)The query, select count(distinct j) from t group by t, i; runs for 35 seconds. However, if I change the query to,select count(distinct j) from t group by i, t; -- note switching the orderingselect count(distinct j) from t group by decode(t, 'escape'), i; -- convert t to byteaRun times are just about 5 and 6.5 seconds. The reason is clear, compare a string with collation is slow, which is well understood by pg hackers. However, here, the sorting order is forced by the planner, not user. Planner can do the following optimizations,1. for the sort we generated for sort agg, planner can switch column ordering, put int before string,2. for the sort we generated for sort agg, use bytea compare instead of string compare.They will bring big improvement to this common query. Is this something reasonable? Thanks,","msg_date":"Fri, 17 Oct 2014 09:10:13 -0700","msg_from":"Feng Tian <fengttt@gmail.com>","msg_from_op":true,"msg_subject":"Optimizer on sort aggregate"},{"msg_contents":"> The query,\n> select count(distinct j) from t group by t, i;\n> \n> runs for 35 seconds. However, if I change the query to,\n> select count(distinct j) from t group by i, t; -- note switching the\n> ordering\n> select count(distinct j) from t group by decode(t, 'escape'), i; -- convert\n> t to bytea\n> \n> Run times are just about 5 and 6.5 seconds. The reason is clear, compare a\n> string with collation is slow, which is well understood by pg hackers.\n> However, here, the sorting order is forced by the planner, not user.\n> Planner can do the following optimizations,\n\nInteresting. I got following result:\n\ntest=# explain analyze select count(distinct j) from t group by t, i;\n QUERY PLAN \n--------------------------------------------------------------------------------------------------------------------------\n GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual time=1332.937..2431.238 rows=1000000 loops=1)\n Group Key: t, i\n -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual time=1332.922..1507.413 rows=1000000 loops=1)\n Sort Key: t, i\n Sort Method: external merge Disk: 33232kB\n -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22) (actual time=0.006..131.406 rows=1000000 loops=1)\n Planning time: 0.031 ms\n Execution time: 2484.271 ms\n(8 rows)\n\nTime: 2484.520 ms\n\ntest=# explain analyze select count(distinct j) from t group by i, t;\n QUERY PLAN \n--------------------------------------------------------------------------------------------------------------------------\n GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual time=602.510..1632.087 rows=1000000 loops=1)\n Group Key: i, t\n -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual time=602.493..703.274 rows=1000000 loops=1)\n Sort Key: i, t\n Sort Method: external sort Disk: 33240kB\n -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22) (actual time=0.014..129.213 rows=1000000 loops=1)\n Planning time: 0.176 ms\n Execution time: 1685.575 ms\n(8 rows)\n\nTime: 1687.641 ms\n\nNot so big difference here (maybe because I use SSD) but there is\nstill about 50% difference in execution time. Note that I disable\nlocale support.\n\n> 1. for the sort we generated for sort agg, planner can switch column\n> ordering, put int before string,\n> 2. for the sort we generated for sort agg, use bytea compare instead of\n> string compare.\n> \n> They will bring big improvement to this common query. Is this something\n> reasonable?\n\nNot looking at sort and planner code closely, I guess planner does\nnot recognize the cost difference between \"external merge\" and\n\"external sort\" because cost estimations for sort step in each plan\nare exactly same (137519.84..140019.84). However I am not sure if we\nshould fix the planner or should fix our sort machinery since it is\npossible that the sort machinery should not expose such a big\ndifference in the two sort methods.\n--\nTatsuo Ishii\nSRA OSS, Inc. Japan\nEnglish: http://www.sraoss.co.jp/index_en.php\nJapanese:http://www.sraoss.co.jp\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 08:35:16 +0900 (JST)","msg_from":"Tatsuo Ishii <ishii@postgresql.org>","msg_from_op":false,"msg_subject":"Re: Optimizer on sort aggregate"},{"msg_contents":"On Sat, Oct 18, 2014 at 5:10 AM, Feng Tian <fengttt@gmail.com> wrote:\n\n> Hi,\n>\n> Consider the following queries.\n>\n> create table t(i int, j int, k int, t text);\n> insert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from\n> generate_series(1, 1000000) i;\n>\n> ftian=# explain select count(distinct j) from t group by t, i;\n> QUERY PLAN\n> ------------------------------------------------------------------------\n> GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22)\n> -> Sort (cost=158029.84..160529.84 rows=1000000 width=22)\n> Sort Key: t, i\n> -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)\n> (4 rows)\n>\n>\n> The query,\n> select count(distinct j) from t group by t, i;\n>\n> runs for 35 seconds. However, if I change the query to,\n> select count(distinct j) from t group by i, t; -- note switching the\n> ordering\n> select count(distinct j) from t group by decode(t, 'escape'), i; --\n> convert t to bytea\n>\n> Run times are just about 5 and 6.5 seconds. The reason is clear, compare\n> a string with collation is slow, which is well understood by pg hackers.\n> However, here, the sorting order is forced by the planner, not user.\n> Planner can do the following optimizations,\n>\n> 1. for the sort we generated for sort agg, planner can switch column\n> ordering, put int before string,\n> 2. for the sort we generated for sort agg, use bytea compare instead of\n> string compare.\n>\n> They will bring big improvement to this common query. Is this something\n> reasonable?\n>\n>\nIt seems like it might be worth looking into, but I think it's likely more\ncomplex than sorting the group by order into narrowest column first.\n\nIf the query was:\n\nselect count(distinct j) from t group by t, i order by t, i;\n\nThen if that was rewritten to make the group by i,t then the planner would\nthen need to perform an additional sort on t,i to get the correct order by\nfor the final result, which may or may not be faster, it would depend on\nhow many groups there were to sort. Though I guess you could possibly just\nskip the optimisation if the next node up didn't need any specific ordering.\n\nI also wonder if taking into account the column's width is not quite\nenough. For example if the 'i' column only had 1 distinct value, then\nperforming the group by on 'i' first wouldn't help at all. Keep in mind\nthat the columns could be much closer in width than in your example, e.g\nint and bigint. Perhaps you'd need to include some sort of heuristics to\nlook at the number of distinct values and the width, and form some sort of\nweighted estimates about which column to put first.\n\nRegards\n\nDavid Rowley\n\nOn Sat, Oct 18, 2014 at 5:10 AM, Feng Tian <fengttt@gmail.com> wrote:Hi, Consider the following queries.create table t(i int, j int, k int, t text);insert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from generate_series(1, 1000000) i;ftian=# explain select count(distinct j) from t group by t, i; QUERY PLAN ------------------------------------------------------------------------ GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22) -> Sort (cost=158029.84..160529.84 rows=1000000 width=22) Sort Key: t, i -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)(4 rows)The query, select count(distinct j) from t group by t, i; runs for 35 seconds. However, if I change the query to,select count(distinct j) from t group by i, t; -- note switching the orderingselect count(distinct j) from t group by decode(t, 'escape'), i; -- convert t to byteaRun times are just about 5 and 6.5 seconds. The reason is clear, compare a string with collation is slow, which is well understood by pg hackers. However, here, the sorting order is forced by the planner, not user. Planner can do the following optimizations,1. for the sort we generated for sort agg, planner can switch column ordering, put int before string,2. for the sort we generated for sort agg, use bytea compare instead of string compare.They will bring big improvement to this common query. Is this something reasonable? It seems like it might be worth looking into, but I think it's likely more complex than sorting the group by order into narrowest column first.If the query was:select count(distinct j) from t group by t, i order by t, i;Then if that was rewritten to make the group by i,t then the planner would then need to perform an additional sort on t,i to get the correct order by for the final result, which may or may not be faster, it would depend on how many groups there were to sort. Though I guess you could possibly just skip the optimisation if the next node up didn't need any specific ordering.I also wonder if taking into account the column's width is not quite enough. For example if the 'i' column only had 1 distinct value, then performing the group by on 'i' first wouldn't help at all. Keep in mind that the columns could be much closer in width than in your example, e.g int and bigint. Perhaps you'd need to include some sort of heuristics to look at the number of distinct values and the width, and form some sort of weighted estimates about which column to put first.RegardsDavid Rowley","msg_date":"Sat, 18 Oct 2014 12:36:30 +1300","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimizer on sort aggregate"},{"msg_contents":"On Sat, Oct 18, 2014 at 12:35 PM, Tatsuo Ishii <ishii@postgresql.org> wrote:\n\n> > The query,\n> > select count(distinct j) from t group by t, i;\n> >\n> > runs for 35 seconds. However, if I change the query to,\n> > select count(distinct j) from t group by i, t; -- note switching the\n> > ordering\n> > select count(distinct j) from t group by decode(t, 'escape'), i; --\n> convert\n> > t to bytea\n> >\n> > Run times are just about 5 and 6.5 seconds. The reason is clear,\n> compare a\n> > string with collation is slow, which is well understood by pg hackers.\n> > However, here, the sorting order is forced by the planner, not user.\n> > Planner can do the following optimizations,\n>\n> Interesting. I got following result:\n>\n> test=# explain analyze select count(distinct j) from t group by t, i;\n> QUERY PLAN\n>\n> --------------------------------------------------------------------------------------------------------------------------\n> GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual\n> time=1332.937..2431.238 rows=1000000 loops=1)\n> Group Key: t, i\n> -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual\n> time=1332.922..1507.413 rows=1000000 loops=1)\n> Sort Key: t, i\n> Sort Method: external merge Disk: 33232kB\n> -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)\n> (actual time=0.006..131.406 rows=1000000 loops=1)\n> Planning time: 0.031 ms\n> Execution time: 2484.271 ms\n> (8 rows)\n>\n> Time: 2484.520 ms\n>\n> test=# explain analyze select count(distinct j) from t group by i, t;\n> QUERY PLAN\n>\n> --------------------------------------------------------------------------------------------------------------------------\n> GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual\n> time=602.510..1632.087 rows=1000000 loops=1)\n> Group Key: i, t\n> -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual\n> time=602.493..703.274 rows=1000000 loops=1)\n> Sort Key: i, t\n> Sort Method: external sort Disk: 33240kB\n> -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)\n> (actual time=0.014..129.213 rows=1000000 loops=1)\n> Planning time: 0.176 ms\n> Execution time: 1685.575 ms\n> (8 rows)\n>\n> Time: 1687.641 ms\n>\n> Not so big difference here (maybe because I use SSD) but there is\n> still about 50% difference in execution time. Note that I disable\n> locale support.\n>\n>\nI think this is more likely your locale settings, as if I do:\n\n create table t(i int, j int, k int, t text collate \"C\");\nThe GROUP BY t,i runs about 25% faster.\n\nI've not looked at it yet, but Peter G's patch here\nhttps://commitfest.postgresql.org/action/patch_view?id=1462 will quite\nlikely narrow the performance gap between the 2 queries.\n\nRegards\n\nDavid Rowley\n\nOn Sat, Oct 18, 2014 at 12:35 PM, Tatsuo Ishii <ishii@postgresql.org> wrote:> The query,\n> select count(distinct j) from t group by t, i;\n>\n> runs for 35 seconds. However, if I change the query to,\n> select count(distinct j) from t group by i, t; -- note switching the\n> ordering\n> select count(distinct j) from t group by decode(t, 'escape'), i; -- convert\n> t to bytea\n>\n> Run times are just about 5 and 6.5 seconds. The reason is clear, compare a\n> string with collation is slow, which is well understood by pg hackers.\n> However, here, the sorting order is forced by the planner, not user.\n> Planner can do the following optimizations,\n\nInteresting. I got following result:\n\ntest=# explain analyze select count(distinct j) from t group by t, i;\n QUERY PLAN\n--------------------------------------------------------------------------------------------------------------------------\n GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual time=1332.937..2431.238 rows=1000000 loops=1)\n Group Key: t, i\n -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual time=1332.922..1507.413 rows=1000000 loops=1)\n Sort Key: t, i\n Sort Method: external merge Disk: 33232kB\n -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22) (actual time=0.006..131.406 rows=1000000 loops=1)\n Planning time: 0.031 ms\n Execution time: 2484.271 ms\n(8 rows)\n\nTime: 2484.520 ms\n\ntest=# explain analyze select count(distinct j) from t group by i, t;\n QUERY PLAN\n--------------------------------------------------------------------------------------------------------------------------\n GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual time=602.510..1632.087 rows=1000000 loops=1)\n Group Key: i, t\n -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual time=602.493..703.274 rows=1000000 loops=1)\n Sort Key: i, t\n Sort Method: external sort Disk: 33240kB\n -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22) (actual time=0.014..129.213 rows=1000000 loops=1)\n Planning time: 0.176 ms\n Execution time: 1685.575 ms\n(8 rows)\n\nTime: 1687.641 ms\n\nNot so big difference here (maybe because I use SSD) but there is\nstill about 50% difference in execution time. Note that I disable\nlocale support.\nI think this is more likely your locale settings, as if I do: create table t(i int, j int, k int, t text collate \"C\");The GROUP BY t,i runs about 25% faster.I've not looked at it yet, but Peter G's patch here https://commitfest.postgresql.org/action/patch_view?id=1462 will quite likely narrow the performance gap between the 2 queries.RegardsDavid Rowley","msg_date":"Sat, 18 Oct 2014 12:53:32 +1300","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimizer on sort aggregate"},{"msg_contents":"Hi, David,\n\nYes, switch sorting order would loose an interesting order so if user\ndictates order by t, i; planner need to resort to its cost model.\nEstimating cardinality of groupby is a much bigger topic than this thread.\n\nI feel sorting string as if it is bytea is particularly interesting. I am\naware Peter G's patch and I think it is great, but for this sort agg case,\nfirst, I believe it is still slower than sorting bytea, and second, Peter\nG's patch depends on data. A common long prefix will make the patch less\neffective, which is probably not so uncommon (for example, URL with a\ndomain prefix). I don't see any downside of sort bytea, other than lost\nthe interest ordering.\n\nThanks,\nFeng\n\n\n\nOn Fri, Oct 17, 2014 at 4:36 PM, David Rowley <dgrowleyml@gmail.com> wrote:\n\n> On Sat, Oct 18, 2014 at 5:10 AM, Feng Tian <fengttt@gmail.com> wrote:\n>\n>> Hi,\n>>\n>> Consider the following queries.\n>>\n>> create table t(i int, j int, k int, t text);\n>> insert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from\n>> generate_series(1, 1000000) i;\n>>\n>> ftian=# explain select count(distinct j) from t group by t, i;\n>> QUERY PLAN\n>> ------------------------------------------------------------------------\n>> GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22)\n>> -> Sort (cost=158029.84..160529.84 rows=1000000 width=22)\n>> Sort Key: t, i\n>> -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)\n>> (4 rows)\n>>\n>>\n>> The query,\n>> select count(distinct j) from t group by t, i;\n>>\n>> runs for 35 seconds. However, if I change the query to,\n>> select count(distinct j) from t group by i, t; -- note switching the\n>> ordering\n>> select count(distinct j) from t group by decode(t, 'escape'), i; --\n>> convert t to bytea\n>>\n>> Run times are just about 5 and 6.5 seconds. The reason is clear, compare\n>> a string with collation is slow, which is well understood by pg hackers.\n>> However, here, the sorting order is forced by the planner, not user.\n>> Planner can do the following optimizations,\n>>\n>> 1. for the sort we generated for sort agg, planner can switch column\n>> ordering, put int before string,\n>> 2. for the sort we generated for sort agg, use bytea compare instead of\n>> string compare.\n>>\n>> They will bring big improvement to this common query. Is this something\n>> reasonable?\n>>\n>>\n> It seems like it might be worth looking into, but I think it's likely more\n> complex than sorting the group by order into narrowest column first.\n>\n> If the query was:\n>\n> select count(distinct j) from t group by t, i order by t, i;\n>\n> Then if that was rewritten to make the group by i,t then the planner\n> would then need to perform an additional sort on t,i to get the correct\n> order by for the final result, which may or may not be faster, it would\n> depend on how many groups there were to sort. Though I guess you could\n> possibly just skip the optimisation if the next node up didn't need any\n> specific ordering.\n>\n> I also wonder if taking into account the column's width is not quite\n> enough. For example if the 'i' column only had 1 distinct value, then\n> performing the group by on 'i' first wouldn't help at all. Keep in mind\n> that the columns could be much closer in width than in your example, e.g\n> int and bigint. Perhaps you'd need to include some sort of heuristics to\n> look at the number of distinct values and the width, and form some sort of\n> weighted estimates about which column to put first.\n>\n> Regards\n>\n> David Rowley\n>\n\nHi, David,Yes, switch sorting order would loose an interesting order so if user dictates order by t, i; planner need to resort to its cost model. Estimating cardinality of groupby is a much bigger topic than this thread.I feel sorting string as if it is bytea is particularly interesting. I am aware Peter G's patch and I think it is great, but for this sort agg case, first, I believe it is still slower than sorting bytea, and second, Peter G's patch depends on data. A common long prefix will make the patch less effective, which is probably not so uncommon (for example, URL with a domain prefix). I don't see any downside of sort bytea, other than lost the interest ordering. Thanks,Feng On Fri, Oct 17, 2014 at 4:36 PM, David Rowley <dgrowleyml@gmail.com> wrote:On Sat, Oct 18, 2014 at 5:10 AM, Feng Tian <fengttt@gmail.com> wrote:Hi, Consider the following queries.create table t(i int, j int, k int, t text);insert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from generate_series(1, 1000000) i;ftian=# explain select count(distinct j) from t group by t, i; QUERY PLAN ------------------------------------------------------------------------ GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22) -> Sort (cost=158029.84..160529.84 rows=1000000 width=22) Sort Key: t, i -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)(4 rows)The query, select count(distinct j) from t group by t, i; runs for 35 seconds. However, if I change the query to,select count(distinct j) from t group by i, t; -- note switching the orderingselect count(distinct j) from t group by decode(t, 'escape'), i; -- convert t to byteaRun times are just about 5 and 6.5 seconds. The reason is clear, compare a string with collation is slow, which is well understood by pg hackers. However, here, the sorting order is forced by the planner, not user. Planner can do the following optimizations,1. for the sort we generated for sort agg, planner can switch column ordering, put int before string,2. for the sort we generated for sort agg, use bytea compare instead of string compare.They will bring big improvement to this common query. Is this something reasonable? It seems like it might be worth looking into, but I think it's likely more complex than sorting the group by order into narrowest column first.If the query was:select count(distinct j) from t group by t, i order by t, i;Then if that was rewritten to make the group by i,t then the planner would then need to perform an additional sort on t,i to get the correct order by for the final result, which may or may not be faster, it would depend on how many groups there were to sort. Though I guess you could possibly just skip the optimisation if the next node up didn't need any specific ordering.I also wonder if taking into account the column's width is not quite enough. For example if the 'i' column only had 1 distinct value, then performing the group by on 'i' first wouldn't help at all. Keep in mind that the columns could be much closer in width than in your example, e.g int and bigint. Perhaps you'd need to include some sort of heuristics to look at the number of distinct values and the width, and form some sort of weighted estimates about which column to put first.RegardsDavid Rowley","msg_date":"Fri, 17 Oct 2014 18:25:14 -0700","msg_from":"Feng Tian <ftian@vitessedata.com>","msg_from_op":false,"msg_subject":"Re: Optimizer on sort aggregate"},{"msg_contents":"On Fri, Oct 17, 2014 at 6:25 PM, Feng Tian <ftian@vitessedata.com> wrote:\n> I feel sorting string as if it is bytea is particularly interesting. I am\n> aware Peter G's patch and I think it is great, but for this sort agg case,\n> first, I believe it is still slower than sorting bytea, and second, Peter\n> G's patch depends on data. A common long prefix will make the patch less\n> effective, which is probably not so uncommon (for example, URL with a domain\n> prefix). I don't see any downside of sort bytea, other than lost the\n> interest ordering.\n\nFWIW, that's probably less true than you'd think. Using Robert's test program:\n\npg@hamster:~/code$ ./strxfrm-binary en_US.UTF-8 \"http://www.something\"\n\"http://www.something\" ->\n131f1f1b2222221e1a18101f131419120109090909090909090909090909090909010909090909090909090909090909090901053d014201420444\n(59 bytes)\npg@hamster:~/code$ ./strxfrm-binary en_US.UTF-8 \"http://www.another\"\n\"http://www.another\" ->\n131f1f1b2222220c191a1f13101d01090909090909090909090909090901090909090909090909090909090901053d014201420444\n(53 bytes)\n\nSo the first eight bytes of the first string is 0x131F1F1B2222221E,\nand the second 0x131F1F1B2222220C. The last byte is different. That's\nbecause the way the Unicode algorithm [1] works, there is often a\nsignificantly greater concentration of entropy in the first 8 bytes as\ncompared to raw C strings compared with memcmp() - punctuation\ncharacters and so on are not actually described at the primary weight\nlevel. If we can get even a single byte to somewhat differentiate each\nstring, we can still win by a very significant amount - just not an\nenormous amount. The break even point is hard to characterize exactly,\nbut I'm quite optimistic that a large majority of real-world text\nsorts will see at least some benefit, while a smaller majority will be\nmuch, much faster.\n\n[1] http://www.unicode.org/reports/tr10/#Notation\n-- \nPeter Geoghegan\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 19:10:34 -0700","msg_from":"Peter Geoghegan <pg@heroku.com>","msg_from_op":false,"msg_subject":"Re: Optimizer on sort aggregate"},{"msg_contents":"On Sat, Oct 18, 2014 at 3:10 AM, Peter Geoghegan <pg@heroku.com> wrote:\n> So the first eight bytes of the first string is 0x131F1F1B2222221E,\n> and the second 0x131F1F1B2222220C. The last byte is different.\n\nThat's interesting but I think it's mostly a quirk of your example.\nAfaics the difference is only that the en_US locale ignores\npunctuation like : and / (or at least treats them as less significant\nthan alphabetic characters). If you had strings that had less\npunctuation or differences that didn't happen to arrive shortly after\nthe 8-byte boundary then it wouldn't make any difference.\n\nAnd we still have to run strfrm at least once, write out the whole\nbinary blob to memory somewhere and if it spills to disk we still have\nto write and read much more data. I think recognizing cases where\nequality is the only thing we're interested in and locale-sensitive\nsorting isn't necessary and using a memcmp would be a clear win.\n\nI'm not immediately clear on what the cleanest way to integrate it\nwould be. A btree opclass function like the cmp function but that\ndoesn't need to be consistent with < and >, only = ? Or perhaps a flag\non the btree opclass that indicates that the data types can safely be\ncompared with memcmp when equality is all that's needed? The latter is\npretty tempting since it would tell code something interesting about\nthe data type's internal storage that may lead to other optimizations.\nOn the other hand the former is nice in that the operator could maybe\nhandle other cases like padding by doing memcmp on only the\nsignificant bits.\n\n\n\n-- \ngreg\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 13:27:22 +0100","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: Optimizer on sort aggregate"},{"msg_contents":"On Sat, Oct 18, 2014 at 5:27 AM, Greg Stark <stark@mit.edu> wrote:\n> That's interesting but I think it's mostly a quirk of your example.\n> Afaics the difference is only that the en_US locale ignores\n> punctuation like : and / (or at least treats them as less significant\n> than alphabetic characters). If you had strings that had less\n> punctuation or differences that didn't happen to arrive shortly after\n> the 8-byte boundary then it wouldn't make any difference.\n\nThe en_US locale treats those kind of punctuation characters as less\nsignificant than alphabetical characters, but is not at all unusual in\ndoing so - all locales I tested do the same. They also do the same for\nspaces. And there is another property of the transformation that is\nvery important for those outside of the English speaking world -\ndiacritics are not represented at the primary weight level. So even\nthough representing certain characters with a diacritic will take 2\nbytes in UTF-8, the corresponding primary weight only takes 1 byte. In\nshort, as I said, the concentration of entropy can easily be a lot\nhigher within the first n bytes of the primary weight level of the\ntransformed blob as compared to the first n bytes of the original\nstring, and frequently *will* be. This will make worst cases for\nabbreviated keys significantly less common than they would otherwise\nbe, while more generally improving performance. Of course, that might\nnot always be as effective as we'd prefer, but that's something that\nyou more or less have to live with in order to have competitive sort\nperformance (even with the HyperLogLog amelioration, you still pay\nsomething for considering the technique). You can always contrive a\ncase that puts things just out of reach, no matter how much entropy\nyou manage to concentrate in the abbreviated key. Fortunately, I don't\nthink those cases are all that representative of what people want from\nsort performance.\n\n> And we still have to run strfrm at least once, write out the whole\n> binary blob to memory somewhere and if it spills to disk we still have\n> to write and read much more data. I think recognizing cases where\n> equality is the only thing we're interested in and locale-sensitive\n> sorting isn't necessary and using a memcmp would be a clear win.\n\nYeah. I was making a point that strictly concerned abbreviated keys as\nproposed in response to Feng's remarks. I am not 100% sure that the\nbenefits of abbreviated keys with locale support aren't worth it here,\nand I say that in full agreement with Feng about locale-aware sorts\nnot actually being necessary. It's clear that cache performance is\nessential to getting good sort performance, which strongly recommends\nabbreviated keys. When we consider the concentration of entropy with\n\"ordinary\" locale-aware abbreviated keys, as compared to abbreviated\nkeys that just use the C locale artificially for this case, it's not\nclear that the concentration of entropy isn't reason enough to prefer\nlocale aware abbreviated keys. Now, it might well be that paying for n\nstrxfrm() operations, rather than doing n straight-up memcpy()\noperations isn't worth it. But I think it might well be worth it -\nparticularly when you factor in the complexity avoided by not\nspecial-casing this. I'm not really sure.\n\nThe general idea of abbreviated keys is almost old hat, to be honest.\nIt just happens to be essential for competitive sort performance.\n-- \nPeter Geoghegan\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 11:01:14 -0700","msg_from":"Peter Geoghegan <pg@heroku.com>","msg_from_op":false,"msg_subject":"Re: Optimizer on sort aggregate"},{"msg_contents":"On Sat, Oct 18, 2014 at 2:25 PM, Feng Tian <ftian@vitessedata.com> wrote:\n\n> Hi, David,\n>\n> Yes, switch sorting order would loose an interesting order so if user\n> dictates order by t, i; planner need to resort to its cost model.\n> Estimating cardinality of groupby is a much bigger topic than this thread.\n>\n>\nWell I don't want to jump in and make the idea more complex than it needs\nto be, More research on this would be really good!\n\nJust to make my thoughts a bit more clear, I had thought that you'd likely\nuse attwidth from pg_statistic to determine the average width of the column\nin order to know if you'd want to play about with the order or not. If not,\nthen how would you know which column to put last in the group by if there\nhappened to be 2 text type columns in the grouping list?\n\nMaybe this could be determined based on some ratio between stadistinct and\nstawidth in pg_statistic. If for example 2 columns had the same width, then\nyou'd likely want to use the one with more distinct values estimated.\nPerhaps the heuristics for determining this would be more complex as, if\nyou had an bigint with 1000 distinct values, then it perhaps would be\nbetter to group by that first before some int with, say 5 distinct values.\nOr maybe not? Some series of benchmarks might some sort of indication if\nthere is any sort of magical tipping point to be sought after here. Of\ncourse the width alone might not be a great thing to base any benchmarks on\nas a multibyte text type with, for example, 6 in the stawidth column, would\nlikely be slower to group by first than a bigint, which would have 8 in the\nstawidth column. People that had carefully written their group bys a\ncertain way for performance might get upset if we made their query slower\nby messing with them too, so I guess the logic should likely only kick in\nif it's a clear win to swap the order.\n\nRegards\n\nDavid Rowley\n\nOn Sat, Oct 18, 2014 at 2:25 PM, Feng Tian <ftian@vitessedata.com> wrote:Hi, David,Yes, switch sorting order would loose an interesting order so if user dictates order by t, i; planner need to resort to its cost model. Estimating cardinality of groupby is a much bigger topic than this thread.Well I don't want to jump in and make the idea more complex than it needs to be, More research on this would be really good!Just to make my thoughts a bit more clear, I had thought that you'd likely use attwidth from pg_statistic to determine the average width of the column in order to know if you'd want to play about with the order or not. If not, then how would you know which column to put last in the group by if there happened to be 2 text type columns in the grouping list? Maybe this could be determined based on some ratio between stadistinct and stawidth in pg_statistic. If for example 2 columns had the same width, then you'd likely want to use the one with more distinct values estimated. Perhaps the heuristics for determining this would be more complex as, if you had an bigint with 1000 distinct values, then it perhaps would be better to group by that first before some int with, say 5 distinct values. Or maybe not? Some series of benchmarks might some sort of indication if there is any sort of magical tipping point to be sought after here. Of course the width alone might not be a great thing to base any benchmarks on as a multibyte text type with, for example, 6 in the stawidth column, would likely be slower to group by first than a bigint, which would have 8 in the stawidth column. People that had carefully written their group bys a certain way for performance might get upset if we made their query slower by messing with them too, so I guess the logic should likely only kick in if it's a clear win to swap the order.RegardsDavid Rowley","msg_date":"Sun, 19 Oct 2014 22:35:51 +1300","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimizer on sort aggregate"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nConsider the following queries.\\n\\ncreate table t(i int, j int, k int, t text);\\ninsert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from\\ngenerate_series(1, 1000000) i;\\n\\nftian=# explain select count(distinct j) from t group by t, i;\\n QUERY PLAN\\n------------------------------------------------------------------------\\n GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22)\\n -> Sort (cost=158029.84..160529.84 rows=1000000 width=22)\\n Sort Key: t, i\\n -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)\\n(4 rows)\\n\\n\\nThe query,\\nselect count(distinct j) from t group by t, i;\\n\\nruns for 35 seconds. However, if I change the query to,\\nselect count(distinct j) from t group by i, t; -- note switching the\\nordering\\nselect count(distinct j) from t group by decode(t, 'escape'), i; -- convert\\nt to bytea\\n\\nRun times are just about 5 and 6.5 seconds. The reason is clear, compare a\\nstring with collation is slow, which is well understood by pg hackers.\\nHowever, here, the sorting order is forced by the planner, not user.\\nPlanner can do the following optimizations,\\n\\n1. for the sort we generated for sort agg, planner can switch column\\nordering, put int before string,\\n2. for the sort we generated for sort agg, use bytea compare instead of\\nstring compare.\\n\\nThey will bring big improvement to this common query. Is this something\\nreasonable?\\n\\nThanks,\\n\\nHi, Consider the following queries.create table t(i int, j int, k int, t text);insert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from generate_series(1, 1000000) i;ftian=# explain select count(distinct j) from t group by t, i; QUERY PLAN ------------------------------------------------------------------------ GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22) -> Sort (cost=158029.84..160529.84 rows=1000000 width=22) Sort Key: t, i -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)(4 rows)The query, select count(distinct j) from t group by t, i; runs for 35 seconds. However, if I change the query to,select count(distinct j) from t group by i, t; -- note switching the orderingselect count(distinct j) from t group by decode(t, 'escape'), i; -- convert t to byteaRun times are just about 5 and 6.5 seconds. The reason is clear, compare a string with collation is slow, which is well understood by pg hackers. However, here, the sorting order is forced by the planner, not user. Planner can do the following optimizations,1. for the sort we generated for sort agg, planner can switch column ordering, put int before string,2. for the sort we generated for sort agg, use bytea compare instead of string compare.They will bring big improvement to this common query. Is this something reasonable? Thanks,\",\n \"msg_date\": \"Fri, 17 Oct 2014 09:10:13 -0700\",\n \"msg_from\": \"Feng Tian <fengttt@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Optimizer on sort aggregate\"\n },\n {\n \"msg_contents\": \"> The query,\\n> select count(distinct j) from t group by t, i;\\n> \\n> runs for 35 seconds. However, if I change the query to,\\n> select count(distinct j) from t group by i, t; -- note switching the\\n> ordering\\n> select count(distinct j) from t group by decode(t, 'escape'), i; -- convert\\n> t to bytea\\n> \\n> Run times are just about 5 and 6.5 seconds. The reason is clear, compare a\\n> string with collation is slow, which is well understood by pg hackers.\\n> However, here, the sorting order is forced by the planner, not user.\\n> Planner can do the following optimizations,\\n\\nInteresting. I got following result:\\n\\ntest=# explain analyze select count(distinct j) from t group by t, i;\\n QUERY PLAN \\n--------------------------------------------------------------------------------------------------------------------------\\n GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual time=1332.937..2431.238 rows=1000000 loops=1)\\n Group Key: t, i\\n -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual time=1332.922..1507.413 rows=1000000 loops=1)\\n Sort Key: t, i\\n Sort Method: external merge Disk: 33232kB\\n -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22) (actual time=0.006..131.406 rows=1000000 loops=1)\\n Planning time: 0.031 ms\\n Execution time: 2484.271 ms\\n(8 rows)\\n\\nTime: 2484.520 ms\\n\\ntest=# explain analyze select count(distinct j) from t group by i, t;\\n QUERY PLAN \\n--------------------------------------------------------------------------------------------------------------------------\\n GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual time=602.510..1632.087 rows=1000000 loops=1)\\n Group Key: i, t\\n -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual time=602.493..703.274 rows=1000000 loops=1)\\n Sort Key: i, t\\n Sort Method: external sort Disk: 33240kB\\n -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22) (actual time=0.014..129.213 rows=1000000 loops=1)\\n Planning time: 0.176 ms\\n Execution time: 1685.575 ms\\n(8 rows)\\n\\nTime: 1687.641 ms\\n\\nNot so big difference here (maybe because I use SSD) but there is\\nstill about 50% difference in execution time. Note that I disable\\nlocale support.\\n\\n> 1. for the sort we generated for sort agg, planner can switch column\\n> ordering, put int before string,\\n> 2. for the sort we generated for sort agg, use bytea compare instead of\\n> string compare.\\n> \\n> They will bring big improvement to this common query. Is this something\\n> reasonable?\\n\\nNot looking at sort and planner code closely, I guess planner does\\nnot recognize the cost difference between \\\"external merge\\\" and\\n\\\"external sort\\\" because cost estimations for sort step in each plan\\nare exactly same (137519.84..140019.84). However I am not sure if we\\nshould fix the planner or should fix our sort machinery since it is\\npossible that the sort machinery should not expose such a big\\ndifference in the two sort methods.\\n--\\nTatsuo Ishii\\nSRA OSS, Inc. Japan\\nEnglish: http://www.sraoss.co.jp/index_en.php\\nJapanese:http://www.sraoss.co.jp\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 08:35:16 +0900 (JST)\",\n \"msg_from\": \"Tatsuo Ishii <ishii@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimizer on sort aggregate\"\n },\n {\n \"msg_contents\": \"On Sat, Oct 18, 2014 at 5:10 AM, Feng Tian <fengttt@gmail.com> wrote:\\n\\n> Hi,\\n>\\n> Consider the following queries.\\n>\\n> create table t(i int, j int, k int, t text);\\n> insert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from\\n> generate_series(1, 1000000) i;\\n>\\n> ftian=# explain select count(distinct j) from t group by t, i;\\n> QUERY PLAN\\n> ------------------------------------------------------------------------\\n> GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22)\\n> -> Sort (cost=158029.84..160529.84 rows=1000000 width=22)\\n> Sort Key: t, i\\n> -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)\\n> (4 rows)\\n>\\n>\\n> The query,\\n> select count(distinct j) from t group by t, i;\\n>\\n> runs for 35 seconds. However, if I change the query to,\\n> select count(distinct j) from t group by i, t; -- note switching the\\n> ordering\\n> select count(distinct j) from t group by decode(t, 'escape'), i; --\\n> convert t to bytea\\n>\\n> Run times are just about 5 and 6.5 seconds. The reason is clear, compare\\n> a string with collation is slow, which is well understood by pg hackers.\\n> However, here, the sorting order is forced by the planner, not user.\\n> Planner can do the following optimizations,\\n>\\n> 1. for the sort we generated for sort agg, planner can switch column\\n> ordering, put int before string,\\n> 2. for the sort we generated for sort agg, use bytea compare instead of\\n> string compare.\\n>\\n> They will bring big improvement to this common query. Is this something\\n> reasonable?\\n>\\n>\\nIt seems like it might be worth looking into, but I think it's likely more\\ncomplex than sorting the group by order into narrowest column first.\\n\\nIf the query was:\\n\\nselect count(distinct j) from t group by t, i order by t, i;\\n\\nThen if that was rewritten to make the group by i,t then the planner would\\nthen need to perform an additional sort on t,i to get the correct order by\\nfor the final result, which may or may not be faster, it would depend on\\nhow many groups there were to sort. Though I guess you could possibly just\\nskip the optimisation if the next node up didn't need any specific ordering.\\n\\nI also wonder if taking into account the column's width is not quite\\nenough. For example if the 'i' column only had 1 distinct value, then\\nperforming the group by on 'i' first wouldn't help at all. Keep in mind\\nthat the columns could be much closer in width than in your example, e.g\\nint and bigint. Perhaps you'd need to include some sort of heuristics to\\nlook at the number of distinct values and the width, and form some sort of\\nweighted estimates about which column to put first.\\n\\nRegards\\n\\nDavid Rowley\\n\\nOn Sat, Oct 18, 2014 at 5:10 AM, Feng Tian <fengttt@gmail.com> wrote:Hi, Consider the following queries.create table t(i int, j int, k int, t text);insert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from generate_series(1, 1000000) i;ftian=# explain select count(distinct j) from t group by t, i; QUERY PLAN ------------------------------------------------------------------------ GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22) -> Sort (cost=158029.84..160529.84 rows=1000000 width=22) Sort Key: t, i -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)(4 rows)The query, select count(distinct j) from t group by t, i; runs for 35 seconds. However, if I change the query to,select count(distinct j) from t group by i, t; -- note switching the orderingselect count(distinct j) from t group by decode(t, 'escape'), i; -- convert t to byteaRun times are just about 5 and 6.5 seconds. The reason is clear, compare a string with collation is slow, which is well understood by pg hackers. However, here, the sorting order is forced by the planner, not user. Planner can do the following optimizations,1. for the sort we generated for sort agg, planner can switch column ordering, put int before string,2. for the sort we generated for sort agg, use bytea compare instead of string compare.They will bring big improvement to this common query. Is this something reasonable? It seems like it might be worth looking into, but I think it's likely more complex than sorting the group by order into narrowest column first.If the query was:select count(distinct j) from t group by t, i order by t, i;Then if that was rewritten to make the group by i,t then the planner would then need to perform an additional sort on t,i to get the correct order by for the final result, which may or may not be faster, it would depend on how many groups there were to sort. Though I guess you could possibly just skip the optimisation if the next node up didn't need any specific ordering.I also wonder if taking into account the column's width is not quite enough. For example if the 'i' column only had 1 distinct value, then performing the group by on 'i' first wouldn't help at all. Keep in mind that the columns could be much closer in width than in your example, e.g int and bigint. Perhaps you'd need to include some sort of heuristics to look at the number of distinct values and the width, and form some sort of weighted estimates about which column to put first.RegardsDavid Rowley\",\n \"msg_date\": \"Sat, 18 Oct 2014 12:36:30 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimizer on sort aggregate\"\n },\n {\n \"msg_contents\": \"On Sat, Oct 18, 2014 at 12:35 PM, Tatsuo Ishii <ishii@postgresql.org> wrote:\\n\\n> > The query,\\n> > select count(distinct j) from t group by t, i;\\n> >\\n> > runs for 35 seconds. However, if I change the query to,\\n> > select count(distinct j) from t group by i, t; -- note switching the\\n> > ordering\\n> > select count(distinct j) from t group by decode(t, 'escape'), i; --\\n> convert\\n> > t to bytea\\n> >\\n> > Run times are just about 5 and 6.5 seconds. The reason is clear,\\n> compare a\\n> > string with collation is slow, which is well understood by pg hackers.\\n> > However, here, the sorting order is forced by the planner, not user.\\n> > Planner can do the following optimizations,\\n>\\n> Interesting. I got following result:\\n>\\n> test=# explain analyze select count(distinct j) from t group by t, i;\\n> QUERY PLAN\\n>\\n> --------------------------------------------------------------------------------------------------------------------------\\n> GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual\\n> time=1332.937..2431.238 rows=1000000 loops=1)\\n> Group Key: t, i\\n> -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual\\n> time=1332.922..1507.413 rows=1000000 loops=1)\\n> Sort Key: t, i\\n> Sort Method: external merge Disk: 33232kB\\n> -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)\\n> (actual time=0.006..131.406 rows=1000000 loops=1)\\n> Planning time: 0.031 ms\\n> Execution time: 2484.271 ms\\n> (8 rows)\\n>\\n> Time: 2484.520 ms\\n>\\n> test=# explain analyze select count(distinct j) from t group by i, t;\\n> QUERY PLAN\\n>\\n> --------------------------------------------------------------------------------------------------------------------------\\n> GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual\\n> time=602.510..1632.087 rows=1000000 loops=1)\\n> Group Key: i, t\\n> -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual\\n> time=602.493..703.274 rows=1000000 loops=1)\\n> Sort Key: i, t\\n> Sort Method: external sort Disk: 33240kB\\n> -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)\\n> (actual time=0.014..129.213 rows=1000000 loops=1)\\n> Planning time: 0.176 ms\\n> Execution time: 1685.575 ms\\n> (8 rows)\\n>\\n> Time: 1687.641 ms\\n>\\n> Not so big difference here (maybe because I use SSD) but there is\\n> still about 50% difference in execution time. Note that I disable\\n> locale support.\\n>\\n>\\nI think this is more likely your locale settings, as if I do:\\n\\n create table t(i int, j int, k int, t text collate \\\"C\\\");\\nThe GROUP BY t,i runs about 25% faster.\\n\\nI've not looked at it yet, but Peter G's patch here\\nhttps://commitfest.postgresql.org/action/patch_view?id=1462 will quite\\nlikely narrow the performance gap between the 2 queries.\\n\\nRegards\\n\\nDavid Rowley\\n\\nOn Sat, Oct 18, 2014 at 12:35 PM, Tatsuo Ishii <ishii@postgresql.org> wrote:> The query,\\n> select count(distinct j) from t group by t, i;\\n>\\n> runs for 35 seconds. However, if I change the query to,\\n> select count(distinct j) from t group by i, t; -- note switching the\\n> ordering\\n> select count(distinct j) from t group by decode(t, 'escape'), i; -- convert\\n> t to bytea\\n>\\n> Run times are just about 5 and 6.5 seconds. The reason is clear, compare a\\n> string with collation is slow, which is well understood by pg hackers.\\n> However, here, the sorting order is forced by the planner, not user.\\n> Planner can do the following optimizations,\\n\\nInteresting. I got following result:\\n\\ntest=# explain analyze select count(distinct j) from t group by t, i;\\n QUERY PLAN\\n--------------------------------------------------------------------------------------------------------------------------\\n GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual time=1332.937..2431.238 rows=1000000 loops=1)\\n Group Key: t, i\\n -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual time=1332.922..1507.413 rows=1000000 loops=1)\\n Sort Key: t, i\\n Sort Method: external merge Disk: 33232kB\\n -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22) (actual time=0.006..131.406 rows=1000000 loops=1)\\n Planning time: 0.031 ms\\n Execution time: 2484.271 ms\\n(8 rows)\\n\\nTime: 2484.520 ms\\n\\ntest=# explain analyze select count(distinct j) from t group by i, t;\\n QUERY PLAN\\n--------------------------------------------------------------------------------------------------------------------------\\n GroupAggregate (cost=137519.84..157519.84 rows=1000000 width=22) (actual time=602.510..1632.087 rows=1000000 loops=1)\\n Group Key: i, t\\n -> Sort (cost=137519.84..140019.84 rows=1000000 width=22) (actual time=602.493..703.274 rows=1000000 loops=1)\\n Sort Key: i, t\\n Sort Method: external sort Disk: 33240kB\\n -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22) (actual time=0.014..129.213 rows=1000000 loops=1)\\n Planning time: 0.176 ms\\n Execution time: 1685.575 ms\\n(8 rows)\\n\\nTime: 1687.641 ms\\n\\nNot so big difference here (maybe because I use SSD) but there is\\nstill about 50% difference in execution time. Note that I disable\\nlocale support.\\nI think this is more likely your locale settings, as if I do: create table t(i int, j int, k int, t text collate \\\"C\\\");The GROUP BY t,i runs about 25% faster.I've not looked at it yet, but Peter G's patch here https://commitfest.postgresql.org/action/patch_view?id=1462 will quite likely narrow the performance gap between the 2 queries.RegardsDavid Rowley\",\n \"msg_date\": \"Sat, 18 Oct 2014 12:53:32 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimizer on sort aggregate\"\n },\n {\n \"msg_contents\": \"Hi, David,\\n\\nYes, switch sorting order would loose an interesting order so if user\\ndictates order by t, i; planner need to resort to its cost model.\\nEstimating cardinality of groupby is a much bigger topic than this thread.\\n\\nI feel sorting string as if it is bytea is particularly interesting. I am\\naware Peter G's patch and I think it is great, but for this sort agg case,\\nfirst, I believe it is still slower than sorting bytea, and second, Peter\\nG's patch depends on data. A common long prefix will make the patch less\\neffective, which is probably not so uncommon (for example, URL with a\\ndomain prefix). I don't see any downside of sort bytea, other than lost\\nthe interest ordering.\\n\\nThanks,\\nFeng\\n\\n\\n\\nOn Fri, Oct 17, 2014 at 4:36 PM, David Rowley <dgrowleyml@gmail.com> wrote:\\n\\n> On Sat, Oct 18, 2014 at 5:10 AM, Feng Tian <fengttt@gmail.com> wrote:\\n>\\n>> Hi,\\n>>\\n>> Consider the following queries.\\n>>\\n>> create table t(i int, j int, k int, t text);\\n>> insert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from\\n>> generate_series(1, 1000000) i;\\n>>\\n>> ftian=# explain select count(distinct j) from t group by t, i;\\n>> QUERY PLAN\\n>> ------------------------------------------------------------------------\\n>> GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22)\\n>> -> Sort (cost=158029.84..160529.84 rows=1000000 width=22)\\n>> Sort Key: t, i\\n>> -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)\\n>> (4 rows)\\n>>\\n>>\\n>> The query,\\n>> select count(distinct j) from t group by t, i;\\n>>\\n>> runs for 35 seconds. However, if I change the query to,\\n>> select count(distinct j) from t group by i, t; -- note switching the\\n>> ordering\\n>> select count(distinct j) from t group by decode(t, 'escape'), i; --\\n>> convert t to bytea\\n>>\\n>> Run times are just about 5 and 6.5 seconds. The reason is clear, compare\\n>> a string with collation is slow, which is well understood by pg hackers.\\n>> However, here, the sorting order is forced by the planner, not user.\\n>> Planner can do the following optimizations,\\n>>\\n>> 1. for the sort we generated for sort agg, planner can switch column\\n>> ordering, put int before string,\\n>> 2. for the sort we generated for sort agg, use bytea compare instead of\\n>> string compare.\\n>>\\n>> They will bring big improvement to this common query. Is this something\\n>> reasonable?\\n>>\\n>>\\n> It seems like it might be worth looking into, but I think it's likely more\\n> complex than sorting the group by order into narrowest column first.\\n>\\n> If the query was:\\n>\\n> select count(distinct j) from t group by t, i order by t, i;\\n>\\n> Then if that was rewritten to make the group by i,t then the planner\\n> would then need to perform an additional sort on t,i to get the correct\\n> order by for the final result, which may or may not be faster, it would\\n> depend on how many groups there were to sort. Though I guess you could\\n> possibly just skip the optimisation if the next node up didn't need any\\n> specific ordering.\\n>\\n> I also wonder if taking into account the column's width is not quite\\n> enough. For example if the 'i' column only had 1 distinct value, then\\n> performing the group by on 'i' first wouldn't help at all. Keep in mind\\n> that the columns could be much closer in width than in your example, e.g\\n> int and bigint. Perhaps you'd need to include some sort of heuristics to\\n> look at the number of distinct values and the width, and form some sort of\\n> weighted estimates about which column to put first.\\n>\\n> Regards\\n>\\n> David Rowley\\n>\\n\\nHi, David,Yes, switch sorting order would loose an interesting order so if user dictates order by t, i; planner need to resort to its cost model. Estimating cardinality of groupby is a much bigger topic than this thread.I feel sorting string as if it is bytea is particularly interesting. I am aware Peter G's patch and I think it is great, but for this sort agg case, first, I believe it is still slower than sorting bytea, and second, Peter G's patch depends on data. A common long prefix will make the patch less effective, which is probably not so uncommon (for example, URL with a domain prefix). I don't see any downside of sort bytea, other than lost the interest ordering. Thanks,Feng On Fri, Oct 17, 2014 at 4:36 PM, David Rowley <dgrowleyml@gmail.com> wrote:On Sat, Oct 18, 2014 at 5:10 AM, Feng Tian <fengttt@gmail.com> wrote:Hi, Consider the following queries.create table t(i int, j int, k int, t text);insert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from generate_series(1, 1000000) i;ftian=# explain select count(distinct j) from t group by t, i; QUERY PLAN ------------------------------------------------------------------------ GroupAggregate (cost=158029.84..178029.84 rows=1000000 width=22) -> Sort (cost=158029.84..160529.84 rows=1000000 width=22) Sort Key: t, i -> Seq Scan on t (cost=0.00..17352.00 rows=1000000 width=22)(4 rows)The query, select count(distinct j) from t group by t, i; runs for 35 seconds. However, if I change the query to,select count(distinct j) from t group by i, t; -- note switching the orderingselect count(distinct j) from t group by decode(t, 'escape'), i; -- convert t to byteaRun times are just about 5 and 6.5 seconds. The reason is clear, compare a string with collation is slow, which is well understood by pg hackers. However, here, the sorting order is forced by the planner, not user. Planner can do the following optimizations,1. for the sort we generated for sort agg, planner can switch column ordering, put int before string,2. for the sort we generated for sort agg, use bytea compare instead of string compare.They will bring big improvement to this common query. Is this something reasonable? It seems like it might be worth looking into, but I think it's likely more complex than sorting the group by order into narrowest column first.If the query was:select count(distinct j) from t group by t, i order by t, i;Then if that was rewritten to make the group by i,t then the planner would then need to perform an additional sort on t,i to get the correct order by for the final result, which may or may not be faster, it would depend on how many groups there were to sort. Though I guess you could possibly just skip the optimisation if the next node up didn't need any specific ordering.I also wonder if taking into account the column's width is not quite enough. For example if the 'i' column only had 1 distinct value, then performing the group by on 'i' first wouldn't help at all. Keep in mind that the columns could be much closer in width than in your example, e.g int and bigint. Perhaps you'd need to include some sort of heuristics to look at the number of distinct values and the width, and form some sort of weighted estimates about which column to put first.RegardsDavid Rowley\",\n \"msg_date\": \"Fri, 17 Oct 2014 18:25:14 -0700\",\n \"msg_from\": \"Feng Tian <ftian@vitessedata.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimizer on sort aggregate\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 6:25 PM, Feng Tian <ftian@vitessedata.com> wrote:\\n> I feel sorting string as if it is bytea is particularly interesting. I am\\n> aware Peter G's patch and I think it is great, but for this sort agg case,\\n> first, I believe it is still slower than sorting bytea, and second, Peter\\n> G's patch depends on data. A common long prefix will make the patch less\\n> effective, which is probably not so uncommon (for example, URL with a domain\\n> prefix). I don't see any downside of sort bytea, other than lost the\\n> interest ordering.\\n\\nFWIW, that's probably less true than you'd think. Using Robert's test program:\\n\\npg@hamster:~/code$ ./strxfrm-binary en_US.UTF-8 \\\"http://www.something\\\"\\n\\\"http://www.something\\\" ->\\n131f1f1b2222221e1a18101f131419120109090909090909090909090909090909010909090909090909090909090909090901053d014201420444\\n(59 bytes)\\npg@hamster:~/code$ ./strxfrm-binary en_US.UTF-8 \\\"http://www.another\\\"\\n\\\"http://www.another\\\" ->\\n131f1f1b2222220c191a1f13101d01090909090909090909090909090901090909090909090909090909090901053d014201420444\\n(53 bytes)\\n\\nSo the first eight bytes of the first string is 0x131F1F1B2222221E,\\nand the second 0x131F1F1B2222220C. The last byte is different. That's\\nbecause the way the Unicode algorithm [1] works, there is often a\\nsignificantly greater concentration of entropy in the first 8 bytes as\\ncompared to raw C strings compared with memcmp() - punctuation\\ncharacters and so on are not actually described at the primary weight\\nlevel. If we can get even a single byte to somewhat differentiate each\\nstring, we can still win by a very significant amount - just not an\\nenormous amount. The break even point is hard to characterize exactly,\\nbut I'm quite optimistic that a large majority of real-world text\\nsorts will see at least some benefit, while a smaller majority will be\\nmuch, much faster.\\n\\n[1] http://www.unicode.org/reports/tr10/#Notation\\n-- \\nPeter Geoghegan\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 19:10:34 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@heroku.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimizer on sort aggregate\"\n },\n {\n \"msg_contents\": \"On Sat, Oct 18, 2014 at 3:10 AM, Peter Geoghegan <pg@heroku.com> wrote:\\n> So the first eight bytes of the first string is 0x131F1F1B2222221E,\\n> and the second 0x131F1F1B2222220C. The last byte is different.\\n\\nThat's interesting but I think it's mostly a quirk of your example.\\nAfaics the difference is only that the en_US locale ignores\\npunctuation like : and / (or at least treats them as less significant\\nthan alphabetic characters). If you had strings that had less\\npunctuation or differences that didn't happen to arrive shortly after\\nthe 8-byte boundary then it wouldn't make any difference.\\n\\nAnd we still have to run strfrm at least once, write out the whole\\nbinary blob to memory somewhere and if it spills to disk we still have\\nto write and read much more data. I think recognizing cases where\\nequality is the only thing we're interested in and locale-sensitive\\nsorting isn't necessary and using a memcmp would be a clear win.\\n\\nI'm not immediately clear on what the cleanest way to integrate it\\nwould be. A btree opclass function like the cmp function but that\\ndoesn't need to be consistent with < and >, only = ? Or perhaps a flag\\non the btree opclass that indicates that the data types can safely be\\ncompared with memcmp when equality is all that's needed? The latter is\\npretty tempting since it would tell code something interesting about\\nthe data type's internal storage that may lead to other optimizations.\\nOn the other hand the former is nice in that the operator could maybe\\nhandle other cases like padding by doing memcmp on only the\\nsignificant bits.\\n\\n\\n\\n-- \\ngreg\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 13:27:22 +0100\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimizer on sort aggregate\"\n },\n {\n \"msg_contents\": \"On Sat, Oct 18, 2014 at 5:27 AM, Greg Stark <stark@mit.edu> wrote:\\n> That's interesting but I think it's mostly a quirk of your example.\\n> Afaics the difference is only that the en_US locale ignores\\n> punctuation like : and / (or at least treats them as less significant\\n> than alphabetic characters). If you had strings that had less\\n> punctuation or differences that didn't happen to arrive shortly after\\n> the 8-byte boundary then it wouldn't make any difference.\\n\\nThe en_US locale treats those kind of punctuation characters as less\\nsignificant than alphabetical characters, but is not at all unusual in\\ndoing so - all locales I tested do the same. They also do the same for\\nspaces. And there is another property of the transformation that is\\nvery important for those outside of the English speaking world -\\ndiacritics are not represented at the primary weight level. So even\\nthough representing certain characters with a diacritic will take 2\\nbytes in UTF-8, the corresponding primary weight only takes 1 byte. In\\nshort, as I said, the concentration of entropy can easily be a lot\\nhigher within the first n bytes of the primary weight level of the\\ntransformed blob as compared to the first n bytes of the original\\nstring, and frequently *will* be. This will make worst cases for\\nabbreviated keys significantly less common than they would otherwise\\nbe, while more generally improving performance. Of course, that might\\nnot always be as effective as we'd prefer, but that's something that\\nyou more or less have to live with in order to have competitive sort\\nperformance (even with the HyperLogLog amelioration, you still pay\\nsomething for considering the technique). You can always contrive a\\ncase that puts things just out of reach, no matter how much entropy\\nyou manage to concentrate in the abbreviated key. Fortunately, I don't\\nthink those cases are all that representative of what people want from\\nsort performance.\\n\\n> And we still have to run strfrm at least once, write out the whole\\n> binary blob to memory somewhere and if it spills to disk we still have\\n> to write and read much more data. I think recognizing cases where\\n> equality is the only thing we're interested in and locale-sensitive\\n> sorting isn't necessary and using a memcmp would be a clear win.\\n\\nYeah. I was making a point that strictly concerned abbreviated keys as\\nproposed in response to Feng's remarks. I am not 100% sure that the\\nbenefits of abbreviated keys with locale support aren't worth it here,\\nand I say that in full agreement with Feng about locale-aware sorts\\nnot actually being necessary. It's clear that cache performance is\\nessential to getting good sort performance, which strongly recommends\\nabbreviated keys. When we consider the concentration of entropy with\\n\\\"ordinary\\\" locale-aware abbreviated keys, as compared to abbreviated\\nkeys that just use the C locale artificially for this case, it's not\\nclear that the concentration of entropy isn't reason enough to prefer\\nlocale aware abbreviated keys. Now, it might well be that paying for n\\nstrxfrm() operations, rather than doing n straight-up memcpy()\\noperations isn't worth it. But I think it might well be worth it -\\nparticularly when you factor in the complexity avoided by not\\nspecial-casing this. I'm not really sure.\\n\\nThe general idea of abbreviated keys is almost old hat, to be honest.\\nIt just happens to be essential for competitive sort performance.\\n-- \\nPeter Geoghegan\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 11:01:14 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@heroku.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimizer on sort aggregate\"\n },\n {\n \"msg_contents\": \"On Sat, Oct 18, 2014 at 2:25 PM, Feng Tian <ftian@vitessedata.com> wrote:\\n\\n> Hi, David,\\n>\\n> Yes, switch sorting order would loose an interesting order so if user\\n> dictates order by t, i; planner need to resort to its cost model.\\n> Estimating cardinality of groupby is a much bigger topic than this thread.\\n>\\n>\\nWell I don't want to jump in and make the idea more complex than it needs\\nto be, More research on this would be really good!\\n\\nJust to make my thoughts a bit more clear, I had thought that you'd likely\\nuse attwidth from pg_statistic to determine the average width of the column\\nin order to know if you'd want to play about with the order or not. If not,\\nthen how would you know which column to put last in the group by if there\\nhappened to be 2 text type columns in the grouping list?\\n\\nMaybe this could be determined based on some ratio between stadistinct and\\nstawidth in pg_statistic. If for example 2 columns had the same width, then\\nyou'd likely want to use the one with more distinct values estimated.\\nPerhaps the heuristics for determining this would be more complex as, if\\nyou had an bigint with 1000 distinct values, then it perhaps would be\\nbetter to group by that first before some int with, say 5 distinct values.\\nOr maybe not? Some series of benchmarks might some sort of indication if\\nthere is any sort of magical tipping point to be sought after here. Of\\ncourse the width alone might not be a great thing to base any benchmarks on\\nas a multibyte text type with, for example, 6 in the stawidth column, would\\nlikely be slower to group by first than a bigint, which would have 8 in the\\nstawidth column. People that had carefully written their group bys a\\ncertain way for performance might get upset if we made their query slower\\nby messing with them too, so I guess the logic should likely only kick in\\nif it's a clear win to swap the order.\\n\\nRegards\\n\\nDavid Rowley\\n\\nOn Sat, Oct 18, 2014 at 2:25 PM, Feng Tian <ftian@vitessedata.com> wrote:Hi, David,Yes, switch sorting order would loose an interesting order so if user dictates order by t, i; planner need to resort to its cost model. Estimating cardinality of groupby is a much bigger topic than this thread.Well I don't want to jump in and make the idea more complex than it needs to be, More research on this would be really good!Just to make my thoughts a bit more clear, I had thought that you'd likely use attwidth from pg_statistic to determine the average width of the column in order to know if you'd want to play about with the order or not. If not, then how would you know which column to put last in the group by if there happened to be 2 text type columns in the grouping list? Maybe this could be determined based on some ratio between stadistinct and stawidth in pg_statistic. If for example 2 columns had the same width, then you'd likely want to use the one with more distinct values estimated. Perhaps the heuristics for determining this would be more complex as, if you had an bigint with 1000 distinct values, then it perhaps would be better to group by that first before some int with, say 5 distinct values. Or maybe not? Some series of benchmarks might some sort of indication if there is any sort of magical tipping point to be sought after here. Of course the width alone might not be a great thing to base any benchmarks on as a multibyte text type with, for example, 6 in the stawidth column, would likely be slower to group by first than a bigint, which would have 8 in the stawidth column. People that had carefully written their group bys a certain way for performance might get upset if we made their query slower by messing with them too, so I guess the logic should likely only kick in if it's a clear win to swap the order.RegardsDavid Rowley\",\n \"msg_date\": \"Sun, 19 Oct 2014 22:35:51 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimizer on sort aggregate\"\n }\n]","isConversational":false}}},{"rowIdx":39678,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Now that we have the create hash index warning in 9.5, I realized that\nwe don't warn about hash indexes with PITR, only crash recovery and\nstreaming. This patch fixes that.\n\nIs the wording \"cannot be used\" too vague. The CREATE INDEX manual\npage has the words \"give wrong answers to queries\", which might be\nbetter, but is kind of long for an error message. Suggestions?\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Fri, 17 Oct 2014 12:22:19 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":true,"msg_subject":"Hash index creation warning"},{"msg_contents":"Bruce Momjian wrote\n> Now that we have the create hash index warning in 9.5, I realized that\n> we don't warn about hash indexes with PITR, only crash recovery and\n> streaming. This patch fixes that.\n> \n> Is the wording \"cannot be used\" too vague. The CREATE INDEX manual\n> page has the words \"give wrong answers to queries\", which might be\n> better, but is kind of long for an error message. Suggestions?\n\nSomething like the following is more specific without being more wordy:\n\n\"hash indexes are not WAL-logged: they are corrupted during recovery and\nchanges do not replicate to standby servers.\"\n\nThe question is whether we explain the implications of not being WAL-logged\nin an error message or simply state the fact and let the documentation\nexplain the hazards - basically just output:\n\n\"hash indexes are not WAL-logged and their use is discouraged\"\n\nDavid J.\n\n\n\n\n--\nView this message in context: http://postgresql.1045698.n5.nabble.com/Hash-index-creation-warning-tp5823443p5823445.html\nSent from the PostgreSQL - hackers mailing list archive at Nabble.com.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 09:43:38 -0700 (PDT)","msg_from":"David G Johnston <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"David G Johnston <david.g.johnston@gmail.com> writes:\n> The question is whether we explain the implications of not being WAL-logged\n> in an error message or simply state the fact and let the documentation\n> explain the hazards - basically just output:\n> \"hash indexes are not WAL-logged and their use is discouraged\"\n\n+1. The warning message is not the place to be trying to explain all the\ndetails.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 12:56:52 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\n> David G Johnston <david.g.johnston@gmail.com> writes:\n> > The question is whether we explain the implications of not being WAL-logged\n> > in an error message or simply state the fact and let the documentation\n> > explain the hazards - basically just output:\n> > \"hash indexes are not WAL-logged and their use is discouraged\"\n> \n> +1. The warning message is not the place to be trying to explain all the\n> details.\n\nOK, updated patch attached.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Fri, 17 Oct 2014 14:36:55 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":true,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On Fri, Oct 17, 2014 at 02:36:55PM -0400, Bruce Momjian wrote:\n> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\n> > David G Johnston <david.g.johnston@gmail.com> writes:\n> > > The question is whether we explain the implications of not being WAL-logged\n> > > in an error message or simply state the fact and let the documentation\n> > > explain the hazards - basically just output:\n> > > \"hash indexes are not WAL-logged and their use is discouraged\"\n> > \n> > +1. The warning message is not the place to be trying to explain all the\n> > details.\n> \n> OK, updated patch attached.\n\nPatch applied.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 10:36:24 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":true,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On 18 October 2014 at 15:36, Bruce Momjian <bruce@momjian.us> wrote:\n> On Fri, Oct 17, 2014 at 02:36:55PM -0400, Bruce Momjian wrote:\n>> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\n>> > David G Johnston <david.g.johnston@gmail.com> writes:\n>> > > The question is whether we explain the implications of not being WAL-logged\n>> > > in an error message or simply state the fact and let the documentation\n>> > > explain the hazards - basically just output:\n>> > > \"hash indexes are not WAL-logged and their use is discouraged\"\n>> >\n>> > +1. The warning message is not the place to be trying to explain all the\n>> > details.\n>>\n>> OK, updated patch attached.\n>\n> Patch applied.\n\nI only just noticed this item when I read the release notes. Should\nwe bother warning when used on an unlogged table?\n\n-- \nThom\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 12 Jun 2015 23:00:54 +0100","msg_from":"Thom Brown <thom@linux.com>","msg_from_op":false,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On 6/12/15 5:00 PM, Thom Brown wrote:\n> On 18 October 2014 at 15:36, Bruce Momjian <bruce@momjian.us> wrote:\n>> On Fri, Oct 17, 2014 at 02:36:55PM -0400, Bruce Momjian wrote:\n>>> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\n>>>> David G Johnston <david.g.johnston@gmail.com> writes:\n>>>>> The question is whether we explain the implications of not being WAL-logged\n>>>>> in an error message or simply state the fact and let the documentation\n>>>>> explain the hazards - basically just output:\n>>>>> \"hash indexes are not WAL-logged and their use is discouraged\"\n>>>>\n>>>> +1. The warning message is not the place to be trying to explain all the\n>>>> details.\n>>>\n>>> OK, updated patch attached.\n>>\n>> Patch applied.\n>\n> I only just noticed this item when I read the release notes. Should\n> we bother warning when used on an unlogged table?\n\nNot really; but I think the bigger question at this point is if we want \nto change it this late in the game.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting, Austin TX\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 22 Jun 2015 19:06:33 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On Tue, Jun 23, 2015 at 9:06 AM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n> On 6/12/15 5:00 PM, Thom Brown wrote:\n>>\n>> On 18 October 2014 at 15:36, Bruce Momjian <bruce@momjian.us> wrote:\n>>>\n>>> On Fri, Oct 17, 2014 at 02:36:55PM -0400, Bruce Momjian wrote:\n>>>>\n>>>> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\n>>>>>\n>>>>> David G Johnston <david.g.johnston@gmail.com> writes:\n>>>>>>\n>>>>>> The question is whether we explain the implications of not being\n>>>>>> WAL-logged\n>>>>>> in an error message or simply state the fact and let the documentation\n>>>>>> explain the hazards - basically just output:\n>>>>>> \"hash indexes are not WAL-logged and their use is discouraged\"\n>>>>>\n>>>>>\n>>>>> +1. The warning message is not the place to be trying to explain all\n>>>>> the\n>>>>> details.\n>>>>\n>>>>\n>>>> OK, updated patch attached.\n>>>\n>>>\n>>> Patch applied.\n>>\n>>\n>> I only just noticed this item when I read the release notes. Should\n>> we bother warning when used on an unlogged table?\n>\n>\n> Not really; but I think the bigger question at this point is if we want to\n> change it this late in the game.\n\nChanging it even during beta looks acceptable to me. I think that it\nis mainly a matter to have a patch (here is one), and someone to push\nit as everybody here seem to agree that for UNLOGGED tables this\nwarning has little sense.\n-- \nMichael\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Tue, 23 Jun 2015 09:46:00 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On Mon, Jun 22, 2015 at 8:46 PM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> On Tue, Jun 23, 2015 at 9:06 AM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n>> On 6/12/15 5:00 PM, Thom Brown wrote:\n>>>\n>>> On 18 October 2014 at 15:36, Bruce Momjian <bruce@momjian.us> wrote:\n>>>>\n>>>> On Fri, Oct 17, 2014 at 02:36:55PM -0400, Bruce Momjian wrote:\n>>>>>\n>>>>> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\n>>>>>>\n>>>>>> David G Johnston <david.g.johnston@gmail.com> writes:\n>>>>>>>\n>>>>>>> The question is whether we explain the implications of not being\n>>>>>>> WAL-logged\n>>>>>>> in an error message or simply state the fact and let the documentation\n>>>>>>> explain the hazards - basically just output:\n>>>>>>> \"hash indexes are not WAL-logged and their use is discouraged\"\n>>>>>>\n>>>>>>\n>>>>>> +1. The warning message is not the place to be trying to explain all\n>>>>>> the\n>>>>>> details.\n>>>>>\n>>>>>\n>>>>> OK, updated patch attached.\n>>>>\n>>>>\n>>>> Patch applied.\n>>>\n>>>\n>>> I only just noticed this item when I read the release notes. Should\n>>> we bother warning when used on an unlogged table?\n>>\n>>\n>> Not really; but I think the bigger question at this point is if we want to\n>> change it this late in the game.\n>\n> Changing it even during beta looks acceptable to me. I think that it\n> is mainly a matter to have a patch (here is one), and someone to push\n> it as everybody here seem to agree that for UNLOGGED tables this\n> warning has little sense.\n\nI think you should be testing RelationNeedsWAL(), not the\nrelpersistence directly. The same point applies for temporary\nindexes.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 23 Jun 2015 11:27:36 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On Wed, Jun 24, 2015 at 12:27 AM, Robert Haas wrote:\n> I think you should be testing RelationNeedsWAL(), not the\n> relpersistence directly. The same point applies for temporary\n> indexes.\n\nIndeed. Patch updated attached.\n-- \nMichael\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Wed, 24 Jun 2015 16:06:19 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On 18 October 2014 at 02:36, Bruce Momjian <bruce@momjian.us> wrote:\n> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\n>> David G Johnston <david.g.johnston@gmail.com> writes:\n>> > The question is whether we explain the implications of not being WAL-logged\n>> > in an error message or simply state the fact and let the documentation\n>> > explain the hazards - basically just output:\n>> > \"hash indexes are not WAL-logged and their use is discouraged\"\n>>\n>> +1. The warning message is not the place to be trying to explain all the\n>> details.\n\nWhile I don't think it should explain all the details, \"WAL-logged\"\nwill mean *nothing* to most users, including most of those who're\nusing streaming replication, PITR, etc.\n\nI would strongly prefer to see something that conveys some meaning to\na user who doesn't know PostgreSQL's innards, since by the time \"WAL\nlogged\" means much to you, you've got a good chance of having already\nlearned that hash indexes aren't crash-safe. Or of reading the manual.\n\nPerhaps:\n\nWARNING: hash indexes are not crash-safe, not replicated, and their\nuse is discouraged\n\n?\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 24 Jun 2015 16:45:52 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On Wed, Jun 24, 2015 at 1:45 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> WARNING: hash indexes are not crash-safe, not replicated, and their\n> use is discouraged\n\n+1\n\n\n-- \nPeter Geoghegan\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 24 Jun 2015 01:53:45 -0700","msg_from":"Peter Geoghegan <pg@heroku.com>","msg_from_op":false,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On Wed, Jun 24, 2015 at 3:06 AM, Michael Paquier\n<michael.paquier@gmail.com> wrote:\n> On Wed, Jun 24, 2015 at 12:27 AM, Robert Haas wrote:\n>> I think you should be testing RelationNeedsWAL(), not the\n>> relpersistence directly. The same point applies for temporary\n>> indexes.\n>\n> Indeed. Patch updated attached.\n\nCommitted.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 26 Jun 2015 11:38:56 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On Wed, Jun 24, 2015 at 4:53 AM, Peter Geoghegan <pg@heroku.com> wrote:\n> On Wed, Jun 24, 2015 at 1:45 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n>> WARNING: hash indexes are not crash-safe, not replicated, and their\n>> use is discouraged\n>\n> +1\n\nI'm not wild about this rewording; I think that if users don't know\nwhat WAL is, they probably need to know that in order to make good\ndecisions about whether to use hash indexes. But I don't feel\nsuper-strongly about it.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 26 Jun 2015 11:40:27 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Hash index creation warning"},{"msg_contents":"On Fri, Jun 26, 2015 at 11:40:27AM -0400, Robert Haas wrote:\n> On Wed, Jun 24, 2015 at 4:53 AM, Peter Geoghegan <pg@heroku.com> wrote:\n> > On Wed, Jun 24, 2015 at 1:45 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> >> WARNING: hash indexes are not crash-safe, not replicated, and their\n> >> use is discouraged\n> >\n> > +1\n> \n> I'm not wild about this rewording; I think that if users don't know\n> what WAL is, they probably need to know that in order to make good\n> decisions about whether to use hash indexes. But I don't feel\n> super-strongly about it.\n\nComing late to this, but I think Robert is right. WAL is used for crash\nrecovery, PITR, and streaming replication, and I am not sure we want to\nspecify all of those in the warning message.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 5 Aug 2015 23:35:15 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":true,"msg_subject":"Re: Hash index creation warning"}],"string":"[\n {\n \"msg_contents\": \"Now that we have the create hash index warning in 9.5, I realized that\\nwe don't warn about hash indexes with PITR, only crash recovery and\\nstreaming. This patch fixes that.\\n\\nIs the wording \\\"cannot be used\\\" too vague. The CREATE INDEX manual\\npage has the words \\\"give wrong answers to queries\\\", which might be\\nbetter, but is kind of long for an error message. Suggestions?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Fri, 17 Oct 2014 12:22:19 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Hash index creation warning\"\n },\n {\n \"msg_contents\": \"Bruce Momjian wrote\\n> Now that we have the create hash index warning in 9.5, I realized that\\n> we don't warn about hash indexes with PITR, only crash recovery and\\n> streaming. This patch fixes that.\\n> \\n> Is the wording \\\"cannot be used\\\" too vague. The CREATE INDEX manual\\n> page has the words \\\"give wrong answers to queries\\\", which might be\\n> better, but is kind of long for an error message. Suggestions?\\n\\nSomething like the following is more specific without being more wordy:\\n\\n\\\"hash indexes are not WAL-logged: they are corrupted during recovery and\\nchanges do not replicate to standby servers.\\\"\\n\\nThe question is whether we explain the implications of not being WAL-logged\\nin an error message or simply state the fact and let the documentation\\nexplain the hazards - basically just output:\\n\\n\\\"hash indexes are not WAL-logged and their use is discouraged\\\"\\n\\nDavid J.\\n\\n\\n\\n\\n--\\nView this message in context: http://postgresql.1045698.n5.nabble.com/Hash-index-creation-warning-tp5823443p5823445.html\\nSent from the PostgreSQL - hackers mailing list archive at Nabble.com.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 09:43:38 -0700 (PDT)\",\n \"msg_from\": \"David G Johnston <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"David G Johnston <david.g.johnston@gmail.com> writes:\\n> The question is whether we explain the implications of not being WAL-logged\\n> in an error message or simply state the fact and let the documentation\\n> explain the hazards - basically just output:\\n> \\\"hash indexes are not WAL-logged and their use is discouraged\\\"\\n\\n+1. The warning message is not the place to be trying to explain all the\\ndetails.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 12:56:52 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\\n> David G Johnston <david.g.johnston@gmail.com> writes:\\n> > The question is whether we explain the implications of not being WAL-logged\\n> > in an error message or simply state the fact and let the documentation\\n> > explain the hazards - basically just output:\\n> > \\\"hash indexes are not WAL-logged and their use is discouraged\\\"\\n> \\n> +1. The warning message is not the place to be trying to explain all the\\n> details.\\n\\nOK, updated patch attached.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Fri, 17 Oct 2014 14:36:55 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 02:36:55PM -0400, Bruce Momjian wrote:\\n> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\\n> > David G Johnston <david.g.johnston@gmail.com> writes:\\n> > > The question is whether we explain the implications of not being WAL-logged\\n> > > in an error message or simply state the fact and let the documentation\\n> > > explain the hazards - basically just output:\\n> > > \\\"hash indexes are not WAL-logged and their use is discouraged\\\"\\n> > \\n> > +1. The warning message is not the place to be trying to explain all the\\n> > details.\\n> \\n> OK, updated patch attached.\\n\\nPatch applied.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 10:36:24 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On 18 October 2014 at 15:36, Bruce Momjian <bruce@momjian.us> wrote:\\n> On Fri, Oct 17, 2014 at 02:36:55PM -0400, Bruce Momjian wrote:\\n>> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\\n>> > David G Johnston <david.g.johnston@gmail.com> writes:\\n>> > > The question is whether we explain the implications of not being WAL-logged\\n>> > > in an error message or simply state the fact and let the documentation\\n>> > > explain the hazards - basically just output:\\n>> > > \\\"hash indexes are not WAL-logged and their use is discouraged\\\"\\n>> >\\n>> > +1. The warning message is not the place to be trying to explain all the\\n>> > details.\\n>>\\n>> OK, updated patch attached.\\n>\\n> Patch applied.\\n\\nI only just noticed this item when I read the release notes. Should\\nwe bother warning when used on an unlogged table?\\n\\n-- \\nThom\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 12 Jun 2015 23:00:54 +0100\",\n \"msg_from\": \"Thom Brown <thom@linux.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On 6/12/15 5:00 PM, Thom Brown wrote:\\n> On 18 October 2014 at 15:36, Bruce Momjian <bruce@momjian.us> wrote:\\n>> On Fri, Oct 17, 2014 at 02:36:55PM -0400, Bruce Momjian wrote:\\n>>> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\\n>>>> David G Johnston <david.g.johnston@gmail.com> writes:\\n>>>>> The question is whether we explain the implications of not being WAL-logged\\n>>>>> in an error message or simply state the fact and let the documentation\\n>>>>> explain the hazards - basically just output:\\n>>>>> \\\"hash indexes are not WAL-logged and their use is discouraged\\\"\\n>>>>\\n>>>> +1. The warning message is not the place to be trying to explain all the\\n>>>> details.\\n>>>\\n>>> OK, updated patch attached.\\n>>\\n>> Patch applied.\\n>\\n> I only just noticed this item when I read the release notes. Should\\n> we bother warning when used on an unlogged table?\\n\\nNot really; but I think the bigger question at this point is if we want \\nto change it this late in the game.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting, Austin TX\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 22 Jun 2015 19:06:33 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On Tue, Jun 23, 2015 at 9:06 AM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n> On 6/12/15 5:00 PM, Thom Brown wrote:\\n>>\\n>> On 18 October 2014 at 15:36, Bruce Momjian <bruce@momjian.us> wrote:\\n>>>\\n>>> On Fri, Oct 17, 2014 at 02:36:55PM -0400, Bruce Momjian wrote:\\n>>>>\\n>>>> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\\n>>>>>\\n>>>>> David G Johnston <david.g.johnston@gmail.com> writes:\\n>>>>>>\\n>>>>>> The question is whether we explain the implications of not being\\n>>>>>> WAL-logged\\n>>>>>> in an error message or simply state the fact and let the documentation\\n>>>>>> explain the hazards - basically just output:\\n>>>>>> \\\"hash indexes are not WAL-logged and their use is discouraged\\\"\\n>>>>>\\n>>>>>\\n>>>>> +1. The warning message is not the place to be trying to explain all\\n>>>>> the\\n>>>>> details.\\n>>>>\\n>>>>\\n>>>> OK, updated patch attached.\\n>>>\\n>>>\\n>>> Patch applied.\\n>>\\n>>\\n>> I only just noticed this item when I read the release notes. Should\\n>> we bother warning when used on an unlogged table?\\n>\\n>\\n> Not really; but I think the bigger question at this point is if we want to\\n> change it this late in the game.\\n\\nChanging it even during beta looks acceptable to me. I think that it\\nis mainly a matter to have a patch (here is one), and someone to push\\nit as everybody here seem to agree that for UNLOGGED tables this\\nwarning has little sense.\\n-- \\nMichael\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Tue, 23 Jun 2015 09:46:00 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On Mon, Jun 22, 2015 at 8:46 PM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n> On Tue, Jun 23, 2015 at 9:06 AM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n>> On 6/12/15 5:00 PM, Thom Brown wrote:\\n>>>\\n>>> On 18 October 2014 at 15:36, Bruce Momjian <bruce@momjian.us> wrote:\\n>>>>\\n>>>> On Fri, Oct 17, 2014 at 02:36:55PM -0400, Bruce Momjian wrote:\\n>>>>>\\n>>>>> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\\n>>>>>>\\n>>>>>> David G Johnston <david.g.johnston@gmail.com> writes:\\n>>>>>>>\\n>>>>>>> The question is whether we explain the implications of not being\\n>>>>>>> WAL-logged\\n>>>>>>> in an error message or simply state the fact and let the documentation\\n>>>>>>> explain the hazards - basically just output:\\n>>>>>>> \\\"hash indexes are not WAL-logged and their use is discouraged\\\"\\n>>>>>>\\n>>>>>>\\n>>>>>> +1. The warning message is not the place to be trying to explain all\\n>>>>>> the\\n>>>>>> details.\\n>>>>>\\n>>>>>\\n>>>>> OK, updated patch attached.\\n>>>>\\n>>>>\\n>>>> Patch applied.\\n>>>\\n>>>\\n>>> I only just noticed this item when I read the release notes. Should\\n>>> we bother warning when used on an unlogged table?\\n>>\\n>>\\n>> Not really; but I think the bigger question at this point is if we want to\\n>> change it this late in the game.\\n>\\n> Changing it even during beta looks acceptable to me. I think that it\\n> is mainly a matter to have a patch (here is one), and someone to push\\n> it as everybody here seem to agree that for UNLOGGED tables this\\n> warning has little sense.\\n\\nI think you should be testing RelationNeedsWAL(), not the\\nrelpersistence directly. The same point applies for temporary\\nindexes.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 23 Jun 2015 11:27:36 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 24, 2015 at 12:27 AM, Robert Haas wrote:\\n> I think you should be testing RelationNeedsWAL(), not the\\n> relpersistence directly. The same point applies for temporary\\n> indexes.\\n\\nIndeed. Patch updated attached.\\n-- \\nMichael\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Wed, 24 Jun 2015 16:06:19 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On 18 October 2014 at 02:36, Bruce Momjian <bruce@momjian.us> wrote:\\n> On Fri, Oct 17, 2014 at 12:56:52PM -0400, Tom Lane wrote:\\n>> David G Johnston <david.g.johnston@gmail.com> writes:\\n>> > The question is whether we explain the implications of not being WAL-logged\\n>> > in an error message or simply state the fact and let the documentation\\n>> > explain the hazards - basically just output:\\n>> > \\\"hash indexes are not WAL-logged and their use is discouraged\\\"\\n>>\\n>> +1. The warning message is not the place to be trying to explain all the\\n>> details.\\n\\nWhile I don't think it should explain all the details, \\\"WAL-logged\\\"\\nwill mean *nothing* to most users, including most of those who're\\nusing streaming replication, PITR, etc.\\n\\nI would strongly prefer to see something that conveys some meaning to\\na user who doesn't know PostgreSQL's innards, since by the time \\\"WAL\\nlogged\\\" means much to you, you've got a good chance of having already\\nlearned that hash indexes aren't crash-safe. Or of reading the manual.\\n\\nPerhaps:\\n\\nWARNING: hash indexes are not crash-safe, not replicated, and their\\nuse is discouraged\\n\\n?\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 24 Jun 2015 16:45:52 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 24, 2015 at 1:45 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> WARNING: hash indexes are not crash-safe, not replicated, and their\\n> use is discouraged\\n\\n+1\\n\\n\\n-- \\nPeter Geoghegan\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 24 Jun 2015 01:53:45 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@heroku.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 24, 2015 at 3:06 AM, Michael Paquier\\n<michael.paquier@gmail.com> wrote:\\n> On Wed, Jun 24, 2015 at 12:27 AM, Robert Haas wrote:\\n>> I think you should be testing RelationNeedsWAL(), not the\\n>> relpersistence directly. The same point applies for temporary\\n>> indexes.\\n>\\n> Indeed. Patch updated attached.\\n\\nCommitted.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 26 Jun 2015 11:38:56 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 24, 2015 at 4:53 AM, Peter Geoghegan <pg@heroku.com> wrote:\\n> On Wed, Jun 24, 2015 at 1:45 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n>> WARNING: hash indexes are not crash-safe, not replicated, and their\\n>> use is discouraged\\n>\\n> +1\\n\\nI'm not wild about this rewording; I think that if users don't know\\nwhat WAL is, they probably need to know that in order to make good\\ndecisions about whether to use hash indexes. But I don't feel\\nsuper-strongly about it.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 26 Jun 2015 11:40:27 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Hash index creation warning\"\n },\n {\n \"msg_contents\": \"On Fri, Jun 26, 2015 at 11:40:27AM -0400, Robert Haas wrote:\\n> On Wed, Jun 24, 2015 at 4:53 AM, Peter Geoghegan <pg@heroku.com> wrote:\\n> > On Wed, Jun 24, 2015 at 1:45 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\\n> >> WARNING: hash indexes are not crash-safe, not replicated, and their\\n> >> use is discouraged\\n> >\\n> > +1\\n> \\n> I'm not wild about this rewording; I think that if users don't know\\n> what WAL is, they probably need to know that in order to make good\\n> decisions about whether to use hash indexes. But I don't feel\\n> super-strongly about it.\\n\\nComing late to this, but I think Robert is right. WAL is used for crash\\nrecovery, PITR, and streaming replication, and I am not sure we want to\\nspecify all of those in the warning message.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 5 Aug 2015 23:35:15 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Hash index creation warning\"\n }\n]","isConversational":false}}},{"rowIdx":39679,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Following up something Pavel wrote, I notice that json_agg() and \njson_object_agg() are both marked as immutable, even though they invoke \nIO functions, while json_object is marked stable, even though it does \nnot, and can probably be marked as immutable. Mea maxima culpa.\n\nI'm not sure what we should do about these things now. Is it a tragedy \nif we let these escape into the 9.4 release that way?\n\ncheers\n\nandrew\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 16:03:10 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":true,"msg_subject":"json function volatility"},{"msg_contents":"On Fri, Oct 17, 2014 at 3:03 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\n> Following up something Pavel wrote, I notice that json_agg() and\n> json_object_agg() are both marked as immutable, even though they invoke IO\n> functions, while json_object is marked stable, even though it does not, and\n> can probably be marked as immutable. Mea maxima culpa.\n>\n> I'm not sure what we should do about these things now. Is it a tragedy if we\n> let these escape into the 9.4 release that way?\n\nIs it too late to change them? Either way, it seems fairly\nimplausible someone would come up with a case to stick json_agg(), or\nany aggregate function really, inside of an index. So it's not exactly\nthe crime of the century.\n\nmerlin\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 15:44:09 -0500","msg_from":"Merlin Moncure <mmoncure@gmail.com>","msg_from_op":false,"msg_subject":"Re: json function volatility"},{"msg_contents":"On Fri, Oct 17, 2014 at 1:44 PM, Merlin Moncure <mmoncure@gmail.com> wrote:\n> Is it too late to change them? Either way, it seems fairly\n> implausible someone would come up with a case to stick json_agg(), or\n> any aggregate function really, inside of an index. So it's not exactly\n> the crime of the century.\n\nIndexes reject aggregates outright as acceptable expressions for\nindexing. That's not the only issue, of course.\n\n-- \nPeter Geoghegan\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 13:52:40 -0700","msg_from":"Peter Geoghegan <pg@heroku.com>","msg_from_op":false,"msg_subject":"Re: json function volatility"},{"msg_contents":"Merlin Moncure wrote:\n> On Fri, Oct 17, 2014 at 3:03 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\n> > Following up something Pavel wrote, I notice that json_agg() and\n> > json_object_agg() are both marked as immutable, even though they invoke IO\n> > functions, while json_object is marked stable, even though it does not, and\n> > can probably be marked as immutable. Mea maxima culpa.\n> >\n> > I'm not sure what we should do about these things now. Is it a tragedy if we\n> > let these escape into the 9.4 release that way?\n> \n> Is it too late to change them?\n\nOne thing to consider is the catversion bump, which we don't want this\nlate in the cycle. Still, you could change the catalogs but not the\nversion, and advise those with the older definitions to tweak the\ncatalogs by hand if they need it. I think we did this once.\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 17:59:49 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: json function volatility"},{"msg_contents":"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n> Merlin Moncure wrote:\n>> On Fri, Oct 17, 2014 at 3:03 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>> Following up something Pavel wrote, I notice that json_agg() and\n>>> json_object_agg() are both marked as immutable, even though they invoke IO\n>>> functions, while json_object is marked stable, even though it does not, and\n>>> can probably be marked as immutable. Mea maxima culpa.\n\n>>> I'm not sure what we should do about these things now. Is it a tragedy if we\n>>> let these escape into the 9.4 release that way?\n\n>> Is it too late to change them?\n\n> One thing to consider is the catversion bump, which we don't want this\n> late in the cycle. Still, you could change the catalogs but not the\n> version, and advise those with the older definitions to tweak the\n> catalogs by hand if they need it. I think we did this once.\n\nI'm fairly sure that the system doesn't actually pay attention to the\nvolatility marking of aggregates, so there's no huge harm done by the\nincorrect markings of those. The incorrect marking of json_object might\nbe a small optimization block.\n\n+1 for changing these in 9.4 without bumping catversion. I don't think\nwe need to give advice for manual corrections, either: the risk of doing\nthat wrong probably outweighs the value of fixing it.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 17:41:09 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: json function volatility"}],"string":"[\n {\n \"msg_contents\": \"Following up something Pavel wrote, I notice that json_agg() and \\njson_object_agg() are both marked as immutable, even though they invoke \\nIO functions, while json_object is marked stable, even though it does \\nnot, and can probably be marked as immutable. Mea maxima culpa.\\n\\nI'm not sure what we should do about these things now. Is it a tragedy \\nif we let these escape into the 9.4 release that way?\\n\\ncheers\\n\\nandrew\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 16:03:10 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"json function volatility\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 3:03 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> Following up something Pavel wrote, I notice that json_agg() and\\n> json_object_agg() are both marked as immutable, even though they invoke IO\\n> functions, while json_object is marked stable, even though it does not, and\\n> can probably be marked as immutable. Mea maxima culpa.\\n>\\n> I'm not sure what we should do about these things now. Is it a tragedy if we\\n> let these escape into the 9.4 release that way?\\n\\nIs it too late to change them? Either way, it seems fairly\\nimplausible someone would come up with a case to stick json_agg(), or\\nany aggregate function really, inside of an index. So it's not exactly\\nthe crime of the century.\\n\\nmerlin\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 15:44:09 -0500\",\n \"msg_from\": \"Merlin Moncure <mmoncure@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: json function volatility\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 1:44 PM, Merlin Moncure <mmoncure@gmail.com> wrote:\\n> Is it too late to change them? Either way, it seems fairly\\n> implausible someone would come up with a case to stick json_agg(), or\\n> any aggregate function really, inside of an index. So it's not exactly\\n> the crime of the century.\\n\\nIndexes reject aggregates outright as acceptable expressions for\\nindexing. That's not the only issue, of course.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 13:52:40 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@heroku.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: json function volatility\"\n },\n {\n \"msg_contents\": \"Merlin Moncure wrote:\\n> On Fri, Oct 17, 2014 at 3:03 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n> > Following up something Pavel wrote, I notice that json_agg() and\\n> > json_object_agg() are both marked as immutable, even though they invoke IO\\n> > functions, while json_object is marked stable, even though it does not, and\\n> > can probably be marked as immutable. Mea maxima culpa.\\n> >\\n> > I'm not sure what we should do about these things now. Is it a tragedy if we\\n> > let these escape into the 9.4 release that way?\\n> \\n> Is it too late to change them?\\n\\nOne thing to consider is the catversion bump, which we don't want this\\nlate in the cycle. Still, you could change the catalogs but not the\\nversion, and advise those with the older definitions to tweak the\\ncatalogs by hand if they need it. I think we did this once.\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 17:59:49 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: json function volatility\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> Merlin Moncure wrote:\\n>> On Fri, Oct 17, 2014 at 3:03 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>> Following up something Pavel wrote, I notice that json_agg() and\\n>>> json_object_agg() are both marked as immutable, even though they invoke IO\\n>>> functions, while json_object is marked stable, even though it does not, and\\n>>> can probably be marked as immutable. Mea maxima culpa.\\n\\n>>> I'm not sure what we should do about these things now. Is it a tragedy if we\\n>>> let these escape into the 9.4 release that way?\\n\\n>> Is it too late to change them?\\n\\n> One thing to consider is the catversion bump, which we don't want this\\n> late in the cycle. Still, you could change the catalogs but not the\\n> version, and advise those with the older definitions to tweak the\\n> catalogs by hand if they need it. I think we did this once.\\n\\nI'm fairly sure that the system doesn't actually pay attention to the\\nvolatility marking of aggregates, so there's no huge harm done by the\\nincorrect markings of those. The incorrect marking of json_object might\\nbe a small optimization block.\\n\\n+1 for changing these in 9.4 without bumping catversion. I don't think\\nwe need to give advice for manual corrections, either: the risk of doing\\nthat wrong probably outweighs the value of fixing it.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 17:41:09 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: json function volatility\"\n }\n]","isConversational":false}}},{"rowIdx":39680,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nThis week we had one of the most annoying problems I've ever encountered \nwith postgres. We had a big index on multiple columns, say, foo(a, b, \nc). According to pg_stat_all_indexes the index was being used *all the \ntime*. However, after looking into our queries more closely, it turns \nout that it was only being used to look up statistics for the foo.a \ncolumn to estimate merge scan viability during planning. But this took \nhours for two people to track down.\n\nSo what I'd like to have is a way to be able to distinguish between \nindexes being used to answer queries, and ones being only used for stats \nlookups during planning. Perhaps the easiest way would be adding a new \ncolumn or two into pg_stat_all_indexes, which we would increment in \nget_actual_variable_range() when fetching data.\n\nAny thoughts?\n\n\n.marko\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 23:43:17 +0200","msg_from":"Marko Tiikkaja <marko@joh.to>","msg_from_op":true,"msg_subject":"get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"Marko Tiikkaja <marko@joh.to> writes:\n> This week we had one of the most annoying problems I've ever encountered \n> with postgres. We had a big index on multiple columns, say, foo(a, b, \n> c). According to pg_stat_all_indexes the index was being used *all the \n> time*. However, after looking into our queries more closely, it turns \n> out that it was only being used to look up statistics for the foo.a \n> column to estimate merge scan viability during planning. But this took \n> hours for two people to track down.\n\n> So what I'd like to have is a way to be able to distinguish between \n> indexes being used to answer queries, and ones being only used for stats \n> lookups during planning.\n\nWhy? Used is used.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 17:47:59 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On 10/17/14, 11:47 PM, Tom Lane wrote:\n> Marko Tiikkaja <marko@joh.to> writes:\n>> This week we had one of the most annoying problems I've ever encountered\n>> with postgres. We had a big index on multiple columns, say, foo(a, b,\n>> c). According to pg_stat_all_indexes the index was being used *all the\n>> time*. However, after looking into our queries more closely, it turns\n>> out that it was only being used to look up statistics for the foo.a\n>> column to estimate merge scan viability during planning. But this took\n>> hours for two people to track down.\n>\n>> So what I'd like to have is a way to be able to distinguish between\n>> indexes being used to answer queries, and ones being only used for stats\n>> lookups during planning.\n>\n> Why? Used is used.\n\nBecause I don't need a 30GB index on foo(a,b,c) to look up statistics. \nIf I ever have a problem, I can replace it with a 5GB one on foo(a).\n\n\n.marko\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 23:49:05 +0200","msg_from":"Marko Tiikkaja <marko@joh.to>","msg_from_op":true,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On 10/17/14, 4:49 PM, Marko Tiikkaja wrote:\n> On 10/17/14, 11:47 PM, Tom Lane wrote:\n>> Marko Tiikkaja <marko@joh.to> writes:\n>>> This week we had one of the most annoying problems I've ever encountered\n>>> with postgres. We had a big index on multiple columns, say, foo(a, b,\n>>> c). According to pg_stat_all_indexes the index was being used *all the\n>>> time*. However, after looking into our queries more closely, it turns\n>>> out that it was only being used to look up statistics for the foo.a\n>>> column to estimate merge scan viability during planning. But this took\n>>> hours for two people to track down.\n>>\n>>> So what I'd like to have is a way to be able to distinguish between\n>>> indexes being used to answer queries, and ones being only used for stats\n>>> lookups during planning.\n>>\n>> Why? Used is used.\n>\n> Because I don't need a 30GB index on foo(a,b,c) to look up statistics. If I ever have a problem, I can replace it with a 5GB one on foo(a).\n\nThat problem can exist with user queries too. Perhaps it would be better to find a way to count scans that didn't use all the fields in the index.\n\nI do also see value in differentiating planning use from real query processing; not doing that can certainly cause confusion. What I don't know is if the added stats bloat is worth it. If we do go down that road, I think it'd be better to add an indicator to EState. Aside from allowing stats for all planning access, it should make it less likely that someone adds a new access path and forgets to mark it as internal (especially if the added field defaults to an invalid value).\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 16:59:10 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"Marko Tiikkaja <marko@joh.to> writes:\n> On 10/17/14, 11:47 PM, Tom Lane wrote:\n>> Marko Tiikkaja <marko@joh.to> writes:\n>>> So what I'd like to have is a way to be able to distinguish between\n>>> indexes being used to answer queries, and ones being only used for stats\n>>> lookups during planning.\n\n>> Why? Used is used.\n\n> Because I don't need a 30GB index on foo(a,b,c) to look up statistics. \n> If I ever have a problem, I can replace it with a 5GB one on foo(a).\n\nWell, the index might've been getting used in queries too in a way that\nreally only involved the first column. I think you're solving the wrong\nproblem here. The right problem is how to identify indexes that are\nbeing used in a way that doesn't exploit all the columns. Which is not\nnecessarily wrong in itself --- what you'd want is to figure out when the\nlast column(s) are *never* used. The existing stats aren't terribly\nhelpful for that, I agree.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 17:59:13 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On 10/17/14, 11:59 PM, Tom Lane wrote:\n> Marko Tiikkaja <marko@joh.to> writes:\n>> On 10/17/14, 11:47 PM, Tom Lane wrote:\n>>> Marko Tiikkaja <marko@joh.to> writes:\n>>>> So what I'd like to have is a way to be able to distinguish between\n>>>> indexes being used to answer queries, and ones being only used for stats\n>>>> lookups during planning.\n>\n>>> Why? Used is used.\n>\n>> Because I don't need a 30GB index on foo(a,b,c) to look up statistics.\n>> If I ever have a problem, I can replace it with a 5GB one on foo(a).\n>\n> Well, the index might've been getting used in queries too in a way that\n> really only involved the first column. I think you're solving the wrong\n> problem here. The right problem is how to identify indexes that are\n> being used in a way that doesn't exploit all the columns.\n\nI'm not sure I agree with that. Even if there was some information the \nplanner could have extracted out of the index by using all columns (thus \nappearing \"fully used\" in these hypothetical new statistics), I still \nwould've wanted the index gone. But in this particular case, an index \non foo(a) alone was not selective enough and it would have been a bad \nchoice for practically every query, so I'm not sure what good those \nstatistics were in the first place.\n\nI think there's a big difference between \"this index was used to look up \nstuff for planning\" and \"this index was used to answer queries quickly\". \n In my mind the first one belongs to the category \"this index was \nconsidered\", and the latter is \"this index was actually useful\". But \nmaybe I'm not seeing the big picture here.\n\n\n.marko\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 00:09:03 +0200","msg_from":"Marko Tiikkaja <marko@joh.to>","msg_from_op":true,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"Marko Tiikkaja <marko@joh.to> writes:\n> On 10/17/14, 11:59 PM, Tom Lane wrote:\n>> Well, the index might've been getting used in queries too in a way that\n>> really only involved the first column. I think you're solving the wrong\n>> problem here. The right problem is how to identify indexes that are\n>> being used in a way that doesn't exploit all the columns.\n\n> I'm not sure I agree with that. Even if there was some information the \n> planner could have extracted out of the index by using all columns (thus \n> appearing \"fully used\" in these hypothetical new statistics), I still \n> would've wanted the index gone. But in this particular case, an index \n> on foo(a) alone was not selective enough and it would have been a bad \n> choice for practically every query, so I'm not sure what good those \n> statistics were in the first place.\n\nThose stats were perfectly valid: what the planner is looking for is\naccurate minimum and maximum values for the index's leading column, and\nthat's what it got. You're correct that a narrower index could have given\nthe same results with a smaller disk footprint, but the planner got the\nresults it needed from the index you provided for it to work with.\n\n> I think there's a big difference between \"this index was used to look up \n> stuff for planning\" and \"this index was used to answer queries quickly\". \n\nI think that's utter nonsense. Even if there were any validity to the\nposition, it wouldn't be enough to justify doubling the stats footprint\nin order to track system-driven accesses separately from query-driven\nones.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 18:15:37 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On 10/18/14, 12:15 AM, Tom Lane wrote:\n> Marko Tiikkaja <marko@joh.to> writes:\n>> I think there's a big difference between \"this index was used to look up\n>> stuff for planning\" and \"this index was used to answer queries quickly\".\n>\n> I think that's utter nonsense.\n\nWell you probably know a bit more about the optimizer than I do. But I \ncan't see a case where the stats provided by the index would be useful \nfor choosing between two (or more) plans that don't use the index in the \nactual query. If you're saying that there are such cases, then clearly \nI don't know something, and my thinking is in the wrong here.\n\n\n.marko\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 00:49:54 +0200","msg_from":"Marko Tiikkaja <marko@joh.to>","msg_from_op":true,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On Fri, Oct 17, 2014 at 06:15:37PM -0400, Tom Lane wrote:\n> Marko Tiikkaja <marko@joh.to> writes:\n> > On 10/17/14, 11:59 PM, Tom Lane wrote:\n> >> Well, the index might've been getting used in queries too in a way that\n> >> really only involved the first column. I think you're solving the wrong\n> >> problem here. The right problem is how to identify indexes that are\n> >> being used in a way that doesn't exploit all the columns.\n> \n> > I'm not sure I agree with that. Even if there was some information the \n> > planner could have extracted out of the index by using all columns (thus \n> > appearing \"fully used\" in these hypothetical new statistics), I still \n> > would've wanted the index gone. But in this particular case, an index \n> > on foo(a) alone was not selective enough and it would have been a bad \n> > choice for practically every query, so I'm not sure what good those \n> > statistics were in the first place.\n> \n> Those stats were perfectly valid: what the planner is looking for is\n> accurate minimum and maximum values for the index's leading column, and\n> that's what it got. You're correct that a narrower index could have given\n> the same results with a smaller disk footprint, but the planner got the\n> results it needed from the index you provided for it to work with.\n\nUh, why is the optimizer looking at the index on a,b,c and not just the\nstats on column a, for example? I am missing something here.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 22:30:01 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"Bruce Momjian <bruce@momjian.us> writes:\n> On Fri, Oct 17, 2014 at 06:15:37PM -0400, Tom Lane wrote:\n>> Those stats were perfectly valid: what the planner is looking for is\n>> accurate minimum and maximum values for the index's leading column, and\n>> that's what it got. You're correct that a narrower index could have given\n>> the same results with a smaller disk footprint, but the planner got the\n>> results it needed from the index you provided for it to work with.\n\n> Uh, why is the optimizer looking at the index on a,b,c and not just the\n> stats on column a, for example? I am missing something here.\n\nBecause it needs up-to-date min/max values in order to avoid being\nseriously misled about selectivities of values near the endpoints.\nSee commit 40608e7f949fb7e4025c0ddd5be01939adc79eec.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 23:03:04 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"I wrote:\n> Because it needs up-to-date min/max values in order to avoid being\n> seriously misled about selectivities of values near the endpoints.\n> See commit 40608e7f949fb7e4025c0ddd5be01939adc79eec.\n\nBTW, on re-reading that code I notice that it will happily seize upon\nthe first suitable index (\"first\" in OID order), regardless of how many\nlower-order columns that index has got. This doesn't make any difference\nI think for get_actual_variable_range's own purposes, because it's only\nexpecting to touch the endmost index page regardless. However, in light\nof Marko's complaint maybe we should teach it to check all the indexes\nand prefer the matching one with fewest columns? It would only take a\ncouple extra lines of code, and probably not that many added cycles\nconsidering we're going to do an index access of some sort. But I'm\nnot sure if it's worth any extra effort --- I think in his example\ncase, there wasn't any narrower index anyway.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 17 Oct 2014 23:16:08 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On Fri, Oct 17, 2014 at 11:03:04PM -0400, Tom Lane wrote:\n> Bruce Momjian <bruce@momjian.us> writes:\n> > On Fri, Oct 17, 2014 at 06:15:37PM -0400, Tom Lane wrote:\n> >> Those stats were perfectly valid: what the planner is looking for is\n> >> accurate minimum and maximum values for the index's leading column, and\n> >> that's what it got. You're correct that a narrower index could have given\n> >> the same results with a smaller disk footprint, but the planner got the\n> >> results it needed from the index you provided for it to work with.\n> \n> > Uh, why is the optimizer looking at the index on a,b,c and not just the\n> > stats on column a, for example? I am missing something here.\n> \n> Because it needs up-to-date min/max values in order to avoid being\n> seriously misled about selectivities of values near the endpoints.\n> See commit 40608e7f949fb7e4025c0ddd5be01939adc79eec.\n\nOh, I had forgotten we did that. It is confusing that there is no way\nvia EXPLAIN to see the access, making the method of consulting pg_stat_*\nand using EXPLAIN unreliable. Should we document this somewhere?\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 09:58:10 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On 10/18/14, 5:16 AM, Tom Lane wrote:\n> BTW, on re-reading that code I notice that it will happily seize upon\n> the first suitable index (\"first\" in OID order), regardless of how many\n> lower-order columns that index has got. This doesn't make any difference\n> I think for get_actual_variable_range's own purposes, because it's only\n> expecting to touch the endmost index page regardless. However, in light\n> of Marko's complaint maybe we should teach it to check all the indexes\n> and prefer the matching one with fewest columns? It would only take a\n> couple extra lines of code, and probably not that many added cycles\n> considering we're going to do an index access of some sort. But I'm\n> not sure if it's worth any extra effort --- I think in his example\n> case, there wasn't any narrower index anyway.\n\nPerhaps accidentally this would have helped in my case, actually, since \nI could have created a new, smaller index CONCURRENTLY and then seen \nthat the usage of the other index stopped increasing. With the \"pick \nthe smallest OID\" behaviour that was not possible. Another idea had was \nsome way to tell the optimizer not to use that particular index for \nstats lookups, but probably the use case for such a feature would be a \nbit narrow.\n\nAll that said, I don't think my struggles justify the change you \ndescribed above. Not sure if it's a good idea or not.\n\n\n.marko\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 16:29:41 +0200","msg_from":"Marko Tiikkaja <marko@joh.to>","msg_from_op":true,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On Sat, Oct 18, 2014 at 04:29:41PM +0200, Marko Tiikkaja wrote:\n> On 10/18/14, 5:16 AM, Tom Lane wrote:\n> >BTW, on re-reading that code I notice that it will happily seize upon\n> >the first suitable index (\"first\" in OID order), regardless of how many\n> >lower-order columns that index has got. This doesn't make any difference\n> >I think for get_actual_variable_range's own purposes, because it's only\n> >expecting to touch the endmost index page regardless. However, in light\n> >of Marko's complaint maybe we should teach it to check all the indexes\n> >and prefer the matching one with fewest columns? It would only take a\n> >couple extra lines of code, and probably not that many added cycles\n> >considering we're going to do an index access of some sort. But I'm\n> >not sure if it's worth any extra effort --- I think in his example\n> >case, there wasn't any narrower index anyway.\n> \n> Perhaps accidentally this would have helped in my case, actually,\n> since I could have created a new, smaller index CONCURRENTLY and\n> then seen that the usage of the other index stopped increasing.\n> With the \"pick the smallest OID\" behaviour that was not possible.\n> Another idea had was some way to tell the optimizer not to use that\n> particular index for stats lookups, but probably the use case for\n> such a feature would be a bit narrow.\n\nWell, if the index is there, why not use it? I thought the problem was\njust that you had no visibility into how those statistics were being\naccessed. Most people think EXPLAIN shows all accesses, but obviously\nnow it doesn't.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 10:33:28 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On 10/18/14, 4:33 PM, Bruce Momjian wrote:\n> On Sat, Oct 18, 2014 at 04:29:41PM +0200, Marko Tiikkaja wrote:\n>> Another idea had was some way to tell the optimizer not to use that\n>> particular index for stats lookups, but probably the use case for\n>> such a feature would be a bit narrow.\n>\n> Well, if the index is there, why not use it? I thought the problem was\n> just that you had no visibility into how those statistics were being\n> accessed.\n\nYes, exactly; if I had had the option to disable the index from the \noptimizer's point of view, I'd have seen that it's not used for looking \nup any data by any queries, and thus I would have known that I can \nsafely drop it without slowing down queries. Which was the only thing I \ncared about, and where the stats we provide failed me.\n\n\n.marko\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 16:38:37 +0200","msg_from":"Marko Tiikkaja <marko@joh.to>","msg_from_op":true,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On Sat, Oct 18, 2014 at 04:38:37PM +0200, Marko Tiikkaja wrote:\n> On 10/18/14, 4:33 PM, Bruce Momjian wrote:\n> >On Sat, Oct 18, 2014 at 04:29:41PM +0200, Marko Tiikkaja wrote:\n> >>Another idea had was some way to tell the optimizer not to use that\n> >>particular index for stats lookups, but probably the use case for\n> >>such a feature would be a bit narrow.\n> >\n> >Well, if the index is there, why not use it? I thought the problem was\n> >just that you had no visibility into how those statistics were being\n> >accessed.\n> \n> Yes, exactly; if I had had the option to disable the index from the\n> optimizer's point of view, I'd have seen that it's not used for\n> looking up any data by any queries, and thus I would have known that\n> I can safely drop it without slowing down queries. Which was the\n> only thing I cared about, and where the stats we provide failed me.\n\nHow many other cases do we have where the statistics are getting\nincremented and there is no user visibility into the operation?\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 10:46:42 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"Bruce Momjian <bruce@momjian.us> writes:\n> How many other cases do we have where the statistics are getting\n> incremented and there is no user visibility into the operation?\n\n* system catalog accesses\n* vacuum/analyze/cluster/etc\n\nThe fact that system-initiated accesses get counted in the statistics\nis a feature, not a bug.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 11:05:48 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"Marko Tiikkaja <marko@joh.to> writes:\n> On 10/18/14, 4:33 PM, Bruce Momjian wrote:\n>> Well, if the index is there, why not use it? I thought the problem was\n>> just that you had no visibility into how those statistics were being\n>> accessed.\n\n> Yes, exactly; if I had had the option to disable the index from the \n> optimizer's point of view, I'd have seen that it's not used for looking \n> up any data by any queries, and thus I would have known that I can \n> safely drop it without slowing down queries. Which was the only thing I \n> cared about, and where the stats we provide failed me.\n\nThis argument is *utterly* wrongheaded, because it assumes that the\nplanner's use of the index provided no benefit to your queries. If the\nplanner was touching the index at all then it was planning queries in\nwhich knowledge of the extremal value was relevant to accurate selectivity\nestimation. So it's quite likely that without the index you'd have gotten\ndifferent and inferior plans, whether or not those plans actually chose to\nuse the index.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 11:46:47 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On 10/18/14, 5:46 PM, Tom Lane wrote:\n> Marko Tiikkaja <marko@joh.to> writes:\n>> Yes, exactly; if I had had the option to disable the index from the\n>> optimizer's point of view, I'd have seen that it's not used for looking\n>> up any data by any queries, and thus I would have known that I can\n>> safely drop it without slowing down queries. Which was the only thing I\n>> cared about, and where the stats we provide failed me.\n>\n> This argument is *utterly* wrongheaded, because it assumes that the\n> planner's use of the index provided no benefit to your queries. If the\n> planner was touching the index at all then it was planning queries in\n> which knowledge of the extremal value was relevant to accurate selectivity\n> estimation. So it's quite likely that without the index you'd have gotten\n> different and inferior plans, whether or not those plans actually chose to\n> use the index.\n\nMaybe. But at the same time that's a big problem: there's no way of \nknowing whether the index is actually useful or not when it's used only \nby the query planner.\n\n\n.marko\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 18:15:03 +0200","msg_from":"Marko Tiikkaja <marko@joh.to>","msg_from_op":true,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On Sat, Oct 18, 2014 at 06:15:03PM +0200, Marko Tiikkaja wrote:\n> On 10/18/14, 5:46 PM, Tom Lane wrote:\n> >Marko Tiikkaja <marko@joh.to> writes:\n> >>Yes, exactly; if I had had the option to disable the index from the\n> >>optimizer's point of view, I'd have seen that it's not used for looking\n> >>up any data by any queries, and thus I would have known that I can\n> >>safely drop it without slowing down queries. Which was the only thing I\n> >>cared about, and where the stats we provide failed me.\n> >\n> >This argument is *utterly* wrongheaded, because it assumes that the\n> >planner's use of the index provided no benefit to your queries. If the\n> >planner was touching the index at all then it was planning queries in\n> >which knowledge of the extremal value was relevant to accurate selectivity\n> >estimation. So it's quite likely that without the index you'd have gotten\n> >different and inferior plans, whether or not those plans actually chose to\n> >use the index.\n> \n> Maybe. But at the same time that's a big problem: there's no way of\n> knowing whether the index is actually useful or not when it's used\n> only by the query planner.\n\nThat is a good point. Without an index, the executor is going to do a\nsequential scan, while a missing index to the optimizer just means worse\nstatistics.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 14:20:45 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On 10/17/14, 10:16 PM, Tom Lane wrote:\n> I wrote:\n>> Because it needs up-to-date min/max values in order to avoid being\n>> seriously misled about selectivities of values near the endpoints.\n>> See commit 40608e7f949fb7e4025c0ddd5be01939adc79eec.\n>\n> BTW, on re-reading that code I notice that it will happily seize upon\n> the first suitable index (\"first\" in OID order), regardless of how many\n> lower-order columns that index has got. This doesn't make any difference\n> I think for get_actual_variable_range's own purposes, because it's only\n> expecting to touch the endmost index page regardless. However, in light\n> of Marko's complaint maybe we should teach it to check all the indexes\n> and prefer the matching one with fewest columns? It would only take a\n> couple extra lines of code, and probably not that many added cycles\n> considering we're going to do an index access of some sort. But I'm\n> not sure if it's worth any extra effort --- I think in his example\n> case, there wasn't any narrower index anyway.\n\nThe real cost here isn't the number of columns, it's the size of the index, no? So shouldn't we look at relpages instead? For example, you'd certainly want to use an index on (field_we_care_about, smallint_field) over an index on (field_we_care_about, big_honking_text_field).\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 20:54:29 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\n> On 10/17/14, 10:16 PM, Tom Lane wrote:\n>> BTW, on re-reading that code I notice that it will happily seize upon\n>> the first suitable index (\"first\" in OID order), regardless of how many\n>> lower-order columns that index has got. This doesn't make any difference\n>> I think for get_actual_variable_range's own purposes, because it's only\n>> expecting to touch the endmost index page regardless. However, in light\n>> of Marko's complaint maybe we should teach it to check all the indexes\n>> and prefer the matching one with fewest columns?\n\n> The real cost here isn't the number of columns, it's the size of the index, no? So shouldn't we look at relpages instead? For example, you'd certainly want to use an index on (field_we_care_about, smallint_field) over an index on (field_we_care_about, big_honking_text_field).\n\nYeah, perhaps. I'd been wondering about adding a tie-breaking rule,\nbut that's a much simpler way to think about it. OTOH, that approach\ncould result in some instability in the choice of index: if you've got\nboth (field_we_care_about, some_int_field) and (field_we_care_about,\nsome_other_int_field) then it might switch between choosing those two\nindexes from day to day depending on basically-chance issues like when\npage splits occur. That would probably annoy Marko even more than the\ncurrent behavior :-(, because it would scatter the planner's usage\nacross multiple indexes for no very good reason.\n\nThe coding I'd been imagining at first would basically break ties in\ncolumn count according to index OID order, so its choices would be stable\nas long as you did not add/drop indexes. That seems like a good property\nto try to preserve.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 19 Oct 2014 15:09:57 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On 10/19/14, 2:09 PM, Tom Lane wrote:\n> Yeah, perhaps. I'd been wondering about adding a tie-breaking rule,\n> but that's a much simpler way to think about it. OTOH, that approach\n> could result in some instability in the choice of index: if you've got\n> both (field_we_care_about, some_int_field) and (field_we_care_about,\n> some_other_int_field) then it might switch between choosing those two\n> indexes from day to day depending on basically-chance issues like when\n> page splits occur. That would probably annoy Marko even more than the\n> current behavior:-(, because it would scatter the planner's usage\n> across multiple indexes for no very good reason.\n>\n> The coding I'd been imagining at first would basically break ties in\n> column count according to index OID order, so its choices would be stable\n> as long as you did not add/drop indexes. That seems like a good property\n> to try to preserve.\n\nMaybe a good alternative is:\n\nORDER BY int( table.reltuples / index.relpages / BLKSZ ) DESC, oid\n\nBy comparing on average tuple size throwing away the fraction presumably we'd throw away noise from page splits too.\n\nWe'd want to use table.reltuples for consistency sake, though theoretically in this case I'd think it should be the same for indexes we care about...\n--\nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 19 Oct 2014 20:52:34 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On 10/18/14, 8:58 AM, Bruce Momjian wrote:\n> On Fri, Oct 17, 2014 at 11:03:04PM -0400, Tom Lane wrote:\n>> Bruce Momjian <bruce@momjian.us> writes:\n>>> On Fri, Oct 17, 2014 at 06:15:37PM -0400, Tom Lane wrote:\n>>>> Those stats were perfectly valid: what the planner is looking for is\n>>>> accurate minimum and maximum values for the index's leading column, and\n>>>> that's what it got. You're correct that a narrower index could have given\n>>>> the same results with a smaller disk footprint, but the planner got the\n>>>> results it needed from the index you provided for it to work with.\n>>\n>>> Uh, why is the optimizer looking at the index on a,b,c and not just the\n>>> stats on column a, for example? I am missing something here.\n>>\n>> Because it needs up-to-date min/max values in order to avoid being\n>> seriously misled about selectivities of values near the endpoints.\n>> See commit 40608e7f949fb7e4025c0ddd5be01939adc79eec.\n>\n> Oh, I had forgotten we did that. It is confusing that there is no way\n> via EXPLAIN to see the access, making the method of consulting pg_stat_*\n> and using EXPLAIN unreliable. Should we document this somewhere?\n\nI think we should. The common (mis)conception is that pg_stats shows *user-driven* access, not access because of stuff the system is doing.\n\nThis is actually a huge problem for anyone who's trying to figure out how useful indexes are; they see usage and thing they have queries that are using the index when in reality they don't.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 19:12:56 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"},{"msg_contents":"On Sat, Oct 18, 2014 at 02:20:45PM -0400, Bruce Momjian wrote:\n> On Sat, Oct 18, 2014 at 06:15:03PM +0200, Marko Tiikkaja wrote:\n> > On 10/18/14, 5:46 PM, Tom Lane wrote:\n> > >Marko Tiikkaja <marko@joh.to> writes:\n> > >>Yes, exactly; if I had had the option to disable the index from the\n> > >>optimizer's point of view, I'd have seen that it's not used for looking\n> > >>up any data by any queries, and thus I would have known that I can\n> > >>safely drop it without slowing down queries. Which was the only thing I\n> > >>cared about, and where the stats we provide failed me.\n> > >\n> > >This argument is *utterly* wrongheaded, because it assumes that the\n> > >planner's use of the index provided no benefit to your queries. If the\n> > >planner was touching the index at all then it was planning queries in\n> > >which knowledge of the extremal value was relevant to accurate selectivity\n> > >estimation. So it's quite likely that without the index you'd have gotten\n> > >different and inferior plans, whether or not those plans actually chose to\n> > >use the index.\n> > \n> > Maybe. But at the same time that's a big problem: there's no way of\n> > knowing whether the index is actually useful or not when it's used\n> > only by the query planner.\n> \n> That is a good point. Without an index, the executor is going to do a\n> sequential scan, while a missing index to the optimizer just means worse\n> statistics.\n\nI have applied the attached patch to document that the optimizer can\nincrease the index usage statistics.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Thu, 19 Mar 2015 22:38:52 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nThis week we had one of the most annoying problems I've ever encountered \\nwith postgres. We had a big index on multiple columns, say, foo(a, b, \\nc). According to pg_stat_all_indexes the index was being used *all the \\ntime*. However, after looking into our queries more closely, it turns \\nout that it was only being used to look up statistics for the foo.a \\ncolumn to estimate merge scan viability during planning. But this took \\nhours for two people to track down.\\n\\nSo what I'd like to have is a way to be able to distinguish between \\nindexes being used to answer queries, and ones being only used for stats \\nlookups during planning. Perhaps the easiest way would be adding a new \\ncolumn or two into pg_stat_all_indexes, which we would increment in \\nget_actual_variable_range() when fetching data.\\n\\nAny thoughts?\\n\\n\\n.marko\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 23:43:17 +0200\",\n \"msg_from\": \"Marko Tiikkaja <marko@joh.to>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"Marko Tiikkaja <marko@joh.to> writes:\\n> This week we had one of the most annoying problems I've ever encountered \\n> with postgres. We had a big index on multiple columns, say, foo(a, b, \\n> c). According to pg_stat_all_indexes the index was being used *all the \\n> time*. However, after looking into our queries more closely, it turns \\n> out that it was only being used to look up statistics for the foo.a \\n> column to estimate merge scan viability during planning. But this took \\n> hours for two people to track down.\\n\\n> So what I'd like to have is a way to be able to distinguish between \\n> indexes being used to answer queries, and ones being only used for stats \\n> lookups during planning.\\n\\nWhy? Used is used.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 17:47:59 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On 10/17/14, 11:47 PM, Tom Lane wrote:\\n> Marko Tiikkaja <marko@joh.to> writes:\\n>> This week we had one of the most annoying problems I've ever encountered\\n>> with postgres. We had a big index on multiple columns, say, foo(a, b,\\n>> c). According to pg_stat_all_indexes the index was being used *all the\\n>> time*. However, after looking into our queries more closely, it turns\\n>> out that it was only being used to look up statistics for the foo.a\\n>> column to estimate merge scan viability during planning. But this took\\n>> hours for two people to track down.\\n>\\n>> So what I'd like to have is a way to be able to distinguish between\\n>> indexes being used to answer queries, and ones being only used for stats\\n>> lookups during planning.\\n>\\n> Why? Used is used.\\n\\nBecause I don't need a 30GB index on foo(a,b,c) to look up statistics. \\nIf I ever have a problem, I can replace it with a 5GB one on foo(a).\\n\\n\\n.marko\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 23:49:05 +0200\",\n \"msg_from\": \"Marko Tiikkaja <marko@joh.to>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On 10/17/14, 4:49 PM, Marko Tiikkaja wrote:\\n> On 10/17/14, 11:47 PM, Tom Lane wrote:\\n>> Marko Tiikkaja <marko@joh.to> writes:\\n>>> This week we had one of the most annoying problems I've ever encountered\\n>>> with postgres. We had a big index on multiple columns, say, foo(a, b,\\n>>> c). According to pg_stat_all_indexes the index was being used *all the\\n>>> time*. However, after looking into our queries more closely, it turns\\n>>> out that it was only being used to look up statistics for the foo.a\\n>>> column to estimate merge scan viability during planning. But this took\\n>>> hours for two people to track down.\\n>>\\n>>> So what I'd like to have is a way to be able to distinguish between\\n>>> indexes being used to answer queries, and ones being only used for stats\\n>>> lookups during planning.\\n>>\\n>> Why? Used is used.\\n>\\n> Because I don't need a 30GB index on foo(a,b,c) to look up statistics. If I ever have a problem, I can replace it with a 5GB one on foo(a).\\n\\nThat problem can exist with user queries too. Perhaps it would be better to find a way to count scans that didn't use all the fields in the index.\\n\\nI do also see value in differentiating planning use from real query processing; not doing that can certainly cause confusion. What I don't know is if the added stats bloat is worth it. If we do go down that road, I think it'd be better to add an indicator to EState. Aside from allowing stats for all planning access, it should make it less likely that someone adds a new access path and forgets to mark it as internal (especially if the added field defaults to an invalid value).\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 16:59:10 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"Marko Tiikkaja <marko@joh.to> writes:\\n> On 10/17/14, 11:47 PM, Tom Lane wrote:\\n>> Marko Tiikkaja <marko@joh.to> writes:\\n>>> So what I'd like to have is a way to be able to distinguish between\\n>>> indexes being used to answer queries, and ones being only used for stats\\n>>> lookups during planning.\\n\\n>> Why? Used is used.\\n\\n> Because I don't need a 30GB index on foo(a,b,c) to look up statistics. \\n> If I ever have a problem, I can replace it with a 5GB one on foo(a).\\n\\nWell, the index might've been getting used in queries too in a way that\\nreally only involved the first column. I think you're solving the wrong\\nproblem here. The right problem is how to identify indexes that are\\nbeing used in a way that doesn't exploit all the columns. Which is not\\nnecessarily wrong in itself --- what you'd want is to figure out when the\\nlast column(s) are *never* used. The existing stats aren't terribly\\nhelpful for that, I agree.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 17:59:13 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On 10/17/14, 11:59 PM, Tom Lane wrote:\\n> Marko Tiikkaja <marko@joh.to> writes:\\n>> On 10/17/14, 11:47 PM, Tom Lane wrote:\\n>>> Marko Tiikkaja <marko@joh.to> writes:\\n>>>> So what I'd like to have is a way to be able to distinguish between\\n>>>> indexes being used to answer queries, and ones being only used for stats\\n>>>> lookups during planning.\\n>\\n>>> Why? Used is used.\\n>\\n>> Because I don't need a 30GB index on foo(a,b,c) to look up statistics.\\n>> If I ever have a problem, I can replace it with a 5GB one on foo(a).\\n>\\n> Well, the index might've been getting used in queries too in a way that\\n> really only involved the first column. I think you're solving the wrong\\n> problem here. The right problem is how to identify indexes that are\\n> being used in a way that doesn't exploit all the columns.\\n\\nI'm not sure I agree with that. Even if there was some information the \\nplanner could have extracted out of the index by using all columns (thus \\nappearing \\\"fully used\\\" in these hypothetical new statistics), I still \\nwould've wanted the index gone. But in this particular case, an index \\non foo(a) alone was not selective enough and it would have been a bad \\nchoice for practically every query, so I'm not sure what good those \\nstatistics were in the first place.\\n\\nI think there's a big difference between \\\"this index was used to look up \\nstuff for planning\\\" and \\\"this index was used to answer queries quickly\\\". \\n In my mind the first one belongs to the category \\\"this index was \\nconsidered\\\", and the latter is \\\"this index was actually useful\\\". But \\nmaybe I'm not seeing the big picture here.\\n\\n\\n.marko\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 00:09:03 +0200\",\n \"msg_from\": \"Marko Tiikkaja <marko@joh.to>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"Marko Tiikkaja <marko@joh.to> writes:\\n> On 10/17/14, 11:59 PM, Tom Lane wrote:\\n>> Well, the index might've been getting used in queries too in a way that\\n>> really only involved the first column. I think you're solving the wrong\\n>> problem here. The right problem is how to identify indexes that are\\n>> being used in a way that doesn't exploit all the columns.\\n\\n> I'm not sure I agree with that. Even if there was some information the \\n> planner could have extracted out of the index by using all columns (thus \\n> appearing \\\"fully used\\\" in these hypothetical new statistics), I still \\n> would've wanted the index gone. But in this particular case, an index \\n> on foo(a) alone was not selective enough and it would have been a bad \\n> choice for practically every query, so I'm not sure what good those \\n> statistics were in the first place.\\n\\nThose stats were perfectly valid: what the planner is looking for is\\naccurate minimum and maximum values for the index's leading column, and\\nthat's what it got. You're correct that a narrower index could have given\\nthe same results with a smaller disk footprint, but the planner got the\\nresults it needed from the index you provided for it to work with.\\n\\n> I think there's a big difference between \\\"this index was used to look up \\n> stuff for planning\\\" and \\\"this index was used to answer queries quickly\\\". \\n\\nI think that's utter nonsense. Even if there were any validity to the\\nposition, it wouldn't be enough to justify doubling the stats footprint\\nin order to track system-driven accesses separately from query-driven\\nones.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 18:15:37 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On 10/18/14, 12:15 AM, Tom Lane wrote:\\n> Marko Tiikkaja <marko@joh.to> writes:\\n>> I think there's a big difference between \\\"this index was used to look up\\n>> stuff for planning\\\" and \\\"this index was used to answer queries quickly\\\".\\n>\\n> I think that's utter nonsense.\\n\\nWell you probably know a bit more about the optimizer than I do. But I \\ncan't see a case where the stats provided by the index would be useful \\nfor choosing between two (or more) plans that don't use the index in the \\nactual query. If you're saying that there are such cases, then clearly \\nI don't know something, and my thinking is in the wrong here.\\n\\n\\n.marko\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 00:49:54 +0200\",\n \"msg_from\": \"Marko Tiikkaja <marko@joh.to>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 06:15:37PM -0400, Tom Lane wrote:\\n> Marko Tiikkaja <marko@joh.to> writes:\\n> > On 10/17/14, 11:59 PM, Tom Lane wrote:\\n> >> Well, the index might've been getting used in queries too in a way that\\n> >> really only involved the first column. I think you're solving the wrong\\n> >> problem here. The right problem is how to identify indexes that are\\n> >> being used in a way that doesn't exploit all the columns.\\n> \\n> > I'm not sure I agree with that. Even if there was some information the \\n> > planner could have extracted out of the index by using all columns (thus \\n> > appearing \\\"fully used\\\" in these hypothetical new statistics), I still \\n> > would've wanted the index gone. But in this particular case, an index \\n> > on foo(a) alone was not selective enough and it would have been a bad \\n> > choice for practically every query, so I'm not sure what good those \\n> > statistics were in the first place.\\n> \\n> Those stats were perfectly valid: what the planner is looking for is\\n> accurate minimum and maximum values for the index's leading column, and\\n> that's what it got. You're correct that a narrower index could have given\\n> the same results with a smaller disk footprint, but the planner got the\\n> results it needed from the index you provided for it to work with.\\n\\nUh, why is the optimizer looking at the index on a,b,c and not just the\\nstats on column a, for example? I am missing something here.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 22:30:01 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"Bruce Momjian <bruce@momjian.us> writes:\\n> On Fri, Oct 17, 2014 at 06:15:37PM -0400, Tom Lane wrote:\\n>> Those stats were perfectly valid: what the planner is looking for is\\n>> accurate minimum and maximum values for the index's leading column, and\\n>> that's what it got. You're correct that a narrower index could have given\\n>> the same results with a smaller disk footprint, but the planner got the\\n>> results it needed from the index you provided for it to work with.\\n\\n> Uh, why is the optimizer looking at the index on a,b,c and not just the\\n> stats on column a, for example? I am missing something here.\\n\\nBecause it needs up-to-date min/max values in order to avoid being\\nseriously misled about selectivities of values near the endpoints.\\nSee commit 40608e7f949fb7e4025c0ddd5be01939adc79eec.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 23:03:04 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Because it needs up-to-date min/max values in order to avoid being\\n> seriously misled about selectivities of values near the endpoints.\\n> See commit 40608e7f949fb7e4025c0ddd5be01939adc79eec.\\n\\nBTW, on re-reading that code I notice that it will happily seize upon\\nthe first suitable index (\\\"first\\\" in OID order), regardless of how many\\nlower-order columns that index has got. This doesn't make any difference\\nI think for get_actual_variable_range's own purposes, because it's only\\nexpecting to touch the endmost index page regardless. However, in light\\nof Marko's complaint maybe we should teach it to check all the indexes\\nand prefer the matching one with fewest columns? It would only take a\\ncouple extra lines of code, and probably not that many added cycles\\nconsidering we're going to do an index access of some sort. But I'm\\nnot sure if it's worth any extra effort --- I think in his example\\ncase, there wasn't any narrower index anyway.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 17 Oct 2014 23:16:08 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 17, 2014 at 11:03:04PM -0400, Tom Lane wrote:\\n> Bruce Momjian <bruce@momjian.us> writes:\\n> > On Fri, Oct 17, 2014 at 06:15:37PM -0400, Tom Lane wrote:\\n> >> Those stats were perfectly valid: what the planner is looking for is\\n> >> accurate minimum and maximum values for the index's leading column, and\\n> >> that's what it got. You're correct that a narrower index could have given\\n> >> the same results with a smaller disk footprint, but the planner got the\\n> >> results it needed from the index you provided for it to work with.\\n> \\n> > Uh, why is the optimizer looking at the index on a,b,c and not just the\\n> > stats on column a, for example? I am missing something here.\\n> \\n> Because it needs up-to-date min/max values in order to avoid being\\n> seriously misled about selectivities of values near the endpoints.\\n> See commit 40608e7f949fb7e4025c0ddd5be01939adc79eec.\\n\\nOh, I had forgotten we did that. It is confusing that there is no way\\nvia EXPLAIN to see the access, making the method of consulting pg_stat_*\\nand using EXPLAIN unreliable. Should we document this somewhere?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 09:58:10 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On 10/18/14, 5:16 AM, Tom Lane wrote:\\n> BTW, on re-reading that code I notice that it will happily seize upon\\n> the first suitable index (\\\"first\\\" in OID order), regardless of how many\\n> lower-order columns that index has got. This doesn't make any difference\\n> I think for get_actual_variable_range's own purposes, because it's only\\n> expecting to touch the endmost index page regardless. However, in light\\n> of Marko's complaint maybe we should teach it to check all the indexes\\n> and prefer the matching one with fewest columns? It would only take a\\n> couple extra lines of code, and probably not that many added cycles\\n> considering we're going to do an index access of some sort. But I'm\\n> not sure if it's worth any extra effort --- I think in his example\\n> case, there wasn't any narrower index anyway.\\n\\nPerhaps accidentally this would have helped in my case, actually, since \\nI could have created a new, smaller index CONCURRENTLY and then seen \\nthat the usage of the other index stopped increasing. With the \\\"pick \\nthe smallest OID\\\" behaviour that was not possible. Another idea had was \\nsome way to tell the optimizer not to use that particular index for \\nstats lookups, but probably the use case for such a feature would be a \\nbit narrow.\\n\\nAll that said, I don't think my struggles justify the change you \\ndescribed above. Not sure if it's a good idea or not.\\n\\n\\n.marko\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 16:29:41 +0200\",\n \"msg_from\": \"Marko Tiikkaja <marko@joh.to>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On Sat, Oct 18, 2014 at 04:29:41PM +0200, Marko Tiikkaja wrote:\\n> On 10/18/14, 5:16 AM, Tom Lane wrote:\\n> >BTW, on re-reading that code I notice that it will happily seize upon\\n> >the first suitable index (\\\"first\\\" in OID order), regardless of how many\\n> >lower-order columns that index has got. This doesn't make any difference\\n> >I think for get_actual_variable_range's own purposes, because it's only\\n> >expecting to touch the endmost index page regardless. However, in light\\n> >of Marko's complaint maybe we should teach it to check all the indexes\\n> >and prefer the matching one with fewest columns? It would only take a\\n> >couple extra lines of code, and probably not that many added cycles\\n> >considering we're going to do an index access of some sort. But I'm\\n> >not sure if it's worth any extra effort --- I think in his example\\n> >case, there wasn't any narrower index anyway.\\n> \\n> Perhaps accidentally this would have helped in my case, actually,\\n> since I could have created a new, smaller index CONCURRENTLY and\\n> then seen that the usage of the other index stopped increasing.\\n> With the \\\"pick the smallest OID\\\" behaviour that was not possible.\\n> Another idea had was some way to tell the optimizer not to use that\\n> particular index for stats lookups, but probably the use case for\\n> such a feature would be a bit narrow.\\n\\nWell, if the index is there, why not use it? I thought the problem was\\njust that you had no visibility into how those statistics were being\\naccessed. Most people think EXPLAIN shows all accesses, but obviously\\nnow it doesn't.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 10:33:28 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On 10/18/14, 4:33 PM, Bruce Momjian wrote:\\n> On Sat, Oct 18, 2014 at 04:29:41PM +0200, Marko Tiikkaja wrote:\\n>> Another idea had was some way to tell the optimizer not to use that\\n>> particular index for stats lookups, but probably the use case for\\n>> such a feature would be a bit narrow.\\n>\\n> Well, if the index is there, why not use it? I thought the problem was\\n> just that you had no visibility into how those statistics were being\\n> accessed.\\n\\nYes, exactly; if I had had the option to disable the index from the \\noptimizer's point of view, I'd have seen that it's not used for looking \\nup any data by any queries, and thus I would have known that I can \\nsafely drop it without slowing down queries. Which was the only thing I \\ncared about, and where the stats we provide failed me.\\n\\n\\n.marko\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 16:38:37 +0200\",\n \"msg_from\": \"Marko Tiikkaja <marko@joh.to>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On Sat, Oct 18, 2014 at 04:38:37PM +0200, Marko Tiikkaja wrote:\\n> On 10/18/14, 4:33 PM, Bruce Momjian wrote:\\n> >On Sat, Oct 18, 2014 at 04:29:41PM +0200, Marko Tiikkaja wrote:\\n> >>Another idea had was some way to tell the optimizer not to use that\\n> >>particular index for stats lookups, but probably the use case for\\n> >>such a feature would be a bit narrow.\\n> >\\n> >Well, if the index is there, why not use it? I thought the problem was\\n> >just that you had no visibility into how those statistics were being\\n> >accessed.\\n> \\n> Yes, exactly; if I had had the option to disable the index from the\\n> optimizer's point of view, I'd have seen that it's not used for\\n> looking up any data by any queries, and thus I would have known that\\n> I can safely drop it without slowing down queries. Which was the\\n> only thing I cared about, and where the stats we provide failed me.\\n\\nHow many other cases do we have where the statistics are getting\\nincremented and there is no user visibility into the operation?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 10:46:42 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"Bruce Momjian <bruce@momjian.us> writes:\\n> How many other cases do we have where the statistics are getting\\n> incremented and there is no user visibility into the operation?\\n\\n* system catalog accesses\\n* vacuum/analyze/cluster/etc\\n\\nThe fact that system-initiated accesses get counted in the statistics\\nis a feature, not a bug.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 11:05:48 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"Marko Tiikkaja <marko@joh.to> writes:\\n> On 10/18/14, 4:33 PM, Bruce Momjian wrote:\\n>> Well, if the index is there, why not use it? I thought the problem was\\n>> just that you had no visibility into how those statistics were being\\n>> accessed.\\n\\n> Yes, exactly; if I had had the option to disable the index from the \\n> optimizer's point of view, I'd have seen that it's not used for looking \\n> up any data by any queries, and thus I would have known that I can \\n> safely drop it without slowing down queries. Which was the only thing I \\n> cared about, and where the stats we provide failed me.\\n\\nThis argument is *utterly* wrongheaded, because it assumes that the\\nplanner's use of the index provided no benefit to your queries. If the\\nplanner was touching the index at all then it was planning queries in\\nwhich knowledge of the extremal value was relevant to accurate selectivity\\nestimation. So it's quite likely that without the index you'd have gotten\\ndifferent and inferior plans, whether or not those plans actually chose to\\nuse the index.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 11:46:47 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On 10/18/14, 5:46 PM, Tom Lane wrote:\\n> Marko Tiikkaja <marko@joh.to> writes:\\n>> Yes, exactly; if I had had the option to disable the index from the\\n>> optimizer's point of view, I'd have seen that it's not used for looking\\n>> up any data by any queries, and thus I would have known that I can\\n>> safely drop it without slowing down queries. Which was the only thing I\\n>> cared about, and where the stats we provide failed me.\\n>\\n> This argument is *utterly* wrongheaded, because it assumes that the\\n> planner's use of the index provided no benefit to your queries. If the\\n> planner was touching the index at all then it was planning queries in\\n> which knowledge of the extremal value was relevant to accurate selectivity\\n> estimation. So it's quite likely that without the index you'd have gotten\\n> different and inferior plans, whether or not those plans actually chose to\\n> use the index.\\n\\nMaybe. But at the same time that's a big problem: there's no way of \\nknowing whether the index is actually useful or not when it's used only \\nby the query planner.\\n\\n\\n.marko\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 18:15:03 +0200\",\n \"msg_from\": \"Marko Tiikkaja <marko@joh.to>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On Sat, Oct 18, 2014 at 06:15:03PM +0200, Marko Tiikkaja wrote:\\n> On 10/18/14, 5:46 PM, Tom Lane wrote:\\n> >Marko Tiikkaja <marko@joh.to> writes:\\n> >>Yes, exactly; if I had had the option to disable the index from the\\n> >>optimizer's point of view, I'd have seen that it's not used for looking\\n> >>up any data by any queries, and thus I would have known that I can\\n> >>safely drop it without slowing down queries. Which was the only thing I\\n> >>cared about, and where the stats we provide failed me.\\n> >\\n> >This argument is *utterly* wrongheaded, because it assumes that the\\n> >planner's use of the index provided no benefit to your queries. If the\\n> >planner was touching the index at all then it was planning queries in\\n> >which knowledge of the extremal value was relevant to accurate selectivity\\n> >estimation. So it's quite likely that without the index you'd have gotten\\n> >different and inferior plans, whether or not those plans actually chose to\\n> >use the index.\\n> \\n> Maybe. But at the same time that's a big problem: there's no way of\\n> knowing whether the index is actually useful or not when it's used\\n> only by the query planner.\\n\\nThat is a good point. Without an index, the executor is going to do a\\nsequential scan, while a missing index to the optimizer just means worse\\nstatistics.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 14:20:45 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On 10/17/14, 10:16 PM, Tom Lane wrote:\\n> I wrote:\\n>> Because it needs up-to-date min/max values in order to avoid being\\n>> seriously misled about selectivities of values near the endpoints.\\n>> See commit 40608e7f949fb7e4025c0ddd5be01939adc79eec.\\n>\\n> BTW, on re-reading that code I notice that it will happily seize upon\\n> the first suitable index (\\\"first\\\" in OID order), regardless of how many\\n> lower-order columns that index has got. This doesn't make any difference\\n> I think for get_actual_variable_range's own purposes, because it's only\\n> expecting to touch the endmost index page regardless. However, in light\\n> of Marko's complaint maybe we should teach it to check all the indexes\\n> and prefer the matching one with fewest columns? It would only take a\\n> couple extra lines of code, and probably not that many added cycles\\n> considering we're going to do an index access of some sort. But I'm\\n> not sure if it's worth any extra effort --- I think in his example\\n> case, there wasn't any narrower index anyway.\\n\\nThe real cost here isn't the number of columns, it's the size of the index, no? So shouldn't we look at relpages instead? For example, you'd certainly want to use an index on (field_we_care_about, smallint_field) over an index on (field_we_care_about, big_honking_text_field).\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 20:54:29 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\\n> On 10/17/14, 10:16 PM, Tom Lane wrote:\\n>> BTW, on re-reading that code I notice that it will happily seize upon\\n>> the first suitable index (\\\"first\\\" in OID order), regardless of how many\\n>> lower-order columns that index has got. This doesn't make any difference\\n>> I think for get_actual_variable_range's own purposes, because it's only\\n>> expecting to touch the endmost index page regardless. However, in light\\n>> of Marko's complaint maybe we should teach it to check all the indexes\\n>> and prefer the matching one with fewest columns?\\n\\n> The real cost here isn't the number of columns, it's the size of the index, no? So shouldn't we look at relpages instead? For example, you'd certainly want to use an index on (field_we_care_about, smallint_field) over an index on (field_we_care_about, big_honking_text_field).\\n\\nYeah, perhaps. I'd been wondering about adding a tie-breaking rule,\\nbut that's a much simpler way to think about it. OTOH, that approach\\ncould result in some instability in the choice of index: if you've got\\nboth (field_we_care_about, some_int_field) and (field_we_care_about,\\nsome_other_int_field) then it might switch between choosing those two\\nindexes from day to day depending on basically-chance issues like when\\npage splits occur. That would probably annoy Marko even more than the\\ncurrent behavior :-(, because it would scatter the planner's usage\\nacross multiple indexes for no very good reason.\\n\\nThe coding I'd been imagining at first would basically break ties in\\ncolumn count according to index OID order, so its choices would be stable\\nas long as you did not add/drop indexes. That seems like a good property\\nto try to preserve.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 19 Oct 2014 15:09:57 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On 10/19/14, 2:09 PM, Tom Lane wrote:\\n> Yeah, perhaps. I'd been wondering about adding a tie-breaking rule,\\n> but that's a much simpler way to think about it. OTOH, that approach\\n> could result in some instability in the choice of index: if you've got\\n> both (field_we_care_about, some_int_field) and (field_we_care_about,\\n> some_other_int_field) then it might switch between choosing those two\\n> indexes from day to day depending on basically-chance issues like when\\n> page splits occur. That would probably annoy Marko even more than the\\n> current behavior:-(, because it would scatter the planner's usage\\n> across multiple indexes for no very good reason.\\n>\\n> The coding I'd been imagining at first would basically break ties in\\n> column count according to index OID order, so its choices would be stable\\n> as long as you did not add/drop indexes. That seems like a good property\\n> to try to preserve.\\n\\nMaybe a good alternative is:\\n\\nORDER BY int( table.reltuples / index.relpages / BLKSZ ) DESC, oid\\n\\nBy comparing on average tuple size throwing away the fraction presumably we'd throw away noise from page splits too.\\n\\nWe'd want to use table.reltuples for consistency sake, though theoretically in this case I'd think it should be the same for indexes we care about...\\n--\\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 19 Oct 2014 20:52:34 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On 10/18/14, 8:58 AM, Bruce Momjian wrote:\\n> On Fri, Oct 17, 2014 at 11:03:04PM -0400, Tom Lane wrote:\\n>> Bruce Momjian <bruce@momjian.us> writes:\\n>>> On Fri, Oct 17, 2014 at 06:15:37PM -0400, Tom Lane wrote:\\n>>>> Those stats were perfectly valid: what the planner is looking for is\\n>>>> accurate minimum and maximum values for the index's leading column, and\\n>>>> that's what it got. You're correct that a narrower index could have given\\n>>>> the same results with a smaller disk footprint, but the planner got the\\n>>>> results it needed from the index you provided for it to work with.\\n>>\\n>>> Uh, why is the optimizer looking at the index on a,b,c and not just the\\n>>> stats on column a, for example? I am missing something here.\\n>>\\n>> Because it needs up-to-date min/max values in order to avoid being\\n>> seriously misled about selectivities of values near the endpoints.\\n>> See commit 40608e7f949fb7e4025c0ddd5be01939adc79eec.\\n>\\n> Oh, I had forgotten we did that. It is confusing that there is no way\\n> via EXPLAIN to see the access, making the method of consulting pg_stat_*\\n> and using EXPLAIN unreliable. Should we document this somewhere?\\n\\nI think we should. The common (mis)conception is that pg_stats shows *user-driven* access, not access because of stuff the system is doing.\\n\\nThis is actually a huge problem for anyone who's trying to figure out how useful indexes are; they see usage and thing they have queries that are using the index when in reality they don't.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 19:12:56 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n },\n {\n \"msg_contents\": \"On Sat, Oct 18, 2014 at 02:20:45PM -0400, Bruce Momjian wrote:\\n> On Sat, Oct 18, 2014 at 06:15:03PM +0200, Marko Tiikkaja wrote:\\n> > On 10/18/14, 5:46 PM, Tom Lane wrote:\\n> > >Marko Tiikkaja <marko@joh.to> writes:\\n> > >>Yes, exactly; if I had had the option to disable the index from the\\n> > >>optimizer's point of view, I'd have seen that it's not used for looking\\n> > >>up any data by any queries, and thus I would have known that I can\\n> > >>safely drop it without slowing down queries. Which was the only thing I\\n> > >>cared about, and where the stats we provide failed me.\\n> > >\\n> > >This argument is *utterly* wrongheaded, because it assumes that the\\n> > >planner's use of the index provided no benefit to your queries. If the\\n> > >planner was touching the index at all then it was planning queries in\\n> > >which knowledge of the extremal value was relevant to accurate selectivity\\n> > >estimation. So it's quite likely that without the index you'd have gotten\\n> > >different and inferior plans, whether or not those plans actually chose to\\n> > >use the index.\\n> > \\n> > Maybe. But at the same time that's a big problem: there's no way of\\n> > knowing whether the index is actually useful or not when it's used\\n> > only by the query planner.\\n> \\n> That is a good point. Without an index, the executor is going to do a\\n> sequential scan, while a missing index to the optimizer just means worse\\n> statistics.\\n\\nI have applied the attached patch to document that the optimizer can\\nincrease the index usage statistics.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Thu, 19 Mar 2015 22:38:52 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: get_actual_variable_range vs idx_scan/idx_tup_fetch\"\n }\n]","isConversational":false}}},{"rowIdx":39681,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"I looked into bug #11703,\nhttp://www.postgresql.org/message-id/20141017162236.2523.74453@wrigleys.postgresql.org\n\nWhat is happening here is that the FieldSelect optimization in\neval_const_expressions fails to play nice with the logic that determines\nwhich Vars need to be available from a relation scan. We start out with,\nsay,\n\nselect x from my_table, func2(my_table) f(x);\n\nwhich inline_set_returning_functions() expands into\n\nselect x from my_table, lateral (select func1(my_table)) f(x);\n\nAt this point we decide that the sub-select can't be flattened into the\nupper query (in this case, because it lacks a FROM clause, but any other\nreason to suppress flattening would have the same result). So we put the\nsub-select on the back burner for the moment and start to plan the outer\nquery. In particular, we determine that the only Var that the scan of\nmy_table has to produce is a whole-row Var (\"my_table.*\") to satisfy the\nlateral reference in the sub-select.\n\nEventually, we get around to recursively planning the sub-query, and at\nthat point we'll inline the func1() call, producing a FieldSelect applied\nto the whole-row Var for \"mytable\". When we do const-simplification on\nthe resulting expression, that pair of nodes is reduced to a simple Var\nreferencing mytable.my_column. Which ought to be fine --- except that we\nalready planned the scan of my_table on the assumption that it should\nproduce \"my_table.*\". So after all the dust settles and setrefs.c is\ndoing the janitorial work, it finds out that the lateral reference to\nmytable.my_column can't be satisfied, and you get the dreaded \"variable\nnot found in subplan target list\" error.\n\nAFAICS, the only way to fix this \"right\" would require major\nrearrangements of the order in which the planner does stuff, for example\ninsisting on doing function inlining and const-simplification all the way\ndown in subqueries before we plan the outer query level. That'd be quite\na lot of work to change, and the odds of introducing new bugs are high\nenough that I think back-patching would be out of the question anyway.\n\nI think the only practical fix is to lobotomize the FieldSelect-collapsing\noptimization. It's probably safe enough to keep allowing it for whole-row\nVars of the current query level, but not those with levelsup > 0.\n\nAnybody have another idea?\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 12:48:16 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"FieldSelect optimization versus overall planner organization"}],"string":"[\n {\n \"msg_contents\": \"I looked into bug #11703,\\nhttp://www.postgresql.org/message-id/20141017162236.2523.74453@wrigleys.postgresql.org\\n\\nWhat is happening here is that the FieldSelect optimization in\\neval_const_expressions fails to play nice with the logic that determines\\nwhich Vars need to be available from a relation scan. We start out with,\\nsay,\\n\\nselect x from my_table, func2(my_table) f(x);\\n\\nwhich inline_set_returning_functions() expands into\\n\\nselect x from my_table, lateral (select func1(my_table)) f(x);\\n\\nAt this point we decide that the sub-select can't be flattened into the\\nupper query (in this case, because it lacks a FROM clause, but any other\\nreason to suppress flattening would have the same result). So we put the\\nsub-select on the back burner for the moment and start to plan the outer\\nquery. In particular, we determine that the only Var that the scan of\\nmy_table has to produce is a whole-row Var (\\\"my_table.*\\\") to satisfy the\\nlateral reference in the sub-select.\\n\\nEventually, we get around to recursively planning the sub-query, and at\\nthat point we'll inline the func1() call, producing a FieldSelect applied\\nto the whole-row Var for \\\"mytable\\\". When we do const-simplification on\\nthe resulting expression, that pair of nodes is reduced to a simple Var\\nreferencing mytable.my_column. Which ought to be fine --- except that we\\nalready planned the scan of my_table on the assumption that it should\\nproduce \\\"my_table.*\\\". So after all the dust settles and setrefs.c is\\ndoing the janitorial work, it finds out that the lateral reference to\\nmytable.my_column can't be satisfied, and you get the dreaded \\\"variable\\nnot found in subplan target list\\\" error.\\n\\nAFAICS, the only way to fix this \\\"right\\\" would require major\\nrearrangements of the order in which the planner does stuff, for example\\ninsisting on doing function inlining and const-simplification all the way\\ndown in subqueries before we plan the outer query level. That'd be quite\\na lot of work to change, and the odds of introducing new bugs are high\\nenough that I think back-patching would be out of the question anyway.\\n\\nI think the only practical fix is to lobotomize the FieldSelect-collapsing\\noptimization. It's probably safe enough to keep allowing it for whole-row\\nVars of the current query level, but not those with levelsup > 0.\\n\\nAnybody have another idea?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 12:48:16 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"FieldSelect optimization versus overall planner organization\"\n }\n]","isConversational":false}}},{"rowIdx":39682,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI've noticed an interesting phenomenon. Look at the following pgbench -c\n32 -j 32 -P 1 output on a scale 200 database. Then guess where I've\nstarted\nSET vacuum_cost_delay = '20ms';\nANALYZE pgbench_accounts;\nand when it finished. The cost_delay is to simulate a autovac triggered\nanalyze.\n\nprogress: 18.0 s, 17018.1 tps, lat 1.879 ms stddev 1.018\nprogress: 19.0 s, 17039.6 tps, lat 1.876 ms stddev 0.806\nprogress: 20.0 s, 16892.1 tps, lat 1.893 ms stddev 0.732\nprogress: 21.0 s, 16636.9 tps, lat 1.922 ms stddev 0.933\nprogress: 22.0 s, 16198.4 tps, lat 1.974 ms stddev 1.180\nprogress: 23.0 s, 16362.1 tps, lat 1.954 ms stddev 0.858\nprogress: 24.0 s, 15926.0 tps, lat 2.007 ms stddev 0.967\nprogress: 25.0 s, 14550.5 tps, lat 2.197 ms stddev 1.870\nprogress: 26.0 s, 15042.2 tps, lat 2.125 ms stddev 1.135\nprogress: 27.0 s, 14579.1 tps, lat 2.196 ms stddev 1.016\nprogress: 28.0 s, 14145.0 tps, lat 2.261 ms stddev 1.160\nprogress: 29.0 s, 13734.1 tps, lat 2.328 ms stddev 0.992\nprogress: 30.0 s, 12127.6 tps, lat 2.565 ms stddev 1.930\nprogress: 31.0 s, 12375.2 tps, lat 1.807 ms stddev 1.556\nprogress: 32.0 s, 11939.9 tps, lat 3.551 ms stddev 45.457\nprogress: 33.0 s, 12466.0 tps, lat 2.566 ms stddev 1.517\nprogress: 34.0 s, 12166.1 tps, lat 2.634 ms stddev 1.309\nprogress: 35.0 s, 11743.7 tps, lat 2.723 ms stddev 1.576\nprogress: 36.0 s, 11623.3 tps, lat 2.752 ms stddev 1.237\nprogress: 37.0 s, 11341.4 tps, lat 2.817 ms stddev 1.413\nprogress: 38.0 s, 11130.2 tps, lat 2.875 ms stddev 1.359\nprogress: 39.0 s, 10988.0 tps, lat 2.910 ms stddev 1.132\nprogress: 40.0 s, 10681.9 tps, lat 2.993 ms stddev 1.355\nprogress: 41.0 s, 10573.1 tps, lat 3.027 ms stddev 1.738\nprogress: 42.0 s, 10174.9 tps, lat 3.143 ms stddev 2.125\nprogress: 43.0 s, 10190.1 tps, lat 3.136 ms stddev 1.892\nprogress: 44.0 s, 9943.2 tps, lat 3.219 ms stddev 1.740\nprogress: 45.0 s, 9792.5 tps, lat 3.265 ms stddev 2.229\nprogress: 46.0 s, 9701.6 tps, lat 3.299 ms stddev 1.834\nprogress: 47.0 s, 9376.6 tps, lat 3.410 ms stddev 2.149\nprogress: 48.0 s, 9217.2 tps, lat 3.471 ms stddev 1.992\nprogress: 49.0 s, 9265.9 tps, lat 3.449 ms stddev 1.819\nprogress: 50.0 s, 9090.9 tps, lat 3.521 ms stddev 1.916\nprogress: 51.0 s, 8933.9 tps, lat 3.578 ms stddev 1.813\nprogress: 52.0 s, 8849.6 tps, lat 3.616 ms stddev 1.913\nprogress: 53.0 s, 8767.5 tps, lat 3.649 ms stddev 1.952\nprogress: 54.0 s, 8521.5 tps, lat 3.753 ms stddev 2.158\nprogress: 55.0 s, 8382.2 tps, lat 3.816 ms stddev 2.383\nprogress: 56.0 s, 8298.1 tps, lat 3.848 ms stddev 2.727\nprogress: 57.0 s, 8369.5 tps, lat 3.825 ms stddev 2.897\nprogress: 58.0 s, 8205.0 tps, lat 3.902 ms stddev 2.669\nprogress: 59.0 s, 8048.1 tps, lat 3.974 ms stddev 2.473\nprogress: 60.0 s, 7886.5 tps, lat 4.050 ms stddev 2.456\nprogress: 61.0 s, 7539.0 tps, lat 4.251 ms stddev 2.992\nprogress: 62.0 s, 7815.4 tps, lat 4.090 ms stddev 2.845\nprogress: 63.0 s, 7863.4 tps, lat 4.062 ms stddev 3.032\nprogress: 64.0 s, 7750.6 tps, lat 4.135 ms stddev 2.677\nprogress: 65.0 s, 7580.7 tps, lat 4.217 ms stddev 2.210\nprogress: 66.0 s, 7164.8 tps, lat 4.466 ms stddev 2.511\nprogress: 67.0 s, 7504.4 tps, lat 4.265 ms stddev 1.343\nprogress: 68.0 s, 9530.0 tps, lat 3.360 ms stddev 1.742\nprogress: 69.0 s, 16751.3 tps, lat 1.909 ms stddev 0.652\nprogress: 70.0 s, 16123.7 tps, lat 1.983 ms stddev 0.751\nprogress: 71.0 s, 15122.7 tps, lat 2.115 ms stddev 1.065\n\nYes. That's right. ANALYZE caused a slowdown of nearly *2. Executing it\non a bigger database where it takes longer makes it much worse. Also, a\nnonuniform access pattern also makes it *much* worse.\n\nSo, you might ask, why is that? It's because ANALYZE keeps a snapshot\naround and thus blocks RecentGlobalXmin from advancing. Leading to hot\npruning being impossible on \"popular\" pages.\n\nNow, the update rates in this example certainly are above what most\nsites see, but in practical sites relations are also much larger. And\nthus autoanalyze also takes longer... Much longer with a large sample\nsize.\n\nInterestingly we already set PROC_IN_ANALYZE - but we never actually\nlook at it. I wonder what it'd take to ignore analyze in\nGetSnapshotData()/GetOldestXmin()... At first sight it looks quite\ndoable to not have a snapshot around (or setup to be ignored) while\nacquire_sample_rows() is running. As that's the expensive bit...\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 19:49:09 +0200","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":true,"msg_subject":"(auto-)analyze causing bloat/load"},{"msg_contents":"Andres Freund <andres@2ndquadrant.com> writes:\n> Interestingly we already set PROC_IN_ANALYZE - but we never actually\n> look at it. I wonder what it'd take to ignore analyze in\n> GetSnapshotData()/GetOldestXmin()... At first sight it looks quite\n> doable to not have a snapshot around (or setup to be ignored) while\n> acquire_sample_rows() is running. As that's the expensive bit...\n\nSay what? How is ANALYZE supposed to know which rows are valid\nwithout a snapshot?\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 14:30:53 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: (auto-)analyze causing bloat/load"},{"msg_contents":"On 2014-10-18 14:30:53 -0400, Tom Lane wrote:\n> Andres Freund <andres@2ndquadrant.com> writes:\n> > Interestingly we already set PROC_IN_ANALYZE - but we never actually\n> > look at it. I wonder what it'd take to ignore analyze in\n> > GetSnapshotData()/GetOldestXmin()... At first sight it looks quite\n> > doable to not have a snapshot around (or setup to be ignored) while\n> > acquire_sample_rows() is running. As that's the expensive bit...\n> \n> Say what? How is ANALYZE supposed to know which rows are valid\n> without a snapshot?\n\nI haven't fully investigated it yet. But it's already not using it's own\nsnapshot to determine visibility. Instead it uses GetOldestXmin() +\nHeapTupleSatisfiesVacuum().\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 20:33:57 +0200","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":true,"msg_subject":"Re: (auto-)analyze causing bloat/load"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI've noticed an interesting phenomenon. Look at the following pgbench -c\\n32 -j 32 -P 1 output on a scale 200 database. Then guess where I've\\nstarted\\nSET vacuum_cost_delay = '20ms';\\nANALYZE pgbench_accounts;\\nand when it finished. The cost_delay is to simulate a autovac triggered\\nanalyze.\\n\\nprogress: 18.0 s, 17018.1 tps, lat 1.879 ms stddev 1.018\\nprogress: 19.0 s, 17039.6 tps, lat 1.876 ms stddev 0.806\\nprogress: 20.0 s, 16892.1 tps, lat 1.893 ms stddev 0.732\\nprogress: 21.0 s, 16636.9 tps, lat 1.922 ms stddev 0.933\\nprogress: 22.0 s, 16198.4 tps, lat 1.974 ms stddev 1.180\\nprogress: 23.0 s, 16362.1 tps, lat 1.954 ms stddev 0.858\\nprogress: 24.0 s, 15926.0 tps, lat 2.007 ms stddev 0.967\\nprogress: 25.0 s, 14550.5 tps, lat 2.197 ms stddev 1.870\\nprogress: 26.0 s, 15042.2 tps, lat 2.125 ms stddev 1.135\\nprogress: 27.0 s, 14579.1 tps, lat 2.196 ms stddev 1.016\\nprogress: 28.0 s, 14145.0 tps, lat 2.261 ms stddev 1.160\\nprogress: 29.0 s, 13734.1 tps, lat 2.328 ms stddev 0.992\\nprogress: 30.0 s, 12127.6 tps, lat 2.565 ms stddev 1.930\\nprogress: 31.0 s, 12375.2 tps, lat 1.807 ms stddev 1.556\\nprogress: 32.0 s, 11939.9 tps, lat 3.551 ms stddev 45.457\\nprogress: 33.0 s, 12466.0 tps, lat 2.566 ms stddev 1.517\\nprogress: 34.0 s, 12166.1 tps, lat 2.634 ms stddev 1.309\\nprogress: 35.0 s, 11743.7 tps, lat 2.723 ms stddev 1.576\\nprogress: 36.0 s, 11623.3 tps, lat 2.752 ms stddev 1.237\\nprogress: 37.0 s, 11341.4 tps, lat 2.817 ms stddev 1.413\\nprogress: 38.0 s, 11130.2 tps, lat 2.875 ms stddev 1.359\\nprogress: 39.0 s, 10988.0 tps, lat 2.910 ms stddev 1.132\\nprogress: 40.0 s, 10681.9 tps, lat 2.993 ms stddev 1.355\\nprogress: 41.0 s, 10573.1 tps, lat 3.027 ms stddev 1.738\\nprogress: 42.0 s, 10174.9 tps, lat 3.143 ms stddev 2.125\\nprogress: 43.0 s, 10190.1 tps, lat 3.136 ms stddev 1.892\\nprogress: 44.0 s, 9943.2 tps, lat 3.219 ms stddev 1.740\\nprogress: 45.0 s, 9792.5 tps, lat 3.265 ms stddev 2.229\\nprogress: 46.0 s, 9701.6 tps, lat 3.299 ms stddev 1.834\\nprogress: 47.0 s, 9376.6 tps, lat 3.410 ms stddev 2.149\\nprogress: 48.0 s, 9217.2 tps, lat 3.471 ms stddev 1.992\\nprogress: 49.0 s, 9265.9 tps, lat 3.449 ms stddev 1.819\\nprogress: 50.0 s, 9090.9 tps, lat 3.521 ms stddev 1.916\\nprogress: 51.0 s, 8933.9 tps, lat 3.578 ms stddev 1.813\\nprogress: 52.0 s, 8849.6 tps, lat 3.616 ms stddev 1.913\\nprogress: 53.0 s, 8767.5 tps, lat 3.649 ms stddev 1.952\\nprogress: 54.0 s, 8521.5 tps, lat 3.753 ms stddev 2.158\\nprogress: 55.0 s, 8382.2 tps, lat 3.816 ms stddev 2.383\\nprogress: 56.0 s, 8298.1 tps, lat 3.848 ms stddev 2.727\\nprogress: 57.0 s, 8369.5 tps, lat 3.825 ms stddev 2.897\\nprogress: 58.0 s, 8205.0 tps, lat 3.902 ms stddev 2.669\\nprogress: 59.0 s, 8048.1 tps, lat 3.974 ms stddev 2.473\\nprogress: 60.0 s, 7886.5 tps, lat 4.050 ms stddev 2.456\\nprogress: 61.0 s, 7539.0 tps, lat 4.251 ms stddev 2.992\\nprogress: 62.0 s, 7815.4 tps, lat 4.090 ms stddev 2.845\\nprogress: 63.0 s, 7863.4 tps, lat 4.062 ms stddev 3.032\\nprogress: 64.0 s, 7750.6 tps, lat 4.135 ms stddev 2.677\\nprogress: 65.0 s, 7580.7 tps, lat 4.217 ms stddev 2.210\\nprogress: 66.0 s, 7164.8 tps, lat 4.466 ms stddev 2.511\\nprogress: 67.0 s, 7504.4 tps, lat 4.265 ms stddev 1.343\\nprogress: 68.0 s, 9530.0 tps, lat 3.360 ms stddev 1.742\\nprogress: 69.0 s, 16751.3 tps, lat 1.909 ms stddev 0.652\\nprogress: 70.0 s, 16123.7 tps, lat 1.983 ms stddev 0.751\\nprogress: 71.0 s, 15122.7 tps, lat 2.115 ms stddev 1.065\\n\\nYes. That's right. ANALYZE caused a slowdown of nearly *2. Executing it\\non a bigger database where it takes longer makes it much worse. Also, a\\nnonuniform access pattern also makes it *much* worse.\\n\\nSo, you might ask, why is that? It's because ANALYZE keeps a snapshot\\naround and thus blocks RecentGlobalXmin from advancing. Leading to hot\\npruning being impossible on \\\"popular\\\" pages.\\n\\nNow, the update rates in this example certainly are above what most\\nsites see, but in practical sites relations are also much larger. And\\nthus autoanalyze also takes longer... Much longer with a large sample\\nsize.\\n\\nInterestingly we already set PROC_IN_ANALYZE - but we never actually\\nlook at it. I wonder what it'd take to ignore analyze in\\nGetSnapshotData()/GetOldestXmin()... At first sight it looks quite\\ndoable to not have a snapshot around (or setup to be ignored) while\\nacquire_sample_rows() is running. As that's the expensive bit...\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 19:49:09 +0200\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"(auto-)analyze causing bloat/load\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@2ndquadrant.com> writes:\\n> Interestingly we already set PROC_IN_ANALYZE - but we never actually\\n> look at it. I wonder what it'd take to ignore analyze in\\n> GetSnapshotData()/GetOldestXmin()... At first sight it looks quite\\n> doable to not have a snapshot around (or setup to be ignored) while\\n> acquire_sample_rows() is running. As that's the expensive bit...\\n\\nSay what? How is ANALYZE supposed to know which rows are valid\\nwithout a snapshot?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 14:30:53 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: (auto-)analyze causing bloat/load\"\n },\n {\n \"msg_contents\": \"On 2014-10-18 14:30:53 -0400, Tom Lane wrote:\\n> Andres Freund <andres@2ndquadrant.com> writes:\\n> > Interestingly we already set PROC_IN_ANALYZE - but we never actually\\n> > look at it. I wonder what it'd take to ignore analyze in\\n> > GetSnapshotData()/GetOldestXmin()... At first sight it looks quite\\n> > doable to not have a snapshot around (or setup to be ignored) while\\n> > acquire_sample_rows() is running. As that's the expensive bit...\\n> \\n> Say what? How is ANALYZE supposed to know which rows are valid\\n> without a snapshot?\\n\\nI haven't fully investigated it yet. But it's already not using it's own\\nsnapshot to determine visibility. Instead it uses GetOldestXmin() +\\nHeapTupleSatisfiesVacuum().\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 20:33:57 +0200\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: (auto-)analyze causing bloat/load\"\n }\n]","isConversational":false}}},{"rowIdx":39683,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nOne of our sites ran into a problem with database installation:\n\ninitdb failed\nFATAL: unexpected data beyond EOF in block 19 of relation base/1/2609\nHINT: This has been seen to occur with buggy kernels; consider updating your system.\nSTATEMENT: COMMENT ON FUNCTION euc_jis_2004_to_shift_jis_2004\n (INTEGER, INTEGER, CSTRING, INTERNAL, INTEGER)\n IS 'internal conversion function for EUC_JIS_2004 to SHIFT_JIS_2004';\n\n\ninitdb is called like this:\n initdb -D <data-dir> -L <input-dir> -E UTF8 --locale=C\n\nThis is Postgres 8.4.4, the installation piece has been stable and always worked, but this time they have a new Red Hat 5.10 server\n\n# uname -a\nLinux <hostname> 2.6.18-371.el5 #1 SMP Thu Sep 5 21:21:44 EDT 2013 x86_64 x86_64 x86_64 GNU/Linux\n\n# cat /etc/redhat-release\nRed Hat Enterprise Linux Server release 5.10 (Tikanga)\n\nI am not sure if this is helpful, but just in case:\n# ldd euc_jis_2004_and_shift_jis_2004.so\n linux-vdso.so.1 => (0x00007fff259fd000)\n libc.so.6 => /lib64/libc.so.6 (0x00002b3d5a756000)\n /lib64/ld-linux-x86-64.so.2 (0x0000003848400000)\n\nAnd these libs are softlinks:\n/lib64/libc.so.6 -> libc-2.5.so\n/lib64/ld-linux-x86-64.so.2 -> ld-2.5.so\n\nThe only thought I have at this point is to run it with strace, but maybe this is a known issue and someone has a better idea?\n\nThank you,\nMichael.\n\n\n\n\n\n\n\nThis email and any attachments are intended solely for the use of the individual or entity to whom it is addressed and may be confidential and/or privileged.\n\nIf you are not one of the named recipients or have received this email in error,\n\n(i) you should not read, disclose, or copy it,\n\n(ii) please notify sender of your receipt by reply email and delete this email and all attachments,\n\n(iii) Dassault Systemes does not accept or assume any liability or responsibility for any use of or reliance on this email.\n\nFor other languages, go to http://www.3ds.com/terms/email-disclaimer\n\n\n\n\n\n\n\n\n\nHi,\n \nOne of our sites ran into a problem with database installation:\n \ninitdb failed \nFATAL: unexpected data beyond EOF in block 19 of relation base/1/2609\n\nHINT: This has been seen to occur with buggy kernels; consider updating your system.\n\nSTATEMENT: COMMENT ON FUNCTION euc_jis_2004_to_shift_jis_2004\n (INTEGER, INTEGER, CSTRING, INTERNAL, INTEGER)\n\n IS 'internal conversion function for EUC_JIS_2004 to SHIFT_JIS_2004';\n\n \n \ninitdb is called like this:\n initdb -D <data-dir> -L <input-dir> -E UTF8 --locale=C\n\n \nThis is Postgres 8.4.4, the installation piece has been stable and always worked, but this time they have a new Red Hat 5.10 server\n \n# uname -a\n\nLinux <hostname> 2.6.18-371.el5 #1 SMP Thu Sep 5 21:21:44 EDT 2013 x86_64 x86_64 x86_64 GNU/Linux\n\n \n# cat /etc/redhat-release\n\nRed Hat Enterprise Linux Server release 5.10 (Tikanga)\n\n \nI am not sure if this is helpful, but just in case:\n# ldd euc_jis_2004_and_shift_jis_2004.so\n\n linux-vdso.so.1 => (0x00007fff259fd000)\n\n libc.so.6 => /lib64/libc.so.6 (0x00002b3d5a756000)\n\n /lib64/ld-linux-x86-64.so.2 (0x0000003848400000)\n\n \nAnd these libs are softlinks:\n/lib64/libc.so.6 -> libc-2.5.so\n/lib64/ld-linux-x86-64.so.2 -> ld-2.5.so\n \nThe only thought I have at this point is to run it with strace, but maybe this is a known issue and someone has a better idea?\n \nThank you,\nMichael.\n \n \n \n \n \n \n\n\nThis email and any attachments are intended solely for the use of the individual or entity to whom it is addressed and may be confidential and/or privileged.\n\nIf you are not one of the named recipients or have received this email in error, \n\n(i) you should not read, disclose, or copy it,\n\n(ii) please notify sender of your receipt by reply email and delete this email and all attachments,\n\n(iii) Dassault Systemes does not accept or assume any liability or responsibility for any use of or reliance on this email.\n\n\n\nFor other languages, go to http://www.3ds.com/terms/email-disclaimer","msg_date":"Sat, 18 Oct 2014 18:21:22 +0000","msg_from":"BRUSSER Michael <Michael.BRUSSER@3ds.com>","msg_from_op":true,"msg_subject":"initdb failure on RH 5.10"},{"msg_contents":"BRUSSER Michael <Michael.BRUSSER@3ds.com> writes:\n> initdb failed\n> FATAL: unexpected data beyond EOF in block 19 of relation base/1/2609\n> HINT: This has been seen to occur with buggy kernels; consider updating your system.\n> STATEMENT: COMMENT ON FUNCTION euc_jis_2004_to_shift_jis_2004\n> (INTEGER, INTEGER, CSTRING, INTERNAL, INTEGER)\n> IS 'internal conversion function for EUC_JIS_2004 to SHIFT_JIS_2004';\n\nThat's ... surprising. We've only ever seen that error message in cases\nwith heavy concurrent updates and wonky underlying storage; initdb is not\nwhere anyone would expect it.\n\n> initdb is called like this:\n> initdb -D <data-dir> -L <input-dir> -E UTF8 --locale=C\n\nIt's not exactly customary to use -L in initdb calls. Is it possible\nyou're pointing it to an incompatible library directory? Not that I see\nhow that would lead to this behavior, but you're definitely dealing with\nsomething pretty weird.\n\n> This is Postgres 8.4.4, the installation piece has been stable and always worked, but this time they have a new Red Hat 5.10 server\n\nWhat in the world are they doing using 8.4.4? The entire 8.4.x release\nseries is out of support anyway, but there is little if any excuse not\nto be using the last minor release, 8.4.22.\n\nI'd call your attention also to the fact that RHEL 5.10 is obsolete.\n5.11 came out last month, and Red Hat are not known for updating back-rev\nrelease series with inessential bug fixes.\n\nIf you can still reproduce this with 5.11 and 8.4.22, people might be\ninterested in looking more closely. Otherwise, well, you're dealing\nwith five-year-old software with a very long list of known bugs.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 18 Oct 2014 14:54:12 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: initdb failure on RH 5.10"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nOne of our sites ran into a problem with database installation:\\n\\ninitdb failed\\nFATAL: unexpected data beyond EOF in block 19 of relation base/1/2609\\nHINT: This has been seen to occur with buggy kernels; consider updating your system.\\nSTATEMENT: COMMENT ON FUNCTION euc_jis_2004_to_shift_jis_2004\\n (INTEGER, INTEGER, CSTRING, INTERNAL, INTEGER)\\n IS 'internal conversion function for EUC_JIS_2004 to SHIFT_JIS_2004';\\n\\n\\ninitdb is called like this:\\n initdb -D <data-dir> -L <input-dir> -E UTF8 --locale=C\\n\\nThis is Postgres 8.4.4, the installation piece has been stable and always worked, but this time they have a new Red Hat 5.10 server\\n\\n# uname -a\\nLinux <hostname> 2.6.18-371.el5 #1 SMP Thu Sep 5 21:21:44 EDT 2013 x86_64 x86_64 x86_64 GNU/Linux\\n\\n# cat /etc/redhat-release\\nRed Hat Enterprise Linux Server release 5.10 (Tikanga)\\n\\nI am not sure if this is helpful, but just in case:\\n# ldd euc_jis_2004_and_shift_jis_2004.so\\n linux-vdso.so.1 => (0x00007fff259fd000)\\n libc.so.6 => /lib64/libc.so.6 (0x00002b3d5a756000)\\n /lib64/ld-linux-x86-64.so.2 (0x0000003848400000)\\n\\nAnd these libs are softlinks:\\n/lib64/libc.so.6 -> libc-2.5.so\\n/lib64/ld-linux-x86-64.so.2 -> ld-2.5.so\\n\\nThe only thought I have at this point is to run it with strace, but maybe this is a known issue and someone has a better idea?\\n\\nThank you,\\nMichael.\\n\\n\\n\\n\\n\\n\\n\\nThis email and any attachments are intended solely for the use of the individual or entity to whom it is addressed and may be confidential and/or privileged.\\n\\nIf you are not one of the named recipients or have received this email in error,\\n\\n(i) you should not read, disclose, or copy it,\\n\\n(ii) please notify sender of your receipt by reply email and delete this email and all attachments,\\n\\n(iii) Dassault Systemes does not accept or assume any liability or responsibility for any use of or reliance on this email.\\n\\nFor other languages, go to http://www.3ds.com/terms/email-disclaimer\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nHi,\\n \\nOne of our sites ran into a problem with database installation:\\n \\ninitdb failed \\nFATAL: unexpected data beyond EOF in block 19 of relation base/1/2609\\n\\nHINT: This has been seen to occur with buggy kernels; consider updating your system.\\n\\nSTATEMENT: COMMENT ON FUNCTION euc_jis_2004_to_shift_jis_2004\\n (INTEGER, INTEGER, CSTRING, INTERNAL, INTEGER)\\n\\n IS 'internal conversion function for EUC_JIS_2004 to SHIFT_JIS_2004';\\n\\n \\n \\ninitdb is called like this:\\n initdb -D <data-dir> -L <input-dir> -E UTF8 --locale=C\\n\\n \\nThis is Postgres 8.4.4, the installation piece has been stable and always worked, but this time they have a new Red Hat 5.10 server\\n \\n# uname -a\\n\\nLinux <hostname> 2.6.18-371.el5 #1 SMP Thu Sep 5 21:21:44 EDT 2013 x86_64 x86_64 x86_64 GNU/Linux\\n\\n \\n# cat /etc/redhat-release\\n\\nRed Hat Enterprise Linux Server release 5.10 (Tikanga)\\n\\n \\nI am not sure if this is helpful, but just in case:\\n# ldd euc_jis_2004_and_shift_jis_2004.so\\n\\n linux-vdso.so.1 => (0x00007fff259fd000)\\n\\n libc.so.6 => /lib64/libc.so.6 (0x00002b3d5a756000)\\n\\n /lib64/ld-linux-x86-64.so.2 (0x0000003848400000)\\n\\n \\nAnd these libs are softlinks:\\n/lib64/libc.so.6 -> libc-2.5.so\\n/lib64/ld-linux-x86-64.so.2 -> ld-2.5.so\\n \\nThe only thought I have at this point is to run it with strace, but maybe this is a known issue and someone has a better idea?\\n \\nThank you,\\nMichael.\\n \\n \\n \\n \\n \\n \\n\\n\\nThis email and any attachments are intended solely for the use of the individual or entity to whom it is addressed and may be confidential and/or privileged.\\n\\nIf you are not one of the named recipients or have received this email in error, \\n\\n(i) you should not read, disclose, or copy it,\\n\\n(ii) please notify sender of your receipt by reply email and delete this email and all attachments,\\n\\n(iii) Dassault Systemes does not accept or assume any liability or responsibility for any use of or reliance on this email.\\n\\n\\n\\nFor other languages, go to http://www.3ds.com/terms/email-disclaimer\",\n \"msg_date\": \"Sat, 18 Oct 2014 18:21:22 +0000\",\n \"msg_from\": \"BRUSSER Michael <Michael.BRUSSER@3ds.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"initdb failure on RH 5.10\"\n },\n {\n \"msg_contents\": \"BRUSSER Michael <Michael.BRUSSER@3ds.com> writes:\\n> initdb failed\\n> FATAL: unexpected data beyond EOF in block 19 of relation base/1/2609\\n> HINT: This has been seen to occur with buggy kernels; consider updating your system.\\n> STATEMENT: COMMENT ON FUNCTION euc_jis_2004_to_shift_jis_2004\\n> (INTEGER, INTEGER, CSTRING, INTERNAL, INTEGER)\\n> IS 'internal conversion function for EUC_JIS_2004 to SHIFT_JIS_2004';\\n\\nThat's ... surprising. We've only ever seen that error message in cases\\nwith heavy concurrent updates and wonky underlying storage; initdb is not\\nwhere anyone would expect it.\\n\\n> initdb is called like this:\\n> initdb -D <data-dir> -L <input-dir> -E UTF8 --locale=C\\n\\nIt's not exactly customary to use -L in initdb calls. Is it possible\\nyou're pointing it to an incompatible library directory? Not that I see\\nhow that would lead to this behavior, but you're definitely dealing with\\nsomething pretty weird.\\n\\n> This is Postgres 8.4.4, the installation piece has been stable and always worked, but this time they have a new Red Hat 5.10 server\\n\\nWhat in the world are they doing using 8.4.4? The entire 8.4.x release\\nseries is out of support anyway, but there is little if any excuse not\\nto be using the last minor release, 8.4.22.\\n\\nI'd call your attention also to the fact that RHEL 5.10 is obsolete.\\n5.11 came out last month, and Red Hat are not known for updating back-rev\\nrelease series with inessential bug fixes.\\n\\nIf you can still reproduce this with 5.11 and 8.4.22, people might be\\ninterested in looking more closely. Otherwise, well, you're dealing\\nwith five-year-old software with a very long list of known bugs.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 18 Oct 2014 14:54:12 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: initdb failure on RH 5.10\"\n }\n]","isConversational":false}}},{"rowIdx":39684,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Buildfarm member hamerkop has been failing in the pg_upgrade regression\ntest for the last several days. The problem looks like this:\n\ncommand: \"C:/buildfarm/build_root/HEAD/pgsql.build/contrib/pg_upgrade/tmp_check/install/bin/pg_restore\" --port 50432 --username \"Administrator\" --exit-on-error --verbose --dbname \"postgres\" \"pg_upgrade_dump_12145.custom\" >> \"pg_upgrade_dump_12145.log\" 2>&1\npg_restore: connecting to database for restore\npg_restore: [archiver (db)] Error while INITIALIZING:\npg_restore: [archiver (db)] could not execute query: ERROR: invalid byte sequence for encoding \"UTF8\": 0x93\n\nI can't help noticing that this started immediately after commit\n0eea804 \"pg_dump: Reduce use of global variables\". No idea why\nthe issue is only showing up on this one animal.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 19 Oct 2014 14:31:13 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"pg_dump/pg_restore seem broken on hamerkop"},{"msg_contents":"> Buildfarm member hamerkop has been failing in the pg_upgrade regression\n> test for the last several days. The problem looks like this:\n> \n> command: \"C:/buildfarm/build_root/HEAD/pgsql.build/contrib/pg_upgrade/tmp_check/install/bin/pg_restore\" --port 50432 --username \"Administrator\" --exit-on-error --verbose --dbname \"postgres\" \"pg_upgrade_dump_12145.custom\" >> \"pg_upgrade_dump_12145.log\" 2>&1\n> pg_restore: connecting to database for restore\n> pg_restore: [archiver (db)] Error while INITIALIZING:\n> pg_restore: [archiver (db)] could not execute query: ERROR: invalid byte sequence for encoding \"UTF8\": 0x93\n> \n> I can't help noticing that this started immediately after commit\n> 0eea804 \"pg_dump: Reduce use of global variables\". No idea why\n> the issue is only showing up on this one animal.\n\nI guess one of possibilities is there's garbage in memory which is\nrelated to restore the process. If so, coverity may be able to find\nsomething. The last coverity analysis was Oct 12. The commit 0eea804\nwas made on Oct 14. So let's see what new coverity scan reports...\n\nBest regards,\n--\nTatsuo Ishii\nSRA OSS, Inc. Japan\nEnglish: http://www.sraoss.co.jp/index_en.php\nJapanese:http://www.sraoss.co.jp\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 08:30:30 +0900 (JST)","msg_from":"Tatsuo Ishii <ishii@postgresql.org>","msg_from_op":false,"msg_subject":"Re: pg_dump/pg_restore seem broken on hamerkop"},{"msg_contents":">> Buildfarm member hamerkop has been failing in the pg_upgrade regression\n>> test for the last several days. The problem looks like this:\n>> \n>> command: \"C:/buildfarm/build_root/HEAD/pgsql.build/contrib/pg_upgrade/tmp_check/install/bin/pg_restore\" --port 50432 --username \"Administrator\" --exit-on-error --verbose --dbname \"postgres\" \"pg_upgrade_dump_12145.custom\" >> \"pg_upgrade_dump_12145.log\" 2>&1\n>> pg_restore: connecting to database for restore\n>> pg_restore: [archiver (db)] Error while INITIALIZING:\n>> pg_restore: [archiver (db)] could not execute query: ERROR: invalid byte sequence for encoding \"UTF8\": 0x93\n>> \n>> I can't help noticing that this started immediately after commit\n>> 0eea804 \"pg_dump: Reduce use of global variables\". No idea why\n>> the issue is only showing up on this one animal.\n> \n> I guess one of possibilities is there's garbage in memory which is\n> related to restore the process. If so, coverity may be able to find\n> something. The last coverity analysis was Oct 12. The commit 0eea804\n> was made on Oct 14. So let's see what new coverity scan reports...\n\nThe latest coverity scan report did not include paticular defects\nrelated to pg_dump/pg_restore...\n\nBest regards,\n--\nTatsuo Ishii\nSRA OSS, Inc. Japan\nEnglish: http://www.sraoss.co.jp/index_en.php\nJapanese:http://www.sraoss.co.jp\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 27 Oct 2014 07:56:33 +0900 (JST)","msg_from":"Tatsuo Ishii <ishii@postgresql.org>","msg_from_op":false,"msg_subject":"Re: pg_dump/pg_restore seem broken on hamerkop"},{"msg_contents":"Tatsuo Ishii wrote:\n> > Buildfarm member hamerkop has been failing in the pg_upgrade regression\n> > test for the last several days. The problem looks like this:\n> > \n> > command: \"C:/buildfarm/build_root/HEAD/pgsql.build/contrib/pg_upgrade/tmp_check/install/bin/pg_restore\" --port 50432 --username \"Administrator\" --exit-on-error --verbose --dbname \"postgres\" \"pg_upgrade_dump_12145.custom\" >> \"pg_upgrade_dump_12145.log\" 2>&1\n> > pg_restore: connecting to database for restore\n> > pg_restore: [archiver (db)] Error while INITIALIZING:\n> > pg_restore: [archiver (db)] could not execute query: ERROR: invalid byte sequence for encoding \"UTF8\": 0x93\n> > \n> > I can't help noticing that this started immediately after commit\n> > 0eea804 \"pg_dump: Reduce use of global variables\". No idea why\n> > the issue is only showing up on this one animal.\n> \n> I guess one of possibilities is there's garbage in memory which is\n> related to restore the process.\n\nThat sounds most likely. The complete error in hamerkop's log is:\n\npg_restore: connecting to database for restore\npg_restore: [archiver (db)] Error while INITIALIZING:\npg_restore: [archiver (db)] could not execute query: ERROR: invalid byte sequence for encoding \"UTF8\": 0x93\n Command was: -- Started on 2014-10-26 03:06:00 “Œ‹ž (•W€Žž)\n\nThis \"Started on\" business comes from pg_backup_archiver.c, which has\n\n if (AH->public.verbose)\n dumpTimestamp(AH, \"Started on\", AH->createDate);\n\nwhere dumpTimestamp is\n\n/*\n * dumpTimestamp\n */\nstatic void\ndumpTimestamp(ArchiveHandle *AH, const char *msg, time_t tim)\n{\n char buf[64];\n\n if (strftime(buf, sizeof(buf), \"%Y-%m-%d %H:%M:%S %z\", localtime(&tim)) != 0)\n ahprintf(AH, \"-- %s %s\\n\\n\", msg, buf);\n}\n\nSo this seems related to the %z part of the strftime() call. I have no\nexplanation for this failure ATM; maybe pg_restore is failing to set the\nlocale properly? I also notice pg_restore.c previously included\npg_backup_archiver.h (which in turn includes <time.h>); strftime\nrequires <time.h> so maybe this is causing a problem, but since\npg_restore.c itself is not calling strftime, I don't see how this would\nbe related.\n\n[Some more code and git-log reading later] I see that the %z is a very\nrecent addition: it only got there as of commit ad5d46a449, of September\n5th ... and now I also see that hamerkop's last green run before the\nfailure, on Oct 13rd, did *not* include the pg_upgrade check. So I'm\nthinking this was broken much earlier than 0eea804.\n\n-- \nÁlvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 26 Oct 2014 20:56:20 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: pg_dump/pg_restore seem broken on hamerkop"},{"msg_contents":"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n> So this seems related to the %z part of the strftime() call. I have no\n> explanation for this failure ATM; maybe pg_restore is failing to set the\n> locale properly? I also notice pg_restore.c previously included\n> pg_backup_archiver.h (which in turn includes <time.h>); strftime\n> requires <time.h> so maybe this is causing a problem, but since\n> pg_restore.c itself is not calling strftime, I don't see how this would\n> be related.\n\nHm. %z ought not be locale-dependent ... however, it has a bigger\nproblem, which is that it's a C99-ism. It's not there in SUSv2,\nwhich is our normal baseline for what's portable. I think we need\nto get rid of that. %Z should be portable.\n\n(Is it possible that Windows' strftime() reads %z as doing something\nother than what C99 says?)\n\n> [Some more code and git-log reading later] I see that the %z is a very\n> recent addition: it only got there as of commit ad5d46a449, of September\n> 5th ... and now I also see that hamerkop's last green run before the\n> failure, on Oct 13rd, did *not* include the pg_upgrade check. So I'm\n> thinking this was broken much earlier than 0eea804.\n\nOoohh ... you are right, the first failing build involved not only\nthe pg_dump refactoring commit, but an upgrade in the buildfarm script\nthat hamerkop was using (from 4.4 to 4.14). So it's entirely possible\nthis issue was already there and we just weren't seeing it tested.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 26 Oct 2014 20:11:49 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: pg_dump/pg_restore seem broken on hamerkop"},{"msg_contents":"I wrote:\n> Hm. %z ought not be locale-dependent ... however, it has a bigger\n> problem, which is that it's a C99-ism. It's not there in SUSv2,\n> which is our normal baseline for what's portable. I think we need\n> to get rid of that. %Z should be portable.\n\n> (Is it possible that Windows' strftime() reads %z as doing something\n> other than what C99 says?)\n\nA bit of googling leads me to Microsoft reference material saying that\ntheir strftime treats %z and %Z alike. So in point of fact, the\nassumption underlying commit ad5d46a4494b0b48 was flat out wrong.\nSwitching to %z doesn't get you out of the problem noted in the\ncomments it removed:\n\n /*\n * We don't print the timezone on Win32, because the names are long and\n * localized, which means they may contain characters in various random\n * encodings; this has been seen to cause encoding errors when reading the\n * dump script.\n */\n\nI'm going to go revert most of that commit and make the code like it\nwas before:\n\n if (strftime(buf, sizeof(buf),\n#ifndef WIN32\n \"%Y-%m-%d %H:%M:%S %Z\",\n#else\n \"%Y-%m-%d %H:%M:%S\",\n#endif\n localtime(&now)) != 0)\n\nIf somebody wants to have timezones printed on Windows, they can\ntry again, but they're gonna have to work harder than this.\n\nIt's possible that fixing this will not fix whatever's bothering\nhamerkop, but given the lack of backward-portability of %z, this\ncode has got to go anyway.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 26 Oct 2014 20:29:19 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: pg_dump/pg_restore seem broken on hamerkop"},{"msg_contents":"I wrote:\n> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n>> [Some more code and git-log reading later] I see that the %z is a very\n>> recent addition: it only got there as of commit ad5d46a449, of September\n>> 5th ... and now I also see that hamerkop's last green run before the\n>> failure, on Oct 13rd, did *not* include the pg_upgrade check. So I'm\n>> thinking this was broken much earlier than 0eea804.\n\n> Ooohh ... you are right, the first failing build involved not only\n> the pg_dump refactoring commit, but an upgrade in the buildfarm script\n> that hamerkop was using (from 4.4 to 4.14). So it's entirely possible\n> this issue was already there and we just weren't seeing it tested.\n\nhamerkop is still failing in its runs done today. However, I'm not sure\nthat that proves anything about our hoped-for fix, because the commit SHAs\nit's showing on the buildfarm status page are still from Oct 13. It looks\nlike hamerkop has failed to pull any git updates since it was migrated to\nthe new buildfarm script version.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 29 Oct 2014 00:26:40 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: pg_dump/pg_restore seem broken on hamerkop"},{"msg_contents":"\nOn 10/29/2014 12:26 AM, Tom Lane wrote:\n> I wrote:\n>> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n>>> [Some more code and git-log reading later] I see that the %z is a very\n>>> recent addition: it only got there as of commit ad5d46a449, of September\n>>> 5th ... and now I also see that hamerkop's last green run before the\n>>> failure, on Oct 13rd, did *not* include the pg_upgrade check. So I'm\n>>> thinking this was broken much earlier than 0eea804.\n>> Ooohh ... you are right, the first failing build involved not only\n>> the pg_dump refactoring commit, but an upgrade in the buildfarm script\n>> that hamerkop was using (from 4.4 to 4.14). So it's entirely possible\n>> this issue was already there and we just weren't seeing it tested.\n> hamerkop is still failing in its runs done today. However, I'm not sure\n> that that proves anything about our hoped-for fix, because the commit SHAs\n> it's showing on the buildfarm status page are still from Oct 13. It looks\n> like hamerkop has failed to pull any git updates since it was migrated to\n> the new buildfarm script version.\n>\n> \t\t\t\n\n\nThis machine appears to be misconfigured for git. I will email the owner \nabout fixing it.\n\ncheers\n\nandrew\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 29 Oct 2014 09:47:14 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: pg_dump/pg_restore seem broken on hamerkop"},{"msg_contents":"On Sun, Oct 26, 2014 at 08:29:19PM -0400, Tom Lane wrote:\n> I wrote:\n> > Hm. %z ought not be locale-dependent ... however, it has a bigger\n> > problem, which is that it's a C99-ism. It's not there in SUSv2,\n> > which is our normal baseline for what's portable. I think we need\n> > to get rid of that. %Z should be portable.\n> \n> > (Is it possible that Windows' strftime() reads %z as doing something\n> > other than what C99 says?)\n> \n> A bit of googling leads me to Microsoft reference material saying that\n> their strftime treats %z and %Z alike. So in point of fact, the\n> assumption underlying commit ad5d46a4494b0b48 was flat out wrong.\n> Switching to %z doesn't get you out of the problem noted in the\n> comments it removed:\n> \n> /*\n> * We don't print the timezone on Win32, because the names are long and\n> * localized, which means they may contain characters in various random\n> * encodings; this has been seen to cause encoding errors when reading the\n> * dump script.\n> */\n> \n> I'm going to go revert most of that commit and make the code like it\n> was before:\n\nThanks. The major goal of the patch was to get a timezone designation\nin there, and you have done that were possible, which is non-Windows. \nYour C comment clearly explained why Windows is a problem, and\ncentralized the code.\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 27 Nov 2014 11:24:11 -0500","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: pg_dump/pg_restore seem broken on hamerkop"}],"string":"[\n {\n \"msg_contents\": \"Buildfarm member hamerkop has been failing in the pg_upgrade regression\\ntest for the last several days. The problem looks like this:\\n\\ncommand: \\\"C:/buildfarm/build_root/HEAD/pgsql.build/contrib/pg_upgrade/tmp_check/install/bin/pg_restore\\\" --port 50432 --username \\\"Administrator\\\" --exit-on-error --verbose --dbname \\\"postgres\\\" \\\"pg_upgrade_dump_12145.custom\\\" >> \\\"pg_upgrade_dump_12145.log\\\" 2>&1\\npg_restore: connecting to database for restore\\npg_restore: [archiver (db)] Error while INITIALIZING:\\npg_restore: [archiver (db)] could not execute query: ERROR: invalid byte sequence for encoding \\\"UTF8\\\": 0x93\\n\\nI can't help noticing that this started immediately after commit\\n0eea804 \\\"pg_dump: Reduce use of global variables\\\". No idea why\\nthe issue is only showing up on this one animal.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 19 Oct 2014 14:31:13 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pg_dump/pg_restore seem broken on hamerkop\"\n },\n {\n \"msg_contents\": \"> Buildfarm member hamerkop has been failing in the pg_upgrade regression\\n> test for the last several days. The problem looks like this:\\n> \\n> command: \\\"C:/buildfarm/build_root/HEAD/pgsql.build/contrib/pg_upgrade/tmp_check/install/bin/pg_restore\\\" --port 50432 --username \\\"Administrator\\\" --exit-on-error --verbose --dbname \\\"postgres\\\" \\\"pg_upgrade_dump_12145.custom\\\" >> \\\"pg_upgrade_dump_12145.log\\\" 2>&1\\n> pg_restore: connecting to database for restore\\n> pg_restore: [archiver (db)] Error while INITIALIZING:\\n> pg_restore: [archiver (db)] could not execute query: ERROR: invalid byte sequence for encoding \\\"UTF8\\\": 0x93\\n> \\n> I can't help noticing that this started immediately after commit\\n> 0eea804 \\\"pg_dump: Reduce use of global variables\\\". No idea why\\n> the issue is only showing up on this one animal.\\n\\nI guess one of possibilities is there's garbage in memory which is\\nrelated to restore the process. If so, coverity may be able to find\\nsomething. The last coverity analysis was Oct 12. The commit 0eea804\\nwas made on Oct 14. So let's see what new coverity scan reports...\\n\\nBest regards,\\n--\\nTatsuo Ishii\\nSRA OSS, Inc. Japan\\nEnglish: http://www.sraoss.co.jp/index_en.php\\nJapanese:http://www.sraoss.co.jp\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 08:30:30 +0900 (JST)\",\n \"msg_from\": \"Tatsuo Ishii <ishii@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_dump/pg_restore seem broken on hamerkop\"\n },\n {\n \"msg_contents\": \">> Buildfarm member hamerkop has been failing in the pg_upgrade regression\\n>> test for the last several days. The problem looks like this:\\n>> \\n>> command: \\\"C:/buildfarm/build_root/HEAD/pgsql.build/contrib/pg_upgrade/tmp_check/install/bin/pg_restore\\\" --port 50432 --username \\\"Administrator\\\" --exit-on-error --verbose --dbname \\\"postgres\\\" \\\"pg_upgrade_dump_12145.custom\\\" >> \\\"pg_upgrade_dump_12145.log\\\" 2>&1\\n>> pg_restore: connecting to database for restore\\n>> pg_restore: [archiver (db)] Error while INITIALIZING:\\n>> pg_restore: [archiver (db)] could not execute query: ERROR: invalid byte sequence for encoding \\\"UTF8\\\": 0x93\\n>> \\n>> I can't help noticing that this started immediately after commit\\n>> 0eea804 \\\"pg_dump: Reduce use of global variables\\\". No idea why\\n>> the issue is only showing up on this one animal.\\n> \\n> I guess one of possibilities is there's garbage in memory which is\\n> related to restore the process. If so, coverity may be able to find\\n> something. The last coverity analysis was Oct 12. The commit 0eea804\\n> was made on Oct 14. So let's see what new coverity scan reports...\\n\\nThe latest coverity scan report did not include paticular defects\\nrelated to pg_dump/pg_restore...\\n\\nBest regards,\\n--\\nTatsuo Ishii\\nSRA OSS, Inc. Japan\\nEnglish: http://www.sraoss.co.jp/index_en.php\\nJapanese:http://www.sraoss.co.jp\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 27 Oct 2014 07:56:33 +0900 (JST)\",\n \"msg_from\": \"Tatsuo Ishii <ishii@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_dump/pg_restore seem broken on hamerkop\"\n },\n {\n \"msg_contents\": \"Tatsuo Ishii wrote:\\n> > Buildfarm member hamerkop has been failing in the pg_upgrade regression\\n> > test for the last several days. The problem looks like this:\\n> > \\n> > command: \\\"C:/buildfarm/build_root/HEAD/pgsql.build/contrib/pg_upgrade/tmp_check/install/bin/pg_restore\\\" --port 50432 --username \\\"Administrator\\\" --exit-on-error --verbose --dbname \\\"postgres\\\" \\\"pg_upgrade_dump_12145.custom\\\" >> \\\"pg_upgrade_dump_12145.log\\\" 2>&1\\n> > pg_restore: connecting to database for restore\\n> > pg_restore: [archiver (db)] Error while INITIALIZING:\\n> > pg_restore: [archiver (db)] could not execute query: ERROR: invalid byte sequence for encoding \\\"UTF8\\\": 0x93\\n> > \\n> > I can't help noticing that this started immediately after commit\\n> > 0eea804 \\\"pg_dump: Reduce use of global variables\\\". No idea why\\n> > the issue is only showing up on this one animal.\\n> \\n> I guess one of possibilities is there's garbage in memory which is\\n> related to restore the process.\\n\\nThat sounds most likely. The complete error in hamerkop's log is:\\n\\npg_restore: connecting to database for restore\\npg_restore: [archiver (db)] Error while INITIALIZING:\\npg_restore: [archiver (db)] could not execute query: ERROR: invalid byte sequence for encoding \\\"UTF8\\\": 0x93\\n Command was: -- Started on 2014-10-26 03:06:00 “Œ‹ž (•W€Žž)\\n\\nThis \\\"Started on\\\" business comes from pg_backup_archiver.c, which has\\n\\n if (AH->public.verbose)\\n dumpTimestamp(AH, \\\"Started on\\\", AH->createDate);\\n\\nwhere dumpTimestamp is\\n\\n/*\\n * dumpTimestamp\\n */\\nstatic void\\ndumpTimestamp(ArchiveHandle *AH, const char *msg, time_t tim)\\n{\\n char buf[64];\\n\\n if (strftime(buf, sizeof(buf), \\\"%Y-%m-%d %H:%M:%S %z\\\", localtime(&tim)) != 0)\\n ahprintf(AH, \\\"-- %s %s\\\\n\\\\n\\\", msg, buf);\\n}\\n\\nSo this seems related to the %z part of the strftime() call. I have no\\nexplanation for this failure ATM; maybe pg_restore is failing to set the\\nlocale properly? I also notice pg_restore.c previously included\\npg_backup_archiver.h (which in turn includes <time.h>); strftime\\nrequires <time.h> so maybe this is causing a problem, but since\\npg_restore.c itself is not calling strftime, I don't see how this would\\nbe related.\\n\\n[Some more code and git-log reading later] I see that the %z is a very\\nrecent addition: it only got there as of commit ad5d46a449, of September\\n5th ... and now I also see that hamerkop's last green run before the\\nfailure, on Oct 13rd, did *not* include the pg_upgrade check. So I'm\\nthinking this was broken much earlier than 0eea804.\\n\\n-- \\nÁlvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 26 Oct 2014 20:56:20 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_dump/pg_restore seem broken on hamerkop\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> So this seems related to the %z part of the strftime() call. I have no\\n> explanation for this failure ATM; maybe pg_restore is failing to set the\\n> locale properly? I also notice pg_restore.c previously included\\n> pg_backup_archiver.h (which in turn includes <time.h>); strftime\\n> requires <time.h> so maybe this is causing a problem, but since\\n> pg_restore.c itself is not calling strftime, I don't see how this would\\n> be related.\\n\\nHm. %z ought not be locale-dependent ... however, it has a bigger\\nproblem, which is that it's a C99-ism. It's not there in SUSv2,\\nwhich is our normal baseline for what's portable. I think we need\\nto get rid of that. %Z should be portable.\\n\\n(Is it possible that Windows' strftime() reads %z as doing something\\nother than what C99 says?)\\n\\n> [Some more code and git-log reading later] I see that the %z is a very\\n> recent addition: it only got there as of commit ad5d46a449, of September\\n> 5th ... and now I also see that hamerkop's last green run before the\\n> failure, on Oct 13rd, did *not* include the pg_upgrade check. So I'm\\n> thinking this was broken much earlier than 0eea804.\\n\\nOoohh ... you are right, the first failing build involved not only\\nthe pg_dump refactoring commit, but an upgrade in the buildfarm script\\nthat hamerkop was using (from 4.4 to 4.14). So it's entirely possible\\nthis issue was already there and we just weren't seeing it tested.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 26 Oct 2014 20:11:49 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_dump/pg_restore seem broken on hamerkop\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Hm. %z ought not be locale-dependent ... however, it has a bigger\\n> problem, which is that it's a C99-ism. It's not there in SUSv2,\\n> which is our normal baseline for what's portable. I think we need\\n> to get rid of that. %Z should be portable.\\n\\n> (Is it possible that Windows' strftime() reads %z as doing something\\n> other than what C99 says?)\\n\\nA bit of googling leads me to Microsoft reference material saying that\\ntheir strftime treats %z and %Z alike. So in point of fact, the\\nassumption underlying commit ad5d46a4494b0b48 was flat out wrong.\\nSwitching to %z doesn't get you out of the problem noted in the\\ncomments it removed:\\n\\n /*\\n * We don't print the timezone on Win32, because the names are long and\\n * localized, which means they may contain characters in various random\\n * encodings; this has been seen to cause encoding errors when reading the\\n * dump script.\\n */\\n\\nI'm going to go revert most of that commit and make the code like it\\nwas before:\\n\\n if (strftime(buf, sizeof(buf),\\n#ifndef WIN32\\n \\\"%Y-%m-%d %H:%M:%S %Z\\\",\\n#else\\n \\\"%Y-%m-%d %H:%M:%S\\\",\\n#endif\\n localtime(&now)) != 0)\\n\\nIf somebody wants to have timezones printed on Windows, they can\\ntry again, but they're gonna have to work harder than this.\\n\\nIt's possible that fixing this will not fix whatever's bothering\\nhamerkop, but given the lack of backward-portability of %z, this\\ncode has got to go anyway.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 26 Oct 2014 20:29:19 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_dump/pg_restore seem broken on hamerkop\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n>> [Some more code and git-log reading later] I see that the %z is a very\\n>> recent addition: it only got there as of commit ad5d46a449, of September\\n>> 5th ... and now I also see that hamerkop's last green run before the\\n>> failure, on Oct 13rd, did *not* include the pg_upgrade check. So I'm\\n>> thinking this was broken much earlier than 0eea804.\\n\\n> Ooohh ... you are right, the first failing build involved not only\\n> the pg_dump refactoring commit, but an upgrade in the buildfarm script\\n> that hamerkop was using (from 4.4 to 4.14). So it's entirely possible\\n> this issue was already there and we just weren't seeing it tested.\\n\\nhamerkop is still failing in its runs done today. However, I'm not sure\\nthat that proves anything about our hoped-for fix, because the commit SHAs\\nit's showing on the buildfarm status page are still from Oct 13. It looks\\nlike hamerkop has failed to pull any git updates since it was migrated to\\nthe new buildfarm script version.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 29 Oct 2014 00:26:40 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: pg_dump/pg_restore seem broken on hamerkop\"\n },\n {\n \"msg_contents\": \"\\nOn 10/29/2014 12:26 AM, Tom Lane wrote:\\n> I wrote:\\n>> Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n>>> [Some more code and git-log reading later] I see that the %z is a very\\n>>> recent addition: it only got there as of commit ad5d46a449, of September\\n>>> 5th ... and now I also see that hamerkop's last green run before the\\n>>> failure, on Oct 13rd, did *not* include the pg_upgrade check. So I'm\\n>>> thinking this was broken much earlier than 0eea804.\\n>> Ooohh ... you are right, the first failing build involved not only\\n>> the pg_dump refactoring commit, but an upgrade in the buildfarm script\\n>> that hamerkop was using (from 4.4 to 4.14). So it's entirely possible\\n>> this issue was already there and we just weren't seeing it tested.\\n> hamerkop is still failing in its runs done today. However, I'm not sure\\n> that that proves anything about our hoped-for fix, because the commit SHAs\\n> it's showing on the buildfarm status page are still from Oct 13. It looks\\n> like hamerkop has failed to pull any git updates since it was migrated to\\n> the new buildfarm script version.\\n>\\n> \\t\\t\\t\\n\\n\\nThis machine appears to be misconfigured for git. I will email the owner \\nabout fixing it.\\n\\ncheers\\n\\nandrew\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 29 Oct 2014 09:47:14 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_dump/pg_restore seem broken on hamerkop\"\n },\n {\n \"msg_contents\": \"On Sun, Oct 26, 2014 at 08:29:19PM -0400, Tom Lane wrote:\\n> I wrote:\\n> > Hm. %z ought not be locale-dependent ... however, it has a bigger\\n> > problem, which is that it's a C99-ism. It's not there in SUSv2,\\n> > which is our normal baseline for what's portable. I think we need\\n> > to get rid of that. %Z should be portable.\\n> \\n> > (Is it possible that Windows' strftime() reads %z as doing something\\n> > other than what C99 says?)\\n> \\n> A bit of googling leads me to Microsoft reference material saying that\\n> their strftime treats %z and %Z alike. So in point of fact, the\\n> assumption underlying commit ad5d46a4494b0b48 was flat out wrong.\\n> Switching to %z doesn't get you out of the problem noted in the\\n> comments it removed:\\n> \\n> /*\\n> * We don't print the timezone on Win32, because the names are long and\\n> * localized, which means they may contain characters in various random\\n> * encodings; this has been seen to cause encoding errors when reading the\\n> * dump script.\\n> */\\n> \\n> I'm going to go revert most of that commit and make the code like it\\n> was before:\\n\\nThanks. The major goal of the patch was to get a timezone designation\\nin there, and you have done that were possible, which is non-Windows. \\nYour C comment clearly explained why Windows is a problem, and\\ncentralized the code.\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 27 Nov 2014 11:24:11 -0500\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_dump/pg_restore seem broken on hamerkop\"\n }\n]","isConversational":false}}},{"rowIdx":39685,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nCommit 32984d8fc3dbb90a3fafb69fece0134f1ea790f9 forgot to change the \nfilename in the comment in contrib/pgcrypto/pgcrypto--1.2.sql. Trivial \npatch attached.\n\n\n.marko\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Sun, 19 Oct 2014 23:24:18 +0200","msg_from":"Marko Tiikkaja <marko@joh.to>","msg_from_op":true,"msg_subject":"Wrong filename in comment"},{"msg_contents":"Marko Tiikkaja <marko@joh.to> writes:\n> Commit 32984d8fc3dbb90a3fafb69fece0134f1ea790f9 forgot to change the \n> filename in the comment in contrib/pgcrypto/pgcrypto--1.2.sql. Trivial \n> patch attached.\n\nPushed, thanks.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 10:54:15 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Wrong filename in comment"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nCommit 32984d8fc3dbb90a3fafb69fece0134f1ea790f9 forgot to change the \\nfilename in the comment in contrib/pgcrypto/pgcrypto--1.2.sql. Trivial \\npatch attached.\\n\\n\\n.marko\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Sun, 19 Oct 2014 23:24:18 +0200\",\n \"msg_from\": \"Marko Tiikkaja <marko@joh.to>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Wrong filename in comment\"\n },\n {\n \"msg_contents\": \"Marko Tiikkaja <marko@joh.to> writes:\\n> Commit 32984d8fc3dbb90a3fafb69fece0134f1ea790f9 forgot to change the \\n> filename in the comment in contrib/pgcrypto/pgcrypto--1.2.sql. Trivial \\n> patch attached.\\n\\nPushed, thanks.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 10:54:15 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Wrong filename in comment\"\n }\n]","isConversational":false}}},{"rowIdx":39686,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Currently, a non-freeze vacuum will punt on any page it can't get a cleanup lock on, with no retry. Presumably this should be a rare occurrence, but I think it's bad that we just assume that and won't warn the user if something bad is going on.\n\nMy thought is that if we skip any pages elog(LOG) how many we skipped. If we skip more than 1% of the pages we visited (not relpages) then elog(WARNING) instead.\n\nComments?\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 19 Oct 2014 20:57:36 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On Mon, Oct 20, 2014 at 2:57 AM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n> Currently, a non-freeze vacuum will punt on any page it can't get a cleanup\n> lock on, with no retry. Presumably this should be a rare occurrence, but I\n> think it's bad that we just assume that and won't warn the user if something\n> bad is going on.\n>\n> My thought is that if we skip any pages elog(LOG) how many we skipped. If we\n> skip more than 1% of the pages we visited (not relpages) then elog(WARNING)\n> instead.\n\nIs there some specific failure you've run into where a page was stuck\nin a pinned state and never got vacuumed?\n\nI would like to see a more systematic way of going about this. What\nLSN or timestamp is associated with the oldest unvacuumed page? How\nmany times have we tried to visit it? What do those numbers look like\noverall -- i.e. what's the median number of times it takes to vacuum a\npage and what does the distribution look like of the unvacuumed ages?\n\nWith that data it should be possible to determine if the behaviour is\nactually working well and where to draw the line to determine outliers\nthat might represent bugs.\n\n-- \ngreg\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 16:29:58 +0100","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 10/20/14, 10:29 AM, Greg Stark wrote:\n> On Mon, Oct 20, 2014 at 2:57 AM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n>> Currently, a non-freeze vacuum will punt on any page it can't get a cleanup\n>> lock on, with no retry. Presumably this should be a rare occurrence, but I\n>> think it's bad that we just assume that and won't warn the user if something\n>> bad is going on.\n>>\n>> My thought is that if we skip any pages elog(LOG) how many we skipped. If we\n>> skip more than 1% of the pages we visited (not relpages) then elog(WARNING)\n>> instead.\n>\n> Is there some specific failure you've run into where a page was stuck\n> in a pinned state and never got vacuumed?\n\nNot that I know of... but how would I actually know? Having that info available is the point of my proposal. :)\n\n> I would like to see a more systematic way of going about this. What\n> LSN or timestamp is associated with the oldest unvacuumed page? How\n> many times have we tried to visit it? What do those numbers look like\n> overall -- i.e. what's the median number of times it takes to vacuum a\n> page and what does the distribution look like of the unvacuumed ages?\n>\n> With that data it should be possible to determine if the behaviour is\n> actually working well and where to draw the line to determine outliers\n> that might represent bugs.\n\nI agree we could use better data about/for vacuum (see http://www.postgresql.org/message-id/544468C1.6050101@BlueTreble.com).\n\nIn the meantime, I think it's worth adding this logging. If in fact this basically never happens (the current assumption), it doesn't hurt anything. If it turns out our assumption is wrong, then we'll actually be able to find that out. :)\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 19:18:31 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-10-20 19:18:31 -0500, Jim Nasby wrote:\n> In the meantime, I think it's worth adding this logging. If in fact this basically never happens (the current assumption), it doesn't hurt anything. If it turns out our assumption is wrong, then we'll actually be able to fin> that out. :)\n\nIt does happen, and not infrequently. Just not enough pages to normally\ncause significant bloat. The most likely place where it happens is very\nsmall tables that all connections hit with a high frequency. Starting to\nissue high volume log spew for a nonexistant problem isn't helping.\n\nIf you're super convinced this is urgent then add it as a *single*\ndatapoint inside the existing messages. But I think there's loads of\nstuff in vacuum logging that are more important this.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 02:31:59 +0200","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 10/20/14, 7:31 PM, Andres Freund wrote:\n> On 2014-10-20 19:18:31 -0500, Jim Nasby wrote:\n>> >In the meantime, I think it's worth adding this logging. If in fact this basically never happens (the current assumption), it doesn't hurt anything. If it turns out our assumption is wrong, then we'll actually be able to fin> that out.:)\n> It does happen, and not infrequently. Just not enough pages to normally\n> cause significant bloat. The most likely place where it happens is very\n> small tables that all connections hit with a high frequency. Starting to\n> issue high volume log spew for a nonexistant problem isn't helping.\n\nHow'd you determine that? Is there some way to measure this? I'm not doubting you; I just don't want to work on a problem that's already solved.\n\n> If you're super convinced this is urgent then add it as a*single*\n> datapoint inside the existing messages. But I think there's loads of\n> stuff in vacuum logging that are more important this.\n\nSee my original proposal; at it's most intrusive this would issue one warning per (auto)vacuum if it was over a certain threshold. I would think that a DBA would really like to know when this happens, but if we think that's too much spew we can limit it to normal vacuum logging.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 19:43:38 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-10-20 19:43:38 -0500, Jim Nasby wrote:\n> On 10/20/14, 7:31 PM, Andres Freund wrote:\n> >On 2014-10-20 19:18:31 -0500, Jim Nasby wrote:\n> >>>In the meantime, I think it's worth adding this logging. If in fact this basically never happens (the current assumption), it doesn't hurt anything. If it turns out our assumption is wrong, then we'll actually be able to fin> that out.:)\n> >It does happen, and not infrequently. Just not enough pages to normally\n> >cause significant bloat. The most likely place where it happens is very\n> >small tables that all connections hit with a high frequency. Starting to\n> >issue high volume log spew for a nonexistant problem isn't helping.\n> \n> How'd you determine that? Is there some way to measure this?\n\nYou'd seen individual pages with too old dead rows in them.\n\n> >If you're super convinced this is urgent then add it as a*single*\n> >datapoint inside the existing messages. But I think there's loads of\n> >stuff in vacuum logging that are more important this.\n> \n> See my original proposal; at it's most intrusive this would issue one\n> warning per (auto)vacuum if it was over a certain threshold.\n\nWhich would vastly increase the log output for setups with small tables\nand a nonzero log_autovacuum_min_duration.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 02:49:55 +0200","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"Jim Nasby wrote:\n\n> Currently, a non-freeze vacuum will punt on any page it can't get a\n> cleanup lock on, with no retry. Presumably this should be a rare\n> occurrence, but I think it's bad that we just assume that and won't\n> warn the user if something bad is going on.\n\nI think if you really want to attack this problem, rather than just\nbeing noisy about it, what you could do is to keep a record of which\npage numbers you had to skip, and then once you're done with your first\nscan you go back and retry the lock on the pages you skipped.\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 19:39:48 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 10/21/14, 5:39 PM, Alvaro Herrera wrote:\n> Jim Nasby wrote:\n>\n>> Currently, a non-freeze vacuum will punt on any page it can't get a\n>> cleanup lock on, with no retry. Presumably this should be a rare\n>> occurrence, but I think it's bad that we just assume that and won't\n>> warn the user if something bad is going on.\n>\n> I think if you really want to attack this problem, rather than just\n> being noisy about it, what you could do is to keep a record of which\n> page numbers you had to skip, and then once you're done with your first\n> scan you go back and retry the lock on the pages you skipped.\n\nI'm OK with that if the community is; I was just trying for minimum invasiveness.\n\nIf I go this route, I'd like some input though...\n\n- How to handle storing the blockIDs. Fixed size array or something fancier? What should we limit it to, especially since we're already allocating maintenance_work_mem for the tid array.\n\n- What happens if we run out of space to remember skipped blocks? I could do something like what we do for running out of space in the dead_tuples array, but I'm worried that will add a serious amount of complexity, especially since re-processing these blocks could be what actually pushes us over the limit.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 18:00:07 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\n> - What happens if we run out of space to remember skipped blocks?\n\nYou forget some, and are no worse off than today. (This might be an\nevent worthy of logging, if the array is large enough that we don't\nexpect it to happen often ...)\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 19:05:23 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 10/21/14, 6:05 PM, Tom Lane wrote:\n> Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\n>> - What happens if we run out of space to remember skipped blocks?\n>\n> You forget some, and are no worse off than today. (This might be an\n> event worthy of logging, if the array is large enough that we don't\n> expect it to happen often ...)\n\nMakes sense. I'll see if there's some reasonable way to retry pages when the array fills up.\n\nI'll make the array 2k in size; that allows for 512 pages without spending a bunch of memory.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 29 Oct 2014 11:49:02 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 10/29/14, 11:49 AM, Jim Nasby wrote:\n> On 10/21/14, 6:05 PM, Tom Lane wrote:\n>> Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\n>>> - What happens if we run out of space to remember skipped blocks?\n>>\n>> You forget some, and are no worse off than today. (This might be an\n>> event worthy of logging, if the array is large enough that we don't\n>> expect it to happen often ...)\n>\n> Makes sense. I'll see if there's some reasonable way to retry pages when the array fills up.\n>\n> I'll make the array 2k in size; that allows for 512 pages without spending a bunch of memory.\n\nAttached is a patch for this. It also adds logging of unobtainable cleanup locks, and refactors scanning a page for vacuum into it's own function.\n\nAnyone reviewing this might want to look at https://github.com/decibel/postgres/commit/69ab22f703d577cbb3d8036e4e42563977bcf74b, which is the refactor with no whitespace changes.\n\nI've verified this works correctly by connecting to a backend with gdb and halting it with a page pinned. Both vacuum and vacuum freeze on that table do what's expected, but I also get this waring (which AFAICT is a false positive):\n\ndecibel@decina.local=# vacuum verbose i;\nINFO: vacuuming \"public.i\"\nINFO: \"i\": found 0 removable, 399774 nonremovable row versions in 1769 out of 1770 pages\nDETAIL: 200000 dead row versions cannot be removed yet.\nThere were 0 unused item pointers.\n0 pages are entirely empty.\nRetried cleanup lock on 0 pages, retry failed on 1, skipped retry on 0.\nCPU 0.00s/0.06u sec elapsed 12.89 sec.\nWARNING: buffer refcount leak: [105] (rel=base/16384/16385, blockNum=0, flags=0x106, refcount=2 1)\nVACUUM\n\nI am doing a simple static allocation of retry_pages[]; my understanding is that will only exist for the duration of this function so it's OK. If not I'll palloc it. If it is OK then I'll do the same for the freeze array.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Thu, 6 Nov 2014 14:55:37 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-11-06 14:55:37 -0600, Jim Nasby wrote:\n> On 10/29/14, 11:49 AM, Jim Nasby wrote:\n> >On 10/21/14, 6:05 PM, Tom Lane wrote:\n> >>Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\n> >>>- What happens if we run out of space to remember skipped blocks?\n> >>\n> >>You forget some, and are no worse off than today. (This might be an\n> >>event worthy of logging, if the array is large enough that we don't\n> >>expect it to happen often ...)\n> >\n> >Makes sense. I'll see if there's some reasonable way to retry pages when the array fills up.\n> >\n> >I'll make the array 2k in size; that allows for 512 pages without spending a bunch of memory.\n> \n> Attached is a patch for this. It also adds logging of unobtainable cleanup locks, and refactors scanning a page for vacuum into it's own function.\n> \n> Anyone reviewing this might want to look at https://github.com/decibel/postgres/commit/69ab22f703d577cbb3d8036e4e42563977bcf74b, which is the refactor with no whitespace changes.\n> \n> I've verified this works correctly by connecting to a backend with gdb and halting it with a page pinned. Both vacuum and vacuum freeze on that table do what's expected, but I also get this waring (which AFAICT is a false positive):\n\nI think the retry logical is a largely pointless complication of already\ncomplex enough code. You're fixing a problem for which there is\nabsolutely no evidence of its existance. Yes, this happens\noccasionally. But it's going to be so absolutely minor in comparison to\njust about every other source of bloat.\n\nSo, I pretty strongly against retrying. I could live with adding logging\nof the number of pages skipped due to not being able to acquire the\ncleanup lock. I don't think that's going to do help many people, but the\ncost is pretty low.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 6 Nov 2014 22:30:15 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On Thu, Nov 6, 2014 at 9:30 PM, Andres Freund <andres@2ndquadrant.com> wrote:\n> I think the retry logical is a largely pointless complication of already\n> complex enough code. You're fixing a problem for which there is\n> absolutely no evidence of its existance. Yes, this happens\n> occasionally. But it's going to be so absolutely minor in comparison to\n> just about every other source of bloat.\n\nI agree bloat isn't really a threat, but what about the relfrozenxid?\nIf we skip even one page we don't get to advance it and retrying could\neliminate those skipped pages and allow us to avoid a vacuum freeze\nwhich can be really painful. Of course that only works if you can be\nsure you haven't overflowed and forgotten any skipped pages and if you\ndon't find the page still pinned every time until you eventually give\nup on it.\n\n-- \ngreg\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 6 Nov 2014 23:40:18 +0000","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 11/6/14, 5:40 PM, Greg Stark wrote:\n> On Thu, Nov 6, 2014 at 9:30 PM, Andres Freund <andres@2ndquadrant.com> wrote:\n>> I think the retry logical is a largely pointless complication of already\n>> complex enough code. You're fixing a problem for which there is\n>> absolutely no evidence of its existance. Yes, this happens\n>> occasionally. But it's going to be so absolutely minor in comparison to\n>> just about every other source of bloat.\n\nFor some reason I don't have Andres' original email, so I'll reply here: I agree with you, and my original proposal was simply to log how many pages were skipped, but that was objected to. Simply logging this extra information would be a patch of a dozen lines or less.\n\nThe problem right now is there's no way to actually obtain evidence that this is (or isn't) something to worry about, because we just silently skip pages. If we had any kind of tracking on this we could stop guessing. :(\n\n> I agree bloat isn't really a threat, but what about the relfrozenxid?\n> If we skip even one page we don't get to advance it and retrying could\n> eliminate those skipped pages and allow us to avoid a vacuum freeze\n> which can be really painful. Of course that only works if you can be\n> sure you haven't overflowed and forgotten any skipped pages and if you\n> don't find the page still pinned every time until you eventually give\n> up on it.\n\nThe overflow part shouldn't be that big a deal. Either we just bump the array size up some more, or worst-case we scan it whenever it fills (like we do when we fill vacrelstats->dead_tuples.\n\nBut like I said above, I think this is already making a mountain out of a mole-hill, until we have evidence there's a real problem.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 6 Nov 2014 19:03:20 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-11-06 23:40:18 +0000, Greg Stark wrote:\n> On Thu, Nov 6, 2014 at 9:30 PM, Andres Freund <andres@2ndquadrant.com> wrote:\n> > I think the retry logical is a largely pointless complication of already\n> > complex enough code. You're fixing a problem for which there is\n> > absolutely no evidence of its existance. Yes, this happens\n> > occasionally. But it's going to be so absolutely minor in comparison to\n> > just about every other source of bloat.\n> \n> I agree bloat isn't really a threat, but what about the relfrozenxid?\n> If we skip even one page we don't get to advance it and retrying could\n> eliminate those skipped pages and allow us to avoid a vacuum freeze\n> which can be really painful. Of course that only works if you can be\n> sure you haven't overflowed and forgotten any skipped pages and if you\n> don't find the page still pinned every time until you eventually give\n> up on it.\n\nI don't buy this argument. Either you're constantly vacuuming the whole\nrelation anyway - in which case we'll acquire the cleanup lock\nunconditionally - or we're doing partial vacuums via the visibility map\nanyway.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 7 Nov 2014 08:27:19 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-11-06 19:03:20 -0600, Jim Nasby wrote:\n> On 11/6/14, 5:40 PM, Greg Stark wrote:\n> >On Thu, Nov 6, 2014 at 9:30 PM, Andres Freund <andres@2ndquadrant.com> wrote:\n> >>I think the retry logical is a largely pointless complication of already\n> >>complex enough code. You're fixing a problem for which there is\n> >>absolutely no evidence of its existance. Yes, this happens\n> >>occasionally. But it's going to be so absolutely minor in comparison to\n> >>just about every other source of bloat.\n> \n> For some reason I don't have Andres' original email, so I'll reply\n> here: I agree with you, and my original proposal was simply to log how\n> many pages were skipped, but that was objected to. Simply logging this\n> extra information would be a patch of a dozen lines or less.\n\nThe objection was that it's unneccessary complexity. So you made the\npatch a magnitude more complex *and* added logging? That doesn't make\nmuch sense.\n\n> The problem right now is there's no way to actually obtain evidence\n> that this is (or isn't) something to worry about, because we just\n> silently skip pages. If we had any kind of tracking on this we could\n> stop guessing. :(\n\nWhat's the worst consequence this could have? A couple pages not marked\nall visible and not immediately cleaned up. That's not particularly\nharmful.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 7 Nov 2014 08:29:19 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"Greg Stark wrote:\n\n> I agree bloat isn't really a threat, but what about the relfrozenxid?\n> If we skip even one page we don't get to advance it and retrying could\n> eliminate those skipped pages and allow us to avoid a vacuum freeze\n> which can be really painful. Of course that only works if you can be\n> sure you haven't overflowed and forgotten any skipped pages and if you\n> don't find the page still pinned every time until you eventually give\n> up on it.\n\nWe never advance relfrozenxid nowadays unless it's a whole-table scan;\nand once we commit to doing those (vacuum_freeze_table_age is past), we\ndon't skip pages anymore.\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 7 Nov 2014 10:32:03 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On Thu, Nov 6, 2014 at 8:03 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n> The problem right now is there's no way to actually obtain evidence that\n> this is (or isn't) something to worry about, because we just silently skip\n> pages. If we had any kind of tracking on this we could stop guessing. :(\n\nI could see logging it, but I agree with Andres and Alvaro that the\nodds are strongly against there being any actual problem here.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 7 Nov 2014 21:21:22 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 11/7/14, 8:21 PM, Robert Haas wrote:\n> On Thu, Nov 6, 2014 at 8:03 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n>> The problem right now is there's no way to actually obtain evidence that\n>> this is (or isn't) something to worry about, because we just silently skip\n>> pages. If we had any kind of tracking on this we could stop guessing. :(\n>\n> I could see logging it, but I agree with Andres and Alvaro that the\n> odds are strongly against there being any actual problem here.\n>\n\nI'm fine with that. Any other objections? Andres?\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 10 Nov 2014 11:09:08 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"Jim Nasby wrote:\n> On 11/7/14, 8:21 PM, Robert Haas wrote:\n> >On Thu, Nov 6, 2014 at 8:03 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n> >>The problem right now is there's no way to actually obtain evidence that\n> >>this is (or isn't) something to worry about, because we just silently skip\n> >>pages. If we had any kind of tracking on this we could stop guessing. :(\n> >\n> >I could see logging it, but I agree with Andres and Alvaro that the\n> >odds are strongly against there being any actual problem here.\n> \n> I'm fine with that. Any other objections? Andres?\n\nIf what we want is to quantify the extent of the issue, would it be more\nconvenient to save counters to pgstat? Vacuum already sends pgstat\nmessages, so there's no additional traffic there.\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 10 Nov 2014 14:28:30 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 11/10/14, 11:28 AM, Alvaro Herrera wrote:\n> Jim Nasby wrote:\n>> On 11/7/14, 8:21 PM, Robert Haas wrote:\n>>> On Thu, Nov 6, 2014 at 8:03 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n>>>> The problem right now is there's no way to actually obtain evidence that\n>>>> this is (or isn't) something to worry about, because we just silently skip\n>>>> pages. If we had any kind of tracking on this we could stop guessing. :(\n>>>\n>>> I could see logging it, but I agree with Andres and Alvaro that the\n>>> odds are strongly against there being any actual problem here.\n>>\n>> I'm fine with that. Any other objections? Andres?\n>\n> If what we want is to quantify the extent of the issue, would it be more\n> convenient to save counters to pgstat? Vacuum already sends pgstat\n> messages, so there's no additional traffic there.\n\nIMHO that would be ideal, but I think Tom was leery of using more space for every table. If we go this route, I'm guessing we should only log pages we skip, and not log pages we had to wait for the lock on (in the case of a freeze). Also, should we still eroprt this even if we are putting it in stats?\n\nIs there a way to avoid duplicating the entire eroprt call? I see I could call errstart & friends manually, but currently that's only done in elog.c.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 10 Nov 2014 12:03:59 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-11-10 14:28:30 -0300, Alvaro Herrera wrote:\n> Jim Nasby wrote:\n> > On 11/7/14, 8:21 PM, Robert Haas wrote:\n> > >On Thu, Nov 6, 2014 at 8:03 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n> > >>The problem right now is there's no way to actually obtain evidence that\n> > >>this is (or isn't) something to worry about, because we just silently skip\n> > >>pages. If we had any kind of tracking on this we could stop guessing. :(\n> > >\n> > >I could see logging it, but I agree with Andres and Alvaro that the\n> > >odds are strongly against there being any actual problem here.\n> > \n> > I'm fine with that. Any other objections? Andres?\n\nIf you feel that strong about the need for logging, go ahead.\n\n> If what we want is to quantify the extent of the issue, would it be more\n> convenient to save counters to pgstat? Vacuum already sends pgstat\n> messages, so there's no additional traffic there.\n\nI'm pretty strongly against that one in isolation. They'd need to be\nstored somewhere and they'd need to be queryable somewhere with enough\ncontext to make sense. To actually make sense of the numbers we'd also\nneed to report all the other datapoints of vacuum in some form. That's\nquite a worthwile project imo - but *much* *much* more work than this.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 10 Nov 2014 19:15:07 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"Andres Freund wrote:\n> On 2014-11-10 14:28:30 -0300, Alvaro Herrera wrote:\n\n> > If what we want is to quantify the extent of the issue, would it be more\n> > convenient to save counters to pgstat? Vacuum already sends pgstat\n> > messages, so there's no additional traffic there.\n> \n> I'm pretty strongly against that one in isolation. They'd need to be\n> stored somewhere and they'd need to be queryable somewhere with enough\n> context to make sense. To actually make sense of the numbers we'd also\n> need to report all the other datapoints of vacuum in some form. That's\n> quite a worthwile project imo - but *much* *much* more work than this.\n\nWe already have \"last_autovacuum\" columns and such in pg_stat_tables et\nal, which only record the last value. My thinking regarding such\nnumbers is that you would save histories and put them in a chart, see\nhow they evolve with time. I doubt the individual numbers are worth\nmuch, but the trends might show something interesting. As far as I\nknow, this is already true for most other pgstat values, with exception\nof things such as live_tuples which are absolute numbers rather than\nrunning counters.\n\nI agree having more vacuuming data in general is a worthwhile project,\nmuch larger than this one. Wasn't Greg Smith working on that?\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 10 Nov 2014 15:36:55 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 11/10/14, 12:15 PM, Andres Freund wrote:\n>> >If what we want is to quantify the extent of the issue, would it be more\n>> >convenient to save counters to pgstat? Vacuum already sends pgstat\n>> >messages, so there's no additional traffic there.\n> I'm pretty strongly against that one in isolation. They'd need to be\n> stored somewhere and they'd need to be queryable somewhere with enough\n> context to make sense. To actually make sense of the numbers we'd also\n> need to report all the other datapoints of vacuum in some form. That's\n> quite a worthwile project imo - but*much* *much* more work than this.\n\nWe already report statistics on vacuums (lazy_vacuum_rel()/pgstat_report_vacuum), so this would just be adding 1 or 2 counters to that. Should we add the other counters from vacuum? That would be significantly more data.\n\nSemi-related, we should probably be reporting stats from heap truncation.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 10 Nov 2014 12:37:29 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-11-10 15:36:55 -0300, Alvaro Herrera wrote:\n> Andres Freund wrote:\n> > On 2014-11-10 14:28:30 -0300, Alvaro Herrera wrote:\n> \n> > > If what we want is to quantify the extent of the issue, would it be more\n> > > convenient to save counters to pgstat? Vacuum already sends pgstat\n> > > messages, so there's no additional traffic there.\n> > \n> > I'm pretty strongly against that one in isolation. They'd need to be\n> > stored somewhere and they'd need to be queryable somewhere with enough\n> > context to make sense. To actually make sense of the numbers we'd also\n> > need to report all the other datapoints of vacuum in some form. That's\n> > quite a worthwile project imo - but *much* *much* more work than this.\n> \n> We already have \"last_autovacuum\" columns and such in pg_stat_tables et\n> al, which only record the last value. My thinking regarding such\n> numbers is that you would save histories and put them in a chart, see\n> how they evolve with time. I doubt the individual numbers are worth\n> much, but the trends might show something interesting.\n\nI don't think they mean anything without also reporting the number of\nbuffers actually scanned and other related stats.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 10 Nov 2014 19:52:21 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-11-10 12:37:29 -0600, Jim Nasby wrote:\n> On 11/10/14, 12:15 PM, Andres Freund wrote:\n> >>>If what we want is to quantify the extent of the issue, would it be more\n> >>>convenient to save counters to pgstat? Vacuum already sends pgstat\n> >>>messages, so there's no additional traffic there.\n> >I'm pretty strongly against that one in isolation. They'd need to be\n> >stored somewhere and they'd need to be queryable somewhere with enough\n> >context to make sense. To actually make sense of the numbers we'd also\n> >need to report all the other datapoints of vacuum in some form. That's\n> >quite a worthwile project imo - but*much* *much* more work than this.\n> \n> We already report statistics on vacuums\n> (lazy_vacuum_rel()/pgstat_report_vacuum), so this would just be adding\n> 1 or 2 counters to that. Should we add the other counters from vacuum?\n> That would be significantly more data.\n\nAt the very least it'd require:\n* The number of buffers skipped due to the vm\n* The number of buffers actually scanned\n* The number of full table in contrast to partial vacuums\n\nI think it'd require a fair amount of thinking about which values are\nrequired to make sense of the number of skipped buffers due to not being\nable to acquire the cleanup lock.\n\nIf you want to do this - and I sure don't want to stop you from it - you\nshould look at it from a general perspective, not from the perspective\nof how skipped cleanup locks are logged.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 10 Nov 2014 19:56:36 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 11/10/14, 12:56 PM, Andres Freund wrote:\n> On 2014-11-10 12:37:29 -0600, Jim Nasby wrote:\n>> On 11/10/14, 12:15 PM, Andres Freund wrote:\n>>>>> If what we want is to quantify the extent of the issue, would it be more\n>>>>> convenient to save counters to pgstat? Vacuum already sends pgstat\n>>>>> messages, so there's no additional traffic there.\n>>> I'm pretty strongly against that one in isolation. They'd need to be\n>>> stored somewhere and they'd need to be queryable somewhere with enough\n>>> context to make sense. To actually make sense of the numbers we'd also\n>>> need to report all the other datapoints of vacuum in some form. That's\n>>> quite a worthwile project imo - but*much* *much* more work than this.\n>>\n>> We already report statistics on vacuums\n>> (lazy_vacuum_rel()/pgstat_report_vacuum), so this would just be adding\n>> 1 or 2 counters to that. Should we add the other counters from vacuum?\n>> That would be significantly more data.\n>\n> At the very least it'd require:\n> * The number of buffers skipped due to the vm\n> * The number of buffers actually scanned\n> * The number of full table in contrast to partial vacuums\n\nIf we're going to track full scan vacuums separately, I think we'd need two separate scan counters. I think (for pgstats) it'd make more sense to just count initial failure to acquire the lock in a full scan in the 'skipped page' counter. In terms of answering the question \"how common is it not to get the lock\", it's really the same event.\n\n> I think it'd require a fair amount of thinking about which values are\n> required to make sense of the number of skipped buffers due to not being\n> able to acquire the cleanup lock.\n>\n> If you want to do this - and I sure don't want to stop you from it - you\n> should look at it from a general perspective, not from the perspective\n> of how skipped cleanup locks are logged.\n\nHonestly, my desire at this point is just to see if there's actually a problem. Many people are asserting that this should be a very rare occurrence, but there's no way to know.\n\nTowards that simple end, I'm a bit torn. My preference would be to simply log, and throw a warning if it's over some threshold. I believe that would give the best odds of getting feedback from users if this isn't as uncommon as we think.\n\nI agree that ideally this would be tracked as another stat, but from that standpoint I think there's other, much more important metrics to track, and AFAIK the only reason we don't have them is that busy systems already push pgstats to it's limits. We should try and fix that, but that's a much bigger issue.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 10 Nov 2014 19:36:18 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\n> On 11/10/14, 12:56 PM, Andres Freund wrote:\n>> If you want to do this - and I sure don't want to stop you from it - you\n>> should look at it from a general perspective, not from the perspective\n>> of how skipped cleanup locks are logged.\n\n> Honestly, my desire at this point is just to see if there's actually a problem. Many people are asserting that this should be a very rare occurrence, but there's no way to know.\n\n> Towards that simple end, I'm a bit torn. My preference would be to simply log, and throw a warning if it's over some threshold. I believe that would give the best odds of getting feedback from users if this isn't as uncommon as we think.\n\n> I agree that ideally this would be tracked as another stat, but from that standpoint I think there's other, much more important metrics to track, and AFAIK the only reason we don't have them is that busy systems already push pgstats to it's limits. We should try and fix that, but that's a much bigger issue.\n\nYeah. We know that per-table pgstat counters are a pretty expensive thing\nin databases with many tables. We should absolutely not be adding them on\nmere speculation that the number might be interesting.\n\nNow, that objection would not apply to a per *database* counter, but I'm\nnot sure if tracking the numbers at that granularity would help anyone.\n\nOn the whole, I'm +1 for just logging the events and seeing what we learn\nthat way. That seems like an appropriate amount of effort for finding out\nwhether there is really an issue.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 10 Nov 2014 20:52:59 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-11-10 19:36:18 -0600, Jim Nasby wrote:\n> On 11/10/14, 12:56 PM, Andres Freund wrote:\n> >On 2014-11-10 12:37:29 -0600, Jim Nasby wrote:\n> >>On 11/10/14, 12:15 PM, Andres Freund wrote:\n> >>>>>If what we want is to quantify the extent of the issue, would it be more\n> >>>>>convenient to save counters to pgstat? Vacuum already sends pgstat\n> >>>>>messages, so there's no additional traffic there.\n> >>>I'm pretty strongly against that one in isolation. They'd need to be\n> >>>stored somewhere and they'd need to be queryable somewhere with enough\n> >>>context to make sense. To actually make sense of the numbers we'd also\n> >>>need to report all the other datapoints of vacuum in some form. That's\n> >>>quite a worthwile project imo - but*much* *much* more work than this.\n> >>\n> >>We already report statistics on vacuums\n> >>(lazy_vacuum_rel()/pgstat_report_vacuum), so this would just be adding\n> >>1 or 2 counters to that. Should we add the other counters from vacuum?\n> >>That would be significantly more data.\n> >\n> >At the very least it'd require:\n> >* The number of buffers skipped due to the vm\n> >* The number of buffers actually scanned\n> >* The number of full table in contrast to partial vacuums\n> \n> If we're going to track full scan vacuums separately, I think we'd\n> need two separate scan counters.\n\nWell, we already have the entire number of vacuums, so we'd have that.\n\n> I think (for pgstats) it'd make more sense to just count initial\n> failure to acquire the lock in a full scan in the 'skipped page'\n> counter. In terms of answering the question \"how common is it not to\n> get the lock\", it's really the same event.\n\nIt's absolutely not. You need to correlate the number of skipped pages\nto the number of vacuumed pages. If you have 100k skipped in 10 billion\ntotal scanned pages it's something entirely different than 100k in 200k\npages.\n\n> Honestly, my desire at this point is just to see if there's actually a\n> problem. Many people are asserting that this should be a very rare\n> occurrence, but there's no way to know.\n\nOk.\n\n> Towards that simple end, I'm a bit torn. My preference would be to\n> simply log, and throw a warning if it's over some threshold. I believe\n> that would give the best odds of getting feedback from users if this\n> isn't as uncommon as we think.\n\nI'm strongly against a warning. We have absolutely no sane way of tuning\nthat. We'll just create a pointless warning that people will get\nconfused about and that they'll have to live with till the next release.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 11 Nov 2014 09:01:58 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 11/11/14, 2:01 AM, Andres Freund wrote:\n> On 2014-11-10 19:36:18 -0600, Jim Nasby wrote:\n>> On 11/10/14, 12:56 PM, Andres Freund wrote:\n>>> On 2014-11-10 12:37:29 -0600, Jim Nasby wrote:\n>>>> On 11/10/14, 12:15 PM, Andres Freund wrote:\n>>>>>>> If what we want is to quantify the extent of the issue, would it be more\n>>>>>>> convenient to save counters to pgstat? Vacuum already sends pgstat\n>>>>>>> messages, so there's no additional traffic there.\n>>>>> I'm pretty strongly against that one in isolation. They'd need to be\n>>>>> stored somewhere and they'd need to be queryable somewhere with enough\n>>>>> context to make sense. To actually make sense of the numbers we'd also\n>>>>> need to report all the other datapoints of vacuum in some form. That's\n>>>>> quite a worthwile project imo - but*much* *much* more work than this.\n>>>>\n>>>> We already report statistics on vacuums\n>>>> (lazy_vacuum_rel()/pgstat_report_vacuum), so this would just be adding\n>>>> 1 or 2 counters to that. Should we add the other counters from vacuum?\n>>>> That would be significantly more data.\n>>>\n>>> At the very least it'd require:\n>>> * The number of buffers skipped due to the vm\n>>> * The number of buffers actually scanned\n>>> * The number of full table in contrast to partial vacuums\n>>\n>> If we're going to track full scan vacuums separately, I think we'd\n>> need two separate scan counters.\n>\n> Well, we already have the entire number of vacuums, so we'd have that.\n\nI mean number of pages scanned, but as I said below I don't think that's really necessary.\n\n>> I think (for pgstats) it'd make more sense to just count initial\n>> failure to acquire the lock in a full scan in the 'skipped page'\n>> counter. In terms of answering the question \"how common is it not to\n>> get the lock\", it's really the same event.\n>\n> It's absolutely not. You need to correlate the number of skipped pages\n> to the number of vacuumed pages. If you have 100k skipped in 10 billion\n> total scanned pages it's something entirely different than 100k in 200k\n> pages.\n\nIf the goal here is to find out if this even is a problem then I think the critical question is not \"did we vacuum\", but \"were we able to acquire the lock on the first try\".\n\nObviously users will care much more about the vacuuming and not so much about the lock; but if this really is a non-issue as most tend to believe I don't think it's worth worrying about any of this (except perhaps putting dtrace/system tap probes in).\n\n>> Honestly, my desire at this point is just to see if there's actually a\n>> problem. Many people are asserting that this should be a very rare\n>> occurrence, but there's no way to know.\n>\n> Ok.\n>\n>> Towards that simple end, I'm a bit torn. My preference would be to\n>> simply log, and throw a warning if it's over some threshold. I believe\n>> that would give the best odds of getting feedback from users if this\n>> isn't as uncommon as we think.\n>\n> I'm strongly against a warning. We have absolutely no sane way of tuning\n> that. We'll just create a pointless warning that people will get\n> confused about and that they'll have to live with till the next release.\n\nTo clarify: I'm only suggesting we issue a warning if we have to skip some significant number of pages; say 5 or 0.01% of the table, whichever is greater. That's aimed directly at the goal of letting us know if this is actually a problem or not.\n\nThe reason I'm inclined to do the warning is because I don't think people will notice this otherwise. If this really isn't a problem then it won't matter; if it's a *big* problem then we'll at least know about it.\n\nI'm thinking of an undocumented GUC to control the threshold, but I assume no one else would be on board with that?\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 12 Nov 2014 11:34:04 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-11-12 11:34:04 -0600, Jim Nasby wrote:\n> On 11/11/14, 2:01 AM, Andres Freund wrote:\n> >On 2014-11-10 19:36:18 -0600, Jim Nasby wrote:\n> >>Towards that simple end, I'm a bit torn. My preference would be to\n> >>simply log, and throw a warning if it's over some threshold. I believe\n> >>that would give the best odds of getting feedback from users if this\n> >>isn't as uncommon as we think.\n> >\n> >I'm strongly against a warning. We have absolutely no sane way of tuning\n> >that. We'll just create a pointless warning that people will get\n> >confused about and that they'll have to live with till the next release.\n> \n> To clarify: I'm only suggesting we issue a warning if we have to skip some significant number of pages; say 5 or 0.01% of the table, whichever is greater. That's aimed directly at the goal of letting us know if this is actually a problem or not.\n\nMeh. You have a 5 page relation and it'll trigger quite easily. And it's\nabsolutely harmless.\n\n> The reason I'm inclined to do the warning is because I don't think people will notice this otherwise. If this really isn't a problem then it won't matter; if it's a *big* problem then we'll at least know about it.\n> \n> I'm thinking of an undocumented GUC to control the threshold, but I assume no one else would be on board with that?\n\nStop overdesigning this.\n\nAdd it to the existing mesage and let us be done with this. This thread\nhas already wasted far too much time.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 12 Nov 2014 18:37:46 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On Wed, Nov 12, 2014 at 12:37 PM, Andres Freund <andres@2ndquadrant.com> wrote:\n> Stop overdesigning this.\n>\n> Add it to the existing mesage and let us be done with this. This thread\n> has already wasted far too much time.\n\nThat's a little harsh, but I agree. Producing a warning here is just\ngoing to be log-spam. We've had this behavior for years and - to my\nknowledge - we have not got one complaint that can be attributed to\nthis feature. What used to happen is that VACUUM would get stuck for\nminutes, hours, or days trying to vacuum a table because of an open\ncursor, and people did complain about that. It was a serious nuisance\nthat is now gone. The entire argument in favor of some change here is\n\"even though a careful theoretical analysis indicates that this is\nsafe, even though it solved a real problem that was hurting our users,\nand even though we have no evidence whatsoever that the change is\nhurting anybody in any way whatsoever, the lack of instrumentation\nmeans that it could possibly be hurting somebody and we wouldn't\nknow.\"\n\nThat is, of course, quite true, but you could apply the same argument\nto lots of patches. It would usually be a waste of time because most\nof the patches we think are good actually ARE good - but of course,\nevery once in a while you would find a real problem that had been\noverlooked.\n\nMaybe the right thing here is not to make any change to the source\ncode *at all* but to patch his own copy and try to come up with a\nreproducible test case where this is actually a problem. Then, if\nthere is a problem, we could actually fix it.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 12 Nov 2014 12:53:41 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 11/10/14, 7:52 PM, Tom Lane wrote:\n> On the whole, I'm +1 for just logging the events and seeing what we learn\n> that way. That seems like an appropriate amount of effort for finding out\n> whether there is really an issue.\n\nAttached is a patch that does this.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Sun, 30 Nov 2014 20:46:51 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 12/1/14, 11:57 AM, Andres Freund wrote:\n> On 2014-11-30 20:46:51 -0600, Jim Nasby wrote:\n>> On 11/10/14, 7:52 PM, Tom Lane wrote:\n>>> On the whole, I'm +1 for just logging the events and seeing what we learn\n>>> that way. That seems like an appropriate amount of effort for finding out\n>>> whether there is really an issue.\n>>\n>> Attached is a patch that does this.\n>\n> Unless somebody protests I plan to push this soon. I'll change two\n> things:\n>\n> * Always use the same message, independent of scan_all. For one Jim's\n> version would be untranslatable, for another it's not actually\n> correct. In most cases we'll *not* wait, even if scan_all is\n> true as we'll often just balk after !lazy_check_needs_freeze().\n\nGood point.\n\n> * Change the new bit in the errdetail. \"could not acquire cleanup lock\"\n> sounds too much like an error to me. \"skipped %u pinned pages\" maybe?\n\nSeems reasonable.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 1 Dec 2014 12:55:43 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 20 October 2014 at 10:57, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n\n> Currently, a non-freeze vacuum will punt on any page it can't get a cleanup\n> lock on, with no retry. Presumably this should be a rare occurrence, but I\n> think it's bad that we just assume that and won't warn the user if something\n> bad is going on.\n\n(I'm having email problems, so I can't see later mails on this thread,\nso replying here.)\n\nLogging patch looks fine, but I would rather not add a line of text\nfor each VACUUM, just in case this is non-zero. I think we should add\nthat log line only if the blocks skipped > 0.\n\nWhat I'm more interested in is what you plan to do with the\ninformation once we get it?\n\nThe assumption that skipping blocks is something bad is strange. I\nadded it because VACUUM could and did regularly hang on busy tables,\nwhich resulted in bloat because other blocks that needed cleaning\ndidn't get any attention.\n\nWhich is better, spend time obsessively trying to vacuum particular\nblocks, or to spend the time on other blocks that are in need of\ncleaning and are available to be cleaned?\n\nWhich is better, have autovacuum or system wide vacuum progress on to\nother tables that need cleaning, or spend lots of effort retrying?\n\nHow do we know what is the best next action?\n\nI'd really want to see some analysis of those things before we spend\neven more cycles on this.\n\n-- \n Simon Riggs http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 8 Dec 2014 09:16:11 +0900","msg_from":"Simon Riggs <simon@2ndQuadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 12/7/14, 6:16 PM, Simon Riggs wrote:\n> On 20 October 2014 at 10:57, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n>\n>> Currently, a non-freeze vacuum will punt on any page it can't get a cleanup\n>> lock on, with no retry. Presumably this should be a rare occurrence, but I\n>> think it's bad that we just assume that and won't warn the user if something\n>> bad is going on.\n>\n> (I'm having email problems, so I can't see later mails on this thread,\n> so replying here.)\n>\n> Logging patch looks fine, but I would rather not add a line of text\n> for each VACUUM, just in case this is non-zero. I think we should add\n> that log line only if the blocks skipped > 0.\n\nI thought about doing that, but I'm loath to duplicate a rather large ereport call. Happy to make the change if that's the consensus though.\n\n> What I'm more interested in is what you plan to do with the\n> information once we get it?\n>\n> The assumption that skipping blocks is something bad is strange. I\n> added it because VACUUM could and did regularly hang on busy tables,\n> which resulted in bloat because other blocks that needed cleaning\n> didn't get any attention.\n>\n> Which is better, spend time obsessively trying to vacuum particular\n> blocks, or to spend the time on other blocks that are in need of\n> cleaning and are available to be cleaned?\n>\n> Which is better, have autovacuum or system wide vacuum progress on to\n> other tables that need cleaning, or spend lots of effort retrying?\n>\n> How do we know what is the best next action?\n>\n> I'd really want to see some analysis of those things before we spend\n> even more cycles on this.\n\nThat's the entire point of logging this information. There is an underlying assumption that we won't actually skip many pages, but there's no data to back that up, nor is there currently any way to get that data.\n\nMy hope is that the logging shows that there isn't anything more that needs to be done here. If this is something that causes problems, at least now DBAs will be aware of it and hopefully we'll be able to identify specific problem scenarios and find a solution.\n\n\n\nBTW, my initial proposal[1] was strictly logging. The only difference was raising it to a warning if a significant portion of the table was skipped. I only investigated retrying locks at the suggestion of others. I never intended this to become a big time sink.\n\n[1]:\n\"Currently, a non-freeze vacuum will punt on any page it can't get a cleanup lock on, with no retry. Presumably this should be a rare occurrence, but I think it's bad that we just assume that and won't warn the user if something bad is going on.\n\n\"My thought is that if we skip any pages elog(LOG) how many we skipped. If we skip more than 1% of the pages we visited (not relpages) then elog(WARNING) instead.\"\n--\nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 8 Dec 2014 15:28:54 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 9 December 2014 at 06:28, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\n> On 12/7/14, 6:16 PM, Simon Riggs wrote:\n\n>> What I'm more interested in is what you plan to do with the\n>> information once we get it?\n>>\n>> The assumption that skipping blocks is something bad is strange. I\n>> added it because VACUUM could and did regularly hang on busy tables,\n>> which resulted in bloat because other blocks that needed cleaning\n>> didn't get any attention.\n>>\n>> Which is better, spend time obsessively trying to vacuum particular\n>> blocks, or to spend the time on other blocks that are in need of\n>> cleaning and are available to be cleaned?\n>>\n>> Which is better, have autovacuum or system wide vacuum progress on to\n>> other tables that need cleaning, or spend lots of effort retrying?\n>>\n>> How do we know what is the best next action?\n>>\n>> I'd really want to see some analysis of those things before we spend\n>> even more cycles on this.\n>\n>\n> That's the entire point of logging this information. There is an underlying\n> assumption that we won't actually skip many pages, but there's no data to\n> back that up, nor is there currently any way to get that data.\n\nThere is no such underlying assumption. You assumed there was one, but\nthere isn't one.\n\nAll I can say for certain is that waiting on a lock for long periods\nwas literally a waste of time. Now it no longer wastes time, it gets\non with vacuuming the pages it can.\n\n-- \n Simon Riggs http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 9 Dec 2014 07:41:06 +0900","msg_from":"Simon Riggs <simon@2ndQuadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 12/01/2014 08:55 PM, Jim Nasby wrote:\n> On 12/1/14, 11:57 AM, Andres Freund wrote:\n>> On 2014-11-30 20:46:51 -0600, Jim Nasby wrote:\n>>> On 11/10/14, 7:52 PM, Tom Lane wrote:\n>>>> On the whole, I'm +1 for just logging the events and seeing what we learn\n>>>> that way. That seems like an appropriate amount of effort for finding out\n>>>> whether there is really an issue.\n>>>\n>>> Attached is a patch that does this.\n>>\n>> Unless somebody protests I plan to push this soon. I'll change two\n>> things:\n>>\n>> * Always use the same message, independent of scan_all. For one Jim's\n>> version would be untranslatable, for another it's not actually\n>> correct. In most cases we'll *not* wait, even if scan_all is\n>> true as we'll often just balk after !lazy_check_needs_freeze().\n>\n> Good point.\n>\n>> * Change the new bit in the errdetail. \"could not acquire cleanup lock\"\n>> sounds too much like an error to me. \"skipped %u pinned pages\" maybe?\n>\n> Seems reasonable.\n\nWell, that's not always true either; when freezing, it doesn't skip the \npinned pages, it waits for them.\n\nHow about the attached (I also edited the comments a bit)? It looks like \nthis:\n\npostgres=# vacuum verbose foo;\nINFO: vacuuming \"public.foo\"\nINFO: \"foo\": found 0 removable, 0 nonremovable row versions in 0 out of \n7256 pages\nDETAIL: 0 dead row versions cannot be removed yet.\nThere were 0 unused item pointers.\n1 pages were pinned.\n0 pages are entirely empty.\nCPU 0.00s/0.00u sec elapsed 0.00 sec.\nINFO: \"foo\": stopping truncate due to conflicting lock request\nINFO: vacuuming \"pg_toast.pg_toast_16384\"\nINFO: index \"pg_toast_16384_index\" now contains 0 row versions in 1 pages\nDETAIL: 0 index row versions were removed.\n0 index pages have been deleted, 0 are currently reusable.\nCPU 0.00s/0.00u sec elapsed 0.00 sec.\n\nand for autovacuum log:\n\nLOG: automatic vacuum of table \"postgres.public.foo\": index scans: 0\n\tpages: 0 removed, 7256 remain, 0 pinned\n\ttuples: 79415 removed, 513156 remain, 0 are dead but not yet removable\n\tbuffer usage: 14532 hits, 6 misses, 6241 dirtied\n\tavg read rate: 0.003 MB/s, avg write rate: 3.413 MB/s\n\tsystem usage: CPU 0.00s/0.30u sec elapsed 14.28 sec\n\nI.e. this just says how many pages were pinned, without saying what was \ndone about them. That's not very meaningful to an average DBA, but \nthat's true for many of the numbers printed in vacuum verbose.\n\n- Heikki\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Wed, 17 Dec 2014 18:20:49 +0200","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 12/17/14, 10:20 AM, Heikki Linnakangas wrote:\n>>\n>>> * Change the new bit in the errdetail. \"could not acquire cleanup lock\"\n>>> sounds too much like an error to me. \"skipped %u pinned pages\" maybe?\n>>\n>> Seems reasonable.\n>\n> Well, that's not always true either; when freezing, it doesn't skip the pinned pages, it waits for them.\n\nOops. :(\n\n> I.e. this just says how many pages were pinned, without saying what was done about them. That's not very meaningful to an average DBA, but that's true for many of the numbers printed in vacuum verbose.\n\nIn this case it'll mean people will be less likely to report that this is happening, but maybe that's OK. At least if someone comes to us with a problem we'll be able to get some info from them. I'll separately look into the vacuum docs and see if we can do a better job explaining the verbose output.\n\nBTW, what is it about a dynamic message that makes it untranslatable? Doesn't the translation happen down-stream, so that at most we'd just need two translation messages? Or worst case we could have two separate elog calls, if we wanted to go that route.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 17 Dec 2014 12:02:40 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 12/17/2014 08:02 PM, Jim Nasby wrote:\n> BTW, what is it about a dynamic message that makes it untranslatable?\n> Doesn't the translation happen down-stream, so that at most we'd just\n> need two translation messages?\n\nI'm not sure what you mean by down-stream. There is some explanation on \nthis in the localization guide in the manual:\n\nhttp://www.postgresql.org/docs/devel/static/nls-programmer.html#NLS-GUIDELINES\n\nprintf(\"Files were %s.\\n\", flag ? \"copied\" : \"removed\");\n\nThe translator will get three strings to translate: \"Files were %s.\\n\", \n\"copied\", and \"removed\".\n\n> Or worst case we could have two\n> separate elog calls, if we wanted to go that route.\n\nYeah, that's one option. I.e:\n\nif (flag)\n printf(\"Files were copied.\\n\");\nelse\n printf(\"Files were removed.\\n\");\n\n- Heikki\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 17 Dec 2014 20:28:39 +0200","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"Jim Nasby wrote:\n> On 12/17/14, 10:20 AM, Heikki Linnakangas wrote:\n> >>\n> >>>* Change the new bit in the errdetail. \"could not acquire cleanup lock\"\n> >>> sounds too much like an error to me. \"skipped %u pinned pages\" maybe?\n> >>\n> >>Seems reasonable.\n> >\n> >Well, that's not always true either; when freezing, it doesn't skip the pinned pages, it waits for them.\n> \n> Oops. :(\n> \n> >I.e. this just says how many pages were pinned, without saying what was done about them. That's not very meaningful to an average DBA, but that's true for many of the numbers printed in vacuum verbose.\n> \n> In this case it'll mean people will be less likely to report that this is happening, but maybe that's OK. At least if someone comes to us with a problem we'll be able to get some info from them. I'll separately look into the vacuum docs and see if we can do a better job explaining the verbose output.\n> \n> BTW, what is it about a dynamic message that makes it untranslatable? Doesn't the translation happen down-stream, so that at most we'd just need two translation messages? Or worst case we could have two separate elog calls, if we wanted to go that route.\n\nSince each part is a complete sentence, you can paste them together.\nFor example you could do something like\n\nmsg1 = _(\"Could not acquire cleanup lock.\");\nmsg2 = _(\"Skipped %u pinned pages.\");\n\nereport(INFO,\n (errcode(ERRCODE_SOME_STUFF),\n errmsg_internal(\"%s\\n%s\", msg1, msg2)));\n\nThe use of errmsg_internal() is to avoid having it be translated again.\n\nFWIW I think the vacuum summary INFO message is pretty terrible already.\nI would support a patch that changed it to be separately translatable\nunits (as long as each unit is a complete sentence.)\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 17 Dec 2014 15:39:39 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On Wed, Dec 17, 2014 at 11:20 AM, Heikki Linnakangas\n<hlinnakangas@vmware.com> wrote:\n> LOG: automatic vacuum of table \"postgres.public.foo\": index scans: 0\n> pages: 0 removed, 7256 remain, 0 pinned\n> tuples: 79415 removed, 513156 remain, 0 are dead but not yet\n> removable\n> buffer usage: 14532 hits, 6 misses, 6241 dirtied\n> avg read rate: 0.003 MB/s, avg write rate: 3.413 MB/s\n> system usage: CPU 0.00s/0.30u sec elapsed 14.28 sec\n>\n> I.e. this just says how many pages were pinned, without saying what was done\n> about them. That's not very meaningful to an average DBA, but that's true\n> for many of the numbers printed in vacuum verbose.\n\nThat message is extremely confusing, to my eyes. If you want to say\n\"pages: 0 removed, 7256 remain, 0 skipped due to pins\", that would\nwork for me, but just say \"0 pinned\" is totally wrong, because vacuum\npinned every page in the table.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 08:56:56 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 12/18/14, 7:56 AM, Robert Haas wrote:\n> On Wed, Dec 17, 2014 at 11:20 AM, Heikki Linnakangas\n> <hlinnakangas@vmware.com> wrote:\n>> LOG: automatic vacuum of table \"postgres.public.foo\": index scans: 0\n>> pages: 0 removed, 7256 remain, 0 pinned\n>> tuples: 79415 removed, 513156 remain, 0 are dead but not yet\n>> removable\n>> buffer usage: 14532 hits, 6 misses, 6241 dirtied\n>> avg read rate: 0.003 MB/s, avg write rate: 3.413 MB/s\n>> system usage: CPU 0.00s/0.30u sec elapsed 14.28 sec\n>>\n>> I.e. this just says how many pages were pinned, without saying what was done\n>> about them. That's not very meaningful to an average DBA, but that's true\n>> for many of the numbers printed in vacuum verbose.\n>\n> That message is extremely confusing, to my eyes. If you want to say\n> \"pages: 0 removed, 7256 remain, 0 skipped due to pins\", that would\n> work for me, but just say \"0 pinned\" is totally wrong, because vacuum\n> pinned every page in the table.\n\nWe have to decide on a tradeoff here. Either we end up with two different log messages (depending on scan_all) that require two different translations, or we end up with a generic message that isn't as clear.\n\nThe best option I can think of for the later is something like \"failed initial lock attempt\". That's the only thing that will be true in all cases.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 11:38:12 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"Jim Nasby wrote:\n\n> We have to decide on a tradeoff here. Either we end up with two\n> different log messages (depending on scan_all) that require two\n> different translations, or we end up with a generic message that isn't\n> as clear.\n> \n> The best option I can think of for the later is something like \"failed\n> initial lock attempt\". That's the only thing that will be true in all\n> cases.\n\nHere's my proposal. Instead of punting, I split the message in\nseparately translatable units, and emit only the ones that apply. The\ncode is messier this way, but I think we can live with that.\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Thu, 18 Dec 2014 16:41:04 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 12/18/2014 09:41 PM, Alvaro Herrera wrote:\n> Jim Nasby wrote:\n>\n>> We have to decide on a tradeoff here. Either we end up with two\n>> different log messages (depending on scan_all) that require two\n>> different translations, or we end up with a generic message that isn't\n>> as clear.\n>>\n>> The best option I can think of for the later is something like \"failed\n>> initial lock attempt\". That's the only thing that will be true in all\n>> cases.\n>\n> Here's my proposal. Instead of punting, I split the message in\n> separately translatable units, and emit only the ones that apply. The\n> code is messier this way, but I think we can live with that.\n\nWorks for me.\n\n- Heikki\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 22:06:57 +0200","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"Heikki Linnakangas wrote:\n> On 12/18/2014 09:41 PM, Alvaro Herrera wrote:\n\n> >Here's my proposal. Instead of punting, I split the message in\n> >separately translatable units, and emit only the ones that apply. The\n> >code is messier this way, but I think we can live with that.\n> \n> Works for me.\n\nGreat, thanks, pushed. I tweaked it a bit more, so that it would say\neither \"skipped N pages\" or \"waited N pins\" in both autovacuum and\nvacuum verbose cases, but only if N > 0.\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 17:21:36 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\n> Great, thanks, pushed. I tweaked it a bit more, so that it would say\n> either \"skipped N pages\" or \"waited N pins\" in both autovacuum and\n> vacuum verbose cases, but only if N > 0.\n\nNot directly relevant but ... I think probably all those BlockNumber\ncounters should be printed with %u not %d. A relation with between\n2G and 4G pages is possible, and it wouldn't look right with %d.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 15:50:31 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-12-18 16:41:04 -0300, Alvaro Herrera wrote:\n> +\t\t\tif (scan_all)\n> +\t\t\t\tappendStringInfo(&buf, _(\"waited for %d buffer pins\\n\"),\n> +\t\t\t\t\t\t\t\t vacrelstats->pinned_pages);\n> +\t\t\telse\n> +\t\t\t\tappendStringInfo(&buf,\n> +\t\t\t\t\t\t\t\t _(\"skipped %d pages due to buffer pins\\n\"),\n> +\t\t\t\t\t\t\t\t vacrelstats->pinned_pages);\n\nUnless I miss something this is, as mentioned before, not\ncorrect. scan_all doesn't imply at all that we waited for buffer\npins. We only do so if lazy_check_needs_freeze(buf). Which usually won't\nbe true for a *significant* number of pages.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 21:55:05 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"Andres Freund wrote:\n> On 2014-12-18 16:41:04 -0300, Alvaro Herrera wrote:\n> > +\t\t\tif (scan_all)\n> > +\t\t\t\tappendStringInfo(&buf, _(\"waited for %d buffer pins\\n\"),\n> > +\t\t\t\t\t\t\t\t vacrelstats->pinned_pages);\n> > +\t\t\telse\n> > +\t\t\t\tappendStringInfo(&buf,\n> > +\t\t\t\t\t\t\t\t _(\"skipped %d pages due to buffer pins\\n\"),\n> > +\t\t\t\t\t\t\t\t vacrelstats->pinned_pages);\n> \n> Unless I miss something this is, as mentioned before, not\n> correct. scan_all doesn't imply at all that we waited for buffer\n> pins. We only do so if lazy_check_needs_freeze(buf). Which usually won't\n> be true for a *significant* number of pages.\n\nAh, interesting, I didn't remember we had that. I guess one possible\ntweak is to discount the pages we skip from pinned_pages; or we could\nkeep a separate count of pages waited for. Jim, up for a patch?\n\n-- \n�lvaro Herrera http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 18:02:54 -0300","msg_from":"Alvaro Herrera <alvherre@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 12/18/14, 3:02 PM, Alvaro Herrera wrote:\n> Andres Freund wrote:\n>> On 2014-12-18 16:41:04 -0300, Alvaro Herrera wrote:\n>>> +\t\t\tif (scan_all)\n>>> +\t\t\t\tappendStringInfo(&buf, _(\"waited for %d buffer pins\\n\"),\n>>> +\t\t\t\t\t\t\t\t vacrelstats->pinned_pages);\n>>> +\t\t\telse\n>>> +\t\t\t\tappendStringInfo(&buf,\n>>> +\t\t\t\t\t\t\t\t _(\"skipped %d pages due to buffer pins\\n\"),\n>>> +\t\t\t\t\t\t\t\t vacrelstats->pinned_pages);\n>>\n>> Unless I miss something this is, as mentioned before, not\n>> correct. scan_all doesn't imply at all that we waited for buffer\n>> pins. We only do so if lazy_check_needs_freeze(buf). Which usually won't\n>> be true for a *significant* number of pages.\n>\n> Ah, interesting, I didn't remember we had that. I guess one possible\n> tweak is to discount the pages we skip from pinned_pages; or we could\n> keep a separate count of pages waited for. Jim, up for a patch?\n\nI would prefer that we at least count if we initially don't get the lock; presumably that number is always low anyway and in that case I think we're done with this. If it turns out it is common to initially miss the pin then we could do something fancier.\n\nSo how about if in the scan_all case we say something like \"unable to initially acquire pin on %d buffers\\n\"?\n\n(Happy to do the patch either way, but I'd like us to decide what we're doing first. ;)\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 18 Dec 2014 16:05:23 -0600","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-12-18 16:05:23 -0600, Jim Nasby wrote:\n> On 12/18/14, 3:02 PM, Alvaro Herrera wrote:\n> >Andres Freund wrote:\n> >>On 2014-12-18 16:41:04 -0300, Alvaro Herrera wrote:\n> >>>+\t\t\tif (scan_all)\n> >>>+\t\t\t\tappendStringInfo(&buf, _(\"waited for %d buffer pins\\n\"),\n> >>>+\t\t\t\t\t\t\t\t vacrelstats->pinned_pages);\n> >>>+\t\t\telse\n> >>>+\t\t\t\tappendStringInfo(&buf,\n> >>>+\t\t\t\t\t\t\t\t _(\"skipped %d pages due to buffer pins\\n\"),\n> >>>+\t\t\t\t\t\t\t\t vacrelstats->pinned_pages);\n> >>\n> >>Unless I miss something this is, as mentioned before, not\n> >>correct. scan_all doesn't imply at all that we waited for buffer\n> >>pins. We only do so if lazy_check_needs_freeze(buf). Which usually won't\n> >>be true for a *significant* number of pages.\n> >\n> >Ah, interesting, I didn't remember we had that. I guess one possible\n> >tweak is to discount the pages we skip from pinned_pages; or we could\n> >keep a separate count of pages waited for. Jim, up for a patch?\n> \n> I would prefer that we at least count if we initially don't get the lock; presumably that number is always low anyway and in that case I think we're done with this. If it turns out it is common to initially miss the pin then we could do something fancier.\n> \n> So how about if in the scan_all case we say something like \"unable to initially acquire pin on %d buffers\\n\"?\n\nI'd just do away with the difference between scan_all/!scan_all and\nalways say the above.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 19 Dec 2014 08:50:35 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2014-12-18 16:05:23 -0600, Jim Nasby wrote:\n> On 12/18/14, 3:02 PM, Alvaro Herrera wrote:\n> >Andres Freund wrote:\n> >>On 2014-12-18 16:41:04 -0300, Alvaro Herrera wrote:\n> >>>+\t\t\tif (scan_all)\n> >>>+\t\t\t\tappendStringInfo(&buf, _(\"waited for %d buffer pins\\n\"),\n> >>>+\t\t\t\t\t\t\t\t vacrelstats->pinned_pages);\n> >>>+\t\t\telse\n> >>>+\t\t\t\tappendStringInfo(&buf,\n> >>>+\t\t\t\t\t\t\t\t _(\"skipped %d pages due to buffer pins\\n\"),\n> >>>+\t\t\t\t\t\t\t\t vacrelstats->pinned_pages);\n> >>\n> >>Unless I miss something this is, as mentioned before, not\n> >>correct. scan_all doesn't imply at all that we waited for buffer\n> >>pins. We only do so if lazy_check_needs_freeze(buf). Which usually won't\n> >>be true for a *significant* number of pages.\n\nAlso, naming the number of pages we could *not* pin, pinned_pages?\nReally?\n\npinskipped_pages,skipped_pages,unpinned_pages,...\n\n> >Ah, interesting, I didn't remember we had that. I guess one possible\n> >tweak is to discount the pages we skip from pinned_pages; or we could\n> >keep a separate count of pages waited for. Jim, up for a patch?\n\nThis is still wrong. I think just counting skipped pages, without\ndistinct messages for waiting/not waiting, is good enough for\nnow. Everything else would only be actually meaningful if we actually\ntracked the waiting time.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 4 Jan 2015 01:53:24 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"},{"msg_contents":"On 2015-01-04 01:53:24 +0100, Andres Freund wrote:\n> > >Ah, interesting, I didn't remember we had that. I guess one possible\n> > >tweak is to discount the pages we skip from pinned_pages; or we could\n> > >keep a separate count of pages waited for. Jim, up for a patch?\n>\n> This is still wrong. I think just counting skipped pages, without\n> distinct messages for waiting/not waiting, is good enough for\n> now. Everything else would only be actually meaningful if we actually\n> tracked the waiting time.\n\nPushed a commit for this, with additional improvements to autovacuum's\nlog output from:\nLOG: automatic vacuum of table \"postgres.public.frak\": index scans: 0\n pages: 0 removed, 1672 remain\n skipped 1 pages due to buffer pins\n tuples: 0 removed, 309959 remain, 309774 are dead but not yet removable\n buffer usage: 4258 hits, 0 misses, 0 dirtied\n avg read rate: 0.000 MB/s, avg write rate: 0.000 MB/s\n system usage: CPU 0.00s/0.04u sec elapsed 0.46 sec\nto:\nLOG: automatic vacuum of table \"postgres.public.frak\": index scans: 0\n pages: 0 removed, 1672 remain, 1 skipped due to pins\n tuples: 0 removed, 309959 remain, 309774 are dead but not yet removable\n buffer usage: 4258 hits, 0 misses, 0 dirtied\n avg read rate: 0.000 MB/s, avg write rate: 0.000 MB/s\n system usage: CPU 0.00s/0.04u sec elapsed 0.46 sec\nas the 'skipped ...' line didn't really look in line with the rest.\n\nGreetings,\n\nAndres Freund\n\n--\n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 8 Jan 2015 13:19:23 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: Proposal: Log inability to lock pages during vacuum"}],"string":"[\n {\n \"msg_contents\": \"Currently, a non-freeze vacuum will punt on any page it can't get a cleanup lock on, with no retry. Presumably this should be a rare occurrence, but I think it's bad that we just assume that and won't warn the user if something bad is going on.\\n\\nMy thought is that if we skip any pages elog(LOG) how many we skipped. If we skip more than 1% of the pages we visited (not relpages) then elog(WARNING) instead.\\n\\nComments?\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 19 Oct 2014 20:57:36 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 20, 2014 at 2:57 AM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n> Currently, a non-freeze vacuum will punt on any page it can't get a cleanup\\n> lock on, with no retry. Presumably this should be a rare occurrence, but I\\n> think it's bad that we just assume that and won't warn the user if something\\n> bad is going on.\\n>\\n> My thought is that if we skip any pages elog(LOG) how many we skipped. If we\\n> skip more than 1% of the pages we visited (not relpages) then elog(WARNING)\\n> instead.\\n\\nIs there some specific failure you've run into where a page was stuck\\nin a pinned state and never got vacuumed?\\n\\nI would like to see a more systematic way of going about this. What\\nLSN or timestamp is associated with the oldest unvacuumed page? How\\nmany times have we tried to visit it? What do those numbers look like\\noverall -- i.e. what's the median number of times it takes to vacuum a\\npage and what does the distribution look like of the unvacuumed ages?\\n\\nWith that data it should be possible to determine if the behaviour is\\nactually working well and where to draw the line to determine outliers\\nthat might represent bugs.\\n\\n-- \\ngreg\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 16:29:58 +0100\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 10/20/14, 10:29 AM, Greg Stark wrote:\\n> On Mon, Oct 20, 2014 at 2:57 AM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n>> Currently, a non-freeze vacuum will punt on any page it can't get a cleanup\\n>> lock on, with no retry. Presumably this should be a rare occurrence, but I\\n>> think it's bad that we just assume that and won't warn the user if something\\n>> bad is going on.\\n>>\\n>> My thought is that if we skip any pages elog(LOG) how many we skipped. If we\\n>> skip more than 1% of the pages we visited (not relpages) then elog(WARNING)\\n>> instead.\\n>\\n> Is there some specific failure you've run into where a page was stuck\\n> in a pinned state and never got vacuumed?\\n\\nNot that I know of... but how would I actually know? Having that info available is the point of my proposal. :)\\n\\n> I would like to see a more systematic way of going about this. What\\n> LSN or timestamp is associated with the oldest unvacuumed page? How\\n> many times have we tried to visit it? What do those numbers look like\\n> overall -- i.e. what's the median number of times it takes to vacuum a\\n> page and what does the distribution look like of the unvacuumed ages?\\n>\\n> With that data it should be possible to determine if the behaviour is\\n> actually working well and where to draw the line to determine outliers\\n> that might represent bugs.\\n\\nI agree we could use better data about/for vacuum (see http://www.postgresql.org/message-id/544468C1.6050101@BlueTreble.com).\\n\\nIn the meantime, I think it's worth adding this logging. If in fact this basically never happens (the current assumption), it doesn't hurt anything. If it turns out our assumption is wrong, then we'll actually be able to find that out. :)\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 19:18:31 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-10-20 19:18:31 -0500, Jim Nasby wrote:\\n> In the meantime, I think it's worth adding this logging. If in fact this basically never happens (the current assumption), it doesn't hurt anything. If it turns out our assumption is wrong, then we'll actually be able to fin> that out. :)\\n\\nIt does happen, and not infrequently. Just not enough pages to normally\\ncause significant bloat. The most likely place where it happens is very\\nsmall tables that all connections hit with a high frequency. Starting to\\nissue high volume log spew for a nonexistant problem isn't helping.\\n\\nIf you're super convinced this is urgent then add it as a *single*\\ndatapoint inside the existing messages. But I think there's loads of\\nstuff in vacuum logging that are more important this.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 02:31:59 +0200\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 10/20/14, 7:31 PM, Andres Freund wrote:\\n> On 2014-10-20 19:18:31 -0500, Jim Nasby wrote:\\n>> >In the meantime, I think it's worth adding this logging. If in fact this basically never happens (the current assumption), it doesn't hurt anything. If it turns out our assumption is wrong, then we'll actually be able to fin> that out.:)\\n> It does happen, and not infrequently. Just not enough pages to normally\\n> cause significant bloat. The most likely place where it happens is very\\n> small tables that all connections hit with a high frequency. Starting to\\n> issue high volume log spew for a nonexistant problem isn't helping.\\n\\nHow'd you determine that? Is there some way to measure this? I'm not doubting you; I just don't want to work on a problem that's already solved.\\n\\n> If you're super convinced this is urgent then add it as a*single*\\n> datapoint inside the existing messages. But I think there's loads of\\n> stuff in vacuum logging that are more important this.\\n\\nSee my original proposal; at it's most intrusive this would issue one warning per (auto)vacuum if it was over a certain threshold. I would think that a DBA would really like to know when this happens, but if we think that's too much spew we can limit it to normal vacuum logging.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 19:43:38 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-10-20 19:43:38 -0500, Jim Nasby wrote:\\n> On 10/20/14, 7:31 PM, Andres Freund wrote:\\n> >On 2014-10-20 19:18:31 -0500, Jim Nasby wrote:\\n> >>>In the meantime, I think it's worth adding this logging. If in fact this basically never happens (the current assumption), it doesn't hurt anything. If it turns out our assumption is wrong, then we'll actually be able to fin> that out.:)\\n> >It does happen, and not infrequently. Just not enough pages to normally\\n> >cause significant bloat. The most likely place where it happens is very\\n> >small tables that all connections hit with a high frequency. Starting to\\n> >issue high volume log spew for a nonexistant problem isn't helping.\\n> \\n> How'd you determine that? Is there some way to measure this?\\n\\nYou'd seen individual pages with too old dead rows in them.\\n\\n> >If you're super convinced this is urgent then add it as a*single*\\n> >datapoint inside the existing messages. But I think there's loads of\\n> >stuff in vacuum logging that are more important this.\\n> \\n> See my original proposal; at it's most intrusive this would issue one\\n> warning per (auto)vacuum if it was over a certain threshold.\\n\\nWhich would vastly increase the log output for setups with small tables\\nand a nonzero log_autovacuum_min_duration.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 02:49:55 +0200\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"Jim Nasby wrote:\\n\\n> Currently, a non-freeze vacuum will punt on any page it can't get a\\n> cleanup lock on, with no retry. Presumably this should be a rare\\n> occurrence, but I think it's bad that we just assume that and won't\\n> warn the user if something bad is going on.\\n\\nI think if you really want to attack this problem, rather than just\\nbeing noisy about it, what you could do is to keep a record of which\\npage numbers you had to skip, and then once you're done with your first\\nscan you go back and retry the lock on the pages you skipped.\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 19:39:48 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 10/21/14, 5:39 PM, Alvaro Herrera wrote:\\n> Jim Nasby wrote:\\n>\\n>> Currently, a non-freeze vacuum will punt on any page it can't get a\\n>> cleanup lock on, with no retry. Presumably this should be a rare\\n>> occurrence, but I think it's bad that we just assume that and won't\\n>> warn the user if something bad is going on.\\n>\\n> I think if you really want to attack this problem, rather than just\\n> being noisy about it, what you could do is to keep a record of which\\n> page numbers you had to skip, and then once you're done with your first\\n> scan you go back and retry the lock on the pages you skipped.\\n\\nI'm OK with that if the community is; I was just trying for minimum invasiveness.\\n\\nIf I go this route, I'd like some input though...\\n\\n- How to handle storing the blockIDs. Fixed size array or something fancier? What should we limit it to, especially since we're already allocating maintenance_work_mem for the tid array.\\n\\n- What happens if we run out of space to remember skipped blocks? I could do something like what we do for running out of space in the dead_tuples array, but I'm worried that will add a serious amount of complexity, especially since re-processing these blocks could be what actually pushes us over the limit.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 18:00:07 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\\n> - What happens if we run out of space to remember skipped blocks?\\n\\nYou forget some, and are no worse off than today. (This might be an\\nevent worthy of logging, if the array is large enough that we don't\\nexpect it to happen often ...)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 19:05:23 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 10/21/14, 6:05 PM, Tom Lane wrote:\\n> Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\\n>> - What happens if we run out of space to remember skipped blocks?\\n>\\n> You forget some, and are no worse off than today. (This might be an\\n> event worthy of logging, if the array is large enough that we don't\\n> expect it to happen often ...)\\n\\nMakes sense. I'll see if there's some reasonable way to retry pages when the array fills up.\\n\\nI'll make the array 2k in size; that allows for 512 pages without spending a bunch of memory.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 29 Oct 2014 11:49:02 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 10/29/14, 11:49 AM, Jim Nasby wrote:\\n> On 10/21/14, 6:05 PM, Tom Lane wrote:\\n>> Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\\n>>> - What happens if we run out of space to remember skipped blocks?\\n>>\\n>> You forget some, and are no worse off than today. (This might be an\\n>> event worthy of logging, if the array is large enough that we don't\\n>> expect it to happen often ...)\\n>\\n> Makes sense. I'll see if there's some reasonable way to retry pages when the array fills up.\\n>\\n> I'll make the array 2k in size; that allows for 512 pages without spending a bunch of memory.\\n\\nAttached is a patch for this. It also adds logging of unobtainable cleanup locks, and refactors scanning a page for vacuum into it's own function.\\n\\nAnyone reviewing this might want to look at https://github.com/decibel/postgres/commit/69ab22f703d577cbb3d8036e4e42563977bcf74b, which is the refactor with no whitespace changes.\\n\\nI've verified this works correctly by connecting to a backend with gdb and halting it with a page pinned. Both vacuum and vacuum freeze on that table do what's expected, but I also get this waring (which AFAICT is a false positive):\\n\\ndecibel@decina.local=# vacuum verbose i;\\nINFO: vacuuming \\\"public.i\\\"\\nINFO: \\\"i\\\": found 0 removable, 399774 nonremovable row versions in 1769 out of 1770 pages\\nDETAIL: 200000 dead row versions cannot be removed yet.\\nThere were 0 unused item pointers.\\n0 pages are entirely empty.\\nRetried cleanup lock on 0 pages, retry failed on 1, skipped retry on 0.\\nCPU 0.00s/0.06u sec elapsed 12.89 sec.\\nWARNING: buffer refcount leak: [105] (rel=base/16384/16385, blockNum=0, flags=0x106, refcount=2 1)\\nVACUUM\\n\\nI am doing a simple static allocation of retry_pages[]; my understanding is that will only exist for the duration of this function so it's OK. If not I'll palloc it. If it is OK then I'll do the same for the freeze array.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Thu, 6 Nov 2014 14:55:37 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-11-06 14:55:37 -0600, Jim Nasby wrote:\\n> On 10/29/14, 11:49 AM, Jim Nasby wrote:\\n> >On 10/21/14, 6:05 PM, Tom Lane wrote:\\n> >>Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\\n> >>>- What happens if we run out of space to remember skipped blocks?\\n> >>\\n> >>You forget some, and are no worse off than today. (This might be an\\n> >>event worthy of logging, if the array is large enough that we don't\\n> >>expect it to happen often ...)\\n> >\\n> >Makes sense. I'll see if there's some reasonable way to retry pages when the array fills up.\\n> >\\n> >I'll make the array 2k in size; that allows for 512 pages without spending a bunch of memory.\\n> \\n> Attached is a patch for this. It also adds logging of unobtainable cleanup locks, and refactors scanning a page for vacuum into it's own function.\\n> \\n> Anyone reviewing this might want to look at https://github.com/decibel/postgres/commit/69ab22f703d577cbb3d8036e4e42563977bcf74b, which is the refactor with no whitespace changes.\\n> \\n> I've verified this works correctly by connecting to a backend with gdb and halting it with a page pinned. Both vacuum and vacuum freeze on that table do what's expected, but I also get this waring (which AFAICT is a false positive):\\n\\nI think the retry logical is a largely pointless complication of already\\ncomplex enough code. You're fixing a problem for which there is\\nabsolutely no evidence of its existance. Yes, this happens\\noccasionally. But it's going to be so absolutely minor in comparison to\\njust about every other source of bloat.\\n\\nSo, I pretty strongly against retrying. I could live with adding logging\\nof the number of pages skipped due to not being able to acquire the\\ncleanup lock. I don't think that's going to do help many people, but the\\ncost is pretty low.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 6 Nov 2014 22:30:15 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 6, 2014 at 9:30 PM, Andres Freund <andres@2ndquadrant.com> wrote:\\n> I think the retry logical is a largely pointless complication of already\\n> complex enough code. You're fixing a problem for which there is\\n> absolutely no evidence of its existance. Yes, this happens\\n> occasionally. But it's going to be so absolutely minor in comparison to\\n> just about every other source of bloat.\\n\\nI agree bloat isn't really a threat, but what about the relfrozenxid?\\nIf we skip even one page we don't get to advance it and retrying could\\neliminate those skipped pages and allow us to avoid a vacuum freeze\\nwhich can be really painful. Of course that only works if you can be\\nsure you haven't overflowed and forgotten any skipped pages and if you\\ndon't find the page still pinned every time until you eventually give\\nup on it.\\n\\n-- \\ngreg\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 6 Nov 2014 23:40:18 +0000\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 11/6/14, 5:40 PM, Greg Stark wrote:\\n> On Thu, Nov 6, 2014 at 9:30 PM, Andres Freund <andres@2ndquadrant.com> wrote:\\n>> I think the retry logical is a largely pointless complication of already\\n>> complex enough code. You're fixing a problem for which there is\\n>> absolutely no evidence of its existance. Yes, this happens\\n>> occasionally. But it's going to be so absolutely minor in comparison to\\n>> just about every other source of bloat.\\n\\nFor some reason I don't have Andres' original email, so I'll reply here: I agree with you, and my original proposal was simply to log how many pages were skipped, but that was objected to. Simply logging this extra information would be a patch of a dozen lines or less.\\n\\nThe problem right now is there's no way to actually obtain evidence that this is (or isn't) something to worry about, because we just silently skip pages. If we had any kind of tracking on this we could stop guessing. :(\\n\\n> I agree bloat isn't really a threat, but what about the relfrozenxid?\\n> If we skip even one page we don't get to advance it and retrying could\\n> eliminate those skipped pages and allow us to avoid a vacuum freeze\\n> which can be really painful. Of course that only works if you can be\\n> sure you haven't overflowed and forgotten any skipped pages and if you\\n> don't find the page still pinned every time until you eventually give\\n> up on it.\\n\\nThe overflow part shouldn't be that big a deal. Either we just bump the array size up some more, or worst-case we scan it whenever it fills (like we do when we fill vacrelstats->dead_tuples.\\n\\nBut like I said above, I think this is already making a mountain out of a mole-hill, until we have evidence there's a real problem.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 6 Nov 2014 19:03:20 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-11-06 23:40:18 +0000, Greg Stark wrote:\\n> On Thu, Nov 6, 2014 at 9:30 PM, Andres Freund <andres@2ndquadrant.com> wrote:\\n> > I think the retry logical is a largely pointless complication of already\\n> > complex enough code. You're fixing a problem for which there is\\n> > absolutely no evidence of its existance. Yes, this happens\\n> > occasionally. But it's going to be so absolutely minor in comparison to\\n> > just about every other source of bloat.\\n> \\n> I agree bloat isn't really a threat, but what about the relfrozenxid?\\n> If we skip even one page we don't get to advance it and retrying could\\n> eliminate those skipped pages and allow us to avoid a vacuum freeze\\n> which can be really painful. Of course that only works if you can be\\n> sure you haven't overflowed and forgotten any skipped pages and if you\\n> don't find the page still pinned every time until you eventually give\\n> up on it.\\n\\nI don't buy this argument. Either you're constantly vacuuming the whole\\nrelation anyway - in which case we'll acquire the cleanup lock\\nunconditionally - or we're doing partial vacuums via the visibility map\\nanyway.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 7 Nov 2014 08:27:19 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-11-06 19:03:20 -0600, Jim Nasby wrote:\\n> On 11/6/14, 5:40 PM, Greg Stark wrote:\\n> >On Thu, Nov 6, 2014 at 9:30 PM, Andres Freund <andres@2ndquadrant.com> wrote:\\n> >>I think the retry logical is a largely pointless complication of already\\n> >>complex enough code. You're fixing a problem for which there is\\n> >>absolutely no evidence of its existance. Yes, this happens\\n> >>occasionally. But it's going to be so absolutely minor in comparison to\\n> >>just about every other source of bloat.\\n> \\n> For some reason I don't have Andres' original email, so I'll reply\\n> here: I agree with you, and my original proposal was simply to log how\\n> many pages were skipped, but that was objected to. Simply logging this\\n> extra information would be a patch of a dozen lines or less.\\n\\nThe objection was that it's unneccessary complexity. So you made the\\npatch a magnitude more complex *and* added logging? That doesn't make\\nmuch sense.\\n\\n> The problem right now is there's no way to actually obtain evidence\\n> that this is (or isn't) something to worry about, because we just\\n> silently skip pages. If we had any kind of tracking on this we could\\n> stop guessing. :(\\n\\nWhat's the worst consequence this could have? A couple pages not marked\\nall visible and not immediately cleaned up. That's not particularly\\nharmful.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 7 Nov 2014 08:29:19 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"Greg Stark wrote:\\n\\n> I agree bloat isn't really a threat, but what about the relfrozenxid?\\n> If we skip even one page we don't get to advance it and retrying could\\n> eliminate those skipped pages and allow us to avoid a vacuum freeze\\n> which can be really painful. Of course that only works if you can be\\n> sure you haven't overflowed and forgotten any skipped pages and if you\\n> don't find the page still pinned every time until you eventually give\\n> up on it.\\n\\nWe never advance relfrozenxid nowadays unless it's a whole-table scan;\\nand once we commit to doing those (vacuum_freeze_table_age is past), we\\ndon't skip pages anymore.\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 7 Nov 2014 10:32:03 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 6, 2014 at 8:03 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n> The problem right now is there's no way to actually obtain evidence that\\n> this is (or isn't) something to worry about, because we just silently skip\\n> pages. If we had any kind of tracking on this we could stop guessing. :(\\n\\nI could see logging it, but I agree with Andres and Alvaro that the\\nodds are strongly against there being any actual problem here.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 7 Nov 2014 21:21:22 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 11/7/14, 8:21 PM, Robert Haas wrote:\\n> On Thu, Nov 6, 2014 at 8:03 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n>> The problem right now is there's no way to actually obtain evidence that\\n>> this is (or isn't) something to worry about, because we just silently skip\\n>> pages. If we had any kind of tracking on this we could stop guessing. :(\\n>\\n> I could see logging it, but I agree with Andres and Alvaro that the\\n> odds are strongly against there being any actual problem here.\\n>\\n\\nI'm fine with that. Any other objections? Andres?\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 10 Nov 2014 11:09:08 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"Jim Nasby wrote:\\n> On 11/7/14, 8:21 PM, Robert Haas wrote:\\n> >On Thu, Nov 6, 2014 at 8:03 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n> >>The problem right now is there's no way to actually obtain evidence that\\n> >>this is (or isn't) something to worry about, because we just silently skip\\n> >>pages. If we had any kind of tracking on this we could stop guessing. :(\\n> >\\n> >I could see logging it, but I agree with Andres and Alvaro that the\\n> >odds are strongly against there being any actual problem here.\\n> \\n> I'm fine with that. Any other objections? Andres?\\n\\nIf what we want is to quantify the extent of the issue, would it be more\\nconvenient to save counters to pgstat? Vacuum already sends pgstat\\nmessages, so there's no additional traffic there.\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 10 Nov 2014 14:28:30 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 11/10/14, 11:28 AM, Alvaro Herrera wrote:\\n> Jim Nasby wrote:\\n>> On 11/7/14, 8:21 PM, Robert Haas wrote:\\n>>> On Thu, Nov 6, 2014 at 8:03 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n>>>> The problem right now is there's no way to actually obtain evidence that\\n>>>> this is (or isn't) something to worry about, because we just silently skip\\n>>>> pages. If we had any kind of tracking on this we could stop guessing. :(\\n>>>\\n>>> I could see logging it, but I agree with Andres and Alvaro that the\\n>>> odds are strongly against there being any actual problem here.\\n>>\\n>> I'm fine with that. Any other objections? Andres?\\n>\\n> If what we want is to quantify the extent of the issue, would it be more\\n> convenient to save counters to pgstat? Vacuum already sends pgstat\\n> messages, so there's no additional traffic there.\\n\\nIMHO that would be ideal, but I think Tom was leery of using more space for every table. If we go this route, I'm guessing we should only log pages we skip, and not log pages we had to wait for the lock on (in the case of a freeze). Also, should we still eroprt this even if we are putting it in stats?\\n\\nIs there a way to avoid duplicating the entire eroprt call? I see I could call errstart & friends manually, but currently that's only done in elog.c.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 10 Nov 2014 12:03:59 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-11-10 14:28:30 -0300, Alvaro Herrera wrote:\\n> Jim Nasby wrote:\\n> > On 11/7/14, 8:21 PM, Robert Haas wrote:\\n> > >On Thu, Nov 6, 2014 at 8:03 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n> > >>The problem right now is there's no way to actually obtain evidence that\\n> > >>this is (or isn't) something to worry about, because we just silently skip\\n> > >>pages. If we had any kind of tracking on this we could stop guessing. :(\\n> > >\\n> > >I could see logging it, but I agree with Andres and Alvaro that the\\n> > >odds are strongly against there being any actual problem here.\\n> > \\n> > I'm fine with that. Any other objections? Andres?\\n\\nIf you feel that strong about the need for logging, go ahead.\\n\\n> If what we want is to quantify the extent of the issue, would it be more\\n> convenient to save counters to pgstat? Vacuum already sends pgstat\\n> messages, so there's no additional traffic there.\\n\\nI'm pretty strongly against that one in isolation. They'd need to be\\nstored somewhere and they'd need to be queryable somewhere with enough\\ncontext to make sense. To actually make sense of the numbers we'd also\\nneed to report all the other datapoints of vacuum in some form. That's\\nquite a worthwile project imo - but *much* *much* more work than this.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 10 Nov 2014 19:15:07 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"Andres Freund wrote:\\n> On 2014-11-10 14:28:30 -0300, Alvaro Herrera wrote:\\n\\n> > If what we want is to quantify the extent of the issue, would it be more\\n> > convenient to save counters to pgstat? Vacuum already sends pgstat\\n> > messages, so there's no additional traffic there.\\n> \\n> I'm pretty strongly against that one in isolation. They'd need to be\\n> stored somewhere and they'd need to be queryable somewhere with enough\\n> context to make sense. To actually make sense of the numbers we'd also\\n> need to report all the other datapoints of vacuum in some form. That's\\n> quite a worthwile project imo - but *much* *much* more work than this.\\n\\nWe already have \\\"last_autovacuum\\\" columns and such in pg_stat_tables et\\nal, which only record the last value. My thinking regarding such\\nnumbers is that you would save histories and put them in a chart, see\\nhow they evolve with time. I doubt the individual numbers are worth\\nmuch, but the trends might show something interesting. As far as I\\nknow, this is already true for most other pgstat values, with exception\\nof things such as live_tuples which are absolute numbers rather than\\nrunning counters.\\n\\nI agree having more vacuuming data in general is a worthwhile project,\\nmuch larger than this one. Wasn't Greg Smith working on that?\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 10 Nov 2014 15:36:55 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 11/10/14, 12:15 PM, Andres Freund wrote:\\n>> >If what we want is to quantify the extent of the issue, would it be more\\n>> >convenient to save counters to pgstat? Vacuum already sends pgstat\\n>> >messages, so there's no additional traffic there.\\n> I'm pretty strongly against that one in isolation. They'd need to be\\n> stored somewhere and they'd need to be queryable somewhere with enough\\n> context to make sense. To actually make sense of the numbers we'd also\\n> need to report all the other datapoints of vacuum in some form. That's\\n> quite a worthwile project imo - but*much* *much* more work than this.\\n\\nWe already report statistics on vacuums (lazy_vacuum_rel()/pgstat_report_vacuum), so this would just be adding 1 or 2 counters to that. Should we add the other counters from vacuum? That would be significantly more data.\\n\\nSemi-related, we should probably be reporting stats from heap truncation.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 10 Nov 2014 12:37:29 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-11-10 15:36:55 -0300, Alvaro Herrera wrote:\\n> Andres Freund wrote:\\n> > On 2014-11-10 14:28:30 -0300, Alvaro Herrera wrote:\\n> \\n> > > If what we want is to quantify the extent of the issue, would it be more\\n> > > convenient to save counters to pgstat? Vacuum already sends pgstat\\n> > > messages, so there's no additional traffic there.\\n> > \\n> > I'm pretty strongly against that one in isolation. They'd need to be\\n> > stored somewhere and they'd need to be queryable somewhere with enough\\n> > context to make sense. To actually make sense of the numbers we'd also\\n> > need to report all the other datapoints of vacuum in some form. That's\\n> > quite a worthwile project imo - but *much* *much* more work than this.\\n> \\n> We already have \\\"last_autovacuum\\\" columns and such in pg_stat_tables et\\n> al, which only record the last value. My thinking regarding such\\n> numbers is that you would save histories and put them in a chart, see\\n> how they evolve with time. I doubt the individual numbers are worth\\n> much, but the trends might show something interesting.\\n\\nI don't think they mean anything without also reporting the number of\\nbuffers actually scanned and other related stats.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 10 Nov 2014 19:52:21 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-11-10 12:37:29 -0600, Jim Nasby wrote:\\n> On 11/10/14, 12:15 PM, Andres Freund wrote:\\n> >>>If what we want is to quantify the extent of the issue, would it be more\\n> >>>convenient to save counters to pgstat? Vacuum already sends pgstat\\n> >>>messages, so there's no additional traffic there.\\n> >I'm pretty strongly against that one in isolation. They'd need to be\\n> >stored somewhere and they'd need to be queryable somewhere with enough\\n> >context to make sense. To actually make sense of the numbers we'd also\\n> >need to report all the other datapoints of vacuum in some form. That's\\n> >quite a worthwile project imo - but*much* *much* more work than this.\\n> \\n> We already report statistics on vacuums\\n> (lazy_vacuum_rel()/pgstat_report_vacuum), so this would just be adding\\n> 1 or 2 counters to that. Should we add the other counters from vacuum?\\n> That would be significantly more data.\\n\\nAt the very least it'd require:\\n* The number of buffers skipped due to the vm\\n* The number of buffers actually scanned\\n* The number of full table in contrast to partial vacuums\\n\\nI think it'd require a fair amount of thinking about which values are\\nrequired to make sense of the number of skipped buffers due to not being\\nable to acquire the cleanup lock.\\n\\nIf you want to do this - and I sure don't want to stop you from it - you\\nshould look at it from a general perspective, not from the perspective\\nof how skipped cleanup locks are logged.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 10 Nov 2014 19:56:36 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 11/10/14, 12:56 PM, Andres Freund wrote:\\n> On 2014-11-10 12:37:29 -0600, Jim Nasby wrote:\\n>> On 11/10/14, 12:15 PM, Andres Freund wrote:\\n>>>>> If what we want is to quantify the extent of the issue, would it be more\\n>>>>> convenient to save counters to pgstat? Vacuum already sends pgstat\\n>>>>> messages, so there's no additional traffic there.\\n>>> I'm pretty strongly against that one in isolation. They'd need to be\\n>>> stored somewhere and they'd need to be queryable somewhere with enough\\n>>> context to make sense. To actually make sense of the numbers we'd also\\n>>> need to report all the other datapoints of vacuum in some form. That's\\n>>> quite a worthwile project imo - but*much* *much* more work than this.\\n>>\\n>> We already report statistics on vacuums\\n>> (lazy_vacuum_rel()/pgstat_report_vacuum), so this would just be adding\\n>> 1 or 2 counters to that. Should we add the other counters from vacuum?\\n>> That would be significantly more data.\\n>\\n> At the very least it'd require:\\n> * The number of buffers skipped due to the vm\\n> * The number of buffers actually scanned\\n> * The number of full table in contrast to partial vacuums\\n\\nIf we're going to track full scan vacuums separately, I think we'd need two separate scan counters. I think (for pgstats) it'd make more sense to just count initial failure to acquire the lock in a full scan in the 'skipped page' counter. In terms of answering the question \\\"how common is it not to get the lock\\\", it's really the same event.\\n\\n> I think it'd require a fair amount of thinking about which values are\\n> required to make sense of the number of skipped buffers due to not being\\n> able to acquire the cleanup lock.\\n>\\n> If you want to do this - and I sure don't want to stop you from it - you\\n> should look at it from a general perspective, not from the perspective\\n> of how skipped cleanup locks are logged.\\n\\nHonestly, my desire at this point is just to see if there's actually a problem. Many people are asserting that this should be a very rare occurrence, but there's no way to know.\\n\\nTowards that simple end, I'm a bit torn. My preference would be to simply log, and throw a warning if it's over some threshold. I believe that would give the best odds of getting feedback from users if this isn't as uncommon as we think.\\n\\nI agree that ideally this would be tracked as another stat, but from that standpoint I think there's other, much more important metrics to track, and AFAIK the only reason we don't have them is that busy systems already push pgstats to it's limits. We should try and fix that, but that's a much bigger issue.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 10 Nov 2014 19:36:18 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"Jim Nasby <Jim.Nasby@BlueTreble.com> writes:\\n> On 11/10/14, 12:56 PM, Andres Freund wrote:\\n>> If you want to do this - and I sure don't want to stop you from it - you\\n>> should look at it from a general perspective, not from the perspective\\n>> of how skipped cleanup locks are logged.\\n\\n> Honestly, my desire at this point is just to see if there's actually a problem. Many people are asserting that this should be a very rare occurrence, but there's no way to know.\\n\\n> Towards that simple end, I'm a bit torn. My preference would be to simply log, and throw a warning if it's over some threshold. I believe that would give the best odds of getting feedback from users if this isn't as uncommon as we think.\\n\\n> I agree that ideally this would be tracked as another stat, but from that standpoint I think there's other, much more important metrics to track, and AFAIK the only reason we don't have them is that busy systems already push pgstats to it's limits. We should try and fix that, but that's a much bigger issue.\\n\\nYeah. We know that per-table pgstat counters are a pretty expensive thing\\nin databases with many tables. We should absolutely not be adding them on\\nmere speculation that the number might be interesting.\\n\\nNow, that objection would not apply to a per *database* counter, but I'm\\nnot sure if tracking the numbers at that granularity would help anyone.\\n\\nOn the whole, I'm +1 for just logging the events and seeing what we learn\\nthat way. That seems like an appropriate amount of effort for finding out\\nwhether there is really an issue.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 10 Nov 2014 20:52:59 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-11-10 19:36:18 -0600, Jim Nasby wrote:\\n> On 11/10/14, 12:56 PM, Andres Freund wrote:\\n> >On 2014-11-10 12:37:29 -0600, Jim Nasby wrote:\\n> >>On 11/10/14, 12:15 PM, Andres Freund wrote:\\n> >>>>>If what we want is to quantify the extent of the issue, would it be more\\n> >>>>>convenient to save counters to pgstat? Vacuum already sends pgstat\\n> >>>>>messages, so there's no additional traffic there.\\n> >>>I'm pretty strongly against that one in isolation. They'd need to be\\n> >>>stored somewhere and they'd need to be queryable somewhere with enough\\n> >>>context to make sense. To actually make sense of the numbers we'd also\\n> >>>need to report all the other datapoints of vacuum in some form. That's\\n> >>>quite a worthwile project imo - but*much* *much* more work than this.\\n> >>\\n> >>We already report statistics on vacuums\\n> >>(lazy_vacuum_rel()/pgstat_report_vacuum), so this would just be adding\\n> >>1 or 2 counters to that. Should we add the other counters from vacuum?\\n> >>That would be significantly more data.\\n> >\\n> >At the very least it'd require:\\n> >* The number of buffers skipped due to the vm\\n> >* The number of buffers actually scanned\\n> >* The number of full table in contrast to partial vacuums\\n> \\n> If we're going to track full scan vacuums separately, I think we'd\\n> need two separate scan counters.\\n\\nWell, we already have the entire number of vacuums, so we'd have that.\\n\\n> I think (for pgstats) it'd make more sense to just count initial\\n> failure to acquire the lock in a full scan in the 'skipped page'\\n> counter. In terms of answering the question \\\"how common is it not to\\n> get the lock\\\", it's really the same event.\\n\\nIt's absolutely not. You need to correlate the number of skipped pages\\nto the number of vacuumed pages. If you have 100k skipped in 10 billion\\ntotal scanned pages it's something entirely different than 100k in 200k\\npages.\\n\\n> Honestly, my desire at this point is just to see if there's actually a\\n> problem. Many people are asserting that this should be a very rare\\n> occurrence, but there's no way to know.\\n\\nOk.\\n\\n> Towards that simple end, I'm a bit torn. My preference would be to\\n> simply log, and throw a warning if it's over some threshold. I believe\\n> that would give the best odds of getting feedback from users if this\\n> isn't as uncommon as we think.\\n\\nI'm strongly against a warning. We have absolutely no sane way of tuning\\nthat. We'll just create a pointless warning that people will get\\nconfused about and that they'll have to live with till the next release.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 11 Nov 2014 09:01:58 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 11/11/14, 2:01 AM, Andres Freund wrote:\\n> On 2014-11-10 19:36:18 -0600, Jim Nasby wrote:\\n>> On 11/10/14, 12:56 PM, Andres Freund wrote:\\n>>> On 2014-11-10 12:37:29 -0600, Jim Nasby wrote:\\n>>>> On 11/10/14, 12:15 PM, Andres Freund wrote:\\n>>>>>>> If what we want is to quantify the extent of the issue, would it be more\\n>>>>>>> convenient to save counters to pgstat? Vacuum already sends pgstat\\n>>>>>>> messages, so there's no additional traffic there.\\n>>>>> I'm pretty strongly against that one in isolation. They'd need to be\\n>>>>> stored somewhere and they'd need to be queryable somewhere with enough\\n>>>>> context to make sense. To actually make sense of the numbers we'd also\\n>>>>> need to report all the other datapoints of vacuum in some form. That's\\n>>>>> quite a worthwile project imo - but*much* *much* more work than this.\\n>>>>\\n>>>> We already report statistics on vacuums\\n>>>> (lazy_vacuum_rel()/pgstat_report_vacuum), so this would just be adding\\n>>>> 1 or 2 counters to that. Should we add the other counters from vacuum?\\n>>>> That would be significantly more data.\\n>>>\\n>>> At the very least it'd require:\\n>>> * The number of buffers skipped due to the vm\\n>>> * The number of buffers actually scanned\\n>>> * The number of full table in contrast to partial vacuums\\n>>\\n>> If we're going to track full scan vacuums separately, I think we'd\\n>> need two separate scan counters.\\n>\\n> Well, we already have the entire number of vacuums, so we'd have that.\\n\\nI mean number of pages scanned, but as I said below I don't think that's really necessary.\\n\\n>> I think (for pgstats) it'd make more sense to just count initial\\n>> failure to acquire the lock in a full scan in the 'skipped page'\\n>> counter. In terms of answering the question \\\"how common is it not to\\n>> get the lock\\\", it's really the same event.\\n>\\n> It's absolutely not. You need to correlate the number of skipped pages\\n> to the number of vacuumed pages. If you have 100k skipped in 10 billion\\n> total scanned pages it's something entirely different than 100k in 200k\\n> pages.\\n\\nIf the goal here is to find out if this even is a problem then I think the critical question is not \\\"did we vacuum\\\", but \\\"were we able to acquire the lock on the first try\\\".\\n\\nObviously users will care much more about the vacuuming and not so much about the lock; but if this really is a non-issue as most tend to believe I don't think it's worth worrying about any of this (except perhaps putting dtrace/system tap probes in).\\n\\n>> Honestly, my desire at this point is just to see if there's actually a\\n>> problem. Many people are asserting that this should be a very rare\\n>> occurrence, but there's no way to know.\\n>\\n> Ok.\\n>\\n>> Towards that simple end, I'm a bit torn. My preference would be to\\n>> simply log, and throw a warning if it's over some threshold. I believe\\n>> that would give the best odds of getting feedback from users if this\\n>> isn't as uncommon as we think.\\n>\\n> I'm strongly against a warning. We have absolutely no sane way of tuning\\n> that. We'll just create a pointless warning that people will get\\n> confused about and that they'll have to live with till the next release.\\n\\nTo clarify: I'm only suggesting we issue a warning if we have to skip some significant number of pages; say 5 or 0.01% of the table, whichever is greater. That's aimed directly at the goal of letting us know if this is actually a problem or not.\\n\\nThe reason I'm inclined to do the warning is because I don't think people will notice this otherwise. If this really isn't a problem then it won't matter; if it's a *big* problem then we'll at least know about it.\\n\\nI'm thinking of an undocumented GUC to control the threshold, but I assume no one else would be on board with that?\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 12 Nov 2014 11:34:04 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-11-12 11:34:04 -0600, Jim Nasby wrote:\\n> On 11/11/14, 2:01 AM, Andres Freund wrote:\\n> >On 2014-11-10 19:36:18 -0600, Jim Nasby wrote:\\n> >>Towards that simple end, I'm a bit torn. My preference would be to\\n> >>simply log, and throw a warning if it's over some threshold. I believe\\n> >>that would give the best odds of getting feedback from users if this\\n> >>isn't as uncommon as we think.\\n> >\\n> >I'm strongly against a warning. We have absolutely no sane way of tuning\\n> >that. We'll just create a pointless warning that people will get\\n> >confused about and that they'll have to live with till the next release.\\n> \\n> To clarify: I'm only suggesting we issue a warning if we have to skip some significant number of pages; say 5 or 0.01% of the table, whichever is greater. That's aimed directly at the goal of letting us know if this is actually a problem or not.\\n\\nMeh. You have a 5 page relation and it'll trigger quite easily. And it's\\nabsolutely harmless.\\n\\n> The reason I'm inclined to do the warning is because I don't think people will notice this otherwise. If this really isn't a problem then it won't matter; if it's a *big* problem then we'll at least know about it.\\n> \\n> I'm thinking of an undocumented GUC to control the threshold, but I assume no one else would be on board with that?\\n\\nStop overdesigning this.\\n\\nAdd it to the existing mesage and let us be done with this. This thread\\nhas already wasted far too much time.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 12 Nov 2014 18:37:46 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 12, 2014 at 12:37 PM, Andres Freund <andres@2ndquadrant.com> wrote:\\n> Stop overdesigning this.\\n>\\n> Add it to the existing mesage and let us be done with this. This thread\\n> has already wasted far too much time.\\n\\nThat's a little harsh, but I agree. Producing a warning here is just\\ngoing to be log-spam. We've had this behavior for years and - to my\\nknowledge - we have not got one complaint that can be attributed to\\nthis feature. What used to happen is that VACUUM would get stuck for\\nminutes, hours, or days trying to vacuum a table because of an open\\ncursor, and people did complain about that. It was a serious nuisance\\nthat is now gone. The entire argument in favor of some change here is\\n\\\"even though a careful theoretical analysis indicates that this is\\nsafe, even though it solved a real problem that was hurting our users,\\nand even though we have no evidence whatsoever that the change is\\nhurting anybody in any way whatsoever, the lack of instrumentation\\nmeans that it could possibly be hurting somebody and we wouldn't\\nknow.\\\"\\n\\nThat is, of course, quite true, but you could apply the same argument\\nto lots of patches. It would usually be a waste of time because most\\nof the patches we think are good actually ARE good - but of course,\\nevery once in a while you would find a real problem that had been\\noverlooked.\\n\\nMaybe the right thing here is not to make any change to the source\\ncode *at all* but to patch his own copy and try to come up with a\\nreproducible test case where this is actually a problem. Then, if\\nthere is a problem, we could actually fix it.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 12 Nov 2014 12:53:41 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 11/10/14, 7:52 PM, Tom Lane wrote:\\n> On the whole, I'm +1 for just logging the events and seeing what we learn\\n> that way. That seems like an appropriate amount of effort for finding out\\n> whether there is really an issue.\\n\\nAttached is a patch that does this.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Sun, 30 Nov 2014 20:46:51 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 12/1/14, 11:57 AM, Andres Freund wrote:\\n> On 2014-11-30 20:46:51 -0600, Jim Nasby wrote:\\n>> On 11/10/14, 7:52 PM, Tom Lane wrote:\\n>>> On the whole, I'm +1 for just logging the events and seeing what we learn\\n>>> that way. That seems like an appropriate amount of effort for finding out\\n>>> whether there is really an issue.\\n>>\\n>> Attached is a patch that does this.\\n>\\n> Unless somebody protests I plan to push this soon. I'll change two\\n> things:\\n>\\n> * Always use the same message, independent of scan_all. For one Jim's\\n> version would be untranslatable, for another it's not actually\\n> correct. In most cases we'll *not* wait, even if scan_all is\\n> true as we'll often just balk after !lazy_check_needs_freeze().\\n\\nGood point.\\n\\n> * Change the new bit in the errdetail. \\\"could not acquire cleanup lock\\\"\\n> sounds too much like an error to me. \\\"skipped %u pinned pages\\\" maybe?\\n\\nSeems reasonable.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 1 Dec 2014 12:55:43 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 20 October 2014 at 10:57, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n\\n> Currently, a non-freeze vacuum will punt on any page it can't get a cleanup\\n> lock on, with no retry. Presumably this should be a rare occurrence, but I\\n> think it's bad that we just assume that and won't warn the user if something\\n> bad is going on.\\n\\n(I'm having email problems, so I can't see later mails on this thread,\\nso replying here.)\\n\\nLogging patch looks fine, but I would rather not add a line of text\\nfor each VACUUM, just in case this is non-zero. I think we should add\\nthat log line only if the blocks skipped > 0.\\n\\nWhat I'm more interested in is what you plan to do with the\\ninformation once we get it?\\n\\nThe assumption that skipping blocks is something bad is strange. I\\nadded it because VACUUM could and did regularly hang on busy tables,\\nwhich resulted in bloat because other blocks that needed cleaning\\ndidn't get any attention.\\n\\nWhich is better, spend time obsessively trying to vacuum particular\\nblocks, or to spend the time on other blocks that are in need of\\ncleaning and are available to be cleaned?\\n\\nWhich is better, have autovacuum or system wide vacuum progress on to\\nother tables that need cleaning, or spend lots of effort retrying?\\n\\nHow do we know what is the best next action?\\n\\nI'd really want to see some analysis of those things before we spend\\neven more cycles on this.\\n\\n-- \\n Simon Riggs http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 8 Dec 2014 09:16:11 +0900\",\n \"msg_from\": \"Simon Riggs <simon@2ndQuadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 12/7/14, 6:16 PM, Simon Riggs wrote:\\n> On 20 October 2014 at 10:57, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n>\\n>> Currently, a non-freeze vacuum will punt on any page it can't get a cleanup\\n>> lock on, with no retry. Presumably this should be a rare occurrence, but I\\n>> think it's bad that we just assume that and won't warn the user if something\\n>> bad is going on.\\n>\\n> (I'm having email problems, so I can't see later mails on this thread,\\n> so replying here.)\\n>\\n> Logging patch looks fine, but I would rather not add a line of text\\n> for each VACUUM, just in case this is non-zero. I think we should add\\n> that log line only if the blocks skipped > 0.\\n\\nI thought about doing that, but I'm loath to duplicate a rather large ereport call. Happy to make the change if that's the consensus though.\\n\\n> What I'm more interested in is what you plan to do with the\\n> information once we get it?\\n>\\n> The assumption that skipping blocks is something bad is strange. I\\n> added it because VACUUM could and did regularly hang on busy tables,\\n> which resulted in bloat because other blocks that needed cleaning\\n> didn't get any attention.\\n>\\n> Which is better, spend time obsessively trying to vacuum particular\\n> blocks, or to spend the time on other blocks that are in need of\\n> cleaning and are available to be cleaned?\\n>\\n> Which is better, have autovacuum or system wide vacuum progress on to\\n> other tables that need cleaning, or spend lots of effort retrying?\\n>\\n> How do we know what is the best next action?\\n>\\n> I'd really want to see some analysis of those things before we spend\\n> even more cycles on this.\\n\\nThat's the entire point of logging this information. There is an underlying assumption that we won't actually skip many pages, but there's no data to back that up, nor is there currently any way to get that data.\\n\\nMy hope is that the logging shows that there isn't anything more that needs to be done here. If this is something that causes problems, at least now DBAs will be aware of it and hopefully we'll be able to identify specific problem scenarios and find a solution.\\n\\n\\n\\nBTW, my initial proposal[1] was strictly logging. The only difference was raising it to a warning if a significant portion of the table was skipped. I only investigated retrying locks at the suggestion of others. I never intended this to become a big time sink.\\n\\n[1]:\\n\\\"Currently, a non-freeze vacuum will punt on any page it can't get a cleanup lock on, with no retry. Presumably this should be a rare occurrence, but I think it's bad that we just assume that and won't warn the user if something bad is going on.\\n\\n\\\"My thought is that if we skip any pages elog(LOG) how many we skipped. If we skip more than 1% of the pages we visited (not relpages) then elog(WARNING) instead.\\\"\\n--\\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 8 Dec 2014 15:28:54 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 9 December 2014 at 06:28, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:\\n> On 12/7/14, 6:16 PM, Simon Riggs wrote:\\n\\n>> What I'm more interested in is what you plan to do with the\\n>> information once we get it?\\n>>\\n>> The assumption that skipping blocks is something bad is strange. I\\n>> added it because VACUUM could and did regularly hang on busy tables,\\n>> which resulted in bloat because other blocks that needed cleaning\\n>> didn't get any attention.\\n>>\\n>> Which is better, spend time obsessively trying to vacuum particular\\n>> blocks, or to spend the time on other blocks that are in need of\\n>> cleaning and are available to be cleaned?\\n>>\\n>> Which is better, have autovacuum or system wide vacuum progress on to\\n>> other tables that need cleaning, or spend lots of effort retrying?\\n>>\\n>> How do we know what is the best next action?\\n>>\\n>> I'd really want to see some analysis of those things before we spend\\n>> even more cycles on this.\\n>\\n>\\n> That's the entire point of logging this information. There is an underlying\\n> assumption that we won't actually skip many pages, but there's no data to\\n> back that up, nor is there currently any way to get that data.\\n\\nThere is no such underlying assumption. You assumed there was one, but\\nthere isn't one.\\n\\nAll I can say for certain is that waiting on a lock for long periods\\nwas literally a waste of time. Now it no longer wastes time, it gets\\non with vacuuming the pages it can.\\n\\n-- \\n Simon Riggs http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 9 Dec 2014 07:41:06 +0900\",\n \"msg_from\": \"Simon Riggs <simon@2ndQuadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 12/01/2014 08:55 PM, Jim Nasby wrote:\\n> On 12/1/14, 11:57 AM, Andres Freund wrote:\\n>> On 2014-11-30 20:46:51 -0600, Jim Nasby wrote:\\n>>> On 11/10/14, 7:52 PM, Tom Lane wrote:\\n>>>> On the whole, I'm +1 for just logging the events and seeing what we learn\\n>>>> that way. That seems like an appropriate amount of effort for finding out\\n>>>> whether there is really an issue.\\n>>>\\n>>> Attached is a patch that does this.\\n>>\\n>> Unless somebody protests I plan to push this soon. I'll change two\\n>> things:\\n>>\\n>> * Always use the same message, independent of scan_all. For one Jim's\\n>> version would be untranslatable, for another it's not actually\\n>> correct. In most cases we'll *not* wait, even if scan_all is\\n>> true as we'll often just balk after !lazy_check_needs_freeze().\\n>\\n> Good point.\\n>\\n>> * Change the new bit in the errdetail. \\\"could not acquire cleanup lock\\\"\\n>> sounds too much like an error to me. \\\"skipped %u pinned pages\\\" maybe?\\n>\\n> Seems reasonable.\\n\\nWell, that's not always true either; when freezing, it doesn't skip the \\npinned pages, it waits for them.\\n\\nHow about the attached (I also edited the comments a bit)? It looks like \\nthis:\\n\\npostgres=# vacuum verbose foo;\\nINFO: vacuuming \\\"public.foo\\\"\\nINFO: \\\"foo\\\": found 0 removable, 0 nonremovable row versions in 0 out of \\n7256 pages\\nDETAIL: 0 dead row versions cannot be removed yet.\\nThere were 0 unused item pointers.\\n1 pages were pinned.\\n0 pages are entirely empty.\\nCPU 0.00s/0.00u sec elapsed 0.00 sec.\\nINFO: \\\"foo\\\": stopping truncate due to conflicting lock request\\nINFO: vacuuming \\\"pg_toast.pg_toast_16384\\\"\\nINFO: index \\\"pg_toast_16384_index\\\" now contains 0 row versions in 1 pages\\nDETAIL: 0 index row versions were removed.\\n0 index pages have been deleted, 0 are currently reusable.\\nCPU 0.00s/0.00u sec elapsed 0.00 sec.\\n\\nand for autovacuum log:\\n\\nLOG: automatic vacuum of table \\\"postgres.public.foo\\\": index scans: 0\\n\\tpages: 0 removed, 7256 remain, 0 pinned\\n\\ttuples: 79415 removed, 513156 remain, 0 are dead but not yet removable\\n\\tbuffer usage: 14532 hits, 6 misses, 6241 dirtied\\n\\tavg read rate: 0.003 MB/s, avg write rate: 3.413 MB/s\\n\\tsystem usage: CPU 0.00s/0.30u sec elapsed 14.28 sec\\n\\nI.e. this just says how many pages were pinned, without saying what was \\ndone about them. That's not very meaningful to an average DBA, but \\nthat's true for many of the numbers printed in vacuum verbose.\\n\\n- Heikki\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Wed, 17 Dec 2014 18:20:49 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 12/17/14, 10:20 AM, Heikki Linnakangas wrote:\\n>>\\n>>> * Change the new bit in the errdetail. \\\"could not acquire cleanup lock\\\"\\n>>> sounds too much like an error to me. \\\"skipped %u pinned pages\\\" maybe?\\n>>\\n>> Seems reasonable.\\n>\\n> Well, that's not always true either; when freezing, it doesn't skip the pinned pages, it waits for them.\\n\\nOops. :(\\n\\n> I.e. this just says how many pages were pinned, without saying what was done about them. That's not very meaningful to an average DBA, but that's true for many of the numbers printed in vacuum verbose.\\n\\nIn this case it'll mean people will be less likely to report that this is happening, but maybe that's OK. At least if someone comes to us with a problem we'll be able to get some info from them. I'll separately look into the vacuum docs and see if we can do a better job explaining the verbose output.\\n\\nBTW, what is it about a dynamic message that makes it untranslatable? Doesn't the translation happen down-stream, so that at most we'd just need two translation messages? Or worst case we could have two separate elog calls, if we wanted to go that route.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 17 Dec 2014 12:02:40 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 12/17/2014 08:02 PM, Jim Nasby wrote:\\n> BTW, what is it about a dynamic message that makes it untranslatable?\\n> Doesn't the translation happen down-stream, so that at most we'd just\\n> need two translation messages?\\n\\nI'm not sure what you mean by down-stream. There is some explanation on \\nthis in the localization guide in the manual:\\n\\nhttp://www.postgresql.org/docs/devel/static/nls-programmer.html#NLS-GUIDELINES\\n\\nprintf(\\\"Files were %s.\\\\n\\\", flag ? \\\"copied\\\" : \\\"removed\\\");\\n\\nThe translator will get three strings to translate: \\\"Files were %s.\\\\n\\\", \\n\\\"copied\\\", and \\\"removed\\\".\\n\\n> Or worst case we could have two\\n> separate elog calls, if we wanted to go that route.\\n\\nYeah, that's one option. I.e:\\n\\nif (flag)\\n printf(\\\"Files were copied.\\\\n\\\");\\nelse\\n printf(\\\"Files were removed.\\\\n\\\");\\n\\n- Heikki\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 17 Dec 2014 20:28:39 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"Jim Nasby wrote:\\n> On 12/17/14, 10:20 AM, Heikki Linnakangas wrote:\\n> >>\\n> >>>* Change the new bit in the errdetail. \\\"could not acquire cleanup lock\\\"\\n> >>> sounds too much like an error to me. \\\"skipped %u pinned pages\\\" maybe?\\n> >>\\n> >>Seems reasonable.\\n> >\\n> >Well, that's not always true either; when freezing, it doesn't skip the pinned pages, it waits for them.\\n> \\n> Oops. :(\\n> \\n> >I.e. this just says how many pages were pinned, without saying what was done about them. That's not very meaningful to an average DBA, but that's true for many of the numbers printed in vacuum verbose.\\n> \\n> In this case it'll mean people will be less likely to report that this is happening, but maybe that's OK. At least if someone comes to us with a problem we'll be able to get some info from them. I'll separately look into the vacuum docs and see if we can do a better job explaining the verbose output.\\n> \\n> BTW, what is it about a dynamic message that makes it untranslatable? Doesn't the translation happen down-stream, so that at most we'd just need two translation messages? Or worst case we could have two separate elog calls, if we wanted to go that route.\\n\\nSince each part is a complete sentence, you can paste them together.\\nFor example you could do something like\\n\\nmsg1 = _(\\\"Could not acquire cleanup lock.\\\");\\nmsg2 = _(\\\"Skipped %u pinned pages.\\\");\\n\\nereport(INFO,\\n (errcode(ERRCODE_SOME_STUFF),\\n errmsg_internal(\\\"%s\\\\n%s\\\", msg1, msg2)));\\n\\nThe use of errmsg_internal() is to avoid having it be translated again.\\n\\nFWIW I think the vacuum summary INFO message is pretty terrible already.\\nI would support a patch that changed it to be separately translatable\\nunits (as long as each unit is a complete sentence.)\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 17 Dec 2014 15:39:39 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 17, 2014 at 11:20 AM, Heikki Linnakangas\\n<hlinnakangas@vmware.com> wrote:\\n> LOG: automatic vacuum of table \\\"postgres.public.foo\\\": index scans: 0\\n> pages: 0 removed, 7256 remain, 0 pinned\\n> tuples: 79415 removed, 513156 remain, 0 are dead but not yet\\n> removable\\n> buffer usage: 14532 hits, 6 misses, 6241 dirtied\\n> avg read rate: 0.003 MB/s, avg write rate: 3.413 MB/s\\n> system usage: CPU 0.00s/0.30u sec elapsed 14.28 sec\\n>\\n> I.e. this just says how many pages were pinned, without saying what was done\\n> about them. That's not very meaningful to an average DBA, but that's true\\n> for many of the numbers printed in vacuum verbose.\\n\\nThat message is extremely confusing, to my eyes. If you want to say\\n\\\"pages: 0 removed, 7256 remain, 0 skipped due to pins\\\", that would\\nwork for me, but just say \\\"0 pinned\\\" is totally wrong, because vacuum\\npinned every page in the table.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 08:56:56 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 12/18/14, 7:56 AM, Robert Haas wrote:\\n> On Wed, Dec 17, 2014 at 11:20 AM, Heikki Linnakangas\\n> <hlinnakangas@vmware.com> wrote:\\n>> LOG: automatic vacuum of table \\\"postgres.public.foo\\\": index scans: 0\\n>> pages: 0 removed, 7256 remain, 0 pinned\\n>> tuples: 79415 removed, 513156 remain, 0 are dead but not yet\\n>> removable\\n>> buffer usage: 14532 hits, 6 misses, 6241 dirtied\\n>> avg read rate: 0.003 MB/s, avg write rate: 3.413 MB/s\\n>> system usage: CPU 0.00s/0.30u sec elapsed 14.28 sec\\n>>\\n>> I.e. this just says how many pages were pinned, without saying what was done\\n>> about them. That's not very meaningful to an average DBA, but that's true\\n>> for many of the numbers printed in vacuum verbose.\\n>\\n> That message is extremely confusing, to my eyes. If you want to say\\n> \\\"pages: 0 removed, 7256 remain, 0 skipped due to pins\\\", that would\\n> work for me, but just say \\\"0 pinned\\\" is totally wrong, because vacuum\\n> pinned every page in the table.\\n\\nWe have to decide on a tradeoff here. Either we end up with two different log messages (depending on scan_all) that require two different translations, or we end up with a generic message that isn't as clear.\\n\\nThe best option I can think of for the later is something like \\\"failed initial lock attempt\\\". That's the only thing that will be true in all cases.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 11:38:12 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"Jim Nasby wrote:\\n\\n> We have to decide on a tradeoff here. Either we end up with two\\n> different log messages (depending on scan_all) that require two\\n> different translations, or we end up with a generic message that isn't\\n> as clear.\\n> \\n> The best option I can think of for the later is something like \\\"failed\\n> initial lock attempt\\\". That's the only thing that will be true in all\\n> cases.\\n\\nHere's my proposal. Instead of punting, I split the message in\\nseparately translatable units, and emit only the ones that apply. The\\ncode is messier this way, but I think we can live with that.\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Thu, 18 Dec 2014 16:41:04 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 12/18/2014 09:41 PM, Alvaro Herrera wrote:\\n> Jim Nasby wrote:\\n>\\n>> We have to decide on a tradeoff here. Either we end up with two\\n>> different log messages (depending on scan_all) that require two\\n>> different translations, or we end up with a generic message that isn't\\n>> as clear.\\n>>\\n>> The best option I can think of for the later is something like \\\"failed\\n>> initial lock attempt\\\". That's the only thing that will be true in all\\n>> cases.\\n>\\n> Here's my proposal. Instead of punting, I split the message in\\n> separately translatable units, and emit only the ones that apply. The\\n> code is messier this way, but I think we can live with that.\\n\\nWorks for me.\\n\\n- Heikki\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 22:06:57 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"Heikki Linnakangas wrote:\\n> On 12/18/2014 09:41 PM, Alvaro Herrera wrote:\\n\\n> >Here's my proposal. Instead of punting, I split the message in\\n> >separately translatable units, and emit only the ones that apply. The\\n> >code is messier this way, but I think we can live with that.\\n> \\n> Works for me.\\n\\nGreat, thanks, pushed. I tweaked it a bit more, so that it would say\\neither \\\"skipped N pages\\\" or \\\"waited N pins\\\" in both autovacuum and\\nvacuum verbose cases, but only if N > 0.\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 17:21:36 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@2ndquadrant.com> writes:\\n> Great, thanks, pushed. I tweaked it a bit more, so that it would say\\n> either \\\"skipped N pages\\\" or \\\"waited N pins\\\" in both autovacuum and\\n> vacuum verbose cases, but only if N > 0.\\n\\nNot directly relevant but ... I think probably all those BlockNumber\\ncounters should be printed with %u not %d. A relation with between\\n2G and 4G pages is possible, and it wouldn't look right with %d.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 15:50:31 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-12-18 16:41:04 -0300, Alvaro Herrera wrote:\\n> +\\t\\t\\tif (scan_all)\\n> +\\t\\t\\t\\tappendStringInfo(&buf, _(\\\"waited for %d buffer pins\\\\n\\\"),\\n> +\\t\\t\\t\\t\\t\\t\\t\\t vacrelstats->pinned_pages);\\n> +\\t\\t\\telse\\n> +\\t\\t\\t\\tappendStringInfo(&buf,\\n> +\\t\\t\\t\\t\\t\\t\\t\\t _(\\\"skipped %d pages due to buffer pins\\\\n\\\"),\\n> +\\t\\t\\t\\t\\t\\t\\t\\t vacrelstats->pinned_pages);\\n\\nUnless I miss something this is, as mentioned before, not\\ncorrect. scan_all doesn't imply at all that we waited for buffer\\npins. We only do so if lazy_check_needs_freeze(buf). Which usually won't\\nbe true for a *significant* number of pages.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 21:55:05 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"Andres Freund wrote:\\n> On 2014-12-18 16:41:04 -0300, Alvaro Herrera wrote:\\n> > +\\t\\t\\tif (scan_all)\\n> > +\\t\\t\\t\\tappendStringInfo(&buf, _(\\\"waited for %d buffer pins\\\\n\\\"),\\n> > +\\t\\t\\t\\t\\t\\t\\t\\t vacrelstats->pinned_pages);\\n> > +\\t\\t\\telse\\n> > +\\t\\t\\t\\tappendStringInfo(&buf,\\n> > +\\t\\t\\t\\t\\t\\t\\t\\t _(\\\"skipped %d pages due to buffer pins\\\\n\\\"),\\n> > +\\t\\t\\t\\t\\t\\t\\t\\t vacrelstats->pinned_pages);\\n> \\n> Unless I miss something this is, as mentioned before, not\\n> correct. scan_all doesn't imply at all that we waited for buffer\\n> pins. We only do so if lazy_check_needs_freeze(buf). Which usually won't\\n> be true for a *significant* number of pages.\\n\\nAh, interesting, I didn't remember we had that. I guess one possible\\ntweak is to discount the pages we skip from pinned_pages; or we could\\nkeep a separate count of pages waited for. Jim, up for a patch?\\n\\n-- \\n�lvaro Herrera http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 18:02:54 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 12/18/14, 3:02 PM, Alvaro Herrera wrote:\\n> Andres Freund wrote:\\n>> On 2014-12-18 16:41:04 -0300, Alvaro Herrera wrote:\\n>>> +\\t\\t\\tif (scan_all)\\n>>> +\\t\\t\\t\\tappendStringInfo(&buf, _(\\\"waited for %d buffer pins\\\\n\\\"),\\n>>> +\\t\\t\\t\\t\\t\\t\\t\\t vacrelstats->pinned_pages);\\n>>> +\\t\\t\\telse\\n>>> +\\t\\t\\t\\tappendStringInfo(&buf,\\n>>> +\\t\\t\\t\\t\\t\\t\\t\\t _(\\\"skipped %d pages due to buffer pins\\\\n\\\"),\\n>>> +\\t\\t\\t\\t\\t\\t\\t\\t vacrelstats->pinned_pages);\\n>>\\n>> Unless I miss something this is, as mentioned before, not\\n>> correct. scan_all doesn't imply at all that we waited for buffer\\n>> pins. We only do so if lazy_check_needs_freeze(buf). Which usually won't\\n>> be true for a *significant* number of pages.\\n>\\n> Ah, interesting, I didn't remember we had that. I guess one possible\\n> tweak is to discount the pages we skip from pinned_pages; or we could\\n> keep a separate count of pages waited for. Jim, up for a patch?\\n\\nI would prefer that we at least count if we initially don't get the lock; presumably that number is always low anyway and in that case I think we're done with this. If it turns out it is common to initially miss the pin then we could do something fancier.\\n\\nSo how about if in the scan_all case we say something like \\\"unable to initially acquire pin on %d buffers\\\\n\\\"?\\n\\n(Happy to do the patch either way, but I'd like us to decide what we're doing first. ;)\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 18 Dec 2014 16:05:23 -0600\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-12-18 16:05:23 -0600, Jim Nasby wrote:\\n> On 12/18/14, 3:02 PM, Alvaro Herrera wrote:\\n> >Andres Freund wrote:\\n> >>On 2014-12-18 16:41:04 -0300, Alvaro Herrera wrote:\\n> >>>+\\t\\t\\tif (scan_all)\\n> >>>+\\t\\t\\t\\tappendStringInfo(&buf, _(\\\"waited for %d buffer pins\\\\n\\\"),\\n> >>>+\\t\\t\\t\\t\\t\\t\\t\\t vacrelstats->pinned_pages);\\n> >>>+\\t\\t\\telse\\n> >>>+\\t\\t\\t\\tappendStringInfo(&buf,\\n> >>>+\\t\\t\\t\\t\\t\\t\\t\\t _(\\\"skipped %d pages due to buffer pins\\\\n\\\"),\\n> >>>+\\t\\t\\t\\t\\t\\t\\t\\t vacrelstats->pinned_pages);\\n> >>\\n> >>Unless I miss something this is, as mentioned before, not\\n> >>correct. scan_all doesn't imply at all that we waited for buffer\\n> >>pins. We only do so if lazy_check_needs_freeze(buf). Which usually won't\\n> >>be true for a *significant* number of pages.\\n> >\\n> >Ah, interesting, I didn't remember we had that. I guess one possible\\n> >tweak is to discount the pages we skip from pinned_pages; or we could\\n> >keep a separate count of pages waited for. Jim, up for a patch?\\n> \\n> I would prefer that we at least count if we initially don't get the lock; presumably that number is always low anyway and in that case I think we're done with this. If it turns out it is common to initially miss the pin then we could do something fancier.\\n> \\n> So how about if in the scan_all case we say something like \\\"unable to initially acquire pin on %d buffers\\\\n\\\"?\\n\\nI'd just do away with the difference between scan_all/!scan_all and\\nalways say the above.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 19 Dec 2014 08:50:35 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2014-12-18 16:05:23 -0600, Jim Nasby wrote:\\n> On 12/18/14, 3:02 PM, Alvaro Herrera wrote:\\n> >Andres Freund wrote:\\n> >>On 2014-12-18 16:41:04 -0300, Alvaro Herrera wrote:\\n> >>>+\\t\\t\\tif (scan_all)\\n> >>>+\\t\\t\\t\\tappendStringInfo(&buf, _(\\\"waited for %d buffer pins\\\\n\\\"),\\n> >>>+\\t\\t\\t\\t\\t\\t\\t\\t vacrelstats->pinned_pages);\\n> >>>+\\t\\t\\telse\\n> >>>+\\t\\t\\t\\tappendStringInfo(&buf,\\n> >>>+\\t\\t\\t\\t\\t\\t\\t\\t _(\\\"skipped %d pages due to buffer pins\\\\n\\\"),\\n> >>>+\\t\\t\\t\\t\\t\\t\\t\\t vacrelstats->pinned_pages);\\n> >>\\n> >>Unless I miss something this is, as mentioned before, not\\n> >>correct. scan_all doesn't imply at all that we waited for buffer\\n> >>pins. We only do so if lazy_check_needs_freeze(buf). Which usually won't\\n> >>be true for a *significant* number of pages.\\n\\nAlso, naming the number of pages we could *not* pin, pinned_pages?\\nReally?\\n\\npinskipped_pages,skipped_pages,unpinned_pages,...\\n\\n> >Ah, interesting, I didn't remember we had that. I guess one possible\\n> >tweak is to discount the pages we skip from pinned_pages; or we could\\n> >keep a separate count of pages waited for. Jim, up for a patch?\\n\\nThis is still wrong. I think just counting skipped pages, without\\ndistinct messages for waiting/not waiting, is good enough for\\nnow. Everything else would only be actually meaningful if we actually\\ntracked the waiting time.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 4 Jan 2015 01:53:24 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n },\n {\n \"msg_contents\": \"On 2015-01-04 01:53:24 +0100, Andres Freund wrote:\\n> > >Ah, interesting, I didn't remember we had that. I guess one possible\\n> > >tweak is to discount the pages we skip from pinned_pages; or we could\\n> > >keep a separate count of pages waited for. Jim, up for a patch?\\n>\\n> This is still wrong. I think just counting skipped pages, without\\n> distinct messages for waiting/not waiting, is good enough for\\n> now. Everything else would only be actually meaningful if we actually\\n> tracked the waiting time.\\n\\nPushed a commit for this, with additional improvements to autovacuum's\\nlog output from:\\nLOG: automatic vacuum of table \\\"postgres.public.frak\\\": index scans: 0\\n pages: 0 removed, 1672 remain\\n skipped 1 pages due to buffer pins\\n tuples: 0 removed, 309959 remain, 309774 are dead but not yet removable\\n buffer usage: 4258 hits, 0 misses, 0 dirtied\\n avg read rate: 0.000 MB/s, avg write rate: 0.000 MB/s\\n system usage: CPU 0.00s/0.04u sec elapsed 0.46 sec\\nto:\\nLOG: automatic vacuum of table \\\"postgres.public.frak\\\": index scans: 0\\n pages: 0 removed, 1672 remain, 1 skipped due to pins\\n tuples: 0 removed, 309959 remain, 309774 are dead but not yet removable\\n buffer usage: 4258 hits, 0 misses, 0 dirtied\\n avg read rate: 0.000 MB/s, avg write rate: 0.000 MB/s\\n system usage: CPU 0.00s/0.04u sec elapsed 0.46 sec\\nas the 'skipped ...' line didn't really look in line with the rest.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n--\\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 8 Jan 2015 13:19:23 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Proposal: Log inability to lock pages during vacuum\"\n }\n]","isConversational":false}}},{"rowIdx":39687,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"I ran into a type \" a a \" in a comment in snapmgr.c, and found that\nthere are four other places that've made the same typo, by the grep\ncommand. And in one of those places, I found yet another typo. Please\nfind attached a patch.\n\nThanks,\n\nBest regards,\nEtsuro Fujita\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Mon, 20 Oct 2014 14:54:05 +0900","msg_from":"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>","msg_from_op":true,"msg_subject":"Typos in comments"},{"msg_contents":"On Mon, Oct 20, 2014 at 6:54 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\nwrote:\n\n> I ran into a type \" a a \" in a comment in snapmgr.c, and found that\n> there are four other places that've made the same typo, by the grep\n> command. And in one of those places, I found yet another typo. Please\n> find attached a patch.\n>\n>\nJust while on this topic, I had a quick look at the results from the regex\n\\b(\\w+)\\s+\\1\\b which finds duplicate words. After sifting through the false\npositives I found a few more. Patch attached.\n\nRegards\n\nDavid Rowley\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Mon, 20 Oct 2014 20:04:47 +1300","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: Typos in comments"},{"msg_contents":"On Mon, Oct 20, 2014 at 3:04 AM, David Rowley <dgrowleyml@gmail.com> wrote:\n> On Mon, Oct 20, 2014 at 6:54 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\n> wrote:\n>> I ran into a type \" a a \" in a comment in snapmgr.c, and found that\n>> there are four other places that've made the same typo, by the grep\n>> command. And in one of those places, I found yet another typo. Please\n>> find attached a patch.\n>\n> Just while on this topic, I had a quick look at the results from the regex\n> \\b(\\w+)\\s+\\1\\b which finds duplicate words. After sifting through the false\n> positives I found a few more. Patch attached.\n\nThanks. I have committed both patches.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 10:34:34 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Typos in comments"}],"string":"[\n {\n \"msg_contents\": \"I ran into a type \\\" a a \\\" in a comment in snapmgr.c, and found that\\nthere are four other places that've made the same typo, by the grep\\ncommand. And in one of those places, I found yet another typo. Please\\nfind attached a patch.\\n\\nThanks,\\n\\nBest regards,\\nEtsuro Fujita\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Mon, 20 Oct 2014 14:54:05 +0900\",\n \"msg_from\": \"Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Typos in comments\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 20, 2014 at 6:54 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\\nwrote:\\n\\n> I ran into a type \\\" a a \\\" in a comment in snapmgr.c, and found that\\n> there are four other places that've made the same typo, by the grep\\n> command. And in one of those places, I found yet another typo. Please\\n> find attached a patch.\\n>\\n>\\nJust while on this topic, I had a quick look at the results from the regex\\n\\\\b(\\\\w+)\\\\s+\\\\1\\\\b which finds duplicate words. After sifting through the false\\npositives I found a few more. Patch attached.\\n\\nRegards\\n\\nDavid Rowley\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Mon, 20 Oct 2014 20:04:47 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typos in comments\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 20, 2014 at 3:04 AM, David Rowley <dgrowleyml@gmail.com> wrote:\\n> On Mon, Oct 20, 2014 at 6:54 PM, Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp>\\n> wrote:\\n>> I ran into a type \\\" a a \\\" in a comment in snapmgr.c, and found that\\n>> there are four other places that've made the same typo, by the grep\\n>> command. And in one of those places, I found yet another typo. Please\\n>> find attached a patch.\\n>\\n> Just while on this topic, I had a quick look at the results from the regex\\n> \\\\b(\\\\w+)\\\\s+\\\\1\\\\b which finds duplicate words. After sifting through the false\\n> positives I found a few more. Patch attached.\\n\\nThanks. I have committed both patches.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 10:34:34 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typos in comments\"\n }\n]","isConversational":false}}},{"rowIdx":39688,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"hello,I am reading the pool manager code,and have some problem:\n\nin poolmgr.c function agent_init():\n\nif (agent->pool)\n> agent_release_connections(agent, false);\n\n......\n\n#ifdef XCP\n> /* find database */\n> agent->pool = find_database_pool(database, user_name);\n> /* create if not found */\n> if (agent->pool == NULL)\n> agent->pool = create_database_pool(database, user_name);\n\n\nin concurrent running,this function with the same dbname and\nusername，create_database_pool(database, user_name) may be run twice,so\nthere may be more than one DatabasePool in the databasePools list? But\nfind_database_pool(database, user_name) will always return the first\nfind,other may be not used,there may be memory leak?\n\nand why this function call agent_release_connections,and then use\nfind_database_pool to get a new pool? For concurrent running,is it possible\none agent release the connections which other agent still is in use?\n\nthanks!\n\nJov\nblog: http:amutu.com/blog <http://amutu.com/blog>\n\nhello,I am reading the pool manager code,and have some problem:in poolmgr.c function agent_init(): if (agent->pool) agent_release_connections(agent, false); ...... #ifdef XCP /* find database */ agent->pool = find_database_pool(database, user_name); /* create if not found */ if (agent->pool == NULL) agent->pool = create_database_pool(database, user_name);in concurrent running,this function with the same dbname and username，create_database_pool(database, user_name) may be run twice,so there may be more than one DatabasePool in the databasePools list? But find_database_pool(database, user_name) will always return the first find,other may be not used,there may be memory leak?and why this function call agent_release_connections,and then use find_database_pool to get a new pool? For concurrent running,is it possible one agent release the connections which other agent still is in use?thanks!Jovblog: http:amutu.com/blog","msg_date":"Mon, 20 Oct 2014 15:29:46 +0800","msg_from":"Jov <amutu@amutu.com>","msg_from_op":true,"msg_subject":"agent_init concurrent running question"},{"msg_contents":"On Mon, Oct 20, 2014 at 4:29 PM, Jov <amutu@amutu.com> wrote:\n\n> in poolmgr.c function agent_init():\n>\nThis code is part of Postgres-XC (or Postgres-XL) and not PostgreSQL\nitself, you should send your questions there:\nhttps://lists.sourceforge.net/lists/listinfo/postgres-xc-developers\nRegards,\n-- \nMichael\n\nOn Mon, Oct 20, 2014 at 4:29 PM, Jov <amutu@amutu.com> wrote:in poolmgr.c function agent_init():This code is part of Postgres-XC (or Postgres-XL) and not PostgreSQL itself, you should send your questions there:https://lists.sourceforge.net/lists/listinfo/postgres-xc-developersRegards,-- Michael","msg_date":"Mon, 20 Oct 2014 16:51:38 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: agent_init concurrent running question"}],"string":"[\n {\n \"msg_contents\": \"hello,I am reading the pool manager code,and have some problem:\\n\\nin poolmgr.c function agent_init():\\n\\nif (agent->pool)\\n> agent_release_connections(agent, false);\\n\\n......\\n\\n#ifdef XCP\\n> /* find database */\\n> agent->pool = find_database_pool(database, user_name);\\n> /* create if not found */\\n> if (agent->pool == NULL)\\n> agent->pool = create_database_pool(database, user_name);\\n\\n\\nin concurrent running,this function with the same dbname and\\nusername，create_database_pool(database, user_name) may be run twice,so\\nthere may be more than one DatabasePool in the databasePools list? But\\nfind_database_pool(database, user_name) will always return the first\\nfind,other may be not used,there may be memory leak?\\n\\nand why this function call agent_release_connections,and then use\\nfind_database_pool to get a new pool? For concurrent running,is it possible\\none agent release the connections which other agent still is in use?\\n\\nthanks!\\n\\nJov\\nblog: http:amutu.com/blog <http://amutu.com/blog>\\n\\nhello,I am reading the pool manager code,and have some problem:in poolmgr.c function agent_init(): if (agent->pool) agent_release_connections(agent, false); ...... #ifdef XCP /* find database */ agent->pool = find_database_pool(database, user_name); /* create if not found */ if (agent->pool == NULL) agent->pool = create_database_pool(database, user_name);in concurrent running,this function with the same dbname and username，create_database_pool(database, user_name) may be run twice,so there may be more than one DatabasePool in the databasePools list? But find_database_pool(database, user_name) will always return the first find,other may be not used,there may be memory leak?and why this function call agent_release_connections,and then use find_database_pool to get a new pool? For concurrent running,is it possible one agent release the connections which other agent still is in use?thanks!Jovblog: http:amutu.com/blog\",\n \"msg_date\": \"Mon, 20 Oct 2014 15:29:46 +0800\",\n \"msg_from\": \"Jov <amutu@amutu.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"agent_init concurrent running question\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 20, 2014 at 4:29 PM, Jov <amutu@amutu.com> wrote:\\n\\n> in poolmgr.c function agent_init():\\n>\\nThis code is part of Postgres-XC (or Postgres-XL) and not PostgreSQL\\nitself, you should send your questions there:\\nhttps://lists.sourceforge.net/lists/listinfo/postgres-xc-developers\\nRegards,\\n-- \\nMichael\\n\\nOn Mon, Oct 20, 2014 at 4:29 PM, Jov <amutu@amutu.com> wrote:in poolmgr.c function agent_init():This code is part of Postgres-XC (or Postgres-XL) and not PostgreSQL itself, you should send your questions there:https://lists.sourceforge.net/lists/listinfo/postgres-xc-developersRegards,-- Michael\",\n \"msg_date\": \"Mon, 20 Oct 2014 16:51:38 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: agent_init concurrent running question\"\n }\n]","isConversational":false}}},{"rowIdx":39689,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\n\nExcept if I am missing something, is there any reason why the sequence used\nin XLogArchiveCheckDone and XLogArchiveIsBusy to check if a XLOG segment\nhas been already archived is duplicated? I guess that doing a little bit of\nrefactoring here would make sense for simplicity, patch is attached.\nRegards,\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Mon, 20 Oct 2014 16:40:58 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":true,"msg_subject":"Refactor status detection for XLOG segment file in xlogarchive.c\n (Was: *FF WALs under 9.2)"}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\n\\nExcept if I am missing something, is there any reason why the sequence used\\nin XLogArchiveCheckDone and XLogArchiveIsBusy to check if a XLOG segment\\nhas been already archived is duplicated? I guess that doing a little bit of\\nrefactoring here would make sense for simplicity, patch is attached.\\nRegards,\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Mon, 20 Oct 2014 16:40:58 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Refactor status detection for XLOG segment file in xlogarchive.c\\n (Was: *FF WALs under 9.2)\"\n }\n]","isConversational":false}}},{"rowIdx":39690,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Dear All,\n L've got a question for you.\nIf the value of the integer type is converted to a value of type Boolean，PostgreSQL will display just like\"The rule has already exist\";\n1. CREATE CAST (integer AS bool) WITH INOUT AS IMPLICIT; \n\nERROR: cast from type integer to type boolean already exists\n If I want to drop the rule,will be told like\"From integer to boolean type rules may not be dropped\";\n2. DROP CAST (integer AS bool) ;\nERROR: cannot drop cast from integer to boolean because it is required by the database system\n Since the rule has already exist,Why execute the following methods will be in error?\n3. insert into testbool(p1,p2)values(20,0::integer);\nERROR: column \"p2\" is of type boolean but expression is of type integer\n So how to deal with this kind of situation if I want a implicit conversion?\n Best Regards,\n rohtodeveloper\n \n \t\t \t \t\t \n\n\n\nDear All,\n L've got a question for you.\nIf the value of the integer type is converted to a value of type Boolean，PostgreSQL will display just like\"The rule has already exist\";\n1. CREATE CAST (integer AS bool) WITH INOUT AS IMPLICIT; \n\nERROR: cast from type integer to type boolean already exists\n If I want to drop the rule,will be told like\"From integer to boolean type rules may not be dropped\";\n2. DROP CAST (integer AS bool) ;\nERROR: cannot drop cast from integer to boolean because it is required by the database system\n Since the rule has already exist,Why execute the following methods will be in error?\n3. insert into testbool(p1,p2)values(20,0::integer);\nERROR: column \"p2\" is of type boolean but expression is of type integer\n So how to deal with this kind of situation if I want a implicit conversion?\n Best Regards,\n rohtodeveloper","msg_date":"Mon, 20 Oct 2014 14:34:25 +0000","msg_from":"rohtodeveloper <rohtodeveloper@outlook.com>","msg_from_op":true,"msg_subject":"Would you help to review our modifications"},{"msg_contents":"rohtodeveloper wrote\n> So how to deal with this kind of situation if I want a implicit\n> conversion?\n\nAs of the out-of-support 8.3 release many of the implicit casts previously\ndefined have been changed to explicit casts. It is a catalog change -\nobviously, since you can still define implicit casts - so if you absolutely\nmust have the pre-existing cast be implicit you can modify the catalog\ndirectly.\n\nYou may wish to describe why you think this is the solution you need - with\nimplicit casting there are generally more downsides that upsides.\n\nDave\n\n\n\n\n\n--\nView this message in context: http://postgresql.1045698.n5.nabble.com/Would-you-help-to-review-our-modifications-tp5823666p5823679.html\nSent from the PostgreSQL - hackers mailing list archive at Nabble.com.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 09:02:32 -0700 (PDT)","msg_from":"David G Johnston <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: Would you help to review our modifications"},{"msg_contents":"On Mon, Oct 20, 2014 at 11:02 AM, David G Johnston\n<david.g.johnston@gmail.com> wrote:\n> rohtodeveloper wrote\n>> So how to deal with this kind of situation if I want a implicit\n>> conversion?\n>\n> As of the out-of-support 8.3 release many of the implicit casts previously\n> defined have been changed to explicit casts. It is a catalog change -\n> obviously, since you can still define implicit casts - so if you absolutely\n> must have the pre-existing cast be implicit you can modify the catalog\n> directly.\n>\n> You may wish to describe why you think this is the solution you need - with\n> implicit casting there are generally more downsides that upsides.\n\nI feel your pain. My company just last year completed a nine month\neffort to validate a sprawling code base for post 8.3 casts. We were\norphaned on 8.1 and were very nearly forced to switch to another\ndatabase.\n\nmerlin\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 08:16:30 -0500","msg_from":"Merlin Moncure <mmoncure@gmail.com>","msg_from_op":false,"msg_subject":"Re: Would you help to review our modifications"}],"string":"[\n {\n \"msg_contents\": \"Dear All,\\n L've got a question for you.\\nIf the value of the integer type is converted to a value of type Boolean，PostgreSQL will display just like\\\"The rule has already exist\\\";\\n1. CREATE CAST (integer AS bool) WITH INOUT AS IMPLICIT; \\n\\nERROR: cast from type integer to type boolean already exists\\n If I want to drop the rule,will be told like\\\"From integer to boolean type rules may not be dropped\\\";\\n2. DROP CAST (integer AS bool) ;\\nERROR: cannot drop cast from integer to boolean because it is required by the database system\\n Since the rule has already exist,Why execute the following methods will be in error?\\n3. insert into testbool(p1,p2)values(20,0::integer);\\nERROR: column \\\"p2\\\" is of type boolean but expression is of type integer\\n So how to deal with this kind of situation if I want a implicit conversion?\\n Best Regards,\\n rohtodeveloper\\n \\n \\t\\t \\t \\t\\t \\n\\n\\n\\nDear All,\\n L've got a question for you.\\nIf the value of the integer type is converted to a value of type Boolean，PostgreSQL will display just like\\\"The rule has already exist\\\";\\n1. CREATE CAST (integer AS bool) WITH INOUT AS IMPLICIT; \\n\\nERROR: cast from type integer to type boolean already exists\\n If I want to drop the rule,will be told like\\\"From integer to boolean type rules may not be dropped\\\";\\n2. DROP CAST (integer AS bool) ;\\nERROR: cannot drop cast from integer to boolean because it is required by the database system\\n Since the rule has already exist,Why execute the following methods will be in error?\\n3. insert into testbool(p1,p2)values(20,0::integer);\\nERROR: column \\\"p2\\\" is of type boolean but expression is of type integer\\n So how to deal with this kind of situation if I want a implicit conversion?\\n Best Regards,\\n rohtodeveloper\",\n \"msg_date\": \"Mon, 20 Oct 2014 14:34:25 +0000\",\n \"msg_from\": \"rohtodeveloper <rohtodeveloper@outlook.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Would you help to review our modifications\"\n },\n {\n \"msg_contents\": \"rohtodeveloper wrote\\n> So how to deal with this kind of situation if I want a implicit\\n> conversion?\\n\\nAs of the out-of-support 8.3 release many of the implicit casts previously\\ndefined have been changed to explicit casts. It is a catalog change -\\nobviously, since you can still define implicit casts - so if you absolutely\\nmust have the pre-existing cast be implicit you can modify the catalog\\ndirectly.\\n\\nYou may wish to describe why you think this is the solution you need - with\\nimplicit casting there are generally more downsides that upsides.\\n\\nDave\\n\\n\\n\\n\\n\\n--\\nView this message in context: http://postgresql.1045698.n5.nabble.com/Would-you-help-to-review-our-modifications-tp5823666p5823679.html\\nSent from the PostgreSQL - hackers mailing list archive at Nabble.com.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 09:02:32 -0700 (PDT)\",\n \"msg_from\": \"David G Johnston <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Would you help to review our modifications\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 20, 2014 at 11:02 AM, David G Johnston\\n<david.g.johnston@gmail.com> wrote:\\n> rohtodeveloper wrote\\n>> So how to deal with this kind of situation if I want a implicit\\n>> conversion?\\n>\\n> As of the out-of-support 8.3 release many of the implicit casts previously\\n> defined have been changed to explicit casts. It is a catalog change -\\n> obviously, since you can still define implicit casts - so if you absolutely\\n> must have the pre-existing cast be implicit you can modify the catalog\\n> directly.\\n>\\n> You may wish to describe why you think this is the solution you need - with\\n> implicit casting there are generally more downsides that upsides.\\n\\nI feel your pain. My company just last year completed a nine month\\neffort to validate a sprawling code base for post 8.3 casts. We were\\norphaned on 8.1 and were very nearly forced to switch to another\\ndatabase.\\n\\nmerlin\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 08:16:30 -0500\",\n \"msg_from\": \"Merlin Moncure <mmoncure@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Would you help to review our modifications\"\n }\n]","isConversational":false}}},{"rowIdx":39691,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"My Salesforce colleague Thomas Fanghaenel observed that the TAP tests\nfor pg_basebackup fail when run in a sufficiently deeply-nested directory\ntree. The cause appears to be that we rely on standard \"tar\" format\nto represent the symlink for a tablespace, and POSIX tar format has a\nhard-wired restriction of 99 bytes in a symlink's expansion.\n\nWhat do we want to do about this? I think a minimum expectation would be\nfor pg_basebackup to notice and complain when it's trying to create an\nunworkably long symlink entry, but it would be far better if we found a\nway to cope instead.\n\nOne thing we could possibly do without reinventing \"tar\" is to avoid using\nabsolute path names if a PGDATA-relative one would do.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 14:59:31 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"pg_basebackup fails with long tablespace paths"},{"msg_contents":"On 10/20/14 2:59 PM, Tom Lane wrote:\n> What do we want to do about this? I think a minimum expectation would be\n> for pg_basebackup to notice and complain when it's trying to create an\n> unworkably long symlink entry, but it would be far better if we found a\n> way to cope instead.\n\nIsn't it the backend that should error out before sending truncated\nfiles names?\n\nsrc/port/tar.c:\n\n /* Name 100 */\n sprintf(&h[0], \"%.99s\", filename);\n\nAnd then do we need to prevent the creation of tablespaces that can't be\nbacked up?\n\n> One thing we could possibly do without reinventing \"tar\" is to avoid >\nusing\n> absolute path names if a PGDATA-relative one would do.\n\nMaybe we could hack up the tar format to store the symlink target as the\nfile body, like cpio does. Of course then we'd lose the property of\nthis actually being tar.\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 16:51:43 -0400","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On Tue, Oct 21, 2014 at 12:29 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> My Salesforce colleague Thomas Fanghaenel observed that the TAP tests\n> for pg_basebackup fail when run in a sufficiently deeply-nested directory\n> tree. The cause appears to be that we rely on standard \"tar\" format\n> to represent the symlink for a tablespace, and POSIX tar format has a\n> hard-wired restriction of 99 bytes in a symlink's expansion.\n>\n> What do we want to do about this? I think a minimum expectation would be\n> for pg_basebackup to notice and complain when it's trying to create an\n> unworkably long symlink entry, but it would be far better if we found a\n> way to cope instead.\n\nOne way to cope with such a situation could be that during backup we create\na backup symlink file which contains listing of symlinks and then archive\nrecovery recreates it. Basically this is the solution (patch), I have\nproposed\nfor Windows [1].\n\n\n[1] - https://commitfest.postgresql.org/action/patch_view?id=1512\n\nWith Regards,\nAmit Kapila.\nEnterpriseDB: http://www.enterprisedb.com\n\nOn Tue, Oct 21, 2014 at 12:29 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:>> My Salesforce colleague Thomas Fanghaenel observed that the TAP tests> for pg_basebackup fail when run in a sufficiently deeply-nested directory> tree. The cause appears to be that we rely on standard \"tar\" format> to represent the symlink for a tablespace, and POSIX tar format has a> hard-wired restriction of 99 bytes in a symlink's expansion.>> What do we want to do about this? I think a minimum expectation would be> for pg_basebackup to notice and complain when it's trying to create an> unworkably long symlink entry, but it would be far better if we found a> way to cope instead.One way to cope with such a situation could be that during backup we createa backup symlink file which contains listing of symlinks and then archiverecovery recreates it. Basically this is the solution (patch), I have proposedfor Windows [1].[1] - https://commitfest.postgresql.org/action/patch_view?id=1512With Regards,Amit Kapila.EnterpriseDB: http://www.enterprisedb.com","msg_date":"Tue, 21 Oct 2014 09:17:44 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On 10/20/14 2:59 PM, Tom Lane wrote:\n> My Salesforce colleague Thomas Fanghaenel observed that the TAP tests\n> for pg_basebackup fail when run in a sufficiently deeply-nested directory\n> tree.\n\nAs for the test, we can do something like the attached to mark the test\nas \"TODO\".\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Tue, 28 Oct 2014 20:29:39 -0400","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On Tue, Oct 28, 2014 at 8:29 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\n> On 10/20/14 2:59 PM, Tom Lane wrote:\n>> My Salesforce colleague Thomas Fanghaenel observed that the TAP tests\n>> for pg_basebackup fail when run in a sufficiently deeply-nested directory\n>> tree.\n>\n> As for the test, we can do something like the attached to mark the test\n> as \"TODO\".\n\nWhat does this actually do? It doesn't appear that it's just\ndisabling the test.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 29 Oct 2014 10:48:00 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On 10/29/14 10:48 AM, Robert Haas wrote:\n> On Tue, Oct 28, 2014 at 8:29 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\n>> On 10/20/14 2:59 PM, Tom Lane wrote:\n>>> My Salesforce colleague Thomas Fanghaenel observed that the TAP tests\n>>> for pg_basebackup fail when run in a sufficiently deeply-nested directory\n>>> tree.\n>>\n>> As for the test, we can do something like the attached to mark the test\n>> as \"TODO\".\n> \n> What does this actually do? It doesn't appear that it's just\n> disabling the test.\n\nIt still runs the tests, but doesn't count the results in whether the\nsuite passes.\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 30 Oct 2014 23:14:04 -0400","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On 10/20/14 4:51 PM, Peter Eisentraut wrote:\n> On 10/20/14 2:59 PM, Tom Lane wrote:\n>> What do we want to do about this? I think a minimum expectation would be\n>> for pg_basebackup to notice and complain when it's trying to create an\n>> unworkably long symlink entry, but it would be far better if we found a\n>> way to cope instead.\n> \n> Isn't it the backend that should error out before sending truncated\n> files names?\n> \n> src/port/tar.c:\n> \n> /* Name 100 */\n> sprintf(&h[0], \"%.99s\", filename);\n\nHere are patches to address that. First, it reports errors when\nattempting to create a tar header that would truncate file or symlink\nnames. Second, it works around the problem in the tests by creating a\nsymlink from the short-name tempdir that we had set up for the\nUnix-socket directory case.\n\nThe first patch can be backpatched to 9.3. The tar code before that is\ndifferent and would need manual adjustments.\n\nIf someone has a too-long tablespace path, I think they can work around\nthat after this patch by creating a shorter symlink and updating the\npg_tblspc symlinks to point there.\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Tue, 04 Nov 2014 15:52:12 -0500","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On 11/4/14 3:52 PM, Peter Eisentraut wrote:\n> Here are patches to address that. First, it reports errors when\n> attempting to create a tar header that would truncate file or symlink\n> names. Second, it works around the problem in the tests by creating a\n> symlink from the short-name tempdir that we had set up for the\n> Unix-socket directory case.\n\nI ended up splitting this up differently. I applied to part of the\nsecond patch that works around the length issue in tablespaces. So the\ntests now pass in 9.4 and up even in working directories with long\nnames. This clears up the regression in 9.4.\n\nThe remaining, not applied patch is attached. It errors when the file\nname is too long and adds tests for that. This could be applied to 9.5\nand backpatched, if we so choose. It might become obsolete if\nhttps://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.\n If that patch doesn't get accepted, I might add my patch to a future\ncommit fest.\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Fri, 07 Nov 2014 21:03:07 -0500","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"08.11.2014, 04:03, Peter Eisentraut kirjoitti:\n> On 11/4/14 3:52 PM, Peter Eisentraut wrote:\n>> > Here are patches to address that. First, it reports errors when\n>> > attempting to create a tar header that would truncate file or symlink\n>> > names. Second, it works around the problem in the tests by creating a\n>> > symlink from the short-name tempdir that we had set up for the\n>> > Unix-socket directory case.\n> I ended up splitting this up differently. I applied to part of the\n> second patch that works around the length issue in tablespaces. So the\n> tests now pass in 9.4 and up even in working directories with long\n> names. This clears up the regression in 9.4.\n> \n> The remaining, not applied patch is attached. It errors when the file\n> name is too long and adds tests for that. This could be applied to 9.5\n> and backpatched, if we so choose. It might become obsolete if\n> https://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.\n> If that patch doesn't get accepted, I might add my patch to a future\n> commit fest.\n\nI think we should just use the UStar tar format\n(http://en.wikipedia.org/wiki/Tar_%28computing%29#UStar_format) and\nallow long file names; all actively used tar implementations should be\nable to handle them. I'll try to write a patch for that soonish.\n\nUntil UStar format is used we should raise an error if a filename is\nbeing truncated by tar instead of creating invalid archives. Also note\nthat Posix tar format allows 100 byte file names as the name doesn't\nhave to be zero terminated, but we may want to stick to 99 bytes in old\ntype tar anyway as using 100 byte filenames has shown bugs in other tar\nimplementations, for example\nhttps://bugs.debian.org/cgi-bin/bugreport.cgi?bug=689582 - and\ntruncating at 100 bytes instead of 99 doesn't help us too much anyway.\n\n/ Oskari\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 23 Dec 2014 00:40:30 +0200","msg_from":"Oskari Saarenmaa <os@ohmu.fi>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On Tue, Dec 23, 2014 at 4:10 AM, Oskari Saarenmaa <os@ohmu.fi> wrote:\n>\n> 08.11.2014, 04:03, Peter Eisentraut kirjoitti:\n> > On 11/4/14 3:52 PM, Peter Eisentraut wrote:\n> >> > Here are patches to address that. First, it reports errors when\n> >> > attempting to create a tar header that would truncate file or symlink\n> >> > names. Second, it works around the problem in the tests by creating\na\n> >> > symlink from the short-name tempdir that we had set up for the\n> >> > Unix-socket directory case.\n> > I ended up splitting this up differently. I applied to part of the\n> > second patch that works around the length issue in tablespaces. So the\n> > tests now pass in 9.4 and up even in working directories with long\n> > names. This clears up the regression in 9.4.\n> >\n> > The remaining, not applied patch is attached. It errors when the file\n> > name is too long and adds tests for that. This could be applied to 9.5\n> > and backpatched, if we so choose. It might become obsolete if\n> > https://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.\n> > If that patch doesn't get accepted, I might add my patch to a future\n> > commit fest.\n>\n> I think we should just use the UStar tar format\n> (http://en.wikipedia.org/wiki/Tar_%28computing%29#UStar_format) and\n> allow long file names; all actively used tar implementations should be\n> able to handle them. I'll try to write a patch for that soonish.\n>\n\nI think even using UStar format won't make it work for Windows where\nthe standard utilities are not able to understand the symlinks in tar.\nThere is already a patch [1] in this CF which will handle both cases, so I\nam\nnot sure if it is very good idea to go with a new tar format to handle this\nissue.\n\n[1] : https://commitfest.postgresql.org/action/patch_view?id=1512\n\nWith Regards,\nAmit Kapila.\nEnterpriseDB: http://www.enterprisedb.com\n\nOn Tue, Dec 23, 2014 at 4:10 AM, Oskari Saarenmaa <os@ohmu.fi> wrote:>> 08.11.2014, 04:03, Peter Eisentraut kirjoitti:> > On 11/4/14 3:52 PM, Peter Eisentraut wrote:> >> > Here are patches to address that. First, it reports errors when> >> > attempting to create a tar header that would truncate file or symlink> >> > names. Second, it works around the problem in the tests by creating a> >> > symlink from the short-name tempdir that we had set up for the> >> > Unix-socket directory case.> > I ended up splitting this up differently. I applied to part of the> > second patch that works around the length issue in tablespaces. So the> > tests now pass in 9.4 and up even in working directories with long> > names. This clears up the regression in 9.4.> >> > The remaining, not applied patch is attached. It errors when the file> > name is too long and adds tests for that. This could be applied to 9.5> > and backpatched, if we so choose. It might become obsolete if> > https://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.> > If that patch doesn't get accepted, I might add my patch to a future> > commit fest.>> I think we should just use the UStar tar format> (http://en.wikipedia.org/wiki/Tar_%28computing%29#UStar_format) and> allow long file names; all actively used tar implementations should be> able to handle them. I'll try to write a patch for that soonish.>I think even using UStar format won't make it work for Windows wherethe standard utilities are not able to understand the symlinks in tar.There is already a patch [1] in this CF which will handle both cases, so I amnot sure if it is very good idea to go with a new tar format to handle thisissue. [1] : https://commitfest.postgresql.org/action/patch_view?id=1512 With Regards,Amit Kapila.EnterpriseDB: http://www.enterprisedb.com","msg_date":"Tue, 23 Dec 2014 08:30:19 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"23.12.2014, 05:00, Amit Kapila kirjoitti:\n> On Tue, Dec 23, 2014 at 4:10 AM, Oskari Saarenmaa wrote:\n>> 08.11.2014, 04:03, Peter Eisentraut kirjoitti:\n>> > It errors when the file\n>> > name is too long and adds tests for that. This could be applied to 9.5\n>> > and backpatched, if we so choose. It might become obsolete if\n>> > https://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.\n>> > If that patch doesn't get accepted, I might add my patch to a future\n>> > commit fest.\n>>\n>> I think we should just use the UStar tar format\n>> (http://en.wikipedia.org/wiki/Tar_%28computing%29#UStar_format) and\n>> allow long file names; all actively used tar implementations should be\n>> able to handle them. I'll try to write a patch for that soonish.\n>>\n> \n> I think even using UStar format won't make it work for Windows where\n> the standard utilities are not able to understand the symlinks in tar.\n> There is already a patch [1] in this CF which will handle both cases, so\n> I am\n> not sure if it is very good idea to go with a new tar format to handle this\n> issue. \n> \n> [1] : https://commitfest.postgresql.org/action/patch_view?id=1512 \n\nThat patch makes sense for 9.5, but I don't think it's going to be\nbackpatched to previous releases? I think we should also apply Peter's\npatch to master and backbranches to avoid creating invalid tar files\nanywhere. And optionally implement and backpatch long filename support\nin tar even if 9.5 no longer creates tar files with long names.\n\n/ Oskari\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 23 Dec 2014 10:33:44 +0200","msg_from":"Oskari Saarenmaa <os@ohmu.fi>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On 12/22/14 5:40 PM, Oskari Saarenmaa wrote:\n> I think we should just use the UStar tar format\n> (http://en.wikipedia.org/wiki/Tar_%28computing%29#UStar_format) and\n> allow long file names; all actively used tar implementations should be\n> able to handle them. I'll try to write a patch for that soonish.\n\nUStar doesn't handle long link targets, only long file names (and then\nonly up to 255 characters, which doesn't seem satisfactory).\n\nAFAICT, to allow long link targets, the available solutions are either\npax extended headers or GNU-specific long-link extra headers.\n\nWhen I create a symlink with a long target and call tar on it, GNU tar\nby default creates the GNU long-link header and BSD tar by default\ncreates a pax header. But they are both able to extract either one.\n\nAs a demo for how this might look, attached is a wildly incomplete patch\nto produce GNU long-link headers.\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Wed, 24 Dec 2014 08:10:46 -0500","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On 12/22/14 10:00 PM, Amit Kapila wrote:\n> There is already a patch [1] in this CF which will handle both cases, so\n> I am\n> not sure if it is very good idea to go with a new tar format to handle this\n> issue. \n\nI think it would still make sense to have proper symlinks in the\nbasebackup if possible, for clarity.\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 24 Dec 2014 08:12:16 -0500","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On Wed, Dec 24, 2014 at 8:12 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\n> On 12/22/14 10:00 PM, Amit Kapila wrote:\n>> There is already a patch [1] in this CF which will handle both cases, so\n>> I am\n>> not sure if it is very good idea to go with a new tar format to handle this\n>> issue.\n>\n> I think it would still make sense to have proper symlinks in the\n> basebackup if possible, for clarity.\n\nI guess I would have assumed it would be more clear to omit the\nsymlinks if we're expecting the server to put them in. Otherwise, the\nserver has to remove the existing symlinks and create new ones, which\nintroduces various possibilities for failure and confusion.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 27 Dec 2014 20:02:52 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On 12/27/14 8:02 PM, Robert Haas wrote:\n> On Wed, Dec 24, 2014 at 8:12 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\n>> On 12/22/14 10:00 PM, Amit Kapila wrote:\n>>> There is already a patch [1] in this CF which will handle both cases, so\n>>> I am\n>>> not sure if it is very good idea to go with a new tar format to handle this\n>>> issue.\n>>\n>> I think it would still make sense to have proper symlinks in the\n>> basebackup if possible, for clarity.\n> \n> I guess I would have assumed it would be more clear to omit the\n> symlinks if we're expecting the server to put them in. Otherwise, the\n> server has to remove the existing symlinks and create new ones, which\n> introduces various possibilities for failure and confusion.\n\nCurrently, when you unpack a tarred basebackup with tablespaces, the\nsymlinks will tell you whether you have unpacked the tablespace tars at\nthe right place. Otherwise, how do you know? Secondly, you also have\nthe option of putting the tablespaces somewhere else by changing the\nsymlinks. Under the new scheme, the existing symlinks would be\noverwritten (or not?). If that is actually correct, then the proposed\nfix doesn't really replicate the required functionality on Windows.\n\nOne way to address this would be to do away with the symlinks altogether\nand have pg_tblspc/12345 be a text file that contains the tablespace\nlocation. Kind of symlinks implemented in user space.\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 06 Jan 2015 16:33:12 -0500","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On Wed, Jan 7, 2015 at 3:03 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\n>\n> On 12/27/14 8:02 PM, Robert Haas wrote:\n> > On Wed, Dec 24, 2014 at 8:12 AM, Peter Eisentraut <peter_e@gmx.net>\nwrote:\n> >> On 12/22/14 10:00 PM, Amit Kapila wrote:\n> >>> There is already a patch [1] in this CF which will handle both cases,\nso\n> >>> I am\n> >>> not sure if it is very good idea to go with a new tar format to\nhandle this\n> >>> issue.\n> >>\n> >> I think it would still make sense to have proper symlinks in the\n> >> basebackup if possible, for clarity.\n> >\n> > I guess I would have assumed it would be more clear to omit the\n> > symlinks if we're expecting the server to put them in. Otherwise, the\n> > server has to remove the existing symlinks and create new ones, which\n> > introduces various possibilities for failure and confusion.\n>\n> Currently, when you unpack a tarred basebackup with tablespaces, the\n> symlinks will tell you whether you have unpacked the tablespace tars at\n> the right place. Otherwise, how do you know?\n\nvia some kind of tablespace map file which will tell us the exact\nlocation where symlink need to be pointed and the same will be used\nto create a symlink. So after you unpack a tarred basebackup with\ntablespaces, there will be no symlinks; when you start the server\n(archive recovery) using base backup, it will create the appropriate\nsymlinks.\n\n> Secondly, you also have\n> the option of putting the tablespaces somewhere else by changing the\n> symlinks. Under the new scheme, the existing symlinks would be\n> overwritten (or not?). If that is actually correct, then the proposed\n> fix doesn't really replicate the required functionality on Windows.\n>\n> One way to address this would be to do away with the symlinks altogether\n> and have pg_tblspc/12345 be a text file that contains the tablespace\n> location. Kind of symlinks implemented in user space.\n>\n\nI think this is somewhat similar to what existing patch [1] does with\nthe different that there is just one file for all the tablespace locations\nrather than individual file in each tablespace directory.\n\n\n[1] : https://commitfest.postgresql.org/action/patch_view?id=1512\nWith Regards,\nAmit Kapila.\nEnterpriseDB: http://www.enterprisedb.com\n\nOn Wed, Jan 7, 2015 at 3:03 AM, Peter Eisentraut <peter_e@gmx.net> wrote:>> On 12/27/14 8:02 PM, Robert Haas wrote:> > On Wed, Dec 24, 2014 at 8:12 AM, Peter Eisentraut <peter_e@gmx.net> wrote:> >> On 12/22/14 10:00 PM, Amit Kapila wrote:> >>> There is already a patch [1] in this CF which will handle both cases, so> >>> I am> >>> not sure if it is very good idea to go with a new tar format to handle this> >>> issue.> >>> >> I think it would still make sense to have proper symlinks in the> >> basebackup if possible, for clarity.> >> > I guess I would have assumed it would be more clear to omit the> > symlinks if we're expecting the server to put them in. Otherwise, the> > server has to remove the existing symlinks and create new ones, which> > introduces various possibilities for failure and confusion.>> Currently, when you unpack a tarred basebackup with tablespaces, the> symlinks will tell you whether you have unpacked the tablespace tars at> the right place. Otherwise, how do you know?via some kind of tablespace map file which will tell us the exactlocation where symlink need to be pointed and the same will be usedto create a symlink. So after you unpack a tarred basebackup withtablespaces, there will be no symlinks; when you start the server(archive recovery) using base backup, it will create the appropriatesymlinks. > Secondly, you also have> the option of putting the tablespaces somewhere else by changing the> symlinks. Under the new scheme, the existing symlinks would be> overwritten (or not?). If that is actually correct, then the proposed> fix doesn't really replicate the required functionality on Windows.>> One way to address this would be to do away with the symlinks altogether> and have pg_tblspc/12345 be a text file that contains the tablespace> location. Kind of symlinks implemented in user space.>I think this is somewhat similar to what existing patch [1] does withthe different that there is just one file for all the tablespace locationsrather than individual file in each tablespace directory. [1] : https://commitfest.postgresql.org/action/patch_view?id=1512With Regards,Amit Kapila.EnterpriseDB: http://www.enterprisedb.com","msg_date":"Wed, 7 Jan 2015 09:06:08 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On Tue, Jan 6, 2015 at 4:33 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\n> Currently, when you unpack a tarred basebackup with tablespaces, the\n> symlinks will tell you whether you have unpacked the tablespace tars at\n> the right place. Otherwise, how do you know? Secondly, you also have\n> the option of putting the tablespaces somewhere else by changing the\n> symlinks.\n\nThat's a good argument for making the tablespace-map file\nhuman-readable and human-editable, but I don't think it's an argument\nfor duplicating its contents inaccurately in the filesystem.\n\n> One way to address this would be to do away with the symlinks altogether\n> and have pg_tblspc/12345 be a text file that contains the tablespace\n> location. Kind of symlinks implemented in user space.\n\nWell, that's just spreading the tablespace-map file out into several\nfiles, and maybe keeping it around after we've restored from backup.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 7 Jan 2015 15:19:58 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On 1/7/15 3:19 PM, Robert Haas wrote:\n> On Tue, Jan 6, 2015 at 4:33 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\n>> Currently, when you unpack a tarred basebackup with tablespaces, the\n>> symlinks will tell you whether you have unpacked the tablespace tars at\n>> the right place. Otherwise, how do you know? Secondly, you also have\n>> the option of putting the tablespaces somewhere else by changing the\n>> symlinks.\n> \n> That's a good argument for making the tablespace-map file\n> human-readable and human-editable, but I don't think it's an argument\n> for duplicating its contents inaccurately in the filesystem.\n> \n>> One way to address this would be to do away with the symlinks altogether\n>> and have pg_tblspc/12345 be a text file that contains the tablespace\n>> location. Kind of symlinks implemented in user space.\n> \n> Well, that's just spreading the tablespace-map file out into several\n> files, and maybe keeping it around after we've restored from backup.\n\nI think the key point I'm approaching is that the information should\nonly ever be in one place, all the time. This is not dissimilar from\nwhy we took the tablespace location out of the system catalogs. Users\nmight have all kinds of workflows for how they back up, restore, and\nmove their tablespaces. This works pretty well right now, because the\nauthoritative configuration information is always in plain view. The\nproposal is essentially that we add another location for this\ninformation, because the existing location is incompatible with some\noperating system tools. And, when considered by a user, that second\nlocation might or might not collide with or overwrite the first location\nat some mysterious times.\n\nSo I think the preferable fix is not to add a second location, but to\nmake the first location compatible with said operating system tools,\npossibly in the way I propose above.\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 13 Jan 2015 16:41:32 -0500","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On Tue, Jan 13, 2015 at 4:41 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\n> I think the key point I'm approaching is that the information should\n> only ever be in one place, all the time. This is not dissimilar from\n> why we took the tablespace location out of the system catalogs. Users\n> might have all kinds of workflows for how they back up, restore, and\n> move their tablespaces. This works pretty well right now, because the\n> authoritative configuration information is always in plain view. The\n> proposal is essentially that we add another location for this\n> information, because the existing location is incompatible with some\n> operating system tools. And, when considered by a user, that second\n> location might or might not collide with or overwrite the first location\n> at some mysterious times.\n>\n> So I think the preferable fix is not to add a second location, but to\n> make the first location compatible with said operating system tools,\n> possibly in the way I propose above.\n\nI see. I'm a little concerned that following symlinks may be cheaper\nthan whatever system we would come up with for caching the\ntablespace-name-to-file-name mappings. But that concern might be\nunfounded, and apart from it I have no reason to oppose your proposal,\nif you want to do the work.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 14 Jan 2015 14:45:38 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"At 2014-12-24 08:10:46 -0500, peter_e@gmx.net wrote:\n>\n> As a demo for how this might look, attached is a wildly incomplete\n> patch to produce GNU long-link headers.\n\nHi Peter.\n\nIn what way exactly is this patch wildly incomplete? (I ask because it's\nbeen added to the current CF).\n\n-- Abhijit\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 23 Jan 2015 13:56:39 +0530","msg_from":"Abhijit Menon-Sen <ams@2ndQuadrant.com>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On Fri, Nov 7, 2014 at 9:03 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\n> On 11/4/14 3:52 PM, Peter Eisentraut wrote:\n>> Here are patches to address that. First, it reports errors when\n>> attempting to create a tar header that would truncate file or symlink\n>> names. Second, it works around the problem in the tests by creating a\n>> symlink from the short-name tempdir that we had set up for the\n>> Unix-socket directory case.\n>\n> I ended up splitting this up differently. I applied to part of the\n> second patch that works around the length issue in tablespaces. So the\n> tests now pass in 9.4 and up even in working directories with long\n> names. This clears up the regression in 9.4.\n>\n> The remaining, not applied patch is attached. It errors when the file\n> name is too long and adds tests for that. This could be applied to 9.5\n> and backpatched, if we so choose. It might become obsolete if\n> https://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.\n> If that patch doesn't get accepted, I might add my patch to a future\n> commit fest.\n\nI think we should commit this, where by \"this\" I mean your patch to\nerror-check the length of filenames and symlinks instead of truncating\nthem. I don't know what will become of Amit's patch, but I think this\nis a good idea anyway. We should perhaps even consider back-patching\nit, because silently eating people's data is generally not cool. It's\npossible that there are people out there who know that their filenames\nand links are being truncated and don't care, and those people would\nbe unhappy to see this back-patched. However, it's also possible that\nthere are people who don't know that this is happening and do care,\nand those people would be happy about a back-patch. I don't know\nwhich group is larger. At the least, I think we should apply it to\nmaster; because whatever we end up doing about Amit's patch, adding\nerror checks for conditions where we're chewing up somebody's\nfilenames and spitting out what's left over has got to be a good\nthing.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 2 Feb 2015 08:58:03 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On 2/2/15 8:58 AM, Robert Haas wrote:\n> I think we should commit this, where by \"this\" I mean your patch to\n> error-check the length of filenames and symlinks instead of truncating\n> them.\n\ndone\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 24 Feb 2015 13:46:30 -0500","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"},{"msg_contents":"On 1/23/15 3:26 AM, Abhijit Menon-Sen wrote:\n> At 2014-12-24 08:10:46 -0500, peter_e@gmx.net wrote:\n>>\n>> As a demo for how this might look, attached is a wildly incomplete\n>> patch to produce GNU long-link headers.\n> \n> Hi Peter.\n> \n> In what way exactly is this patch wildly incomplete? (I ask because it's\n> been added to the current CF).\n\nThis patch is not in the commit fest. It's just the most recently\nposted patch-like attachment in this thread, which confuses the new CP app.\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 24 Feb 2015 13:56:50 -0500","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: pg_basebackup fails with long tablespace paths"}],"string":"[\n {\n \"msg_contents\": \"My Salesforce colleague Thomas Fanghaenel observed that the TAP tests\\nfor pg_basebackup fail when run in a sufficiently deeply-nested directory\\ntree. The cause appears to be that we rely on standard \\\"tar\\\" format\\nto represent the symlink for a tablespace, and POSIX tar format has a\\nhard-wired restriction of 99 bytes in a symlink's expansion.\\n\\nWhat do we want to do about this? I think a minimum expectation would be\\nfor pg_basebackup to notice and complain when it's trying to create an\\nunworkably long symlink entry, but it would be far better if we found a\\nway to cope instead.\\n\\nOne thing we could possibly do without reinventing \\\"tar\\\" is to avoid using\\nabsolute path names if a PGDATA-relative one would do.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 14:59:31 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On 10/20/14 2:59 PM, Tom Lane wrote:\\n> What do we want to do about this? I think a minimum expectation would be\\n> for pg_basebackup to notice and complain when it's trying to create an\\n> unworkably long symlink entry, but it would be far better if we found a\\n> way to cope instead.\\n\\nIsn't it the backend that should error out before sending truncated\\nfiles names?\\n\\nsrc/port/tar.c:\\n\\n /* Name 100 */\\n sprintf(&h[0], \\\"%.99s\\\", filename);\\n\\nAnd then do we need to prevent the creation of tablespaces that can't be\\nbacked up?\\n\\n> One thing we could possibly do without reinventing \\\"tar\\\" is to avoid >\\nusing\\n> absolute path names if a PGDATA-relative one would do.\\n\\nMaybe we could hack up the tar format to store the symlink target as the\\nfile body, like cpio does. Of course then we'd lose the property of\\nthis actually being tar.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 16:51:43 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 21, 2014 at 12:29 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> My Salesforce colleague Thomas Fanghaenel observed that the TAP tests\\n> for pg_basebackup fail when run in a sufficiently deeply-nested directory\\n> tree. The cause appears to be that we rely on standard \\\"tar\\\" format\\n> to represent the symlink for a tablespace, and POSIX tar format has a\\n> hard-wired restriction of 99 bytes in a symlink's expansion.\\n>\\n> What do we want to do about this? I think a minimum expectation would be\\n> for pg_basebackup to notice and complain when it's trying to create an\\n> unworkably long symlink entry, but it would be far better if we found a\\n> way to cope instead.\\n\\nOne way to cope with such a situation could be that during backup we create\\na backup symlink file which contains listing of symlinks and then archive\\nrecovery recreates it. Basically this is the solution (patch), I have\\nproposed\\nfor Windows [1].\\n\\n\\n[1] - https://commitfest.postgresql.org/action/patch_view?id=1512\\n\\nWith Regards,\\nAmit Kapila.\\nEnterpriseDB: http://www.enterprisedb.com\\n\\nOn Tue, Oct 21, 2014 at 12:29 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:>> My Salesforce colleague Thomas Fanghaenel observed that the TAP tests> for pg_basebackup fail when run in a sufficiently deeply-nested directory> tree. The cause appears to be that we rely on standard \\\"tar\\\" format> to represent the symlink for a tablespace, and POSIX tar format has a> hard-wired restriction of 99 bytes in a symlink's expansion.>> What do we want to do about this? I think a minimum expectation would be> for pg_basebackup to notice and complain when it's trying to create an> unworkably long symlink entry, but it would be far better if we found a> way to cope instead.One way to cope with such a situation could be that during backup we createa backup symlink file which contains listing of symlinks and then archiverecovery recreates it. Basically this is the solution (patch), I have proposedfor Windows [1].[1] - https://commitfest.postgresql.org/action/patch_view?id=1512With Regards,Amit Kapila.EnterpriseDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Tue, 21 Oct 2014 09:17:44 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On 10/20/14 2:59 PM, Tom Lane wrote:\\n> My Salesforce colleague Thomas Fanghaenel observed that the TAP tests\\n> for pg_basebackup fail when run in a sufficiently deeply-nested directory\\n> tree.\\n\\nAs for the test, we can do something like the attached to mark the test\\nas \\\"TODO\\\".\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Tue, 28 Oct 2014 20:29:39 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 28, 2014 at 8:29 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n> On 10/20/14 2:59 PM, Tom Lane wrote:\\n>> My Salesforce colleague Thomas Fanghaenel observed that the TAP tests\\n>> for pg_basebackup fail when run in a sufficiently deeply-nested directory\\n>> tree.\\n>\\n> As for the test, we can do something like the attached to mark the test\\n> as \\\"TODO\\\".\\n\\nWhat does this actually do? It doesn't appear that it's just\\ndisabling the test.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 29 Oct 2014 10:48:00 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On 10/29/14 10:48 AM, Robert Haas wrote:\\n> On Tue, Oct 28, 2014 at 8:29 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n>> On 10/20/14 2:59 PM, Tom Lane wrote:\\n>>> My Salesforce colleague Thomas Fanghaenel observed that the TAP tests\\n>>> for pg_basebackup fail when run in a sufficiently deeply-nested directory\\n>>> tree.\\n>>\\n>> As for the test, we can do something like the attached to mark the test\\n>> as \\\"TODO\\\".\\n> \\n> What does this actually do? It doesn't appear that it's just\\n> disabling the test.\\n\\nIt still runs the tests, but doesn't count the results in whether the\\nsuite passes.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 30 Oct 2014 23:14:04 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On 10/20/14 4:51 PM, Peter Eisentraut wrote:\\n> On 10/20/14 2:59 PM, Tom Lane wrote:\\n>> What do we want to do about this? I think a minimum expectation would be\\n>> for pg_basebackup to notice and complain when it's trying to create an\\n>> unworkably long symlink entry, but it would be far better if we found a\\n>> way to cope instead.\\n> \\n> Isn't it the backend that should error out before sending truncated\\n> files names?\\n> \\n> src/port/tar.c:\\n> \\n> /* Name 100 */\\n> sprintf(&h[0], \\\"%.99s\\\", filename);\\n\\nHere are patches to address that. First, it reports errors when\\nattempting to create a tar header that would truncate file or symlink\\nnames. Second, it works around the problem in the tests by creating a\\nsymlink from the short-name tempdir that we had set up for the\\nUnix-socket directory case.\\n\\nThe first patch can be backpatched to 9.3. The tar code before that is\\ndifferent and would need manual adjustments.\\n\\nIf someone has a too-long tablespace path, I think they can work around\\nthat after this patch by creating a shorter symlink and updating the\\npg_tblspc symlinks to point there.\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Tue, 04 Nov 2014 15:52:12 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On 11/4/14 3:52 PM, Peter Eisentraut wrote:\\n> Here are patches to address that. First, it reports errors when\\n> attempting to create a tar header that would truncate file or symlink\\n> names. Second, it works around the problem in the tests by creating a\\n> symlink from the short-name tempdir that we had set up for the\\n> Unix-socket directory case.\\n\\nI ended up splitting this up differently. I applied to part of the\\nsecond patch that works around the length issue in tablespaces. So the\\ntests now pass in 9.4 and up even in working directories with long\\nnames. This clears up the regression in 9.4.\\n\\nThe remaining, not applied patch is attached. It errors when the file\\nname is too long and adds tests for that. This could be applied to 9.5\\nand backpatched, if we so choose. It might become obsolete if\\nhttps://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.\\n If that patch doesn't get accepted, I might add my patch to a future\\ncommit fest.\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Fri, 07 Nov 2014 21:03:07 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"08.11.2014, 04:03, Peter Eisentraut kirjoitti:\\n> On 11/4/14 3:52 PM, Peter Eisentraut wrote:\\n>> > Here are patches to address that. First, it reports errors when\\n>> > attempting to create a tar header that would truncate file or symlink\\n>> > names. Second, it works around the problem in the tests by creating a\\n>> > symlink from the short-name tempdir that we had set up for the\\n>> > Unix-socket directory case.\\n> I ended up splitting this up differently. I applied to part of the\\n> second patch that works around the length issue in tablespaces. So the\\n> tests now pass in 9.4 and up even in working directories with long\\n> names. This clears up the regression in 9.4.\\n> \\n> The remaining, not applied patch is attached. It errors when the file\\n> name is too long and adds tests for that. This could be applied to 9.5\\n> and backpatched, if we so choose. It might become obsolete if\\n> https://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.\\n> If that patch doesn't get accepted, I might add my patch to a future\\n> commit fest.\\n\\nI think we should just use the UStar tar format\\n(http://en.wikipedia.org/wiki/Tar_%28computing%29#UStar_format) and\\nallow long file names; all actively used tar implementations should be\\nable to handle them. I'll try to write a patch for that soonish.\\n\\nUntil UStar format is used we should raise an error if a filename is\\nbeing truncated by tar instead of creating invalid archives. Also note\\nthat Posix tar format allows 100 byte file names as the name doesn't\\nhave to be zero terminated, but we may want to stick to 99 bytes in old\\ntype tar anyway as using 100 byte filenames has shown bugs in other tar\\nimplementations, for example\\nhttps://bugs.debian.org/cgi-bin/bugreport.cgi?bug=689582 - and\\ntruncating at 100 bytes instead of 99 doesn't help us too much anyway.\\n\\n/ Oskari\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 23 Dec 2014 00:40:30 +0200\",\n \"msg_from\": \"Oskari Saarenmaa <os@ohmu.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 23, 2014 at 4:10 AM, Oskari Saarenmaa <os@ohmu.fi> wrote:\\n>\\n> 08.11.2014, 04:03, Peter Eisentraut kirjoitti:\\n> > On 11/4/14 3:52 PM, Peter Eisentraut wrote:\\n> >> > Here are patches to address that. First, it reports errors when\\n> >> > attempting to create a tar header that would truncate file or symlink\\n> >> > names. Second, it works around the problem in the tests by creating\\na\\n> >> > symlink from the short-name tempdir that we had set up for the\\n> >> > Unix-socket directory case.\\n> > I ended up splitting this up differently. I applied to part of the\\n> > second patch that works around the length issue in tablespaces. So the\\n> > tests now pass in 9.4 and up even in working directories with long\\n> > names. This clears up the regression in 9.4.\\n> >\\n> > The remaining, not applied patch is attached. It errors when the file\\n> > name is too long and adds tests for that. This could be applied to 9.5\\n> > and backpatched, if we so choose. It might become obsolete if\\n> > https://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.\\n> > If that patch doesn't get accepted, I might add my patch to a future\\n> > commit fest.\\n>\\n> I think we should just use the UStar tar format\\n> (http://en.wikipedia.org/wiki/Tar_%28computing%29#UStar_format) and\\n> allow long file names; all actively used tar implementations should be\\n> able to handle them. I'll try to write a patch for that soonish.\\n>\\n\\nI think even using UStar format won't make it work for Windows where\\nthe standard utilities are not able to understand the symlinks in tar.\\nThere is already a patch [1] in this CF which will handle both cases, so I\\nam\\nnot sure if it is very good idea to go with a new tar format to handle this\\nissue.\\n\\n[1] : https://commitfest.postgresql.org/action/patch_view?id=1512\\n\\nWith Regards,\\nAmit Kapila.\\nEnterpriseDB: http://www.enterprisedb.com\\n\\nOn Tue, Dec 23, 2014 at 4:10 AM, Oskari Saarenmaa <os@ohmu.fi> wrote:>> 08.11.2014, 04:03, Peter Eisentraut kirjoitti:> > On 11/4/14 3:52 PM, Peter Eisentraut wrote:> >> > Here are patches to address that. First, it reports errors when> >> > attempting to create a tar header that would truncate file or symlink> >> > names. Second, it works around the problem in the tests by creating a> >> > symlink from the short-name tempdir that we had set up for the> >> > Unix-socket directory case.> > I ended up splitting this up differently. I applied to part of the> > second patch that works around the length issue in tablespaces. So the> > tests now pass in 9.4 and up even in working directories with long> > names. This clears up the regression in 9.4.> >> > The remaining, not applied patch is attached. It errors when the file> > name is too long and adds tests for that. This could be applied to 9.5> > and backpatched, if we so choose. It might become obsolete if> > https://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.> > If that patch doesn't get accepted, I might add my patch to a future> > commit fest.>> I think we should just use the UStar tar format> (http://en.wikipedia.org/wiki/Tar_%28computing%29#UStar_format) and> allow long file names; all actively used tar implementations should be> able to handle them. I'll try to write a patch for that soonish.>I think even using UStar format won't make it work for Windows wherethe standard utilities are not able to understand the symlinks in tar.There is already a patch [1] in this CF which will handle both cases, so I amnot sure if it is very good idea to go with a new tar format to handle thisissue. [1] : https://commitfest.postgresql.org/action/patch_view?id=1512 With Regards,Amit Kapila.EnterpriseDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Tue, 23 Dec 2014 08:30:19 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"23.12.2014, 05:00, Amit Kapila kirjoitti:\\n> On Tue, Dec 23, 2014 at 4:10 AM, Oskari Saarenmaa wrote:\\n>> 08.11.2014, 04:03, Peter Eisentraut kirjoitti:\\n>> > It errors when the file\\n>> > name is too long and adds tests for that. This could be applied to 9.5\\n>> > and backpatched, if we so choose. It might become obsolete if\\n>> > https://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.\\n>> > If that patch doesn't get accepted, I might add my patch to a future\\n>> > commit fest.\\n>>\\n>> I think we should just use the UStar tar format\\n>> (http://en.wikipedia.org/wiki/Tar_%28computing%29#UStar_format) and\\n>> allow long file names; all actively used tar implementations should be\\n>> able to handle them. I'll try to write a patch for that soonish.\\n>>\\n> \\n> I think even using UStar format won't make it work for Windows where\\n> the standard utilities are not able to understand the symlinks in tar.\\n> There is already a patch [1] in this CF which will handle both cases, so\\n> I am\\n> not sure if it is very good idea to go with a new tar format to handle this\\n> issue. \\n> \\n> [1] : https://commitfest.postgresql.org/action/patch_view?id=1512 \\n\\nThat patch makes sense for 9.5, but I don't think it's going to be\\nbackpatched to previous releases? I think we should also apply Peter's\\npatch to master and backbranches to avoid creating invalid tar files\\nanywhere. And optionally implement and backpatch long filename support\\nin tar even if 9.5 no longer creates tar files with long names.\\n\\n/ Oskari\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 23 Dec 2014 10:33:44 +0200\",\n \"msg_from\": \"Oskari Saarenmaa <os@ohmu.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On 12/22/14 5:40 PM, Oskari Saarenmaa wrote:\\n> I think we should just use the UStar tar format\\n> (http://en.wikipedia.org/wiki/Tar_%28computing%29#UStar_format) and\\n> allow long file names; all actively used tar implementations should be\\n> able to handle them. I'll try to write a patch for that soonish.\\n\\nUStar doesn't handle long link targets, only long file names (and then\\nonly up to 255 characters, which doesn't seem satisfactory).\\n\\nAFAICT, to allow long link targets, the available solutions are either\\npax extended headers or GNU-specific long-link extra headers.\\n\\nWhen I create a symlink with a long target and call tar on it, GNU tar\\nby default creates the GNU long-link header and BSD tar by default\\ncreates a pax header. But they are both able to extract either one.\\n\\nAs a demo for how this might look, attached is a wildly incomplete patch\\nto produce GNU long-link headers.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Wed, 24 Dec 2014 08:10:46 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On 12/22/14 10:00 PM, Amit Kapila wrote:\\n> There is already a patch [1] in this CF which will handle both cases, so\\n> I am\\n> not sure if it is very good idea to go with a new tar format to handle this\\n> issue. \\n\\nI think it would still make sense to have proper symlinks in the\\nbasebackup if possible, for clarity.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 24 Dec 2014 08:12:16 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 24, 2014 at 8:12 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n> On 12/22/14 10:00 PM, Amit Kapila wrote:\\n>> There is already a patch [1] in this CF which will handle both cases, so\\n>> I am\\n>> not sure if it is very good idea to go with a new tar format to handle this\\n>> issue.\\n>\\n> I think it would still make sense to have proper symlinks in the\\n> basebackup if possible, for clarity.\\n\\nI guess I would have assumed it would be more clear to omit the\\nsymlinks if we're expecting the server to put them in. Otherwise, the\\nserver has to remove the existing symlinks and create new ones, which\\nintroduces various possibilities for failure and confusion.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 27 Dec 2014 20:02:52 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On 12/27/14 8:02 PM, Robert Haas wrote:\\n> On Wed, Dec 24, 2014 at 8:12 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n>> On 12/22/14 10:00 PM, Amit Kapila wrote:\\n>>> There is already a patch [1] in this CF which will handle both cases, so\\n>>> I am\\n>>> not sure if it is very good idea to go with a new tar format to handle this\\n>>> issue.\\n>>\\n>> I think it would still make sense to have proper symlinks in the\\n>> basebackup if possible, for clarity.\\n> \\n> I guess I would have assumed it would be more clear to omit the\\n> symlinks if we're expecting the server to put them in. Otherwise, the\\n> server has to remove the existing symlinks and create new ones, which\\n> introduces various possibilities for failure and confusion.\\n\\nCurrently, when you unpack a tarred basebackup with tablespaces, the\\nsymlinks will tell you whether you have unpacked the tablespace tars at\\nthe right place. Otherwise, how do you know? Secondly, you also have\\nthe option of putting the tablespaces somewhere else by changing the\\nsymlinks. Under the new scheme, the existing symlinks would be\\noverwritten (or not?). If that is actually correct, then the proposed\\nfix doesn't really replicate the required functionality on Windows.\\n\\nOne way to address this would be to do away with the symlinks altogether\\nand have pg_tblspc/12345 be a text file that contains the tablespace\\nlocation. Kind of symlinks implemented in user space.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 06 Jan 2015 16:33:12 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 7, 2015 at 3:03 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n>\\n> On 12/27/14 8:02 PM, Robert Haas wrote:\\n> > On Wed, Dec 24, 2014 at 8:12 AM, Peter Eisentraut <peter_e@gmx.net>\\nwrote:\\n> >> On 12/22/14 10:00 PM, Amit Kapila wrote:\\n> >>> There is already a patch [1] in this CF which will handle both cases,\\nso\\n> >>> I am\\n> >>> not sure if it is very good idea to go with a new tar format to\\nhandle this\\n> >>> issue.\\n> >>\\n> >> I think it would still make sense to have proper symlinks in the\\n> >> basebackup if possible, for clarity.\\n> >\\n> > I guess I would have assumed it would be more clear to omit the\\n> > symlinks if we're expecting the server to put them in. Otherwise, the\\n> > server has to remove the existing symlinks and create new ones, which\\n> > introduces various possibilities for failure and confusion.\\n>\\n> Currently, when you unpack a tarred basebackup with tablespaces, the\\n> symlinks will tell you whether you have unpacked the tablespace tars at\\n> the right place. Otherwise, how do you know?\\n\\nvia some kind of tablespace map file which will tell us the exact\\nlocation where symlink need to be pointed and the same will be used\\nto create a symlink. So after you unpack a tarred basebackup with\\ntablespaces, there will be no symlinks; when you start the server\\n(archive recovery) using base backup, it will create the appropriate\\nsymlinks.\\n\\n> Secondly, you also have\\n> the option of putting the tablespaces somewhere else by changing the\\n> symlinks. Under the new scheme, the existing symlinks would be\\n> overwritten (or not?). If that is actually correct, then the proposed\\n> fix doesn't really replicate the required functionality on Windows.\\n>\\n> One way to address this would be to do away with the symlinks altogether\\n> and have pg_tblspc/12345 be a text file that contains the tablespace\\n> location. Kind of symlinks implemented in user space.\\n>\\n\\nI think this is somewhat similar to what existing patch [1] does with\\nthe different that there is just one file for all the tablespace locations\\nrather than individual file in each tablespace directory.\\n\\n\\n[1] : https://commitfest.postgresql.org/action/patch_view?id=1512\\nWith Regards,\\nAmit Kapila.\\nEnterpriseDB: http://www.enterprisedb.com\\n\\nOn Wed, Jan 7, 2015 at 3:03 AM, Peter Eisentraut <peter_e@gmx.net> wrote:>> On 12/27/14 8:02 PM, Robert Haas wrote:> > On Wed, Dec 24, 2014 at 8:12 AM, Peter Eisentraut <peter_e@gmx.net> wrote:> >> On 12/22/14 10:00 PM, Amit Kapila wrote:> >>> There is already a patch [1] in this CF which will handle both cases, so> >>> I am> >>> not sure if it is very good idea to go with a new tar format to handle this> >>> issue.> >>> >> I think it would still make sense to have proper symlinks in the> >> basebackup if possible, for clarity.> >> > I guess I would have assumed it would be more clear to omit the> > symlinks if we're expecting the server to put them in. Otherwise, the> > server has to remove the existing symlinks and create new ones, which> > introduces various possibilities for failure and confusion.>> Currently, when you unpack a tarred basebackup with tablespaces, the> symlinks will tell you whether you have unpacked the tablespace tars at> the right place. Otherwise, how do you know?via some kind of tablespace map file which will tell us the exactlocation where symlink need to be pointed and the same will be usedto create a symlink. So after you unpack a tarred basebackup withtablespaces, there will be no symlinks; when you start the server(archive recovery) using base backup, it will create the appropriatesymlinks. > Secondly, you also have> the option of putting the tablespaces somewhere else by changing the> symlinks. Under the new scheme, the existing symlinks would be> overwritten (or not?). If that is actually correct, then the proposed> fix doesn't really replicate the required functionality on Windows.>> One way to address this would be to do away with the symlinks altogether> and have pg_tblspc/12345 be a text file that contains the tablespace> location. Kind of symlinks implemented in user space.>I think this is somewhat similar to what existing patch [1] does withthe different that there is just one file for all the tablespace locationsrather than individual file in each tablespace directory. [1] : https://commitfest.postgresql.org/action/patch_view?id=1512With Regards,Amit Kapila.EnterpriseDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Wed, 7 Jan 2015 09:06:08 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 6, 2015 at 4:33 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n> Currently, when you unpack a tarred basebackup with tablespaces, the\\n> symlinks will tell you whether you have unpacked the tablespace tars at\\n> the right place. Otherwise, how do you know? Secondly, you also have\\n> the option of putting the tablespaces somewhere else by changing the\\n> symlinks.\\n\\nThat's a good argument for making the tablespace-map file\\nhuman-readable and human-editable, but I don't think it's an argument\\nfor duplicating its contents inaccurately in the filesystem.\\n\\n> One way to address this would be to do away with the symlinks altogether\\n> and have pg_tblspc/12345 be a text file that contains the tablespace\\n> location. Kind of symlinks implemented in user space.\\n\\nWell, that's just spreading the tablespace-map file out into several\\nfiles, and maybe keeping it around after we've restored from backup.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 7 Jan 2015 15:19:58 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On 1/7/15 3:19 PM, Robert Haas wrote:\\n> On Tue, Jan 6, 2015 at 4:33 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n>> Currently, when you unpack a tarred basebackup with tablespaces, the\\n>> symlinks will tell you whether you have unpacked the tablespace tars at\\n>> the right place. Otherwise, how do you know? Secondly, you also have\\n>> the option of putting the tablespaces somewhere else by changing the\\n>> symlinks.\\n> \\n> That's a good argument for making the tablespace-map file\\n> human-readable and human-editable, but I don't think it's an argument\\n> for duplicating its contents inaccurately in the filesystem.\\n> \\n>> One way to address this would be to do away with the symlinks altogether\\n>> and have pg_tblspc/12345 be a text file that contains the tablespace\\n>> location. Kind of symlinks implemented in user space.\\n> \\n> Well, that's just spreading the tablespace-map file out into several\\n> files, and maybe keeping it around after we've restored from backup.\\n\\nI think the key point I'm approaching is that the information should\\nonly ever be in one place, all the time. This is not dissimilar from\\nwhy we took the tablespace location out of the system catalogs. Users\\nmight have all kinds of workflows for how they back up, restore, and\\nmove their tablespaces. This works pretty well right now, because the\\nauthoritative configuration information is always in plain view. The\\nproposal is essentially that we add another location for this\\ninformation, because the existing location is incompatible with some\\noperating system tools. And, when considered by a user, that second\\nlocation might or might not collide with or overwrite the first location\\nat some mysterious times.\\n\\nSo I think the preferable fix is not to add a second location, but to\\nmake the first location compatible with said operating system tools,\\npossibly in the way I propose above.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 13 Jan 2015 16:41:32 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 13, 2015 at 4:41 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n> I think the key point I'm approaching is that the information should\\n> only ever be in one place, all the time. This is not dissimilar from\\n> why we took the tablespace location out of the system catalogs. Users\\n> might have all kinds of workflows for how they back up, restore, and\\n> move their tablespaces. This works pretty well right now, because the\\n> authoritative configuration information is always in plain view. The\\n> proposal is essentially that we add another location for this\\n> information, because the existing location is incompatible with some\\n> operating system tools. And, when considered by a user, that second\\n> location might or might not collide with or overwrite the first location\\n> at some mysterious times.\\n>\\n> So I think the preferable fix is not to add a second location, but to\\n> make the first location compatible with said operating system tools,\\n> possibly in the way I propose above.\\n\\nI see. I'm a little concerned that following symlinks may be cheaper\\nthan whatever system we would come up with for caching the\\ntablespace-name-to-file-name mappings. But that concern might be\\nunfounded, and apart from it I have no reason to oppose your proposal,\\nif you want to do the work.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 14 Jan 2015 14:45:38 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"At 2014-12-24 08:10:46 -0500, peter_e@gmx.net wrote:\\n>\\n> As a demo for how this might look, attached is a wildly incomplete\\n> patch to produce GNU long-link headers.\\n\\nHi Peter.\\n\\nIn what way exactly is this patch wildly incomplete? (I ask because it's\\nbeen added to the current CF).\\n\\n-- Abhijit\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 23 Jan 2015 13:56:39 +0530\",\n \"msg_from\": \"Abhijit Menon-Sen <ams@2ndQuadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 7, 2014 at 9:03 PM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n> On 11/4/14 3:52 PM, Peter Eisentraut wrote:\\n>> Here are patches to address that. First, it reports errors when\\n>> attempting to create a tar header that would truncate file or symlink\\n>> names. Second, it works around the problem in the tests by creating a\\n>> symlink from the short-name tempdir that we had set up for the\\n>> Unix-socket directory case.\\n>\\n> I ended up splitting this up differently. I applied to part of the\\n> second patch that works around the length issue in tablespaces. So the\\n> tests now pass in 9.4 and up even in working directories with long\\n> names. This clears up the regression in 9.4.\\n>\\n> The remaining, not applied patch is attached. It errors when the file\\n> name is too long and adds tests for that. This could be applied to 9.5\\n> and backpatched, if we so choose. It might become obsolete if\\n> https://commitfest.postgresql.org/action/patch_view?id=1512 is accepted.\\n> If that patch doesn't get accepted, I might add my patch to a future\\n> commit fest.\\n\\nI think we should commit this, where by \\\"this\\\" I mean your patch to\\nerror-check the length of filenames and symlinks instead of truncating\\nthem. I don't know what will become of Amit's patch, but I think this\\nis a good idea anyway. We should perhaps even consider back-patching\\nit, because silently eating people's data is generally not cool. It's\\npossible that there are people out there who know that their filenames\\nand links are being truncated and don't care, and those people would\\nbe unhappy to see this back-patched. However, it's also possible that\\nthere are people who don't know that this is happening and do care,\\nand those people would be happy about a back-patch. I don't know\\nwhich group is larger. At the least, I think we should apply it to\\nmaster; because whatever we end up doing about Amit's patch, adding\\nerror checks for conditions where we're chewing up somebody's\\nfilenames and spitting out what's left over has got to be a good\\nthing.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 2 Feb 2015 08:58:03 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On 2/2/15 8:58 AM, Robert Haas wrote:\\n> I think we should commit this, where by \\\"this\\\" I mean your patch to\\n> error-check the length of filenames and symlinks instead of truncating\\n> them.\\n\\ndone\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 24 Feb 2015 13:46:30 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n },\n {\n \"msg_contents\": \"On 1/23/15 3:26 AM, Abhijit Menon-Sen wrote:\\n> At 2014-12-24 08:10:46 -0500, peter_e@gmx.net wrote:\\n>>\\n>> As a demo for how this might look, attached is a wildly incomplete\\n>> patch to produce GNU long-link headers.\\n> \\n> Hi Peter.\\n> \\n> In what way exactly is this patch wildly incomplete? (I ask because it's\\n> been added to the current CF).\\n\\nThis patch is not in the commit fest. It's just the most recently\\nposted patch-like attachment in this thread, which confuses the new CP app.\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 24 Feb 2015 13:56:50 -0500\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_basebackup fails with long tablespace paths\"\n }\n]","isConversational":false}}},{"rowIdx":39692,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi.\n\nOne of our \"production issues\" is that the system generates lots of\nwal-files, lots is like 151952 files over the last 24h, which is about\n2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\nbut the system does indeed work fine. We do subsequently gzip the files to\nlimit actual disk-usage, this makes the files roughly 30-50% in size.\n\nThat being said, along comes the backup, scheduled ones a day and tries to\nread off these wal-files, which to the backup looks like \"an awfull lot of\nsmall files\", our backup utillized a single thread to read of those files\nand levels of at reading through 30-40MB/s from a 21 drive Raid50 of\nrotating drives, which is quite bad. That causes a daily incremental run\nto take in the order of 24h. Differential picking up larger deltas and\nfull are even worse.\n\nOne could of-course question a lot of the other things here, but\ncurrently the only limiting factor is actually the backup being\nable to keep up transporting the wal-log away from the system due\nto the small wal size.\n\nA short test like:\nfind . -type f -ctime -1 | tail -n 50 | xargs cat | pipebench > /dev/null\nconfirms the backup speed to be roughly the same as seen by the backup\nsoftware.\nAnother test from the same volume doing:\nfind . -type f -ctime -1 | tail -n 50 | xargs cat > largefile\nAnd then wait for the fs to not cache the file any more and\ncat largefile | pipebench > /dev/null\nconfirms that the disk-subsystem can do way (150-200MB/s) better on larger\nfiles.\n\nThread here around the same topic:\nhttp://postgresql.1045698.n5.nabble.com/How-do-you-change-the-size-of-the-WAL-files-td3425516.html\nBut not a warm welcoming workaround.\n\nSuggestions are welcome. An archive-command/restore command that could\ncombine/split wal-segments might be the easiest workaround, but how about\ncrash-safeness?\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 21:03:59 +0200","msg_from":"jesper@krogh.cc","msg_from_op":true,"msg_subject":"wal-size limited to 16MB - Performance issue for subsequent backup"},{"msg_contents":"jesper@krogh.cc writes:\n> Thread here around the same topic:\n> http://postgresql.1045698.n5.nabble.com/How-do-you-change-the-size-of-the-WAL-files-td3425516.html\n> But not a warm welcoming workaround.\n\nconfigure --with-wal-segsize=something ?\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 15:08:29 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: wal-size limited to 16MB - Performance issue for subsequent\n backup"},{"msg_contents":"On Mon, Oct 20, 2014 at 09:03:59PM +0200, jesper@krogh.cc wrote:\n> Hi.\n> \n> One of our \"production issues\" is that the system generates lots of\n> wal-files, lots is like 151952 files over the last 24h, which is about\n> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\n> but the system does indeed work fine. We do subsequently gzip the files to\n> limit actual disk-usage, this makes the files roughly 30-50% in size.\n> \n> ...\n> \n> Suggestions are welcome. An archive-command/restore command that could\n> combine/split wal-segments might be the easiest workaround, but how about\n> crash-safeness?\n> \n\nHi,\n\nHave you considered using something like tar/star in the archive command\nto pack them into much larger tar archives?\n\nRegards,\nKen\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 14:27:21 -0500","msg_from":"\"ktm@rice.edu\" <ktm@rice.edu>","msg_from_op":false,"msg_subject":"Re: wal-size limited to 16MB - Performance issue for subsequent\n backup"},{"msg_contents":"> jesper@krogh.cc writes:\n>> Thread here around the same topic:\n>> http://postgresql.1045698.n5.nabble.com/How-do-you-change-the-size-of-the-WAL-files-td3425516.html\n>> But not a warm welcoming workaround.\n>\n> configure --with-wal-segsize=something ?\n\nYes, but there are good reasons not to go down that route. One is that:\n\n1) It looks like I'am going to be the only one, beware of the dragons.\n2) It requires apparently a re-initdb, thus dump/restore of 4.5TB of data\n\nThe latter can absolutely be done by scheduling downtime, but the former\nwould compromise any level of good practice around our DB-operations.\n\nJesper\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 21:27:39 +0200","msg_from":"jesper@krogh.cc","msg_from_op":true,"msg_subject":"Re: wal-size limited to 16MB - Performance issue for subsequent\n backup"},{"msg_contents":"Hi,\n\nOn 2014-10-20 21:03:59 +0200, jesper@krogh.cc wrote:\n> One of our \"production issues\" is that the system generates lots of\n> wal-files, lots is like 151952 files over the last 24h, which is about\n> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\n> but the system does indeed work fine. We do subsequently gzip the files to\n> limit actual disk-usage, this makes the files roughly 30-50% in size.\n\nHave you analyzed what the source of that volume is? Which version of\npostgres are you using? What's your checkpoint_timeout/segments\nsettings?\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 21:35:08 +0200","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: wal-size limited to 16MB - Performance issue for subsequent\n backup"},{"msg_contents":"jesper@krogh.cc writes:\n>> configure --with-wal-segsize=something ?\n\n> Yes, but there are good reasons not to go down that route. One is that:\n\n> 1) It looks like I'am going to be the only one, beware of the dragons.\n> 2) It requires apparently a re-initdb, thus dump/restore of 4.5TB of data\n\nI think a clean shutdown and pg_resetxlog would be sufficient. I agree\nyou'd need to do some testing though ...\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 15:39:03 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: wal-size limited to 16MB - Performance issue for subsequent\n backup"},{"msg_contents":"\n> On 2014-10-20 21:03:59 +0200, jesper@krogh.cc wrote:\n>> One of our \"production issues\" is that the system generates lots of\n>> wal-files, lots is like 151952 files over the last 24h, which is about\n>> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\n>> but the system does indeed work fine. We do subsequently gzip the files\n>> to\n>> limit actual disk-usage, this makes the files roughly 30-50% in size.\n>\n> Have you analyzed what the source of that volume is? Which version of\n> postgres are you using? What's your checkpoint_timeout/segments\n> settings?\n\nSuggestions are surely welcome. I do suspect the majority is from 30\nconcurrent processes updating an 506GB GIN index, but it would be nice to\nconfirm that. There is also a message-queue in the DB with a fairly high\nturnaround.\n\nCurrently PG 9.2 moving to 9.3 hopefully before end-of-year,\ncheckpoint_timeout = 30min, checkpoint_segments = 4096.\n\nAccording to logs checkpoints are roughly 15 minutes apart.\n\nJesper\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 21:41:26 +0200","msg_from":"jesper@krogh.cc","msg_from_op":true,"msg_subject":"Re: wal-size limited to 16MB - Performance issue for subsequent\n backup"},{"msg_contents":"On 2014-10-20 21:41:26 +0200, jesper@krogh.cc wrote:\n> \n> > On 2014-10-20 21:03:59 +0200, jesper@krogh.cc wrote:\n> >> One of our \"production issues\" is that the system generates lots of\n> >> wal-files, lots is like 151952 files over the last 24h, which is about\n> >> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\n> >> but the system does indeed work fine. We do subsequently gzip the files\n> >> to limit actual disk-usage, this makes the files roughly 30-50% in size.\n\nI'm a bit doubtful that 16MB vs., say, 64MB files really changes\nanything substantial for you. If it indeed is a problem, it's simple\nenough to join the files temporarily?\n\n> > Have you analyzed what the source of that volume is? Which version of\n> > postgres are you using? What's your checkpoint_timeout/segments\n> > settings?\n> \n> Suggestions are surely welcome.\n\nOnce you're on 9.3 I'd suggest using pg_xlogdump --stats on it. There's\na backport of the facility for 9.3 (looking somewhat different than what\nis now in 9.5) at\nhttp://archives.postgresql.org/message-id/CABRT9RAzGowqLFcEE8aF6VdPoFEy%2BP9gmu7ktGRzw0dgRwVr9Q%40mail.gmail.com\n\nThat'd tell you a fair bit more. It's noticeably harder to backport to <\n9.3.\n\n> I do suspect the majority is from 30 concurrent processes updating an\n> 506GB GIN index, but it would be nice to confirm that. There is also a\n> message-queue in the DB with a fairly high turnaround.\n\nA 506GB GIN index? Uh, interesting :). What's it used for? Trigrams?\n\nI'd suspect that the message queue isn't the primary culprit, but it's\nhard to say for sure.\n\n> Currently PG 9.2 moving to 9.3 hopefully before end-of-year,\n> checkpoint_timeout = 30min, checkpoint_segments = 4096.\n\nGenerally a high checkpoint_timeout can significantly reduce the WAL\nvolume because of fewer full page writes. I've seen cases where spacing\ncheckpoint further apart by a factor of two reduced the overall WAL\nvolume by more than two.\n\n> According to logs checkpoints are roughly 15 minutes apart.\n\nCan you show log_checkpoints output?\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 21:49:40 +0200","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: wal-size limited to 16MB - Performance issue for subsequent\n backup"},{"msg_contents":"> On 2014-10-20 21:41:26 +0200, jesper@krogh.cc wrote:\n>>\n>> > On 2014-10-20 21:03:59 +0200, jesper@krogh.cc wrote:\n>> >> One of our \"production issues\" is that the system generates lots of\n>> >> wal-files, lots is like 151952 files over the last 24h, which is\n>> about\n>> >> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by\n>> itself,\n>> >> but the system does indeed work fine. We do subsequently gzip the\n>> files\n>> >> to limit actual disk-usage, this makes the files roughly 30-50% in\n>> size.\n>\n> I'm a bit doubtful that 16MB vs., say, 64MB files really changes\n> anything substantial for you. If it indeed is a problem, it's simple\n> enough to join the files temporarily?\n\nI am trying to get my head around a good way to do that. 64MB probably\nisn't a silverbullet. But it would definetely benefit the backup in terms\nof single thread access to data on rotating drives.\n\n>> > Have you analyzed what the source of that volume is? Which version of\n>> > postgres are you using? What's your checkpoint_timeout/segments\n>> > settings?\n>>\n>> Suggestions are surely welcome.\n>\n> Once you're on 9.3 I'd suggest using pg_xlogdump --stats on it. There's\n> a backport of the facility for 9.3 (looking somewhat different than what\n> is now in 9.5) at\n> http://archives.postgresql.org/message-id/CABRT9RAzGowqLFcEE8aF6VdPoFEy%2BP9gmu7ktGRzw0dgRwVr9Q%40mail.gmail.com\n>\n> That'd tell you a fair bit more. It's noticeably harder to backport to <\n> 9.3.\n\nI'll bookmark that one.\n\n>> I do suspect the majority is from 30 concurrent processes updating an\n>> 506GB GIN index, but it would be nice to confirm that. There is also a\n>> message-queue in the DB with a fairly high turnaround.\n>\n> A 506GB GIN index? Uh, interesting :). What's it used for? Trigrams?\n\nIt is for full-text-search, but it is being updated entirely regulary,\n~100M records. A dump/restore cycle typically reduces the size to 30-40%\nof current size.\n\n\n>> Currently PG 9.2 moving to 9.3 hopefully before end-of-year,\n>> checkpoint_timeout = 30min, checkpoint_segments = 4096.\n>\n> Generally a high checkpoint_timeout can significantly reduce the WAL\n> volume because of fewer full page writes. I've seen cases where spacing\n> checkpoint further apart by a factor of two reduced the overall WAL\n> volume by more than two.\n\nI'll work with that, I was just uncomfortable bumping checkpoint_segments\nup much higher, any field experience in that corner?\n\n>> According to logs checkpoints are roughly 15 minutes apart.\n>\n> Can you show log_checkpoints output?\n\n2014-10-20 18:10:22 CEST LOG: checkpoint starting: time\n2014-10-20 18:15:44 CEST LOG: checkpoint complete: wrote 76851 buffers\n(7.3%); 0 transaction log file(s) added, 0 removed, 3238 recycled;\nwrite=295.834 s, sync=23.903 s, total=322.011 s; sync files=2115,\nlongest=0.278 s, average=0.011 s\n2014-10-20 18:40:22 CEST LOG: checkpoint starting: time\n2014-10-20 18:44:30 CEST LOG: checkpoint complete: wrote 60550 buffers\n(5.8%); 0 transaction log file(s) added, 0 removed, 3460 recycled;\nwrite=224.678 s, sync=21.795 s, total=248.340 s; sync files=2090,\nlongest=0.963 s, average=0.010 s\n2014-10-20 19:10:22 CEST LOG: checkpoint starting: time\n2014-10-20 19:14:11 CEST LOG: checkpoint complete: wrote 42720 buffers\n(4.1%); 0 transaction log file(s) added, 0 removed, 3598 recycled;\nwrite=206.259 s, sync=21.185 s, total=229.254 s; sync files=2065,\nlongest=0.945 s, average=0.010 s\n2014-10-20 19:40:22 CEST LOG: checkpoint starting: time\n2014-10-20 19:43:31 CEST LOG: checkpoint complete: wrote 32897 buffers\n(3.1%); 0 transaction log file(s) added, 0 removed, 3626 recycled;\nwrite=161.801 s, sync=26.936 s, total=189.635 s; sync files=2115,\nlongest=0.458 s, average=0.012 s\n2014-10-20 20:10:22 CEST LOG: checkpoint starting: time\n2014-10-20 20:14:04 CEST LOG: checkpoint complete: wrote 37557 buffers\n(3.6%); 0 transaction log file(s) added, 0 removed, 3285 recycled;\nwrite=205.011 s, sync=16.550 s, total=222.442 s; sync files=2113,\nlongest=0.935 s, average=0.007 s\n2014-10-20 20:40:22 CEST LOG: checkpoint starting: time\n2014-10-20 20:45:18 CEST LOG: checkpoint complete: wrote 58012 buffers\n(5.5%); 0 transaction log file(s) added, 0 removed, 3678 recycled;\nwrite=252.750 s, sync=39.178 s, total=295.107 s; sync files=2075,\nlongest=0.990 s, average=0.018 s\n2014-10-20 21:10:22 CEST LOG: checkpoint starting: time\n2014-10-20 21:13:31 CEST LOG: checkpoint complete: wrote 40530 buffers\n(3.9%); 0 transaction log file(s) added, 0 removed, 3652 recycled;\nwrite=167.925 s, sync=19.719 s, total=189.057 s; sync files=2077,\nlongest=0.470 s, average=0.009 s\n2014-10-20 21:40:22 CEST LOG: checkpoint starting: time\n2014-10-20 21:44:20 CEST LOG: checkpoint complete: wrote 45158 buffers\n(4.3%); 0 transaction log file(s) added, 0 removed, 3449 recycled;\nwrite=202.986 s, sync=32.564 s, total=237.441 s; sync files=2100,\nlongest=0.445 s, average=0.015 s\n\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 22:02:05 +0200","msg_from":"jesper@krogh.cc","msg_from_op":true,"msg_subject":"Re: wal-size limited to 16MB - Performance issue for subsequent\n backup"},{"msg_contents":"On Mon, Oct 20, 2014 at 12:03 PM, <jesper@krogh.cc> wrote:\n\n> Hi.\n>\n> One of our \"production issues\" is that the system generates lots of\n> wal-files, lots is like 151952 files over the last 24h, which is about\n> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\n> but the system does indeed work fine. We do subsequently gzip the files to\n> limit actual disk-usage, this makes the files roughly 30-50% in size.\n>\n> That being said, along comes the backup, scheduled ones a day and tries to\n> read off these wal-files, which to the backup looks like \"an awfull lot of\n> small files\", our backup utillized a single thread to read of those files\n> and levels of at reading through 30-40MB/s from a 21 drive Raid50 of\n> rotating drives, which is quite bad. That causes a daily incremental run\n> to take in the order of 24h. Differential picking up larger deltas and\n> full are even worse.\n>\n\nWhy not have archive_command (which gets the files while they are still\ncached) put the files directly into their final destination on the backup\nserver?\n\n\n> Suggestions are welcome. An archive-command/restore command that could\n> combine/split wal-segments might be the easiest workaround, but how about\n> crash-safeness?\n>\n\nI think you would just have to combine them by looking at the file name and\nseeking to a specific spot in the large file (rather than just appending to\nit) so that if the archive_command fails and gets rerun, it will still end\nup in the correct place. I don't see what other crash-safeness issues you\nwould have, other than the ones you already have. You would want to do the\ncompression afterward combining, not before, so that all segments are of\npredictable size.\n\nIt should be pretty easy as long as want your combined files to consist of\neither 16 or 256 (or 255 in older versions) WAL files.\n\nYou would have to pass through directly any files not matching the filename\npattern of ordinary WAL files.\n\nCheers,\n\nJeff\n\nOn Mon, Oct 20, 2014 at 12:03 PM, <jesper@krogh.cc> wrote:Hi.\n\nOne of our \"production issues\" is that the system generates lots of\nwal-files, lots is like 151952 files over the last 24h, which is about\n2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\nbut the system does indeed work fine. We do subsequently gzip the files to\nlimit actual disk-usage, this makes the files roughly 30-50% in size.\n\nThat being said, along comes the backup, scheduled ones a day and tries to\nread off these wal-files, which to the backup looks like \"an awfull lot of\nsmall files\", our backup utillized a single thread to read of those files\nand levels of at reading through 30-40MB/s from a 21 drive Raid50 of\nrotating drives, which is quite bad. That causes a daily incremental run\nto take in the order of 24h. Differential picking up larger deltas and\nfull are even worse.Why not have archive_command (which gets the files while they are still cached) put the files directly into their final destination on the backup server?\nSuggestions are welcome. An archive-command/restore command that could\ncombine/split wal-segments might be the easiest workaround, but how about\ncrash-safeness?I think you would just have to combine them by looking at the file name and seeking to a specific spot in the large file (rather than just appending to it) so that if the archive_command fails and gets rerun, it will still end up in the correct place. I don't see what other crash-safeness issues you would have, other than the ones you already have. You would want to do the compression afterward combining, not before, so that all segments are of predictable size.It should be pretty easy as long as want your combined files to consist of either 16 or 256 (or 255 in older versions) WAL files.You would have to pass through directly any files not matching the filename pattern of ordinary WAL files.Cheers,Jeff","msg_date":"Tue, 21 Oct 2014 14:12:11 -0700","msg_from":"Jeff Janes <jeff.janes@gmail.com>","msg_from_op":false,"msg_subject":"Re: wal-size limited to 16MB - Performance issue for subsequent\n backup"},{"msg_contents":"On 10/21/2014 03:03 AM, jesper@krogh.cc wrote:\n\n> That being said, along comes the backup, scheduled ones a day and tries to\n> read off these wal-files, which to the backup looks like \"an awfull lot of\n> small files\", our backup utillized a single thread to read of those files\n> and levels of at reading through 30-40MB/s from a 21 drive Raid50 of\n> rotating drives, which is quite bad. That causes a daily incremental run\n> to take in the order of 24h. Differential picking up larger deltas and\n> full are even worse.\n\nWhat's the backup system?\n\n151952 files should be a trivial matter for any backup system. I'm very\nsurprised you're seeing those kind of run times for 2TB of WAL, and\nthink it's worth investigating just why the backup system is behaving\nthis way.\n\nWhat does 'filefrag' say about the WAL segments? Are they generally a\nsingle extent each? If not, how many extents?\n\nIt'd be useful to know the kernel version, file system, RAID controller,\nwhether you use LVM, and other relevant details? What's your RAID\narray's stripe size?\n\n> A short test like:\n> find . -type f -ctime -1 | tail -n 50 | xargs cat | pipebench > /dev/null\n> confirms the backup speed to be roughly the same as seen by the backup\n> software.\n> Another test from the same volume doing:\n> find . -type f -ctime -1 | tail -n 50 | xargs cat > largefile\n> And then wait for the fs to not cache the file any more and\n> cat largefile | pipebench > /dev/null\n> confirms that the disk-subsystem can do way (150-200MB/s) better on larger\n> files.\n\nOK, so a larger contiguously allocated file looks like it's probably\nread faster. That doesn't mean there's any guarantee that big WAL\nsegment would be allocated contiguously if there are lots of other\nwrites interspersed, but the FS will try.\n\n(What does 'filefrag' say about your 'largefile'?)\n\nI'm wondering if you're having issues related to a RAID stripe size that\nis close to, or bigger than, your WAL segment size. So each segment is\nonly being read from one disk or a couple of disks. If that's the case\nyou're probably not getting ideal write performance either.\n\nThat said, I don't see any particular reason why readahead wouldn't\nresult in you getting similar results from multiple smaller WAL segments\nthat're allocated contiguously, and they usually would be if they're\ncreated one after the other. What are your readahead settings? (There\nare often several at different levels; what exists depends on how your\nstorage is configured, use of LVM, use of SW RAID, etc).\n\nIn my opinion RAID 50, or RAID 5, are generally pretty poor options for\na database file system in performance terms anyway. Especially for\ntransaction logs. RAID 50 is also not wonderfully durable for arrays of\nlarger numbers of bigger disks given modern disks' sizes, even with the\nlow block error rates and relatively low disk failure rates. I\npersonally tend to consider two parity disks the minimum acceptable for\narrays of more than four or five disks. I'd certainly want continuous\narchiving or streaming replication in place if I was running RAID 50 on\na big array.\n\n-- \n Craig Ringer http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 22 Oct 2014 08:43:20 +0800","msg_from":"Craig Ringer <craig@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: wal-size limited to 16MB - Performance issue for subsequent\n backup"},{"msg_contents":"On 10/20/2014 11:02 PM, jesper@krogh.cc wrote:\n>>> >>I do suspect the majority is from 30 concurrent processes updating an\n>>> >>506GB GIN index, but it would be nice to confirm that. There is also a\n>>> >>message-queue in the DB with a fairly high turnaround.\n>> >\n>> >A 506GB GIN index? Uh, interesting :). What's it used for? Trigrams?\n> It is for full-text-search, but it is being updated entirely regulary,\n> ~100M records. A dump/restore cycle typically reduces the size to 30-40%\n> of current size.\n\nTry 9.4 beta. The on-disk format of GIN indexes was rewritten in 9.4, \nmaking them a lot smaller. That might help with WAL volume too. Or not, \nbut I'd love to hear what the impact is, in a real-life database :-).\n\n- Heikki\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 22 Oct 2014 09:56:02 +0300","msg_from":"Heikki Linnakangas <hlinnakangas@vmware.com>","msg_from_op":false,"msg_subject":"Re: wal-size limited to 16MB - Performance issue for subsequent\n backup"}],"string":"[\n {\n \"msg_contents\": \"Hi.\\n\\nOne of our \\\"production issues\\\" is that the system generates lots of\\nwal-files, lots is like 151952 files over the last 24h, which is about\\n2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\\nbut the system does indeed work fine. We do subsequently gzip the files to\\nlimit actual disk-usage, this makes the files roughly 30-50% in size.\\n\\nThat being said, along comes the backup, scheduled ones a day and tries to\\nread off these wal-files, which to the backup looks like \\\"an awfull lot of\\nsmall files\\\", our backup utillized a single thread to read of those files\\nand levels of at reading through 30-40MB/s from a 21 drive Raid50 of\\nrotating drives, which is quite bad. That causes a daily incremental run\\nto take in the order of 24h. Differential picking up larger deltas and\\nfull are even worse.\\n\\nOne could of-course question a lot of the other things here, but\\ncurrently the only limiting factor is actually the backup being\\nable to keep up transporting the wal-log away from the system due\\nto the small wal size.\\n\\nA short test like:\\nfind . -type f -ctime -1 | tail -n 50 | xargs cat | pipebench > /dev/null\\nconfirms the backup speed to be roughly the same as seen by the backup\\nsoftware.\\nAnother test from the same volume doing:\\nfind . -type f -ctime -1 | tail -n 50 | xargs cat > largefile\\nAnd then wait for the fs to not cache the file any more and\\ncat largefile | pipebench > /dev/null\\nconfirms that the disk-subsystem can do way (150-200MB/s) better on larger\\nfiles.\\n\\nThread here around the same topic:\\nhttp://postgresql.1045698.n5.nabble.com/How-do-you-change-the-size-of-the-WAL-files-td3425516.html\\nBut not a warm welcoming workaround.\\n\\nSuggestions are welcome. An archive-command/restore command that could\\ncombine/split wal-segments might be the easiest workaround, but how about\\ncrash-safeness?\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 21:03:59 +0200\",\n \"msg_from\": \"jesper@krogh.cc\",\n \"msg_from_op\": true,\n \"msg_subject\": \"wal-size limited to 16MB - Performance issue for subsequent backup\"\n },\n {\n \"msg_contents\": \"jesper@krogh.cc writes:\\n> Thread here around the same topic:\\n> http://postgresql.1045698.n5.nabble.com/How-do-you-change-the-size-of-the-WAL-files-td3425516.html\\n> But not a warm welcoming workaround.\\n\\nconfigure --with-wal-segsize=something ?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 15:08:29 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: wal-size limited to 16MB - Performance issue for subsequent\\n backup\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 20, 2014 at 09:03:59PM +0200, jesper@krogh.cc wrote:\\n> Hi.\\n> \\n> One of our \\\"production issues\\\" is that the system generates lots of\\n> wal-files, lots is like 151952 files over the last 24h, which is about\\n> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\\n> but the system does indeed work fine. We do subsequently gzip the files to\\n> limit actual disk-usage, this makes the files roughly 30-50% in size.\\n> \\n> ...\\n> \\n> Suggestions are welcome. An archive-command/restore command that could\\n> combine/split wal-segments might be the easiest workaround, but how about\\n> crash-safeness?\\n> \\n\\nHi,\\n\\nHave you considered using something like tar/star in the archive command\\nto pack them into much larger tar archives?\\n\\nRegards,\\nKen\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 14:27:21 -0500\",\n \"msg_from\": \"\\\"ktm@rice.edu\\\" <ktm@rice.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: wal-size limited to 16MB - Performance issue for subsequent\\n backup\"\n },\n {\n \"msg_contents\": \"> jesper@krogh.cc writes:\\n>> Thread here around the same topic:\\n>> http://postgresql.1045698.n5.nabble.com/How-do-you-change-the-size-of-the-WAL-files-td3425516.html\\n>> But not a warm welcoming workaround.\\n>\\n> configure --with-wal-segsize=something ?\\n\\nYes, but there are good reasons not to go down that route. One is that:\\n\\n1) It looks like I'am going to be the only one, beware of the dragons.\\n2) It requires apparently a re-initdb, thus dump/restore of 4.5TB of data\\n\\nThe latter can absolutely be done by scheduling downtime, but the former\\nwould compromise any level of good practice around our DB-operations.\\n\\nJesper\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 21:27:39 +0200\",\n \"msg_from\": \"jesper@krogh.cc\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: wal-size limited to 16MB - Performance issue for subsequent\\n backup\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2014-10-20 21:03:59 +0200, jesper@krogh.cc wrote:\\n> One of our \\\"production issues\\\" is that the system generates lots of\\n> wal-files, lots is like 151952 files over the last 24h, which is about\\n> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\\n> but the system does indeed work fine. We do subsequently gzip the files to\\n> limit actual disk-usage, this makes the files roughly 30-50% in size.\\n\\nHave you analyzed what the source of that volume is? Which version of\\npostgres are you using? What's your checkpoint_timeout/segments\\nsettings?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 21:35:08 +0200\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: wal-size limited to 16MB - Performance issue for subsequent\\n backup\"\n },\n {\n \"msg_contents\": \"jesper@krogh.cc writes:\\n>> configure --with-wal-segsize=something ?\\n\\n> Yes, but there are good reasons not to go down that route. One is that:\\n\\n> 1) It looks like I'am going to be the only one, beware of the dragons.\\n> 2) It requires apparently a re-initdb, thus dump/restore of 4.5TB of data\\n\\nI think a clean shutdown and pg_resetxlog would be sufficient. I agree\\nyou'd need to do some testing though ...\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 15:39:03 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: wal-size limited to 16MB - Performance issue for subsequent\\n backup\"\n },\n {\n \"msg_contents\": \"\\n> On 2014-10-20 21:03:59 +0200, jesper@krogh.cc wrote:\\n>> One of our \\\"production issues\\\" is that the system generates lots of\\n>> wal-files, lots is like 151952 files over the last 24h, which is about\\n>> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\\n>> but the system does indeed work fine. We do subsequently gzip the files\\n>> to\\n>> limit actual disk-usage, this makes the files roughly 30-50% in size.\\n>\\n> Have you analyzed what the source of that volume is? Which version of\\n> postgres are you using? What's your checkpoint_timeout/segments\\n> settings?\\n\\nSuggestions are surely welcome. I do suspect the majority is from 30\\nconcurrent processes updating an 506GB GIN index, but it would be nice to\\nconfirm that. There is also a message-queue in the DB with a fairly high\\nturnaround.\\n\\nCurrently PG 9.2 moving to 9.3 hopefully before end-of-year,\\ncheckpoint_timeout = 30min, checkpoint_segments = 4096.\\n\\nAccording to logs checkpoints are roughly 15 minutes apart.\\n\\nJesper\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 21:41:26 +0200\",\n \"msg_from\": \"jesper@krogh.cc\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: wal-size limited to 16MB - Performance issue for subsequent\\n backup\"\n },\n {\n \"msg_contents\": \"On 2014-10-20 21:41:26 +0200, jesper@krogh.cc wrote:\\n> \\n> > On 2014-10-20 21:03:59 +0200, jesper@krogh.cc wrote:\\n> >> One of our \\\"production issues\\\" is that the system generates lots of\\n> >> wal-files, lots is like 151952 files over the last 24h, which is about\\n> >> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\\n> >> but the system does indeed work fine. We do subsequently gzip the files\\n> >> to limit actual disk-usage, this makes the files roughly 30-50% in size.\\n\\nI'm a bit doubtful that 16MB vs., say, 64MB files really changes\\nanything substantial for you. If it indeed is a problem, it's simple\\nenough to join the files temporarily?\\n\\n> > Have you analyzed what the source of that volume is? Which version of\\n> > postgres are you using? What's your checkpoint_timeout/segments\\n> > settings?\\n> \\n> Suggestions are surely welcome.\\n\\nOnce you're on 9.3 I'd suggest using pg_xlogdump --stats on it. There's\\na backport of the facility for 9.3 (looking somewhat different than what\\nis now in 9.5) at\\nhttp://archives.postgresql.org/message-id/CABRT9RAzGowqLFcEE8aF6VdPoFEy%2BP9gmu7ktGRzw0dgRwVr9Q%40mail.gmail.com\\n\\nThat'd tell you a fair bit more. It's noticeably harder to backport to <\\n9.3.\\n\\n> I do suspect the majority is from 30 concurrent processes updating an\\n> 506GB GIN index, but it would be nice to confirm that. There is also a\\n> message-queue in the DB with a fairly high turnaround.\\n\\nA 506GB GIN index? Uh, interesting :). What's it used for? Trigrams?\\n\\nI'd suspect that the message queue isn't the primary culprit, but it's\\nhard to say for sure.\\n\\n> Currently PG 9.2 moving to 9.3 hopefully before end-of-year,\\n> checkpoint_timeout = 30min, checkpoint_segments = 4096.\\n\\nGenerally a high checkpoint_timeout can significantly reduce the WAL\\nvolume because of fewer full page writes. I've seen cases where spacing\\ncheckpoint further apart by a factor of two reduced the overall WAL\\nvolume by more than two.\\n\\n> According to logs checkpoints are roughly 15 minutes apart.\\n\\nCan you show log_checkpoints output?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 21:49:40 +0200\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: wal-size limited to 16MB - Performance issue for subsequent\\n backup\"\n },\n {\n \"msg_contents\": \"> On 2014-10-20 21:41:26 +0200, jesper@krogh.cc wrote:\\n>>\\n>> > On 2014-10-20 21:03:59 +0200, jesper@krogh.cc wrote:\\n>> >> One of our \\\"production issues\\\" is that the system generates lots of\\n>> >> wal-files, lots is like 151952 files over the last 24h, which is\\n>> about\\n>> >> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by\\n>> itself,\\n>> >> but the system does indeed work fine. We do subsequently gzip the\\n>> files\\n>> >> to limit actual disk-usage, this makes the files roughly 30-50% in\\n>> size.\\n>\\n> I'm a bit doubtful that 16MB vs., say, 64MB files really changes\\n> anything substantial for you. If it indeed is a problem, it's simple\\n> enough to join the files temporarily?\\n\\nI am trying to get my head around a good way to do that. 64MB probably\\nisn't a silverbullet. But it would definetely benefit the backup in terms\\nof single thread access to data on rotating drives.\\n\\n>> > Have you analyzed what the source of that volume is? Which version of\\n>> > postgres are you using? What's your checkpoint_timeout/segments\\n>> > settings?\\n>>\\n>> Suggestions are surely welcome.\\n>\\n> Once you're on 9.3 I'd suggest using pg_xlogdump --stats on it. There's\\n> a backport of the facility for 9.3 (looking somewhat different than what\\n> is now in 9.5) at\\n> http://archives.postgresql.org/message-id/CABRT9RAzGowqLFcEE8aF6VdPoFEy%2BP9gmu7ktGRzw0dgRwVr9Q%40mail.gmail.com\\n>\\n> That'd tell you a fair bit more. It's noticeably harder to backport to <\\n> 9.3.\\n\\nI'll bookmark that one.\\n\\n>> I do suspect the majority is from 30 concurrent processes updating an\\n>> 506GB GIN index, but it would be nice to confirm that. There is also a\\n>> message-queue in the DB with a fairly high turnaround.\\n>\\n> A 506GB GIN index? Uh, interesting :). What's it used for? Trigrams?\\n\\nIt is for full-text-search, but it is being updated entirely regulary,\\n~100M records. A dump/restore cycle typically reduces the size to 30-40%\\nof current size.\\n\\n\\n>> Currently PG 9.2 moving to 9.3 hopefully before end-of-year,\\n>> checkpoint_timeout = 30min, checkpoint_segments = 4096.\\n>\\n> Generally a high checkpoint_timeout can significantly reduce the WAL\\n> volume because of fewer full page writes. I've seen cases where spacing\\n> checkpoint further apart by a factor of two reduced the overall WAL\\n> volume by more than two.\\n\\nI'll work with that, I was just uncomfortable bumping checkpoint_segments\\nup much higher, any field experience in that corner?\\n\\n>> According to logs checkpoints are roughly 15 minutes apart.\\n>\\n> Can you show log_checkpoints output?\\n\\n2014-10-20 18:10:22 CEST LOG: checkpoint starting: time\\n2014-10-20 18:15:44 CEST LOG: checkpoint complete: wrote 76851 buffers\\n(7.3%); 0 transaction log file(s) added, 0 removed, 3238 recycled;\\nwrite=295.834 s, sync=23.903 s, total=322.011 s; sync files=2115,\\nlongest=0.278 s, average=0.011 s\\n2014-10-20 18:40:22 CEST LOG: checkpoint starting: time\\n2014-10-20 18:44:30 CEST LOG: checkpoint complete: wrote 60550 buffers\\n(5.8%); 0 transaction log file(s) added, 0 removed, 3460 recycled;\\nwrite=224.678 s, sync=21.795 s, total=248.340 s; sync files=2090,\\nlongest=0.963 s, average=0.010 s\\n2014-10-20 19:10:22 CEST LOG: checkpoint starting: time\\n2014-10-20 19:14:11 CEST LOG: checkpoint complete: wrote 42720 buffers\\n(4.1%); 0 transaction log file(s) added, 0 removed, 3598 recycled;\\nwrite=206.259 s, sync=21.185 s, total=229.254 s; sync files=2065,\\nlongest=0.945 s, average=0.010 s\\n2014-10-20 19:40:22 CEST LOG: checkpoint starting: time\\n2014-10-20 19:43:31 CEST LOG: checkpoint complete: wrote 32897 buffers\\n(3.1%); 0 transaction log file(s) added, 0 removed, 3626 recycled;\\nwrite=161.801 s, sync=26.936 s, total=189.635 s; sync files=2115,\\nlongest=0.458 s, average=0.012 s\\n2014-10-20 20:10:22 CEST LOG: checkpoint starting: time\\n2014-10-20 20:14:04 CEST LOG: checkpoint complete: wrote 37557 buffers\\n(3.6%); 0 transaction log file(s) added, 0 removed, 3285 recycled;\\nwrite=205.011 s, sync=16.550 s, total=222.442 s; sync files=2113,\\nlongest=0.935 s, average=0.007 s\\n2014-10-20 20:40:22 CEST LOG: checkpoint starting: time\\n2014-10-20 20:45:18 CEST LOG: checkpoint complete: wrote 58012 buffers\\n(5.5%); 0 transaction log file(s) added, 0 removed, 3678 recycled;\\nwrite=252.750 s, sync=39.178 s, total=295.107 s; sync files=2075,\\nlongest=0.990 s, average=0.018 s\\n2014-10-20 21:10:22 CEST LOG: checkpoint starting: time\\n2014-10-20 21:13:31 CEST LOG: checkpoint complete: wrote 40530 buffers\\n(3.9%); 0 transaction log file(s) added, 0 removed, 3652 recycled;\\nwrite=167.925 s, sync=19.719 s, total=189.057 s; sync files=2077,\\nlongest=0.470 s, average=0.009 s\\n2014-10-20 21:40:22 CEST LOG: checkpoint starting: time\\n2014-10-20 21:44:20 CEST LOG: checkpoint complete: wrote 45158 buffers\\n(4.3%); 0 transaction log file(s) added, 0 removed, 3449 recycled;\\nwrite=202.986 s, sync=32.564 s, total=237.441 s; sync files=2100,\\nlongest=0.445 s, average=0.015 s\\n\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 22:02:05 +0200\",\n \"msg_from\": \"jesper@krogh.cc\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: wal-size limited to 16MB - Performance issue for subsequent\\n backup\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 20, 2014 at 12:03 PM, <jesper@krogh.cc> wrote:\\n\\n> Hi.\\n>\\n> One of our \\\"production issues\\\" is that the system generates lots of\\n> wal-files, lots is like 151952 files over the last 24h, which is about\\n> 2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\\n> but the system does indeed work fine. We do subsequently gzip the files to\\n> limit actual disk-usage, this makes the files roughly 30-50% in size.\\n>\\n> That being said, along comes the backup, scheduled ones a day and tries to\\n> read off these wal-files, which to the backup looks like \\\"an awfull lot of\\n> small files\\\", our backup utillized a single thread to read of those files\\n> and levels of at reading through 30-40MB/s from a 21 drive Raid50 of\\n> rotating drives, which is quite bad. That causes a daily incremental run\\n> to take in the order of 24h. Differential picking up larger deltas and\\n> full are even worse.\\n>\\n\\nWhy not have archive_command (which gets the files while they are still\\ncached) put the files directly into their final destination on the backup\\nserver?\\n\\n\\n> Suggestions are welcome. An archive-command/restore command that could\\n> combine/split wal-segments might be the easiest workaround, but how about\\n> crash-safeness?\\n>\\n\\nI think you would just have to combine them by looking at the file name and\\nseeking to a specific spot in the large file (rather than just appending to\\nit) so that if the archive_command fails and gets rerun, it will still end\\nup in the correct place. I don't see what other crash-safeness issues you\\nwould have, other than the ones you already have. You would want to do the\\ncompression afterward combining, not before, so that all segments are of\\npredictable size.\\n\\nIt should be pretty easy as long as want your combined files to consist of\\neither 16 or 256 (or 255 in older versions) WAL files.\\n\\nYou would have to pass through directly any files not matching the filename\\npattern of ordinary WAL files.\\n\\nCheers,\\n\\nJeff\\n\\nOn Mon, Oct 20, 2014 at 12:03 PM, <jesper@krogh.cc> wrote:Hi.\\n\\nOne of our \\\"production issues\\\" is that the system generates lots of\\nwal-files, lots is like 151952 files over the last 24h, which is about\\n2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\\nbut the system does indeed work fine. We do subsequently gzip the files to\\nlimit actual disk-usage, this makes the files roughly 30-50% in size.\\n\\nThat being said, along comes the backup, scheduled ones a day and tries to\\nread off these wal-files, which to the backup looks like \\\"an awfull lot of\\nsmall files\\\", our backup utillized a single thread to read of those files\\nand levels of at reading through 30-40MB/s from a 21 drive Raid50 of\\nrotating drives, which is quite bad. That causes a daily incremental run\\nto take in the order of 24h. Differential picking up larger deltas and\\nfull are even worse.Why not have archive_command (which gets the files while they are still cached) put the files directly into their final destination on the backup server?\\nSuggestions are welcome. An archive-command/restore command that could\\ncombine/split wal-segments might be the easiest workaround, but how about\\ncrash-safeness?I think you would just have to combine them by looking at the file name and seeking to a specific spot in the large file (rather than just appending to it) so that if the archive_command fails and gets rerun, it will still end up in the correct place. I don't see what other crash-safeness issues you would have, other than the ones you already have. You would want to do the compression afterward combining, not before, so that all segments are of predictable size.It should be pretty easy as long as want your combined files to consist of either 16 or 256 (or 255 in older versions) WAL files.You would have to pass through directly any files not matching the filename pattern of ordinary WAL files.Cheers,Jeff\",\n \"msg_date\": \"Tue, 21 Oct 2014 14:12:11 -0700\",\n \"msg_from\": \"Jeff Janes <jeff.janes@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: wal-size limited to 16MB - Performance issue for subsequent\\n backup\"\n },\n {\n \"msg_contents\": \"On 10/21/2014 03:03 AM, jesper@krogh.cc wrote:\\n\\n> That being said, along comes the backup, scheduled ones a day and tries to\\n> read off these wal-files, which to the backup looks like \\\"an awfull lot of\\n> small files\\\", our backup utillized a single thread to read of those files\\n> and levels of at reading through 30-40MB/s from a 21 drive Raid50 of\\n> rotating drives, which is quite bad. That causes a daily incremental run\\n> to take in the order of 24h. Differential picking up larger deltas and\\n> full are even worse.\\n\\nWhat's the backup system?\\n\\n151952 files should be a trivial matter for any backup system. I'm very\\nsurprised you're seeing those kind of run times for 2TB of WAL, and\\nthink it's worth investigating just why the backup system is behaving\\nthis way.\\n\\nWhat does 'filefrag' say about the WAL segments? Are they generally a\\nsingle extent each? If not, how many extents?\\n\\nIt'd be useful to know the kernel version, file system, RAID controller,\\nwhether you use LVM, and other relevant details? What's your RAID\\narray's stripe size?\\n\\n> A short test like:\\n> find . -type f -ctime -1 | tail -n 50 | xargs cat | pipebench > /dev/null\\n> confirms the backup speed to be roughly the same as seen by the backup\\n> software.\\n> Another test from the same volume doing:\\n> find . -type f -ctime -1 | tail -n 50 | xargs cat > largefile\\n> And then wait for the fs to not cache the file any more and\\n> cat largefile | pipebench > /dev/null\\n> confirms that the disk-subsystem can do way (150-200MB/s) better on larger\\n> files.\\n\\nOK, so a larger contiguously allocated file looks like it's probably\\nread faster. That doesn't mean there's any guarantee that big WAL\\nsegment would be allocated contiguously if there are lots of other\\nwrites interspersed, but the FS will try.\\n\\n(What does 'filefrag' say about your 'largefile'?)\\n\\nI'm wondering if you're having issues related to a RAID stripe size that\\nis close to, or bigger than, your WAL segment size. So each segment is\\nonly being read from one disk or a couple of disks. If that's the case\\nyou're probably not getting ideal write performance either.\\n\\nThat said, I don't see any particular reason why readahead wouldn't\\nresult in you getting similar results from multiple smaller WAL segments\\nthat're allocated contiguously, and they usually would be if they're\\ncreated one after the other. What are your readahead settings? (There\\nare often several at different levels; what exists depends on how your\\nstorage is configured, use of LVM, use of SW RAID, etc).\\n\\nIn my opinion RAID 50, or RAID 5, are generally pretty poor options for\\na database file system in performance terms anyway. Especially for\\ntransaction logs. RAID 50 is also not wonderfully durable for arrays of\\nlarger numbers of bigger disks given modern disks' sizes, even with the\\nlow block error rates and relatively low disk failure rates. I\\npersonally tend to consider two parity disks the minimum acceptable for\\narrays of more than four or five disks. I'd certainly want continuous\\narchiving or streaming replication in place if I was running RAID 50 on\\na big array.\\n\\n-- \\n Craig Ringer http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 22 Oct 2014 08:43:20 +0800\",\n \"msg_from\": \"Craig Ringer <craig@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: wal-size limited to 16MB - Performance issue for subsequent\\n backup\"\n },\n {\n \"msg_contents\": \"On 10/20/2014 11:02 PM, jesper@krogh.cc wrote:\\n>>> >>I do suspect the majority is from 30 concurrent processes updating an\\n>>> >>506GB GIN index, but it would be nice to confirm that. There is also a\\n>>> >>message-queue in the DB with a fairly high turnaround.\\n>> >\\n>> >A 506GB GIN index? Uh, interesting :). What's it used for? Trigrams?\\n> It is for full-text-search, but it is being updated entirely regulary,\\n> ~100M records. A dump/restore cycle typically reduces the size to 30-40%\\n> of current size.\\n\\nTry 9.4 beta. The on-disk format of GIN indexes was rewritten in 9.4, \\nmaking them a lot smaller. That might help with WAL volume too. Or not, \\nbut I'd love to hear what the impact is, in a real-life database :-).\\n\\n- Heikki\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 22 Oct 2014 09:56:02 +0300\",\n \"msg_from\": \"Heikki Linnakangas <hlinnakangas@vmware.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: wal-size limited to 16MB - Performance issue for subsequent\\n backup\"\n }\n]","isConversational":false}}},{"rowIdx":39693,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hackers,\n\nIn Mac OS X 10.10 “Yosemite,” Apple removed SystemStarter, upon which our OS X start script has relied since 2007. So here is a patch that adds support for its replacement, launchd. It includes 7 day log rotation like the old script did. The install script still prefers the SystemStarter approach for older versions of the OS, for the sake of easier backward compatibility. We could change that if we wanted, since launchd has been part of the OS for around a decade.","msg_date":"Mon, 20 Oct 2014 15:59:01 -0700","msg_from":"\"David E. Wheeler\" <david@justatheory.com>","msg_from_op":true,"msg_subject":"Patch: Add launchd Support"},{"msg_contents":"\"David E. Wheeler\" <david@justatheory.com> writes:\n> In Mac OS X 10.10 “Yosemite,” Apple removed SystemStarter, upon which our OS X start script has relied since 2007. So here is a patch that adds support for its replacement, launchd. It includes 7 day log rotation like the old script did. The install script still prefers the SystemStarter approach for older versions of the OS, for the sake of easier backward compatibility. We could change that if we wanted, since launchd has been part of the OS for around a decade.\n\n(1) I'd vote for just removing the SystemStarter stuff: it complicates\nunderstanding what's happening, to no very good end. We can easily\ncheck that the launchd way works back to whatever we think our oldest\nsupported OS X release is. (10.4.x according to the buildfarm, at least;\nand I think SystemStarter was deprecated even then ...)\n\n(2) AFAICS, this .plist file doesn't do anything about launchd's habit of\nnot waiting for the network to come up. See my comments in today's thread\nin -general:\nhttp://www.postgresql.org/message-id/1239.1413823243@sss.pgh.pa.us\n\n(3) I don't think you want Disabled = true.\n\n(4) I'm suspicious of all the -c arguments in the .plist file. In general\nI'm not a fan of specifying GUCs on the postmaster command line; that\nmakes it impossible to override their values via \"normal\" methods like\npostgresql.conf or ALTER SYSTEM.\n\n(5) According to the launchd.plist man page, there are options for\nredirecting stdout and stderr to someplace useful. It might be worth\nexercising those ...\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 19:36:56 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On 10/20/14, 5:59 PM, David E. Wheeler wrote:\n> In Mac OS X 10.10 �Yosemite,� Apple removed SystemStarter, upon which our OS X start script has relied since 2007. So here is a patch that adds support for its replacement, launchd. It includes 7 day log rotation like the old script did. The install script still prefers the SystemStarter approach for older versions of the OS, for the sake of easier backward compatibility. We could change that if we wanted, since launchd has been part of the OS for around a decade.\n\nYou're enabling POSTGRESQL in /etc/hostconfig before any of the files are copied over... what happens if we puke before the files get copied? Would it be better to enable after the scripts are in place?\n\nBTW, Mavericks has a comment that /etc/hostconfig is going away, but google isn't telling me what's replacing it...\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 18:58:34 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On Oct 20, 2014, at 4:36 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> (1) I'd vote for just removing the SystemStarter stuff: it complicates\n> understanding what's happening, to no very good end. We can easily\n> check that the launchd way works back to whatever we think our oldest\n> supported OS X release is. (10.4.x according to the buildfarm, at least;\n> and I think SystemStarter was deprecated even then ...)\n\nOkay. Might have to use OnDemand instead of KeepAlive on 10.4. The former was deprecated in 10.5, but I’m not sure when the former was added.\n\n> (2) AFAICS, this .plist file doesn't do anything about launchd's habit of\n> not waiting for the network to come up. See my comments in today's thread\n> in -general:\n> http://www.postgresql.org/message-id/1239.1413823243@sss.pgh.pa.us\n\nHa! How funny you posted a call for a patch today. I didn’t see that, just needed to get it working today myself.\n\nAnyway, I knew there was a reason I didn’t bother with this years ago: launchd does not support dependencies. From the launchd.plist(5)\n\n> DEPENDENCIES\n> Unlike many bootstrapping daemons, launchd has no explicit dependency model.\n> Interdependencies are expected to be solved through the use of IPC. It is\n> therefore in the best interest of a job developer who expects dependents to\n> define all of the sockets in the configuration file. This has the added ben-\n> efit of making it possible to start the job based on demand instead of imme-\n> diately. launchd will continue to place as many restrictions on jobs that\n> do not conform to this model as possible.\n\nThis another reason not to use KeepAlive, I guess. OnDemand is supposed to fire up a job only when it’s needed. No idea what that means. We might be able to put something in LaunchEvents that gets it to fire when the network launches, but documentation is hella thin (and may only be supported on Yosemite, where there are a bunch of poorly-documented launchd changes).\n\n> (3) I don't think you want Disabled = true.\n\nIt’s the default. When you run `launchctl load -w` it overrides it to false in its database. I’m fine to have it be less opaque, though.\n\n> (4) I'm suspicious of all the -c arguments in the .plist file. In general\n> I'm not a fan of specifying GUCs on the postmaster command line; that\n> makes it impossible to override their values via \"normal\" methods like\n> postgresql.conf or ALTER SYSTEM.\n\nYeah, I am okay with removing those; they weren’t in the SystemStarter script. Was the only way to replicate the log rotation stuff, but probably best not to do that in the start script, anyway.\n\n> (5) According to the launchd.plist man page, there are options for\n> redirecting stdout and stderr to someplace useful. It might be worth\n> exercising those ...\n\nSuggestions?\n\nBest,\n\nDavid","msg_date":"Mon, 20 Oct 2014 17:03:24 -0700","msg_from":"\"David E. Wheeler\" <david@justatheory.com>","msg_from_op":true,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On Oct 20, 2014, at 4:58 PM, Jim Nasby <Jim.Nasby@BlueTreble.com> wrote:\n\n> You're enabling POSTGRESQL in /etc/hostconfig before any of the files are copied over... what happens if we puke before the files get copied? Would it be better to enable after the scripts are in place?\n\nThat code was there; I just indented it in an if/then block.\n\n> BTW, Mavericks has a comment that /etc/hostconfig is going away, but google isn't telling me what's replacing it...\n\nlaunchd.\n\nBest,\n\nDavid","msg_date":"Mon, 20 Oct 2014 17:03:58 -0700","msg_from":"\"David E. Wheeler\" <david@justatheory.com>","msg_from_op":true,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"\"David E. Wheeler\" <david@justatheory.com> writes:\n> On Oct 20, 2014, at 4:36 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> (1) I'd vote for just removing the SystemStarter stuff: it complicates\n>> understanding what's happening, to no very good end. We can easily\n>> check that the launchd way works back to whatever we think our oldest\n>> supported OS X release is. (10.4.x according to the buildfarm, at least;\n>> and I think SystemStarter was deprecated even then ...)\n\n> Okay. Might have to use OnDemand instead of KeepAlive on 10.4. The former was deprecated in 10.5, but I’m not sure when the former was added.\n\n[ looks ... ] Yeah, there's no mention of KeepAlive in 10.4's\nlaunchd.plist man page. It does have a convenient example\nsaying that OnDemand = false does what we want:\n\n The following XML Property List simply keeps \"exampled\" running continu-\n ously:\n\n <?xml version=\"1.0\" encoding=\"UTF-8\"?>\n <!DOCTYPE plist PUBLIC -//Apple Computer//DTD PLIST 1.0//EN\n http://www.apple.com/DTDs/PropertyList-1.0.dtd >\n <plist version=\"1.0\">\n <dict>\n <key>Label</key>\n <string>com.example.exampled</string>\n <key>ProgramArguments</key>\n <array>\n <string>exampled</string>\n </array>\n <key>OnDemand</key>\n <false/>\n </dict>\n </plist>\n\n>> (5) According to the launchd.plist man page, there are options for\n>> redirecting stdout and stderr to someplace useful. It might be worth\n>> exercising those ...\n\n> Suggestions?\n\nI'd just drop them into files in the data directory; we're still going\nto recommend that people use the logging_collector, so this is just a\nstopgap to collect startup errors.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 20:17:12 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On Oct 20, 2014, at 5:17 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n> [ looks ... ] Yeah, there's no mention of KeepAlive in 10.4's\n> launchd.plist man page. It does have a convenient example\n> saying that OnDemand = false does what we want:\n\nYeah, let’s see if we can cover both.\n\n> I'd just drop them into files in the data directory; we're still going\n> to recommend that people use the logging_collector, so this is just a\n> stopgap to collect startup errors.\n\nHow about this?\n\n<plist version=\"1.0\">\n<dict>\n\t<key>Disabled</key>\n\t<false/>\n\t<key>Label</key>\n\t<string>org.postgresql.postgresql</string>\n\t<key>UserName</key>\n\t<string>postgres</string>\n\t<key>GroupName</key>\n\t<string>postgres</string>\n\t<key>ProgramArguments</key>\n\t<array>\n\t\t<string>/usr/local/pgsql/bin/postgres</string>\n\t\t<string>-D</string>\n\t\t<string>/usr/local/pgsql/data</string>\n\t</array>\n <key>StandardOutPath</key>\n <string>/usr/local/pgsql/data/launchd.log</string>\n <key>StandardErrorPath</key>\n <string>/usr/local/pgsql/data/launchd.log</string>\n\t<key>OnDemand</key><!-- OS X 10.4 -->\n\t<false/>\n\t<key>KeepAlive</key><!-- OS X 10.5+ -->\n\t<true/>\n</dict>\n</plist>\n\nNo fix for the networking issue, of course.\n\nBest,\n\nDavid","msg_date":"Mon, 20 Oct 2014 17:34:02 -0700","msg_from":"\"David E. Wheeler\" <david@justatheory.com>","msg_from_op":true,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"\nOn Oct 20, 2014, at 5:03 PM, David E. Wheeler <david@justatheory.com> wrote:\n> This another reason not to use KeepAlive, I guess. OnDemand is supposed to fire up a job only when it’s needed. No idea what that means.\n\nI think the idea of OnDemand is for launchd items to act a bit like inetd does: launchd creates the listening socket (or mach port or file-change notification) on the port specified in the plist, and only starts the process when someone tries to connect to it. This might not be a terrible thing to support, but it would require more changes throughout postgres (postmaster would have to check in with launchd at start time to get the listen socket; ports and socket paths would no longer be specifiable in postgres’ config, etc) and I’m skeptical that it’d be worth the work without a more concrete motivation.\n\nApple has published their changes to Postgres (since they ship it in recent versions of OSX) here, fwiw, including the launchd plist they use: http://www.opensource.apple.com/source/PostgreSQL/\n\nOne thing I noticed is that Apple also used the label “org.postgres.postgres” for their launchd job. I don’t know if that will collide in some way with a second job with the same label. Launchctl load/unload takes a pathname, not a job label, so I don’t think it’d be a problem unless you actually do want to run both copies of postgres at the same time.\n\nMacPorts also has a launchd job for their postgresql port, which invokes daemondo, which invokes a wrapper script, which invokes postgres. I’m not sure why they did it that way.\n\n> 2) AFAICS, this .plist file doesn't do anything about launchd's habit of not waiting for the network to come up. \n\nHave you experimented with this setting?:\n\n <key>KeepAlive</key>\n <dict><key>NetworkState</key><true/></dict>\n\nThe launchd.plist man page claims that if you set that key in the sub-dictionary:\n> If true, the job will be kept alive as long as the network is up, where up is defined as at least one non-loopback interface being up and having IPv4 or IPv6 addresses assigned to them. If false, the job will be kept alive in the inverse condition.\n\nOn the other hand, it’s not unreasonable to have postgres running on a machine with only a loopback interface, depending on the use.\n\n> We might be able to put something in LaunchEvents that gets it to fire when the network launches, but documentation is hella thin (and may only be supported on Yosemite, where there are a bunch of poorly-documented launchd changes).\n\nIf one were desperate enough... it’s possible to dig through the launchd sources to make up for the gaps in the documentation (also on opensource.apple.com; there used to be a community-ish site for it at macosforge.org as well). It’s rough going, though, IIRC.\n\n>> (3) I don't think you want Disabled = true.\n> \n> It’s the default. When you run `launchctl load -w` it overrides it to false in its database. I’m fine to have it be less opaque, though.\n\nYes, AFAICT it’s conventional to specify Disabled=true in a launchd plist and use launchctl to enable the item.\n\n> BTW, Mavericks has a comment that /etc/hostconfig is going away, but google isn't telling me what's replacing it…\n\nI think that’s been “going away” for a decade now.\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 17:53:11 -0700","msg_from":"Wim Lewis <wiml@omnigroup.com>","msg_from_op":false,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On Tue, Oct 21, 2014 at 3:53 AM, Wim Lewis <wiml@omnigroup.com> wrote:\n> I think the idea of OnDemand is for launchd items to act a bit like inetd\n> does: launchd creates the listening socket (or mach port or file-change\n> notification) on the port specified in the plist, and only starts the\n> process when someone tries to connect to it. This might not be a terrible\n> thing to support, but it would require more changes throughout postgres\n> (postmaster would have to check in with launchd at start time to get the\n> listen socket; ports and socket paths would no longer be specifiable in\n> postgres’ config, etc) and I’m skeptical that it’d be worth the work\n> without a more concrete motivation.\n\nsystemd socket activation on Linux works pretty much the same way. If\nthis will ever be implemented, it's most reasonable to tackle both\nlaunchd and systemd together.\n\nRegards,\nMarti\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 12:30:33 +0300","msg_from":"Marti Raudsepp <marti@juffo.org>","msg_from_op":false,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On Oct21, 2014, at 02:53 , Wim Lewis <wiml@omnigroup.com> wrote:\n> \n>> 2) AFAICS, this .plist file doesn't do anything about launchd's habit of\n>> not waiting for the network to come up. \n\n>> If true, the job will be kept alive as long as the network is up, where\n>> up is defined as at least one non-loopback interface being up and having\n>> IPv4 or IPv6 addresses assigned to them. If false, the job will be kept\n>> alive in the inverse condition.\n> \n> On the other hand, it’s not unreasonable to have postgres running on a\n> machine with only a loopback interface, depending on the use.\n\nThis, I think, shows the gist of the problem -- \"Networking is up\" is not\na clearly defined stated on modern desktop machines. Interfaces can come and\ngo all the time, and even things like ethernet may take a while to come up\nduring boot, or even depend on a user logging in if something like 802.11X\nis used.\n\n>> We might be able to put something in LaunchEvents that gets it to fire\n>> when the network launches, but documentation is hella thin (and may only\n>> be supported on Yosemite, where there are a bunch of poorly-documented\n>> launchd changes).\n> \n> If one were desperate enough... it’s possible to dig through the launchd\n> sources to make up for the gaps in the documentation (also on\n> opensource.apple.com; there used to be a community-ish site for it at\n> macosforge.org as well). It’s rough going, though, IIRC.\n\nThe correct way to deal with this would be to teach postgres to react to\nconnectivity changes dynamically. During postmaster launch, we'd ignore all\n\"listen\" entries which mention non-bindable addresses. If the network\nconfiguration changes, we'd rescan the \"listen\" entries, and attempt to\ncreate any missing sockets.\n\nOS X provides an API (via the SystemConfiguration framework IIRC) which allows\nto subscribe to networking-changed notifications. I believe systemd provides\nsomething similar via dbus or some such.\n\nWhether or not doing this is worth the effort though, I don't know. I guess\nit depends on whether postgres on OS X is actually used for productions --\nmy guess is that most installations are for development and testing purposes,\nbut maybe my view's biased there.\n\n>>> (3) I don't think you want Disabled = true.\n>> \n>> It’s the default. When you run `launchctl load -w` it overrides it to false in\n>> its database. I’m fine to have it be less opaque, though.\n> \n> Yes, AFAICT it’s conventional to specify Disabled=true in a launchd plist and\n> use launchctl to enable the item.\n\nYup, macports also has Disabled=true in the launchd items they install.\n\nbest regards,\nFlorian Pflug\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 15:14:00 +0200","msg_from":"Florian Pflug <fgp@phlo.org>","msg_from_op":false,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On 10/20/14 8:53 PM, Wim Lewis wrote:\n> Apple has published their changes to Postgres (since they ship it in recent versions of OSX) here, fwiw, including the launchd plist they use: http://www.opensource.apple.com/source/PostgreSQL/\n> \n> One thing I noticed is that Apple also used the label �org.postgres.postgres� for their launchd job. I don�t know if that will collide in some way with a second job with the same label. Launchctl load/unload takes a pathname, not a job label, so I don�t think it�d be a problem unless you actually do want to run both copies of postgres at the same time.\n\nI have a file /System/Library/LaunchAgents/com.apple.postgres.plist that\nuses the label com.apple.postgres instead.\n\n> MacPorts also has a launchd job for their postgresql port, which invokes daemondo, which invokes a wrapper script, which invokes postgres. I�m not sure why they did it that way.\n\nProbably because it's apparently really difficult to do it in a\ndifferent way.\n\n>> 2) AFAICS, this .plist file doesn't do anything about launchd's habit of not waiting for the network to come up. \n> \n> Have you experimented with this setting?:\n> \n> <key>KeepAlive</key>\n> <dict><key>NetworkState</key><true/></dict>\n> \n> The launchd.plist man page claims that if you set that key in the sub-dictionary:\n>> If true, the job will be kept alive as long as the network is up, where up is defined as at least one non-loopback interface being up and having IPv4 or IPv6 addresses assigned to them. If false, the job will be kept alive in the inverse condition.\n\nMy launchd.plist man page says:\n\n NetworkState <boolean>\n This key is no longer implemented as it never acted how most users\nexpected.\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 01 Nov 2014 07:16:23 -0400","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On Mon, Oct 20, 2014 at 03:59:01PM -0700, David E. Wheeler wrote:\n> Hackers,\n>\n> In Mac OS X 10.10 “Yosemite,” Apple removed SystemStarter, upon\n> which our OS X start script has relied since 2007. So here is a patch\n> that adds support for its replacement, launchd. It includes 7 day log\n> rotation like the old script did. The install script still prefers\n> the SystemStarter approach for older versions of the OS, for the sake\n> of easier backward compatibility. We could change that if we wanted,\n> since launchd has been part of the OS for around a decade.\n\nWhere are we on this?\n\n-- \n Bruce Momjian <bruce@momjian.us> http://momjian.us\n EnterpriseDB http://enterprisedb.com\n\n + Everyone has their own god. +\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 19 Mar 2015 23:12:56 -0400","msg_from":"Bruce Momjian <bruce@momjian.us>","msg_from_op":false,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On Mar 19, 2015, at 8:12 PM, Bruce Momjian <bruce@momjian.us> wrote:\n\n> Where are we on this?\n\nI suggested this plist:\n\n<dict>\n\t<key>Disabled</key>\n\t<false/>\n\t<key>Label</key>\n\t<string>org.postgresql.postgresql</string>\n\t<key>UserName</key>\n\t<string>postgres</string>\n\t<key>GroupName</key>\n\t<string>postgres</string>\n\t<key>ProgramArguments</key>\n\t<array>\n\t\t<string>/usr/local/pgsql/bin/postgres</string>\n\t\t<string>-D</string>\n\t\t<string>/usr/local/pgsql/data</string>\n\t</array>\n <key>StandardOutPath</key>\n <string>/usr/local/pgsql/data/launchd.log</string>\n <key>StandardErrorPath</key>\n <string>/usr/local/pgsql/data/launchd.log</string>\n\t<key>OnDemand</key><!-- OS X 10.4 -->\n\t<false/>\n\t<key>KeepAlive</key><!-- OS X 10.5+ -->\n\t<true/>\n</dict>\n</plist>\n\nNo one replied. Want a new patch with that?\n\nDavid","msg_date":"Fri, 20 Mar 2015 16:11:14 -0700","msg_from":"\"David E. Wheeler\" <david@justatheory.com>","msg_from_op":true,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On 3/20/15 6:11 PM, David E. Wheeler wrote:\n> \t<key>ProgramArguments</key>\n> \t<array>\n> \t\t<string>/usr/local/pgsql/bin/postgres</string>\n> \t\t<string>-D</string>\n> \t\t<string>/usr/local/pgsql/data</string>\n> \t</array>\n\nHrm, would /var/db/postgres be better? I'm not sure if the stuff Apple \ndoes with /private/ would cause problems though. (In any case, I think \npostgres is better than pgsql.)\n\n> <key>StandardOutPath</key>\n> <string>/usr/local/pgsql/data/launchd.log</string>\n> <key>StandardErrorPath</key>\n> <string>/usr/local/pgsql/data/launchd.log</string>\n\nWouldn't /var/log be better?\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 20 Mar 2015 18:21:33 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":false,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On Mar 20, 2015, at 4:21 PM, Jim Nasby <Jim.Nasby@BlueTreble.com> wrote:\n> \n> On 3/20/15 6:11 PM, David E. Wheeler wrote:\n>> \t<key>ProgramArguments</key>\n>> \t<array>\n>> \t\t<string>/usr/local/pgsql/bin/postgres</string>\n>> \t\t<string>-D</string>\n>> \t\t<string>/usr/local/pgsql/data</string>\n>> \t</array>\n> \n> Hrm, would /var/db/postgres be better? I'm not sure if the stuff Apple does with /private/ would cause problems though. (In any case, I think postgres is better than pgsql.)\n> \n>> <key>StandardOutPath</key>\n>> <string>/usr/local/pgsql/data/launchd.log</string>\n>> <key>StandardErrorPath</key>\n>> <string>/usr/local/pgsql/data/launchd.log</string>\n> \n> Wouldn't /var/log be better?\n\n/usr/local/pgsql has been the standard install location for the PostgreSQL core distribution for as long as I can remember, including on OS X. Our original OS X startup script refers to it. I figure it’s best to keep it consistent.\n\nBest,\n\nDavid","msg_date":"Fri, 20 Mar 2015 16:25:07 -0700","msg_from":"\"David E. Wheeler\" <david@justatheory.com>","msg_from_op":true,"msg_subject":"Re: Patch: Add launchd Support"},{"msg_contents":"On Mar 20, 2015, at 4:11 PM, David E. Wheeler <david@justatheory.com> wrote:\n\n> No one replied. Want a new patch with that?\n\nHere it is.\n\nBest,\n\nDavid","msg_date":"Sun, 22 Mar 2015 11:03:37 -0700","msg_from":"\"David E. Wheeler\" <david@justatheory.com>","msg_from_op":true,"msg_subject":"Re: Patch: Add launchd Support"}],"string":"[\n {\n \"msg_contents\": \"Hackers,\\n\\nIn Mac OS X 10.10 “Yosemite,” Apple removed SystemStarter, upon which our OS X start script has relied since 2007. So here is a patch that adds support for its replacement, launchd. It includes 7 day log rotation like the old script did. The install script still prefers the SystemStarter approach for older versions of the OS, for the sake of easier backward compatibility. We could change that if we wanted, since launchd has been part of the OS for around a decade.\",\n \"msg_date\": \"Mon, 20 Oct 2014 15:59:01 -0700\",\n \"msg_from\": \"\\\"David E. Wheeler\\\" <david@justatheory.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"\\\"David E. Wheeler\\\" <david@justatheory.com> writes:\\n> In Mac OS X 10.10 “Yosemite,” Apple removed SystemStarter, upon which our OS X start script has relied since 2007. So here is a patch that adds support for its replacement, launchd. It includes 7 day log rotation like the old script did. The install script still prefers the SystemStarter approach for older versions of the OS, for the sake of easier backward compatibility. We could change that if we wanted, since launchd has been part of the OS for around a decade.\\n\\n(1) I'd vote for just removing the SystemStarter stuff: it complicates\\nunderstanding what's happening, to no very good end. We can easily\\ncheck that the launchd way works back to whatever we think our oldest\\nsupported OS X release is. (10.4.x according to the buildfarm, at least;\\nand I think SystemStarter was deprecated even then ...)\\n\\n(2) AFAICS, this .plist file doesn't do anything about launchd's habit of\\nnot waiting for the network to come up. See my comments in today's thread\\nin -general:\\nhttp://www.postgresql.org/message-id/1239.1413823243@sss.pgh.pa.us\\n\\n(3) I don't think you want Disabled = true.\\n\\n(4) I'm suspicious of all the -c arguments in the .plist file. In general\\nI'm not a fan of specifying GUCs on the postmaster command line; that\\nmakes it impossible to override their values via \\\"normal\\\" methods like\\npostgresql.conf or ALTER SYSTEM.\\n\\n(5) According to the launchd.plist man page, there are options for\\nredirecting stdout and stderr to someplace useful. It might be worth\\nexercising those ...\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 19:36:56 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On 10/20/14, 5:59 PM, David E. Wheeler wrote:\\n> In Mac OS X 10.10 �Yosemite,� Apple removed SystemStarter, upon which our OS X start script has relied since 2007. So here is a patch that adds support for its replacement, launchd. It includes 7 day log rotation like the old script did. The install script still prefers the SystemStarter approach for older versions of the OS, for the sake of easier backward compatibility. We could change that if we wanted, since launchd has been part of the OS for around a decade.\\n\\nYou're enabling POSTGRESQL in /etc/hostconfig before any of the files are copied over... what happens if we puke before the files get copied? Would it be better to enable after the scripts are in place?\\n\\nBTW, Mavericks has a comment that /etc/hostconfig is going away, but google isn't telling me what's replacing it...\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 18:58:34 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On Oct 20, 2014, at 4:36 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> (1) I'd vote for just removing the SystemStarter stuff: it complicates\\n> understanding what's happening, to no very good end. We can easily\\n> check that the launchd way works back to whatever we think our oldest\\n> supported OS X release is. (10.4.x according to the buildfarm, at least;\\n> and I think SystemStarter was deprecated even then ...)\\n\\nOkay. Might have to use OnDemand instead of KeepAlive on 10.4. The former was deprecated in 10.5, but I’m not sure when the former was added.\\n\\n> (2) AFAICS, this .plist file doesn't do anything about launchd's habit of\\n> not waiting for the network to come up. See my comments in today's thread\\n> in -general:\\n> http://www.postgresql.org/message-id/1239.1413823243@sss.pgh.pa.us\\n\\nHa! How funny you posted a call for a patch today. I didn’t see that, just needed to get it working today myself.\\n\\nAnyway, I knew there was a reason I didn’t bother with this years ago: launchd does not support dependencies. From the launchd.plist(5)\\n\\n> DEPENDENCIES\\n> Unlike many bootstrapping daemons, launchd has no explicit dependency model.\\n> Interdependencies are expected to be solved through the use of IPC. It is\\n> therefore in the best interest of a job developer who expects dependents to\\n> define all of the sockets in the configuration file. This has the added ben-\\n> efit of making it possible to start the job based on demand instead of imme-\\n> diately. launchd will continue to place as many restrictions on jobs that\\n> do not conform to this model as possible.\\n\\nThis another reason not to use KeepAlive, I guess. OnDemand is supposed to fire up a job only when it’s needed. No idea what that means. We might be able to put something in LaunchEvents that gets it to fire when the network launches, but documentation is hella thin (and may only be supported on Yosemite, where there are a bunch of poorly-documented launchd changes).\\n\\n> (3) I don't think you want Disabled = true.\\n\\nIt’s the default. When you run `launchctl load -w` it overrides it to false in its database. I’m fine to have it be less opaque, though.\\n\\n> (4) I'm suspicious of all the -c arguments in the .plist file. In general\\n> I'm not a fan of specifying GUCs on the postmaster command line; that\\n> makes it impossible to override their values via \\\"normal\\\" methods like\\n> postgresql.conf or ALTER SYSTEM.\\n\\nYeah, I am okay with removing those; they weren’t in the SystemStarter script. Was the only way to replicate the log rotation stuff, but probably best not to do that in the start script, anyway.\\n\\n> (5) According to the launchd.plist man page, there are options for\\n> redirecting stdout and stderr to someplace useful. It might be worth\\n> exercising those ...\\n\\nSuggestions?\\n\\nBest,\\n\\nDavid\",\n \"msg_date\": \"Mon, 20 Oct 2014 17:03:24 -0700\",\n \"msg_from\": \"\\\"David E. Wheeler\\\" <david@justatheory.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On Oct 20, 2014, at 4:58 PM, Jim Nasby <Jim.Nasby@BlueTreble.com> wrote:\\n\\n> You're enabling POSTGRESQL in /etc/hostconfig before any of the files are copied over... what happens if we puke before the files get copied? Would it be better to enable after the scripts are in place?\\n\\nThat code was there; I just indented it in an if/then block.\\n\\n> BTW, Mavericks has a comment that /etc/hostconfig is going away, but google isn't telling me what's replacing it...\\n\\nlaunchd.\\n\\nBest,\\n\\nDavid\",\n \"msg_date\": \"Mon, 20 Oct 2014 17:03:58 -0700\",\n \"msg_from\": \"\\\"David E. Wheeler\\\" <david@justatheory.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"\\\"David E. Wheeler\\\" <david@justatheory.com> writes:\\n> On Oct 20, 2014, at 4:36 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> (1) I'd vote for just removing the SystemStarter stuff: it complicates\\n>> understanding what's happening, to no very good end. We can easily\\n>> check that the launchd way works back to whatever we think our oldest\\n>> supported OS X release is. (10.4.x according to the buildfarm, at least;\\n>> and I think SystemStarter was deprecated even then ...)\\n\\n> Okay. Might have to use OnDemand instead of KeepAlive on 10.4. The former was deprecated in 10.5, but I’m not sure when the former was added.\\n\\n[ looks ... ] Yeah, there's no mention of KeepAlive in 10.4's\\nlaunchd.plist man page. It does have a convenient example\\nsaying that OnDemand = false does what we want:\\n\\n The following XML Property List simply keeps \\\"exampled\\\" running continu-\\n ously:\\n\\n <?xml version=\\\"1.0\\\" encoding=\\\"UTF-8\\\"?>\\n <!DOCTYPE plist PUBLIC -//Apple Computer//DTD PLIST 1.0//EN\\n http://www.apple.com/DTDs/PropertyList-1.0.dtd >\\n <plist version=\\\"1.0\\\">\\n <dict>\\n <key>Label</key>\\n <string>com.example.exampled</string>\\n <key>ProgramArguments</key>\\n <array>\\n <string>exampled</string>\\n </array>\\n <key>OnDemand</key>\\n <false/>\\n </dict>\\n </plist>\\n\\n>> (5) According to the launchd.plist man page, there are options for\\n>> redirecting stdout and stderr to someplace useful. It might be worth\\n>> exercising those ...\\n\\n> Suggestions?\\n\\nI'd just drop them into files in the data directory; we're still going\\nto recommend that people use the logging_collector, so this is just a\\nstopgap to collect startup errors.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 20:17:12 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On Oct 20, 2014, at 5:17 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n> [ looks ... ] Yeah, there's no mention of KeepAlive in 10.4's\\n> launchd.plist man page. It does have a convenient example\\n> saying that OnDemand = false does what we want:\\n\\nYeah, let’s see if we can cover both.\\n\\n> I'd just drop them into files in the data directory; we're still going\\n> to recommend that people use the logging_collector, so this is just a\\n> stopgap to collect startup errors.\\n\\nHow about this?\\n\\n<plist version=\\\"1.0\\\">\\n<dict>\\n\\t<key>Disabled</key>\\n\\t<false/>\\n\\t<key>Label</key>\\n\\t<string>org.postgresql.postgresql</string>\\n\\t<key>UserName</key>\\n\\t<string>postgres</string>\\n\\t<key>GroupName</key>\\n\\t<string>postgres</string>\\n\\t<key>ProgramArguments</key>\\n\\t<array>\\n\\t\\t<string>/usr/local/pgsql/bin/postgres</string>\\n\\t\\t<string>-D</string>\\n\\t\\t<string>/usr/local/pgsql/data</string>\\n\\t</array>\\n <key>StandardOutPath</key>\\n <string>/usr/local/pgsql/data/launchd.log</string>\\n <key>StandardErrorPath</key>\\n <string>/usr/local/pgsql/data/launchd.log</string>\\n\\t<key>OnDemand</key><!-- OS X 10.4 -->\\n\\t<false/>\\n\\t<key>KeepAlive</key><!-- OS X 10.5+ -->\\n\\t<true/>\\n</dict>\\n</plist>\\n\\nNo fix for the networking issue, of course.\\n\\nBest,\\n\\nDavid\",\n \"msg_date\": \"Mon, 20 Oct 2014 17:34:02 -0700\",\n \"msg_from\": \"\\\"David E. Wheeler\\\" <david@justatheory.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"\\nOn Oct 20, 2014, at 5:03 PM, David E. Wheeler <david@justatheory.com> wrote:\\n> This another reason not to use KeepAlive, I guess. OnDemand is supposed to fire up a job only when it’s needed. No idea what that means.\\n\\nI think the idea of OnDemand is for launchd items to act a bit like inetd does: launchd creates the listening socket (or mach port or file-change notification) on the port specified in the plist, and only starts the process when someone tries to connect to it. This might not be a terrible thing to support, but it would require more changes throughout postgres (postmaster would have to check in with launchd at start time to get the listen socket; ports and socket paths would no longer be specifiable in postgres’ config, etc) and I’m skeptical that it’d be worth the work without a more concrete motivation.\\n\\nApple has published their changes to Postgres (since they ship it in recent versions of OSX) here, fwiw, including the launchd plist they use: http://www.opensource.apple.com/source/PostgreSQL/\\n\\nOne thing I noticed is that Apple also used the label “org.postgres.postgres” for their launchd job. I don’t know if that will collide in some way with a second job with the same label. Launchctl load/unload takes a pathname, not a job label, so I don’t think it’d be a problem unless you actually do want to run both copies of postgres at the same time.\\n\\nMacPorts also has a launchd job for their postgresql port, which invokes daemondo, which invokes a wrapper script, which invokes postgres. I’m not sure why they did it that way.\\n\\n> 2) AFAICS, this .plist file doesn't do anything about launchd's habit of not waiting for the network to come up. \\n\\nHave you experimented with this setting?:\\n\\n <key>KeepAlive</key>\\n <dict><key>NetworkState</key><true/></dict>\\n\\nThe launchd.plist man page claims that if you set that key in the sub-dictionary:\\n> If true, the job will be kept alive as long as the network is up, where up is defined as at least one non-loopback interface being up and having IPv4 or IPv6 addresses assigned to them. If false, the job will be kept alive in the inverse condition.\\n\\nOn the other hand, it’s not unreasonable to have postgres running on a machine with only a loopback interface, depending on the use.\\n\\n> We might be able to put something in LaunchEvents that gets it to fire when the network launches, but documentation is hella thin (and may only be supported on Yosemite, where there are a bunch of poorly-documented launchd changes).\\n\\nIf one were desperate enough... it’s possible to dig through the launchd sources to make up for the gaps in the documentation (also on opensource.apple.com; there used to be a community-ish site for it at macosforge.org as well). It’s rough going, though, IIRC.\\n\\n>> (3) I don't think you want Disabled = true.\\n> \\n> It’s the default. When you run `launchctl load -w` it overrides it to false in its database. I’m fine to have it be less opaque, though.\\n\\nYes, AFAICT it’s conventional to specify Disabled=true in a launchd plist and use launchctl to enable the item.\\n\\n> BTW, Mavericks has a comment that /etc/hostconfig is going away, but google isn't telling me what's replacing it…\\n\\nI think that’s been “going away” for a decade now.\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 17:53:11 -0700\",\n \"msg_from\": \"Wim Lewis <wiml@omnigroup.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 21, 2014 at 3:53 AM, Wim Lewis <wiml@omnigroup.com> wrote:\\n> I think the idea of OnDemand is for launchd items to act a bit like inetd\\n> does: launchd creates the listening socket (or mach port or file-change\\n> notification) on the port specified in the plist, and only starts the\\n> process when someone tries to connect to it. This might not be a terrible\\n> thing to support, but it would require more changes throughout postgres\\n> (postmaster would have to check in with launchd at start time to get the\\n> listen socket; ports and socket paths would no longer be specifiable in\\n> postgres’ config, etc) and I’m skeptical that it’d be worth the work\\n> without a more concrete motivation.\\n\\nsystemd socket activation on Linux works pretty much the same way. If\\nthis will ever be implemented, it's most reasonable to tackle both\\nlaunchd and systemd together.\\n\\nRegards,\\nMarti\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 12:30:33 +0300\",\n \"msg_from\": \"Marti Raudsepp <marti@juffo.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On Oct21, 2014, at 02:53 , Wim Lewis <wiml@omnigroup.com> wrote:\\n> \\n>> 2) AFAICS, this .plist file doesn't do anything about launchd's habit of\\n>> not waiting for the network to come up. \\n\\n>> If true, the job will be kept alive as long as the network is up, where\\n>> up is defined as at least one non-loopback interface being up and having\\n>> IPv4 or IPv6 addresses assigned to them. If false, the job will be kept\\n>> alive in the inverse condition.\\n> \\n> On the other hand, it’s not unreasonable to have postgres running on a\\n> machine with only a loopback interface, depending on the use.\\n\\nThis, I think, shows the gist of the problem -- \\\"Networking is up\\\" is not\\na clearly defined stated on modern desktop machines. Interfaces can come and\\ngo all the time, and even things like ethernet may take a while to come up\\nduring boot, or even depend on a user logging in if something like 802.11X\\nis used.\\n\\n>> We might be able to put something in LaunchEvents that gets it to fire\\n>> when the network launches, but documentation is hella thin (and may only\\n>> be supported on Yosemite, where there are a bunch of poorly-documented\\n>> launchd changes).\\n> \\n> If one were desperate enough... it’s possible to dig through the launchd\\n> sources to make up for the gaps in the documentation (also on\\n> opensource.apple.com; there used to be a community-ish site for it at\\n> macosforge.org as well). It’s rough going, though, IIRC.\\n\\nThe correct way to deal with this would be to teach postgres to react to\\nconnectivity changes dynamically. During postmaster launch, we'd ignore all\\n\\\"listen\\\" entries which mention non-bindable addresses. If the network\\nconfiguration changes, we'd rescan the \\\"listen\\\" entries, and attempt to\\ncreate any missing sockets.\\n\\nOS X provides an API (via the SystemConfiguration framework IIRC) which allows\\nto subscribe to networking-changed notifications. I believe systemd provides\\nsomething similar via dbus or some such.\\n\\nWhether or not doing this is worth the effort though, I don't know. I guess\\nit depends on whether postgres on OS X is actually used for productions --\\nmy guess is that most installations are for development and testing purposes,\\nbut maybe my view's biased there.\\n\\n>>> (3) I don't think you want Disabled = true.\\n>> \\n>> It’s the default. When you run `launchctl load -w` it overrides it to false in\\n>> its database. I’m fine to have it be less opaque, though.\\n> \\n> Yes, AFAICT it’s conventional to specify Disabled=true in a launchd plist and\\n> use launchctl to enable the item.\\n\\nYup, macports also has Disabled=true in the launchd items they install.\\n\\nbest regards,\\nFlorian Pflug\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 15:14:00 +0200\",\n \"msg_from\": \"Florian Pflug <fgp@phlo.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On 10/20/14 8:53 PM, Wim Lewis wrote:\\n> Apple has published their changes to Postgres (since they ship it in recent versions of OSX) here, fwiw, including the launchd plist they use: http://www.opensource.apple.com/source/PostgreSQL/\\n> \\n> One thing I noticed is that Apple also used the label �org.postgres.postgres� for their launchd job. I don�t know if that will collide in some way with a second job with the same label. Launchctl load/unload takes a pathname, not a job label, so I don�t think it�d be a problem unless you actually do want to run both copies of postgres at the same time.\\n\\nI have a file /System/Library/LaunchAgents/com.apple.postgres.plist that\\nuses the label com.apple.postgres instead.\\n\\n> MacPorts also has a launchd job for their postgresql port, which invokes daemondo, which invokes a wrapper script, which invokes postgres. I�m not sure why they did it that way.\\n\\nProbably because it's apparently really difficult to do it in a\\ndifferent way.\\n\\n>> 2) AFAICS, this .plist file doesn't do anything about launchd's habit of not waiting for the network to come up. \\n> \\n> Have you experimented with this setting?:\\n> \\n> <key>KeepAlive</key>\\n> <dict><key>NetworkState</key><true/></dict>\\n> \\n> The launchd.plist man page claims that if you set that key in the sub-dictionary:\\n>> If true, the job will be kept alive as long as the network is up, where up is defined as at least one non-loopback interface being up and having IPv4 or IPv6 addresses assigned to them. If false, the job will be kept alive in the inverse condition.\\n\\nMy launchd.plist man page says:\\n\\n NetworkState <boolean>\\n This key is no longer implemented as it never acted how most users\\nexpected.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 01 Nov 2014 07:16:23 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 20, 2014 at 03:59:01PM -0700, David E. Wheeler wrote:\\n> Hackers,\\n>\\n> In Mac OS X 10.10 “Yosemite,” Apple removed SystemStarter, upon\\n> which our OS X start script has relied since 2007. So here is a patch\\n> that adds support for its replacement, launchd. It includes 7 day log\\n> rotation like the old script did. The install script still prefers\\n> the SystemStarter approach for older versions of the OS, for the sake\\n> of easier backward compatibility. We could change that if we wanted,\\n> since launchd has been part of the OS for around a decade.\\n\\nWhere are we on this?\\n\\n-- \\n Bruce Momjian <bruce@momjian.us> http://momjian.us\\n EnterpriseDB http://enterprisedb.com\\n\\n + Everyone has their own god. +\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 19 Mar 2015 23:12:56 -0400\",\n \"msg_from\": \"Bruce Momjian <bruce@momjian.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On Mar 19, 2015, at 8:12 PM, Bruce Momjian <bruce@momjian.us> wrote:\\n\\n> Where are we on this?\\n\\nI suggested this plist:\\n\\n<dict>\\n\\t<key>Disabled</key>\\n\\t<false/>\\n\\t<key>Label</key>\\n\\t<string>org.postgresql.postgresql</string>\\n\\t<key>UserName</key>\\n\\t<string>postgres</string>\\n\\t<key>GroupName</key>\\n\\t<string>postgres</string>\\n\\t<key>ProgramArguments</key>\\n\\t<array>\\n\\t\\t<string>/usr/local/pgsql/bin/postgres</string>\\n\\t\\t<string>-D</string>\\n\\t\\t<string>/usr/local/pgsql/data</string>\\n\\t</array>\\n <key>StandardOutPath</key>\\n <string>/usr/local/pgsql/data/launchd.log</string>\\n <key>StandardErrorPath</key>\\n <string>/usr/local/pgsql/data/launchd.log</string>\\n\\t<key>OnDemand</key><!-- OS X 10.4 -->\\n\\t<false/>\\n\\t<key>KeepAlive</key><!-- OS X 10.5+ -->\\n\\t<true/>\\n</dict>\\n</plist>\\n\\nNo one replied. Want a new patch with that?\\n\\nDavid\",\n \"msg_date\": \"Fri, 20 Mar 2015 16:11:14 -0700\",\n \"msg_from\": \"\\\"David E. Wheeler\\\" <david@justatheory.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On 3/20/15 6:11 PM, David E. Wheeler wrote:\\n> \\t<key>ProgramArguments</key>\\n> \\t<array>\\n> \\t\\t<string>/usr/local/pgsql/bin/postgres</string>\\n> \\t\\t<string>-D</string>\\n> \\t\\t<string>/usr/local/pgsql/data</string>\\n> \\t</array>\\n\\nHrm, would /var/db/postgres be better? I'm not sure if the stuff Apple \\ndoes with /private/ would cause problems though. (In any case, I think \\npostgres is better than pgsql.)\\n\\n> <key>StandardOutPath</key>\\n> <string>/usr/local/pgsql/data/launchd.log</string>\\n> <key>StandardErrorPath</key>\\n> <string>/usr/local/pgsql/data/launchd.log</string>\\n\\nWouldn't /var/log be better?\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 20 Mar 2015 18:21:33 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On Mar 20, 2015, at 4:21 PM, Jim Nasby <Jim.Nasby@BlueTreble.com> wrote:\\n> \\n> On 3/20/15 6:11 PM, David E. Wheeler wrote:\\n>> \\t<key>ProgramArguments</key>\\n>> \\t<array>\\n>> \\t\\t<string>/usr/local/pgsql/bin/postgres</string>\\n>> \\t\\t<string>-D</string>\\n>> \\t\\t<string>/usr/local/pgsql/data</string>\\n>> \\t</array>\\n> \\n> Hrm, would /var/db/postgres be better? I'm not sure if the stuff Apple does with /private/ would cause problems though. (In any case, I think postgres is better than pgsql.)\\n> \\n>> <key>StandardOutPath</key>\\n>> <string>/usr/local/pgsql/data/launchd.log</string>\\n>> <key>StandardErrorPath</key>\\n>> <string>/usr/local/pgsql/data/launchd.log</string>\\n> \\n> Wouldn't /var/log be better?\\n\\n/usr/local/pgsql has been the standard install location for the PostgreSQL core distribution for as long as I can remember, including on OS X. Our original OS X startup script refers to it. I figure it’s best to keep it consistent.\\n\\nBest,\\n\\nDavid\",\n \"msg_date\": \"Fri, 20 Mar 2015 16:25:07 -0700\",\n \"msg_from\": \"\\\"David E. Wheeler\\\" <david@justatheory.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n },\n {\n \"msg_contents\": \"On Mar 20, 2015, at 4:11 PM, David E. Wheeler <david@justatheory.com> wrote:\\n\\n> No one replied. Want a new patch with that?\\n\\nHere it is.\\n\\nBest,\\n\\nDavid\",\n \"msg_date\": \"Sun, 22 Mar 2015 11:03:37 -0700\",\n \"msg_from\": \"\\\"David E. Wheeler\\\" <david@justatheory.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Patch: Add launchd Support\"\n }\n]","isConversational":false}}},{"rowIdx":39694,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"I'm trying to create what amounts to a new type. This would be rather easy if I could perform a CHECK on a composite type, which I could do if I could create a domain on top of a composite. Is there any reason in particular that hasn't been done?\n\nAs an alternative, I tried accomplishing this with a straight domain. That would work, except for this:\n\nWARNING: cast will be ignored because the source data type is a domain\n\nWhy do we ignore casts from domains to other data types? I'm guessing because it's simply not what domains were meant for?\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 18:51:48 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Questions on domain on composite / casts ignoring domains"},{"msg_contents":"Jim Nasby-5 wrote\n> I'm trying to create what amounts to a new type. This would be rather easy\n> if I could perform a CHECK on a composite type, which I could do if I\n> could create a domain on top of a composite. Is there any reason in\n> particular that hasn't been done?\n> \n> As an alternative, I tried accomplishing this with a straight domain. That\n> would work, except for this:\n> \n> WARNING: cast will be ignored because the source data type is a domain\n> \n> Why do we ignore casts from domains to other data types? I'm guessing\n> because it's simply not what domains were meant for?\n\nA domain is a base type with a constraint. When you cast you already know\nthe existing value is valid and the system simply uses the cast available\nfor the base type instead. i.e., You cannot have a domain with a different\ncast rule than the base type over which it is defined.\n\nLikely the lack of capability is simply a matter of complexity in the face\nof somewhat uncommon usage and limited resources.\n\nDavid J.\n\n\n\n\n--\nView this message in context: http://postgresql.1045698.n5.nabble.com/Questions-on-domain-on-composite-casts-ignoring-domains-tp5823745p5823763.html\nSent from the PostgreSQL - hackers mailing list archive at Nabble.com.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 20 Oct 2014 18:18:38 -0700 (PDT)","msg_from":"David G Johnston <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: Questions on domain on composite / casts ignoring domains"}],"string":"[\n {\n \"msg_contents\": \"I'm trying to create what amounts to a new type. This would be rather easy if I could perform a CHECK on a composite type, which I could do if I could create a domain on top of a composite. Is there any reason in particular that hasn't been done?\\n\\nAs an alternative, I tried accomplishing this with a straight domain. That would work, except for this:\\n\\nWARNING: cast will be ignored because the source data type is a domain\\n\\nWhy do we ignore casts from domains to other data types? I'm guessing because it's simply not what domains were meant for?\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 18:51:48 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Questions on domain on composite / casts ignoring domains\"\n },\n {\n \"msg_contents\": \"Jim Nasby-5 wrote\\n> I'm trying to create what amounts to a new type. This would be rather easy\\n> if I could perform a CHECK on a composite type, which I could do if I\\n> could create a domain on top of a composite. Is there any reason in\\n> particular that hasn't been done?\\n> \\n> As an alternative, I tried accomplishing this with a straight domain. That\\n> would work, except for this:\\n> \\n> WARNING: cast will be ignored because the source data type is a domain\\n> \\n> Why do we ignore casts from domains to other data types? I'm guessing\\n> because it's simply not what domains were meant for?\\n\\nA domain is a base type with a constraint. When you cast you already know\\nthe existing value is valid and the system simply uses the cast available\\nfor the base type instead. i.e., You cannot have a domain with a different\\ncast rule than the base type over which it is defined.\\n\\nLikely the lack of capability is simply a matter of complexity in the face\\nof somewhat uncommon usage and limited resources.\\n\\nDavid J.\\n\\n\\n\\n\\n--\\nView this message in context: http://postgresql.1045698.n5.nabble.com/Questions-on-domain-on-composite-casts-ignoring-domains-tp5823745p5823763.html\\nSent from the PostgreSQL - hackers mailing list archive at Nabble.com.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 20 Oct 2014 18:18:38 -0700 (PDT)\",\n \"msg_from\": \"David G Johnston <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Questions on domain on composite / casts ignoring domains\"\n }\n]","isConversational":false}}},{"rowIdx":39695,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"If you do any Postgres development on OS X, you've probably gotten\nseriously annoyed by the way that, every single time you reinstall the\npostmaster executable, you get a dialog box asking whether you'd like\nto allow it to accept incoming network connections. (At least, you\ndo unless you disable the OS firewall, which is not a great idea.)\nIt's particularly awful to run \"make check-world\" in this environment,\nbecause you get a pop-up for each test install.\n\nMy Salesforce colleagues researched how to fix this, and found out\nthat it can be suppressed if you sign the postgres executable, which\nyou can easily do with a self-signed certificate. Once you've allowed\nor denied network connections for a signed executable, you don't get\nprompted again when the executable is replaced, so long as it's at\nthe same file path and signed with the same certificate. So you only\nhave to dismiss the dialogs once more during a check-world run, and\nyou're done seeing them. (Tested on Mavericks and Yosemite, have not\ntried anything older.)\n\nAccordingly, we'd like to propose something like the attached patch\nto add an optional signing step to the build process. It lacks any\ndocumentation ATM, but if there are not objections to the basic idea\nI'll write some.\n\n\t\t\tregards, tom lane\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Tue, 21 Oct 2014 13:16:46 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Getting rid of \"accept incoming network connections\" prompts on OS X"},{"msg_contents":"One of the queries in ri_triggers.c has be a little baffled.\n\nFor (relatively) obvious reasons, a FK insert triggers a SELECT 1 FROM\npk_rel ... FOR KEY SHARE.\nFor not-so-obvious reasons, a PK delete triggers a SELECT 1 FROM fk_rel ...\nFOR KEY SHARE.\n\nI can't see what the lock on fk_rel achieves. Both operations are already\ncontending for the lock on the PK row, which seems like enough to cover\nevery eventuality.\n\nAnd even if the lock serves a purpose, KEY SHARE is an odd choice, since the\nreferencing field is, in general, not a \"key\" in this sense.\n\n\n\n\n-----\naaaaaaa\n--\nView this message in context: http://postgresql.1045698.n5.nabble.com/Getting-rid-of-accept-incoming-network-connections-prompts-on-OS-X-tp5823819p5823890.html\nSent from the PostgreSQL - hackers mailing list archive at Nabble.com.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 23:14:29 -0700 (PDT)","msg_from":"edward745 <cedward345@gmail.com>","msg_from_op":false,"msg_subject":"Re: Getting rid of \"accept incoming network connections\" prompts on\n OS X"},{"msg_contents":"On Tue, Oct 21, 2014 at 1:16 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> If you do any Postgres development on OS X, you've probably gotten\n> seriously annoyed by the way that, every single time you reinstall the\n> postmaster executable, you get a dialog box asking whether you'd like\n> to allow it to accept incoming network connections. (At least, you\n> do unless you disable the OS firewall, which is not a great idea.)\n> It's particularly awful to run \"make check-world\" in this environment,\n> because you get a pop-up for each test install.\n\nUgh. This must be new in Mavericks, because I don't get any such\nbehavior on 10.8.5.\n\nWhat an awful, awful behavior.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 22 Oct 2014 20:32:10 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Getting rid of \"accept incoming network connections\"\n prompts on OS X"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Tue, Oct 21, 2014 at 1:16 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> If you do any Postgres development on OS X, you've probably gotten\n>> seriously annoyed by the way that, every single time you reinstall the\n>> postmaster executable, you get a dialog box asking whether you'd like\n>> to allow it to accept incoming network connections.\n\n> Ugh. This must be new in Mavericks, because I don't get any such\n> behavior on 10.8.5.\n\nHm, I thought it went further back than that ... I remember having put up\nwith it for some time now.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 22 Oct 2014 21:02:43 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Getting rid of \"accept incoming network connections\" prompts on\n OS X"},{"msg_contents":"On 10/21/14 1:16 PM, Tom Lane wrote:\n> If you do any Postgres development on OS X, you've probably gotten\n> seriously annoyed by the way that, every single time you reinstall the\n> postmaster executable, you get a dialog box asking whether you'd like\n> to allow it to accept incoming network connections.\n\nI used to, but somehow I don't see this anymore. Just to be sure, I\nmade sure the firewall is on, checked that postgres is not in the\nexception list, rebooted, built postgresql from scratch, ran make check,\nbut no pop-up.\n\nI'm on Yosemite. Maybe this was changed.\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 24 Oct 2014 01:18:31 -0400","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: Getting rid of \"accept incoming network connections\"\n prompts on OS X"},{"msg_contents":"On Fri, Oct 24, 2014 at 7:18 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\n> On 10/21/14 1:16 PM, Tom Lane wrote:\n>> If you do any Postgres development on OS X, you've probably gotten\n>> seriously annoyed by the way that, every single time you reinstall the\n>> postmaster executable, you get a dialog box asking whether you'd like\n>> to allow it to accept incoming network connections.\n>\n> I used to, but somehow I don't see this anymore. Just to be sure, I\n> made sure the firewall is on, checked that postgres is not in the\n> exception list, rebooted, built postgresql from scratch, ran make check,\n> but no pop-up.\n>\n> I'm on Yosemite. Maybe this was changed.\n\nI've never seen it on any version of OS X (I've worked my way from\nPanther to Yosemite). There must be more to it...\n\n-- \nDave Page\nBlog: http://pgsnake.blogspot.com\nTwitter: @pgsnake\n\nEnterpriseDB UK: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 24 Oct 2014 08:26:27 +0200","msg_from":"Dave Page <dpage@pgadmin.org>","msg_from_op":false,"msg_subject":"Re: Getting rid of \"accept incoming network connections\"\n prompts on OS X"},{"msg_contents":"On Fri, Oct 24, 2014 at 8:26 AM, Dave Page <dpage@pgadmin.org> wrote:\n\n> On Fri, Oct 24, 2014 at 7:18 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\n> > On 10/21/14 1:16 PM, Tom Lane wrote:\n> >> If you do any Postgres development on OS X, you've probably gotten\n> >> seriously annoyed by the way that, every single time you reinstall the\n> >> postmaster executable, you get a dialog box asking whether you'd like\n> >> to allow it to accept incoming network connections.\n> >\n> > I used to, but somehow I don't see this anymore. Just to be sure, I\n> > made sure the firewall is on, checked that postgres is not in the\n> > exception list, rebooted, built postgresql from scratch, ran make check,\n> > but no pop-up.\n> >\n> > I'm on Yosemite. Maybe this was changed.\n>\n> I've never seen it on any version of OS X (I've worked my way from\n> Panther to Yosemite). There must be more to it...\n>\nFWIW, with firewall at on, I am used to see this annoying popup window when\nstarting an instance manually, make check never complains though.\n-- \nMichael\n\nOn Fri, Oct 24, 2014 at 8:26 AM, Dave Page <dpage@pgadmin.org> wrote:On Fri, Oct 24, 2014 at 7:18 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\n> On 10/21/14 1:16 PM, Tom Lane wrote:\n>> If you do any Postgres development on OS X, you've probably gotten\n>> seriously annoyed by the way that, every single time you reinstall the\n>> postmaster executable, you get a dialog box asking whether you'd like\n>> to allow it to accept incoming network connections.\n>\n> I used to, but somehow I don't see this anymore. Just to be sure, I\n> made sure the firewall is on, checked that postgres is not in the\n> exception list, rebooted, built postgresql from scratch, ran make check,\n> but no pop-up.\n>\n> I'm on Yosemite. Maybe this was changed.\n\nI've never seen it on any version of OS X (I've worked my way from\nPanther to Yosemite). There must be more to it...FWIW, with firewall at on, I am used to see this annoying popup window when starting an instance manually, make check never complains though.-- Michael","msg_date":"Fri, 24 Oct 2014 09:10:54 +0200","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Getting rid of \"accept incoming network connections\"\n prompts on OS X"},{"msg_contents":"Michael Paquier <michael.paquier@gmail.com> writes:\n> On Fri, Oct 24, 2014 at 8:26 AM, Dave Page <dpage@pgadmin.org> wrote:\n>> On Fri, Oct 24, 2014 at 7:18 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\n>>> On 10/21/14 1:16 PM, Tom Lane wrote:\n>>>> If you do any Postgres development on OS X, you've probably gotten\n>>>> seriously annoyed by the way that, every single time you reinstall the\n>>>> postmaster executable, you get a dialog box asking whether you'd like\n>>>> to allow it to accept incoming network connections.\n\n>>> I used to, but somehow I don't see this anymore. Just to be sure, I\n>>> made sure the firewall is on, checked that postgres is not in the\n>>> exception list, rebooted, built postgresql from scratch, ran make check,\n>>> but no pop-up.\n>>> \n>>> I'm on Yosemite. Maybe this was changed.\n\n>> I've never seen it on any version of OS X (I've worked my way from\n>> Panther to Yosemite). There must be more to it...\n\nI see it every darn time I've changed the postmaster executable.\nMaybe there is a difference in security settings? I have the firewall\nenabled and in Settings->Security->General, \"Allow apps downloaded from:\nMac App Store and identified developers\", which I think is the default.\n[ experiments... ] Hm, setting that to \"Anywhere\" doesn't change the\nresults anyway.\n\n> FWIW, with firewall at on, I am used to see this annoying popup window when\n> starting an instance manually, make check never complains though.\n\nAh. pg_regress sets listen_addresses to empty so that no TCP ports\nare opened, hence no firewall complaints from \"make check\". However,\nas soon as you start a normal installation, you get the complaint,\nas even an open port on 127.0.0.1 is enough to provoke it.\n\nPeter, Dave: maybe you have tweaked things to keep listen_addresses\nempty and rely only on Unix-socket connections?\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 24 Oct 2014 09:39:59 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Getting rid of \"accept incoming network connections\" prompts on\n OS X"},{"msg_contents":"On Wed, Oct 22, 2014 at 2:14 AM, edward745 <cedward345@gmail.com> wrote:\n> One of the queries in ri_triggers.c has be a little baffled.\n>\n> For (relatively) obvious reasons, a FK insert triggers a SELECT 1 FROM\n> pk_rel ... FOR KEY SHARE.\n> For not-so-obvious reasons, a PK delete triggers a SELECT 1 FROM fk_rel ...\n> FOR KEY SHARE.\n>\n> I can't see what the lock on fk_rel achieves. Both operations are already\n> contending for the lock on the PK row, which seems like enough to cover\n> every eventuality.\n>\n> And even if the lock serves a purpose, KEY SHARE is an odd choice, since the\n> referencing field is, in general, not a \"key\" in this sense.\n\nPlease don't post unrelated questions onto existing mailing list\nthreads. Start a new thread for a new topic.\n\nThanks,\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 24 Oct 2014 09:49:36 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Re: Getting rid of \"accept incoming network\n connections\" prompts on OS X"},{"msg_contents":"On Fri, Oct 24, 2014 at 3:39 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Peter, Dave: maybe you have tweaked things to keep listen_addresses\n> empty and rely only on Unix-socket connections?\nShould be so. The target of this feature is development on OSX, right?\nAnd most of the time development would be done only on the local\nmachine, machine being most of the time a laptop. So instead of adding\nan optional step in configure to enforce the creation of a\ncertificate, why not simply encourage people to use listen_addresses =\n'' on OSX by documenting it? Even when working on replication or\nrelated things on a local machine, it is possible to simply pass the\nsocket directory...\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 24 Oct 2014 16:05:16 +0200","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: Getting rid of \"accept incoming network connections\"\n prompts on OS X"},{"msg_contents":"Michael Paquier <michael.paquier@gmail.com> writes:\n> On Fri, Oct 24, 2014 at 3:39 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> Peter, Dave: maybe you have tweaked things to keep listen_addresses\n>> empty and rely only on Unix-socket connections?\n\n> Should be so. The target of this feature is development on OSX, right?\n> And most of the time development would be done only on the local\n> machine, machine being most of the time a laptop. So instead of adding\n> an optional step in configure to enforce the creation of a\n> certificate, why not simply encourage people to use listen_addresses =\n> '' on OSX by documenting it? Even when working on replication or\n> related things on a local machine, it is possible to simply pass the\n> socket directory...\n\nSome clients (eg JDBC) don't support Unix-socket connections AFAIK, so\nthis seems like a rather restricted solution.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 24 Oct 2014 10:27:15 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Getting rid of \"accept incoming network connections\" prompts on\n OS X"},{"msg_contents":"On 10/24/14 10:27 AM, Tom Lane wrote:\n> Some clients (eg JDBC) don't support Unix-socket connections AFAIK, so\n> this seems like a rather restricted solution.\n\nWhile this is a valid point, they're actually working on fixing that.\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 25 Oct 2014 09:06:38 -0400","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: Getting rid of \"accept incoming network connections\"\n prompts on OS X"},{"msg_contents":"On 10/24/14 9:39 AM, Tom Lane wrote:\n> Peter, Dave: maybe you have tweaked things to keep listen_addresses\n> empty and rely only on Unix-socket connections?\n\nI can confirm that I do get the popup when starting an installed\npostmaster with the default settings.\n\nGiven that this doesn't affect \"make check\" anymore, I'm unsure about\nthis patch. There is a lot of magic in the configure change. I don't\nknow what to pass as the configure option argument, so can't really\nevaluate that. I'd like to see an explanation for what is done there.\n\nI'm afraid there is security ridicule potential. We are essentially\nadding an option to patch out an operating system security feature that\nthe user chose. Some might find that neat and ship binaries built that\nway. Because it's \"--with-codesign\" and not\n\"--with-codesign-for-devel-dont-use-in-production\".\n\nHave we dug deep enough into the firewall configuration to evaluate\nother options? Can we, for example, exclude a port range?\n\nI could see adding this as a contrib script if we don't find a better way.\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 25 Oct 2014 09:22:30 -0400","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: Getting rid of \"accept incoming network connections\"\n prompts on OS X"},{"msg_contents":"Peter Eisentraut <peter_e@gmx.net> writes:\n> Given that this doesn't affect \"make check\" anymore, I'm unsure about\n> this patch. There is a lot of magic in the configure change. I don't\n> know what to pass as the configure option argument, so can't really\n> evaluate that. I'd like to see an explanation for what is done there.\n\nAs I said, I'd not written any docs. The argument that would get passed\nthere is just a name identifying the signing certificate you want to use.\nMost of the documentation would be about how to create such a cert, no\ndoubt. (It's pretty simple to make a self-signed cert using Apple's\nkeychain utility, but it would require some explanation.)\n\n> I'm afraid there is security ridicule potential. We are essentially\n> adding an option to patch out an operating system security feature that\n> the user chose. Some might find that neat and ship binaries built that\n> way. Because it's \"--with-codesign\" and not\n> \"--with-codesign-for-devel-dont-use-in-production\".\n\nYeah, that would be a risk :-(. However, for the typical case of a\nself-signed certificate, nothing much would happen because no one\nelse's machine would even have the same certificate let alone trust it.\n\n> Have we dug deep enough into the firewall configuration to evaluate\n> other options? Can we, for example, exclude a port range?\n\nNot that I've been able to detect. Any simple way to do that would\npresumably open up exactly the security hole Apple is trying to close,\nso I'd bet against there being one. (It is annoying that the firewall\ntriggers on ports bound to 127.0.0.1, though --- it's not apparent why\nthat's a security risk. Perhaps there's some way to adjust that choice?)\n\n> I could see adding this as a contrib script if we don't find a better way.\n\nMeh. That's just a less convenient packaging of the same code, with\nthe same potential for misuse.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 25 Oct 2014 14:26:01 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Getting rid of \"accept incoming network connections\" prompts on\n OS X"},{"msg_contents":"I wrote:\n> Peter Eisentraut <peter_e@gmx.net> writes:\n>> Have we dug deep enough into the firewall configuration to evaluate\n>> other options? Can we, for example, exclude a port range?\n\n> Not that I've been able to detect. Any simple way to do that would\n> presumably open up exactly the security hole Apple is trying to close,\n> so I'd bet against there being one. (It is annoying that the firewall\n> triggers on ports bound to 127.0.0.1, though --- it's not apparent why\n> that's a security risk. Perhaps there's some way to adjust that choice?)\n\nAnd a bit of experimentation later: it seems that on Yosemite (and\nprobably earlier OS X versions), \"localhost\" maps to all three of these\naddresses:\n\t127.0.0.1\n\t::1\n\tfe80:1::1\nBinding to 127.0.0.1 does not trigger the firewall popup. Binding\nto ::1 doesn't, either. But binding to fe80:1::1 does. So the\neasy fix, for a default installation, is to keep the postmaster\nfrom binding to that last address.\n\nI'm not sufficiently up on my IPv6 to be sure exactly what that third\naddress does. Perhaps it is a bug in the firewall logic that it\nconsiders that address external? If it *is* externally accessible,\nwhat the heck is the OS doing including it in \"localhost\"?\n\n(Not sure if it's relevant, but I've got IPv6 set to \"link-local only\"\nin network preferences.)\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 25 Oct 2014 14:52:36 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Getting rid of \"accept incoming network connections\" prompts on\n OS X"},{"msg_contents":"On 10/25/14 2:52 PM, Tom Lane wrote:\n> And a bit of experimentation later: it seems that on Yosemite (and\n> probably earlier OS X versions), \"localhost\" maps to all three of these\n> addresses:\n> \t127.0.0.1\n> \t::1\n> \tfe80:1::1\n> Binding to 127.0.0.1 does not trigger the firewall popup. Binding\n> to ::1 doesn't, either. But binding to fe80:1::1 does. So the\n> easy fix, for a default installation, is to keep the postmaster\n> from binding to that last address.\n> \n> I'm not sufficiently up on my IPv6 to be sure exactly what that third\n> address does. Perhaps it is a bug in the firewall logic that it\n> considers that address external?\n\nI think that's exactly it. I have filed a bug with Apple about it.\n\nFor the time begin, I think it's a reasonable workaround to comment out\nthe line in /etc/hosts.\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 25 Oct 2014 22:02:05 -0400","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":false,"msg_subject":"Re: Getting rid of \"accept incoming network connections\"\n prompts on OS X"},{"msg_contents":"Peter Eisentraut <peter_e@gmx.net> writes:\n> On 10/25/14 2:52 PM, Tom Lane wrote:\n>> And a bit of experimentation later: it seems that on Yosemite (and\n>> probably earlier OS X versions), \"localhost\" maps to all three of these\n>> addresses:\n>> 127.0.0.1\n>> ::1\n>> fe80:1::1\n>> Binding to 127.0.0.1 does not trigger the firewall popup. Binding\n>> to ::1 doesn't, either. But binding to fe80:1::1 does. So the\n>> easy fix, for a default installation, is to keep the postmaster\n>> from binding to that last address.\n>> \n>> I'm not sufficiently up on my IPv6 to be sure exactly what that third\n>> address does. Perhaps it is a bug in the firewall logic that it\n>> considers that address external?\n\n> I think that's exactly it. I have filed a bug with Apple about it.\n\n> For the time begin, I think it's a reasonable workaround to comment out\n> the line in /etc/hosts.\n\nHmm ... I was about to object that that seemed likely to break other\nstuff, but on poking around I notice that my non-laptop Yosemite machine\nhas no such entry at all. (That one was originally Mountain Lion, and was\nupgraded to Mavericks and then Yosemite, whereas my laptop was Mavericks\non delivery.) Even more interesting, there's a \"/etc/hosts~orig\" file on\nmy laptop that does not have the entry.\n\nA little bit of data gathering later:\nwife's laptop (10.9.5): has it\ndromedary (10.6.8): has it\nprairiedog (10.4.11): doesn't have it\n\nSo it looks like Apple has been using this for awhile but it's not really\nessential to anything. What worries me a bit is that the evidence on my\nlaptop suggests there may be code somewhere (like System Preferences) that\nwill edit the file, so that a manual removal might not \"keep\".\nWe shall see.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 25 Oct 2014 22:40:12 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Getting rid of \"accept incoming network connections\" prompts on\n OS X"},{"msg_contents":"[ blast-from-the-past department ]\n\nPeter Eisentraut <peter_e@gmx.net> writes:\n> On 10/25/14 2:52 PM, Tom Lane wrote:\n>> And a bit of experimentation later: it seems that on Yosemite (and\n>> probably earlier OS X versions), \"localhost\" maps to all three of these\n>> addresses:\n>> 127.0.0.1\n>> ::1\n>> fe80:1::1\n>> Binding to 127.0.0.1 does not trigger the firewall popup. Binding\n>> to ::1 doesn't, either. But binding to fe80:1::1 does. So the\n>> easy fix, for a default installation, is to keep the postmaster\n>> from binding to that last address.\n>> \n>> I'm not sufficiently up on my IPv6 to be sure exactly what that third\n>> address does. Perhaps it is a bug in the firewall logic that it\n>> considers that address external?\n\n> I think that's exactly it. I have filed a bug with Apple about it.\n\nThis came up again today, and I've confirmed that the issue still exists\nin current macOS. Did you get any response to your bug report, and if\nso what did they say?\n\nBTW, I seem to have typo'd in the above-quoted message --- what's\nactually in the problematic line, at least on my Mac, is\n\nfe80::1%lo0 localhost\n\nwhich is not quite the same address I wrote above. It's definitely\na link-local address though, and the binding to %lo0 does seem to\nbe understood by macOS, at least to the extent that netstat reports\nit:\n\n$ netstat -a -n\nActive Internet connections (including servers)\nProto Recv-Q Send-Q Local Address Foreign Address (state) \ntcp4 0 0 127.0.0.1.5432 *.* LISTEN \ntcp6 0 0 fe80::1%lo0.5432 *.* LISTEN \ntcp6 0 0 ::1.5432 *.* LISTEN \n... unrelated stuff ...\n\nSo either the kernel isn't really treating it as link-local on the\nloopback interface, or the firewall doesn't understand that it's\nlink-local. Either way, sure seems like an Apple bug.\n\n\t\t\tregards, tom lane\n\n","msg_date":"Thu, 26 Jul 2018 17:45:17 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: [HACKERS] Getting rid of \"accept incoming network connections\"\n prompts on OS X"},{"msg_contents":"On 26/07/2018 23:45, Tom Lane wrote:\n> This came up again today, and I've confirmed that the issue still exists\n> in current macOS. Did you get any response to your bug report, and if\n> so what did they say?\n\nThere hasn't been any response to the radar. I think our analysis is\ncorrect, it's an OS bug.\n\n-- \nPeter Eisentraut http://www.2ndQuadrant.com/\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\n\n","msg_date":"Mon, 30 Jul 2018 21:09:06 +0200","msg_from":"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: [HACKERS] Getting rid of \"accept incoming network connections\"\n prompts on OS X"}],"string":"[\n {\n \"msg_contents\": \"If you do any Postgres development on OS X, you've probably gotten\\nseriously annoyed by the way that, every single time you reinstall the\\npostmaster executable, you get a dialog box asking whether you'd like\\nto allow it to accept incoming network connections. (At least, you\\ndo unless you disable the OS firewall, which is not a great idea.)\\nIt's particularly awful to run \\\"make check-world\\\" in this environment,\\nbecause you get a pop-up for each test install.\\n\\nMy Salesforce colleagues researched how to fix this, and found out\\nthat it can be suppressed if you sign the postgres executable, which\\nyou can easily do with a self-signed certificate. Once you've allowed\\nor denied network connections for a signed executable, you don't get\\nprompted again when the executable is replaced, so long as it's at\\nthe same file path and signed with the same certificate. So you only\\nhave to dismiss the dialogs once more during a check-world run, and\\nyou're done seeing them. (Tested on Mavericks and Yosemite, have not\\ntried anything older.)\\n\\nAccordingly, we'd like to propose something like the attached patch\\nto add an optional signing step to the build process. It lacks any\\ndocumentation ATM, but if there are not objections to the basic idea\\nI'll write some.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Tue, 21 Oct 2014 13:16:46 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Getting rid of \\\"accept incoming network connections\\\" prompts on OS X\"\n },\n {\n \"msg_contents\": \"One of the queries in ri_triggers.c has be a little baffled.\\n\\nFor (relatively) obvious reasons, a FK insert triggers a SELECT 1 FROM\\npk_rel ... FOR KEY SHARE.\\nFor not-so-obvious reasons, a PK delete triggers a SELECT 1 FROM fk_rel ...\\nFOR KEY SHARE.\\n\\nI can't see what the lock on fk_rel achieves. Both operations are already\\ncontending for the lock on the PK row, which seems like enough to cover\\nevery eventuality.\\n\\nAnd even if the lock serves a purpose, KEY SHARE is an odd choice, since the\\nreferencing field is, in general, not a \\\"key\\\" in this sense.\\n\\n\\n\\n\\n-----\\naaaaaaa\\n--\\nView this message in context: http://postgresql.1045698.n5.nabble.com/Getting-rid-of-accept-incoming-network-connections-prompts-on-OS-X-tp5823819p5823890.html\\nSent from the PostgreSQL - hackers mailing list archive at Nabble.com.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 23:14:29 -0700 (PDT)\",\n \"msg_from\": \"edward745 <cedward345@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\" prompts on\\n OS X\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 21, 2014 at 1:16 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> If you do any Postgres development on OS X, you've probably gotten\\n> seriously annoyed by the way that, every single time you reinstall the\\n> postmaster executable, you get a dialog box asking whether you'd like\\n> to allow it to accept incoming network connections. (At least, you\\n> do unless you disable the OS firewall, which is not a great idea.)\\n> It's particularly awful to run \\\"make check-world\\\" in this environment,\\n> because you get a pop-up for each test install.\\n\\nUgh. This must be new in Mavericks, because I don't get any such\\nbehavior on 10.8.5.\\n\\nWhat an awful, awful behavior.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 22 Oct 2014 20:32:10 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\"\\n prompts on OS X\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Tue, Oct 21, 2014 at 1:16 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> If you do any Postgres development on OS X, you've probably gotten\\n>> seriously annoyed by the way that, every single time you reinstall the\\n>> postmaster executable, you get a dialog box asking whether you'd like\\n>> to allow it to accept incoming network connections.\\n\\n> Ugh. This must be new in Mavericks, because I don't get any such\\n> behavior on 10.8.5.\\n\\nHm, I thought it went further back than that ... I remember having put up\\nwith it for some time now.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 22 Oct 2014 21:02:43 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\" prompts on\\n OS X\"\n },\n {\n \"msg_contents\": \"On 10/21/14 1:16 PM, Tom Lane wrote:\\n> If you do any Postgres development on OS X, you've probably gotten\\n> seriously annoyed by the way that, every single time you reinstall the\\n> postmaster executable, you get a dialog box asking whether you'd like\\n> to allow it to accept incoming network connections.\\n\\nI used to, but somehow I don't see this anymore. Just to be sure, I\\nmade sure the firewall is on, checked that postgres is not in the\\nexception list, rebooted, built postgresql from scratch, ran make check,\\nbut no pop-up.\\n\\nI'm on Yosemite. Maybe this was changed.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 24 Oct 2014 01:18:31 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\"\\n prompts on OS X\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 24, 2014 at 7:18 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n> On 10/21/14 1:16 PM, Tom Lane wrote:\\n>> If you do any Postgres development on OS X, you've probably gotten\\n>> seriously annoyed by the way that, every single time you reinstall the\\n>> postmaster executable, you get a dialog box asking whether you'd like\\n>> to allow it to accept incoming network connections.\\n>\\n> I used to, but somehow I don't see this anymore. Just to be sure, I\\n> made sure the firewall is on, checked that postgres is not in the\\n> exception list, rebooted, built postgresql from scratch, ran make check,\\n> but no pop-up.\\n>\\n> I'm on Yosemite. Maybe this was changed.\\n\\nI've never seen it on any version of OS X (I've worked my way from\\nPanther to Yosemite). There must be more to it...\\n\\n-- \\nDave Page\\nBlog: http://pgsnake.blogspot.com\\nTwitter: @pgsnake\\n\\nEnterpriseDB UK: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 24 Oct 2014 08:26:27 +0200\",\n \"msg_from\": \"Dave Page <dpage@pgadmin.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\"\\n prompts on OS X\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 24, 2014 at 8:26 AM, Dave Page <dpage@pgadmin.org> wrote:\\n\\n> On Fri, Oct 24, 2014 at 7:18 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n> > On 10/21/14 1:16 PM, Tom Lane wrote:\\n> >> If you do any Postgres development on OS X, you've probably gotten\\n> >> seriously annoyed by the way that, every single time you reinstall the\\n> >> postmaster executable, you get a dialog box asking whether you'd like\\n> >> to allow it to accept incoming network connections.\\n> >\\n> > I used to, but somehow I don't see this anymore. Just to be sure, I\\n> > made sure the firewall is on, checked that postgres is not in the\\n> > exception list, rebooted, built postgresql from scratch, ran make check,\\n> > but no pop-up.\\n> >\\n> > I'm on Yosemite. Maybe this was changed.\\n>\\n> I've never seen it on any version of OS X (I've worked my way from\\n> Panther to Yosemite). There must be more to it...\\n>\\nFWIW, with firewall at on, I am used to see this annoying popup window when\\nstarting an instance manually, make check never complains though.\\n-- \\nMichael\\n\\nOn Fri, Oct 24, 2014 at 8:26 AM, Dave Page <dpage@pgadmin.org> wrote:On Fri, Oct 24, 2014 at 7:18 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n> On 10/21/14 1:16 PM, Tom Lane wrote:\\n>> If you do any Postgres development on OS X, you've probably gotten\\n>> seriously annoyed by the way that, every single time you reinstall the\\n>> postmaster executable, you get a dialog box asking whether you'd like\\n>> to allow it to accept incoming network connections.\\n>\\n> I used to, but somehow I don't see this anymore. Just to be sure, I\\n> made sure the firewall is on, checked that postgres is not in the\\n> exception list, rebooted, built postgresql from scratch, ran make check,\\n> but no pop-up.\\n>\\n> I'm on Yosemite. Maybe this was changed.\\n\\nI've never seen it on any version of OS X (I've worked my way from\\nPanther to Yosemite). There must be more to it...FWIW, with firewall at on, I am used to see this annoying popup window when starting an instance manually, make check never complains though.-- Michael\",\n \"msg_date\": \"Fri, 24 Oct 2014 09:10:54 +0200\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\"\\n prompts on OS X\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael.paquier@gmail.com> writes:\\n> On Fri, Oct 24, 2014 at 8:26 AM, Dave Page <dpage@pgadmin.org> wrote:\\n>> On Fri, Oct 24, 2014 at 7:18 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n>>> On 10/21/14 1:16 PM, Tom Lane wrote:\\n>>>> If you do any Postgres development on OS X, you've probably gotten\\n>>>> seriously annoyed by the way that, every single time you reinstall the\\n>>>> postmaster executable, you get a dialog box asking whether you'd like\\n>>>> to allow it to accept incoming network connections.\\n\\n>>> I used to, but somehow I don't see this anymore. Just to be sure, I\\n>>> made sure the firewall is on, checked that postgres is not in the\\n>>> exception list, rebooted, built postgresql from scratch, ran make check,\\n>>> but no pop-up.\\n>>> \\n>>> I'm on Yosemite. Maybe this was changed.\\n\\n>> I've never seen it on any version of OS X (I've worked my way from\\n>> Panther to Yosemite). There must be more to it...\\n\\nI see it every darn time I've changed the postmaster executable.\\nMaybe there is a difference in security settings? I have the firewall\\nenabled and in Settings->Security->General, \\\"Allow apps downloaded from:\\nMac App Store and identified developers\\\", which I think is the default.\\n[ experiments... ] Hm, setting that to \\\"Anywhere\\\" doesn't change the\\nresults anyway.\\n\\n> FWIW, with firewall at on, I am used to see this annoying popup window when\\n> starting an instance manually, make check never complains though.\\n\\nAh. pg_regress sets listen_addresses to empty so that no TCP ports\\nare opened, hence no firewall complaints from \\\"make check\\\". However,\\nas soon as you start a normal installation, you get the complaint,\\nas even an open port on 127.0.0.1 is enough to provoke it.\\n\\nPeter, Dave: maybe you have tweaked things to keep listen_addresses\\nempty and rely only on Unix-socket connections?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 24 Oct 2014 09:39:59 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\" prompts on\\n OS X\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 22, 2014 at 2:14 AM, edward745 <cedward345@gmail.com> wrote:\\n> One of the queries in ri_triggers.c has be a little baffled.\\n>\\n> For (relatively) obvious reasons, a FK insert triggers a SELECT 1 FROM\\n> pk_rel ... FOR KEY SHARE.\\n> For not-so-obvious reasons, a PK delete triggers a SELECT 1 FROM fk_rel ...\\n> FOR KEY SHARE.\\n>\\n> I can't see what the lock on fk_rel achieves. Both operations are already\\n> contending for the lock on the PK row, which seems like enough to cover\\n> every eventuality.\\n>\\n> And even if the lock serves a purpose, KEY SHARE is an odd choice, since the\\n> referencing field is, in general, not a \\\"key\\\" in this sense.\\n\\nPlease don't post unrelated questions onto existing mailing list\\nthreads. Start a new thread for a new topic.\\n\\nThanks,\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 24 Oct 2014 09:49:36 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Re: Getting rid of \\\"accept incoming network\\n connections\\\" prompts on OS X\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 24, 2014 at 3:39 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Peter, Dave: maybe you have tweaked things to keep listen_addresses\\n> empty and rely only on Unix-socket connections?\\nShould be so. The target of this feature is development on OSX, right?\\nAnd most of the time development would be done only on the local\\nmachine, machine being most of the time a laptop. So instead of adding\\nan optional step in configure to enforce the creation of a\\ncertificate, why not simply encourage people to use listen_addresses =\\n'' on OSX by documenting it? Even when working on replication or\\nrelated things on a local machine, it is possible to simply pass the\\nsocket directory...\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 24 Oct 2014 16:05:16 +0200\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\"\\n prompts on OS X\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael.paquier@gmail.com> writes:\\n> On Fri, Oct 24, 2014 at 3:39 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> Peter, Dave: maybe you have tweaked things to keep listen_addresses\\n>> empty and rely only on Unix-socket connections?\\n\\n> Should be so. The target of this feature is development on OSX, right?\\n> And most of the time development would be done only on the local\\n> machine, machine being most of the time a laptop. So instead of adding\\n> an optional step in configure to enforce the creation of a\\n> certificate, why not simply encourage people to use listen_addresses =\\n> '' on OSX by documenting it? Even when working on replication or\\n> related things on a local machine, it is possible to simply pass the\\n> socket directory...\\n\\nSome clients (eg JDBC) don't support Unix-socket connections AFAIK, so\\nthis seems like a rather restricted solution.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 24 Oct 2014 10:27:15 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\" prompts on\\n OS X\"\n },\n {\n \"msg_contents\": \"On 10/24/14 10:27 AM, Tom Lane wrote:\\n> Some clients (eg JDBC) don't support Unix-socket connections AFAIK, so\\n> this seems like a rather restricted solution.\\n\\nWhile this is a valid point, they're actually working on fixing that.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 25 Oct 2014 09:06:38 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\"\\n prompts on OS X\"\n },\n {\n \"msg_contents\": \"On 10/24/14 9:39 AM, Tom Lane wrote:\\n> Peter, Dave: maybe you have tweaked things to keep listen_addresses\\n> empty and rely only on Unix-socket connections?\\n\\nI can confirm that I do get the popup when starting an installed\\npostmaster with the default settings.\\n\\nGiven that this doesn't affect \\\"make check\\\" anymore, I'm unsure about\\nthis patch. There is a lot of magic in the configure change. I don't\\nknow what to pass as the configure option argument, so can't really\\nevaluate that. I'd like to see an explanation for what is done there.\\n\\nI'm afraid there is security ridicule potential. We are essentially\\nadding an option to patch out an operating system security feature that\\nthe user chose. Some might find that neat and ship binaries built that\\nway. Because it's \\\"--with-codesign\\\" and not\\n\\\"--with-codesign-for-devel-dont-use-in-production\\\".\\n\\nHave we dug deep enough into the firewall configuration to evaluate\\nother options? Can we, for example, exclude a port range?\\n\\nI could see adding this as a contrib script if we don't find a better way.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 25 Oct 2014 09:22:30 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\"\\n prompts on OS X\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter_e@gmx.net> writes:\\n> Given that this doesn't affect \\\"make check\\\" anymore, I'm unsure about\\n> this patch. There is a lot of magic in the configure change. I don't\\n> know what to pass as the configure option argument, so can't really\\n> evaluate that. I'd like to see an explanation for what is done there.\\n\\nAs I said, I'd not written any docs. The argument that would get passed\\nthere is just a name identifying the signing certificate you want to use.\\nMost of the documentation would be about how to create such a cert, no\\ndoubt. (It's pretty simple to make a self-signed cert using Apple's\\nkeychain utility, but it would require some explanation.)\\n\\n> I'm afraid there is security ridicule potential. We are essentially\\n> adding an option to patch out an operating system security feature that\\n> the user chose. Some might find that neat and ship binaries built that\\n> way. Because it's \\\"--with-codesign\\\" and not\\n> \\\"--with-codesign-for-devel-dont-use-in-production\\\".\\n\\nYeah, that would be a risk :-(. However, for the typical case of a\\nself-signed certificate, nothing much would happen because no one\\nelse's machine would even have the same certificate let alone trust it.\\n\\n> Have we dug deep enough into the firewall configuration to evaluate\\n> other options? Can we, for example, exclude a port range?\\n\\nNot that I've been able to detect. Any simple way to do that would\\npresumably open up exactly the security hole Apple is trying to close,\\nso I'd bet against there being one. (It is annoying that the firewall\\ntriggers on ports bound to 127.0.0.1, though --- it's not apparent why\\nthat's a security risk. Perhaps there's some way to adjust that choice?)\\n\\n> I could see adding this as a contrib script if we don't find a better way.\\n\\nMeh. That's just a less convenient packaging of the same code, with\\nthe same potential for misuse.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 25 Oct 2014 14:26:01 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\" prompts on\\n OS X\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Peter Eisentraut <peter_e@gmx.net> writes:\\n>> Have we dug deep enough into the firewall configuration to evaluate\\n>> other options? Can we, for example, exclude a port range?\\n\\n> Not that I've been able to detect. Any simple way to do that would\\n> presumably open up exactly the security hole Apple is trying to close,\\n> so I'd bet against there being one. (It is annoying that the firewall\\n> triggers on ports bound to 127.0.0.1, though --- it's not apparent why\\n> that's a security risk. Perhaps there's some way to adjust that choice?)\\n\\nAnd a bit of experimentation later: it seems that on Yosemite (and\\nprobably earlier OS X versions), \\\"localhost\\\" maps to all three of these\\naddresses:\\n\\t127.0.0.1\\n\\t::1\\n\\tfe80:1::1\\nBinding to 127.0.0.1 does not trigger the firewall popup. Binding\\nto ::1 doesn't, either. But binding to fe80:1::1 does. So the\\neasy fix, for a default installation, is to keep the postmaster\\nfrom binding to that last address.\\n\\nI'm not sufficiently up on my IPv6 to be sure exactly what that third\\naddress does. Perhaps it is a bug in the firewall logic that it\\nconsiders that address external? If it *is* externally accessible,\\nwhat the heck is the OS doing including it in \\\"localhost\\\"?\\n\\n(Not sure if it's relevant, but I've got IPv6 set to \\\"link-local only\\\"\\nin network preferences.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 25 Oct 2014 14:52:36 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\" prompts on\\n OS X\"\n },\n {\n \"msg_contents\": \"On 10/25/14 2:52 PM, Tom Lane wrote:\\n> And a bit of experimentation later: it seems that on Yosemite (and\\n> probably earlier OS X versions), \\\"localhost\\\" maps to all three of these\\n> addresses:\\n> \\t127.0.0.1\\n> \\t::1\\n> \\tfe80:1::1\\n> Binding to 127.0.0.1 does not trigger the firewall popup. Binding\\n> to ::1 doesn't, either. But binding to fe80:1::1 does. So the\\n> easy fix, for a default installation, is to keep the postmaster\\n> from binding to that last address.\\n> \\n> I'm not sufficiently up on my IPv6 to be sure exactly what that third\\n> address does. Perhaps it is a bug in the firewall logic that it\\n> considers that address external?\\n\\nI think that's exactly it. I have filed a bug with Apple about it.\\n\\nFor the time begin, I think it's a reasonable workaround to comment out\\nthe line in /etc/hosts.\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 25 Oct 2014 22:02:05 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\"\\n prompts on OS X\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter_e@gmx.net> writes:\\n> On 10/25/14 2:52 PM, Tom Lane wrote:\\n>> And a bit of experimentation later: it seems that on Yosemite (and\\n>> probably earlier OS X versions), \\\"localhost\\\" maps to all three of these\\n>> addresses:\\n>> 127.0.0.1\\n>> ::1\\n>> fe80:1::1\\n>> Binding to 127.0.0.1 does not trigger the firewall popup. Binding\\n>> to ::1 doesn't, either. But binding to fe80:1::1 does. So the\\n>> easy fix, for a default installation, is to keep the postmaster\\n>> from binding to that last address.\\n>> \\n>> I'm not sufficiently up on my IPv6 to be sure exactly what that third\\n>> address does. Perhaps it is a bug in the firewall logic that it\\n>> considers that address external?\\n\\n> I think that's exactly it. I have filed a bug with Apple about it.\\n\\n> For the time begin, I think it's a reasonable workaround to comment out\\n> the line in /etc/hosts.\\n\\nHmm ... I was about to object that that seemed likely to break other\\nstuff, but on poking around I notice that my non-laptop Yosemite machine\\nhas no such entry at all. (That one was originally Mountain Lion, and was\\nupgraded to Mavericks and then Yosemite, whereas my laptop was Mavericks\\non delivery.) Even more interesting, there's a \\\"/etc/hosts~orig\\\" file on\\nmy laptop that does not have the entry.\\n\\nA little bit of data gathering later:\\nwife's laptop (10.9.5): has it\\ndromedary (10.6.8): has it\\nprairiedog (10.4.11): doesn't have it\\n\\nSo it looks like Apple has been using this for awhile but it's not really\\nessential to anything. What worries me a bit is that the evidence on my\\nlaptop suggests there may be code somewhere (like System Preferences) that\\nwill edit the file, so that a manual removal might not \\\"keep\\\".\\nWe shall see.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 25 Oct 2014 22:40:12 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Getting rid of \\\"accept incoming network connections\\\" prompts on\\n OS X\"\n },\n {\n \"msg_contents\": \"[ blast-from-the-past department ]\\n\\nPeter Eisentraut <peter_e@gmx.net> writes:\\n> On 10/25/14 2:52 PM, Tom Lane wrote:\\n>> And a bit of experimentation later: it seems that on Yosemite (and\\n>> probably earlier OS X versions), \\\"localhost\\\" maps to all three of these\\n>> addresses:\\n>> 127.0.0.1\\n>> ::1\\n>> fe80:1::1\\n>> Binding to 127.0.0.1 does not trigger the firewall popup. Binding\\n>> to ::1 doesn't, either. But binding to fe80:1::1 does. So the\\n>> easy fix, for a default installation, is to keep the postmaster\\n>> from binding to that last address.\\n>> \\n>> I'm not sufficiently up on my IPv6 to be sure exactly what that third\\n>> address does. Perhaps it is a bug in the firewall logic that it\\n>> considers that address external?\\n\\n> I think that's exactly it. I have filed a bug with Apple about it.\\n\\nThis came up again today, and I've confirmed that the issue still exists\\nin current macOS. Did you get any response to your bug report, and if\\nso what did they say?\\n\\nBTW, I seem to have typo'd in the above-quoted message --- what's\\nactually in the problematic line, at least on my Mac, is\\n\\nfe80::1%lo0 localhost\\n\\nwhich is not quite the same address I wrote above. It's definitely\\na link-local address though, and the binding to %lo0 does seem to\\nbe understood by macOS, at least to the extent that netstat reports\\nit:\\n\\n$ netstat -a -n\\nActive Internet connections (including servers)\\nProto Recv-Q Send-Q Local Address Foreign Address (state) \\ntcp4 0 0 127.0.0.1.5432 *.* LISTEN \\ntcp6 0 0 fe80::1%lo0.5432 *.* LISTEN \\ntcp6 0 0 ::1.5432 *.* LISTEN \\n... unrelated stuff ...\\n\\nSo either the kernel isn't really treating it as link-local on the\\nloopback interface, or the firewall doesn't understand that it's\\nlink-local. Either way, sure seems like an Apple bug.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\",\n \"msg_date\": \"Thu, 26 Jul 2018 17:45:17 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [HACKERS] Getting rid of \\\"accept incoming network connections\\\"\\n prompts on OS X\"\n },\n {\n \"msg_contents\": \"On 26/07/2018 23:45, Tom Lane wrote:\\n> This came up again today, and I've confirmed that the issue still exists\\n> in current macOS. Did you get any response to your bug report, and if\\n> so what did they say?\\n\\nThere hasn't been any response to the radar. I think our analysis is\\ncorrect, it's an OS bug.\\n\\n-- \\nPeter Eisentraut http://www.2ndQuadrant.com/\\nPostgreSQL Development, 24x7 Support, Remote DBA, Training & Services\\n\\n\",\n \"msg_date\": \"Mon, 30 Jul 2018 21:09:06 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [HACKERS] Getting rid of \\\"accept incoming network connections\\\"\\n prompts on OS X\"\n }\n]","isConversational":false}}},{"rowIdx":39696,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\n\n\nI am looking to patch the LDAP Service Name feature and would like some feedback prior to doing so. In general, while I personally view this as a bug fix, it could also easily be considered to simply be an enhancement to current functionality and therefore the below is written as a proposed enhancement.\n\n\n\nNote: This will also potentially serve as my first patch submission, so this will be a \"smoke test\" of my direct interaction with the mailing list as a Goldman Sachs Employee for proposal/feedback as well as following up with a potential patch/participation in commit processes, etc. As such my replies and related code submissions may be delayed for processes to happen on my end, with my associated apologies in advance. I hope to have this process completed for contribution to the 9.5 release.\n\n\n\n################## Currently:\n\n\n* Service names are provided as an available abstraction mechanism for a host:port (and user, dbname, ssl requirements, etc.) information for connection parameters via libpq.\n\n* If service names are used, they must each individually have an entry in pg_service.conf.\n\n* In a service name entry, host:port entries, hosts can either be directly provided or their values may be referenced via an LDAP location that provides equivalent name/value pairs for connection information.\n\n\n\nThe service name feature facilitates the use of dynamic/managed host entries or allows for active connection configuration management with the goal in either case of separating such overhead from application or other process configuration.\n\n\n\nTherefore:\n\n* Dynamic DNS entries require a fixed port or active management of the file (DNS resolution can happen elsewhere)\n\n* Non-Dynamic DNS entries require active management of pg_service.conf, but may still be desirable for configuration management reasons - version control?\n\n* LDAP entries are largely static, but do require active management of pg_service.conf if new data servers are to be added (or old ones decommissioned)\n\n\n\nIn all cases described above, a newly provisioned data server requires a new service name entry, which requires the file to be actively managed. This mitigates the value of using LDAP or Dynamic DNS entries (with fixed ports in the DNS case) since an active distribution mechanism must still be in place, especially for deployments of many servers and perhaps this useful mechanism is used less often as a result.\n\n\n\n################## Proposal:\n\n\n\n* Have a wildcard entry that allows string substitution of the service name into either the host field or an LDAP record.\n\n\n\nI do not want to over complicate the existing configuration file or parsing logic as it is pretty lightweight currently. As such, I am looking for feedback related to if it would be acceptable to have a \"wildcard\" entry for service names as well as what that wildcard should be. I do *not* propose that complex string substitution, including regular expressions, etc. be utilized for service name matching at this time.\n\n\n\nPer 31.16 in the 9.4 Documentation, a valid service entry is of the following format:\n\n\n\n # comment\n\n [mydb]\n\n host=somehost.domain.com\n\n port=5433\n\n user=admin\n\n\n\nWould specifying a special value for the service name, perhaps [%], be an acceptable implementation of this enhancement/fix to my above concerns?\n\n\n\nExample:\n\n\n\n # comment\n\n [%]\n\n host=%.domain.com\n\n port=5433\n\n user=admin\n\n\n\n...or more interestingly, per 31.17:\n\n\n\n # comment\n\n [%]\n\n ldap=ldap://ldap.mycompany.com/dc=mycompany,dc=com?uniqueMember?one?(cn=%)\n\n\n\nThe location of [%] and therefore order of the service names would still be honored. I believe this allows for useful functionality:\n\n* LDAP wildcard listed first - libpq can always try an LDAP lookup and proceed with further service names if one is not found in LDAP location\n\n* LDAP wildcard listed last - only look up LDAP entries if a service name doesn't match prior service names\n\n* LDAP wildcard in the middle - combination of first 2 behaviors\n\n* DNS wildcard listed last - try connecting with service name in well-formed dns-based host entry (this likely can't be first as it would always match)\n\n* Combination of the LDAP/keyword = value entries per 31.17:\n\n\n\n From the existing documentation: \"Processing of pg_service.conf is terminated after a successful LDAP lookup, but is continued if the LDAP server cannot be contacted. This is to provide a fallback with further LDAP URL lines that point to different LDAP servers, classical keyword = value pairs, or default connection options. If you would rather get an error message in this case, add a syntactically incorrect line after the LDAP URL.\"\n\n\n\nThe \"%\" in the ldap, host entry would replace with the service name in use - also possible to make it less likely to match a future valid \"%\" by providing it as [%] in the string, though I don't know why a % would reasonably show up in an ldap or host record.\n\n\n\nI have not made the patch yet, but it looks like this change is isolated to the parseServiceFile subroutine, currently starting on line 3867 in https://github.com/postgres/postgres/blob/master/src/interfaces/libpq/fe-connect.c (and associated documentation).\n\n\n\n################## Scope:\n\n\n\nNote that this discussion is scoped only for LDAP directory services and not for LDAP authentication/authorization. Again, I do *not* propose that complex string substitution, including regular expressions, etc. be utilized for service name matching at this time. Just looking for a simple [%] entry, which could be extended for flexibility in a consistent way, if people find this general case useful.\n\n\n\nI am looking forward to feedback and our first interaction on the distribution list.\n\n\n\nThanks,\n\n\n\nBryan\n\nHi, I am looking to patch the LDAP Service Name feature and would like some feedback prior to doing so. In general, while I personally view this as a bug fix, it could also easily be considered to simply be an enhancement to current functionality and therefore the below is written as a proposed enhancement. Note: This will also potentially serve as my first patch submission, so this will be a \"smoke test\" of my direct interaction with the mailing list as a Goldman Sachs Employee for proposal/feedback as well as following up with a potential patch/participation in commit processes, etc. As such my replies and related code submissions may be delayed for processes to happen on my end, with my associated apologies in advance. I hope to have this process completed for contribution to the 9.5 release. ################## Currently:* Service names are provided as an available abstraction mechanism for a host:port (and user, dbname, ssl requirements, etc.) information for connection parameters via libpq.* If service names are used, they must each individually have an entry in pg_service.conf.* In a service name entry, host:port entries, hosts can either be directly provided or their values may be referenced via an LDAP location that provides equivalent name/value pairs for connection information. The service name feature facilitates the use of dynamic/managed host entries or allows for active connection configuration management with the goal in either case of separating such overhead from application or other process configuration. Therefore:* Dynamic DNS entries require a fixed port or active management of the file (DNS resolution can happen elsewhere)* Non-Dynamic DNS entries require active management of pg_service.conf, but may still be desirable for configuration management reasons - version control?* LDAP entries are largely static, but do require active management of pg_service.conf if new data servers are to be added (or old ones decommissioned) In all cases described above, a newly provisioned data server requires a new service name entry, which requires the file to be actively managed. This mitigates the value of using LDAP or Dynamic DNS entries (with fixed ports in the DNS case) since an active distribution mechanism must still be in place, especially for deployments of many servers and perhaps this useful mechanism is used less often as a result. ################## Proposal: * Have a wildcard entry that allows string substitution of the service name into either the host field or an LDAP record. I do not want to over complicate the existing configuration file or parsing logic as it is pretty lightweight currently. As such, I am looking for feedback related to if it would be acceptable to have a \"wildcard\" entry for service names as well as what that wildcard should be. I do *not* propose that complex string substitution, including regular expressions, etc. be utilized for service name matching at this time. Per 31.16 in the 9.4 Documentation, a valid service entry is of the following format: # comment [mydb] host=somehost.domain.com port=5433 user=admin Would specifying a special value for the service name, perhaps [%], be an acceptable implementation of this enhancement/fix to my above concerns? Example: # comment [%] host=%.domain.com port=5433 user=admin ...or more interestingly, per 31.17: # comment [%] ldap=ldap://ldap.mycompany.com/dc=mycompany,dc=com?uniqueMember?one?(cn=%) The location of [%] and therefore order of the service names would still be honored. I believe this allows for useful functionality:* LDAP wildcard listed first - libpq can always try an LDAP lookup and proceed with further service names if one is not found in LDAP location* LDAP wildcard listed last - only look up LDAP entries if a service name doesn’t match prior service names* LDAP wildcard in the middle - combination of first 2 behaviors* DNS wildcard listed last - try connecting with service name in well-formed dns-based host entry (this likely can't be first as it would always match)* Combination of the LDAP/keyword = value entries per 31.17: From the existing documentation: \"Processing of pg_service.conf is terminated after a successful LDAP lookup, but is continued if the LDAP server cannot be contacted. This is to provide a fallback with further LDAP URL lines that point to different LDAP servers, classical keyword = value pairs, or default connection options. If you would rather get an error message in this case, add a syntactically incorrect line after the LDAP URL.\" The \"%\" in the ldap, host entry would replace with the service name in use - also possible to make it less likely to match a future valid \"%\" by providing it as [%] in the string, though I don’t know why a % would reasonably show up in an ldap or host record. I have not made the patch yet, but it looks like this change is isolated to the parseServiceFile subroutine, currently starting on line 3867 in https://github.com/postgres/postgres/blob/master/src/interfaces/libpq/fe-connect.c (and associated documentation). ################## Scope: Note that this discussion is scoped only for LDAP directory services and not for LDAP authentication/authorization. Again, I do *not* propose that complex string substitution, including regular expressions, etc. be utilized for service name matching at this time. Just looking for a simple [%] entry, which could be extended for flexibility in a consistent way, if people find this general case useful. I am looking forward to feedback and our first interaction on the distribution list. Thanks, Bryan","msg_date":"Tue, 21 Oct 2014 14:51:34 -0400","msg_from":"\"Doyle, Bryan\" <Bryan.Doyle@gs.com>","msg_from_op":true,"msg_subject":"PostgreSQL Service Name Enhancement - Wildcard support for LDAP/DNS\n lookup"},{"msg_contents":"\"Doyle, Bryan\" <Bryan.Doyle@gs.com> writes:\n> Would specifying a special value for the service name, perhaps [%], be an acceptable implementation of this enhancement/fix to my above concerns?\n\n> Example:\n> # comment\n> [%]\n> host=%.domain.com\n> port=5433\n> user=admin\n\nThis doesn't seem like a terribly good idea, because such an entry would\ncapture *any* service name whatsoever. And, since we check service names\nbefore other possibilities such as host/database names, the entry would\nthen proceed to capture every possible connection request.\n\nI follow what you're trying to do, but it needs to be a more constrained\nsyntax. One possibility is to insist that the wildcard be only a part\nof the name string, eg\n\n [myservers-%]\n host=%.domain.com\n port=5433\n user=admin\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 15:15:48 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: PostgreSQL Service Name Enhancement - Wildcard support for\n LDAP/DNS lookup"},{"msg_contents":"Tom, \n\nI believe there are two main concerns that you raise, addressed below:\n\nFirst:\n\n> It needs to be a more constrained syntax. \n> One possibility is to insist that the wildcard be only a part\n> of the name string, eg\n>\n> [myservers-%]\n> host=%.domain.com\n> port=5433\n> user=admin\n\n* This counter-proposal is getting closer to the more complex matching requirements that I was attempting to de-scope, but it should certainly be considered. I can see where someone may want to have a different LDAP/DNS domain or something like that through prefix convention, though we would likely want to restrict to one \"%\" symbol in the service definition, yes? I am fine with including this capability in the patch, provided the general [%] case is still supported (see below for expanded reasons).\n\nSecond:\n\n> since we check service names\n> before other possibilities such as host/database names, the entry would\n> then proceed to capture every possible connection request\n\n* I should have explicitly covered the case where no service name is provided at connection time - If a service name is not specified in the connection string/connection parameters at all, I would propose that this wildcard entry not match (even if service names are processed first) and normal processing proceed. As a comparable, the '%' in a like statement doesn't match a NULL after all. I don't think having a blank replacement value would make much sense either. Please inform me if I am not addressing some part of your concern with this mindset.\n\n\nSome additional comments:\n\n* It is in fact is desirable for us (and likely others) to capture all service names in one entry; I expect to utilize it exclusively once implemented. I would like to look up all service name entries from a single LDAP location if a previous entry in the file does not short circuit the match. To do this, I am explicitly looking to use a [%] entry in our service name file - the prefix requirement is not consistent with our environmental requirements. The service name file is a client construct that can be overridden by the caller if they desire, but keep in mind that this type of feature is targeted for controlled/managed environments.\n\n* Adding additional processing logic for 'myserver-%' would only make this more flexible for other use cases and would certainly meet the goals of this proposal, so I am fine including it in scope if the [%] is also allowed per above. When these other wildcard w/ prefix entries can match above the general [%] entry, it could be compelling (see my first email regarding entry order considerations). Those not wanting the [%] could choose to not implement it and stick to something closer to the prefix approach you have in mind.\n\n\nSummary of Open Questions:\n\n* (From Above) For prefix wildcards, OK to restrict to one % replacement?\n\n* Do the above points address initial concerns regarding service names being processed before host/db names?\n\n* If both prefix/non-prefix are allowed, what should be the behavior for cases where [prefix-%] matches and fails to connect/lookup and then [%] is also located further down?\n** Without additional discussion, I would assume that it would attempt a connection for consistency. Again, people can choose to not use both features together. Another option would be to somehow introduce a \"stop processing\" flag in the service entries on connection/lookup failures, which may be generally useful even when wild cards are not in use.\n\n\nThanks for the reply and your thoughts on this proposal so far. I am looking forward to the continued conversation.\n\nBryan\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 25 Oct 2014 22:51:50 -0400","msg_from":"\"Doyle, Bryan\" <Bryan.Doyle@gs.com>","msg_from_op":true,"msg_subject":"Re: PostgreSQL Service Name Enhancement - Wildcard\n support for LDAP/DNS lookup"},{"msg_contents":"I have suggested a similar feature before and met with little enthusiasm:\r\nhttp://www.postgresql.org/message-id/D960CB61B694CF459DCFB4B0128514C2F3442B@exadv11.host.magwien.gv.at\r\n\r\nI still think it would be a nice feature and would make pg_service.conf\r\nmore useful than it is now.\r\n\r\nYours,\r\nLaurenz Albe\r\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 29 Oct 2014 11:49:35 +0000","msg_from":"Albe Laurenz <laurenz.albe@wien.gv.at>","msg_from_op":false,"msg_subject":"Re: PostgreSQL Service Name Enhancement - Wildcard\n support for LDAP/DNS lookup"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\n\\n\\nI am looking to patch the LDAP Service Name feature and would like some feedback prior to doing so. In general, while I personally view this as a bug fix, it could also easily be considered to simply be an enhancement to current functionality and therefore the below is written as a proposed enhancement.\\n\\n\\n\\nNote: This will also potentially serve as my first patch submission, so this will be a \\\"smoke test\\\" of my direct interaction with the mailing list as a Goldman Sachs Employee for proposal/feedback as well as following up with a potential patch/participation in commit processes, etc. As such my replies and related code submissions may be delayed for processes to happen on my end, with my associated apologies in advance. I hope to have this process completed for contribution to the 9.5 release.\\n\\n\\n\\n################## Currently:\\n\\n\\n* Service names are provided as an available abstraction mechanism for a host:port (and user, dbname, ssl requirements, etc.) information for connection parameters via libpq.\\n\\n* If service names are used, they must each individually have an entry in pg_service.conf.\\n\\n* In a service name entry, host:port entries, hosts can either be directly provided or their values may be referenced via an LDAP location that provides equivalent name/value pairs for connection information.\\n\\n\\n\\nThe service name feature facilitates the use of dynamic/managed host entries or allows for active connection configuration management with the goal in either case of separating such overhead from application or other process configuration.\\n\\n\\n\\nTherefore:\\n\\n* Dynamic DNS entries require a fixed port or active management of the file (DNS resolution can happen elsewhere)\\n\\n* Non-Dynamic DNS entries require active management of pg_service.conf, but may still be desirable for configuration management reasons - version control?\\n\\n* LDAP entries are largely static, but do require active management of pg_service.conf if new data servers are to be added (or old ones decommissioned)\\n\\n\\n\\nIn all cases described above, a newly provisioned data server requires a new service name entry, which requires the file to be actively managed. This mitigates the value of using LDAP or Dynamic DNS entries (with fixed ports in the DNS case) since an active distribution mechanism must still be in place, especially for deployments of many servers and perhaps this useful mechanism is used less often as a result.\\n\\n\\n\\n################## Proposal:\\n\\n\\n\\n* Have a wildcard entry that allows string substitution of the service name into either the host field or an LDAP record.\\n\\n\\n\\nI do not want to over complicate the existing configuration file or parsing logic as it is pretty lightweight currently. As such, I am looking for feedback related to if it would be acceptable to have a \\\"wildcard\\\" entry for service names as well as what that wildcard should be. I do *not* propose that complex string substitution, including regular expressions, etc. be utilized for service name matching at this time.\\n\\n\\n\\nPer 31.16 in the 9.4 Documentation, a valid service entry is of the following format:\\n\\n\\n\\n # comment\\n\\n [mydb]\\n\\n host=somehost.domain.com\\n\\n port=5433\\n\\n user=admin\\n\\n\\n\\nWould specifying a special value for the service name, perhaps [%], be an acceptable implementation of this enhancement/fix to my above concerns?\\n\\n\\n\\nExample:\\n\\n\\n\\n # comment\\n\\n [%]\\n\\n host=%.domain.com\\n\\n port=5433\\n\\n user=admin\\n\\n\\n\\n...or more interestingly, per 31.17:\\n\\n\\n\\n # comment\\n\\n [%]\\n\\n ldap=ldap://ldap.mycompany.com/dc=mycompany,dc=com?uniqueMember?one?(cn=%)\\n\\n\\n\\nThe location of [%] and therefore order of the service names would still be honored. I believe this allows for useful functionality:\\n\\n* LDAP wildcard listed first - libpq can always try an LDAP lookup and proceed with further service names if one is not found in LDAP location\\n\\n* LDAP wildcard listed last - only look up LDAP entries if a service name doesn't match prior service names\\n\\n* LDAP wildcard in the middle - combination of first 2 behaviors\\n\\n* DNS wildcard listed last - try connecting with service name in well-formed dns-based host entry (this likely can't be first as it would always match)\\n\\n* Combination of the LDAP/keyword = value entries per 31.17:\\n\\n\\n\\n From the existing documentation: \\\"Processing of pg_service.conf is terminated after a successful LDAP lookup, but is continued if the LDAP server cannot be contacted. This is to provide a fallback with further LDAP URL lines that point to different LDAP servers, classical keyword = value pairs, or default connection options. If you would rather get an error message in this case, add a syntactically incorrect line after the LDAP URL.\\\"\\n\\n\\n\\nThe \\\"%\\\" in the ldap, host entry would replace with the service name in use - also possible to make it less likely to match a future valid \\\"%\\\" by providing it as [%] in the string, though I don't know why a % would reasonably show up in an ldap or host record.\\n\\n\\n\\nI have not made the patch yet, but it looks like this change is isolated to the parseServiceFile subroutine, currently starting on line 3867 in https://github.com/postgres/postgres/blob/master/src/interfaces/libpq/fe-connect.c (and associated documentation).\\n\\n\\n\\n################## Scope:\\n\\n\\n\\nNote that this discussion is scoped only for LDAP directory services and not for LDAP authentication/authorization. Again, I do *not* propose that complex string substitution, including regular expressions, etc. be utilized for service name matching at this time. Just looking for a simple [%] entry, which could be extended for flexibility in a consistent way, if people find this general case useful.\\n\\n\\n\\nI am looking forward to feedback and our first interaction on the distribution list.\\n\\n\\n\\nThanks,\\n\\n\\n\\nBryan\\n\\nHi, I am looking to patch the LDAP Service Name feature and would like some feedback prior to doing so. In general, while I personally view this as a bug fix, it could also easily be considered to simply be an enhancement to current functionality and therefore the below is written as a proposed enhancement. Note: This will also potentially serve as my first patch submission, so this will be a \\\"smoke test\\\" of my direct interaction with the mailing list as a Goldman Sachs Employee for proposal/feedback as well as following up with a potential patch/participation in commit processes, etc. As such my replies and related code submissions may be delayed for processes to happen on my end, with my associated apologies in advance. I hope to have this process completed for contribution to the 9.5 release. ################## Currently:* Service names are provided as an available abstraction mechanism for a host:port (and user, dbname, ssl requirements, etc.) information for connection parameters via libpq.* If service names are used, they must each individually have an entry in pg_service.conf.* In a service name entry, host:port entries, hosts can either be directly provided or their values may be referenced via an LDAP location that provides equivalent name/value pairs for connection information. The service name feature facilitates the use of dynamic/managed host entries or allows for active connection configuration management with the goal in either case of separating such overhead from application or other process configuration. Therefore:* Dynamic DNS entries require a fixed port or active management of the file (DNS resolution can happen elsewhere)* Non-Dynamic DNS entries require active management of pg_service.conf, but may still be desirable for configuration management reasons - version control?* LDAP entries are largely static, but do require active management of pg_service.conf if new data servers are to be added (or old ones decommissioned) In all cases described above, a newly provisioned data server requires a new service name entry, which requires the file to be actively managed. This mitigates the value of using LDAP or Dynamic DNS entries (with fixed ports in the DNS case) since an active distribution mechanism must still be in place, especially for deployments of many servers and perhaps this useful mechanism is used less often as a result. ################## Proposal: * Have a wildcard entry that allows string substitution of the service name into either the host field or an LDAP record. I do not want to over complicate the existing configuration file or parsing logic as it is pretty lightweight currently. As such, I am looking for feedback related to if it would be acceptable to have a \\\"wildcard\\\" entry for service names as well as what that wildcard should be. I do *not* propose that complex string substitution, including regular expressions, etc. be utilized for service name matching at this time. Per 31.16 in the 9.4 Documentation, a valid service entry is of the following format: # comment [mydb] host=somehost.domain.com port=5433 user=admin Would specifying a special value for the service name, perhaps [%], be an acceptable implementation of this enhancement/fix to my above concerns? Example: # comment [%] host=%.domain.com port=5433 user=admin ...or more interestingly, per 31.17: # comment [%] ldap=ldap://ldap.mycompany.com/dc=mycompany,dc=com?uniqueMember?one?(cn=%) The location of [%] and therefore order of the service names would still be honored. I believe this allows for useful functionality:* LDAP wildcard listed first - libpq can always try an LDAP lookup and proceed with further service names if one is not found in LDAP location* LDAP wildcard listed last - only look up LDAP entries if a service name doesn’t match prior service names* LDAP wildcard in the middle - combination of first 2 behaviors* DNS wildcard listed last - try connecting with service name in well-formed dns-based host entry (this likely can't be first as it would always match)* Combination of the LDAP/keyword = value entries per 31.17: From the existing documentation: \\\"Processing of pg_service.conf is terminated after a successful LDAP lookup, but is continued if the LDAP server cannot be contacted. This is to provide a fallback with further LDAP URL lines that point to different LDAP servers, classical keyword = value pairs, or default connection options. If you would rather get an error message in this case, add a syntactically incorrect line after the LDAP URL.\\\" The \\\"%\\\" in the ldap, host entry would replace with the service name in use - also possible to make it less likely to match a future valid \\\"%\\\" by providing it as [%] in the string, though I don’t know why a % would reasonably show up in an ldap or host record. I have not made the patch yet, but it looks like this change is isolated to the parseServiceFile subroutine, currently starting on line 3867 in https://github.com/postgres/postgres/blob/master/src/interfaces/libpq/fe-connect.c (and associated documentation). ################## Scope: Note that this discussion is scoped only for LDAP directory services and not for LDAP authentication/authorization. Again, I do *not* propose that complex string substitution, including regular expressions, etc. be utilized for service name matching at this time. Just looking for a simple [%] entry, which could be extended for flexibility in a consistent way, if people find this general case useful. I am looking forward to feedback and our first interaction on the distribution list. Thanks, Bryan\",\n \"msg_date\": \"Tue, 21 Oct 2014 14:51:34 -0400\",\n \"msg_from\": \"\\\"Doyle, Bryan\\\" <Bryan.Doyle@gs.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"PostgreSQL Service Name Enhancement - Wildcard support for LDAP/DNS\\n lookup\"\n },\n {\n \"msg_contents\": \"\\\"Doyle, Bryan\\\" <Bryan.Doyle@gs.com> writes:\\n> Would specifying a special value for the service name, perhaps [%], be an acceptable implementation of this enhancement/fix to my above concerns?\\n\\n> Example:\\n> # comment\\n> [%]\\n> host=%.domain.com\\n> port=5433\\n> user=admin\\n\\nThis doesn't seem like a terribly good idea, because such an entry would\\ncapture *any* service name whatsoever. And, since we check service names\\nbefore other possibilities such as host/database names, the entry would\\nthen proceed to capture every possible connection request.\\n\\nI follow what you're trying to do, but it needs to be a more constrained\\nsyntax. One possibility is to insist that the wildcard be only a part\\nof the name string, eg\\n\\n [myservers-%]\\n host=%.domain.com\\n port=5433\\n user=admin\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 15:15:48 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PostgreSQL Service Name Enhancement - Wildcard support for\\n LDAP/DNS lookup\"\n },\n {\n \"msg_contents\": \"Tom, \\n\\nI believe there are two main concerns that you raise, addressed below:\\n\\nFirst:\\n\\n> It needs to be a more constrained syntax. \\n> One possibility is to insist that the wildcard be only a part\\n> of the name string, eg\\n>\\n> [myservers-%]\\n> host=%.domain.com\\n> port=5433\\n> user=admin\\n\\n* This counter-proposal is getting closer to the more complex matching requirements that I was attempting to de-scope, but it should certainly be considered. I can see where someone may want to have a different LDAP/DNS domain or something like that through prefix convention, though we would likely want to restrict to one \\\"%\\\" symbol in the service definition, yes? I am fine with including this capability in the patch, provided the general [%] case is still supported (see below for expanded reasons).\\n\\nSecond:\\n\\n> since we check service names\\n> before other possibilities such as host/database names, the entry would\\n> then proceed to capture every possible connection request\\n\\n* I should have explicitly covered the case where no service name is provided at connection time - If a service name is not specified in the connection string/connection parameters at all, I would propose that this wildcard entry not match (even if service names are processed first) and normal processing proceed. As a comparable, the '%' in a like statement doesn't match a NULL after all. I don't think having a blank replacement value would make much sense either. Please inform me if I am not addressing some part of your concern with this mindset.\\n\\n\\nSome additional comments:\\n\\n* It is in fact is desirable for us (and likely others) to capture all service names in one entry; I expect to utilize it exclusively once implemented. I would like to look up all service name entries from a single LDAP location if a previous entry in the file does not short circuit the match. To do this, I am explicitly looking to use a [%] entry in our service name file - the prefix requirement is not consistent with our environmental requirements. The service name file is a client construct that can be overridden by the caller if they desire, but keep in mind that this type of feature is targeted for controlled/managed environments.\\n\\n* Adding additional processing logic for 'myserver-%' would only make this more flexible for other use cases and would certainly meet the goals of this proposal, so I am fine including it in scope if the [%] is also allowed per above. When these other wildcard w/ prefix entries can match above the general [%] entry, it could be compelling (see my first email regarding entry order considerations). Those not wanting the [%] could choose to not implement it and stick to something closer to the prefix approach you have in mind.\\n\\n\\nSummary of Open Questions:\\n\\n* (From Above) For prefix wildcards, OK to restrict to one % replacement?\\n\\n* Do the above points address initial concerns regarding service names being processed before host/db names?\\n\\n* If both prefix/non-prefix are allowed, what should be the behavior for cases where [prefix-%] matches and fails to connect/lookup and then [%] is also located further down?\\n** Without additional discussion, I would assume that it would attempt a connection for consistency. Again, people can choose to not use both features together. Another option would be to somehow introduce a \\\"stop processing\\\" flag in the service entries on connection/lookup failures, which may be generally useful even when wild cards are not in use.\\n\\n\\nThanks for the reply and your thoughts on this proposal so far. I am looking forward to the continued conversation.\\n\\nBryan\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 25 Oct 2014 22:51:50 -0400\",\n \"msg_from\": \"\\\"Doyle, Bryan\\\" <Bryan.Doyle@gs.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PostgreSQL Service Name Enhancement - Wildcard\\n support for LDAP/DNS lookup\"\n },\n {\n \"msg_contents\": \"I have suggested a similar feature before and met with little enthusiasm:\\r\\nhttp://www.postgresql.org/message-id/D960CB61B694CF459DCFB4B0128514C2F3442B@exadv11.host.magwien.gv.at\\r\\n\\r\\nI still think it would be a nice feature and would make pg_service.conf\\r\\nmore useful than it is now.\\r\\n\\r\\nYours,\\r\\nLaurenz Albe\\r\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 29 Oct 2014 11:49:35 +0000\",\n \"msg_from\": \"Albe Laurenz <laurenz.albe@wien.gv.at>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PostgreSQL Service Name Enhancement - Wildcard\\n support for LDAP/DNS lookup\"\n }\n]","isConversational":false}}},{"rowIdx":39697,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"expected/sequence_1.out hasn't been updated at least since\n\ncommit d90ced8bb22194cbb45f58beb0961251103aeff5\nDate: Thu Oct 3 16:17:18 2013 -0400\n\nand nobody appears to have complained.\n\nCan it be removed?\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 15:46:33 -0400","msg_from":"Peter Eisentraut <peter_e@gmx.net>","msg_from_op":true,"msg_subject":"expected/sequence_1.out obsolete?"},{"msg_contents":"Peter Eisentraut <peter_e@gmx.net> writes:\n> expected/sequence_1.out hasn't been updated at least since\n> commit d90ced8bb22194cbb45f58beb0961251103aeff5\n> Date: Thu Oct 3 16:17:18 2013 -0400\n> and nobody appears to have complained.\n\n> Can it be removed?\n\nNo, it needs to be fixed. The alternate output from \"select * from\nfoo_seq\" will appear occasionally (from memory, if a checkpoint starts\nor perhaps completes during the test).\n\nWe might have noticed this already if the buildfarm weren't so noisy\nlately :-(\n\nAn alternative answer would be to remove log_cnt from the set of\ncolumns displayed, but I think that would be taking away part of the\npoint of that particular test.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 18:10:54 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: expected/sequence_1.out obsolete?"},{"msg_contents":"On Wed, Oct 22, 2014 at 4:46 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\n\n> expected/sequence_1.out hasn't been updated at least since\n>\n> commit d90ced8bb22194cbb45f58beb0961251103aeff5\n> Date: Thu Oct 3 16:17:18 2013 -0400\n>\n> and nobody appears to have complained.\n>\n> Can it be removed?\n>\nI don't think that this is a good idea, have a look at bb3f839 explaining\nthat this alternate output may happen when a checkpoint kicks in. Simple\npatch is attached.\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Wed, 22 Oct 2014 07:15:32 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: expected/sequence_1.out obsolete?"},{"msg_contents":"Michael Paquier <michael.paquier@gmail.com> writes:\n> I don't think that this is a good idea, have a look at bb3f839 explaining\n> that this alternate output may happen when a checkpoint kicks in. Simple\n> patch is attached.\n\nPushed, thanks. (The 9.4 branch was broken too, but differently :-()\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 21 Oct 2014 18:27:20 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: expected/sequence_1.out obsolete?"}],"string":"[\n {\n \"msg_contents\": \"expected/sequence_1.out hasn't been updated at least since\\n\\ncommit d90ced8bb22194cbb45f58beb0961251103aeff5\\nDate: Thu Oct 3 16:17:18 2013 -0400\\n\\nand nobody appears to have complained.\\n\\nCan it be removed?\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 15:46:33 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter_e@gmx.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"expected/sequence_1.out obsolete?\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter_e@gmx.net> writes:\\n> expected/sequence_1.out hasn't been updated at least since\\n> commit d90ced8bb22194cbb45f58beb0961251103aeff5\\n> Date: Thu Oct 3 16:17:18 2013 -0400\\n> and nobody appears to have complained.\\n\\n> Can it be removed?\\n\\nNo, it needs to be fixed. The alternate output from \\\"select * from\\nfoo_seq\\\" will appear occasionally (from memory, if a checkpoint starts\\nor perhaps completes during the test).\\n\\nWe might have noticed this already if the buildfarm weren't so noisy\\nlately :-(\\n\\nAn alternative answer would be to remove log_cnt from the set of\\ncolumns displayed, but I think that would be taking away part of the\\npoint of that particular test.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 18:10:54 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: expected/sequence_1.out obsolete?\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 22, 2014 at 4:46 AM, Peter Eisentraut <peter_e@gmx.net> wrote:\\n\\n> expected/sequence_1.out hasn't been updated at least since\\n>\\n> commit d90ced8bb22194cbb45f58beb0961251103aeff5\\n> Date: Thu Oct 3 16:17:18 2013 -0400\\n>\\n> and nobody appears to have complained.\\n>\\n> Can it be removed?\\n>\\nI don't think that this is a good idea, have a look at bb3f839 explaining\\nthat this alternate output may happen when a checkpoint kicks in. Simple\\npatch is attached.\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Wed, 22 Oct 2014 07:15:32 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: expected/sequence_1.out obsolete?\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael.paquier@gmail.com> writes:\\n> I don't think that this is a good idea, have a look at bb3f839 explaining\\n> that this alternate output may happen when a checkpoint kicks in. Simple\\n> patch is attached.\\n\\nPushed, thanks. (The 9.4 branch was broken too, but differently :-()\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 21 Oct 2014 18:27:20 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: expected/sequence_1.out obsolete?\"\n }\n]","isConversational":false}}},{"rowIdx":39698,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"In heap_prune_chain():\n\n\ttup.t_tableOid = RelationGetRelid(relation);\n\n\trootlp = PageGetItemId(dp, rootoffnum);\n\n\t/*\n\t * If it's a heap-only tuple, then it is not the start of a HOT chain.\n\t */\n\tif (ItemIdIsNormal(rootlp))\n\t{\n\t\thtup = (HeapTupleHeader) PageGetItem(dp, rootlp);\n\n\t\ttup.t_data = htup;\n\t\ttup.t_len = ItemIdGetLength(rootlp);\n\t\ttup.t_tableOid = RelationGetRelid(relation);\n\nAFAICT the second case of setting tup.t_tableOid is pointless. Attached patch removes it. Passes make check.\n-- \nJim Nasby, Data Architect, Blue Treble Consulting\nData in Trouble? Get it in Treble! http://BlueTreble.com\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Tue, 21 Oct 2014 17:51:08 -0500","msg_from":"Jim Nasby <Jim.Nasby@BlueTreble.com>","msg_from_op":true,"msg_subject":"Spurious set in heap_prune_chain()"}],"string":"[\n {\n \"msg_contents\": \"In heap_prune_chain():\\n\\n\\ttup.t_tableOid = RelationGetRelid(relation);\\n\\n\\trootlp = PageGetItemId(dp, rootoffnum);\\n\\n\\t/*\\n\\t * If it's a heap-only tuple, then it is not the start of a HOT chain.\\n\\t */\\n\\tif (ItemIdIsNormal(rootlp))\\n\\t{\\n\\t\\thtup = (HeapTupleHeader) PageGetItem(dp, rootlp);\\n\\n\\t\\ttup.t_data = htup;\\n\\t\\ttup.t_len = ItemIdGetLength(rootlp);\\n\\t\\ttup.t_tableOid = RelationGetRelid(relation);\\n\\nAFAICT the second case of setting tup.t_tableOid is pointless. Attached patch removes it. Passes make check.\\n-- \\nJim Nasby, Data Architect, Blue Treble Consulting\\nData in Trouble? Get it in Treble! http://BlueTreble.com\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Tue, 21 Oct 2014 17:51:08 -0500\",\n \"msg_from\": \"Jim Nasby <Jim.Nasby@BlueTreble.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Spurious set in heap_prune_chain()\"\n }\n]","isConversational":false}}},{"rowIdx":39699,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI have been thinking about why we need to grab an AccessExclusiveLock on \nthe table with the PK when we add a foreign key. Adding new tables with \nforeign keys to old ones is common so it would be really nice if the \nlock strength could be reduced.\n\nA comment in the code claims that we need to lock so no rows are deleted \nunder us and that adding a trigger will lock in AccessExclusive anyway. \nBut with MVCC catalog access and the removal of SnapshotNow I do not see \nwhy adding a trigger would require an exclusive lock. Just locking for \ndata changes should be enough.\n\nLooking at the code the only see the duplicate name check use the fact \nthat we have grabbed an exclusive lock and that should work anyway due \nto the unique constraint, but since I am pretty new to the code base I \ncould be missing something obvious. I have attached a proof of concept \npatch which reduces the lock strength to ShareLock.\n\nWhat do you say? Am I missing something?\n\nMy gut feel also says that if we know it is a RI trigger we are adding \nthen AccessShare should be enough for the PK table, since we could rely \non row locks to prevent rows from being deleted. It would be really nice \nthough if this was possible since this would make it possible to add a \nnew table with foreign keys and data without locking anything more than \nthe referred rows.\n\nAndreas\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Wed, 22 Oct 2014 09:06:52 +0200","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Reducing lock strength of adding foreign keys"},{"msg_contents":"Andreas Karlsson <andreas@proxel.se> writes:\n> I have attached a proof of concept \n> patch which reduces the lock strength to ShareLock.\n\nYou're kidding, right? ShareLock isn't even self-exclusive.\n\n\t\t\tregards, tom lane\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 22 Oct 2014 10:13:58 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"On 10/22/2014 04:13 PM, Tom Lane wrote:\n> Andreas Karlsson <andreas@proxel.se> writes:\n>> I have attached a proof of concept\n>> patch which reduces the lock strength to ShareLock.\n>\n> You're kidding, right? ShareLock isn't even self-exclusive.\n\nWhy would it have to be self-exclusive? As far as I know we only need to \nensure that nobody changes the rows while we add the trigger. Adding \nmultiple triggers concurrently should not pose a problem unless I am \nmissing something (which I probably am).\n\nAndreas\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 23 Oct 2014 09:25:05 +0200","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"On Wed, Oct 22, 2014 at 3:06 AM, Andreas Karlsson <andreas@proxel.se> wrote:\n> I have been thinking about why we need to grab an AccessExclusiveLock on the\n> table with the PK when we add a foreign key. Adding new tables with foreign\n> keys to old ones is common so it would be really nice if the lock strength\n> could be reduced.\n>\n> A comment in the code claims that we need to lock so no rows are deleted\n> under us and that adding a trigger will lock in AccessExclusive anyway. But\n> with MVCC catalog access and the removal of SnapshotNow I do not see why\n> adding a trigger would require an exclusive lock. Just locking for data\n> changes should be enough.\n\nThe use of MVCC catalog access doesn't necessarily mean that adding a\ntrigger doesn't require an AccessExclusive lock. Those changes - if I\ndare to say so myself - solved a complex and important problem, but\nonly one of many problems in this area, and people seem prone to\nthinking that they solved more problems than they in fact did.\n\nI think instead of focusing on foreign keys, we should rewind a bit\nand think about the locking level required to add a trigger. If we\nfigure something out there, then we can consider how it affects\nforeign keys. I went looking for previous discussions of remaining\nhazards and found these postings:\n\nhttp://www.postgresql.org/message-id/CA+TgmoY4GLsXZk0tAO29-LJtcuj0SL1xWCwQ51xb-HFYsgi5RQ@mail.gmail.com\nhttp://www.postgresql.org/message-id/20893.1393892127@sss.pgh.pa.us\nhttp://www.postgresql.org/message-id/20140306224340.GA3551655@tornado.leadboat.com\n\nAs far as triggers are concerned, the issue of skew between the\ntransaction snapshot and what the ruleutils.c snapshots do seems to be\nthe principal issue. Commit e5550d5fec66aa74caad1f79b79826ec64898688\nchanged pg_get_constraintdef() to use an MVCC snapshot rather than a\ncurrent MVCC snapshot; if that change is safe, I am not aware of any\nreason why we couldn't change pg_get_triggerdef() similarly. Barring\nfurther hazards I haven't thought of, I would expect that we could add\na trigger to a relation with only ShareRowExclusiveLock. Anything\nless than ShareRowExclusiveLock would open up strange timing races\naround the firing of triggers by transactions writing the table: they\nmight or might not notice that a trigger had been added before\nend-of-transaction, depending on the timing of cache flushes, which\ncertainly seems no good. But even RowExclusiveLock seems like a large\nimprovement over AccessExclusiveLock.\n\nWhen a constraint trigger - which is used to implement a foreign key -\nis added, there are actually TWO tables involved: the table upon which\nthe trigger will actually fire, and some other table which is\nmentioned in passing in the trigger definition. It's possible that\nthe locking requirements for the secondary table are weaker since I\ndon't think the presence of the trigger actually affects runtime\nbehavior there. However, there's probably little point in try to\nweaken the lock to less than the level required for the main table\nbecause a foreign key involves adding referential integrity triggers\nto both tables.\n\nSo I tentatively propose (and with due regard for the possibility\nothers may see dangers that I've missed) that a reasonable goal would\nbe to lower the lock strength required for both CREATE TRIGGER and ADD\nFOREIGN KEY from AccessExclusiveLock to ShareRowExclusiveLock,\nallowing concurrent SELECT and SELECT FOR SHARE against the tables,\nbut not any actual write operations.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 24 Oct 2014 12:07:42 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"On Fri, Oct 24, 2014 at 12:07:42PM -0400, Robert Haas wrote:\n> I think instead of focusing on foreign keys, we should rewind a bit\n> and think about the locking level required to add a trigger.\n\nAgreed.\n\n> http://www.postgresql.org/message-id/CA+TgmoY4GLsXZk0tAO29-LJtcuj0SL1xWCwQ51xb-HFYsgi5RQ@mail.gmail.com\n> http://www.postgresql.org/message-id/20893.1393892127@sss.pgh.pa.us\n> http://www.postgresql.org/message-id/20140306224340.GA3551655@tornado.leadboat.com\n> \n> As far as triggers are concerned, the issue of skew between the\n> transaction snapshot and what the ruleutils.c snapshots do seems to be\n> the principal issue. Commit e5550d5fec66aa74caad1f79b79826ec64898688\n> changed pg_get_constraintdef() to use an MVCC snapshot rather than a\n> current MVCC snapshot; if that change is safe, I am not aware of any\n> reason why we couldn't change pg_get_triggerdef() similarly.\n\npg_get_triggerdef() is fine as-is with concurrent CREATE TRIGGER. The\npg_get_constraintdef() change arose to ensure a consistent result when\nconcurrent ALTER TABLE VALIDATE CONSTRAINT mutates a constraint definition.\n(Reducing the lock level of DROP TRIGGER or ALTER TRIGGER, however, would\ncreate the analogous problem for pg_get_triggerdef().)\n\n> So I tentatively propose (and with due regard for the possibility\n> others may see dangers that I've missed) that a reasonable goal would\n> be to lower the lock strength required for both CREATE TRIGGER and ADD\n> FOREIGN KEY from AccessExclusiveLock to ShareRowExclusiveLock,\n> allowing concurrent SELECT and SELECT FOR SHARE against the tables,\n> but not any actual write operations.\n\n+1\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 25 Oct 2014 14:00:17 -0400","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":false,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"On 10/24/2014 06:07 PM, Robert Haas wrote:\n> I think instead of focusing on foreign keys, we should rewind a bit\n> and think about the locking level required to add a trigger.\n\nAgreed.\n\n> As far as triggers are concerned, the issue of skew between the\n> transaction snapshot and what the ruleutils.c snapshots do seems to be\n> the principal issue. Commit e5550d5fec66aa74caad1f79b79826ec64898688\n> changed pg_get_constraintdef() to use an MVCC snapshot rather than a\n> current MVCC snapshot; if that change is safe, I am not aware of any\n> reason why we couldn't change pg_get_triggerdef() similarly. Barring\n> further hazards I haven't thought of, I would expect that we could add\n> a trigger to a relation with only ShareRowExclusiveLock.\n\nThanks for the info. This is just the kind of issues I was worrying about.\n\n> Anything\n> less than ShareRowExclusiveLock would open up strange timing races\n> around the firing of triggers by transactions writing the table: they\n> might or might not notice that a trigger had been added before\n> end-of-transaction, depending on the timing of cache flushes, which\n> certainly seems no good. But even RowExclusiveLock seems like a large\n> improvement over AccessExclusiveLock.\n\nWould not ShareLock give the same result, except for also allowing \nconcurrent CREATE INDEX and concurrent other CREATE TRIGGER which does \nnot look dangerous to me?\n\n From a user point of view ShareRowExclusiveLock should be as useful as \nShareLock.\n\n> When a constraint trigger - which is used to implement a foreign key -\n> is added, there are actually TWO tables involved: the table upon which\n> the trigger will actually fire, and some other table which is\n> mentioned in passing in the trigger definition. It's possible that\n> the locking requirements for the secondary table are weaker since I\n> don't think the presence of the trigger actually affects runtime\n> behavior there. However, there's probably little point in try to\n> weaken the lock to less than the level required for the main table\n> because a foreign key involves adding referential integrity triggers\n> to both tables.\n>\n> So I tentatively propose (and with due regard for the possibility\n> others may see dangers that I've missed) that a reasonable goal would\n> be to lower the lock strength required for both CREATE TRIGGER and ADD\n> FOREIGN KEY from AccessExclusiveLock to ShareRowExclusiveLock,\n> allowing concurrent SELECT and SELECT FOR SHARE against the tables,\n> but not any actual write operations.\n\nAgreed.. But I think reducing the lock level of the secondary table is \nmuch more important than doing the same for the primary table due to the \ncase where the secondary table is an existing table which is hit by a \nworkload of long running queries and DML while the primary is a new \ntable which is added now. In my dream world I could add the new table \nwithout any disruption at all of queries using the secondary table, no \nmatter the duration of the transaction adding the table (barring \ninsertion of actual data into the primary table, which would take row \nlocks).\n\nThis is just a dream scenario though, and focusing on triggers is indeed \nthe reasonable goal for 9.5.\n\n-- \nAndreas Karlsson\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 27 Oct 2014 02:48:55 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"There are actually TWO tables involved: the table upon which \nthe trigger will actually fire, and some other table which is \nmentioned in passing in the trigger definition. It's possible that \nthe locking requirements for the secondary table are weaker since I \ndon't think the presence of the trigger actually affects runtime \nbehavior there. However, there's probably little point in try to \nweaken the lock to less than the level required for the main table \nbecause a foreign key involves adding referential integrity triggers \nto both tables. \n\n\n\n\n-----\nGUL\n--\nView this message in context: http://postgresql.1045698.n5.nabble.com/Reducing-lock-strength-of-adding-foreign-keys-tp5823894p5824376.html\nSent from the PostgreSQL - hackers mailing list archive at Nabble.com.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 27 Oct 2014 01:44:23 -0700 (PDT)","msg_from":"adamrose045 <adamrose045@gmail.com>","msg_from_op":false,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"Thanks for weighing in, Noah.\n\nOn Sat, Oct 25, 2014 at 2:00 PM, Noah Misch <noah@leadboat.com> wrote:\n>> http://www.postgresql.org/message-id/CA+TgmoY4GLsXZk0tAO29-LJtcuj0SL1xWCwQ51xb-HFYsgi5RQ@mail.gmail.com\n>> http://www.postgresql.org/message-id/20893.1393892127@sss.pgh.pa.us\n>> http://www.postgresql.org/message-id/20140306224340.GA3551655@tornado.leadboat.com\n>>\n>> As far as triggers are concerned, the issue of skew between the\n>> transaction snapshot and what the ruleutils.c snapshots do seems to be\n>> the principal issue. Commit e5550d5fec66aa74caad1f79b79826ec64898688\n>> changed pg_get_constraintdef() to use an MVCC snapshot rather than a\n>> current MVCC snapshot; if that change is safe, I am not aware of any\n>> reason why we couldn't change pg_get_triggerdef() similarly.\n>\n> pg_get_triggerdef() is fine as-is with concurrent CREATE TRIGGER. The\n> pg_get_constraintdef() change arose to ensure a consistent result when\n> concurrent ALTER TABLE VALIDATE CONSTRAINT mutates a constraint definition.\n> (Reducing the lock level of DROP TRIGGER or ALTER TRIGGER, however, would\n> create the analogous problem for pg_get_triggerdef().)\n\nMaybe so, but I'd favor changing it anyway and getting it over with.\nThe current situation seems to have little to recommend it; moreover,\nit would be nice, if it's possible and safe, to weaken the lock levels\nfor all three of those commands at the same time. Do you see any\nhazards for ALTER or DROP that do not exist for CREATE?\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 27 Oct 2014 08:24:15 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"On Sun, Oct 26, 2014 at 9:48 PM, Andreas Karlsson <andreas@proxel.se> wrote:\n> Agreed.. But I think reducing the lock level of the secondary table is much\n> more important than doing the same for the primary table due to the case\n> where the secondary table is an existing table which is hit by a workload of\n> long running queries and DML while the primary is a new table which is added\n> now. In my dream world I could add the new table without any disruption at\n> all of queries using the secondary table, no matter the duration of the\n> transaction adding the table (barring insertion of actual data into the\n> primary table, which would take row locks).\n\nThat would indeed be nice, but it doesn't seem very practical, because\nthe parent needs triggers, too: if you try to delete a row from the\nparent, or update the key, it's got to go look at the child and see\nwhether there are rows against the old value. Then it's got to either\nupdate those rows, or null out the value, or throw an error.\nRegardless of which of those things it does (which depends on the ON\nDELETE and ON UPDATE settings you choose), it's hard to imagine that\nit would be a good idea for any of those things to start happening in\nthe middle of a transaction or statement.\n\nSo, let's take what we can get. :-)\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 27 Oct 2014 08:35:59 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"On Mon, Oct 27, 2014 at 08:24:15AM -0400, Robert Haas wrote:\n> On Sat, Oct 25, 2014 at 2:00 PM, Noah Misch <noah@leadboat.com> wrote:\n> >> http://www.postgresql.org/message-id/CA+TgmoY4GLsXZk0tAO29-LJtcuj0SL1xWCwQ51xb-HFYsgi5RQ@mail.gmail.com\n> >> http://www.postgresql.org/message-id/20893.1393892127@sss.pgh.pa.us\n> >> http://www.postgresql.org/message-id/20140306224340.GA3551655@tornado.leadboat.com\n> >>\n> >> As far as triggers are concerned, the issue of skew between the\n> >> transaction snapshot and what the ruleutils.c snapshots do seems to be\n> >> the principal issue. Commit e5550d5fec66aa74caad1f79b79826ec64898688\n> >> changed pg_get_constraintdef() to use an MVCC snapshot rather than a\n> >> current MVCC snapshot; if that change is safe, I am not aware of any\n> >> reason why we couldn't change pg_get_triggerdef() similarly.\n> >\n> > pg_get_triggerdef() is fine as-is with concurrent CREATE TRIGGER. The\n> > pg_get_constraintdef() change arose to ensure a consistent result when\n> > concurrent ALTER TABLE VALIDATE CONSTRAINT mutates a constraint definition.\n> > (Reducing the lock level of DROP TRIGGER or ALTER TRIGGER, however, would\n> > create the analogous problem for pg_get_triggerdef().)\n> \n> Maybe so, but I'd favor changing it anyway and getting it over with.\n> The current situation seems to have little to recommend it; moreover,\n> it would be nice, if it's possible and safe, to weaken the lock levels\n> for all three of those commands at the same time. Do you see any\n> hazards for ALTER or DROP that do not exist for CREATE?\n\nALTER TRIGGER is not bad; like you say, change pg_get_triggerdef_worker() the\nway commit e5550d5 changed pg_get_constraintdef_worker(). DROP TRIGGER is\nmore difficult. pg_constraint.tgqual of a dropped trigger may reference other\ndropped objects, which calls for equipping get_rule_expr() to use the\ntransaction snapshot. That implicates quite a bit of ruleutils.c code.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 27 Oct 2014 20:33:56 -0400","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":false,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"On 10/28/2014 01:33 AM, Noah Misch wrote:\n> ALTER TRIGGER is not bad; like you say, change pg_get_triggerdef_worker() the\n> way commit e5550d5 changed pg_get_constraintdef_worker(). DROP TRIGGER is\n> more difficult. pg_constraint.tgqual of a dropped trigger may reference other\n> dropped objects, which calls for equipping get_rule_expr() to use the\n> transaction snapshot. That implicates quite a bit of ruleutils.c code.\n\nI started looking into this again and fixed \npg_get_constraintdef_worker() as suggested.\n\nBut I have no idea how to fix get_rule_expr() since it relies on doing \nlookups in the catcache. Replacing these with uncached lookups sounds \nlike it could cause quite some slowdown. Any ideas?\n\nIndexes should suffer from the same problems since they too have emay \ncontain expressions but they seem to solve this by having a higher lock \nlevel on DROP INDEX, but I do wonder about the CONCURRENTLY case.\n\nBy the way, unless I am mistaken there is currently no protection \nagainst having a concurrent ALTER FUNCTION ... RENAME mess up what is \ndumped in by pg_get_triggerdef().\n\n-- \nAndreas Karlsson\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sat, 13 Dec 2014 11:42:13 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"Hi,\n\nI have attached a patch with the current status of my work on reducing \nthe lock level of trigger and foreign key related DDL.\n\nThis commit reduces the lock level of the following commands from ACCESS \nEXCLUSIVE to SHARE ROW EXCLUSIVE, plus that it does the same for the \nreferred table of the foreign key.\n\nADD TRIGGER\nALTER TRIGGER\nALTER TABLE ... ADD FOREIGN KEY\nALTER TABLE ... DISABLE TRIGGER\nALTER TABLE ... ENABLE TRIGGER\n\nThe patch does currently not reducing the lock level of the following \ntwo commands, but I think it would be really nice to fix those too and \nhere I would like some feedback and ideas.\n\nDROP TRIGGER\nALTER TABLE ... DROP CONSTRAINT -- For foreign keys\n\nForeign keys and triggers are fixed at the same time because foreign \nkeys are implemented with two triggers, one at each of the involved tables.\n\nThe idea behind the patch is that since we started using MVCC snapshots \nwe no longer need to hold an exclusive lock on the table when adding a \ntrigger. Theoretically we only need to lock out writers of the rows of \nthe table (SHARE ROW EXCLUSIVE/SHARE), but as noted by Robert and Noah \njust reducing the lock level will break pg_dump since \npg_get_triggerdef() does not use the current snapshot when reading the \ncatalog.\n\nI have fixed pg_get_triggerdef() to use the snapshot like how \ne5550d5fec66aa74caad1f79b79826ec64898688 fixed pg_get_constraintdef(), \nand this fixes the code for the ALTER TRIGGER case (ADD TRIGGER was \nalready safe). But to fix it for the DROP TRIGGER case we need to also \nmake the dumping of the WHEN clause (which is dumped by get_rule_expr()) \nuse the current snapshot.\n\nget_rule_expr() relies heavily on the catcache which to me does not look \nlike it could easily be (and probably not even should be) made to use \nthe current snapshot. Refactoring ruleutils.c to rely less no the \ncatcache seems like a reasonable thing to do if we want to reduce \nweirdness of how it ignores MVCC but it is quite a bit of work and I \nfear it could give us performance regressions.\n\nDo you have any ideas for how to fix get_rule_expr()? Is this patch \nworthwhile even without reducing the lock levels of the drop commands?\n\nAndreas\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Sat, 13 Dec 2014 21:45:54 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"PATCH: Reducing lock strength of trigger and foreign key DDL"},{"msg_contents":"I'm not sure about the rest of this but...\n\nOn Sat, Dec 13, 2014 at 3:45 PM, Andreas Karlsson <andreas@proxel.se> wrote:\n> Is this patch\n> worthwhile even without reducing the lock levels of the drop commands?\n\nYes! It certainly makes more sense to reduce the lock levels where we\ncan do that relatively easily, and postpone work on related projects\nthat are harder, rather than waiting until it all seems to work before\ndoing anything at all.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 15 Dec 2014 11:24:24 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign key DDL"},{"msg_contents":"On 25 October 2014 19:00, Noah Misch <noah@leadboat.com> wrote:\n\n>> So I tentatively propose (and with due regard for the possibility\n>> others may see dangers that I've missed) that a reasonable goal would\n>> be to lower the lock strength required for both CREATE TRIGGER and ADD\n>> FOREIGN KEY from AccessExclusiveLock to ShareRowExclusiveLock,\n>> allowing concurrent SELECT and SELECT FOR SHARE against the tables,\n>> but not any actual write operations.\n>\n> +1\n\nAll of this is just a replay of the earlier conversations about\nreducing lock levels.\n\nMy original patch did reduce lock levels for CREATE TRIGGER, but we\nstepped back from doing that in 9.4 until we had feedback as to\nwhether the whole idea of reducing lock levels was safe, which Robert,\nTom and others were slightly uncertain about.\n\nIs there anything different here from work in my original patch series?\n\n-- \n Simon Riggs http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 17 Dec 2014 14:41:39 +0000","msg_from":"Simon Riggs <simon@2ndQuadrant.com>","msg_from_op":false,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"On Wed, Dec 17, 2014 at 02:41:39PM +0000, Simon Riggs wrote:\n> Is there anything different here from work in my original patch series?\n\nNot to my knowledge.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 19 Dec 2014 21:26:58 -0500","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":false,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"On 12/17/2014 03:41 PM, Simon Riggs wrote:\n> All of this is just a replay of the earlier conversations about\n> reducing lock levels.\n>\n> My original patch did reduce lock levels for CREATE TRIGGER, but we\n> stepped back from doing that in 9.4 until we had feedback as to\n> whether the whole idea of reducing lock levels was safe, which Robert,\n> Tom and others were slightly uncertain about.\n>\n> Is there anything different here from work in my original patch series?\n\nAs far as I know, no.\n\n-- \nAndreas Karlsson\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 23 Dec 2014 19:59:29 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: Reducing lock strength of adding foreign keys"},{"msg_contents":"On Sun, Dec 14, 2014 at 5:45 AM, Andreas Karlsson <andreas@proxel.se> wrote:\n> get_rule_expr() relies heavily on the catcache which to me does not look\n> like it could easily be (and probably not even should be) made to use the\n> current snapshot. Refactoring ruleutils.c to rely less no the catcache seems\n> like a reasonable thing to do if we want to reduce weirdness of how it\n> ignores MVCC but it is quite a bit of work and I fear it could give us\n> performance regressions.\n> Do you have any ideas for how to fix get_rule_expr()?\nAgreed. There are 20 calls to SearchSysCache in ruleutils.c, let's not\nfocus on that for now though and get things right for this patch.\n\n> Is this patch worthwhile even without reducing the lock levels of the drop commands?\nYes. IMV, it is better to do this work slowly and incrementally.\n\nHere are some comments about this patch:\n1) There is no documentation. Could you update mvcc.sgml for SHARE ROW\nEXCLUSIVE?\n2) Some isolation tests would be welcome, the point of the feature is\nto test if SELECT and SELECT FOR SHARE/UPDATE are allowed while\nrunning one of the command mentioned above.\n3) This patch breaks the isolation test alter-table-1.\nRegards,\n--\nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Wed, 24 Dec 2014 17:27:51 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"Thanks for the review.\n\nHere is a new version of the patch with updated isolation tests and \ndocumentation.\n\n-- \nAndreas Karlsson\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Tue, 13 Jan 2015 09:21:47 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On Tue, Jan 13, 2015 at 5:21 PM, Andreas Karlsson <andreas@proxel.se> wrote:\n> Here is a new version of the patch with updated isolation tests and\n> documentation.\nThanks.\n\nSome more comments:\n- It would be good to have as well tests for ENABLE/DISABLE TRIGGER,\nthe locks of those commands being reduced as well.\n- Patch has some typos, like \"an share\".\n- Documentation can be simplified. Listing the flavors of ALTER TABLE\ntaking ROW SHARE EXCLUSIVE in mvcc/sgml is clumsy IMO, it is better to\nmention that in alter_table.sgml and have a link pointing to the page\nof ALTER TABLE in mvcc.sgml.\n- In ATAddForeignKeyConstraint, I think that it would be better to not\nremove the comment, adding instead that ShareRowExclusiveLock is\nactually safe because CREATE TRIGGER uses this level of lock.\nAll those things gathered give the patch attached. Andreas, if you are\nfine with it I think that we could pass it to a committer.\nRegards,\n-- \nMichael\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Wed, 14 Jan 2015 16:48:00 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"I wrote:\n> I think that we could pass it to a committer.\nMarked as such.\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Thu, 15 Jan 2015 16:00:55 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 01/14/2015 08:48 AM, Michael Paquier wrote:\n> All those things gathered give the patch attached. Andreas, if you are\n> fine with it I think that we could pass it to a committer.\n\nExcellent changes. Thanks for the patch and the reviews.\n\n-- \nAndreas Karlsson\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 16 Jan 2015 00:36:55 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"Hi,\n\n> \t/*\n> -\t * Grab an exclusive lock on the pk table, so that someone doesn't delete\n> -\t * rows out from under us. (Although a lesser lock would do for that\n> -\t * purpose, we'll need exclusive lock anyway to add triggers to the pk\n> -\t * table; trying to start with a lesser lock will just create a risk of\n> -\t * deadlock.)\n> +\t * Grab ShareRowExclusiveLock on the pk table, so that someone doesn't\n> +\t * delete rows out from under us. Note that this does not create risks\n> +\t * of deadlocks as triggers add added to the pk table using the same\n> +\t * lock.\n> \t */\n\n\"add added\" doesn't look intended. The rest of the sentence doesn't look\nentirely right either.\n\n> \t/*\n> \t * Triggers must be on tables or views, and there are additional\n> @@ -526,8 +526,7 @@ CreateTrigger(CreateTrigStmt *stmt, const char *queryString,\n> \t * can skip this for internally generated triggers, since the name\n> \t * modification above should be sufficient.\n> \t *\n> -\t * NOTE that this is cool only because we have AccessExclusiveLock on the\n> -\t * relation, so the trigger set won't be changing underneath us.\n> +\t * NOTE that this is cool only because of the unique contraint.\n\nI fail to see what the unique constraint has to do with this? The\nprevious comment refers to the fact that the AccessExclusiveLock is what\nprevents a race where another transaction adds a trigger with the same\nname already exists. Yes, the unique index would, as noted earlier in\nthe comment, catch the error. But that's not the point of the\ncheck. Unless I miss something the comment is just as true if you\nreplace the access exclusive with share row exlusive as it's also self\nconflicting.\n\n> @@ -1272,8 +1271,7 @@ renametrig(RenameStmt *stmt)\n> \t * on tgrelid/tgname would complain anyway) and to ensure a trigger does\n> \t * exist with oldname.\n> \t *\n> -\t * NOTE that this is cool only because we have AccessExclusiveLock on the\n> -\t * relation, so the trigger set won't be changing underneath us.\n> +\t * NOTE that this is cool only because there is a unique constraint.\n> \t */\n\nSame as above.\n\n> \ttgrel = heap_open(TriggerRelationId, RowExclusiveLock);\n> \n> diff --git a/src/backend/utils/adt/ruleutils.c b/src/backend/utils/adt/ruleutils.c\n> index dd748ac..8eeccf2 100644\n> --- a/src/backend/utils/adt/ruleutils.c\n> +++ b/src/backend/utils/adt/ruleutils.c\n> @@ -699,7 +699,8 @@ pg_get_triggerdef_worker(Oid trigid, bool pretty)\n> \tHeapTuple\tht_trig;\n> \tForm_pg_trigger trigrec;\n> \tStringInfoData buf;\n> -\tRelation\ttgrel;\n> +\tSnapshot\tsnapshot = RegisterSnapshot(GetTransactionSnapshot());\n> +\tRelation\ttgrel = heap_open(TriggerRelationId, AccessShareLock);\n> \tScanKeyData skey[1];\n> \tSysScanDesc tgscan;\n> \tint\t\t\tfindx = 0;\n> @@ -710,18 +711,18 @@ pg_get_triggerdef_worker(Oid trigid, bool pretty)\n> \t/*\n> \t * Fetch the pg_trigger tuple by the Oid of the trigger\n> \t */\n> -\ttgrel = heap_open(TriggerRelationId, AccessShareLock);\n> -\n> \tScanKeyInit(&skey[0],\n> \t\t\t\tObjectIdAttributeNumber,\n> \t\t\t\tBTEqualStrategyNumber, F_OIDEQ,\n> \t\t\t\tObjectIdGetDatum(trigid));\n> \n> \ttgscan = systable_beginscan(tgrel, TriggerOidIndexId, true,\n> -\t\t\t\t\t\t\t\tNULL, 1, skey);\n> +\t\t\t\t\t\t\t\tsnapshot, 1, skey);\n> \n> \tht_trig = systable_getnext(tgscan);\n> \n> +\tUnregisterSnapshot(snapshot);\n> +\n> \tif (!HeapTupleIsValid(ht_trig))\n> \t\telog(ERROR, \"could not find tuple for trigger %u\", trigid);\n>\n\nHm. Pushing the snapshot is supposed to make this fully mvcc? Idon't\nthink that's actually sufficient because of the deparsing of the WHEN\nclause and of the function name.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 16 Jan 2015 15:01:06 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 01/16/2015 03:01 PM, Andres Freund wrote:\n> Hi,\n>\n>> \t/*\n>> -\t * Grab an exclusive lock on the pk table, so that someone doesn't delete\n>> -\t * rows out from under us. (Although a lesser lock would do for that\n>> -\t * purpose, we'll need exclusive lock anyway to add triggers to the pk\n>> -\t * table; trying to start with a lesser lock will just create a risk of\n>> -\t * deadlock.)\n>> +\t * Grab ShareRowExclusiveLock on the pk table, so that someone doesn't\n>> +\t * delete rows out from under us. Note that this does not create risks\n>> +\t * of deadlocks as triggers add added to the pk table using the same\n>> +\t * lock.\n>> \t */\n>\n> \"add added\" doesn't look intended. The rest of the sentence doesn't look\n> entirely right either.\n\nIt was intended to be \"are added\", but the sentence is pretty awful \nanyway. I am not sure the sentence is really necessary anyway.\n\n>> \t/*\n>> \t * Triggers must be on tables or views, and there are additional\n>> @@ -526,8 +526,7 @@ CreateTrigger(CreateTrigStmt *stmt, const char *queryString,\n>> \t * can skip this for internally generated triggers, since the name\n>> \t * modification above should be sufficient.\n>> \t *\n>> -\t * NOTE that this is cool only because we have AccessExclusiveLock on the\n>> -\t * relation, so the trigger set won't be changing underneath us.\n>> +\t * NOTE that this is cool only because of the unique contraint.\n>\n> I fail to see what the unique constraint has to do with this? The\n> previous comment refers to the fact that the AccessExclusiveLock is what\n> prevents a race where another transaction adds a trigger with the same\n> name already exists. Yes, the unique index would, as noted earlier in\n> the comment, catch the error. But that's not the point of the\n> check. Unless I miss something the comment is just as true if you\n> replace the access exclusive with share row exlusive as it's also self\n> conflicting.\n\nYeah, this must have been a remainder from the version where I only \ngrabbed a ShareLock. The comment should be restored.\n\n> Hm. Pushing the snapshot is supposed to make this fully mvcc? Idon't\n> think that's actually sufficient because of the deparsing of the WHEN\n> clause and of the function name.\n\nIndeed. As Noah and I discussed previously in this thread we would need \nto do quite a bit of refactoring of ruleutils.c to make it fully MVCC. \nFor this reason I opted to only lower the lock levels of ADD and ALTER \nTRIGGER, and not DROP TRIGGER. Neither of those require MVCC of then \nWHEN clause.\n\nAndreas\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 16 Jan 2015 15:16:20 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 2015-01-16 15:16:20 +0100, Andreas Karlsson wrote:\n> Indeed. As Noah and I discussed previously in this thread we would need to\n> do quite a bit of refactoring of ruleutils.c to make it fully MVCC.\n\nRight.\n\n> For this reason I opted to only lower the lock levels of ADD and ALTER\n> TRIGGER, and not DROP TRIGGER. Neither of those require MVCC of then\n> WHEN clause.\n\nI'm unconvinced that this is true. Using a snapshot for part of getting\na definition certainly opens the door for getting strange\nresults.\n\nAcquiring a lock that prevents schema changes on the table and then\ngetting the definition using the syscaches guarantees that that\ndefinition is at least self consistent because no further schema changes\nare possible and the catalog caches will be up2date.\n\nWhat you're doing though is doing part of the scan using the\ntransaction's snapshot (as used by pg_dump that will usually be a\nrepeatable read snapshot and possibly quite old) and the other using a\nfresh catalog snapshot. This can result in rather odd things.\n\nJust consider:\nS1: CREATE TABLE flubber(id serial primary key, data text);\nS1: CREATE FUNCTION blarg() RETURNS TRIGGER LANGUAGE plpgsql AS $$BEGIN RETURN NEW; END;$$;\nS1: CREATE TRIGGER flubber_blarg BEFORE INSERT ON flubber FOR EACH ROW WHEN (NEW.data IS NOT NULL) EXECUTE PROCEDURE blarg();\nS2: BEGIN TRANSACTION ISOLATION LEVEL SERIALIZABLE;\nS2: SELECT 'dosomethingelse';\nS1: ALTER TABLE flubber RENAME TO wasflubber;\nS1: ALTER TABLE wasflubber RENAME COLUMN data TO wasdata;\nS1: ALTER TRIGGER flubber_blarg ON wasflubber RENAME TO wasflubber_blarg;\nS1: ALTER FUNCTION blarg() RENAME TO wasblarg;\nS2: SELECT pg_get_triggerdef(oid) FROM pg_trigger;\n\nThis will give you: The old trigger name. The new table name. The new\ncolumn name. The new function name.\n\nI don't think using a snapshot for tiny parts of these functions\nactually buys anything. Now, this isn't a pattern you introduced. But I\nthink we should think about this for a second before expanding it\nfurther.\n\nBefore you argue that this isn't relevant for pg_dump: It is. Precisely\nthe above can happen - just replace the 'dosomethingelse' with several\nLOCK TABLEs as pg_dump does. The first blocks after a snapshot has been\nacquired. While waiting, all the ALTERs happen.\n\nArguably the benefit here is that the window for this issue is becoming\nsmaller as pg_dump (and hopefully other possible callers) acquire\nexclusive locks on the table. I.e. that the lowering of the lock level\ndoesn't introduce new races. But on the other side of the coin, this\nmakes the result of pg_get_triggerdef() even *more* inaccurate in many\ncases.\n\nGreetings,\n\nAndres Freund\n\n-- \n Andres Freund\t http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 16 Jan 2015 16:59:56 +0100","msg_from":"Andres Freund <andres@2ndquadrant.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On Fri, Jan 16, 2015 at 10:59 AM, Andres Freund <andres@2ndquadrant.com> wrote:\n> Just consider:\n> S1: CREATE TABLE flubber(id serial primary key, data text);\n> S1: CREATE FUNCTION blarg() RETURNS TRIGGER LANGUAGE plpgsql AS $$BEGIN RETURN NEW; END;$$;\n> S1: CREATE TRIGGER flubber_blarg BEFORE INSERT ON flubber FOR EACH ROW WHEN (NEW.data IS NOT NULL) EXECUTE PROCEDURE blarg();\n> S2: BEGIN TRANSACTION ISOLATION LEVEL SERIALIZABLE;\n> S2: SELECT 'dosomethingelse';\n> S1: ALTER TABLE flubber RENAME TO wasflubber;\n> S1: ALTER TABLE wasflubber RENAME COLUMN data TO wasdata;\n> S1: ALTER TRIGGER flubber_blarg ON wasflubber RENAME TO wasflubber_blarg;\n> S1: ALTER FUNCTION blarg() RENAME TO wasblarg;\n> S2: SELECT pg_get_triggerdef(oid) FROM pg_trigger;\n>\n> This will give you: The old trigger name. The new table name. The new\n> column name. The new function name.\n\nOuch. That's clearly no good. I'm struggling to understand whether\nthis is a problem with our previous analysis, or a problem with this\npatch:\n\nhttp://www.postgresql.org/message-id/20141028003356.GA387814@tornado.leadboat.com\n\npg_get_triggerdef_worker() relies on generate_function_name(), which\nuses the system caches, and on get_rule_expr(), for deparsing the WHEN\nclause. If we allowed only ADDING triggers with a lesser lock and\nnever modifying or dropping them with a lesser lock, then changing the\ninitial scan of pg_trigger at the top of pg_get_triggerdef_worker() to\nuse the transaction snapshot might be OK; if we can see the trigger\nwith the transaction snapshot at all, we know it can't have\nsubsequently changed. But allowing alterations of any kind isn't\ngoing to work, so I think our previous analysis on that point was\nincorrect.\n\nI *think* we could fix that if generate_function_name() and\nget_rule_expr() had an option to use the active snapshot instead of a\nfresh snapshot. The former doesn't look too hard to arrange, but the\nlatter is a tougher nut to crack.\n\n-- \nRobert Haas\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 19 Jan 2015 12:14:31 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 01/19/2015 06:14 PM, Robert Haas wrote:\n> On Fri, Jan 16, 2015 at 10:59 AM, Andres Freund <andres@2ndquadrant.com> wrote:\n>> This will give you: The old trigger name. The new table name. The new\n>> column name. The new function name.\n>\n> Ouch. That's clearly no good. I'm struggling to understand whether\n> this is a problem with our previous analysis, or a problem with this\n> patch:\n>\n> http://www.postgresql.org/message-id/20141028003356.GA387814@tornado.leadboat.com\n>\n> pg_get_triggerdef_worker() relies on generate_function_name(), which\n> uses the system caches, and on get_rule_expr(), for deparsing the WHEN\n> clause. If we allowed only ADDING triggers with a lesser lock and\n> never modifying or dropping them with a lesser lock, then changing the\n> initial scan of pg_trigger at the top of pg_get_triggerdef_worker() to\n> use the transaction snapshot might be OK; if we can see the trigger\n> with the transaction snapshot at all, we know it can't have\n> subsequently changed. But allowing alterations of any kind isn't\n> going to work, so I think our previous analysis on that point was\n> incorrect.\n>\n> I *think* we could fix that if generate_function_name() and\n> get_rule_expr() had an option to use the active snapshot instead of a\n> fresh snapshot. The former doesn't look too hard to arrange, but the\n> latter is a tougher nut to crack.\n\nThe function name in the trigger should already be broken in master \nsince changing the name of a function does not lock the table. Would be \nreally neat to fix though.\n\nA possible way forward I see here is to undo my changes to the lock \nlevel of ALTER TRIGGER and of pg_get_triggerdef_worker(). This should \nwork if Noah is correct about his analysis in where he came to the \nconclusion: \"pg_get_triggerdef() is fine as-is with concurrent CREATE \nTRIGGER.\". Is this still true with the new information?\n\n-- \nAndreas Karlsson\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 20 Jan 2015 00:06:04 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On Fri, Jan 16, 2015 at 04:59:56PM +0100, Andres Freund wrote:\n> On 2015-01-16 15:16:20 +0100, Andreas Karlsson wrote:\n> > For this reason I opted to only lower the lock levels of ADD and ALTER\n> > TRIGGER, and not DROP TRIGGER. Neither of those require MVCC of then\n> > WHEN clause.\n> \n> I'm unconvinced that this is true. Using a snapshot for part of getting\n> a definition certainly opens the door for getting strange\n> results.\n> \n> Acquiring a lock that prevents schema changes on the table and then\n> getting the definition using the syscaches guarantees that that\n> definition is at least self consistent because no further schema changes\n> are possible and the catalog caches will be up2date.\n> \n> What you're doing though is doing part of the scan using the\n> transaction's snapshot (as used by pg_dump that will usually be a\n> repeatable read snapshot and possibly quite old) and the other using a\n> fresh catalog snapshot. This can result in rather odd things.\n> \n> Just consider:\n> S1: CREATE TABLE flubber(id serial primary key, data text);\n> S1: CREATE FUNCTION blarg() RETURNS TRIGGER LANGUAGE plpgsql AS $$BEGIN RETURN NEW; END;$$;\n> S1: CREATE TRIGGER flubber_blarg BEFORE INSERT ON flubber FOR EACH ROW WHEN (NEW.data IS NOT NULL) EXECUTE PROCEDURE blarg();\n> S2: BEGIN TRANSACTION ISOLATION LEVEL SERIALIZABLE;\n> S2: SELECT 'dosomethingelse';\n> S1: ALTER TABLE flubber RENAME TO wasflubber;\n> S1: ALTER TABLE wasflubber RENAME COLUMN data TO wasdata;\n> S1: ALTER TRIGGER flubber_blarg ON wasflubber RENAME TO wasflubber_blarg;\n> S1: ALTER FUNCTION blarg() RENAME TO wasblarg;\n> S2: SELECT pg_get_triggerdef(oid) FROM pg_trigger;\n> \n> This will give you: The old trigger name. The new table name. The new\n> column name. The new function name.\n> \n> I don't think using a snapshot for tiny parts of these functions\n> actually buys anything. Now, this isn't a pattern you introduced. But I\n> think we should think about this for a second before expanding it\n> further.\n\nFair enough. It did reinforce pg_get_constraintdef() as a subroutine of\npg_dump rather than an independent, rigorous interface. It perhaps made the\nfunction worse for non-pg_dump callers. In that vein, each one of these hacks\nhas a cost. One could make a reasonable argument that ALTER TRIGGER RENAME\nlocking is not important enough to justify spreading the hack from\npg_get_constraintdef() to pg_get_triggerdef(). Lowering the CREATE TRIGGER\nlock level does not require any ruleutils.c change for the benefit of pg_dump,\nbecause pg_dump won't see the pg_trigger row of a too-recent trigger.\n\n> Before you argue that this isn't relevant for pg_dump: It is. Precisely\n> the above can happen - just replace the 'dosomethingelse' with several\n> LOCK TABLEs as pg_dump does. The first blocks after a snapshot has been\n> acquired. While waiting, all the ALTERs happen.\n\nWe wish pg_dump would take a snapshot of the database; that is, we wish its\noutput always matched some serial execution of transactions. pg_dump has,\nsince ancient times, failed to achieve that if non-table DDL commits during\nthe dump or if table DDL commits between acquiring the dump transaction\nsnapshot and acquiring the last table lock. My reviews have defended the\nstandard that table DDL issued after pg_dump has acquired locks does not\nchange the dump. That's what we bought with pg_get_constraintdef()'s use of\nthe transaction snapshot and would buy with the same in pg_get_triggerdef().\nMy reviews have deliberately ignored effects on scenarios where pg_dump\nalready fails to guarantee snapshot-like output.\n\n> Arguably the benefit here is that the window for this issue is becoming\n> smaller as pg_dump (and hopefully other possible callers) acquire\n> exclusive locks on the table. I.e. that the lowering of the lock level\n> doesn't introduce new races. But on the other side of the coin, this\n> makes the result of pg_get_triggerdef() even *more* inaccurate in many\n> cases.\n\nWhat is this about pg_dump acquiring exclusive locks?\n\nTo summarize, the problem you raise has been out of scope, because it affects\npg_dump only at times when pg_dump is already wrong.\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Tue, 20 Jan 2015 04:08:25 -0500","msg_from":"Noah Misch <noah@leadboat.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 01/20/2015 10:08 AM, Noah Misch wrote:\n> Fair enough. It did reinforce pg_get_constraintdef() as a subroutine of\n> pg_dump rather than an independent, rigorous interface. It perhaps made the\n> function worse for non-pg_dump callers. In that vein, each one of these hacks\n> has a cost. One could make a reasonable argument that ALTER TRIGGER RENAME\n> locking is not important enough to justify spreading the hack from\n> pg_get_constraintdef() to pg_get_triggerdef(). Lowering the CREATE TRIGGER\n> lock level does not require any ruleutils.c change for the benefit of pg_dump,\n> because pg_dump won't see the pg_trigger row of a too-recent trigger.\n\nI agree with this view, and am not sure myself that it is worth lowering \nthe lock level of ALTER TRIGGER RENAME. I have attached a patch without \nthe changes to ALTER TRIGGER and ruleutils.c and also fixes the comment \nissues noted by Andres.\n\n-- \nAndreas Karlsson\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Thu, 22 Jan 2015 22:31:44 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 01/22/2015 10:31 PM, Andreas Karlsson wrote:\n> I agree with this view, and am not sure myself that it is worth lowering\n> the lock level of ALTER TRIGGER RENAME. I have attached a patch without\n> the changes to ALTER TRIGGER and ruleutils.c and also fixes the comment\n> issues noted by Andres.\n\nWhops, forgot to include the isolation tests.\n\n-- \nAndreas Karlsson\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Fri, 23 Jan 2015 00:55:21 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On Fri, Jan 23, 2015 at 8:55 AM, Andreas Karlsson <andreas@proxel.se> wrote:\n> On 01/22/2015 10:31 PM, Andreas Karlsson wrote:\n>>\n>> I agree with this view, and am not sure myself that it is worth lowering\n>> the lock level of ALTER TRIGGER RENAME. I have attached a patch without\n>> the changes to ALTER TRIGGER and ruleutils.c and also fixes the comment\n>> issues noted by Andres.\n>\n>\n> Whops, forgot to include the isolation tests.\nOk, so the deal is to finally reduce the locks to\nShareRowExclusiveLock for the following commands :\n- CREATE TRIGGER\n- ALTER TABLE ENABLE/DISABLE\n- ALTER TABLE ADD CONSTRAINT\n\nLooking at the latest patch, it seems that in\nAlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\nAT_AddConstraintRecurse and AT_ProcessedConstraint under the same\nbanner as AT_AddConstraint. Thoughts?\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 30 Jan 2015 15:48:25 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 01/30/2015 07:48 AM, Michael Paquier wrote:\n> Ok, so the deal is to finally reduce the locks to\n> ShareRowExclusiveLock for the following commands :\n> - CREATE TRIGGER\n> - ALTER TABLE ENABLE/DISABLE\n> - ALTER TABLE ADD CONSTRAINT\n\nCorrect. I personally still find this useful enough to justify a patch.\n\n> Looking at the latest patch, it seems that in\n> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\n> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\n> banner as AT_AddConstraint. Thoughts?\n\nGood point. I think moving those would be a good thing even though it is \ntechnically not necessary for AT_AddConstraintRecurse, since that one \nshould only be related to check constraints.\n\n-- \nAndreas\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 30 Jan 2015 19:26:28 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On Fri, Jan 30, 2015 at 10:26 PM, Andreas Karlsson wrote:\n> On 01/30/2015 07:48 AM, Michael Paquier wrote:\n>> Looking at the latest patch, it seems that in\n>> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\n>> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\n>> banner as AT_AddConstraint. Thoughts?\n>\n> Good point. I think moving those would be a good thing even though it is\n> technically not necessary for AT_AddConstraintRecurse, since that one should\n> only be related to check constraints.\n\nAndreas, are you planning to send an updated patch?\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 6 Feb 2015 11:16:33 +0400","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 02/06/2015 08:16 AM, Michael Paquier wrote:\n> On Fri, Jan 30, 2015 at 10:26 PM, Andreas Karlsson wrote:\n>> On 01/30/2015 07:48 AM, Michael Paquier wrote:\n>>> Looking at the latest patch, it seems that in\n>>> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\n>>> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\n>>> banner as AT_AddConstraint. Thoughts?\n>>\n>> Good point. I think moving those would be a good thing even though it is\n>> technically not necessary for AT_AddConstraintRecurse, since that one should\n>> only be related to check constraints.\n>\n> Andreas, are you planning to send an updated patch?\n\nYes, I will hopefully send it later today or tomorrow.\n\nAndreas\n\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 06 Feb 2015 10:38:21 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 01/30/2015 07:48 AM, Michael Paquier wrote:\n> Looking at the latest patch, it seems that in\n> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\n> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\n> banner as AT_AddConstraint. Thoughts?\n\nA new version of the patch is attached which treats them as the same for \nlocking. I think it is correct and improves readability to do so.\n\n-- \nAndreas Karlsson\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers","msg_date":"Sun, 08 Feb 2015 02:05:46 +0100","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On Sun, Feb 8, 2015 at 10:05 AM, Andreas Karlsson <andreas@proxel.se> wrote:\n> On 01/30/2015 07:48 AM, Michael Paquier wrote:\n>>\n>> Looking at the latest patch, it seems that in\n>> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\n>> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\n>> banner as AT_AddConstraint. Thoughts?\n>\n>\n> A new version of the patch is attached which treats them as the same for\n> locking. I think it is correct and improves readability to do so.\n\nWell then, let's switch it to \"Ready for committer\". I am moving as\nwell this entry to the next CF with the same status.\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Fri, 13 Feb 2015 10:10:44 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 7 February 2015 at 20:05, Andreas Karlsson <andreas@proxel.se> wrote:\n> On 01/30/2015 07:48 AM, Michael Paquier wrote:\n>>\n>> Looking at the latest patch, it seems that in\n>> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\n>> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\n>> banner as AT_AddConstraint. Thoughts?\n>\n>\n> A new version of the patch is attached which treats them as the same for\n> locking. I think it is correct and improves readability to do so.\n\nCommitted. We move forwards, slowly but surely. Thanks for the patch.\n\n-- \n Simon Riggs http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, RemoteDBA, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 5 Apr 2015 11:56:42 -0400","msg_from":"Simon Riggs <simon@2ndQuadrant.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On Mon, Apr 6, 2015 at 12:56 AM, Simon Riggs <simon@2ndquadrant.com> wrote:\n> On 7 February 2015 at 20:05, Andreas Karlsson <andreas@proxel.se> wrote:\n>> On 01/30/2015 07:48 AM, Michael Paquier wrote:\n>>>\n>>> Looking at the latest patch, it seems that in\n>>> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\n>>> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\n>>> banner as AT_AddConstraint. Thoughts?\n>>\n>>\n>> A new version of the patch is attached which treats them as the same for\n>> locking. I think it is correct and improves readability to do so.\n>\n> Committed. We move forwards, slowly but surely. Thanks for the patch.\n\nCool! Thanks for showing up.\n-- \nMichael\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 6 Apr 2015 08:19:37 +0900","msg_from":"Michael Paquier <michael.paquier@gmail.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 04/05/2015 05:56 PM, Simon Riggs wrote:\n> Committed. We move forwards, slowly but surely. Thanks for the patch.\n\nThanks to you and the reviewers for helping me out with this patch!\n\n-- \nAndreas Karlsson\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Mon, 06 Apr 2015 01:36:49 +0200","msg_from":"Andreas Karlsson <andreas@proxel.se>","msg_from_op":true,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"},{"msg_contents":"On 5 April 2015 at 19:19, Michael Paquier <michael.paquier@gmail.com> wrote:\n\n> Cool! Thanks for showing up.\n\nVisibility <> Activity. How is REINDEX CONCURRENTLY doing?\n\n-- \n Simon Riggs http://www.2ndQuadrant.com/\n PostgreSQL Development, 24x7 Support, RemoteDBA, Training & Services\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers\n","msg_date":"Sun, 5 Apr 2015 19:37:46 -0400","msg_from":"Simon Riggs <simon@2ndQuadrant.com>","msg_from_op":false,"msg_subject":"Re: PATCH: Reducing lock strength of trigger and foreign\n key DDL"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI have been thinking about why we need to grab an AccessExclusiveLock on \\nthe table with the PK when we add a foreign key. Adding new tables with \\nforeign keys to old ones is common so it would be really nice if the \\nlock strength could be reduced.\\n\\nA comment in the code claims that we need to lock so no rows are deleted \\nunder us and that adding a trigger will lock in AccessExclusive anyway. \\nBut with MVCC catalog access and the removal of SnapshotNow I do not see \\nwhy adding a trigger would require an exclusive lock. Just locking for \\ndata changes should be enough.\\n\\nLooking at the code the only see the duplicate name check use the fact \\nthat we have grabbed an exclusive lock and that should work anyway due \\nto the unique constraint, but since I am pretty new to the code base I \\ncould be missing something obvious. I have attached a proof of concept \\npatch which reduces the lock strength to ShareLock.\\n\\nWhat do you say? Am I missing something?\\n\\nMy gut feel also says that if we know it is a RI trigger we are adding \\nthen AccessShare should be enough for the PK table, since we could rely \\non row locks to prevent rows from being deleted. It would be really nice \\nthough if this was possible since this would make it possible to add a \\nnew table with foreign keys and data without locking anything more than \\nthe referred rows.\\n\\nAndreas\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Wed, 22 Oct 2014 09:06:52 +0200\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"Andreas Karlsson <andreas@proxel.se> writes:\\n> I have attached a proof of concept \\n> patch which reduces the lock strength to ShareLock.\\n\\nYou're kidding, right? ShareLock isn't even self-exclusive.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 22 Oct 2014 10:13:58 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"On 10/22/2014 04:13 PM, Tom Lane wrote:\\n> Andreas Karlsson <andreas@proxel.se> writes:\\n>> I have attached a proof of concept\\n>> patch which reduces the lock strength to ShareLock.\\n>\\n> You're kidding, right? ShareLock isn't even self-exclusive.\\n\\nWhy would it have to be self-exclusive? As far as I know we only need to \\nensure that nobody changes the rows while we add the trigger. Adding \\nmultiple triggers concurrently should not pose a problem unless I am \\nmissing something (which I probably am).\\n\\nAndreas\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 23 Oct 2014 09:25:05 +0200\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 22, 2014 at 3:06 AM, Andreas Karlsson <andreas@proxel.se> wrote:\\n> I have been thinking about why we need to grab an AccessExclusiveLock on the\\n> table with the PK when we add a foreign key. Adding new tables with foreign\\n> keys to old ones is common so it would be really nice if the lock strength\\n> could be reduced.\\n>\\n> A comment in the code claims that we need to lock so no rows are deleted\\n> under us and that adding a trigger will lock in AccessExclusive anyway. But\\n> with MVCC catalog access and the removal of SnapshotNow I do not see why\\n> adding a trigger would require an exclusive lock. Just locking for data\\n> changes should be enough.\\n\\nThe use of MVCC catalog access doesn't necessarily mean that adding a\\ntrigger doesn't require an AccessExclusive lock. Those changes - if I\\ndare to say so myself - solved a complex and important problem, but\\nonly one of many problems in this area, and people seem prone to\\nthinking that they solved more problems than they in fact did.\\n\\nI think instead of focusing on foreign keys, we should rewind a bit\\nand think about the locking level required to add a trigger. If we\\nfigure something out there, then we can consider how it affects\\nforeign keys. I went looking for previous discussions of remaining\\nhazards and found these postings:\\n\\nhttp://www.postgresql.org/message-id/CA+TgmoY4GLsXZk0tAO29-LJtcuj0SL1xWCwQ51xb-HFYsgi5RQ@mail.gmail.com\\nhttp://www.postgresql.org/message-id/20893.1393892127@sss.pgh.pa.us\\nhttp://www.postgresql.org/message-id/20140306224340.GA3551655@tornado.leadboat.com\\n\\nAs far as triggers are concerned, the issue of skew between the\\ntransaction snapshot and what the ruleutils.c snapshots do seems to be\\nthe principal issue. Commit e5550d5fec66aa74caad1f79b79826ec64898688\\nchanged pg_get_constraintdef() to use an MVCC snapshot rather than a\\ncurrent MVCC snapshot; if that change is safe, I am not aware of any\\nreason why we couldn't change pg_get_triggerdef() similarly. Barring\\nfurther hazards I haven't thought of, I would expect that we could add\\na trigger to a relation with only ShareRowExclusiveLock. Anything\\nless than ShareRowExclusiveLock would open up strange timing races\\naround the firing of triggers by transactions writing the table: they\\nmight or might not notice that a trigger had been added before\\nend-of-transaction, depending on the timing of cache flushes, which\\ncertainly seems no good. But even RowExclusiveLock seems like a large\\nimprovement over AccessExclusiveLock.\\n\\nWhen a constraint trigger - which is used to implement a foreign key -\\nis added, there are actually TWO tables involved: the table upon which\\nthe trigger will actually fire, and some other table which is\\nmentioned in passing in the trigger definition. It's possible that\\nthe locking requirements for the secondary table are weaker since I\\ndon't think the presence of the trigger actually affects runtime\\nbehavior there. However, there's probably little point in try to\\nweaken the lock to less than the level required for the main table\\nbecause a foreign key involves adding referential integrity triggers\\nto both tables.\\n\\nSo I tentatively propose (and with due regard for the possibility\\nothers may see dangers that I've missed) that a reasonable goal would\\nbe to lower the lock strength required for both CREATE TRIGGER and ADD\\nFOREIGN KEY from AccessExclusiveLock to ShareRowExclusiveLock,\\nallowing concurrent SELECT and SELECT FOR SHARE against the tables,\\nbut not any actual write operations.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 24 Oct 2014 12:07:42 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 24, 2014 at 12:07:42PM -0400, Robert Haas wrote:\\n> I think instead of focusing on foreign keys, we should rewind a bit\\n> and think about the locking level required to add a trigger.\\n\\nAgreed.\\n\\n> http://www.postgresql.org/message-id/CA+TgmoY4GLsXZk0tAO29-LJtcuj0SL1xWCwQ51xb-HFYsgi5RQ@mail.gmail.com\\n> http://www.postgresql.org/message-id/20893.1393892127@sss.pgh.pa.us\\n> http://www.postgresql.org/message-id/20140306224340.GA3551655@tornado.leadboat.com\\n> \\n> As far as triggers are concerned, the issue of skew between the\\n> transaction snapshot and what the ruleutils.c snapshots do seems to be\\n> the principal issue. Commit e5550d5fec66aa74caad1f79b79826ec64898688\\n> changed pg_get_constraintdef() to use an MVCC snapshot rather than a\\n> current MVCC snapshot; if that change is safe, I am not aware of any\\n> reason why we couldn't change pg_get_triggerdef() similarly.\\n\\npg_get_triggerdef() is fine as-is with concurrent CREATE TRIGGER. The\\npg_get_constraintdef() change arose to ensure a consistent result when\\nconcurrent ALTER TABLE VALIDATE CONSTRAINT mutates a constraint definition.\\n(Reducing the lock level of DROP TRIGGER or ALTER TRIGGER, however, would\\ncreate the analogous problem for pg_get_triggerdef().)\\n\\n> So I tentatively propose (and with due regard for the possibility\\n> others may see dangers that I've missed) that a reasonable goal would\\n> be to lower the lock strength required for both CREATE TRIGGER and ADD\\n> FOREIGN KEY from AccessExclusiveLock to ShareRowExclusiveLock,\\n> allowing concurrent SELECT and SELECT FOR SHARE against the tables,\\n> but not any actual write operations.\\n\\n+1\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 25 Oct 2014 14:00:17 -0400\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"On 10/24/2014 06:07 PM, Robert Haas wrote:\\n> I think instead of focusing on foreign keys, we should rewind a bit\\n> and think about the locking level required to add a trigger.\\n\\nAgreed.\\n\\n> As far as triggers are concerned, the issue of skew between the\\n> transaction snapshot and what the ruleutils.c snapshots do seems to be\\n> the principal issue. Commit e5550d5fec66aa74caad1f79b79826ec64898688\\n> changed pg_get_constraintdef() to use an MVCC snapshot rather than a\\n> current MVCC snapshot; if that change is safe, I am not aware of any\\n> reason why we couldn't change pg_get_triggerdef() similarly. Barring\\n> further hazards I haven't thought of, I would expect that we could add\\n> a trigger to a relation with only ShareRowExclusiveLock.\\n\\nThanks for the info. This is just the kind of issues I was worrying about.\\n\\n> Anything\\n> less than ShareRowExclusiveLock would open up strange timing races\\n> around the firing of triggers by transactions writing the table: they\\n> might or might not notice that a trigger had been added before\\n> end-of-transaction, depending on the timing of cache flushes, which\\n> certainly seems no good. But even RowExclusiveLock seems like a large\\n> improvement over AccessExclusiveLock.\\n\\nWould not ShareLock give the same result, except for also allowing \\nconcurrent CREATE INDEX and concurrent other CREATE TRIGGER which does \\nnot look dangerous to me?\\n\\n From a user point of view ShareRowExclusiveLock should be as useful as \\nShareLock.\\n\\n> When a constraint trigger - which is used to implement a foreign key -\\n> is added, there are actually TWO tables involved: the table upon which\\n> the trigger will actually fire, and some other table which is\\n> mentioned in passing in the trigger definition. It's possible that\\n> the locking requirements for the secondary table are weaker since I\\n> don't think the presence of the trigger actually affects runtime\\n> behavior there. However, there's probably little point in try to\\n> weaken the lock to less than the level required for the main table\\n> because a foreign key involves adding referential integrity triggers\\n> to both tables.\\n>\\n> So I tentatively propose (and with due regard for the possibility\\n> others may see dangers that I've missed) that a reasonable goal would\\n> be to lower the lock strength required for both CREATE TRIGGER and ADD\\n> FOREIGN KEY from AccessExclusiveLock to ShareRowExclusiveLock,\\n> allowing concurrent SELECT and SELECT FOR SHARE against the tables,\\n> but not any actual write operations.\\n\\nAgreed.. But I think reducing the lock level of the secondary table is \\nmuch more important than doing the same for the primary table due to the \\ncase where the secondary table is an existing table which is hit by a \\nworkload of long running queries and DML while the primary is a new \\ntable which is added now. In my dream world I could add the new table \\nwithout any disruption at all of queries using the secondary table, no \\nmatter the duration of the transaction adding the table (barring \\ninsertion of actual data into the primary table, which would take row \\nlocks).\\n\\nThis is just a dream scenario though, and focusing on triggers is indeed \\nthe reasonable goal for 9.5.\\n\\n-- \\nAndreas Karlsson\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 27 Oct 2014 02:48:55 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"There are actually TWO tables involved: the table upon which \\nthe trigger will actually fire, and some other table which is \\nmentioned in passing in the trigger definition. It's possible that \\nthe locking requirements for the secondary table are weaker since I \\ndon't think the presence of the trigger actually affects runtime \\nbehavior there. However, there's probably little point in try to \\nweaken the lock to less than the level required for the main table \\nbecause a foreign key involves adding referential integrity triggers \\nto both tables. \\n\\n\\n\\n\\n-----\\nGUL\\n--\\nView this message in context: http://postgresql.1045698.n5.nabble.com/Reducing-lock-strength-of-adding-foreign-keys-tp5823894p5824376.html\\nSent from the PostgreSQL - hackers mailing list archive at Nabble.com.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 27 Oct 2014 01:44:23 -0700 (PDT)\",\n \"msg_from\": \"adamrose045 <adamrose045@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"Thanks for weighing in, Noah.\\n\\nOn Sat, Oct 25, 2014 at 2:00 PM, Noah Misch <noah@leadboat.com> wrote:\\n>> http://www.postgresql.org/message-id/CA+TgmoY4GLsXZk0tAO29-LJtcuj0SL1xWCwQ51xb-HFYsgi5RQ@mail.gmail.com\\n>> http://www.postgresql.org/message-id/20893.1393892127@sss.pgh.pa.us\\n>> http://www.postgresql.org/message-id/20140306224340.GA3551655@tornado.leadboat.com\\n>>\\n>> As far as triggers are concerned, the issue of skew between the\\n>> transaction snapshot and what the ruleutils.c snapshots do seems to be\\n>> the principal issue. Commit e5550d5fec66aa74caad1f79b79826ec64898688\\n>> changed pg_get_constraintdef() to use an MVCC snapshot rather than a\\n>> current MVCC snapshot; if that change is safe, I am not aware of any\\n>> reason why we couldn't change pg_get_triggerdef() similarly.\\n>\\n> pg_get_triggerdef() is fine as-is with concurrent CREATE TRIGGER. The\\n> pg_get_constraintdef() change arose to ensure a consistent result when\\n> concurrent ALTER TABLE VALIDATE CONSTRAINT mutates a constraint definition.\\n> (Reducing the lock level of DROP TRIGGER or ALTER TRIGGER, however, would\\n> create the analogous problem for pg_get_triggerdef().)\\n\\nMaybe so, but I'd favor changing it anyway and getting it over with.\\nThe current situation seems to have little to recommend it; moreover,\\nit would be nice, if it's possible and safe, to weaken the lock levels\\nfor all three of those commands at the same time. Do you see any\\nhazards for ALTER or DROP that do not exist for CREATE?\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 27 Oct 2014 08:24:15 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"On Sun, Oct 26, 2014 at 9:48 PM, Andreas Karlsson <andreas@proxel.se> wrote:\\n> Agreed.. But I think reducing the lock level of the secondary table is much\\n> more important than doing the same for the primary table due to the case\\n> where the secondary table is an existing table which is hit by a workload of\\n> long running queries and DML while the primary is a new table which is added\\n> now. In my dream world I could add the new table without any disruption at\\n> all of queries using the secondary table, no matter the duration of the\\n> transaction adding the table (barring insertion of actual data into the\\n> primary table, which would take row locks).\\n\\nThat would indeed be nice, but it doesn't seem very practical, because\\nthe parent needs triggers, too: if you try to delete a row from the\\nparent, or update the key, it's got to go look at the child and see\\nwhether there are rows against the old value. Then it's got to either\\nupdate those rows, or null out the value, or throw an error.\\nRegardless of which of those things it does (which depends on the ON\\nDELETE and ON UPDATE settings you choose), it's hard to imagine that\\nit would be a good idea for any of those things to start happening in\\nthe middle of a transaction or statement.\\n\\nSo, let's take what we can get. :-)\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 27 Oct 2014 08:35:59 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 27, 2014 at 08:24:15AM -0400, Robert Haas wrote:\\n> On Sat, Oct 25, 2014 at 2:00 PM, Noah Misch <noah@leadboat.com> wrote:\\n> >> http://www.postgresql.org/message-id/CA+TgmoY4GLsXZk0tAO29-LJtcuj0SL1xWCwQ51xb-HFYsgi5RQ@mail.gmail.com\\n> >> http://www.postgresql.org/message-id/20893.1393892127@sss.pgh.pa.us\\n> >> http://www.postgresql.org/message-id/20140306224340.GA3551655@tornado.leadboat.com\\n> >>\\n> >> As far as triggers are concerned, the issue of skew between the\\n> >> transaction snapshot and what the ruleutils.c snapshots do seems to be\\n> >> the principal issue. Commit e5550d5fec66aa74caad1f79b79826ec64898688\\n> >> changed pg_get_constraintdef() to use an MVCC snapshot rather than a\\n> >> current MVCC snapshot; if that change is safe, I am not aware of any\\n> >> reason why we couldn't change pg_get_triggerdef() similarly.\\n> >\\n> > pg_get_triggerdef() is fine as-is with concurrent CREATE TRIGGER. The\\n> > pg_get_constraintdef() change arose to ensure a consistent result when\\n> > concurrent ALTER TABLE VALIDATE CONSTRAINT mutates a constraint definition.\\n> > (Reducing the lock level of DROP TRIGGER or ALTER TRIGGER, however, would\\n> > create the analogous problem for pg_get_triggerdef().)\\n> \\n> Maybe so, but I'd favor changing it anyway and getting it over with.\\n> The current situation seems to have little to recommend it; moreover,\\n> it would be nice, if it's possible and safe, to weaken the lock levels\\n> for all three of those commands at the same time. Do you see any\\n> hazards for ALTER or DROP that do not exist for CREATE?\\n\\nALTER TRIGGER is not bad; like you say, change pg_get_triggerdef_worker() the\\nway commit e5550d5 changed pg_get_constraintdef_worker(). DROP TRIGGER is\\nmore difficult. pg_constraint.tgqual of a dropped trigger may reference other\\ndropped objects, which calls for equipping get_rule_expr() to use the\\ntransaction snapshot. That implicates quite a bit of ruleutils.c code.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 27 Oct 2014 20:33:56 -0400\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"On 10/28/2014 01:33 AM, Noah Misch wrote:\\n> ALTER TRIGGER is not bad; like you say, change pg_get_triggerdef_worker() the\\n> way commit e5550d5 changed pg_get_constraintdef_worker(). DROP TRIGGER is\\n> more difficult. pg_constraint.tgqual of a dropped trigger may reference other\\n> dropped objects, which calls for equipping get_rule_expr() to use the\\n> transaction snapshot. That implicates quite a bit of ruleutils.c code.\\n\\nI started looking into this again and fixed \\npg_get_constraintdef_worker() as suggested.\\n\\nBut I have no idea how to fix get_rule_expr() since it relies on doing \\nlookups in the catcache. Replacing these with uncached lookups sounds \\nlike it could cause quite some slowdown. Any ideas?\\n\\nIndexes should suffer from the same problems since they too have emay \\ncontain expressions but they seem to solve this by having a higher lock \\nlevel on DROP INDEX, but I do wonder about the CONCURRENTLY case.\\n\\nBy the way, unless I am mistaken there is currently no protection \\nagainst having a concurrent ALTER FUNCTION ... RENAME mess up what is \\ndumped in by pg_get_triggerdef().\\n\\n-- \\nAndreas Karlsson\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sat, 13 Dec 2014 11:42:13 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nI have attached a patch with the current status of my work on reducing \\nthe lock level of trigger and foreign key related DDL.\\n\\nThis commit reduces the lock level of the following commands from ACCESS \\nEXCLUSIVE to SHARE ROW EXCLUSIVE, plus that it does the same for the \\nreferred table of the foreign key.\\n\\nADD TRIGGER\\nALTER TRIGGER\\nALTER TABLE ... ADD FOREIGN KEY\\nALTER TABLE ... DISABLE TRIGGER\\nALTER TABLE ... ENABLE TRIGGER\\n\\nThe patch does currently not reducing the lock level of the following \\ntwo commands, but I think it would be really nice to fix those too and \\nhere I would like some feedback and ideas.\\n\\nDROP TRIGGER\\nALTER TABLE ... DROP CONSTRAINT -- For foreign keys\\n\\nForeign keys and triggers are fixed at the same time because foreign \\nkeys are implemented with two triggers, one at each of the involved tables.\\n\\nThe idea behind the patch is that since we started using MVCC snapshots \\nwe no longer need to hold an exclusive lock on the table when adding a \\ntrigger. Theoretically we only need to lock out writers of the rows of \\nthe table (SHARE ROW EXCLUSIVE/SHARE), but as noted by Robert and Noah \\njust reducing the lock level will break pg_dump since \\npg_get_triggerdef() does not use the current snapshot when reading the \\ncatalog.\\n\\nI have fixed pg_get_triggerdef() to use the snapshot like how \\ne5550d5fec66aa74caad1f79b79826ec64898688 fixed pg_get_constraintdef(), \\nand this fixes the code for the ALTER TRIGGER case (ADD TRIGGER was \\nalready safe). But to fix it for the DROP TRIGGER case we need to also \\nmake the dumping of the WHEN clause (which is dumped by get_rule_expr()) \\nuse the current snapshot.\\n\\nget_rule_expr() relies heavily on the catcache which to me does not look \\nlike it could easily be (and probably not even should be) made to use \\nthe current snapshot. Refactoring ruleutils.c to rely less no the \\ncatcache seems like a reasonable thing to do if we want to reduce \\nweirdness of how it ignores MVCC but it is quite a bit of work and I \\nfear it could give us performance regressions.\\n\\nDo you have any ideas for how to fix get_rule_expr()? Is this patch \\nworthwhile even without reducing the lock levels of the drop commands?\\n\\nAndreas\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Sat, 13 Dec 2014 21:45:54 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"PATCH: Reducing lock strength of trigger and foreign key DDL\"\n },\n {\n \"msg_contents\": \"I'm not sure about the rest of this but...\\n\\nOn Sat, Dec 13, 2014 at 3:45 PM, Andreas Karlsson <andreas@proxel.se> wrote:\\n> Is this patch\\n> worthwhile even without reducing the lock levels of the drop commands?\\n\\nYes! It certainly makes more sense to reduce the lock levels where we\\ncan do that relatively easily, and postpone work on related projects\\nthat are harder, rather than waiting until it all seems to work before\\ndoing anything at all.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 15 Dec 2014 11:24:24 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign key DDL\"\n },\n {\n \"msg_contents\": \"On 25 October 2014 19:00, Noah Misch <noah@leadboat.com> wrote:\\n\\n>> So I tentatively propose (and with due regard for the possibility\\n>> others may see dangers that I've missed) that a reasonable goal would\\n>> be to lower the lock strength required for both CREATE TRIGGER and ADD\\n>> FOREIGN KEY from AccessExclusiveLock to ShareRowExclusiveLock,\\n>> allowing concurrent SELECT and SELECT FOR SHARE against the tables,\\n>> but not any actual write operations.\\n>\\n> +1\\n\\nAll of this is just a replay of the earlier conversations about\\nreducing lock levels.\\n\\nMy original patch did reduce lock levels for CREATE TRIGGER, but we\\nstepped back from doing that in 9.4 until we had feedback as to\\nwhether the whole idea of reducing lock levels was safe, which Robert,\\nTom and others were slightly uncertain about.\\n\\nIs there anything different here from work in my original patch series?\\n\\n-- \\n Simon Riggs http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 17 Dec 2014 14:41:39 +0000\",\n \"msg_from\": \"Simon Riggs <simon@2ndQuadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"On Wed, Dec 17, 2014 at 02:41:39PM +0000, Simon Riggs wrote:\\n> Is there anything different here from work in my original patch series?\\n\\nNot to my knowledge.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 19 Dec 2014 21:26:58 -0500\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"On 12/17/2014 03:41 PM, Simon Riggs wrote:\\n> All of this is just a replay of the earlier conversations about\\n> reducing lock levels.\\n>\\n> My original patch did reduce lock levels for CREATE TRIGGER, but we\\n> stepped back from doing that in 9.4 until we had feedback as to\\n> whether the whole idea of reducing lock levels was safe, which Robert,\\n> Tom and others were slightly uncertain about.\\n>\\n> Is there anything different here from work in my original patch series?\\n\\nAs far as I know, no.\\n\\n-- \\nAndreas Karlsson\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 23 Dec 2014 19:59:29 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Reducing lock strength of adding foreign keys\"\n },\n {\n \"msg_contents\": \"On Sun, Dec 14, 2014 at 5:45 AM, Andreas Karlsson <andreas@proxel.se> wrote:\\n> get_rule_expr() relies heavily on the catcache which to me does not look\\n> like it could easily be (and probably not even should be) made to use the\\n> current snapshot. Refactoring ruleutils.c to rely less no the catcache seems\\n> like a reasonable thing to do if we want to reduce weirdness of how it\\n> ignores MVCC but it is quite a bit of work and I fear it could give us\\n> performance regressions.\\n> Do you have any ideas for how to fix get_rule_expr()?\\nAgreed. There are 20 calls to SearchSysCache in ruleutils.c, let's not\\nfocus on that for now though and get things right for this patch.\\n\\n> Is this patch worthwhile even without reducing the lock levels of the drop commands?\\nYes. IMV, it is better to do this work slowly and incrementally.\\n\\nHere are some comments about this patch:\\n1) There is no documentation. Could you update mvcc.sgml for SHARE ROW\\nEXCLUSIVE?\\n2) Some isolation tests would be welcome, the point of the feature is\\nto test if SELECT and SELECT FOR SHARE/UPDATE are allowed while\\nrunning one of the command mentioned above.\\n3) This patch breaks the isolation test alter-table-1.\\nRegards,\\n--\\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Wed, 24 Dec 2014 17:27:51 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"Thanks for the review.\\n\\nHere is a new version of the patch with updated isolation tests and \\ndocumentation.\\n\\n-- \\nAndreas Karlsson\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Tue, 13 Jan 2015 09:21:47 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 13, 2015 at 5:21 PM, Andreas Karlsson <andreas@proxel.se> wrote:\\n> Here is a new version of the patch with updated isolation tests and\\n> documentation.\\nThanks.\\n\\nSome more comments:\\n- It would be good to have as well tests for ENABLE/DISABLE TRIGGER,\\nthe locks of those commands being reduced as well.\\n- Patch has some typos, like \\\"an share\\\".\\n- Documentation can be simplified. Listing the flavors of ALTER TABLE\\ntaking ROW SHARE EXCLUSIVE in mvcc/sgml is clumsy IMO, it is better to\\nmention that in alter_table.sgml and have a link pointing to the page\\nof ALTER TABLE in mvcc.sgml.\\n- In ATAddForeignKeyConstraint, I think that it would be better to not\\nremove the comment, adding instead that ShareRowExclusiveLock is\\nactually safe because CREATE TRIGGER uses this level of lock.\\nAll those things gathered give the patch attached. Andreas, if you are\\nfine with it I think that we could pass it to a committer.\\nRegards,\\n-- \\nMichael\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Wed, 14 Jan 2015 16:48:00 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> I think that we could pass it to a committer.\\nMarked as such.\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Thu, 15 Jan 2015 16:00:55 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 01/14/2015 08:48 AM, Michael Paquier wrote:\\n> All those things gathered give the patch attached. Andreas, if you are\\n> fine with it I think that we could pass it to a committer.\\n\\nExcellent changes. Thanks for the patch and the reviews.\\n\\n-- \\nAndreas Karlsson\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 16 Jan 2015 00:36:55 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\n> \\t/*\\n> -\\t * Grab an exclusive lock on the pk table, so that someone doesn't delete\\n> -\\t * rows out from under us. (Although a lesser lock would do for that\\n> -\\t * purpose, we'll need exclusive lock anyway to add triggers to the pk\\n> -\\t * table; trying to start with a lesser lock will just create a risk of\\n> -\\t * deadlock.)\\n> +\\t * Grab ShareRowExclusiveLock on the pk table, so that someone doesn't\\n> +\\t * delete rows out from under us. Note that this does not create risks\\n> +\\t * of deadlocks as triggers add added to the pk table using the same\\n> +\\t * lock.\\n> \\t */\\n\\n\\\"add added\\\" doesn't look intended. The rest of the sentence doesn't look\\nentirely right either.\\n\\n> \\t/*\\n> \\t * Triggers must be on tables or views, and there are additional\\n> @@ -526,8 +526,7 @@ CreateTrigger(CreateTrigStmt *stmt, const char *queryString,\\n> \\t * can skip this for internally generated triggers, since the name\\n> \\t * modification above should be sufficient.\\n> \\t *\\n> -\\t * NOTE that this is cool only because we have AccessExclusiveLock on the\\n> -\\t * relation, so the trigger set won't be changing underneath us.\\n> +\\t * NOTE that this is cool only because of the unique contraint.\\n\\nI fail to see what the unique constraint has to do with this? The\\nprevious comment refers to the fact that the AccessExclusiveLock is what\\nprevents a race where another transaction adds a trigger with the same\\nname already exists. Yes, the unique index would, as noted earlier in\\nthe comment, catch the error. But that's not the point of the\\ncheck. Unless I miss something the comment is just as true if you\\nreplace the access exclusive with share row exlusive as it's also self\\nconflicting.\\n\\n> @@ -1272,8 +1271,7 @@ renametrig(RenameStmt *stmt)\\n> \\t * on tgrelid/tgname would complain anyway) and to ensure a trigger does\\n> \\t * exist with oldname.\\n> \\t *\\n> -\\t * NOTE that this is cool only because we have AccessExclusiveLock on the\\n> -\\t * relation, so the trigger set won't be changing underneath us.\\n> +\\t * NOTE that this is cool only because there is a unique constraint.\\n> \\t */\\n\\nSame as above.\\n\\n> \\ttgrel = heap_open(TriggerRelationId, RowExclusiveLock);\\n> \\n> diff --git a/src/backend/utils/adt/ruleutils.c b/src/backend/utils/adt/ruleutils.c\\n> index dd748ac..8eeccf2 100644\\n> --- a/src/backend/utils/adt/ruleutils.c\\n> +++ b/src/backend/utils/adt/ruleutils.c\\n> @@ -699,7 +699,8 @@ pg_get_triggerdef_worker(Oid trigid, bool pretty)\\n> \\tHeapTuple\\tht_trig;\\n> \\tForm_pg_trigger trigrec;\\n> \\tStringInfoData buf;\\n> -\\tRelation\\ttgrel;\\n> +\\tSnapshot\\tsnapshot = RegisterSnapshot(GetTransactionSnapshot());\\n> +\\tRelation\\ttgrel = heap_open(TriggerRelationId, AccessShareLock);\\n> \\tScanKeyData skey[1];\\n> \\tSysScanDesc tgscan;\\n> \\tint\\t\\t\\tfindx = 0;\\n> @@ -710,18 +711,18 @@ pg_get_triggerdef_worker(Oid trigid, bool pretty)\\n> \\t/*\\n> \\t * Fetch the pg_trigger tuple by the Oid of the trigger\\n> \\t */\\n> -\\ttgrel = heap_open(TriggerRelationId, AccessShareLock);\\n> -\\n> \\tScanKeyInit(&skey[0],\\n> \\t\\t\\t\\tObjectIdAttributeNumber,\\n> \\t\\t\\t\\tBTEqualStrategyNumber, F_OIDEQ,\\n> \\t\\t\\t\\tObjectIdGetDatum(trigid));\\n> \\n> \\ttgscan = systable_beginscan(tgrel, TriggerOidIndexId, true,\\n> -\\t\\t\\t\\t\\t\\t\\t\\tNULL, 1, skey);\\n> +\\t\\t\\t\\t\\t\\t\\t\\tsnapshot, 1, skey);\\n> \\n> \\tht_trig = systable_getnext(tgscan);\\n> \\n> +\\tUnregisterSnapshot(snapshot);\\n> +\\n> \\tif (!HeapTupleIsValid(ht_trig))\\n> \\t\\telog(ERROR, \\\"could not find tuple for trigger %u\\\", trigid);\\n>\\n\\nHm. Pushing the snapshot is supposed to make this fully mvcc? Idon't\\nthink that's actually sufficient because of the deparsing of the WHEN\\nclause and of the function name.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 16 Jan 2015 15:01:06 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 01/16/2015 03:01 PM, Andres Freund wrote:\\n> Hi,\\n>\\n>> \\t/*\\n>> -\\t * Grab an exclusive lock on the pk table, so that someone doesn't delete\\n>> -\\t * rows out from under us. (Although a lesser lock would do for that\\n>> -\\t * purpose, we'll need exclusive lock anyway to add triggers to the pk\\n>> -\\t * table; trying to start with a lesser lock will just create a risk of\\n>> -\\t * deadlock.)\\n>> +\\t * Grab ShareRowExclusiveLock on the pk table, so that someone doesn't\\n>> +\\t * delete rows out from under us. Note that this does not create risks\\n>> +\\t * of deadlocks as triggers add added to the pk table using the same\\n>> +\\t * lock.\\n>> \\t */\\n>\\n> \\\"add added\\\" doesn't look intended. The rest of the sentence doesn't look\\n> entirely right either.\\n\\nIt was intended to be \\\"are added\\\", but the sentence is pretty awful \\nanyway. I am not sure the sentence is really necessary anyway.\\n\\n>> \\t/*\\n>> \\t * Triggers must be on tables or views, and there are additional\\n>> @@ -526,8 +526,7 @@ CreateTrigger(CreateTrigStmt *stmt, const char *queryString,\\n>> \\t * can skip this for internally generated triggers, since the name\\n>> \\t * modification above should be sufficient.\\n>> \\t *\\n>> -\\t * NOTE that this is cool only because we have AccessExclusiveLock on the\\n>> -\\t * relation, so the trigger set won't be changing underneath us.\\n>> +\\t * NOTE that this is cool only because of the unique contraint.\\n>\\n> I fail to see what the unique constraint has to do with this? The\\n> previous comment refers to the fact that the AccessExclusiveLock is what\\n> prevents a race where another transaction adds a trigger with the same\\n> name already exists. Yes, the unique index would, as noted earlier in\\n> the comment, catch the error. But that's not the point of the\\n> check. Unless I miss something the comment is just as true if you\\n> replace the access exclusive with share row exlusive as it's also self\\n> conflicting.\\n\\nYeah, this must have been a remainder from the version where I only \\ngrabbed a ShareLock. The comment should be restored.\\n\\n> Hm. Pushing the snapshot is supposed to make this fully mvcc? Idon't\\n> think that's actually sufficient because of the deparsing of the WHEN\\n> clause and of the function name.\\n\\nIndeed. As Noah and I discussed previously in this thread we would need \\nto do quite a bit of refactoring of ruleutils.c to make it fully MVCC. \\nFor this reason I opted to only lower the lock levels of ADD and ALTER \\nTRIGGER, and not DROP TRIGGER. Neither of those require MVCC of then \\nWHEN clause.\\n\\nAndreas\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 16 Jan 2015 15:16:20 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 2015-01-16 15:16:20 +0100, Andreas Karlsson wrote:\\n> Indeed. As Noah and I discussed previously in this thread we would need to\\n> do quite a bit of refactoring of ruleutils.c to make it fully MVCC.\\n\\nRight.\\n\\n> For this reason I opted to only lower the lock levels of ADD and ALTER\\n> TRIGGER, and not DROP TRIGGER. Neither of those require MVCC of then\\n> WHEN clause.\\n\\nI'm unconvinced that this is true. Using a snapshot for part of getting\\na definition certainly opens the door for getting strange\\nresults.\\n\\nAcquiring a lock that prevents schema changes on the table and then\\ngetting the definition using the syscaches guarantees that that\\ndefinition is at least self consistent because no further schema changes\\nare possible and the catalog caches will be up2date.\\n\\nWhat you're doing though is doing part of the scan using the\\ntransaction's snapshot (as used by pg_dump that will usually be a\\nrepeatable read snapshot and possibly quite old) and the other using a\\nfresh catalog snapshot. This can result in rather odd things.\\n\\nJust consider:\\nS1: CREATE TABLE flubber(id serial primary key, data text);\\nS1: CREATE FUNCTION blarg() RETURNS TRIGGER LANGUAGE plpgsql AS $$BEGIN RETURN NEW; END;$$;\\nS1: CREATE TRIGGER flubber_blarg BEFORE INSERT ON flubber FOR EACH ROW WHEN (NEW.data IS NOT NULL) EXECUTE PROCEDURE blarg();\\nS2: BEGIN TRANSACTION ISOLATION LEVEL SERIALIZABLE;\\nS2: SELECT 'dosomethingelse';\\nS1: ALTER TABLE flubber RENAME TO wasflubber;\\nS1: ALTER TABLE wasflubber RENAME COLUMN data TO wasdata;\\nS1: ALTER TRIGGER flubber_blarg ON wasflubber RENAME TO wasflubber_blarg;\\nS1: ALTER FUNCTION blarg() RENAME TO wasblarg;\\nS2: SELECT pg_get_triggerdef(oid) FROM pg_trigger;\\n\\nThis will give you: The old trigger name. The new table name. The new\\ncolumn name. The new function name.\\n\\nI don't think using a snapshot for tiny parts of these functions\\nactually buys anything. Now, this isn't a pattern you introduced. But I\\nthink we should think about this for a second before expanding it\\nfurther.\\n\\nBefore you argue that this isn't relevant for pg_dump: It is. Precisely\\nthe above can happen - just replace the 'dosomethingelse' with several\\nLOCK TABLEs as pg_dump does. The first blocks after a snapshot has been\\nacquired. While waiting, all the ALTERs happen.\\n\\nArguably the benefit here is that the window for this issue is becoming\\nsmaller as pg_dump (and hopefully other possible callers) acquire\\nexclusive locks on the table. I.e. that the lowering of the lock level\\ndoesn't introduce new races. But on the other side of the coin, this\\nmakes the result of pg_get_triggerdef() even *more* inaccurate in many\\ncases.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n-- \\n Andres Freund\\t http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 16 Jan 2015 16:59:56 +0100\",\n \"msg_from\": \"Andres Freund <andres@2ndquadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 16, 2015 at 10:59 AM, Andres Freund <andres@2ndquadrant.com> wrote:\\n> Just consider:\\n> S1: CREATE TABLE flubber(id serial primary key, data text);\\n> S1: CREATE FUNCTION blarg() RETURNS TRIGGER LANGUAGE plpgsql AS $$BEGIN RETURN NEW; END;$$;\\n> S1: CREATE TRIGGER flubber_blarg BEFORE INSERT ON flubber FOR EACH ROW WHEN (NEW.data IS NOT NULL) EXECUTE PROCEDURE blarg();\\n> S2: BEGIN TRANSACTION ISOLATION LEVEL SERIALIZABLE;\\n> S2: SELECT 'dosomethingelse';\\n> S1: ALTER TABLE flubber RENAME TO wasflubber;\\n> S1: ALTER TABLE wasflubber RENAME COLUMN data TO wasdata;\\n> S1: ALTER TRIGGER flubber_blarg ON wasflubber RENAME TO wasflubber_blarg;\\n> S1: ALTER FUNCTION blarg() RENAME TO wasblarg;\\n> S2: SELECT pg_get_triggerdef(oid) FROM pg_trigger;\\n>\\n> This will give you: The old trigger name. The new table name. The new\\n> column name. The new function name.\\n\\nOuch. That's clearly no good. I'm struggling to understand whether\\nthis is a problem with our previous analysis, or a problem with this\\npatch:\\n\\nhttp://www.postgresql.org/message-id/20141028003356.GA387814@tornado.leadboat.com\\n\\npg_get_triggerdef_worker() relies on generate_function_name(), which\\nuses the system caches, and on get_rule_expr(), for deparsing the WHEN\\nclause. If we allowed only ADDING triggers with a lesser lock and\\nnever modifying or dropping them with a lesser lock, then changing the\\ninitial scan of pg_trigger at the top of pg_get_triggerdef_worker() to\\nuse the transaction snapshot might be OK; if we can see the trigger\\nwith the transaction snapshot at all, we know it can't have\\nsubsequently changed. But allowing alterations of any kind isn't\\ngoing to work, so I think our previous analysis on that point was\\nincorrect.\\n\\nI *think* we could fix that if generate_function_name() and\\nget_rule_expr() had an option to use the active snapshot instead of a\\nfresh snapshot. The former doesn't look too hard to arrange, but the\\nlatter is a tougher nut to crack.\\n\\n-- \\nRobert Haas\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 19 Jan 2015 12:14:31 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 01/19/2015 06:14 PM, Robert Haas wrote:\\n> On Fri, Jan 16, 2015 at 10:59 AM, Andres Freund <andres@2ndquadrant.com> wrote:\\n>> This will give you: The old trigger name. The new table name. The new\\n>> column name. The new function name.\\n>\\n> Ouch. That's clearly no good. I'm struggling to understand whether\\n> this is a problem with our previous analysis, or a problem with this\\n> patch:\\n>\\n> http://www.postgresql.org/message-id/20141028003356.GA387814@tornado.leadboat.com\\n>\\n> pg_get_triggerdef_worker() relies on generate_function_name(), which\\n> uses the system caches, and on get_rule_expr(), for deparsing the WHEN\\n> clause. If we allowed only ADDING triggers with a lesser lock and\\n> never modifying or dropping them with a lesser lock, then changing the\\n> initial scan of pg_trigger at the top of pg_get_triggerdef_worker() to\\n> use the transaction snapshot might be OK; if we can see the trigger\\n> with the transaction snapshot at all, we know it can't have\\n> subsequently changed. But allowing alterations of any kind isn't\\n> going to work, so I think our previous analysis on that point was\\n> incorrect.\\n>\\n> I *think* we could fix that if generate_function_name() and\\n> get_rule_expr() had an option to use the active snapshot instead of a\\n> fresh snapshot. The former doesn't look too hard to arrange, but the\\n> latter is a tougher nut to crack.\\n\\nThe function name in the trigger should already be broken in master \\nsince changing the name of a function does not lock the table. Would be \\nreally neat to fix though.\\n\\nA possible way forward I see here is to undo my changes to the lock \\nlevel of ALTER TRIGGER and of pg_get_triggerdef_worker(). This should \\nwork if Noah is correct about his analysis in where he came to the \\nconclusion: \\\"pg_get_triggerdef() is fine as-is with concurrent CREATE \\nTRIGGER.\\\". Is this still true with the new information?\\n\\n-- \\nAndreas Karlsson\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 20 Jan 2015 00:06:04 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 16, 2015 at 04:59:56PM +0100, Andres Freund wrote:\\n> On 2015-01-16 15:16:20 +0100, Andreas Karlsson wrote:\\n> > For this reason I opted to only lower the lock levels of ADD and ALTER\\n> > TRIGGER, and not DROP TRIGGER. Neither of those require MVCC of then\\n> > WHEN clause.\\n> \\n> I'm unconvinced that this is true. Using a snapshot for part of getting\\n> a definition certainly opens the door for getting strange\\n> results.\\n> \\n> Acquiring a lock that prevents schema changes on the table and then\\n> getting the definition using the syscaches guarantees that that\\n> definition is at least self consistent because no further schema changes\\n> are possible and the catalog caches will be up2date.\\n> \\n> What you're doing though is doing part of the scan using the\\n> transaction's snapshot (as used by pg_dump that will usually be a\\n> repeatable read snapshot and possibly quite old) and the other using a\\n> fresh catalog snapshot. This can result in rather odd things.\\n> \\n> Just consider:\\n> S1: CREATE TABLE flubber(id serial primary key, data text);\\n> S1: CREATE FUNCTION blarg() RETURNS TRIGGER LANGUAGE plpgsql AS $$BEGIN RETURN NEW; END;$$;\\n> S1: CREATE TRIGGER flubber_blarg BEFORE INSERT ON flubber FOR EACH ROW WHEN (NEW.data IS NOT NULL) EXECUTE PROCEDURE blarg();\\n> S2: BEGIN TRANSACTION ISOLATION LEVEL SERIALIZABLE;\\n> S2: SELECT 'dosomethingelse';\\n> S1: ALTER TABLE flubber RENAME TO wasflubber;\\n> S1: ALTER TABLE wasflubber RENAME COLUMN data TO wasdata;\\n> S1: ALTER TRIGGER flubber_blarg ON wasflubber RENAME TO wasflubber_blarg;\\n> S1: ALTER FUNCTION blarg() RENAME TO wasblarg;\\n> S2: SELECT pg_get_triggerdef(oid) FROM pg_trigger;\\n> \\n> This will give you: The old trigger name. The new table name. The new\\n> column name. The new function name.\\n> \\n> I don't think using a snapshot for tiny parts of these functions\\n> actually buys anything. Now, this isn't a pattern you introduced. But I\\n> think we should think about this for a second before expanding it\\n> further.\\n\\nFair enough. It did reinforce pg_get_constraintdef() as a subroutine of\\npg_dump rather than an independent, rigorous interface. It perhaps made the\\nfunction worse for non-pg_dump callers. In that vein, each one of these hacks\\nhas a cost. One could make a reasonable argument that ALTER TRIGGER RENAME\\nlocking is not important enough to justify spreading the hack from\\npg_get_constraintdef() to pg_get_triggerdef(). Lowering the CREATE TRIGGER\\nlock level does not require any ruleutils.c change for the benefit of pg_dump,\\nbecause pg_dump won't see the pg_trigger row of a too-recent trigger.\\n\\n> Before you argue that this isn't relevant for pg_dump: It is. Precisely\\n> the above can happen - just replace the 'dosomethingelse' with several\\n> LOCK TABLEs as pg_dump does. The first blocks after a snapshot has been\\n> acquired. While waiting, all the ALTERs happen.\\n\\nWe wish pg_dump would take a snapshot of the database; that is, we wish its\\noutput always matched some serial execution of transactions. pg_dump has,\\nsince ancient times, failed to achieve that if non-table DDL commits during\\nthe dump or if table DDL commits between acquiring the dump transaction\\nsnapshot and acquiring the last table lock. My reviews have defended the\\nstandard that table DDL issued after pg_dump has acquired locks does not\\nchange the dump. That's what we bought with pg_get_constraintdef()'s use of\\nthe transaction snapshot and would buy with the same in pg_get_triggerdef().\\nMy reviews have deliberately ignored effects on scenarios where pg_dump\\nalready fails to guarantee snapshot-like output.\\n\\n> Arguably the benefit here is that the window for this issue is becoming\\n> smaller as pg_dump (and hopefully other possible callers) acquire\\n> exclusive locks on the table. I.e. that the lowering of the lock level\\n> doesn't introduce new races. But on the other side of the coin, this\\n> makes the result of pg_get_triggerdef() even *more* inaccurate in many\\n> cases.\\n\\nWhat is this about pg_dump acquiring exclusive locks?\\n\\nTo summarize, the problem you raise has been out of scope, because it affects\\npg_dump only at times when pg_dump is already wrong.\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Tue, 20 Jan 2015 04:08:25 -0500\",\n \"msg_from\": \"Noah Misch <noah@leadboat.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 01/20/2015 10:08 AM, Noah Misch wrote:\\n> Fair enough. It did reinforce pg_get_constraintdef() as a subroutine of\\n> pg_dump rather than an independent, rigorous interface. It perhaps made the\\n> function worse for non-pg_dump callers. In that vein, each one of these hacks\\n> has a cost. One could make a reasonable argument that ALTER TRIGGER RENAME\\n> locking is not important enough to justify spreading the hack from\\n> pg_get_constraintdef() to pg_get_triggerdef(). Lowering the CREATE TRIGGER\\n> lock level does not require any ruleutils.c change for the benefit of pg_dump,\\n> because pg_dump won't see the pg_trigger row of a too-recent trigger.\\n\\nI agree with this view, and am not sure myself that it is worth lowering \\nthe lock level of ALTER TRIGGER RENAME. I have attached a patch without \\nthe changes to ALTER TRIGGER and ruleutils.c and also fixes the comment \\nissues noted by Andres.\\n\\n-- \\nAndreas Karlsson\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Thu, 22 Jan 2015 22:31:44 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 01/22/2015 10:31 PM, Andreas Karlsson wrote:\\n> I agree with this view, and am not sure myself that it is worth lowering\\n> the lock level of ALTER TRIGGER RENAME. I have attached a patch without\\n> the changes to ALTER TRIGGER and ruleutils.c and also fixes the comment\\n> issues noted by Andres.\\n\\nWhops, forgot to include the isolation tests.\\n\\n-- \\nAndreas Karlsson\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Fri, 23 Jan 2015 00:55:21 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 23, 2015 at 8:55 AM, Andreas Karlsson <andreas@proxel.se> wrote:\\n> On 01/22/2015 10:31 PM, Andreas Karlsson wrote:\\n>>\\n>> I agree with this view, and am not sure myself that it is worth lowering\\n>> the lock level of ALTER TRIGGER RENAME. I have attached a patch without\\n>> the changes to ALTER TRIGGER and ruleutils.c and also fixes the comment\\n>> issues noted by Andres.\\n>\\n>\\n> Whops, forgot to include the isolation tests.\\nOk, so the deal is to finally reduce the locks to\\nShareRowExclusiveLock for the following commands :\\n- CREATE TRIGGER\\n- ALTER TABLE ENABLE/DISABLE\\n- ALTER TABLE ADD CONSTRAINT\\n\\nLooking at the latest patch, it seems that in\\nAlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\\nAT_AddConstraintRecurse and AT_ProcessedConstraint under the same\\nbanner as AT_AddConstraint. Thoughts?\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 30 Jan 2015 15:48:25 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 01/30/2015 07:48 AM, Michael Paquier wrote:\\n> Ok, so the deal is to finally reduce the locks to\\n> ShareRowExclusiveLock for the following commands :\\n> - CREATE TRIGGER\\n> - ALTER TABLE ENABLE/DISABLE\\n> - ALTER TABLE ADD CONSTRAINT\\n\\nCorrect. I personally still find this useful enough to justify a patch.\\n\\n> Looking at the latest patch, it seems that in\\n> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\\n> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\\n> banner as AT_AddConstraint. Thoughts?\\n\\nGood point. I think moving those would be a good thing even though it is \\ntechnically not necessary for AT_AddConstraintRecurse, since that one \\nshould only be related to check constraints.\\n\\n-- \\nAndreas\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 30 Jan 2015 19:26:28 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 30, 2015 at 10:26 PM, Andreas Karlsson wrote:\\n> On 01/30/2015 07:48 AM, Michael Paquier wrote:\\n>> Looking at the latest patch, it seems that in\\n>> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\\n>> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\\n>> banner as AT_AddConstraint. Thoughts?\\n>\\n> Good point. I think moving those would be a good thing even though it is\\n> technically not necessary for AT_AddConstraintRecurse, since that one should\\n> only be related to check constraints.\\n\\nAndreas, are you planning to send an updated patch?\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 6 Feb 2015 11:16:33 +0400\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 02/06/2015 08:16 AM, Michael Paquier wrote:\\n> On Fri, Jan 30, 2015 at 10:26 PM, Andreas Karlsson wrote:\\n>> On 01/30/2015 07:48 AM, Michael Paquier wrote:\\n>>> Looking at the latest patch, it seems that in\\n>>> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\\n>>> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\\n>>> banner as AT_AddConstraint. Thoughts?\\n>>\\n>> Good point. I think moving those would be a good thing even though it is\\n>> technically not necessary for AT_AddConstraintRecurse, since that one should\\n>> only be related to check constraints.\\n>\\n> Andreas, are you planning to send an updated patch?\\n\\nYes, I will hopefully send it later today or tomorrow.\\n\\nAndreas\\n\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 06 Feb 2015 10:38:21 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 01/30/2015 07:48 AM, Michael Paquier wrote:\\n> Looking at the latest patch, it seems that in\\n> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\\n> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\\n> banner as AT_AddConstraint. Thoughts?\\n\\nA new version of the patch is attached which treats them as the same for \\nlocking. I think it is correct and improves readability to do so.\\n\\n-- \\nAndreas Karlsson\\n\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\",\n \"msg_date\": \"Sun, 08 Feb 2015 02:05:46 +0100\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 8, 2015 at 10:05 AM, Andreas Karlsson <andreas@proxel.se> wrote:\\n> On 01/30/2015 07:48 AM, Michael Paquier wrote:\\n>>\\n>> Looking at the latest patch, it seems that in\\n>> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\\n>> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\\n>> banner as AT_AddConstraint. Thoughts?\\n>\\n>\\n> A new version of the patch is attached which treats them as the same for\\n> locking. I think it is correct and improves readability to do so.\\n\\nWell then, let's switch it to \\\"Ready for committer\\\". I am moving as\\nwell this entry to the next CF with the same status.\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Fri, 13 Feb 2015 10:10:44 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 7 February 2015 at 20:05, Andreas Karlsson <andreas@proxel.se> wrote:\\n> On 01/30/2015 07:48 AM, Michael Paquier wrote:\\n>>\\n>> Looking at the latest patch, it seems that in\\n>> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\\n>> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\\n>> banner as AT_AddConstraint. Thoughts?\\n>\\n>\\n> A new version of the patch is attached which treats them as the same for\\n> locking. I think it is correct and improves readability to do so.\\n\\nCommitted. We move forwards, slowly but surely. Thanks for the patch.\\n\\n-- \\n Simon Riggs http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, RemoteDBA, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 5 Apr 2015 11:56:42 -0400\",\n \"msg_from\": \"Simon Riggs <simon@2ndQuadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On Mon, Apr 6, 2015 at 12:56 AM, Simon Riggs <simon@2ndquadrant.com> wrote:\\n> On 7 February 2015 at 20:05, Andreas Karlsson <andreas@proxel.se> wrote:\\n>> On 01/30/2015 07:48 AM, Michael Paquier wrote:\\n>>>\\n>>> Looking at the latest patch, it seems that in\\n>>> AlterTableGetLockLevel@tablecmds.c we ought to put AT_ReAddConstraint,\\n>>> AT_AddConstraintRecurse and AT_ProcessedConstraint under the same\\n>>> banner as AT_AddConstraint. Thoughts?\\n>>\\n>>\\n>> A new version of the patch is attached which treats them as the same for\\n>> locking. I think it is correct and improves readability to do so.\\n>\\n> Committed. We move forwards, slowly but surely. Thanks for the patch.\\n\\nCool! Thanks for showing up.\\n-- \\nMichael\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 6 Apr 2015 08:19:37 +0900\",\n \"msg_from\": \"Michael Paquier <michael.paquier@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 04/05/2015 05:56 PM, Simon Riggs wrote:\\n> Committed. We move forwards, slowly but surely. Thanks for the patch.\\n\\nThanks to you and the reviewers for helping me out with this patch!\\n\\n-- \\nAndreas Karlsson\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Mon, 06 Apr 2015 01:36:49 +0200\",\n \"msg_from\": \"Andreas Karlsson <andreas@proxel.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n },\n {\n \"msg_contents\": \"On 5 April 2015 at 19:19, Michael Paquier <michael.paquier@gmail.com> wrote:\\n\\n> Cool! Thanks for showing up.\\n\\nVisibility <> Activity. How is REINDEX CONCURRENTLY doing?\\n\\n-- \\n Simon Riggs http://www.2ndQuadrant.com/\\n PostgreSQL Development, 24x7 Support, RemoteDBA, Training & Services\\n\\n\\n-- \\nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\\nTo make changes to your subscription:\\nhttp://www.postgresql.org/mailpref/pgsql-hackers\\n\",\n \"msg_date\": \"Sun, 5 Apr 2015 19:37:46 -0400\",\n \"msg_from\": \"Simon Riggs <simon@2ndQuadrant.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: Reducing lock strength of trigger and foreign\\n key DDL\"\n }\n]","isConversational":false}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":396,"numItemsPerPage":100,"numTotalItems":58371,"offset":39600,"length":100}},"jwt":"eyJhbGciOiJFZERTQSIsImtpZCI6IjVHZDBvd0g5MTM2eDZjc1FvbE1zcktNWUZoRVFoUm5PVVVybEpjOUhaUEEifQ.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc4NTYyMzU4NywianRpIjoiOWZiODYzYzctOTRkNi00MDFhLWIxNDQtOGM5M2I3MmRlMTEwIiwic3ViIjoiL2RhdGFzZXRzL3dhbnNoZW5sL3Bnc3FsLWhhY2tlcnMiLCJleHAiOjE3ODU2MjcxODcsImlzcyI6Imh0dHBzOi8vaHVnZ2luZ2ZhY2UuY28ifQ.Wh3sU_w_xJDb1mxUonQrKGdYpANmKES3au6sHIDpwMpn4w-5g7mri4heoV2Df_qztz9URN-SvFHUK-EOu3dMDQ","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\nWe already have IF NOT EXISTS for CREATE TABLE. There are some reason to\ndon't have to CREATE TABLE AS and CREATE MATERIALIZED VIEW??\n\nRegards,\n\n-- \nFabrízio de Royes Mello\nConsultoria/Coaching PostgreSQL\n>> Timbira: http://www.timbira.com.br\n>> Blog: http://fabriziomello.... |
[
{
"msg_contents": "Hi,\n\nI have a patch which is actually not commitfest-ready now, but it\nalways better to start discussing proof of concept having some patch\ninstead of just an idea.\n\nSince I'am a DBA rather than C programmer, I will appreciate any\nsuggestions/critics about the patch and code quality to... |
[
{
"msg_contents": "Hi\n\nSQL:2003 introduced the function NEXT VALUE FOR <sequence>. Google\ntells me that at least DB2, SQL Server and a few niche databases\nunderstand it so far. As far as I can tell there is no standardised\nequivalent of currval and setval (but I only have access to second\nhand informatio... |
[
{
"msg_contents": "Having been surprisingly successful at advancing our understanding of\narguments for and against various approaches to \"value locking\", I\ndecided to try the same thing out elsewhere. I have created a\ngeneral-purpose UPSERT wiki page.\n\nThe page is: https://wiki.postgresql.org/wiki/UPSERT... |
[
{
"msg_contents": "Here is the comments in process_matched_tle() in rewriteHandler.c.\n\n883 * such nodes; consider\n884 * UPDATE tab SET col.fld1.subfld1 = x, col.fld2.subfld2 = y\n885 * The two expressions produced by the parser will look like\n886 * FieldStore(col, fld1, FieldSt... |
[
{
"msg_contents": "On Thu, Oct 2, 2014 at 5:25 AM, Craig Ringer <craig@2ndquadrant.com> wrote:\n> It's not at all clear to me that a DTrace-like (or perf-based, rather)\n> approach is unsafe, slow, or unsuitable for production use.\n\n> With appropriate wrapper tools I think we could have quite a useful\n> libr... |
[
{
"msg_contents": "Hi\n\nThere are few less readable examples of dynamic SQL in plpgsql doc\n\nlike:\n\nEXECUTE 'SELECT count(*) FROM '\n || tabname::regclass\n || ' WHERE inserted_by = $1 AND inserted <= $2'\n INTO c\n USING checked_user, checked_date;\n\nor\n\nEXECUTE 'UPDATE tbl SET '\n || q... |
[
{
"msg_contents": "Hi,\n\nwe have seen repeatedly that users can be confused about why PostgreSQL\nis not shutting down even though they requested it. Usually, this is\nbecause `log_checkpoints' is not enabled and the final checkpoint is\nbeing written, delaying shutdown. As no message besides \"shutting down\... |
[
{
"msg_contents": "I got the following error when I try to build my extension\ntowards the latest master branch.\n\nIs the \"port/atomics/*.h\" files forgotten on make install?\n\n[kaigai@magro pg_strom]$ make\ngcc -O2 -Wall -Wmissing-prototypes -Wpointer-arith\n-Wdeclaration-after-statement -Wendif-labels\n-Wm... |
[
{
"msg_contents": "Hi fellow hackers,\n\nI would like to work on a new feature allowing our users to assess the\namount of trouble they will run into when running a DDL script on their\nproduction setups, *before* actually getting their services down.\n\nThe main practical example I can offer here is the ALTER ... |
[
{
"msg_contents": "make check-world dies ingloriously for me, like this:\n\n/bin/sh ../../../config/install-sh -c -d tmp_check/log\nmake -C ../../..\nDESTDIR='/Users/rhaas/pgsql/src/bin/initdb'/tmp_check/install install\n>'/Users/rhaas/pgsql/src/bin/initdb'/tmp_check/log/install.log 2>&1\ncd . && TESTDIR='/User... |
[
{
"msg_contents": "Hi list,\n\nAttached patch allows semijoin/antijoin/hashed SubPlan optimization\nwhen an EXISTS subquery contains a LIMIT clause with a positive\nconstant. It seems to be a fairly common meme to put LIMIT 1 into\nEXISTS() subqueries, and it even makes sense when you're not aware\nthat the dat... |
[
{
"msg_contents": "I did a fair dive into double-checking the decision to just leave \nxloginsert_locks fixed at 8 for 9.4. My conclusion: good call, move \nalong. Further improvements beyond what the 8-way split gives sure are \npossible. But my guess from chasing them a little is that additional \nplaces ... |
[
{
"msg_contents": "This is from Bug #11555, which is still in moderation as I type this\n(analysis was done via IRC).\n\nThe GiST insertion code appears to have no length checks at all on the\ninserted entry. index_form_tuple checks for length <= 8191, with the\ndefault blocksize, but obviously a tuple less tha... |
[
{
"msg_contents": "Hello,\n\nI recently got a trouble on development of my extension that utilizes\nthe shared buffer when it released each buffer page.\n\nThis extension transfers contents of the shared buffers to GPU device\nusing DMA feature, then kicks a device kernel code.\nUsually 8KB (= BLCKSZ) is too sm... |
[
{
"msg_contents": "Hi Hackers,\n\nI've updated the wiki page\nhttps://wiki.postgresql.org/wiki/Incremental_backup following the result\nof discussion on hackers.\n\nCompared to first version, we switched from a timestamp+checksum based\napproach to one based on LSN.\n\nThis patch adds an option to pg_basebackup... |
[
{
"msg_contents": "As discussed recently, here is an undocumented patch for \njson_strip_nulls and jsonb_strip_nulls.\n\ncheers\n\nandrew\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers",
... |
[
{
"msg_contents": "Hello\n\nI am playing with RLS. I created simple table\n\ntable_data (inserted_by text, v integer);\n\nI created two policies\n\ncreate policy p1 on data with check (inserted_by = session_user);\ncreate policy p2 on data with check (v between 10 and 1000);\n\nI was surprised so p2 effectively... |
[
{
"msg_contents": "I got interested in the problem discussed in\nhttp://www.postgresql.org/message-id/20714.1412456604@sss.pgh.pa.us\nto wit:\n\n> It's becoming clear to me that our existing design whereby zone\n> abbreviations represent fixed GMT offsets isn't really good enough.\n> I've been wondering whether... |
[
{
"msg_contents": "Hi all\n\nWhile looking at an unrelated issue in PgJDBC I noticed that it's\ndifficult for users and the driver to tell in advance if a given\nstatement will support bind parameters.\n\nPostgreSQL just treats placeholders as syntax errors for non-plannable\nstatements at parse time.\n\nThis f... |
[
{
"msg_contents": "Hi,\n\nThis morning while running make check-world on my OSX Mavericks laptop, I\nfound the following failure:\ntest pgtypeslib/dt_test2 ... stderr FAILED (test process was\nterminated by signal 6: Abort trap)\n(lldb) bt\n* thread #1: tid = 0x0000, 0x00007fff8052c866\nlibsystem_kernel.dy... |
[
{
"msg_contents": "In addition to the few patches left in Needs Review state, we have six \npatches marked as Ready for Committer.\n\nCommitters: Could you please pick a patch, and commit if appropriate? Or \nif there's a patch there that you think should *not* be committed, \nplease speak up.\n\n- Heikki\n\n\n... |
[
{
"msg_contents": "I noticed that analyze messages shown by autovacuum don't discriminate\nbetween non-inherited cases and inherited cases, as shown in the below\nexample:\n\nLOG: automatic analyze of table \"postgres.public.pt\" system usage: CPU\n0.00s/0.01u sec elapsed 0.06 sec\nLOG: automatic analyze of t... |
[
{
"msg_contents": "Hi all,\n\nLately I have come across two inconveniences/bugs related to running\nthe autocommit-off mode in psql.\n\nThese are:\n- BUG #11524: Unable to add value to ENUM when having AUTOCOMMIT\ndisabled in psql\n- BUG #10822: \"ALTER SYSTEM cannot run inside a transaction block\"\nwhen havin... |
[
{
"msg_contents": "Hi all,\n\npg_receivexlog always handles argument of -d option as connstr formatted value.\nWe can doubly specify host name, port number.\nThe other client tools handles -d option as connstr value only if\nargument has \"=\" character.\nThe document says that pg_receivexlog ignores database ... |
[
{
"msg_contents": "In DoCopy, some RLS-specific code constructs a SelectStmt to handle\nthe case where COPY TO is invoked on an RLS-protected relation. But I\nthink this step is bogus in two ways:\n\n /* Build FROM clause */\n from = makeRangeVar(NULL, RelationGetRelationName(rel), 1);\n\... |
[
{
"msg_contents": "Hi folks,\n\nMy corp (CSC) OSS division requires CLAs to be signed for OSS project \nparticipation to begin. I need to fix a few problems in PGJDBC driver \nand am unable to start without them. Neither Google nor PG Wiki contain \nCLA licenses and I have no idea where else to look.\nPostgreSQ... |
[
{
"msg_contents": "While reviewing the patch myself, i spotted some formatting problems in the\ncode. Fixed in this v5 patch.\n\nAlso, this patch uses context patch format (in first versions, because of\nthe small modification, context patch format obfucates the changes. After\nreimplementation this isn't the c... |
[
{
"msg_contents": "Hi,\n\nIt appears that I'm not the only person who finds it somewhat\nunintuitive for overlapping RLS policies to be permissive rather than\nrestrictive (OR vs AND) (at least 3 others seem to expect AND\nbehaviour), although I understand the reasoning behind\nit. And I've since discovered th... |
[
{
"msg_contents": "Hi,\n\nThe code in there doesn't look very interesting - and very unlikely to\nrun these days. Notably it relies on a binary called 'postmaster' et\nal...\n\nThe last realy changes are from a long time ago:\n\ncommit 142d42f9386ed81a4f0779ec8a0cad1254173b5e\nAuthor: Vadim B. Mikheev <vadim4o@... |
[
{
"msg_contents": "While looking at bug #11431, I noticed that pg_upgrade still seems to \nthink that encoding and locale are cluster-wide properties. We got \nper-database locale support in 8.4, and encoding has been per-database \nmuch longer than that.\n\npg_upgrade checks the encoding and locale of template... |
[
{
"msg_contents": "The RLS patch added OCLASS_ROWSECURITY but it seems that not enough\neffort was made to grep for places that might require adjustment as a\nresult.\n\nIn objectaddress.c, getObjectDescription() was updated, but\ngetObjectTypeDescription() and getObjectIdentity() were not.\n\nIn dependency.c, ... |
[
{
"msg_contents": "Some time ago, when debugging a GIN bug, I wrote these pageinspect \nfunctions to inspect GIN indexes. They were very useful; we should add them.\n\n- Heikki\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postg... |
[
{
"msg_contents": "Hello.\n\nI am interested in the json type on postgresql. I would like to implement\nadditional operations on the json structure that may extract/insert table\nlike information from the json tree structure.\nI have a implementation on javascript that shows this type of operations.\nYou can se... |
[
{
"msg_contents": "Our CRC algorithm is a bit weird. It's supposedly CRC-32, with the same \npolynomial as used in Ethernet et al, but it actually is not. The \ncomments refer to \"Painless Guide to CRC Error Detection Algorithms\" by \nRoss N. Williams [1] (http://www.ross.net/crc/download/crc_v3.txt), but \nI... |
[
{
"msg_contents": "Hi there!\n\nI'm trying to implement a custom indexing system using the GiST index \nframework, and I have a few questions.\nBasically, I'm trying to implement a system that allows me to search \nacross a set of 64 bit integers by hamming distance. This is intended to \nbe used in searching f... |
[
{
"msg_contents": "Hello!\n\nAutovacuum daemon performs vacuum when the number of rows \nupdated/deleted (n_dead_tuples) reaches some threshold.\nSimilarly it performs analyze when the number of rows changed in any way \n(incl. inserted).\nWhen a table is mostly insert-only, its visibility map is not updated as... |
[
{
"msg_contents": "Hello,\n\nOne user reported a hang problem with 9.4 beta2 on Windows. The PostgreSQL \nis 64-bit version. I couldn't find the cause, but want to solve the \nproblem. Could you help with this?\n\nI heard that the user had run 16 concurrent psql sessions which executes \nINSERT and UPDATE st... |
[
{
"msg_contents": "\nSalut!\n\nFellow volunteers, I request assistance in understanding the following:\n\nWhen I explain a query I can get the following information:\n\n\n | \"I/O Read Time\": 0.000,\n | \"I/O Write Time\": 0.000\n\nI know why it is 0. My question is this, can ... |
[
{
"msg_contents": "Summary: pg_restore -n attempts to restore objects to pg_catalog schema\nVersions Tested: 9.3.5, 9.3.0, 9.2.4\nSeverity: Failure\nDescription:\n\nThe -n option (or --schema) for pg_restore is supposed to allow you to\nrestore a single schema from a custom-format pg_dump file. Instead, it\nat... |
[
{
"msg_contents": "Having just git pulled from orgin/master:\n\n$ ./configure\n$ make\n\nMileage:\n\n...\nmake -C bootstrap all\nmake[3]: Entering directory `/path/to/postgresql/src/backend/bootstrap'\ngcc -O2 -Wall -Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement -Wendif-labels -Wmissing-form... |
[
{
"msg_contents": "I found a reference made obsolete by commit 9e85183b, which is from\nway back in 2000.\n\ncomparetup_index_btree() says:\n\n/*\n* This is similar to _bt_tuplecompare(), but we have already done the\n* index_getattr calls for the first column, and we need to keep track of\n* whether any null f... |
[
{
"msg_contents": "Hello, on the way considering alter role set .., I found that\nALTER ROLE/USER cannot take CURRENT_USER as the role name.\n\nIn addition to that, documents of ALTER ROLE / USER are\ninconsistent with each other in spite of ALTER USER is said to be\nan alias for ALTER ROLE. Plus, ALTER USER ca... |
[
{
"msg_contents": "Postgres currently supports column level SELECT privileges.\n\n1. If we want to confirm a credit card number, we can issue SELECT 1\nFROM customer WHERE stored_card_number = '1234 5678 5344 7733'\n\n2. If we want to look for card fraud, we need to be able to use the\nfull card number to join ... |
[
{
"msg_contents": "Hi all,\n\nCurrently all the row-level lock modes are described in the page for\nSELECT query:\nhttp://www.postgresql.org/docs/devel/static/explicit-locking.html#LOCKING-ROWS\nHowever, after browsing the documentation, I noticed in the page\ndescribing all the explicit locks of the system tha... |
[
{
"msg_contents": "Hi,\n\nI've been testing out some of the new materialized view functionality\nin 9.4 and noticed that they don't show up in the information_schema\ndata dictionary views, specifically information_schema.tables or\ninformation_schema.views (sample output is below and it's the same for\n9.3 as ... |
[
{
"msg_contents": "Both Robert and Heikki expressed dissatisfaction with the fact that\nB-Tree index builds don't use sortsupport. Because B-Tree index builds\ncannot really use the \"onlyKey\" optimization, the historic oversight\nof not supporting B-Tree builds (and CLUSTER) might have been at least\na little... |
[
{
"msg_contents": "Started a new thread to raise awareness.\n\nMichael Paquier wrote:\n> On Thu, Oct 9, 2014 at 8:29 AM, Michael Paquier\n> <michael.paquier@gmail.com> wrote:\n> > On Thu, Oct 9, 2014 at 6:26 AM, Alvaro Herrera <alvherre@2ndquadrant.com>\n> > wrote:\n> >> However, if I were to do it, I would ins... |
[
{
"msg_contents": "Greetings,\n\nWhile looking for easier items in PostgreSQL Wiki's Todo List (this will be\nmy 3rd patch), i found this TODO:\n\nAdd a built-in array_agg(anyarray) or similar, that can aggregate\n> 1-dimensional arrays into a 2-dimensional array.\n>\n\nI've stumbled by this lack of array_agg(a... |
[
{
"msg_contents": "Hi all,\n\nAttached WIP patch adds new syntax REINEX SCHEMA which does reindexing\nall table of specified schema.\nThere are syntax dose reindexing specified index, per table and per database,\nbut we can not do reindexing per schema for now.\nSo we must use reindexdb command if we want to do... |
[
{
"msg_contents": "Hi,\n\nattached is a WIP patch implementing multivariate statistics. The code\ncertainly is not \"ready\" - parts of it look as if written by a rogue\nchimp who got bored of attempts to type the complete works of William\nShakespeare, and decided to try something different.\n\nI also cut some... |
[
{
"msg_contents": "Hi all!\n\nI've been working on implementing a custom index using the Index Access Method API and have the need for custom reloptions that are \"complex\" strings (ie, also contain non-alphaumerics).\n\npg_get_indexdef() and pg_dump don't quote the reloption values, making a restore (or cut-n... |
[
{
"msg_contents": "The August commitfest is still Open, with a few more patches left. The \npatches that remain have stayed in limbo for a long time. It's not \nrealistic to expect anything to happen to them.\n\nI'm going to move the remaining patches to the next commitfest, and \nclose the August one. I hate t... |
[
{
"msg_contents": "CacheInvalidateHeapTuple currently does the following tests first; would there be a performance improvement to testing the system relation case first? We're almost never in bootstrap mode, so that test is almost always a waste. Is there any reason not to switch the two?\n\n\t/* Do nothing dur... |
[
{
"msg_contents": "I found a minor typo in bgworker.sgml. Patch attached. It should be\napplied to 9.4 too.\n\n-- \nShigeru HANADA\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscription:\nhttp://www.postgresql.org/mailpref/pgsql-hackers",
"msg_... |
[
{
"msg_contents": "A few years ago I started working on a concurrent hash table for\nPostgreSQL. The hash table part of it worked, but I never did\nanything with it, really. Amit mentioned to me earlier this week that\nhe was seeing contention inside the dynahash machinery, which inspired\nme to go back and u... |
[
{
"msg_contents": "Hi all,\n\nI found that the statistics are still remained after it's truncated.\nIn addition, the analyzing ignores table does not have any tuple.\nAfter table truncated, the entry of statistics continues to remain\nunless insertion and analyzing are executed.\nThere is reason why statistics ... |
[
{
"msg_contents": "Hello,\n\nI changed the buffer manager code in order to generate a trace of page\nrequests from the buffer manager perspective. In summary, whenever\nReleaseBuffer() or ReleaseAndReadBuffer() are called, I print the page\ncurrently being released which is identified by the tuple (tableSpace,\... |
[
{
"msg_contents": "Hi Hackers,\n\nfollowing the advices gathered on the list I've prepared a third partial\npatch on the way of implementing incremental pg_basebackup as described\nhere https://wiki.postgresql.org/wiki/Incremental_backup\n\n\n== Changes\n\nCompared to the previous version I've made the followin... |
[
{
"msg_contents": "On 14 October 2014 13:57, Stephen Frost <sfrost@snowman.net> wrote:\n\n> Create an 'audit' role.\n>\n> Every command run by roles which are granted to the 'audit' role are\n> audited.\n>\n> Every 'select' against tables which the 'audit' role has 'select' rights\n> on are audited. Similairly... |
[
{
"msg_contents": "Hi fellow hackers,\n\nPlease find attached to this email a patch to implement a new Event\nTrigger, fired on the the \"table_rewrite\" event. As attached, it's meant\nas a discussion enabler and only supports ALTER TABLE (and maybe not in\nall forms of it). It will need to grow support for VA... |
[
{
"msg_contents": "Hi list,\n\nI happened to notice that there are no less than 14 places in the code\nthat check whether a relation is a system catalog and throwing the\nerror \"permission denied: \"foo\" is a system catalog\"\n\nThe attached patch factors all of those into a single\nForbidSystemTableMods() fu... |
[
{
"msg_contents": "For parallelism, I think we need a concept of group locking. That is,\nsuppose we have a user backend and N worker backends collaborating to\nexecute some query. For the sake of argument, let's say it's a\nparallel CLUSTER, requiring a full table lock. We need all of the\nbackends to be ab... |
[
{
"msg_contents": "Greetings,\n\n The attached patch for review implements a few additional role\n attributes (all requested by users or clients in various forums) for\n common operations which currently require superuser privileges. This\n is not a complete solution for all of the superuser-only privilege... |
[
{
"msg_contents": "Hi All,\n\nPlease find attached a patch which implements support for UPDATE table1\nSET(*)=...\nThe patch supports both UPDATE table SET(*)=(a,b,c) and UPDATE table1\nSET(*)=(SELECT a,b,c FROM...).\nIt solves the problem of doing UPDATE from a record variable of the same\ntype as the table e.... |
[
{
"msg_contents": "I have observed that for renaming some of the objects\nAccessExclusiveLock is taken on object whereas for\nother kind of objects no lock is taken on object before\nrenaming the object.\n\nThe object's that are renamed via AlterObjectRename_internal()\ntakes the lock (during get_object_address... |
[
{
"msg_contents": "Hackers,\n\nPostgres was initially designed to support access methods extendability.\nThis extendability lives to present day. However, this is mostly internal\nin-core extendability. One can quite easily add new access method into\nPostgreSQL core. But if one try to implement access method a... |
[
{
"msg_contents": "All,\n\nThe attached patch for review implements a directory permission system that\nallows for providing a directory read/write capability to directories for\nCOPY TO/FROM and Generic File Access Functions to non-superusers. This is\nnot a complete solution as it does not currently contain ... |
[
{
"msg_contents": "Hi,\n\nWhile I was checking the behavior of RLS, I found that the policy for SELECT\ndoesn't seem to be applied to RETURNING. Is this intentional? Please see\nthe following example.\n\nCREATE ROLE foo LOGIN NOSUPERUSER;\nCREATE TABLE hoge AS SELECT col FROM generate_series(1,10) col;\nALTER T... |
[
{
"msg_contents": "Hi All,\n\npg_dump binary-upgrade fail with segmentation fault for type without\nelement.\n\nConsider the following testcase:\n\nrushabh@postgresql$ ./db/bin/psql postgres\npsql (9.5devel)\nType \"help\" for help.\n\npostgres=# drop type typ;\nDROP TYPE\npostgres=# create type typ as ();\nCRE... |
[
{
"msg_contents": "Hi,\n\nPostgreSQL has for ages defined INT64_FORMAT and UINT64_FORMAT in \npg_config.h. This commit\n\nhttp://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=ce486056ecd28050\n\nmoved those definitions to c.h, which breaks compilation of all released \nSlony-I versions against current ... |
[
{
"msg_contents": "One of the queries in ri_triggers.c has be a little baffled.\n\nFor (relatively) obvious reasons, a FK insert triggers a SELECT 1 FROM\npk_rel ... FOR KEY SHARE.\nFor not-so-obvious reasons, a PK delete triggers a SELECT 1 FROM fk_rel ...\nFOR KEY SHARE.\n\nI can't see what the lock on fk_rel... |
[
{
"msg_contents": "I just tripped over that CommitFest mace that went missing after\nHeikki put it down from the last CF. Unless someone else wants to\npick it up, I'll manage this one.\n\nJust to get the ball rolling, I note that 19 patches were moved to\nthis CF from the last one. These are in the following... |
[
{
"msg_contents": "Over in the \"Log notice that checkpoint is to be written on shutdown\" thread...\n\nOn 10/16/14, 2:31 PM, Michael Banck wrote:\n > There were some comments that this might not actually be the case and/or\n > that the postmaster was simply waiting for clients to disconnect due to\n > smart sh... |
[
{
"msg_contents": "A little while back some users started complaining that the contrib module\nI develop (MADlib) was failing to build with the following error:\n\n------\n/usr/include/postgresql/9.2/server/port.h:480:32: error: declaration of\n'char* mkdtemp(char*)' has a different exception specifier\n/usr/in... |
[
{
"msg_contents": "The attached patches add an ssl_protocols configuration option which\ncontrol which versions of SSL or TLS the server will use. The syntax is\nsimilar to Apache's SSLProtocols directive, except that the list is\ncolon-separated instead of whitespace-separated, although that is easy\nto chang... |
[
{
"msg_contents": "Hi everyone,\n\nVitesse DB 9.3.5.S is Postgres 9.3.5 with a LLVM-JIT query executor\ndesigned for compute intensive OLAP workload. We have gotten it to a\nreasonable state and would like to open it up to the pg hackers\ncommunity for testing and suggestions.\n\nVitesse DB offers\n-- JIT Compi... |
[
{
"msg_contents": "Hi,\n\nConsider the following queries.\n\ncreate table t(i int, j int, k int, t text);\ninsert into t select i, i % 100, i % 1000, 'AABBCCDD' || i from\ngenerate_series(1, 1000000) i;\n\nftian=# explain select count(distinct j) from t group by t, i;\n QUERY PLAN\... |
[
{
"msg_contents": "Now that we have the create hash index warning in 9.5, I realized that\nwe don't warn about hash indexes with PITR, only crash recovery and\nstreaming. This patch fixes that.\n\nIs the wording \"cannot be used\" too vague. The CREATE INDEX manual\npage has the words \"give wrong answers to ... |
[
{
"msg_contents": "Following up something Pavel wrote, I notice that json_agg() and \njson_object_agg() are both marked as immutable, even though they invoke \nIO functions, while json_object is marked stable, even though it does \nnot, and can probably be marked as immutable. Mea maxima culpa.\n\nI'm not sure ... |
[
{
"msg_contents": "Hi,\n\nThis week we had one of the most annoying problems I've ever encountered \nwith postgres. We had a big index on multiple columns, say, foo(a, b, \nc). According to pg_stat_all_indexes the index was being used *all the \ntime*. However, after looking into our queries more closely, i... |
[
{
"msg_contents": "I looked into bug #11703,\nhttp://www.postgresql.org/message-id/20141017162236.2523.74453@wrigleys.postgresql.org\n\nWhat is happening here is that the FieldSelect optimization in\neval_const_expressions fails to play nice with the logic that determines\nwhich Vars need to be available from a... |
[
{
"msg_contents": "Hi,\n\nI've noticed an interesting phenomenon. Look at the following pgbench -c\n32 -j 32 -P 1 output on a scale 200 database. Then guess where I've\nstarted\nSET vacuum_cost_delay = '20ms';\nANALYZE pgbench_accounts;\nand when it finished. The cost_delay is to simulate a autovac triggered\na... |
[
{
"msg_contents": "Hi,\n\nOne of our sites ran into a problem with database installation:\n\ninitdb failed\nFATAL: unexpected data beyond EOF in block 19 of relation base/1/2609\nHINT: This has been seen to occur with buggy kernels; consider updating your system.\nSTATEMENT: COMMENT ON FUNCTION euc_jis_2004_to_... |
[
{
"msg_contents": "Buildfarm member hamerkop has been failing in the pg_upgrade regression\ntest for the last several days. The problem looks like this:\n\ncommand: \"C:/buildfarm/build_root/HEAD/pgsql.build/contrib/pg_upgrade/tmp_check/install/bin/pg_restore\" --port 50432 --username \"Administrator\" --exit-... |
[
{
"msg_contents": "Hi,\n\nCommit 32984d8fc3dbb90a3fafb69fece0134f1ea790f9 forgot to change the \nfilename in the comment in contrib/pgcrypto/pgcrypto--1.2.sql. Trivial \npatch attached.\n\n\n.marko\n\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo make changes to your subscript... |
[
{
"msg_contents": "Currently, a non-freeze vacuum will punt on any page it can't get a cleanup lock on, with no retry. Presumably this should be a rare occurrence, but I think it's bad that we just assume that and won't warn the user if something bad is going on.\n\nMy thought is that if we skip any pages elog(... |
[
{
"msg_contents": "I ran into a type \" a a \" in a comment in snapmgr.c, and found that\nthere are four other places that've made the same typo, by the grep\ncommand. And in one of those places, I found yet another typo. Please\nfind attached a patch.\n\nThanks,\n\nBest regards,\nEtsuro Fujita\n\n\n\n-- \nSe... |
[
{
"msg_contents": "hello,I am reading the pool manager code,and have some problem:\n\nin poolmgr.c function agent_init():\n\nif (agent->pool)\n> agent_release_connections(agent, false);\n\n......\n\n#ifdef XCP\n> /* find database */\n> agent->pool = find_database_pool(database, user_name);\n> /* create if not f... |
[
{
"msg_contents": "Hi all,\n\nExcept if I am missing something, is there any reason why the sequence used\nin XLogArchiveCheckDone and XLogArchiveIsBusy to check if a XLOG segment\nhas been already archived is duplicated? I guess that doing a little bit of\nrefactoring here would make sense for simplicity, patc... |
[
{
"msg_contents": "Dear All,\n L've got a question for you.\nIf the value of the integer type is converted to a value of type Boolean，PostgreSQL will display just like\"The rule has already exist\";\n1. CREATE CAST (integer AS bool) WITH INOUT AS IMPLICIT; \n\nERROR: cast from type in... |
[
{
"msg_contents": "My Salesforce colleague Thomas Fanghaenel observed that the TAP tests\nfor pg_basebackup fail when run in a sufficiently deeply-nested directory\ntree. The cause appears to be that we rely on standard \"tar\" format\nto represent the symlink for a tablespace, and POSIX tar format has a\nhard... |
[
{
"msg_contents": "Hi.\n\nOne of our \"production issues\" is that the system generates lots of\nwal-files, lots is like 151952 files over the last 24h, which is about\n2.4TB worth of WAL files. I wouldn't say that isn't an issue by itself,\nbut the system does indeed work fine. We do subsequently gzip the file... |
[
{
"msg_contents": "Hackers,\n\nIn Mac OS X 10.10 “Yosemite,” Apple removed SystemStarter, upon which our OS X start script has relied since 2007. So here is a patch that adds support for its replacement, launchd. It includes 7 day log rotation like the old script did. The install script still prefers the System... |
[
{
"msg_contents": "I'm trying to create what amounts to a new type. This would be rather easy if I could perform a CHECK on a composite type, which I could do if I could create a domain on top of a composite. Is there any reason in particular that hasn't been done?\n\nAs an alternative, I tried accomplishing th... |
[
{
"msg_contents": "If you do any Postgres development on OS X, you've probably gotten\nseriously annoyed by the way that, every single time you reinstall the\npostmaster executable, you get a dialog box asking whether you'd like\nto allow it to accept incoming network connections. (At least, you\ndo unless you... |
[
{
"msg_contents": "Hi,\n\n\n\nI am looking to patch the LDAP Service Name feature and would like some feedback prior to doing so. In general, while I personally view this as a bug fix, it could also easily be considered to simply be an enhancement to current functionality and therefore the below is written as a... |
[
{
"msg_contents": "expected/sequence_1.out hasn't been updated at least since\n\ncommit d90ced8bb22194cbb45f58beb0961251103aeff5\nDate: Thu Oct 3 16:17:18 2013 -0400\n\nand nobody appears to have complained.\n\nCan it be removed?\n\n\n-- \nSent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)\nTo... |
[
{
"msg_contents": "In heap_prune_chain():\n\n\ttup.t_tableOid = RelationGetRelid(relation);\n\n\trootlp = PageGetItemId(dp, rootoffnum);\n\n\t/*\n\t * If it's a heap-only tuple, then it is not the start of a HOT chain.\n\t */\n\tif (ItemIdIsNormal(rootlp))\n\t{\n\t\thtup = (HeapTupleHeader) PageGetItem(dp, root... |
[
{
"msg_contents": "Hi,\n\nI have been thinking about why we need to grab an AccessExclusiveLock on \nthe table with the PK when we add a foreign key. Adding new tables with \nforeign keys to old ones is common so it would be really nice if the \nlock strength could be reduced.\n\nA comment in the code claims th... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.